i i

i i

Color Imaging

i i

i i
 i i

i i

Color Imaging
Fundamentals and Applications

Erik Reinhard
Erum Arif Khan
Ahmet Oğuz Akyüz
Garrett Johnson

A K Peters, Ltd.
Wellesley, Massachusetts

i i

i i
 i i

i i

Editorial, Sales, and Customer Service Office

A K Peters, Ltd.
888 Worcester Street, Suite 230
Wellesley, MA 02482
www.akpeters.com

Copyright 2008 by A K Peters, Ltd.

All rights reserved. No part of the material protected by this copyright notice may
be reproduced or utilized in any form, electronic or mechanical, including photo-
copying, recording, or by any information storage and retrieval system, without
written permission from the copyright owner.

Library of Congress Cataloging-in-Publication Data

Reinhard, Erik, 1968–

Color imaging : fundamentals and applications / Erik Reinhard . . . [et al.].
p. cm.
Includes bibliographical references and index.
ISBN: 978-1-56881-344-8 (alk. paper)
1. Computer vision. 2. Image processing. 3. Color display systems. 4. Color
separation. I. Title.

TA1634.R45 2007
621.36’7--dc22
2007015704

Printed in India
12 11 10 09 08 10 9 8 7 6 5 4 3 2 1

i i

i i
 i i

i i

Contents

Preface xiii

I Principles 1

1 Introduction 3
1.1 Color in Nature . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Color in Society . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3 In this Book . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 14

2 Physics of Light 17
2.1 Electromagnetic Theory . . . . . . . . . . . . . . . . . . . . 18
2.2 Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3 Polarization . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.4 Spectral Irradiance . . . . . . . . . . . . . . . . . . . . . . 45
2.5 Reflection and Refraction . . . . . . . . . . . . . . . . . . . 47
2.6 Birefringence . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.7 Interference and Diffraction . . . . . . . . . . . . . . . . . . 66
2.8 Scattering . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
2.9 Geometrical Optics . . . . . . . . . . . . . . . . . . . . . . 84
2.10 Application: Image Synthesis . . . . . . . . . . . . . . . . . 96
2.11 Application: Modeling the Atmosphere . . . . . . . . . . . 104
2.12 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
2.13 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 120

3 Chemistry of Matter 121

3.1 Classical Physics . . . . . . . . . . . . . . . . . . . . . . . 122
3.2 Quantum Mechanics . . . . . . . . . . . . . . . . . . . . . 124
3.3 Atoms and Ions . . . . . . . . . . . . . . . . . . . . . . . . 138
v

i i

i i
 i i

i i

vi Contents

3.4 Molecules . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

3.5 Sources of Radiation . . . . . . . . . . . . . . . . . . . . . 159
3.6 Polarization in Dielectric Materials . . . . . . . . . . . . . . 182
3.7 Dichroism . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
3.8 Application: Modeling of Fire and Flames . . . . . . . . . . 191
3.9 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 197

4 Human Vision 199

4.1 Osteology of the Skull . . . . . . . . . . . . . . . . . . . . 200
4.2 Anatomy of the Eye . . . . . . . . . . . . . . . . . . . . . . 201
4.3 The Retina . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
4.4 The Lateral Geniculate Nucleus . . . . . . . . . . . . . . . . 228
4.5 The Visual Cortex . . . . . . . . . . . . . . . . . . . . . . . 230
4.6 A Multi-Stage Color Model . . . . . . . . . . . . . . . . . . 237
4.7 Alternative Theory of Color Vision . . . . . . . . . . . . . . 245
4.8 Application: Modeling a Human Retina . . . . . . . . . . . 247
4.9 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 250

5 Perception 251
5.1 Lightness, Brightness, and Related Definitions . . . . . . . . 252
5.2 Reflectance and Illumination . . . . . . . . . . . . . . . . . 254
5.3 Models of Color Processing . . . . . . . . . . . . . . . . . . 256
5.4 Visual Illusions . . . . . . . . . . . . . . . . . . . . . . . . 259
5.5 Adaptation and Sensitivity . . . . . . . . . . . . . . . . . . 270
5.6 Visual Acuity . . . . . . . . . . . . . . . . . . . . . . . . . 279
5.7 Simultaneous Contrast . . . . . . . . . . . . . . . . . . . . 282
5.8 Lightness Constancy . . . . . . . . . . . . . . . . . . . . . 286
5.9 Color Constancy . . . . . . . . . . . . . . . . . . . . . . . . 295
5.10 Category-Based Processing . . . . . . . . . . . . . . . . . . 298
5.11 Color Anomalies . . . . . . . . . . . . . . . . . . . . . . . 302
5.12 Application: Shadow Removal from Images . . . . . . . . . 309
5.13 Application: Graphical Design . . . . . . . . . . . . . . . . 312
5.14 Application: Telling Humans and Computers Apart . . . . . 314
5.15 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 314

II Color Models 317

6 Radiometry and Photometry 319

6.1 The Sensitivity of the Human Eye . . . . . . . . . . . . . . 320
6.2 Radiometric and Photometric Quantities . . . . . . . . . . . 322
6.3 The Efficacy of Optical Radiation . . . . . . . . . . . . . . 337
6.4 Luminance, Brightness, and Contrast . . . . . . . . . . . . . 340

i i

i i
 i i

i i

Contents vii

6.5 Optical Detectors . . . . . . . . . . . . . . . . . . . . . . . 342

6.6 Light Standards . . . . . . . . . . . . . . . . . . . . . . . . 345
6.7 Detector Standards . . . . . . . . . . . . . . . . . . . . . . 346
6.8 Measurement of Optical Radiation . . . . . . . . . . . . . . 347
6.9 Visual Photometry . . . . . . . . . . . . . . . . . . . . . . . 356
6.10 Application: Measuring Materials . . . . . . . . . . . . . . 359
6.11 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 362

7 Colorimetry 363
7.1 Grassmann’s Laws . . . . . . . . . . . . . . . . . . . . . . 364
7.2 Visual Color Matching . . . . . . . . . . . . . . . . . . . . 366
7.3 Color-Matching Functions . . . . . . . . . . . . . . . . . . 373
7.4 CIE 1931 and 1964 Standard Observers . . . . . . . . . . . 375
7.5 Calculating Tristimulus Values and Chromaticities . . . . . . 378
7.6 Practical Applications of Colorimetry . . . . . . . . . . . . 387
7.7 Application: Iso-Luminant Color Maps . . . . . . . . . . . 397
7.8 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 403

8 Color Spaces 405

8.1 RGB Color Spaces . . . . . . . . . . . . . . . . . . . . . . 411
8.2 Printers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 418
8.3 Luminance-Chrominance Color Spaces . . . . . . . . . . . 427
8.4 Television and Video . . . . . . . . . . . . . . . . . . . . . 430
8.5 Hue-Saturation-Lightness Spaces . . . . . . . . . . . . . . . 439
8.6 HVS Derived Color Spaces . . . . . . . . . . . . . . . . . . 444
8.7 Color Opponent Spaces . . . . . . . . . . . . . . . . . . . . 448
8.8 Color Difference Metrics . . . . . . . . . . . . . . . . . . . 459
8.9 Color Order Systems . . . . . . . . . . . . . . . . . . . . . 465
8.10 Application: Color Transfer between Images . . . . . . . . . 467
8.11 Application: Color-to-Gray Conversion . . . . . . . . . . . 474
8.12 Application: Rendering . . . . . . . . . . . . . . . . . . . . 478
8.13 Application: Rendering and Color-Matching Paints . . . . . 480
8.14 Application: Classification of Edges . . . . . . . . . . . . . 484
8.15 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 490

9 Illuminants 491
9.1 CIE Standard Illuminants and Sources . . . . . . . . . . . . 491
9.2 Color Temperature . . . . . . . . . . . . . . . . . . . . . . 503
9.3 Color-Rendering Index . . . . . . . . . . . . . . . . . . . . 508
9.4 CIE Metamerism Index . . . . . . . . . . . . . . . . . . . . 512
9.5 Dominant Wavelength . . . . . . . . . . . . . . . . . . . . . 514
9.6 Excitation Purity . . . . . . . . . . . . . . . . . . . . . . . 517
9.7 Colorimetric Purity . . . . . . . . . . . . . . . . . . . . . . 517

i i

i i
 i i

i i

viii Contents

9.8 Application: Modeling Light-Emitting Diodes . . . . . . . . 518

9.9 Application: Estimating the Illuminant in an Image . . . . . 520
9.10 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 524

10 Chromatic Adaptation 525

10.1 Changes in Illumination . . . . . . . . . . . . . . . . . . . . 526
10.2 Measuring Chromatic Adaptation . . . . . . . . . . . . . . . 530
10.3 Mechanisms of Chromatic Adaptation . . . . . . . . . . . . 532
10.4 Models of Chromatic Adaptation . . . . . . . . . . . . . . . 538
10.5 Application: Transforming sRGB Colors to D50 for an ICC
Workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . 553
10.6 Application: White Balancing a Digital Camera . . . . . . . 555
10.7 Application: Color-Accurate Rendering . . . . . . . . . . . 562
10.8 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 564

11 Color and Image Appearance Models 565

11.1 Vocabulary . . . . . . . . . . . . . . . . . . . . . . . . . . 566
11.2 Color Appearance Phenomena . . . . . . . . . . . . . . . . 582
11.3 Color Appearance Modeling . . . . . . . . . . . . . . . . . 591
11.4 Image Appearance Modeling . . . . . . . . . . . . . . . . . 605
11.5 Applications of Color and Image Appearance Models . . . . 620
11.6 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 629

III Digital Color Imaging 631

12 Image Capture 633

12.1 Optical Image Formation . . . . . . . . . . . . . . . . . . . 635
12.2 Lenses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649
12.3 Aberrations . . . . . . . . . . . . . . . . . . . . . . . . . . 654
12.4 The Diaphragm . . . . . . . . . . . . . . . . . . . . . . . . 667
12.5 The Shutter . . . . . . . . . . . . . . . . . . . . . . . . . . 668
12.6 Filters and Coatings . . . . . . . . . . . . . . . . . . . . . . 669
12.7 Solid-State Sensors . . . . . . . . . . . . . . . . . . . . . . 672
12.8 In-Camera Signal Processing . . . . . . . . . . . . . . . . . 678
12.9 A Camera Model . . . . . . . . . . . . . . . . . . . . . . . 682
12.10 Sensor Noise Characteristics . . . . . . . . . . . . . . . . . 683
12.11 Measuring Camera Noise . . . . . . . . . . . . . . . . . . . 688
12.12 Radiometric Camera Calibration . . . . . . . . . . . . . . . 694
12.13 Light Field Data . . . . . . . . . . . . . . . . . . . . . . . . 697
12.14 Holography . . . . . . . . . . . . . . . . . . . . . . . . . . 701
12.15 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 706

i i

i i
 i i

i i

Contents ix

13 High Dynamic Range Image Capture 709

13.1 Multi-Exposure Techniques . . . . . . . . . . . . . . . . . . 710
13.2 Response Curve Recovery . . . . . . . . . . . . . . . . . . 715
13.3 Noise Removal . . . . . . . . . . . . . . . . . . . . . . . . 722
13.4 Ghost Removal . . . . . . . . . . . . . . . . . . . . . . . . 726
13.5 Image Alignment . . . . . . . . . . . . . . . . . . . . . . . 733
13.6 Single Capture High Dynamic Range Images . . . . . . . . 734
13.7 Direct High Dynamic Range Capture . . . . . . . . . . . . . 737
13.8 Application: Drawing Programs . . . . . . . . . . . . . . . 739
13.9 Application: Image-Based Material Editing . . . . . . . . . 740
13.10 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 741

14 Display Technologies 743

14.1 Cathode-Ray Tubes (CRTs) . . . . . . . . . . . . . . . . . . 743
14.2 Liquid Crystal Displays (LCDs) . . . . . . . . . . . . . . . 746
14.3 Transflective Liquid Crystal Displays . . . . . . . . . . . . . 767
14.4 Plasma Display Panels (PDPs) . . . . . . . . . . . . . . . . 768
14.5 Light-Emitting Diode (LED) Displays . . . . . . . . . . . . 770
14.6 Organic Light-Emitting Diode Displays . . . . . . . . . . . 772
14.7 Field Emission Displays . . . . . . . . . . . . . . . . . . . 775
14.8 Surface-Conduction Electron-Emitter Displays . . . . . . . 776
14.9 Microcavity Plasma Devices . . . . . . . . . . . . . . . . . 777
14.10 Interferometric Modulator (IMOD) Displays . . . . . . . . . 777
14.11 Projection Displays . . . . . . . . . . . . . . . . . . . . . . 779
14.12 Liquid Crystal Display (LCD) Projectors . . . . . . . . . . . 781
14.13 Digital Light Processing (DLP
R
) Projectors . . . . . . . . . 782
14.14 Liquid Crystal on Silicon (LCoS) Projectors . . . . . . . . . 785
14.15 Multi-Primary Display Devices . . . . . . . . . . . . . . . . 787
14.16 High Dynamic Range Display Devices . . . . . . . . . . . . 791
14.17 Electronic Ink . . . . . . . . . . . . . . . . . . . . . . . . . 794
14.18 Display Characterization . . . . . . . . . . . . . . . . . . . 794
14.19 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 803

15 Image Properties and Image Display 805

15.1 Natural Image Statistics . . . . . . . . . . . . . . . . . . . . 806
15.2 Dynamic Range . . . . . . . . . . . . . . . . . . . . . . . . 816
15.3 Cross-Media Display . . . . . . . . . . . . . . . . . . . . . 827
15.4 Gamut Mapping . . . . . . . . . . . . . . . . . . . . . . . . 833
15.5 Gamma Correction . . . . . . . . . . . . . . . . . . . . . . 841
15.6 Ambient Light . . . . . . . . . . . . . . . . . . . . . . . . . 843

i i

i i
 i i

i i

x Contents

16 Color Management 849

16.1 A Generic Color Management System . . . . . . . . . . . . 849
16.2 ICC Color Management . . . . . . . . . . . . . . . . . . . . 851
16.3 Practical Applications . . . . . . . . . . . . . . . . . . . . . 874

17 Dynamic Range Reduction 881

17.1 Spatial Operators . . . . . . . . . . . . . . . . . . . . . . . 885
17.2 Sigmoidal Compression . . . . . . . . . . . . . . . . . . . . 888
17.3 Local Neighborhoods . . . . . . . . . . . . . . . . . . . . . 892
17.4 Sub-Band Systems . . . . . . . . . . . . . . . . . . . . . . 895
17.5 Edge-Preserving Smoothing Operators . . . . . . . . . . . . 897
17.6 Gradient-Domain Operators . . . . . . . . . . . . . . . . . . 899
17.7 Histogram Adjustment . . . . . . . . . . . . . . . . . . . . 900
17.8 Lightness Perception . . . . . . . . . . . . . . . . . . . . . 901
17.9 Counter Shading . . . . . . . . . . . . . . . . . . . . . . . 905
17.10 Post-Processing . . . . . . . . . . . . . . . . . . . . . . . . 906
17.11 Validation and Comparison . . . . . . . . . . . . . . . . . . 910
17.12 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 926

IV Appendices 929

A Vectors and Matrices 931

A.1 Cross and Dot Product . . . . . . . . . . . . . . . . . . . . 931
A.2 Vector Differentiation . . . . . . . . . . . . . . . . . . . . . 933
A.3 Gradient of a Scalar Function . . . . . . . . . . . . . . . . . 933
A.4 Divergence . . . . . . . . . . . . . . . . . . . . . . . . . . 934
A.5 Gauss’ Theorem . . . . . . . . . . . . . . . . . . . . . . . . 934
A.6 Curl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 935
A.7 Stokes’ Theorem . . . . . . . . . . . . . . . . . . . . . . . 936
A.8 Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . . 937
A.9 Vector Identities . . . . . . . . . . . . . . . . . . . . . . . . 937
A.10 Homogeneous Coordinates . . . . . . . . . . . . . . . . . . 937

B Trigonometry 939
B.1 Sum and Difference Formulae . . . . . . . . . . . . . . . . 939
B.2 Product Identities . . . . . . . . . . . . . . . . . . . . . . . 940
B.3 Double-Angle Formulae . . . . . . . . . . . . . . . . . . . 941
B.4 Half-Angle Formulae . . . . . . . . . . . . . . . . . . . . . 941
B.5 Sum Identities . . . . . . . . . . . . . . . . . . . . . . . . . 941
B.6 Solid Angle . . . . . . . . . . . . . . . . . . . . . . . . . . 942

i i

i i
 i i

i i

Contents xi

C Complex Numbers 945

C.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . 945
C.2 Euler’s Formula . . . . . . . . . . . . . . . . . . . . . . . . 946
C.3 Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . 947
C.4 Time-Harmonic Quantities . . . . . . . . . . . . . . . . . . 948

D Units and Constants 949

E The CIE Luminous Efficiency Functions 951

F CIE Illuminants 955

G Chromaticity Coordinates of Paints 959

Bibliography 961

Index 1027

i i

i i
 i i

i i

Preface

Color is one of the most fascinating areas to study. Color forms an integral part
of nature, and we humans are exposed to it every day. We all have an intuitive
understanding of what color is, but by studying the underlying physics, chemistry,
optics, and human visual perception, the true beauty and complexity of color can
be appreciated—at least to some extent. Such understanding is not just important
in these areas of research, but also for fields such as color reproduction, vision
science, atmospheric modeling, image archiving, art, photography, and the like.
Many of these application areas are served very well by several specifically
targeted books. These books do an excellent job of explaining in detail some as-
pect of color that happens to be most important for the target audience. This is
understandable as our knowledge of color spans many disciplines and can there-
fore be difficult to fathom.
It is our opinion that in application areas of computer science and computer
engineering, including such exciting fields as computer graphics, computer vi-
sion, high dynamic range imaging, image processing and game development, the
role of color is not yet fully appreciated. We have come across several applications
as well as research papers where color is added as an afterthought, and frequently
wrongly too. The dreaded RGB color space, which is really a collection of loosely
similar color spaces, is one of the culprits.
With this book, we hope to give a deep understanding of what color is, and
where color comes from. We also aim to show how color can be used correctly
in many different applications. Where appropriate, we include at the end of each
chapter sections on applications that exploit the material covered. While the book
is primarily aimed at computer-science and computer-engineering related areas,
as mentioned above, it is suitable for any technically minded reader with an in-
terest in color. In addition, the book can also be used as a text book serving a
graduate-level course on color theory. In any case, we believe that to be useful in
any engineering-related discipline, the theories should be presented in an intuitive
manner, while also presenting all of the mathematics in a form that allows both a
deeper understanding, as well as its implementation.
xiii

i i

i i
 i i

i i

xiv Preface

Most of the behavior of light and color can be demonstrated with simple ex-
periments that can be replicated at home. To add to the appeal of this book, where
possible, we show how to set-up such experiments that frequently require no more
than ordinary household objects. For instance, the wave-like behavior of light is
easily demonstrated with a laser pointer and a knife. Also, several visual illusions
can be replicated at home. We have shied away from such simple experiments
only when unavoidable.
The life cycle of images starts with either photography or rendering, and in-
volves image processing, storage, and display. After the introduction of digital
imaging, the imaging pipeline has remained essentially the same for more than
two decades. The phosphors of conventional CRT devices are such that in the
operating range of the human visual system only a small number of discernible
intensity levels can be reproduced. As a result, there was never a need to capture
and store images with a fidelity greater than can be displayed. Hence the immense
legacy of eight-bit images.
High dynamic range display devices have effectively lifted this restriction, and
this has caused a rethinking of the imaging pipeline. Image capturing techniques
can and should record the full dynamic range of the scene, rather than just the
restricted range that can be reproduced on older display devices. In this book,
the vast majority of the photography was done in high dynamic range (HDR),
with each photograph tone-mapped for reproduction on paper. In addition, high
dynamic range imaging (HDRI) is integral to the writing of the text, with excep-
tions only made in specific places to highlight the differences between conven-
tional imaging and HDRI. Thus, the book is as future-proof as we could possibly
make it.

Acknowledgments
Numerous people have contributed to this book with their expertise and help.
In particular, we would like to thank Eric van Stryland, Dean and Director of
CREOL, who has given access to many optics labs, introduced us to his col-
leagues, and allowed us to photograph some of the exciting research undertaken
at the School of Optics, University of Central Florida. Karen Louden, Curator
and Director of Education of the Albin Polasek Museum, Winter Park, Florida,
has given us free access to photograph in the Albin Polasek collection.
We have sourced many images from various researchers. In particular, we
are grateful for the spectacular renderings given to us by Diego Gutierrez and
his colleagues from the University of Zaragoza. The professional photographs
donated by Kirt Witte (Savannah College of Art and Design) grace several pages,
and we gratefully acknowledge his help. Several interesting weather phenomena
were photographed by Timo Kunkel, and he has kindly allowed us to reproduce

i i

i i
 i i

i i

Preface xv

some of them. We also thank him for carefully proofreading an early draft of the
manuscript.
We have had stimulating discussions with Karol Myszkowski, Grzegorz
Krawczyk, Rafał Mantiuk, Kaleigh Smith, Edward Adelson and Yuanzhen Li,
the results of which have become part of the chapter on tone reproduction. This
chapter also benefitted from the source code of Yuanzhen Li’s tone-reproduction
operator, made available by Li and her colleagues, Edward Adelson and Lavanya
Sharan. We are extremely grateful for the feedback received from Charles Poyn-
ton, which helped improve the manuscript throughout in both form and substance.
We have received a lot of help in various ways, both direct and indirect, from
many people. In no particular order, we gratefully acknowledge the help from
Janet Milliez, Vasile Rotar, Eric G Johnson, Claudiu Cirloganu, Kadi Bouatouch,
Dani Lischinski, Ranaan Fattal, Alice Peters, Franz and Ineke Reinhard, Gordon
Kindlmann, Sarah Creem-Regehr, Charles Hughes, Mark Colbert, Jared Johnson,
Jaakko Konttinen, Veronica Sundstedt, Greg Ward, Mashhuda Glencross, Helge
Seetzen, Mahdi Nezamabadi, Paul Debevec, Tim Cox, Jessie Evans, Michelle
Ward, Denise Penrose, Tiffany Gasbarrini, Aaron Hertzmann, Kevin Suffern,
Guoping Qiu, Graham Finlayson, Peter Shirley, Michael Ashikhmin, Wolfgang
Heidrich, Karol Myszkowski, Grzegorz Krawczyk, Rafal Mantiuk, Kaleigh
Smith, Majid Mirmehdi, Louis Silverstein, Mark Fairchild, Nan Schaller, Walt
Bankes, Tom Troscianko, Heinrich Bülthoff, Roland Fleming, Bernard Riecke,
Kate Devlin, David Ebert, Francisco Seron, Drew Hess, Gary McTaggart, Habib
Zargarpour, Peter Hall, Maureen Stone, Holly Rushmeier, Narantuja Bujantog-
toch, Margarita Bratkova, Tania Pouli, Ben Long, Native Visions Art Gallery
(Winter Park, Florida), the faculty, staff, and students of the Munsell Color Sci-
ence Laboratory, Lawrence Taplin, Ethan Montag, Roy Berns, Val Helmink,
Colleen Desimone, Sheila Brady, Angus Taggart, Ron Brinkmann, Melissa An-
sweeney, Bryant Johnson, and Paul and Linda Johnson.

i i

i i
 i i

i i

Part I
Principles

i i

i i
 i i

i i

Chapter 1
Introduction

Color is a phenomenon that relates to the physics of light, chemistry of matter,

geometric properties of objects, as well as to human visual perception and cog-
nition. We may call sunlight yellow, and in this case we refer to a property of
light. A car may be painted red, in which case the color red is an attribute of the
car. When light enters the eye, a complex chain of events leads to the sensation of
color, a perceived quantity. Finally, color may be remembered, associated with
events, and reasoned about. These are cognitive aspects of color. Color means
different things under different circumstances [814].
At the same time, it is clear that an understanding of color will involve each
of these aspects. Thus, the study of color theory and its applications necessar-
ily spans several different fields, including physics, chemistry, optics, radiometry,
photometry, colorimetry, physiology, vision science, color appearance modeling,
and image processing. As a result, what on the surface appears to be a rela-
tively simple subject, turns out to have many hidden depths. Perhaps this is the
reason that in practical applications found in computer science—computer vision,
graphics and image processing—the use of color is often under-explored and mis-
understood.
In our view, to understand color with sufficient depth, and to be able to apply
this knowledge to your own area of interest, it is not enough to read the literature
in any one specific discipline, be it computer graphics, computer vision, photog-
raphy, art, etc. Instead, it is necessary to step outside one’s own field in order to
appreciate the subtleties and complexities of color.
The purpose of this book is therefore to explain color theory, its development
and current state-of-the-art, as well as its practical use in engineering-oriented
disciplines such as computer graphics, computer vision, photography, and film.
Along the way, we delve into the physics of light, and its interaction with matter
3

i i

i i
 i i

i i

4 1. Introduction

at the atomic level, such that the origins of color can be appreciated. We find that
the intimate relationship between energy levels, orbital states, and electromag-
netic waves helps to understand why diamonds shimmer, rubies are red, and the
feathers of the blue jay are blue. Even before light enters the eye, a lot has already
happened.
The complexities of color multiply when perception is taken into account. The
human eye is not a simple light detector by any stretch of the imagination. Human
vision is able to solve an inherently under-constrained problem: it tries to make
sense out of a 3D world using optical projections that are two-dimensional. To
reconstruct a three-dimensional world, the human visual system needs to make
a great many assumptions about the structure of the world. It is quite remark-
able how well this system works, given how difficult it is to find a computational
solution that only partially replicates these achievements.
When these assumptions are violated, the human visual system can be fooled
into perceiving the wrong thing. For instance, if a human face is lit from above, it
is instantly recognizable. If the same face is lit from below, it is almost impossi-
ble to determine whose face it is. It can be argued that whenever an assumption is
broken, a visual illusion emerges. Visual illusions are therefore important to learn
about how the human visual system operates. At the same time, they are impor-
tant, for instance in computer graphics, to understand which image features need
to be rendered correctly and which ones can be approximated while maintaining
realism.
Color theory is at the heart of this book. All other topics serve to underpin
the importance of using color correctly in engineering applications. We find that
too often color is taken for granted, and engineering solutions, particularly in
computer graphics and computer vision, therefore appear suboptimal. To redress
the balance, we provide chapters detailing all important issues governing color
and its perception, along with many examples of applications.
We begin this book with a brief assessment of the roles color plays in different
contexts, including nature and society.

1.1 Color in Nature

Living organisms are embedded in an environment with which they interact. To
maximize survival, they must be in tune with this environment. Color plays an
important role in three ways:

• Organisms may be colored by default without this giving them a specific

advantage for survival. An example is the green color of most plants (Fig-

i i

i i
 i i

i i

1.1. Color in Nature 5

Figure 1.1. The chlorophyll in leaves causes most plants to be colored green.

ure 1.1), which is due to chlorophyll, a pigment that plays a role in photo-
synthesis (see Section 3.4.1).

• Color has evolved in many species in conjunction with the color vision
of the same or other species, for instance for camouflage (Figure 1.2), for
attracting partners (Figure 1.3), for attracting pollinators (Figure 1.4), or for
appearing unappetizing to potential predators (Figure 1.5).

• Biochemical substances may be colored as a result of being optimized to

serve a specific function unrelated to color. An example is hemoglobin
which colors blood red due to its iron content. Such functional colors are
normally found inside the body, rather than at the surface.
Plants reflect green and absorb all other colors of light. In particular, plants
absorb red and blue light and use the energy gained to drive the production of
carbohydrates. It has been postulated that these two colors are being absorbed as
a result of two individual mechanisms that allow plants to photosynthesize with
maximum efficiency under different lighting conditions [518, 769].

i i

i i
 i i

i i

6 1. Introduction

Figure 1.2. Many animals are colored similar to their environment to evade predators.

Figure 1.3. This peacock uses bright colors to attract a mate; Paignton Zoo, Devon, UK.
(Photo by Brett Burridge (www.brettb.com).)

i i

i i
 i i

i i

1.1. Color in Nature 7

Figure 1.4. Many plant species grow brightly colored flowers to attract pollinators such
as insects and bees; Rennes, France, June 2005.

Figure 1.5. These beetles have a metallic color, presumably to discourage predators.

i i

i i
 i i

i i

8 1. Introduction

Figure 1.6. This desert rose is colored to reflect light and thereby better control its
temperature.

Color in plants also aids other functions such as the regulation of temperature.
In arid climates plants frequently reflect light of all colors, thus appearing lightly
colored such as the desert rose shown in Figure 1.6.
In humans, color vision is said to have co-evolved with the color of fruit [787,
944]. In industrialized Caucasians, color deficiencies occur relatively often, while
color vision is better developed in people who work the land. Thus, on average,
color vision is diminished in people who do not depend on it for survival [216].
Human skin color is largely due to pigments such as eumelanin and phaeome-
lanin. The former is brown to black, whereas the latter is yellow to reddish-brown.
Deeper layers contain yellow carotenoids. Some color in human skin is derived
from scattering, as well as the occurrence of blood vessels [520].
Light scattering in combination with melanin pigmentation is also the mech-
anism that determines eye color in humans and mammals. There appears to be
a correlation between eye color and reactive skills. In the animal world, hunters
who stalk their prey tend to have light eye colors, whereas hunters who obtain
their prey in a reactive manner, such as birds that catch insects in flight, have dark
eyes. This difference, as yet to be fully explained, extends to humankind where
dark-eyed people tend to have faster reaction times to both visual and auditory
stimuli than light-eyed people [968].

i i

i i
 i i

i i

1.1. Color in Nature 9

Figure 1.7. Color plays an important role in art. An example is this photograph of the
Tybee Light House, taken by Kirt Witte, which won the 2006 International Color Award’s
Masters of Color Photography award in the abstract category for professional photogra-
phers (see also www.theothersavannah.com).

i i

i i
 i i

i i

10 1. Introduction

Figure 1.8. Statue by Albin Polasek; Albin Polasek Museum, Winter Park, FL, 2004.

1.2 Color in Society

The use of color by man to communicate is probably as old as mankind itself. Ar-
chaeologists have found colored materials, in particular different shades of ochre,
at sites occupied some 300,000 years ago by Homo erectus [519]. Approximately
70,000 years ago, Homo sapiens neanderthalensis used ochre in burials, a tradi-
tion followed later by Homo sapiens sapiens.
Color, of course, remains an important means of communication in art [695,
1257,1291] (Figure 1.7), even in cases where color is created by subtle reflections
(Figure 1.8).
Color also plays an important role in religion, where the staining of church
windows is used to impress churchgoers. The interior of a church reflects the
times in which it was built, ranging from light and airy to dark and somber
(Figure 1.9).

i i

i i
 i i

i i

1.2. Color in Society 11

Figure 1.9. Color in different church interiors; Rennes, France, June 2005 (top); Mainau,
Germany, July 2005 (bottom).

With modern technology, new uses of color have come into fashion. Impor-
tant areas include the reproduction of color [509], lighting design [941], photogra-
phy [697], and television and video [923]. In computer graphics, computer vision,
and other engineering disciplines, color also plays a crucial role, but perhaps less
so than it should. Throughout this book, examples of color use in these fields are
provided.

i i

i i
 i i

i i

12 1. Introduction

1.3 In this Book

This book offers an in-depth treatment of various topics related to color. Our
intention is to explain only a subset of color science, but to treat each of the topics
we have chosen in depth. Our choice of topics is intended to be relevant to those
who require a more-than-casual understanding of color for their work, which we
envisage to be in computer graphics, computer vision, animation, photography,
image processing, and related disciplines.
All use of color in any discipline is either explicitly or implicitly based on
theories developed to describe the physics of light, as well as the perception of it
by humans. We therefore offer in the first chapters a reasonably detailed account
of light, its propagation through vacuum as well as other media, and its interaction
with boundaries.
The physics of light is governed by Maxwell’s equations, which form the
basis of our thinking about the wave nature of light. Without these equations,
there would not be any physical optics. Since almost everything in radiometry
and geometric physics is derived from Maxwell’s equations, we expect that areas
of study further afield would also look very different without them. Fields affected
include photometry, lighting design, computer graphics, and computer vision. As
such, we begin this book with a detailed description of electromagnetic waves
and show how radiometry, geometric optics, and much of computer graphics are
derived from them.
While the wave nature of light constitutes a powerful model in understand-
ing the properties of light, the theory of electromagnetic radiation is not able to
explain all measurable light behavior. In particular, light sometimes behaves as
particles, and this fact is not captured by electromagnetic theory. Hence, in Chap-
ter 3, we briefly introduce quantum mechanics as well as molecular orbital theory,
as these form the basic tools for understanding how light interacts with matter at
the atomic scale. Such interaction is the cause of various behaviors such as ab-
sorption, emission, diffraction, and dispersion. These theories afford insight into
questions such as “why is water pale blue” and “why is a ruby red and can sap-
phires be blue”. Thus, Chapter 3 is largely concerned with providing answers to
questions regarding the causes of color. Whereas Chapter 2 deals with the prop-
agation of light, Chapter 3 is largely informed by the interaction of light with
matter.
Chapter 4 provides a brief introduction to human vision. The optics of the eye,
as well as the neurophysiology of what is collectively known as the human visual
system, are described. The scientific literature on this topic is vast, and a full
account of the neuronal processing of visual systems is beyond the scope of this

i i

i i
 i i

i i

1.3. In this Book 13

book. However, it is clear that with the advent of more sophisticated techniques,
the once relatively straightforward theories of color processing in the visual cor-
tex, have progressed to be significantly less straightforward. This trend is contin-
uing to this day. The early inferences made on the basis of single-cell recordings
have been replaced with a vast amount of knowledge that is often contradictory,
and every new study that becomes available poses intriguing new questions. On
the whole, however, it appears that color is not processed as a separate image
attribute, but is processed together with other attributes such as position, size,
frequency, direction, and orientation.
Color can also be surmised from a perceptual point of view. Here, the human
visual system is treated as a black box, with outputs that can be measured. In
psychophysical tests, participants are set a task which must be completed in re-
sponse to the presentation of visual stimuli. By correlating the task response to
the stimuli that are presented, important conclusions regarding the human visual
system may be drawn. Chapter 5 describes some of the findings from this field of
study, as it pertains to theories of color. This includes visual illusions, adaptation,
visual acuity, contrast sensitivity, and constancy.
In the following chapters, we build upon the fundamentals underlying color
theory. Chapter 6 deals with radiometry and photometry, whereas Chapter 7 dis-
cusses colorimetry. Much research has been devoted to color spaces that are de-
signed for different purposes. Chapter 8 introduces many of the currently-used
color spaces and explains the strengths and weaknesses of each color space. The
purpose of this chapter is to give transformations between existing color spaces
and to enable the selection of an appropriate color space for specific tasks, realiz-
ing that each task may require a different color space.
Light sources, and their theoretical formulations (called illuminants), are dis-
cussed in Chapter 9. Chapter 10 introduces chromatic adaptation, showing that the
perception of a colored object does not only depend on the object’s reflectance,
but also on its illumination and the state of adaptation of the observer. While
colorimetry is sufficient for describing colors, an extended model is required to
account for the environment in which the color is observed. Color appearance
models take as input the color of a patch, as well as a parameterized description
of the environment. These models then compute appearance correlates that de-
scribe perceived attributes of the color, given the environment. Color appearance
models are presented in Chapter 11.
In Part III, the focus is on images, and in particular their capture and display.
Much of this part of the book deals with the capture of high dynamic range im-
ages, as we feel that such images are gaining importance and may well become
the de-facto norm in all applications that deal with images.

i i

i i
 i i

i i

14 1. Introduction

Chapter 12 deals with the capture of images and includes an in-depth descrip-
tion of the optical processes involved in image formation, as well as issues related
to digital sensors. This chapter also includes sections on camera characterization,
and more specialized capture techniques such as holography and light field data.
Techniques for the capture of high dynamic range images are discussed in Chap-
ter 13. The emphasis is on multiple exposure techniques, as these are currently
most cost effective, requiring only a standard camera and appropriate software.
Display hardware is discussed in Chapter 14, including conventional and
emerging display hardware. Here, the focus is on liquid crystal display devices, as
these currently form the dominant display technology. Further, display calibration
techniques are discussed.
Chapter 15 is devoted to a discussion on natural image statistics, a field im-
portant as a tool both to help understand the human visual system, and to help
structure and improve image processing algorithms. This chapter also includes
sections on techniques to measure the dynamic range of images, and discusses
cross-media display technology, gamut mapping, gamma correction, and algo-
rithms for correcting for light reflected off display devices. Color management for
images is treated in Chapter 16, with a strong emphasis on ICC profiles. Finally,
Chapter 17 presents current issues in tone reproduction, a collection of algorithms
required to prepare a high dynamic range image for display on a conventional dis-
play device.
For each of the topics presented in the third part of the book, the emphasis
is on color management, rather than spatial processing. As such, these chapters
augment, rather than replace, current books on image processing.
The book concludes with a set of appendices, which are designed to help clar-
ify the mathematics used throughout the book (vectors and matrices, trigonome-
try, and complex numbers), and to provide tables of units and constants for easy
reference. We also refer to the DVD-ROM included with the book, which contains
a large collection of images in high dynamic range format, as well as tonemapped
versions of these images (in JPEG-HDR format for backward compatibility), in-
cluded for experimentation. The DVD-ROM also contains a range of spectral
functions, a metameric spectral image, as well as links to various resources on the
Internet.

1.4 Further Reading

The history of color is described in Nassau’s book [814], whereas some of the
history of color science is collected in MacAdam’s books [717, 720]. A historical

i i

i i
 i i

i i

1.4. Further Reading 15

overview of dyes and pigments is available in Colors: The Story of Dyes and Pig-
ments [245]. An overview of color in art and science is presented in a collection
of papers edited by Lamb and Bourriau [642]. Finally, a history of color order,
including practical applications, is collected in Rolf Kuehni’s Color Space and its
Divisions: Color Order from Antiquity to Present [631].

i i

i i
 i i

i i

Chapter 2
Physics of Light

Light travels through environments as electromagnetic energy that may interact

with surfaces and volumes of matter. Ultimately, some of that light reaches the
human eye which triggers a complicated chain of events leading to the perception,
cognition, and understanding of these environments.
To understand the physical aspects of light, i.e., everything that happens to
light before it reaches the eye, we have to study electromagnetism. To understand
the various ways by which colored light may be formed, we also need to know a
little about quantum mechanics and molecular orbital theory (discussed in Chapter
3). These topics are not particularly straightforward, but they are nonetheless well
worth studying. They afford insight into the foundations of many fields related
to color theory, such as optics, radiometry (and therefore photometry), as well as
computer graphics.
The physical properties of light are well modeled by Maxwell’s equations. We
therefore begin this chapter with a brief discussion of Maxwell’s equations. We
then discuss various optical phenomena that may be explained by the theory of
electromagnetic radiation. These include scattering, polarization, reflection, and
refraction.
There are other optical phenomena that involve the interaction between light
and materials at the atomic structure. Examples of these phenomena are diffrac-
tion, interference, and dispersion, each capable of separating light into different
wavelengths—and are therefore perceived as producing different colors. With
the exception of interference and diffraction, the explanation of these phenom-
ena requires some insight into the chemistry of matter. We therefore defer their
description until Chapter 3.
In addition to presenting electromagnetic theory, in this chapter we also in-
troduce the concept of geometrical optics, which provides a simplified view of
17

i i

i i
 i i

i i

18 2. Physics of Light

the theory of light. It gives rise to various applications, including ray tracing in
optics. It is also the foundation for all image synthesis as practiced in the field of
computer graphics. We show this by example in Section 2.10.
Thus, the purpose of this chapter is to present the theory of electromagnetic
waves and to show how light propagates through different media and behaves near
boundaries and obstacles. This behavior by itself gives rise to color. In Chapter 3,
we explain how light interacts with matter at the atomic level, which gives rise to
several further causes for color, including dispersion and absorption.

2.1 Electromagnetic Theory

Light may be modeled by a transverse electromagnetic (TEM) wave traveling
through a medium. This suggests that there is an interaction between electric
and magnetic fields and their sources of charge and current. A moving electric
charge is the source of an electric field. At the same time, electric currents pro-
duce a magnetic field. The relationship between electric and magnetic fields are
governed by Maxwell’s equations which consist of four laws:

• Gauss’ law for electric fields;

• Gauss’ law for magnetic fields;

• Faraday’s law;

• Ampere’s circuital law.

We present each of these laws in integral form first, followed by their equiv-
alent differential form. The integral form has a more intuitive meaning but is
restricted to simple geometric cases, whereas the differential form is valid for any
point in space where the vector fields are continuous.
There are several systems of units and dimensions used in Maxwell’s equa-
tions, including Gaussian units, Heaviside-Lorentz units, electrostatic units, elec-
tromagnetic units, and SI units [379]. There is no specific reason to prefer one
system over another. Since the SI system is favored in engineering-oriented dis-
ciplines, we present all equations in this system. In the SI system, the basic quan-
tities are the meter (m) for length, the kilogram (kg) for mass, the second (s)
for time, the ampere (A) for electric current, the kelvin (K) for thermodynamic
temperature, the mole (mol) for amount of substance, and the candela (cd) for
luminous intensity. (see Table D.3 in Appendix D).

i i

i i
 i i

i i

2.1. Electromagnetic Theory 19

-F

eQ

-e Q
R
F q

Figure 2.1. Two charges Q and q exert equal, but opposite force F upon each other
(assuming the two charges have equal sign).

2.1.1 Electric Fields

Given a three-dimensional space, we may associate attributes with each point in
this space. For instance, the temperature in a room may vary by location in the
room. With heat tending to rise, the temperature near the ceiling is usually higher
than near the floor.
Similarly, it is possible to associate other attributes to regions in space. If
these attributes have physical meaning, we may speak of a field, which simply
indicates that there exists a description of how these physical phenomena change
with position in space. In a time-varying field, these phenomena also change with
time.
For instance, we could place an electrical charge in space and a second charge
some distance away. These two charges exert a force on each other dependent
upon the magnitude of their respective charges and their distance. Thus, if we
move the second charge around in space, the force exerted on it by the first charge
changes (and vice versa). The two charges thus create a force field.
The force F that two charges Q and q exert on each other is given by
Coulomb’s law (see Figure 2.1):

1 Qq
F= eQ . (2.1)
4 π ε 0 R2

In this equation, R is the distance between the charges, and the constant ε0 =
−9 (in Farad/meter) is called the permittivity of vacuum. The vector e is
36 π × 10
1
Q
a unit vector pointing from the position of one charge to the position of the other.
If we assume that Q is an arbitrary charge and that q is a unit charge, then
we can compute the electric field intensity E by dividing the left- and right-hand

i i

i i
 i i

i i

20 2. Physics of Light

Q2 Q1
R2
-e Q
2

-e Q
q 1
R1
E1

E2
E

Figure 2.2. The electric field intensity E at the position of charge q is due to multiple
charges located in space.

sides of Equation (2.1) by q:

F Q
E= = eQ . (2.2)
q 4 π ε 0 R2

If there are multiple charges present, then the electric field intensity at the
position of the test charge is given by the sum of the individual field intensities
(Figure 2.2):
N
Qi
E=∑ e .
2 Qi
(2.3)
i=1 4 π ε0 Ri

Finally, if we scale the electric field intensity E by the permittivity of vacuum

(ε0 ), we obtain what is called the electric flux density, indicated by the vector D
which points in the same direction as E:
N
Qi
D = ε0 E = ∑ e .
2 Qi
(2.4)
i=1 4 π Ri

For media other than vacuum, the permittivity will generally have a different
value. In that case, we drop the subscript 0, so that, in general, we have

D = ε E. (2.5)

This equation may be seen as relating the electric field intensity to the electric
flux density, whereby the difference between the two vectors is in most cases
determined by a constant unique to the specific medium. This constant, ε , is
called the material’s permittivity or dielectric constant. Equation (2.5) is one of

i i

i i
 i i

i i

2.1. Electromagnetic Theory 21

three so-called material equations. The remaining two material equations will be
discussed in Sections 2.1.3 and 2.1.8.
If, instead of a finite number of separate charges, we have a distribution of
charges over space, the electric flux density is governed by Gauss’ law for electric
fields.

2.1.2 Gauss’ Law for Electric Fields

In a static or time-varying electric field there is a distribution of charges. The rela-
tionship between an electric field and a charge distribution is quantified by Gauss’
law for electric fields. It states that the total electric flux, D = ε E, emanating from
a closed surface s is equal to the electric charge Q enclosed by that surface:1

D · n ds = Q. (2.6)
s

The integral is over the closed surface s and n denotes the outward facing surface
normal. If the charge Q is distributed over the volume v according to a charge
distribution function ρ (also known as electric charge density), we may rewrite
Gauss’ law as follows:  
D · n ds = ρ dv. (2.7)
s v
Thus, a distribution of charges over a volume gives rise to an electric field that may
be measured over a surface that bounds that volume. In other words, the electric
flux emanating from an enclosing surface is related to the charge contained by
that surface.

2.1.3 Magnetic Fields

While charges may create an electric field, electric currents may create a magnetic
field. Thus, analogous to electric flux, we may speak of magnetic flux that has the
ability to exert a force on either a magnet or another electric current.
Given that an electric current is nothing more than a flow of moving charges,
it is apparent that a magnetic field can only be produced from moving charges;
stationary charges do not produce a magnetic field. Conversely, a magnetic field
has the ability to exert a force on moving charged particles.
The magnetic flux density associated with a magnetic field is indicated by B.
A charged particle with a charge Q moving with velocity v through a magnetic
field with a flux density of B is pulled by a force F which is given by

F = Q v × B. (2.8)
1 See also Appendix A which provides the fundamentals of vector algebra and includes further

detail about the relationship between integrals over contours, surfaces, and volumes.

i i

i i
 i i

i i

22 2. Physics of Light

v B
B

E+v B v
Q

Figure 2.3. The Lorentz force equation: F is the sum of E and the cross product of the
particle’s velocity v and the magnetic flux density B at the position of the particle.

However, according to (2.2), a particle with charge Q is also pulled by a force

equal to F = Q E. The total force exerted on such a particle is given by the su-
perposition of electric and magnetic forces, which is known as the Lorentz force
equation (Figure 2.3):
F = Q (E + v × B) . (2.9)
This equation thus provides a method to determine the motion of a charged par-
ticle as it moves through a combined electric and magnetic field.
Given the cross-product between the magnetic flux density and the velocity
vector in the above equation, we deduce that the direction of the magnetic force is
perpendicular to both the direction of movement and the direction of the magnetic
flux. As such, this force has the ability to change the direction of the motion, but
not the magnitude of the charged particle’s velocity. It also does not change the
energy associated with this particle.
On the other hand, the electric field exerts a force that is independent of the
motion of the particle. As a result, energy may be transferred between the field
and the charged particle.
The magnetic flux density B has a related quantity called the magnetic vector,
indicated by H. The relationship between B and H is governed by a material
constant μ called the magnetic permeability:

B = μ H. (2.10)

This equation is the second of three material equations and will be discussed
further in Section 2.1.8.

2.1.4 Gauss’ Law for Magnetic Fields

While electric fields are governed by Gauss’ law for electric fields (the electric
flux emanating from a closed surface depends on the charge present in the volume

i i

i i
 i i

i i

2.1. Electromagnetic Theory 23

enclosed by that surface), magnetic fields behave somewhat differently. The total
magnetic flux emanating from a closed surface bounding a magnetic field is equal
to zero; this constitutes Gauss’ law for magnetic fields:

B · n ds = 0. (2.11)
s

This equation implies that no free magnetic poles exist. As an example, the mag-
netic flux emanating from a magnetic dipole at its north pole is matched by the
flux directed inward towards its south pole.

2.1.5 Faraday’s Law

As shown earlier, currents produce a magnetic field. Conversely, a time-varying
magnetic field is capable of producing a current. Faraday’s law states that the
electric field induced by a time-varying magnetic field is given by
 
d
E · n dc = − B · n ds. (2.12)
c dt s

Here, the left-hand side is an integral over the contour c that encloses an open
surface s. The quantity integrated is the component of the electric field intensity E
normal to the contour. The right-hand side integrates the normal component of B
over the surface s. Note that the right-hand side integrates over an open surface,
whereas the integral in (2.11) integrates over a closed surface.

2.1.6 Ampere’s Circuital Law

Given a surface area s enclosed by a contour c, the magnetic flux density along
this contour is related to the total current passing through area s. The total current
is composed of two components, namely a current as a result of moving charged
particles and a current related to changes in electric flux density. The latter current
is also known as displacement current.
Moving charged particles may be characterized by the current flux density j,
which is given in ampere per square meter (A/m2 ). If electric charges with a
density of ρ are moving with a velocity v, the current flux density j is given by

j = ρ v. (2.13)

If the current flux density j is integrated over surface area, then we find the total
charge passing through this surface per second (coulomb/second = ampere; C/s =
A). Thus, the current resulting from a flow of charges is given by

j · n ds. (2.14)
s

i i

i i
 i i

i i

24 2. Physics of Light

The displacement current depends on the electric flux density. If we in-

tegrate the electric flux density over the same surface area, we obtain charge
(coulomb; C): 
D · n ds. (2.15)
s
If we differentiate this quantity by time, the result is a charge passing through
surface s per second, i.e., current:

d
D · n ds. (2.16)
dt s

The units in (2.14) and (2.16) are now both in coulomb per second and are thus
measures of current. Both types of current are related to the magnetic flux density
according to Ampere’s circuital law:
  
d
H · n dc = j · n ds + D · nds. (2.17)
c s dt s

This law states that a time-varying magnetic field can be produced by the flow of
charges (a current), as well as by a displacement current.

2.1.7 Maxwell’s Equations

Both of Gauss’ laws, Faraday’s law, and Ampere’s circuital law together form a
set of equations that are collectively known as Maxwell’s equations. For conve-
nience we repeat them here:
 
D · n ds = ρ dv; Gauss’ law for electric fields (2.18a)
s v

B · n ds = 0; Gauss’ law for magnetic fields

s
(2.18b)
 
d
E · n dc = − B · n ds; Faraday’s law (2.18c)
c dt s
  
d
H · n dc = j · n ds + D · n ds. Ampere’s circuital law (2.18d)
c s dt s

The above equations are given in integral form. They may be rewritten in dif-
ferential form, after which these equations hold for points in space where both
electric and magnetic fields are continuous. This facilitates solving these four
simultaneous equations.
Starting with Gauss’ law for electric fields, we see that the left-hand side
of (2.18a) is an integral over a surface, whereas the right-hand side is an integral

i i

i i
 i i

i i

2.1. Electromagnetic Theory 25

over a volume. As we are interested in a form of Maxwell’s equations that is valid

for points in space, we would like to replace the left-hand side with an integral
over the volume under consideration. This may be accomplished by applying
Gauss’ theorem (see Appendix A), yielding
 
D · n ds = ∇·D dv. (2.19)
s v

Combining this result with (2.18a) yields

 
∇·D dv = ρ dv, (2.20)
v v

and, therefore, in the limit when the volume v goes to zero, we have

∇·D = ρ . (2.21)

A similar set of steps may be applied to Gauss’ law for magnetic fields, and this
will result in a similar differential form:

∇·B = 0. (2.22)

Ampere’s law, given in (2.18d), is stated in terms of a contour integral on the left-
hand side and a surface integral on the right-hand side. Here, it is appropriate to
apply Stokes’ theorem to bring both sides into the same domain (see Appendix A):
 
H · n dc = ∇× H · n ds. (2.23)
c s

Substituting this result into (2.18d) yields a form where all integrals are over the
same surface. In the limit when this surface area s goes to zero, Equation (2.18d)
becomes
∂D
∇× H = j + , (2.24)
∂t
which is the desired differential form. Finally, Faraday’s law may also be rewrit-
ten by applying Stokes’ theorem applied to the left-hand side of (2.18c):
  
∂B
E · n dc = ∇× E · n ds = − · n ds. (2.25)
c s s ∂t

Under the assumption that the area s becomes vanishingly small, this equation
yields
∂B
∇× E = − . (2.26)
∂t
Faraday’s law and Ampere’s law indicate that a time-varying magnetic field
has the ability to generate an electric field, and that a time-varying electric field

i i

i i
 i i

i i

26 2. Physics of Light

generates a magnetic field. Thus, time-varying electric and magnetic fields can
generate each other. This property forms the basis for wave propagation, allow-
ing electric and magnetic fields to propagate away from their source. As light can
be considered to consist of waves propagating through space, Maxwell’s equa-
tions are fundamental to all disciplines involved with the analysis, modeling, and
synthesis of light.

2.1.8 Material Equations

The four laws that comprise Maxwell’s equations are normally complemented
with three material equations (also known as constitutive relations). Two of
these were presented in Equations (2.5) and (2.10) and are repeated here for
convenience:

D = ε E; (2.27a)
B = μ H. (2.27b)

The third material equation relates the current flux density j to the electric field
intensity E according to a constant σ and is known as Ohm’s law. It is given here
in differential form:

j = σ E. (2.27c)

Material σ Material σ
Good conductors
Silver 6.17 × 107 Tungsten 1.82 × 107
Copper 5.8 × 107 Brass 1.5 × 107
Gold 4.1 × 107 Bronze 1.0 × 107
Aluminium 3.82 × 107 Iron 1.0 × 107
Poor conductors
Water (fresh) 1.0 × 10−3 Earth (dry) 1.0 × 10−3
Water (sea) 4.0 × 100 Earth (wet) 3.0 × 10−2
Insulators
Diamond 2.0 × 10−13 Porcelain 1.0 × 10−10
Glass 1.0 × 10−12 Quartz 1.0 × 10−17
Polystyrene 1.0 × 10−16 Rubber 1.0 × 10−15

Table 2.1. Conductivity constants σ for several materials [532].

i i

i i
 i i

i i

2.1. Electromagnetic Theory 27

The constant σ is called the specific conductivity, and is determined by the

medium. Values of σ for some materials are listed in Table 2.1. Materials with a
conductivity significantly different from 0 are called conductors. Metals, such as
gold, silver, and copper are good conductors. On the other hand, materials with
low conductivity (σ ≈ 0) are called insulators or dielectrics . Finally, for some
materials, (semiconductors) the conductivity increases with increasing tempera-
ture. See also Section 3.3.3.
For most materials, the magnetic permeability μ will be close to 1. Some
materials, however, have a permeability significantly different from 1, and these
are then called magnetic.
The speed v at which light travels through a medium is related to the material
constants as follows:
1
v= √ . (2.28)
εμ
We will derive this result in Section 2.2.4. The symbol μ0 is reserved for the
permeability of vacuum. The value of μ0 is related to both permittivity ε0 and the
speed of light c in vacuum as follows:
1
c= √ . (2.29)
ε0 μ0
Values for all three constants are given in Table 2.2. The permittivity and perme-
ability of materials is normally given relative to those of vacuum:

ε = ε0 εr ; (2.30a)
μ = μ0 μr . (2.30b)

Values of εr and μr are given for several materials in Tables 2.3 and 2.4.
Normally, the three material constants, σ , ε , and μ , are independent of the
field strengths. However, this is not always the case. For some materials these
values also depend on past values of E or B. In this book, we will not consider
such effects of hysteresis. Similarly, unless indicated otherwise, the material con-
stants are considered to be isotropic, which means that their values do not change

Constant Value Unit

c 3 · 108 m s−1
1
ε0 · 10−9 C2 s2 kg−1 m−3 = Fm−1
36 π
μ0 4 π · 10−7 kg m C−2 = Hm−1

Table 2.2. The speed of light c, permittivity ε0 , and permeability μ0 (all in vacuum).

i i

i i
 i i

i i

28 2. Physics of Light

Material εr Material εr
Air 1.0006 Paper 2−4
Alcohol 25 Polystyrene 2.56
Earth (dry) 7 Porcelain 6
Earth (wet) 30 Quartz 3.8
Glass 4 − 10 Snow 3.3
Ice 4.2 Water (distilled) 81
Nylon 4 Water (sea) 70

Table 2.3. Dielectric constants εr for several materials [532].

Material μr Material μr
Aluminium 1.000021 Nickel 600.0
Cobalt 250.0 Platinum 1.0003
Copper 0.99999 Silver 0.9999976
Gold 0.99996 Tungsten 1.00008
Iron 5000.0 Water 0.9999901

Table 2.4. Permeability constants μr for several materials [532].

with spatial orientation. Materials which do exhibit a variation in material con-

stants with spatial orientation are called anisotropic.

2.2 Waves
Maxwell’s equations form a set of simultaneous equations that are normally diffi-
cult to solve. In this section, we are concerned with finding solutions to Maxwell’s
equations, and to accomplish this, we may apply simplifying assumptions. The
assumptions outlined in the previous section, that material constants are isotropic
and independent of time, are the first simplifications. To get closer to an appro-
priate solution with which we can model light and optical phenomena, further
simplifications are necessary.
In particular, a reasonable class of models that are solutions to Maxwell’s
equations is formed by time-harmonic plane waves. With these waves, we can ex-
plain optical phenomena such as polarization, reflection, and refraction. We first
derive the wave equation, which enables the decoupling of Maxwell’s equations,
and therefore simplifies the solution. We then discuss plane waves, followed by
time-harmonic fields, and time-harmonic plane waves. Each of these steps con-
stitutes a further specialization of Maxwell’s equations.

i i

i i
 i i

i i

2.2. Waves 29

2.2.1 The Wave Equation

We assume that the source of an electromagnetic wave is sufficiently far away
from a given region of interest. In this case, the region is called source-free. In
such a region, the charge ρ and current distribution j will be 0, and as a result
Maxwell’s equations reduce to a simpler form:

∇·D = 0; (2.31a)
∇·B = 0; (2.31b)
∂B
∇× E = − ; (2.31c)
∂t
∂D
∇× H = . (2.31d)
∂t
We may apply the material equations for D and H to yield a set of Maxwell’s
equations in E and B only:

ε ∇·E = 0; (2.32a)
∇·B = 0; (2.32b)
∂B
∇× E = − ; (2.32c)
∂t
1 ∂E
∇× B = ε . (2.32d)
μ ∂t
This set of equations still expresses E in terms of B and B in terms of E. Never-
theless, this result may be decoupled by applying the curl operator to (2.32c) (see
Appendix A):
∂
∇× ∇× E = − ∇× B. (2.33)
∂t
Substituting (2.32d) into this equation then yields

∂ 2E
∇× ∇× E = −μ ε . (2.34)
∂ t2
To simplify this equation, we may apply identity (A.23) from Appendix A:

∂ 2E
∇(∇·E) − ∇2 E = − μ ε . (2.35)
∂ t2
From (2.32a), we know that we may set ∇·E to zero in (2.35) to yield the standard
equation for wave motion of an electric field,

∂ 2E
∇2 E − μ ε = 0. (2.36)
∂ t2

i i

i i
 i i

i i

30 2. Physics of Light

For a magnetic field, a similar wave equation may be derived:

∂ 2B
∇2 B − μ ε = 0. (2.37)
∂ t2
The two wave equations do not depend on each other, thereby simplifying the
solution of Maxwell’s equations. This result is possible, because the charge and
current distributions are zero in source-free regions, and, therefore, we could sub-
stitute ∇·E = 0 in (2.35) to produce (2.36) (and similarly for (2.37)). Thus, both
wave equations are valid for wave propagation problems in regions of space that
do not generate radiation (i.e., they are source-free). Alternatively, we may as-
sume that the source of the wave is sufficiently far away.
Under these conditions, we are therefore looking for solutions to the wave
equations, rather than solutions to Maxwell’s equations. One such solution is
afforded by plane waves, which we discuss next.

2.2.2 Plane Waves

A plane wave may be thought of as an infinitely large plane traveling in a given
direction. We will show that plane waves form a solution to the wave equations
derived in the preceding section. For a position vector r(x, y, z) and unit direction
vector s = (sx , sy , sz ), the solution will be of the form E = f (r · s,t). A plane is
then defined as
r · s = constant. (2.38)

For convenience, we choose a coordinate system such that one of the axes, say z,
is aligned with the direction of propagation, i.e., the surface normal of the plane is
s = (0, 0, 1). As a result, we have r·s = z. Thus, if we consider a wave modeled by
an infinitely large plane propagating through space in the direction of its normal
r · s, we are looking for a solution whereby the spatial derivatives in the plane
are zero and the spatial derivatives along the surface normal are non-zero. The
Laplacian operators in (2.36) and (2.37) then simplify to2

∂ 2E
∇2 E = . (2.39)
∂ z2
Substituting into (2.36) yields

∂ 2E ∂ 2E
− μ ε 2 = 0. (2.40)
∂z 2 ∂t
2 In this section we show results for E and note that similar results may be derived for B.

i i

i i
 i i

i i

2.2. Waves 31

A general solution to this equation is given by

   
1 1
E = E1 r · s − √ t + E2 r · s + √ t (2.41a)
με με
= E1 (r · s − vt) + E2 (r · s + vt) , (2.41b)
with E1 and E2 arbitrary functions.3 The general solution to the wave equations
therefore consists of two planes propagating into opposite directions (+r · s and
−r · s).
For simplicity, we consider only one of the planes and momentarily assume
that (a partial) solution is given by a single plane traveling in the +z direction:
E = E1 (r · s − vt). (2.42)
We will now show that for plane waves the vectors E and H are both perpendicular
to the direction of propagation s, and also that E and H are orthogonal to each
other. We begin by differentiating the above expression with respect to time,
which produces
∂E
= −vE , (2.43)
∂t
where the prime indicates differentiation with respect to the argument r · s − vt
of E. Next, we consider the curl of E. The x-component of the curl of E is given
by
∂ Ez ∂ Ey
(∇× E)x = − (2.44a)
∂y ∂z
= Ez sy − Ey sz (2.44b)
 
= s × E x . (2.44c)
The curl for the y- and z-components can be derived similarly. By substitution
of (2.43) and (2.44) into Maxwell’s equations (2.31) and applying the material
equations (2.5) and (2.10), we find
s × H + v ε E = 0, (2.45a)
 
s × E − v μ H = 0. (2.45b)
 
Using (2.29), we find that v ε = ε / μ and v μ = μ /ε . Integrating the previous
equation then gives

μ
E=− s × H, (2.46a)
ε

ε
H= s × E. (2.46b)
μ
3 We will derive expressions for these functions in the following sections.

i i

i i
 i i

i i

32 2. Physics of Light

To show that plane waves are transversal, i.e., both E and H are perpendicular
to the direction of propagation s, we take the dot product between E and s and H
and s:
  
μ
E·s = − s × H · s, (2.47a)
ε
 
ε
H·s = s × E · s. (2.47b)
μ

Using (A.6) in Appendix A, we find that E · s = H · s = 0, thereby proving that E

and H are perpendicular to s. Such waves are called transversal electromagnetic
(TEM) waves.
These equations also imply that s = E × H. The vector s is generally known
as the Poynting vector or power density vector and plays an important role in
quantifying energy. This can be seen from the units employed in vectors E (V /m)
and H (A/m), so that s is specified in watts per square meter (V /m·A/m = W /m2 ).
For planar waves in homogeneous media, the Poynting vector thus points in the
direction of wave propagation, and its length is a measure of the energy flow
induced by the electromagnetic field (discussed further in Section 2.4). In keeping
with general practice, we will use the symbol S when referring to the Poynting
vector in the remainder of this book.

2.2.3 Time-Harmonic Fields

A reasonable approach to modeling the oscillating wave behavior of light in free
space is by using sinusoidal functions. Thus, we assume that the current and
charge distributions, j and ρ , at the source vary with time as

a cos(ω t + ϕ ), (2.48)

where a is the amplitude and ω t + ϕ is the phase. The current and charge sources
vary with time t as well as space r and are therefore written as j(r,t) and ρ (r,t).
Using the results from Appendix C.4, we may separate these fields into a
spatial component and a time-varying component:

j(r,t) = j(r) eiω t , (2.49a)

iω t
ρ (r,t) = ρ (r) e . (2.49b)

i i

i i
 i i

i i

2.2. Waves 33

Since Maxwell’s equations are linear, this results in the following set of equations:

∇·Deiω t = ρ eiω t , (2.50a)

iω t
∇·Be = 0, (2.50b)
 
∇× Eeiω t = − iω Beiω t , (2.50c)
 
∇× Heiω t = iω Deiω t + jeiω t . (2.50d)

Note that all the underlined quantities are (complex) functions of space r only.
In addition, the time-dependent quantity eiω t cancels everywhere. For a homoge-
neous field, we may set j and ρ to zero. For a homogeneous field in steady state
we therefore obtain the following set of equations:

∇·D = 0, (2.51a)
∇·B = 0, (2.51b)
∇× E = − iω B, (2.51c)
∇× H = iω D. (2.51d)

By a procedure similar to the one shown in Section 2.2.1, wave equations for the
electric field E and the magnetic field E may be derived:

∇2 E − ω 2 μ ε E = 0, (2.52a)
∇ B − ω μ ε B = 0.
2 2
(2.52b)

2.2.4 Harmonic Plane Waves

In this section we will find a solution to the plane wave equations of harmonic
waves in homogeneous media, i.e., a solution to Equations (2.52). We will start
by assuming that the wave is traveling through a dielectric medium, i.e., the con-
ductivity of the material is close to zero (σ ≈ 0). Later in this section, we will
consider conductive materials.
For general plane waves we assumed that the planes are traveling in the +z
and −z directions. As both E and H (and thereby D and B) are transversal to the
direction or propagation, we find that

∂E ∂B
= 0, = 0, (2.53a)
∂x ∂x
∂E ∂B
= 0, = 0. (2.53b)
∂y ∂y

i i

i i
 i i

i i

34 2. Physics of Light

From Faraday’s law, we find identities for the partial derivatives in the z directions
for E:
∂ Ey
− = −i ω Bx , (2.54a)
∂z
∂ Ex
= −i ω By , (2.54b)
∂z
0 = −i ω Bz . (2.54c)

Similar solutions are found from Ampere’s law for B:

∂ By
− = i ω ε μ Ex , (2.54d)
∂z
∂ Bx
= i ω ε μ Ey , (2.54e)
∂z
0 = i ω ε μ Ez . (2.54f)
From these equations, we find that the components of E and B in the direction of
propagation (z) are zero. Differentiation of (2.54b) and substitution from (2.54d)
leads to
∂ 2 Ex ∂ By
= −i ω = −ω 2 μ ε Ex . (2.55)
∂z 2 ∂z
As we assume that the field is uniform, Ex is a function of z only, and we may
therefore replace the partial derivatives with ordinary derivatives, yielding the
wave equation for harmonic plane waves ,
d 2 Ex
+ ω 2 μ ε Ex = 0. (2.56)
dz2
Similar equations may be set up for Ey , Bx , and By . By letting
β 2 = ω2 μ ε, (2.57)
a general solution for (2.56) is given by
−i β z iβ z
Ex = E+
me + E−
me . (2.58)
+
For argument’s sake, we will assume that the newly introduced constants Em and
− + −
Em are real, and therefore we replace them with Em and Em . The solution of the
wave equation is then the real part of Ex ei ω t :
Ex (z,t) = Re{Ex ei ω t } (2.59a)
i(ω t−β z) i(ω t+β z)
= Re{E+
me + E−
me } (2.59b)
= E+
m cos(ω t − β z) + E−
m cos(ω t + β z). (2.59c)

i i

i i
 i i

i i

2.2. Waves 35

Thus, the solution to the wave equation for harmonic plane waves in homogeneous
media may be modeled by a pair of waves, one propagating in the +z direction
and the other traveling in the −z direction.
Turning our attention to only one of the waves, for instance E+
m cos (ω t − β z),
it is clear that by keeping to a single position z, the wave produces an oscillation
with angular frequency ω . The frequency f of the wave and its period T may be
derived from ω as follows:
ω 1
f= = . (2.60)
2π T
At the same time, the wave E+m cos(ω t − β z) travels through space in the positive
z direction, which follows from the −β z component of the wave’s phase. The
value of the cosine does not alter if we add or subtract multiples of 2 π to the
phase. Hence, we have

ω t + β z = ω t + β (z + λ ) + 2 π . (2.61)

Solving for λ , which is called wavelength, and combining with (2.57) we get
2π 2π
λ= = √ . (2.62)
β ω με

With the help of Equations (2.29) and (2.60), we find the well-known result that
the wavelength λ of a harmonic plane wave relates to its frequency by means of
the speed of light:
vp
λ= . (2.63)
f
The phase velocity v p may be viewed as the speed with which the wave prop-
agates. This velocity may be derived by setting the phase value to a constant:
ωt C
ωt − β z = C ⇒ z= − . (2.64)
β β
For a wave traveling in the z direction through vacuum, the velocity of the wave
equals the time derivative in z:
dz ω 1
vp = = =√ . (2.65)
dt β με

Thus, we have derived the result of (2.28). In vacuum, the phase velocity is v p = c.
For conductive media, the derivation is somewhat different because the elec-
tric field intensity j = σ E is now not zero. Thus, Ampere’s law is given by

∇× H = σ E + iωε E. (2.66)

i i

i i
 i i

i i

36 2. Physics of Light

Rewriting this equation, we have

 
iσ
∇× H = iω ε − E. (2.67)
ω
The previous solution for dielectric materials may therefore be extended to in-
iσ
clude conductive media by substitution of ε by ε − . As a result, the solution
ω
of (2.58) should be rewritten as
−i γ z iγ z
Ex = E+
me + E−
me , (2.68)
where γ is complex:
 
iσ
γ = iω μ ε − . (2.69)
ω
The real and complex parts of γ are given by

√
σ 2
ω με
α = Re{γ } = √ 1+ − 1, (2.70a)
2 ωε

√
σ 2
ω με
β = Im{γ } = √ 1+ + 1. (2.70b)
2 ωε
The complex form of the electric field is therefore given by (compare with (2.58))
−α z −i β z α z iβ z
Ex = E+
me e + E−
me e . (2.71)
As before, the real part of this complex solution is of interest. By splitting the
+
into a real part and a phase part, i.e., E+ + i φ + , and
complex amplitude Em m = Em e
−
doing the same for Em , the real part of the solution is given by
 
Ex (z,t) = Em+ e−α z cos ω t − β z + φ + (2.72a)
− αz
 −

+ Em e cos ω t + β z + φ (2.72b)
It is easy to observe that this result is an extension of the dielectric case discussed
earlier in this section. In particular, for dielectrics with σ = 0, the solution is real
because α = 0. In addition, for dielectrics the value of β is the same as before.
For a wave propagating along the +z-axis, we see that its amplitude is mod-
ulated by a factor e−α z . This means that its amplitude becomes smaller as the
wave progresses through the material. A consequence of this is that waves can
only penetrate a conductive medium by a short distance. As such, we note that
conductive materials are generally not transparent.
Conversely, for dielectric materials with very low conductivity, the value of
alpha tends to zero, so that the amplitude of waves propagating through dielectrics
is not significantly attenuated. Hence, dielectric materials tend to be transparent.

i i

i i
 i i

i i

2.2. Waves 37

{
1Å 1 nm
-10 -9 1m
λ -13 10 10 -6 0 5
10 10 10 10

{
1 THz 1 GHz 1 MHz 1 kHz
f
22 15 12 9 6 3
10 10 10 10 10 10

γ-r X- ult vis inf mi rad

ays ray rav ibl rar cro iow
s iol e li ed wa ave
et gh ves s
t

Figure 2.4. The electromagnetic spectrum.

2.2.5 The Electromagnetic Spectrum and Superposition

So far, we have discussed harmonic plane waves with a specific given wavelength.
Such waves are called monochromatic. The wavelength can be nearly anything,
ranging for instance from 10−13 m for gamma rays to 105 m for radio frequen-
cies [447]. This range of wavelengths is known as the electromagnetic spectrum
and is presented in Figure 2.4.
The range of wavelengths that the human eye is sensitive to is called the visible
spectrum and ranges from roughly 380 nm to 780 nm. This range constitutes only
a very small portion of the electromagnetic spectrum and coincides with the range
of wavelengths that interacts sensibly with materials. Longer wavelengths have
lower energies associated with them, and their interaction with materials is usually
limited to the formation of heat. At wavelengths shorter than the visible range, the
energy is strong enough to ionize atoms and permanently destroy molecules [815].
As such, only waves with wavelengths in the visible range interact with the atomic
structure of matter in such a way that it produces color.
So far, our discussion has centered on monochromatic light; the shape of the
wave was considered to be sinusoidal. In practice, this is rarely the case. Ac-
cording to the Fourier theorem, more complicated waves may be constructed by
a superposition of sinusoidal waves of different wavelengths (or, equivalently, of
different frequencies). This leads to the superposition principle, which is given
here for a dielectric material (compare with (2.59c)):
 ∞  
Ex (z,t) = E+ m,ω
−
cos(ω t − β z) + Em,ω cos(ω t + β z) d ω . (2.73)
ω =0

In addition, radiators tend to emit light at many different wavelengths. In Fig-

ure 2.5, an example is shown of the wavelength composition of the light emitted
by a tungsten radiator heated to 2000 K.

i i

i i
 i i

i i

38 2. Physics of Light

Emissivity
0.48
Emissivity of tungsten
0.47 at 2000 (K)
0.46

0.45

0.44

0.43

0.42

0.41
300 350 400 450 500 550 600 650 700 750 800
Wavelength λ (nm)

Figure 2.5. The relative contribution of each wavelength to the light emitted by a tungsten
radiator at a temperature of 2000 K [653].

2.3 Polarization
We have already shown that harmonic waves are transversal: both E and H lie
in a plane perpendicular to the direction of propagation. This still leaves some
degrees of freedom. First, both vectors may be oriented in any direction in this
plane (albeit with the caveat that they are orthogonal to one another). Further,
the orientation of these vectors may change with time. Third, their magnitude
may vary with time. In all, the time-dependent variation of E and H leads to
polarization, as we will discuss in this section.
We continue to assume without loss of generality that a harmonic plane wave
is traveling along the positive z-axis. This means that the vectors E and H may be
decomposed into constituent components in the x- and y-directions:
 
E = E x ex + E y ey e−i β z . (2.74)

Here, ex and ey are unit normal vectors along the x- and y-axes. The complex
amplitudes E x and E y are defined as

E x = |E x | ei ϕx , (2.75a)
i ϕy
E y = |E y | e . (2.75b)

The phase angles are therefore given by ϕx and ϕy . For a given point in space z =
r · s, as time progresses the orientation and magnitude of the electric field intensity
vector E will generally vary. This can be seen by writing Equations (2.75) in their

i i

i i
 i i

i i

2.3. Polarization 39

real form:
Ex
= cos(ω t − β z + ϕx ), (2.76a)
|E x |
Ey
= cos(ω t − β z + ϕy ). (2.76b)
|E y |
It is now possible to eliminate the component of the phase that is common to both
of these equations, i.e., ω t − β z, by rewriting them in the following form (using
identity (B.7a); see Appendix B):
Ex
= cos(ω t − β z) cos(ϕx ) − sin(ω t − β z)sin(ϕx ), (2.77a)
|E x |
Ey
= cos(ω t − β z) cos(ϕy ) − sin(ω t − β z)sin(ϕy ). (2.77b)
|E y |
If we solve both equations for cos(ω t − β z) and equate them, we get
Ex Ey
cos(ϕy ) − cos(ϕx ) = sin(ω t − β z)
|E x | |E y |
× (sin(ϕy ) cos(ϕx ) − cos(ϕy ) sin(ϕx )) (2.78a)
= sin(ω t − β z) sin(ϕy − ϕx ). (2.78b)
Repeating this, but now solving for sin(ω t − β z) and equating the results, we find
Ex Ey
sin(ϕy ) − sin(ϕx ) = cos(ω t − β z) sin(ϕy − ϕx ). (2.79)
|E x | |E y |
By squaring and adding these equations, we obtain
   
Ex 2 Ey 2 Ex Ey
+ −2 cos(ϕy − ϕx ) = sin2 (ϕy − ϕx ). (2.80)
|E x | |E y | |E x | |E y |
This equation shows that the vector E rotates around the z-axis describing an
ellipse. The wave is therefore elliptically polarized. The axes of the ellipse do not
need to be aligned with the x- and y-axes, but could be oriented at an angle.
Two special cases exist; the first is when the phase angles ϕx and ϕy are sepa-
rated by multiples of π :
ϕy − ϕx = m π (m = 0, ±1, ±2, . . .). (2.81)
For integer values of m, the sine operator is 0 and the cosine operator is either +1
or −1 dependent on whether m is even or odd. Therefore, Equation (2.80) reduces
to    
Ex 2 Ey 2 Ex E y
+ = 2 (−1)m . (2.82)
|E x | |E y | |E x | |E y |

i i

i i
 i i

i i

40 2. Physics of Light

B
Wave propagation
E

Figure 2.6. An electromagnetic wave with wavelength λ is defined by electric vector E

and magnetic induction B, which are both orthogonal to the direction of propagation. In
this case, both vectors are linearly polarized.

The general form of this equation is either x2 +y2 = 2xy or x2 +y2 = −2xy. We are
interested in the ratio between x and y, as this determines the level of eccentricity
of the ellipse. We find this as follows:

x2 + y2 = 2xy. (2.83a)

By dividing both sides by y2 we get

x2 x
2
+1 = 2 . (2.83b)
y y

Solving for the ratio x/y yields

x
= 1. (2.83c)
y

Solving x2 + y2 = −2xy in a similar manner yields x/y = −1. We therefore find

that the ratio between Ex and Ey is

Ex Ey
= (−1)m . (2.84)
Ey Ex

As such, the ratio between x- and y-components of E are constant for fixed m.
This means that instead of inscribing an ellipse, this vector oscillates along a line.
Thus, when the phase angles ϕx and ϕy are in phase, the electric field intensity
vector is linearly polarized, as shown in Figure 2.6. The same is then true for the
magnetic vector H.

i i

i i
 i i

i i

2.3. Polarization 41

E
y

Wave
x propagation

Figure 2.7. For a circularly polarized wave, the electric vector E rotates around the Poynt-
ing vector while propagating. Not shown is the magnetic vector, which also rotates around
the Poynting vector while remaining orthogonal to E.

The second special case occurs when the amplitudes |E x | and |E y | are equal
and the phase angles differ by either π /2 ± 2 m π or −π /2 ± 2 m π . In this case,
(2.80) reduces to

Ex2 + Ey2 = ±|E x |2 . (2.85)

This is the equation of a circle, and this type of polarization is therefore called
circular polarization. If ϕy − ϕx = π /2 ± 2 m π the wave is called a right-handed
circularly polarized wave. Conversely, if ϕy − ϕx = −π /2 ± 2 m π the wave is
called left-handed circularly polarized. In either case, the field vectors inscribe a
circle, as shown for E in Figure 2.7.
The causes of polarization include reflection of waves off surfaces or scat-
tering by particles suspended in a medium. For instance, sunlight entering the
Earth’s atmosphere undergoes scattering by small particles, which causes the sky
to be polarized.
Polarization can also be induced by employing polarization filters. These fil-
ters are frequently used in photography to reduce glare from reflecting surfaces.
Such filters create linearly polarized light. As a consequence, a pair of such filters
can be stacked such that together they block all light.
This is achieved if the two filters polarize light in orthogonal directions, as
shown in the overlapping region of the two sheets in Figure 2.8. If the two filters
are aligned, then linearly polarized light will emerge, as if only one filter were
present. The amount of light Ee transmitted through the pair of polarizing filters

i i

i i
 i i

i i

42 2. Physics of Light

Figure 2.8. Two sheets of polarizing material are oriented such that together they block
light, whereas each single sheet transmits light.

Figure 2.9. A polarizing sheet in front of an LCD screen can be oriented such that all light
is blocked.

i i

i i
 i i

i i

2.3. Polarization 43

Figure 2.10. A polarizing sheet is oriented such that polarized laser light is transmitted.

is a function of the angle θ between the two polarizers and the amount of incident
light Ee,0 :
Ee = Ee,0 cos2 (θ ) . (2.86)
This relation is known as the Law of Malus [447, 730, 1128].
The same effect is achieved by placing a single polarizing filter in front of an
LCD screen, as shown in Figure 2.9 (see also Section 14.2). Here, a backlight
emits non-polarized light, which is first linearly polarized in one direction. Then
the intensity of each pixel is adjusted by means of a second variable polarization in
the orthogonal direction. Thus, the light that is transmitted through this sequence
of filters is linearly polarized. As demonstrated in Figure 2.9, placing one further
polarizing filter in front of the screen thus blocks the remainder of the light.
In addition, laser light is polarized. This can be shown by using a single
polarization filter to block laser light. In Figure 2.10, a sheet of polarizing material
is placed in the path of a laser. The sheet is oriented such that most of the light is
transmitted. Some of the light is also reflected. By changing the orientation of the
sheet, the light can be blocked, as shown in Figure 2.11. Here, only the reflecting
component remains.
Polarization is extensively used in photography in the form of filters that can
be attached to the camera. This procedure allows unwanted reflections to be re-

i i

i i
 i i

i i

44 2. Physics of Light

Figure 2.11. A polarizing sheet is oriented such that polarized laser light is blocked.

Figure 2.12. The LCD screen emits polarized white light, which undergoes further polar-
ization upon reflection dependent on the amount of stress in the reflective object. Thus, the
colorization of the polarized reflected light follows the stress patterns in the material.

i i

i i
 i i

i i

2.4. Spectral Irradiance 45

moved from scenes. For instance the glint induced by reflections off water can be
removed. It can also be used to darken the sky and improve its contrast.
In computer vision, polarization can be used to infer material properties us-
ing a sequence of images taken with a polarization filter oriented at different an-
gles [1250]. This technique is based on the fact that specular materials partially
polarize light.
Polarization also finds uses in material science, where analysis of the polar-
izing properties of a material provides information of its internal stresses. An
example is shown in Figure 2.12 where the colored patterns on the CD case are
due to stresses in the material induced during fabrication. They become visible
by using the polarized light of an LCD screen.

2.4 Spectral Irradiance

In Section 2.2.2, we argued that the Poynting vector of a plane wave points in
the direction of propagation. However, we did not discuss its magnitude. In
general, the vectors E and H vary extremely rapidly in both direction and time.
For harmonic plane waves, the magnitude of the Poynting vector may therefore
be seen as giving the instantaneous energy density.
In the case of a linearly polarized harmonic plane wave, the E and H fields
may be modeled as

E = E0 cos (r · s − ω t) , (2.87a)
H = H0 cos (r · s − ω t) . (2.87b)

The Poynting vector, given by S = E × H, is then

S = E0 × H0 cos2 (r · s − ω t) . (2.88)

The instantaneous energy density thus varies with double the angular frequency
(because the cosine is squared). At this rate of change, the Poynting vector does
not constitute a practical measure of energy flow.
However, it is possible to average the magnitude of the Poynting vector over
time. The time average  f (t)T of a function f (t) over a time interval T is given
by
 t+T /2
1
 f (t)T = f (t) dt. (2.89)
T t−T /2

i i

i i
 i i

i i

46 2. Physics of Light

For a harmonic function eiω t , such a time average becomes

  t+T /2
1
eiω t T
= eiω t dt (2.90a)
T t−T /2
1 iω t
iω T /2 
= e e − e−iω T /2 (2.90b)
iω T 
ωT
sin
2
= eiω t (2.90c)
ωT
2 
ωT
= sinc ei ω t (2.90d)
2
 
ωT
= sinc (cos(ω t) + i sin(ω t)) . (2.90e)
2
The sinc function tends to zero for suitably large time intervals, and the cosine
and sine terms average to zero as well, i.e., cos(ω t)T = sin(ω t)T = 0.
However, the Poynting vector for a harmonic plane wave varies with cos2 . It
can be shown that the time average of such a function is 1/2 for intervals of T
chosen to be large with respect to the period τ of a single oscillation:
 2 1
cos (ω t) T = T  τ. (2.91)
2
Applying this result to the Poynting vector of (2.88), we find that the magnitude
of the time-averaged Poynting vector is given by
1
 S T =
E0 × H0 . (2.92)
2
To derive an expression for the time-averaged Poynting vector, we will first relate
the magnitude of E to the magnitude of H. From (2.46) we have that

μ
E = s × H , (2.93a)
ε

ε
H = s × E . (2.93b)
μ
The fact that all three vectors involved in the above equations are orthogonal
means we may apply Equation (A.7), yielding

μ
E = s H , (2.93c)
ε

ε
H = s E . (2.93d)
μ

i i

i i
 i i

i i

2.5. Reflection and Refraction 47

By solving both expressions for s and equating them, we find the following
relation:
√ √
μ H = ε E . (2.94)

Using Equation (A.7) once more, the magnitude of the time-averaged Poynting
vector follows from (2.92):

1 ε0
 S T = E0 2 (2.95a)
2 μ0
ε0 c
= E0 2 . (2.95b)
2

An important quantity proportional to this magnitude is called irradiance4 Ee and

is given by [447]
ε0 c  2
Ee = E0 T . (2.96)
2
This quantity is also known as radiant flux density and is measured in W/m2 .
Under the assumption that the material is dielectric, linear, homogeneous, and
isotropic, this expression becomes

εv 2
Ee = E0 T . (2.97)
2

The irradiance is a measure of the amount of light illuminating a surface. It can be

seen as the average energy reaching a surface unit per unit of time. Irradiance is
one of the core concepts of the field of radiometry, and therefore lies at the heart
of all color theory. It is discussed further in Chapter 6.

2.5 Reflection and Refraction

So far, we have discussed harmonic plane waves traveling in free space. For many
applications in optics, computer graphics, and for the discussion of the causes of
colored materials following in Chapter 3, the behavior of light at the boundary of
two different media constitutes an important and interesting case. In the following
sections, we discuss the direction of propagation as well as the strength of the
reflected and refracted waves.
4 In the physics literature, irradiance is often indicated with the symbol I. However, we are using

the radiometric symbol Ee for consistency with later chapters.

i i

i i
 i i

i i

48 2. Physics of Light

n
s(i) s(i)
C
θi
CB
Medium 1 θi z=0 sin (θi ) = sin (θi ) =
AC
Medium 2 A B AD
θt sin (θt ) = sin (θt ) =
θt AB
z D
sin (θi ) sin (θt )
x -n s(t) v1
=
v2

Figure 2.13. Geometry associated with refraction.

2.5.1 Direction of Propagation

For now, we assume that a harmonic plane wave is incident upon a planar bound-
ary between two materials. At this boundary, the wave will be split into two
waves, one which is propagated back into the medium and one that enters the
second medium. The first of these two is thus reflected, whereas the second is
refracted (or transmitted).
We will also assume that we have a coordinate system such that the boundary
is located at the z = 0 plane, and that the incident plane travels in a direction
indicated by S(i) . The direction vectors of the reflected and transmitted waves are
named S(r) and S(t) , respectively.
At any given location r = (x y 0) on the boundary, the time variation of each
of the three fields will be identical:

r · S(i) r · S(r) r · S(t)

t− =t− =t− . (2.98)
v1 v1 v2

As both the incident and the reflected wave propagate through the same medium,
their velocities will be identical. However, the refracted wave will have a different
velocity, since the medium has different permittivity and permeability constants.
The above equalities should hold for any point on the boundary. In particular, the
equalities hold for locations r1 = (1 0 0) and r2 = (0 1 0). These two locations
give us the following set of equalities:

(i) (r) (t)

sx sx sx
= = , (2.99a)
v1 v1 v2
(i) (r) (t)
sy sy sy
= = . (2.99b)
v1 v1 v2

i i

i i
 i i

i i

2.5. Reflection and Refraction 49

n
s(i) s(r)
θi θr
Medium 1 z=0
sin (θi ) sin (θr)
Medium 2 =
θt v1 v1
z
sin (θi ) sin (θt )
x -n s(t) v1
=
v2

Figure 2.14. Geometry associated with reflection.

If we assume that the Poynting vectors for the incident, reflected, and trans-
mitted waves lie in the x-z plane, then by referring to Figures 2.13 and 2.14, we
have
⎛ ⎞
sin(θi )
S(i) = ⎝ 0 ⎠ , (2.100a)
cos(θi )
⎛ ⎞
sin(θr )
S(r) = ⎝ 0 ⎠ , (2.100b)
cos(θr )
⎛ ⎞
sin(θt )
S(t) = ⎝ 0 ⎠ . (2.100c)
cos(θt )

The z-coordinates are positive for S(i) and S(t) and negative for S(r) . By combin-
ing (2.99) and (2.100) we find that

sin(θi ) sin(θr ) sin(θt )

= = . (2.101)
v1 v1 v2

Since sin(θi ) = sin(θr ) and the z-coordinates of S(i) and S(r) are of opposite sign,
the angle of incidence and the angle of reflection are related by θr = π − θi . This
relation is known as the law of reflection.
Since the speed of light in a given medium is related to the permittivity and
permeability of the material according to (2.28), we may rewrite (2.101) as

sin(θi ) v1 μ2 ε2 n2
= = = = n. (2.102)
sin(θt ) v2 μ1 ε1 n1

i i

i i
 i i

i i

50 2. Physics of Light

Figure 2.15. Refraction demonstrated by means of a laser beam making the transition
from smoke-filled air to water.

Figure 2.16. Laser light reflecting and refracting off an air-to-water boundary.

i i

i i
 i i

i i

2.5. Reflection and Refraction 51

Figure 2.17. Laser light interacting with a magnifying lens. Note the displacement of the
transmitted light, as well as the secondary reflections of front and back surfaces of the lens.

√ √
The values n1 = μ1 ε1 and n2 = μ2 ε2 are called the absolute refractive indices
of the two media, whereas n is the relative refractive index for refraction from the
first into the second medium. The relations given in (2.102) constitute Snell’s
law.5
An example of refraction is shown in Figure 2.15, where a laser was aimed
at a tank filled with water. The laser is a typical consumer-grade device normally
used as part of a light show in discotheques and clubs. A smoke machine was used
to produce smoke and allow the laser light to scatter towards the camera. For the
same reason a few drops of milk were added to the tank. Figure 2.16 shows a
close-up with a shallower angle of incidence. This figure shows that light is both
reflected and refracted.
A second example of Snell’s law at work is shown in Figure 2.17 where a
laser beam was aimed at a magnifying lens. As the beam was aimed at the center
of the lens, the transmitted light is parallel to the incident light, albeit displaced.
The displacement is due to the double refraction at either boundary of the lens.
Similarly, the figure shows light being reflected on the left side of the lens. Two
beams are visible here: one reflected off the front surface of the lens and the
5 Named after the Dutch scientist Willebrord van Roijen Snell (1580–1626).

i i

i i
 i i

i i

52 2. Physics of Light

second beam reflected off of the back surface. These beams are predicted by a
straightforward repeated application of Snell’s law.
The refraction diagram shown in Figure 2.13 shows a wave traveling from an
optically less dense medium into a more dense medium, and hence θt < θi . If
we increase the angle of incidence until, in the limit, we reach θi = π /2, then the
wave is traveling parallel to the boundary and S(i) · n = 0. For this situation, there
will be a maximum angle of refraction, θt = θc . This angle is called the critical
angle.
A wave traveling in the opposite direction, i.e., approaching the boundary
from the denser medium towards the less dense medium, refraction into the less
dense medium will only occur if the angle of incidence is smaller than the crit-
ical angle. For angles of incidence larger than θc , the wave will not refract but
will only reflect. This phenomenon is called total reflection, or sometimes total
internal reflection, and is illustrated in Figure 2.18. It can be demonstrated by
punching a hole in a plastic bottle and draining water through it. By aligning a
laser with the opening in the bottle and adding a drop of milk (to make the laser
beam visible), total internal reflection such as that seen in Figure 2.19 can be ob-
tained. The light stays within the water stream through multiple reflections. The
transmittance of light through strands of fiber, known as fiberoptics, works on the
same principle (Figure 2.20) .
While Snell’s law predicts the angles of reflection and transmission, it does
not give insight into the amplitude of the reflected and refracted fields, E and B,
respectively. To evaluate these amplitudes, we first need to look at the boundary
conditions related to the interface between the two media, which are most easily
evaluated by means of Maxwell’s equations in integral form. The components
of D and B normal to the boundary are assessed first, followed by the tangential
components of these fields.6
n θi < θc : Reflection + Refraction
s(t)
θi > θc : Total Reflection

θt
Medium 1 z=0
Medium 2 θi
θr θc
s(r) θr θi s(i)
z -n
x s(r) s(i)

Figure 2.18. Geometry associated with total reflection.

6 As D = ε E, the analysis is equivalent.

i i

i i
 i i

i i

2.5. Reflection and Refraction 53

Figure 2.19. A stream of water exits a plastic bottle. Coming from the right, a laser beam
enters the bottle and exits through the same hole as the water. Multiple reflections inside
the stream of water are visible, essentially keeping the light trapped inside the stream.

Figure 2.20. Fiberoptic strands transmitting light with very little loss of energy due to
total internal reflection inside the strands.

2.5.2 Strength of Reflected and Refracted Fields

We assume that the electric and magnetic fields across the boundary are not dis-
continuous, but merely have a sharp gradient. If we construct a pill box that in-

i i

i i
 i i

i i

54 2. Physics of Light

δA1
n

δh

Medium 1
Medium 2 -n
δA 2

Figure 2.21. Pill box intersecting the boundary between two media.

tersects the boundary between two dielectric materials, as shown in Figure 2.21,
then the fields gradually change from the top of the volume to the bottom, which
is located on the other side of the boundary.
By Gauss’ theorem (see Appendix A.5), the divergence of B integrated over
the volume of the pill box is related to the integral of the normal component of B,
when integrated over the surface of the pill box:
 
∇·B dv = B · n ds = 0. (2.103)
v s

In the limit that the height of the pill box goes to zero, contribution of the cylin-
drical side of the pill box also goes to zero. If the areas of the top and bottom
of the pill box (δ A1 and δ A2 ) are small and there is no flux or induction at the

surface, then the above surface integral simplifies to

B(1) · nδ A1 − B(2) · nδ A2 = 0, (2.104a)


B(1) − B(2) · n = 0, (2.104b)

where δ A = δ A1 = δ A2 . This equation shows that the component of the magnetic

induction B normal to the boundary between the two media is continuous across
the media.
The normal component of the electric displacement D is analyzed similarly:
  
D dv = D · n ds = ρ dv. (2.105)
v s v

In the limit that the pill box is shrunk to zero height, the charge density ρ , i.e., the
charge per unit of volume, goes to infinity. For this reason, the concept of surface

i i

i i
 i i

i i

2.5. Reflection and Refraction 55

δs1
n
t
δh
t -t
Medium 1 b
Medium 2 δs2

Figure 2.22. Plane intersecting the boundary between two media.

charge density ρ̂ is introduced as follows:

 
lim ρ dv = ρ̂ dA. (2.106)
δ h→0 v s

The normal component of the electric displacement D may therefore be derived

from

D(1) − D(2) · n = ρ̂ . (2.107)

This means that upon crossing a boundary between two dielectrics, the surface
charge density causes the normal component of D to be discontinuous and jump
by ρ̂ .
To analyze the tangential components of E and H, we can make use of Stokes’
theorem (Appendix A.7), which relates line integrals to surface integrals. We
place a plane perpendicular to the surface boundary that intersects the bound-
ary, as shown in Figure 2.22. Note also that the incident, reflected and refracted
Poynting vectors are assumed to lie in this plane, as shown in Figure 2.23.
In Figure 2.22, n is the normal of the boundary between the two dielectrics, b
is the normal of the newly created plane, and t = b × n is a vector that lies both

Plane of incidence
s (r) z=0
n s (i)

t
Medium 1 b
Medium 2 s (t)

Figure 2.23. Plane (green) intersecting the boundary between two media. The Poynting
vectors of incident, reflected, and refracted TEM waves all lie in this plane.

i i

i i
 i i

i i

56 2. Physics of Light

in the new plane and in the plane of the surface boundary. All three vectors are
assumed to be unit length. Stokes’ theorem then gives
  
∂B
∇× E · b ds = E · b dc = − · b ds, (2.108)
s c s ∂t
where the middle integral is a line integral over the contour c that surrounds the
plane. The other two integrals are surface integrals over the area of the plane. If
δ s1 and δ s2 are small, (2.108) simplifies as follows:
∂B
E(2) · t δ s2 − E(1) · t δ s1 = − · b δ s δ h. (2.109)
∂t
In the limit that the height δh goes to zero, we get


E(2) − E(1) · t δ s = 0. (2.110)

By noting that t = b × n and −t = −b × n, this equation can be rewritten as


b · n × E(2) − E(1) = 0. (2.111)

This result indicates that at the boundary between two dielectrics, the tangential
component of the electric vector E is continuous. A similar derivation can be
made for the magnetic vector, which yields


n × H(2) − H(2) = ĵ, (2.112)

where ĵ is the surface current density (introduced for the same reason ρ̂ was
above). Thus, the tangential component of the magnetic vector H is discontin-
uous across a surface and jumps by ĵ. Of course, in the absence of a surface
current density, the tangential component of vector H becomes continuous.
The above results for the continuity and discontinuity of the various fields will
now be used to help derive the Fresnel equations, which predict the amplitudes
of the transmitted and reflected waves. To facilitate the derivation, the vector E at
the boundary between the two dielectrics may be decomposed into a component
parallel to the plane of incidence, and a component perpendicular to the plane of
incidence, as indicated in Figure 2.24. The magnitudes of these components are
termed A and A⊥ . The x and y components of the incident electric field are then
given by
(i)
Ex = − A cos(Θi ) e−i τi , (2.113a)
(i)
Ey = A⊥ e−i τi , (2.113b)
(i)
Ez = A sin(Θi ) e−i τi , (2.113c)

i i

i i
 i i

i i

2.5. Reflection and Refraction 57

E s (i)
A = E projected onto P
Perpendicular plane Q A = E projected onto Q
Plane of incidence P
s (i)
n E z=0
A A
t
b
Medium 1
Medium 2

Figure 2.24. The electric vector E can be projected onto the plane of incidence as well
as onto a perpendicular plane. The magnitudes of the projected vector are indicated by A
and A⊥ .

with τi equal to the variable part of the wave function:

   
r · S(i) x sin(Θi ) + y cos(Θi )
τi = ω t − =ω t− . (2.114)
v1 v1

As for dielectric materials, the magnetic permeability is close to one (μ = 1),

and therefore we can derive the components of the magnetic vector H from E by
applying (2.46b):
√
H = ε S × E, (2.115)
so that we have
(i) √
Hx = − A⊥ cos(Θi ) ε1 , e−i τi , (2.116a)
(i) √
Hy = − A ε1 e−i τi , (2.116b)
(i) √
Hz = A⊥ sin(Θi ) ε1 e−i τi . (2.116c)

With τr and τt equal to

 
x sin(Θr ) + y cos(Θr )
τr = ω t − , (2.117a)
v1
 
x sin(Θt ) + y cos(Θt )
τt = ω t − , (2.117b)
v2

and R and T the complex amplitudes of the reflected and refracted waves, ex-
pressions for the reflected and transmitted fields may be derived analogously.

i i

i i
 i i

i i

58 2. Physics of Light

The reflected field is then given by

(r) (r) √
Ex = − R cos(Θr ) e−i τr Hx = − R⊥ cos(Θr ) ε1 , e−i τr , (2.118a)
(r) (r) √
Ey = R⊥ e−i τr Hy = − R ε1 e−i τr , (2.118b)
(r) (r) √
Ez = R sin(Θr ) e−i τr Hz = R⊥ sin(Θr ) ε1 e−i τr , (2.118c)

and the transmitted field is

(t) (t) √
Ex = − T cos(Θt ) e−i τt Hx = − T ⊥ cos(Θt ) ε2 , e−i τt , (2.119a)
(t) (t) √
Ey = T ⊥ e−i τt Hy = − T ε2 e−i τt , (2.119b)
(t) (t) √
Ez = T sin(Θt ) e−i τt Hz = T ⊥ sin(Θt ) ε2 e−i τt . (2.119c)

According to (2.111) and (2.112) and under the assumption that the surface cur-
rent density is zero7 , the tangential components of E and H should be zero at the
surface boundary. We may therefore write

(i) (r) (t) (i) (r) (t)

E x + E x = Ex Hx + Hx = Hx , (2.120a)
(i) (r) (t) (i) (r) (t)
E y + Ey = Ey Hy + Hy = Hy . (2.120b)

By substitution of this equation, and using the fact that cos(Θr ) = cos(π − Θr ) =
− cos(Θr ), we get


cos(Θi ) A − R = cos(Θt ) T , (2.121a)

ε2
A + R = T , (2.121b)
ε1

ε2
cos(Θi ) (A⊥ − R⊥ ) = cos(Θt ) T ⊥ , (2.121c)
ε1
A⊥ + R⊥ = T ⊥ . (2.121d)

7 This is a reasonable assumption for dielectric materials.

i i

i i
 i i

i i

2.5. Reflection and Refraction 59

√ √
We use Maxwell’s relation n = μ ε = ε and solve the above set of equations
for the components of the reflected and transmitted waves, yielding

n2 cos(Θi ) − n1 cos(Θt )
R = A , (2.122a)
n2 cos(Θi ) + n1 cos(Θt )
n1 cos(Θi ) − n2 cos(Θt )
R⊥ = A⊥ , (2.122b)
n1 cos(Θi ) + n2 cos(Θt )
2n1 cos(Θi )
T = A , (2.122c)
n2 cos(Θi ) + n1 cos(Θt )
2n1 cos(Θi )
T ⊥ = A⊥ . (2.122d)
n1 cos(Θi ) + n2 cos(Θt )

These equations are known as the Fresnel equations; they allow us to compute
the transmission coefficients t and t⊥ as follows:

T 2 cos(Θi )
t = = n2 , (2.123a)
A cos(Θi ) + cos(Θt )
n1
T⊥ 2 cos(Θi )
t⊥ = = n2 . (2.123b)
A⊥ cos(Θi ) + cos(Θt )
n1
The reflection coefficients r and r⊥ are computed similarly:
n1
R cos(Θi ) −
cos(Θt )
n2
r = = n1 , (2.124a)
A cos(Θi ) + cos(Θt )
n2
n2
cos(Θi ) − cos(Θt )
R⊥ n1
r⊥ = = n2 . (2.124b)
A⊥ cos(Θi ) + cos(Θt )
n1
We may rewrite both the reflection and the transmission coefficients in terms of
the angle of incidence alone. This can be achieved by applying Snell’s law (2.102),
so that
n1
sin (Θt ) = sin (Θi ) (2.125)
n2
and therefore, using Equation (B.11), we have
 2
n1
cos (Θt ) = 1 − sin2 (Θi ) (2.126)
n2

i i

i i
 i i

i i

60 2. Physics of Light

By substitution into (2.123) and (2.124) expressions in n1 , n2 , and cos (Θi ) can
be obtained. The reflection and transmission coefficients give the percentage of
incident light that is reflected or refracted:
E0,r
r= (2.127a)
E0,i
E0,t
t= (2.127b)
E0,i
In the case of a wave incident upon a conducting (metal) surface, light is both
reflected and refracted as above, although the refracted wave is strongly attenuated
by the metal. This attenuation causes the material to appear opaque.
A water surface is an example of a reflecting and refracting boundary that
occurs in nature. While all water reflects and refracts, on a wind-still day the

Figure 2.25. On a wind-still morning, the water surface is smooth enough to reflect light
specularly on the macroscopic scale; Konstanz, Bodensee, Germany, June 2005.

i i

i i
 i i

i i

2.5. Reflection and Refraction 61

boundary between air and water becomes smooth on the macroscopic scale, so
that the effects discussed in this section may be observed with the naked eye, as
shown in Figure 2.25.

2.5.3 Reflectance and Transmittance

If we assume that a beam of light is incident upon a dielectric surface at a given
angle Θi , and that the cross-section of the beam has an area A, then the incident
energy per unit of time Pe,i is given by

Pe,i = Ee,i A cos (Θi ) , (2.128)

where Ee,i is the radiant flux density of Equation (2.96). Energy per unit of time
is known as radiant flux and is discussed further in Section 6.2.2. Similarly, we
can compute the reflected and transmitted flux (Pe,r and Pe,t ) with

Pe,r = Ee,r A cos (Θr ) , (2.129a)

Pe,t = Ee,t A cos (Θt ) . (2.129b)

The reflectance R is defined as the ratio of reflected to incident flux:

Pe,r Ee,r A cos (Θr )
R= = . (2.130)
Pe,i Ee,i A cos (Θi )
Given that the angle of incidence and the angle of reflectance are identical for
specular reflection, this expression simplifies to
Ee,r
R= . (2.131)
Ee,i
Using (2.96), this expression can be rewritten as
εr vr 2
E0,r
R = ε2v . (2.132)
i i
E20,i
2
As the incident and reflected beams are in the same medium, we have εi = εr and
vi = vr , so that the reflectance becomes
 
E0,r 2
R= = r2 . (2.133)
E0,i
The reflectance can be split into its component forms, leading to

R = r 2 , (2.134a)
R⊥ = r⊥
2
. (2.134b)

i i

i i
 i i

i i

62 2. Physics of Light

The transmittance T is defined as the ratio between incident flux to transmitted

flux:
Pe,i Ee,i A cos (Θi )
T= = . (2.135)
Pe,t Ee,t A cos (Θt )
Through similar arguments, the transmittance can be derived to be
n2 cos (Θt ) 2
T= t , (2.136a)
n1 cos (Θi )
n2 cos (Θt ) 2
T = t , (2.136b)
n1 cos (Θi )
n2 cos (Θt ) 2
T⊥ = t . (2.136c)
n1 cos (Θi ) ⊥
Under the assumption that no absorbtion takes place, all light is either reflected or
transmitted, so that we have
R + T = 1, (2.137a)
R + T = 1, (2.137b)
R⊥ + T⊥ = 1. (2.137c)
An especially important case arises when the angle of incidence is perpen-
dicular to the boundary between the two media. In this case Θi becomes 0, and
therefore Θt also approaches 0. In that case, the distinction between parallel and
perpendicular components vanishes, so that R = R⊥ = R and T = T⊥ = T . The
resulting reflectance and transmittance is then [447]
 
n2 − n1 2
R= , (2.138a)
n2 + n1
4 n1 n2
T= . (2.138b)
(n1 + n2 )2
Finally, in the case that the surrounding medium is air, the index of refraction may
be taken to approximate n1 ≈ 1, so that the reflectance at normal angles is
 
n2 − 1 2
R≈ , (2.139a)
n2 + 1
4 n2
T≈ . (2.139b)
(n2 + 1)2
For instance, for glass having an index of refraction of around 1.5, this means
that approximately 4% of light is reflected. This can become a problem in lens
design, where compound lenses are made of several air-glass interfaces. The
amount of light reflected by a dielectric can be reduced by applying a thin coat of
anti-reflective material. This is discussed further in Section 12.6.

i i

i i
 i i

i i

2.6. Birefringence 63

Figure 2.26. A calcite crystal causes birefringence which depends on the orientation of
the crystal.

2.6 Birefringence
Some materials have an index of refraction that depends on the polarization of the
light. This means that unpolarized light can be split into two or more polarized
light paths. This effect is called birefringence or double refraction and is demon-
strated in Figure 2.26 by means of a calcite crystal (also known as iceland spar).
A second example of a birefringent material is the cornea of the human eye (see
Section 4.2.2) [755].
To see that light becomes linearly polarized into two orthogonal directions,
Figure 2.27 contains photographs of the same calcite crystal placed upon a detail
of Figure 1.7. The light is linearly polarized, which is demonstrated by placing
a linear polarization filter on top of the crystal. This allows one of the two lin-
early polarized light directions to pass through, while blocking the other. As a
consequence, slowly rotating either the filter or the crystal, will suddenly change
the path selected to pass through the polarizer. The combination of linear polariz-
ers and liquid birefringent materials, which have polarizing behavior that can be
controlled by applying an electric field, forms the basis of liquid crystal displays
(LCDs), which are discussed further in Section 14.2.
Molecules with an elongated shape tend to give rise to birefringence. Such
rod-like molecules are called calamitic. To create birefringence in bulk, these
molecules must be aligned, so that the material itself will exhibit birefringence.
The vector pointing along the long axis of the molecules is called the director.
In that case, the relative permittivity (i.e., the dielectric constant) of the material

i i

i i
 i i

i i

64 2. Physics of Light

Figure 2.27. The block of birefringent calcite is placed on a print of Figure 1.7 (top),
showing double refraction. The bottom two images are photographs of the same configu-
ration, except that now a linear polarizing filter is placed on top of the crystal. Dependent
on the orientation (which is the difference between the bottom two photographs), one or
the other refraction remains. A polarizer used in this manner is called an analyzer.

is dependent on direction, yielding ε for directions parallel to the director and

ε⊥ for all directions orthogonal to the director. The dielectric anisotropy is then
given by Δε :
Δε = ε − ε⊥ . (2.140)
As the index of refraction of a material depends on its permittivity, in the di-
rection of the director the material will have an index of refraction of n , whereas
in orthogonal directions it is given by n⊥ . The optical anisotropy Δn is then

Δn = n − n⊥ . (2.141)

We now consider the amplitude of the electric field intensity in two orthogonal
directions x and y, as given in (2.76), with x aligned with the director. Recall that
the distance z in this equation is only a convenient notation representing a point r
in space under a plane wave traveling in direction S. In other words, in the more
general case we have β z = β r · S. With β given by
ω 2π n
β= = , (2.142)
c λ

i i

i i
 i i

i i

2.6. Birefringence 65

the wave vector k is given by

2π n
k= S. (2.143)
λ
Thus, the wave vector k is nothing more than S scaled by an amount dependent
on wavelength λ and index of refraction n.
With indices of refraction that are different for the x and y directions, (2.76)
becomes
Ex
= cos(ω t − k r + ϕx ), (2.144a)
|E x |
Ey
= cos(ω t − k⊥ r + ϕy ), (2.144b)
|E y |

with
2 π n
k = S, (2.145a)
λ
2 π n⊥
k⊥ = S. (2.145b)
λ
In the case that the director makes an angle, say ψ , with our choice of coordinate
system, a further scaling of the electric field magnitudes is required:
Ex
= cos(ψ ) cos(ω t − k r + ϕx ), (2.146a)
|E x |
Ey
= sin(ψ ) cos(ω t − k⊥ r + ϕy ). (2.146b)
|E y |

For computations involving liquid crystals, it is often convenient to remove

all components that do not involve r, and add them back in afterwards. Thus, we
may analyze the complex phasors Px and Py :
 
Px = Ex cos(ψ ) exp −ik r , (2.147a)
Py = Ey sin(ψ ) exp (−ik⊥ r) , (2.147b)

whereby we momentarily ignore the factors exp(iω t) and exp(iϕ ). Absorbing the
scaling according to ψ into the constants Ex and Ey and placing the two phasors
in a column vector, we obtain the so-called Jones vector J:
    
Jx Ex,ψ exp −ik r
J= = . (2.148)
Jy Ey,ψ exp (−ik⊥ r)

The use of the Jones vector will return in the analysis of liquid crystals in
Section 14.2.

i i

i i
 i i

i i

66 2. Physics of Light

2.7 Interference and Diffraction

The emphasis of this chapter is on the wave nature of light. One of the most
compelling demonstrations of wave behavior is afforded by two closely related
phenomena, interference and diffraction.
We are modeling light with time-harmonic electromagnetic waves. The su-
perposition principle discussed in Section 2.2.5 shows that multiple waves may
be summed to determine a compound solution at any given position in space. For
waves with random phases, the result yields an average intensity that does not
vary appreciably with space, except for the attenuation caused by the medium of
propagation. This assumption is, for instance, made in geometrical optics, which
will be discussed in Section 2.9.

2.7.1 Interference
If the phase function of each of the waves is correlated, we speak of coherent
light. When such waves are superposed, the result shows peaks and troughs as a
result of the sinusoidal variation of both electric and magnetic vectors, as shown
in Figure 2.28. If the two waves are out of phase by π , then the two waves exactly
cancel. Thus, dependent on the phase difference at a given point in space, the
two waves may reinforce each other or cancel each other. It is said that coherent
waves interfere, and the resulting pattern is called an interference pattern.

The irradiance at a point in space is given by (2.96), where E20 T is the time
average of the magnitude of the electric field vector. Applying the superposition
principle for a pair of coherent monochromatic time-harmonic plane waves with

2.0
cos (x)
1.5 cos (x+0.5)
cos (x) + cos (x+0.5)
1.0
0.5
0.0
-0.5
−1.0
−1.5
−2.0
0 2 4 6 8 10 12 14
x

Figure 2.28. Two waves superposed forming a new wave, in this case with a larger ampli-
tude than either of the constituent waves.

i i

i i
 i i

i i

2.7. Interference and Diffraction 67

electric field vectors E1 and E2 , the irradiance becomes

 
Ee = c ε0 (E1 + E2 )2 . (2.149)
T

Since the electric vector is now the sum of two components, the square of its
magnitude is given by

(E1 + E2 )2 = E21 + E22 + 2 E1 · E2 . (2.150)

The implications for the irradiance are that this quantity is composed of three
terms, namely the irradiance due to either wave as well as a cross term:
  
Ee = ε0 c E21 T + E22 T + 2 E1 · E2 T . (2.151)

The cross term 2 E1 · E2 T , known as the interference term, explains the patterns
that may be observed when two coherent bundles of light are superposed.
We assume that the electric vectors of the two waves are modeled by

E1 (r,t) = a1 cos (r · S1 − ω t + ϕ1 ) , (2.152a)

E2 (r,t) = a2 cos (r · S2 − ω t + ϕ2 ) , (2.152b)

where ai are the amplitudes and ϕi are the initial phases of the two (linearly po-
larized) waves. The interference term is then

a1 · a2 [cos(r · S1 + ϕ1 ) cos(ω t) + sin(r · S1 + ϕ1 ) sin(ω t)]

× [cos(r · S2 + ϕ2 ) cos(ω t) + sin(r · S2 + ϕ2 ) sin(ω t)] . (2.153)

It can be shown that the time-averaged quantity of the interference term becomes

2 E1 · E2 T = a1 · a2 cos (r · S1 + ϕ1 − r · S2 − ϕ2 ) . (2.154)

The interference term can then be rewritten in the following form to highlight the
fact that its oscillations depend on the phase difference between the two waves:

2 E1 · E2 T = a1 · a2 cos (δ ) , (2.155)

where the phase difference δ equals r · S1 + ϕ1 − r · S2 − ϕ2 . This result shows
that the phase difference is due to a difference in initial phase, ϕ1 − ϕ2 , as well as
a difference in path length between the origins of the two waves and the point of
interest r (Figure 2.29).
The irradiance at r is maximum when cos(δ ) = 1 and occurs when the phase
difference is a multiple of 2 π :

δ = 0, ±2π , ±4π , . . . . (2.156)

i i

i i
 i i

i i

68 2. Physics of Light

s2
ϕ2

ϕ1 r
s1

Figure 2.29. The phase at point r is determined by the phases at the origins of the two
waves, as well as the distance traveled.

In this situation, we have total constructive interference. The irradiance attains a

minimum when the phase difference is

δ = ±π , ±3π , ±5π , . . . . (2.157)

In this case, there is total destructive interference.

So far we are assuming that the waves are planar and that they correspond to a
source that is infinitely far away. This, of course, is rarely the case although many
interesting theoretical results may be derived in this way. However, if the source
of radiation is not infinitely far away, it is better to assume a spherical wavefront,
i.e., a wave originating at a point in space and traveling outward in all directions,
as shown in Figure 2.30. Such a wavefront may be modeled by
 
2π r
E(r,t) = a(r) exp i − ωt + ϕ . (2.158a)
λ

Here, r is the distance between the point of interest and the source of the wave-
front, i.e., the radius. A pair of spherical wavefronts may show interference just
as planar waves would. The phase difference δ of two spherical wavefronts E1

Figure 2.30. Radial wavefront due to a point source.

i i

i i
 i i

i i

2.7. Interference and Diffraction 69

and E2 is then
2π
δ= (r1 − r2 ) + (ϕ1 − ϕ2 ) . (2.159)
λ
If for simplicity we assume the amplitudes a1 and a2 to be identical and constant
over the region of interest, it can be shown that the irradiance depends on the
phase difference as follows [447]:
 
δ
Ee = 4Ee,0 cos2 , (2.160)
2
where Ee,0 is the irradiance due to either source (which in this case are identical).
Maxima in the irradiance Ee occur at intervals,

δ = 0, ±2π , ±4π , . . . , (2.161)

and minima occur at

δ = ±π , ±3π , ±5π , . . . . (2.162)
Thus, maxima occur for integer multiples of 2π , i.e., when
2π
2π m = (r1 − r2 ) + (ϕ1 − ϕ2 ) m = 0, ±1, ±2, . . . . (2.163)
λ
This expression can be simplified to
λ (ϕ1 − ϕ2 )
r1 − r2 = − mλ . (2.164)
2π
If the two sources are emitting light in phase, then ϕ1 = ϕ2 , and we obtain

r2 − r 1 = mλ . (2.165)

This result indicates that the location of peaks and troughs in the interference pat-
tern depends on the wavelength of the two sources, which in our simplified anal-
ysis are assumed to be in phase and emit light of the same wavelengths. Thus, if
the phase difference between two bundles of light can be coherently manipulated
by fractions of a wavelength, we can expect white light to be broken up into its
constituent parts.
Further, a thin layer which reflects off its front surface as well as its back sur-
face, can create interference patterns. This occurs because the path length from
the light source to the human observer is slightly shorter for the light reflected
off the front surface than it is for the light reflected off the back surface. As a
result, according to (2.164) this results in small phase differences. If the spacing
of the reflecting layers causes a phase difference in the order of a visible wave-
length, this is seen as interference color. As shown in Figure 2.31, the viewing

i i

i i
 i i

i i

70 2. Physics of Light

Incoming light VD 1 VD 2

r1 r2 The pathlengths of the two paths

r1 r2
r1 and r2 reaching view direction 1
(VD 1) are slightly different,
causing interference colors.
Front surface
Idem for VD 2, but the path-
length difference varies with
Back surface viewing angle, i.e. r2-r1 ≠ r2-r1
leading to different interference
Dark backing
colors.

Figure 2.31. Interference colors, which depend on viewing angle, are caused by a semi-
transparent layer creating two thinly spaced reflective layers. The path lengths of light
reflected off each surface are slightly different, causing interference colors.

angle has an impact on the relative path lengths, and therefore one can expect
the color of the surface to vary with viewing angle. The dark substrate serves to
create brighter and more brilliant colors. In practice, this effect is not limited to a
single transparent layer, but further reflections can be created by a stack of layers.

Figure 2.32. This Morpho zephyritis shows iridescent blue colors as a result of interfer-
ence in its wings.

i i

i i
 i i

i i

2.7. Interference and Diffraction 71

Figure 2.33. Iridescence in sea shells. Shown here is a close-up of the inside of an abalone
shell.

Recently, display devices have been proposed which are based on this principle,
as discussed in Section 14.10.
Interference colors occur in nature, for instance, in the wings of certain species
of butterfly such as the Morpho zephyritis shown in Figure 2.32. Colors resulting
from interference at wavelength scales are usually termed iridescent. Multiple
colors may be present and normally vary with viewing angle. In elytra, the wing-
cases of beetles (see Figure 1.5), as well as in butterfly wings, the layers causing
interference are backed by a dark layer, usually made of melanin, which strength-
ens the color. Iridescence is also present in many birds, including the peacock
(Figure 1.3) and hummingbirds, as well as sea shells such as abalones and oysters
(Figure 2.33).

Figure 2.34. The number 20 from a US $20 bill photographed under different viewing
angles. The particles in the ink are aligned to produce a distinct change in color with
viewing angle.

i i

i i
 i i

i i

72 2. Physics of Light

Figure 2.35. An interference pattern is etched into a small piece of plastic, which only
reveals itself when hit by a laser beam. (Etching courtesy of Eric G. Johnson, School of
Optics, University of Central Florida.)

A practical application of interference is shown in Figure 2.34. The ink used

to print the number 20 on US currency contains small particles that produce inter-
ference colors [447]. The color changes with viewing angle due to the structured
orientation of the particles. A second application of interference in optics is shown
in Figure 2.35. Here, a small piece of plastic is etched with a pattern that does not
appear to contain information (left side of figure). When a laser is aimed at it, an
entirely different pattern is formed, as seen in the same figure (right). Finally, an
algorithm for simulating iridescence was recently proposed [1107]. It allows the
simulation of, for instance, butterfly wings.

2.7.2 Diffraction
We have thus far discussed interference in the form of a pair of hypothetical light
sources that emit light of the same wavelength and are in phase. In this section
we discuss diffraction, which is the behavior of light in the presence of obstacles.
Diffraction may cause interference patterns.
In Section 2.5 the behavior of light near a boundary was discussed. In the case
of a boundary between a dielectric and a metal, the wave will not propagate very
far into the metal. The theory of diffraction deals with opaque boundaries with
edges and holes in different configurations.
If we assume that we have a thin opaque sheet of material with a hole in it,
then the interaction of the wave with the sheet away from the hole will be as
discussed in Section 2.5. In the center of the hole, and assuming that the hole
is large enough, the wave will propagate as if there were no opaque sheet there.
However, near the edge of the hole, diffraction occurs. Here the wave interacts
with the material, causing new wavefronts. A detailed treatment of this interaction
would involve atomic theory (touched upon in Chapter 3) and is beyond the scope

i i

i i
 i i

i i

2.7. Interference and Diffraction 73

Screen with slit Wall

D θ
R

Plane waves

Figure 2.36. Light passing through a slit is projected onto a wall. The distribution of
irradiance values is observed for points P along the wall.

of this book. We briefly outline the behavior of light near edges using a simplified
wave theory of light.
For a sheet of material with a hole, the visible manifestation of diffraction
depends on the size of the hole in relation to the wavelength of the incident light.
If the hole is very large, the sheet casts a shadow and the result is much like
everyday experience predicts. When the hole becomes small, the edge of the hole
will diffract light and cause interference patterns.
For a rectangular hole, i.e., a slit, we are interested in the irradiance distri-
bution as a result of diffraction. If the distance R between the slit and the line
of observation on a wall is large in comparison with the size of the slit (R  D,
see Figure 2.36), we speak of Fraunhofer diffraction. It can be shown that the
irradiance distribution along the wall as a function of angle θ is given by [447]
 
2 Dπ
Ee (θ ) = Ee (0) sinc sin(θ ) . (2.166)
λ

1.0
Ee(θ)/Ee(0)

0.8

0.6

0.4

0.2

0.0
−2 −1 0 1 2
θ

Figure 2.37. Irradiance as function of eccentricity angle θ at a distance R from the slit.
The size of the slit was chosen to be D = 10λ /π .

i i

i i
 i i

i i

74 2. Physics of Light

A plot of this function for D = 10λ /π is shown in Figure 2.37. The equation
shows that the diffraction pattern will become very weak as the size of the slit
becomes large relative to the wavelength λ . Diffraction therefore occurs most
strongly when the slit is narrow.
The irradiance of the center spot is given by Ee (0) and may be expressed in
terms of the strength of the electric field at the slit Ξ0 :
 2
1 Ξ0 D
Ee (0) = . (2.167)
2 R
Fraunhofer diffraction can also be shown for other configurations. An often-
seen demonstration is that of a double slit. Here, two parallel slits each cause an
interference pattern; these patterns are superposed. The irradiance is then
⎛  ⎞
2 Dπ
⎜ sin sin(θ ) ⎟

⎜ λ ⎟ 2 aπ
Ee (θ ) = Ee (0) ⎜  2 ⎟ cos , (2.168)
⎝ Dπ ⎠ λ
sin(θ )
λ
with a the separation distance between the two slits, as shown in Figure 2.38.
Thus, a diffraction effect may be obtained by aiming a coherent bundle of light
at a single or double slit. Such an experiment can be readily devised from a laser
pointer, a comb, and pieces of cardboard. By using the cardboard to mask off all
the gaps in the comb bar, a single or double slit can be fabricated (Figure 2.39).
The projection of the light on a wall is as shown in Figure 2.40. The horizontal
bright-and-dark pattern is caused by interference of the light passing through the
slits. For comparison, the cross section of the laser beam is shown in the same
figure. A good quality beam should have a Gaussian radial fall-off. The cheap
laser used here has a much more erratic pattern.
To show diffraction occurring near an edge, a (near-) monochromatic laser is
a good source of light, since it emits a coherent bundle of light. By shining this

Screen with double slit Wall

D P

a θ
R
D
Plane waves

Figure 2.38. Light passing through a double slit is projected onto a wall.

i i

i i
 i i

i i

2.7. Interference and Diffraction 75

Figure 2.39. With a source of laser light, a comb, and two bits of cardboard, a setup may
be created to demonstrate interference patterns. The cardboard is used to mask off most of
the comb such that only one or two slits are available to let light through. This is the setup
used to demonstrate single- and double-slit diffraction in Figure 2.40.

Figure 2.40. Interference patterns created by passing laser light through a single slit
(middle) and a double slit (bottom), as shown in Figure 2.39. The laser’s cross section is
shown at the top.

i i

i i
 i i

i i

76 2. Physics of Light

Figure 2.41. A laser is shone just over a knife’s edge (left). This causes an interference
pattern (right).

source over the edge of a knife, as shown in Figure 2.41, a diffraction pattern can
be created. The right panel in this figure shows the pattern of light created on a
wall several meters away from the laser and the knife. Below the center spot there
is a rapid fall-off, whereas above the center spot a series of additional spots are
visible.
This interference pattern is due to small changes in phase, which themselves
are due to the light interacting with the knife’s edge in a manner dependent on its
minimum distance to the edge.
While pointing a laser at the edge of a knife is not a common everyday oc-
currence, light streaking past an edge is very common. The above experiment
shows that shadow edges in practice are never completely sharp. This is in part
because light sources are never infinitely far away, nor infinitely small, although
even in such an idealized case, the diffraction around the edge will cause some
interference and therefore softening of the edge.

2.7.3 Diffraction of Polychromatic Light

It can be seen that the interference patterns caused by edges, slits, and apertures
of other shapes are dependent on the size of the aperture in relation to the wave-
length. In the case of an edge, there is a dependency between wavelength and
closest distance between the wavefront and the edge.
Rather than study interference patterns with single or double slits, we may
construct a large number of parallel slits. Such an array of diffracting obstacles
is called a diffraction grating. Another way of creating a diffraction grating is
by creating parallel grooves in a transparent sheet of material. If the grooves are

i i

i i
 i i

i i

2.7. Interference and Diffraction 77

Figure 2.42. A diffraction grating lit by a D65 daylight simulator (left). The rainbow
patterns on both the reflected (middle) and transmitted sides (right) are shown.

close enough together (D ≈ λ ), light is diffracted into different directions based

on its wavelength.
Polychromatic light sources consist of wavelengths distributed over the entire
visible spectrum. Most common light sources, including daylight, are of this
variety. Passing such light through a diffraction grating causes a rainbow pattern,
as observed in Figure 2.42. As the sheet shown in this figure both transmits and
reflects light, the rainbow pattern is seen on both sides.

Figure 2.43. A DVD lit by a D65 daylight simulator. The small pits encoding the data are
of the order of a wavelength, and therefore cause a diffraction pattern resulting in a set of
rainbow colors.

i i

i i
 i i

i i

78 2. Physics of Light

A more common occurrence of color caused by diffraction is shown in Fig-

ure 2.43, where a white light source is shone at a DVD. The pits encoding binary
information in the reflecting layer are of the order of a wavelength and, therefore,
cause a rainbow pattern.

2.8 Scattering
Many optical phenomena may be described as scattering, although it is often more
convenient to use specialized theories. For instance, we have described reflection
and refraction of electro-magnetic waves at a boundary between two dielectrics
with different indices of refraction in Section 2.5. Similarly diffraction and inter-
ference can be seen as special cases of scattering [111].
Scattering is an interaction of light with particles such as single atoms,
molecules, or clusters of molecules. Light can be absorbed by such particles and
almost instantaneously be re-emitted. This repeated absorption and re-emission
of light is called scattering. The direction of scattering may be different from the
direction of the incident light, leading for instance to the translucent appearance
of some materials as shown in Figure 2.44. Scattering events are called inelastic if
the wavelength of the re-emitted light is similar to the wavelength of the incident
light. If the wavelengths alter significantly, the scattering is called elastic.

Figure 2.44. This statue scatters light internally, resulting in a translucent appearance.

i i

i i
 i i

i i

2.8. Scattering 79

While a full account of scattering theory would be too involved for a book on
color theory, in the following subsections we offer a simplified explanation of dif-
ferent types of scattering. More detailed accounts of light scattering can be found
in “Light Scattering by Small Particles” [499], “Dynamic Light Scattering” [88],
“Scattering Theory of Waves and Particles” [832] and “Electromagnetic Scatter-
ing in Disperse Media” [54].

2.8.1 Rayleigh Scattering

Consider a linearly polarized electric dipole with charge q. The magnitude of its
moment p as a function of time t is then described by

p(t) = q d cos(ω t), (2.169)

where d is the initial maximum distance between the positive and negative charges.
The oscillation of this dipole gives rise to an electro-magnetic field with a rela-
tively complicated structure near the dipole. At some distance away from the
dipole, the field has a simpler structure whereby the E and H fields are transverse,
perpendicular, and in phase [447]. The magnitude of the electric vector is given
by
qd π sin(θ ) cos(r · s − ω t)
E = , (2.170)
λ 2 ε0 r
where position r is a distance r away from the dipole. The irradiance Ee associated
with this field is obtained by applying (2.96) and noting that the squared cosine
term becomes 1/2 after time-averaging (see Section 2.4):

q2 d 2 π 2 c sin2 (θ )
Ee = . (2.171)
2λ 4 r 2
This equation shows that a dipole oscillating with an angular frequency ω , gener-
ates an electro-magnetic field with an irradiance that falls off with the square of
the distance and, as such, obeys the inverse-square law. We therefore speak of a
dipole radiator.
In addition, the irradiance depends inversely on the fourth power of wave-
length. Thus, the higher the frequency of the oscillations (smaller λ ), the stronger
the resulting field.
Dipole radiation can be induced in small particles in the presence of an electro-
magnetic field. Such particles thus scatter light at the blue end of the visible spec-
trum more than light with longer wavelengths. If these scattering particles are
spaced in a medium at distances of at least one wavelength (in the visible region),
then the scattered waves are essentially independent. Light passing through such

i i

i i
 i i

i i

80 2. Physics of Light

Figure 2.45. Rayleigh scattering demonstrated with a tank of water and milk. On the left,
a D65 daylight simulator is placed behind the tank. The scattering causes a color that tends
towards orange. The image on the right shows the same tank and contents lit from the side
by the same light source. The substance now has a bluish tinge.

a medium will be absorbed and re-emitted in all directions. As a result, waves are
independent from each other in all directions, except forward. Light scattered in
the forward direction will generally maintain its phase coherence. This type of
scattering is called Rayleigh scattering.
Rayleigh scattering can be demonstrated by shining light through a tank of
water which contains a small amount of milk. The milk solution acts as a scatter-
ing medium. The light passing through in the forward direction will take on an
orange tint, since longer wavelengths are scattered less than shorter wavelengths.
Light scattered towards the side will have a blue tinge, as shown in Figure 2.45.
Another example of Rayleigh scattering is the blue color of some people’s
eyes (Figure 2.46). Both blue skies and orange and red sunrises and sunsets are
explained by Rayleigh scattering as well. The latter example is discussed further

Figure 2.46. The blue color in the iris shown on the left is due to Rayleigh scattering.
Green eyes, as shown on the right, occur due to a combination of Rayleigh scattering and
staining due to melanin. (Photographs adapted from [667].)

i i

i i
 i i

i i

2.8. Scattering 81

Figure 2.47. The blue coloring of the feathers of this blue jay are caused by Rayleigh
scattering; Yosemite National Park, California, 1999.

in Section 2.11. Rayleigh scattering is rare in the plant world, but surprisingly
common among animals. It is for instance the cause of the blue color in blue jays
(Figure 2.47).

2.8.2 Mie Scattering

While Rayleigh scattering affords a relatively simple model for small scattering
particles, for media containing particles larger than about a wavelength this model
is not appropriate. Instead, the more complicated Mie scattering model may be
used [111, 112, 436, 779]. The complexity of this model is such that it is beyond
the scope of this book to give a detailed mathematical account.
This complexity arises as follows. The model assumes that the particle is
spherical, with a planar wave-front scattering off of it. The wave equations (2.36)
and (2.37) are most conveniently restated in spherical coordinates. However, the
plane wave is most naturally represented in Cartesian coordinates. Converting the
plane wave representation to spherical coordinates is somewhat involved. Con-
versely, expressing the spherical boundary conditions in a Cartesian coordinate
system is similarly involved [1233].

i i

i i
 i i

i i

82 2. Physics of Light

Figure 2.48. Large particles, such as water droplets, cause light to be scattered in a largely
wavelength-independent manner. Hence, clouds appear a neutral gray or white; A Coruña,
Spain, August 2005.

Figure 2.49. Early morning fog causing Mie scattering; Casselberry, FL, 2004.

i i

i i
 i i

i i

2.8. Scattering 83

Figure 2.50. Smoke machines used for visual effect at concerts and in clubs produce fine
droplets that remain suspended in air, causing Mie scattering.

Mie scattering is only weakly dependent on wavelength and may scatter red
and green wavelengths in specific directions [815]. This is called polychroism and
occurs when the particles are larger than those causing Rayleigh scattering, but
still relatively small. For large particles, the scattering is largely independent of
wavelength, causing white light to remain white after scattering. As an example,
Mie scattering causes clouds to appear white or gray, as shown in Figure 2.48.
Mist and fog are other examples of Mie scattering (Figure 2.49), including artifi-
cially created fog such as that used at concerts (Figure 2.50).
A further natural diffraction-related phenomenon explained by Mie’s theory,
is the back scattering that occurs in clouds near the anti-solar point. A viewer
observing clouds with the sun directly behind, most commonly observed from an
airplane, may see a circular pattern called a glory. Dependent on the distance
between the observer and the clouds, the center of the glory will be occupied by
the shadow of the airplane. Figure 2.51 shows an example of a glory. Here the
distance between the clouds and the plane is large, and therefore the shadow is
negligibly small.

i i

i i
 i i

i i

84 2. Physics of Light

Figure 2.51. A glory; New Zealand, 2002. (Photograph courtesy of Timo Kunkel
(www.timo-kunkel.com).)

2.9 Geometrical Optics

The wavelength of light in the visible range is normally very small (on the order
of 10−7 meters) in relation to the size of the optical features of interest, which
include surface boundaries, volumes of space, and lenses. In these cases, to a first
approximation, it is possible to model the behavior of wavefronts by neglecting
wavelengths. Thus, in the limit that λ goes to zero, light may be modeled in terms
of geometrical behavior. When this is done, we speak of geometrical optics.
In this section, we assume that regions are free of currents and charges, i.e.
j = 0 and ρ = 0. The general form of the fields considered here is

E = e(r)ei β S(r) , (2.172a)

i β S(r)
H = h(r)e , (2.172b)

where S(r) is a real scalar function of space r. The function S(r) is known as
the optical path or the eikonal. Vectors e(r) and h(r) are complex vector valued
functions of space. Maxwell’s equations for this particular solution can then be

i i

i i
 i i

i i

2.9. Geometrical Optics 85

rewritten using the identities given in Section A.9 as

1
(∇ S) × h + ε e = − ∇× h, (2.173a)
iβ
1
(∇ S) × e − μ h = − ∇× e, (2.173b)
iβ
1
e · (∇ S) = − (e · ∇ ln ε + ∇·e) , (2.173c)
iβ
1
h · (∇ S) = − (h · ∇ ln μ + ∇·h) . (2.173d)
iβ

In the limit that the wavelength goes to zero, the value of β goes to infinity (be-
cause β = 2λπ ). As a result Maxwell’s equations simplify to

(∇ S) × h + ε e = 0, (2.174a)
(∇ S) × e − μ h = 0, (2.174b)
e · (∇ S) = 0, (2.174c)
h · (∇ S) = 0. (2.174d)

If h from (2.174b) is substituted into (2.174a), the latter equation yields

1

(e · (∇ S)) ∇ S − e (∇ S)2 + ε e = 0. (2.175)
μ

Substituting (2.174c) into this equation gives

(∇ S)2 = μ ε , (2.176a)
(∇ S)2 = n2 , (2.176b)

or equivalently
 2  2  2
∂S ∂S ∂S
+ + = n2 (x, y, z). (2.177)
∂x ∂y ∂z

This equation is known as the eikonal equation. The gradient of S can be seen
as the normal vector of the wavefront. It can be shown that the average Poynting
vector S is in the direction of the wavefront:

∇S
S= . (2.178)
n

i i

i i
 i i

i i

86 2. Physics of Light

2.9.1 Fermat’s Principle

The eikonal equation provides a description of geometrical optics. An alternative
description is afforded by Fermat’s principle, which states that light follows a
trajectory (a ray) such that the optical path length is an extremum. The optical
path length is defined as
 b
n ds, (2.179)
a
where ds is an element of arc length, and n is the index of refraction. The path
has fixed end points a and b. Minimizing this integral can be achieved using the
calculus of variations [736, 1082] resulting in
 
d dr
n(r) = ∇ n(r). (2.180)
dl dl

This equation is known as the ray equation [466, 617, 650, 1081] and is valid
for inhomogeneous isotropic media that are stationary over time. This equation
expresses the fact that at every point in the medium, the tangent and the normal
vector associated with the ray path span a plane called the osculating plane, and
the gradient of the refractive index must lie in this plane.
A consequence of Fermat’s principle is that in a homogeneous medium where
n(r) is constant, light travels in a straight line. In inhomogeneous media, however,
light may travel along curved arcs according to the above ray equation. This
happens, for instance, in the atmosphere, as shown in Section 2.10.4. Finally,
Snell’s law for reflection and refraction can be derived from Fermat’s principle
(Section 2.5.1).

2.9.2 The Cosine Law of Illumination

The cosine law of illumination states that a surface can receive radiation only in
proportion to its area projected in the direction of the light source. If the direction
of the light source is given by L and the surface normal is N, then the cosine of
the angle of incidence θ is given by their dot product:

cos(θ ) = N · L. (2.181)

The projected area is then

A = A N · L = A cos(θ ), (2.182)

as illustrated in Figure 2.52. In this figure, a large surface is angled away from the
light source, whereas a smaller surface is aimed directly at the light source. The

i i

i i
 i i

i i

2.9. Geometrical Optics 87

A’ = A cos(θ)

θ
A
A’

Figure 2.52. A surface can only receive radiation proportional to its projected area in the
direction of the light source.

projected area of the larger surface in the direction of the light source is identical
to the projected area of the smaller surface, and therefore both surfaces receive
the same amount of radiation.
A similar law called the cosine law of emission states that radiation emit-
ted from iso-radiant or Lambertian surfaces 8 decreases with a factor of cos(θ ),
where θ is the angle of exitance. An example is shown in Figure 2.53, where the
radiation emitted from a point decreases as the angle of exitance increases.

θ2 θ1

Figure 2.53. The radiation emitted from a uniform diffuser decreases as the angle of
exitance Θ increases.

Surfaces that obey the cosine law of emission appear equally light regardless
of the viewing direction. Although no perfect Lambertian surfaces exist in prac-
tice, many surfaces are well approximated by this law. These are called matte or
diffuse surfaces.

2.9.3 The Inverse Square Law

In Section 2.1.1, we saw that according to Coulomb’s law two charges exert a
force on each other that diminishes with the square of their distance. The cause
8 Strictly speaking, these surfaces are iso-radiant. Although a surface might be Lambertian at each

point individually, there may be spatial variation in the amount that is emitted. Iso-radiance requires
uniformity across the entire surface [839].

i i

i i
 i i

i i

88 2. Physics of Light

Wavefronts

Point source

Figure 2.54. Wavefronts diverging spherically from a point source.

of this relation lies in a more general law which stems from geometrical consid-
erations; this more general law is called the inverse square law9 .
Consider a point source that emits light. This light propagates through space
as spherical wavefronts. The farther the wave has traveled, the greater the surface
area spanned by the wave. The energy carried by the wave is thus distributed over
increasingly large areas, and therefore the energy density is reduced (Figure 2.54).

Figure 2.55. The objects farther from the light source (a 25W omni-directional incandes-
cent light) are dimmer than the objects closer to it due to the inverse square law.

9 This law applies to other modalities as well, including gravity and sound.

i i

i i
 i i

i i

2.9. Geometrical Optics 89

The inverse square law states that the surface density of radiant energy emitted
by a point source decreases with the squared distance d between the surface and
the source:
1
Ee ∝ 2 . (2.183)
d
An example is shown in Figure 2.55 where the mugs farther from the light source
are dimmer than the objects nearer to it.
As a general rule of thumb, if the distance of an observer to a source is greater
than 10 times the largest dimension of the source, it can be approximated as a
point source with respect to the observer [397]. The error associated with this
approximation is important in radiometry and will be further discussed in Sec-
tion 6.2.11.

2.9.4 Bouguer, Beer, and Lambert’s Laws

Light traveling through a dielectric medium will occasionally collide with a par-
ticle and either undergo scattering, or be absorbed and turned into a small amount
of heat. Both phenomena are discussed in detail in the following chapter. How-
ever, the effect of either result of the collision is that a photon has a probability of
interacting with the medium for every unit of distance traveled.
Hence, the irradiance after traveling a distance Δs through the dielectric ma-
terial will be attenuated by some amount:
Ee (s + Δs) = Ee (s) (1 − σa Δs) . (2.184)
Here, σa is a wavelength-dependent quantity modeling the attenuation induced by
the medium. It is known as the spectral absorbtivity. Rewriting this equation and
taking the limit Δs → 0, we find
Ee (s + Δs) − Ee (s)
lim = −σa Ee (s) , (2.185a)
Δs→0 Δs
d Ee (s)
= −σa Ee (s) . (2.185b)
ds
Assuming that the medium is homogeneous and isotropic, this equation can be
solved to yield
Ee (s) = Ee (0) e−σa s . (2.186)
Thus, the attenuation in an absorbing dielectric material is exponential. This result
is known as Bouguer’s law. The ratio Ee (s)/Ee (0) is called the transmittance T :
Ee (s)
T= (2.187a)
Ee (0)
= e−σa s . (2.187b)

i i

i i
 i i

i i

90 2. Physics of Light

A related quantity is absorbance A, written as

Ee (s)
A = − ln (2.188a)
Ee (0)
= σa s. (2.188b)

In spectroscopy and spectrophotometry, transmittance and absorbance are often

expressed in powers of 10. The conversion between the two forms is a factor of
ln(10) ≈ 2.3:

T = 10−σa s/ ln(10)
, (2.189a)
A = σa s/ ln(10). (2.189b)

The attenuation σa can be split into a material- dependent constant ε , which is

called either molar absorbtivity, extinction coefficient, or the concentration of the
absorbing material c. The absorbance is then given as

A = ε cs, (2.190)

and the transmittance is

T = 10−ε cs . (2.191)

This relation is known as Beer’s law. Here, we have merged the ln(10) factor
with the constant ε . As both concentration c and pathlength s have the same
effect on absorbance and transmittance, it is clear that a given percent change
in absorbance or transmittance can be effected by a change in pathlength, or a
change in concentration, or a combination of both.
Beer’s law is valid for filters and other materials where the concentration c
is low to moderate. For higher concentrations, the attenuation will deviate from
that predicted by Beer’s law. In addition, if scattering particles are present in the
medium, Beer’s law is not appropriate. This is further discussed in Section 3.4.6.
Finally, if the pathlength is changed by a factor of k1 from s to k1 s, then the
transmittance changes from T to T  as follows:

T  = T k1 , (2.192)

known as Lambert’s law. Similarly, if the concentration is changed from c to k2 c,

the transmittance goes from T to T  :

T  = T k2 . (2.193)

i i

i i
 i i

i i

2.9. Geometrical Optics 91

Surface normals
Light paths

Specular surface Rough surface

Figure 2.56. A perfect specular reflector (left) and a rough surface (right). Light incident
from a single direction is reflected into a single direction for specular surfaces, whereas
rough surfaces reflect light into different directions.

2.9.5 Surface Reflectance

We have shown that when a plane wave is incident upon a boundary between two
dielectrics, two new waves are created. These new waves travel in the reflected
direction and the transmitted direction. For a boundary between a dielectric and
a conductive object, such as a metal, the refracted wave is dampened and, there-
fore, travels into the material for a short distance. Reflected and refracted waves
traveling in directions as predicted by Snell’s law are called specular.
There are several aspects of material appearance that are not explained by this
model. For instance, it is not clear why shining a white light at a tomato produces
a red color; nor are diffuse, glossy, or translucent materials explained. While the
color of an object depends on the atomic structure of the material (this is the topic
of Chapter 3), non-specular surface reflectance and transmittance are the cause of
microscopic variations in surface orientation.
Thus, the occurrence of different levels of shininess depends on surface fea-
tures that are larger than single atoms and molecules, but still below the resolving
power of the human visual system. Specular reflection can, therefore, only occur
when the boundary is perfectly smooth at the microscopic level, i.e., the surface
normal is constant over small surface areas. At the macroscopic level, the surface
normal may still vary; this is then interpreted by the human visual system as a
curving but smooth surface.

Surface normals
Light paths

Multiple reflections

Figure 2.57. Self-occlusion occurs when reflected rays are incident upon the same surface
and cause a secondary reflection.

i i

i i
 i i

i i

92 2. Physics of Light

A rough surface has surface normals pointing in different directions. Since at

each point of the surface, light is still reflected specularly, shown in Figure 2.56,
a beam of light is reflected into many different directions. The appearance of
such an object is glossy. In the limit that the surface normals have a uniform
distribution over all possible directions facing away from the object, light will
be reflected in all directions equally. Such surfaces are perfectly diffuse and are
called Lambertian surfaces.
The distribution of surface undulations can be such that reflected rays are in-
cident upon the same surface before being reflected again away from the surface.
Such self-occlusion is shown in Figure 2.57.

2.9.6 Micro-Facets
The most direct way to model the macroscopic reflectance behavior of surfaces
is to treat a surface as a large collection of very small flat surfaces, called micro-
facets. It is generally assumed that each facet is perfectly specular, as outlined in
the preceding section. The orientation of each facet could be explicitly modeled,
but it is more efficient to use a statistical model to describe the distribution of
orientations. An example is the first micro-facet model, now named the Torrance-
Sparrow model [1140].
Blinn proposed to model the distribution of orientations D(Θi , Θo ) as an expo-
nential function of the cosine between the half-angle and the aggregated surface
normal [105]. The half-angle vector Θh is taken between the angle of incidence
Θi and the angle of reflection Θo :

D(Θi , Θo ) = (Θh · n)e (2.194a)

Θi + Θo
Θh = . (2.194b)
Θi + Θo

The exponent e is a user parameter.

Different micro-facet models make different assumptions, for instance, on the
existence of self-occlusion. Another common assumption is that micro-facets
form a regular grid of V-shaped grooves. One model, namely Oren-Nayar’s, as-
sumes that each facet is a perfectly diffuse reflector, rather than a specular reflec-
tor [859]. In this model, the amount of light reflected off a surface in a particular
direction Θo as function of light incident from direction Θi is given by
ρ
fr (Θi , Θo ) = (A + B max (0, cos(φi − φo ) sin(α ) tan(β ))) , (2.195a)
π
σ2
A = 1− 2 , (2.195b)
2σ + 0.66

i i

i i
 i i

i i

2.9. Geometrical Optics 93

0.45σ 2
B= , (2.195c)
σ 2 + 0.09
α = max(θi , θo ), (2.195d)
β = min(θi , θo ). (2.195e)

In this set of equations, σ is a free parameter modeling the roughness of the sur-
face. The factor ρ is the (diffuse) reflectance factor of each facet. The directions
of incidence and reflection (Θi and Θo ) are decomposed into polar coordinates,
where polar angle θ is the elevation above the horizon and azimuth angle φ is in
the plane of the facet:
Θ = (θ , φ ) . (2.196)
The above micro-facet models are isotropic, i.e., they are radially symmetric.
This is explicitly visible in the Oren-Nayar model where the result depends only
on the difference between φi and φo , but not on the actual values of these two
angles. This precludes modeling of anisotropic materials, where the amount re-
flected depends on the angles themselves. A characteristic feature of anisotropic
materials is, therefore, that rotation around the surface normal may alter the
amount of light reflected. An often-quoted example of an anisotropic material
is brushed aluminium, although it is also seen at a much larger scale for exam-
ple in pressed hay as shown in Figure 2.58. Several micro-facet models that take
anisotropy into account are available [45, 46, 702, 921, 1005, 1212].

Figure 2.58. Pressing hay into rolls, as shown here, produces an anisotropic effect on
the right side due to the circular orientation of each of the grass leaves; Rennes, France,
June 2005.

i i

i i
 i i

i i

94 2. Physics of Light

2.9.7 Bi-Directional Reflectance Distribution Functions

A further abstraction in the modeling of surface reflectance is afforded by bi-
directional reflectance distribution functions, or BRDFs for short [838]. The
micro-facet models presented in the preceding section can be seen as a specific
class of BRDF. In general, BRDFs are functions that return for each possible an-
gle of incidence and angle of reflectance, the fraction of light that is reflected at
the surface.
For BRDFs to be plausible models of physical reality, at the very least they
should be reciprocal as well as energy conserving. Reciprocal means that if the
angle of incidence and reflectance are reversed, the function returns the same
value:
fr (Θi , Θo ) = fr (Θo , Θi ). (2.197)
The energy-conserving requirement refers to the fact that light incident upon a sur-
face is either absorbed, transmitted, or reflected. The sum of these three compo-
nents can not be larger than the incident amount. Light emitted by self-luminous
surfaces is modeled separately and does not form part of a BRDF.
For computational convenience, the reflection of light may be classified into
four broad categories: diffuse, glossy specular, perfect specular, and retro-
reflective, although most surfaces exhibit a combination of all four types [895].
Fully specular BRDFs behave like the reflection detailed in Section 2.5, i.e.,
the Poynting vector of incident light is mirrored in the surface normal to determine
the outgoing direction [666]:

δ (θo − θi ) δ (φo − φr − π )
fr (Θi , Θo ) = , (2.198)
sin(θr ) cos(θr )

where we have used (2.196), and δ () is the Kronecker delta function. At the other
extreme, Lambertian surface models assume that light incident from all different
directions is reflected equally [645]:

1
fr (Θi , Θo ) = . (2.199)
π
Although this is not a physically plausible model, it is a reasonable approximation
for several materials, including matte paint.
An early model of glossy reflection was presented by Phong [897] and is given
here in modified form with the coordinate system chosen such that x and y are in
the plane of the surface and the z-coordinate is along the surface normal [895]:

e
fr (Θi , Θo ) = Θi · (−Θox , −Θoy , Θoz )T . (2.200)

i i

i i
 i i

i i

2.9. Geometrical Optics 95

Here, the outgoing direction Θo is scaled first by (−1, −1, 1)T , thus mirroring
this vector in the surface normal. It is now possible to make this scaling a user
parameter, allowing the modeling of various types of off-specular reflection:
  e
fr (Θi , Θo , s) = Θi · s ΘTo . (2.201)
While this BRDF could be used to model glossy materials, there is a further re-
finement that can be made by realizing that this function specifies a single lobe
around the angle of incidence. More complicated BRDFs which are still physi-
cally plausible can be created by modeling multiple lobes and summing them:
n   e
ρd
fr (Θi , Θo ) = + ∑ Θi · si ΘTo i . (2.202)
π i=1
ρd
The term models a diffuse component. Each of the n lobes is modeled by a
π
different scaling vector si as well as a different exponent ei . The result is called
the Lafortune model [637] .
An empirical anisotropic reflectance model takes into account the orientation
of the surface. The amount of light reflected into Θo depends on the angle of rota-
tion φn around the surface normal. Defining directional vectors for the incoming
and outgoing directions (vΘi and vΘo ), the half-vector vh is defined by
vΘi + vΘo
vh = . (2.203)
vΘi + vΘo
The angle Θnh is between the half-vector and the surface normal. Ward’s aniso-
tropic BRDF is then given by [269, 1200, 1212]
ρd ρs A
fr (Θi , Θo , s) = + , (2.204a)
π cos(θi ) cos(θo ) 4 π σx σy

where
      
φn φn
A = exp − tan Θnh cos 2
+ sin 2
. (2.204b)
σx2 σy2
The specular and diffuse components are modeled by ρs and ρd . Their sum will
be less than 1. The function is parameterized in σx and σy , each of which are
typically less than 0.2. These two parameters quantify the level of anisotropy
in the model. If they are given identical values, i.e., σ = σx = σy , the model
simplifies to an isotropic model:
  
Θnh
exp − tan
ρd ρs σ2
fr (Θi , Θo , s) = + . (2.205)
π cos(θi ) cos(θo ) 4π σ 2

i i

i i
 i i

i i

96 2. Physics of Light

Material ρd ρs σx σy
Lightly brushed aluminium 0.15 0.19 0.088 0.13
Rolled aluminium 0.10 0.21 0.04 0.09
Rolled brass 0.10 0.33 0.05 0.16
Enamel finished metal 0.25 0.047 0.080 0.096

Table 2.5. Example parameter settings for Ward’s model of anisotropic reflection
(After [1212]).

Parameters measured for several materials are listed in Table 2.5.

Although BRDFs are typically modeled as the ratio of incident to reflected
light and depend on incoming and outgoing directions, they can be extended
to be wavelength dependent, i.e., fr (Θi , Θo , λ ). In addition, polarization can
be accounted for in formulations related to BRDFs; these formulations are then
called bi-directional surface scattering reflectance distribution functions (BSS-
RDF) [36].

2.10 Application: Image Synthesis

Light travels normally in straight lines until it reaches a boundary. When such a
boundary is encountered, BRDFs describe the aggregate behavior of a differential
surface area, which may be thought of consisting of a large number of differently
oriented boundaries.
These two observations together may be used to compute the distribution of
light over a scene. Within a closed environment that contains both light sources as
well as reflective materials (such as a room, or approximately any outdoor scene),
light is at equilibrium, i.e., the amount of light emitted by all the light sources
together equals the amount of light absorbed by all the surfaces. Equilibrium is
reached after a very short period of time after switching on a light source.
For any point P on a surface, we already know how it reflects light; this is
given by the BRDF. To compute how the ability to reflect light contributes to the
light incident upon another point in the scene, we need to compute from where in
the scene light arrives at P. In other words, to compute how much light is reflected
in a given Θo direction, we need to integrate over all angles of incidence10 :

Lo (P, Θo ) = fr (P, Θi , Θo ) Li (P , Θi ) cos(θi )d ωi , (2.206)
Ωi
10 We follow the notation used by Arjan Kok in his PhD thesis [603]. We use the non-descript term

“light” here, indicated with the symbol L, since the appropriate radiometric term, radiance, is not
defined yet. Its definition can be found in Section 6.2.7.

i i

i i
 i i

i i

2.10. Application: Image Synthesis 97

where P are points on other surfaces in the scene that contribute to the incident
light. The total amount of light traveling from point P in direction Θo is then
simply the sum of all reflected light into that direction plus the light emitted at
surface point P. The latter value is only non-zero for self-luminous surfaces. The
equation governing light transport in a scene is known as the rendering equation
and is given by [568]

Lo (P, Θo ) = Le (P, Θo ) + fr (P, Θi , Θo )Li (P , Θi ) cos(θi )d ωi . (2.207)
Ωi

Although not explicitly shown, the terms in this integral, specifically Lo , Li , Le ,

and fr , are wavelength-dependent. In later chapters, it is shown that human vision
encodes color as a triplet of values, and therefore renderers can be optimized with
little harm to sample three color bands, rather than many separate wavelengths.
This does reduce the repertoire of visual phenomena that can be rendered, though.
For instance, dispersion that occurs due to a wavelength-dependent index of re-
fraction can only be modeled within a spectral renderer.
Evaluation of (2.207) is difficult in practice for several reasons. The quan-
tity computed also appears as a factor under the integral. This makes (2.207) a
Fredholm integral of the second kind, a type of equation that needs to be solved
numerically for all but the simplest integration domains.
However, in practice, the domain of integration is over a hemisphere placed
over point P which is visible by a good part of the scene. Thus, numerical inte-
gration involves sampling the environment to compute the factor Li (P , Θi ) for a
large set of directions Θi and then multiplying by the BRDF fr .
Since most of the environment contributes some quantity of light to the reflec-
tion in point P, rendering algorithms that compute the light interaction between a
point in the scene with every other point in the scene are called global illumina-
tion algorithms. Global illumination solutions differ in the way that the rendering
equation is approximated, where each affords a different trade-off between ren-
dering accuracy and computational cost [370, 895, 1043].
Common techniques include Monte Carlo sampling [271], ray tracing [370,
1042,1102,1242], radiosity [380,1046], and photon mapping [549]. None of these
techniques currently run in real time, perhaps with the exception of ray tracing,
which is on the brink of achieving interactive rates [875, 876, 1192, 1193].

2.10.1 Ray Tracing

The ray-tracing algorithm simulates the flux of light by recursively tracing rays
through an environment [1242]. A view point and view plane with associated
parameters is specified within the 3D environment that is to be rendered. This

i i

i i
 i i

i i

98 2. Physics of Light

serves to specify what will be visible in the image, much like holding a photo
camera in a real environment determines the composition of a photograph.
Rather than start at the light sources and follow the path that photons take
through the scene until they hit the view plane, the process is usually reversed
and light is traced backwards from the view point through the view plane into the
scene. This optimization helps reduce the number of rays that needs to be traced,
as there is no guarantee that any ray started at a light source will ever reach the
view plane.
When a ray hits a surface at a point P, called the intersection point, the ren-
dering equation can be evaluated by sampling the environment with further (sec-
ondary) rays. This sampling is potentially costly, as most of the environment
would have to be sampled to determine which parts of the scene illuminated the
intersection point.
To reduce the cost of sampling the environment, it is noted that some parts of
the scene contribute more to the illumination of a point than others. In particular,
nearby light sources that are directly visible from P are likely to contribute the
bulk of illumination. Similarly, unless the material is Lambertian, light arriving
from the reflected direction is more likely to contribute than that from other direc-
tions. By splitting the integral in (2.207) into separate components, an efficient
approximate solution can be found:

Lo (P, Θo ) = Le (P, Θo ) (2.208)

  
+∑ v (P, PS ) frdiff (P) Le PS , Θo cos (θS ) d ωS
S PS ∈S

+ f rspec (P, ΘR , Θo ) Lo (PR , ΘR ) cos (θR ) d ωR
ΘR ∈ΩR

+ ρd (P) La (P) .

In this equation, the first term signifies the light emitted from point P, as
before. The second term samples all the light sources directly. In it, v (P, PS )
is the visibility term which equals 1 if there is a direct line of sight between P
and the sample point PS on the light source. The visibility term is 0 otherwise.
In (2.208), the third term samples a small set of directions centered around the
reflected direction to account for specular and possibly glossy reflection. All other
light directions are not sampled, but are approximated with a constant, the fourth
term in (2.208). This last term is named the ambient term. It should also be noted
spec
that the BRDF is split into a specular and diffuse component ( fr and f rdiff ) for
the sake of convenience.

i i

i i
 i i

i i

2.10. Application: Image Synthesis 99

Figure 2.59. Diffuse inter-reflection between the rocks creates an intense orange
glow; Bryce Canyon, Southern Utah. (Image courtesy of the National Park Service
(http://www.nps.gov/archive/brca/photogalleryarchive.htm).)

2.10.2 Radiosity
While ray tracing is able to produce serviceable images in many applications, the
approximations are such that many light paths are not sampled. For instance, light
reflected from a Lambertian surface incident upon a second Lambertian surface,
i.e., diffuse inter-reflection, is not accounted for. Instead, in ray tracing this com-
ponent is approximated with a single constant, the fourth term in (2.208). An
example of diffuse inter-reflection in nature is shown in Figure 2.59.
Under the assumption that an environment consists of Lambertian surfaces,
BRDFs for all surfaces are simplified to be constant:

ρd (P)
fr (P) = . (2.209)
π
The rendering equation 2.207 then simplifies to become the radiosity equation
[200]:

L (P) = Le (P) (2.210)

    
ρd (P) cos (Θi ) cos (Θo )
+ L P v P, P dA ,
π all P P − P 2

i i

i i
 i i

i i

100 2. Physics of Light

where the integration is over differential surface areas dA . To make the solution
to this equation tractable, the scene is usually subdivided into small patches Ai ,
where the energy associated with each patch is assumed to be constant. For a
patch i this yields the following equation:
 
ρd,i Lj cos (Θi ) cos (Θ j )
π ∑
Li = Le,i + δi j dA j dAi . (2.211)
j Ai Ai A j π r2

In this equation, r is the distance between the patches i and j and δi j gives the
mutual visibility between the delta areas of the patches i and j. This equation can
be rewritten as

Li = Le,i + ρd,i ∑ L j f i→ j , (2.212a)

j
 
1 cos (Θi ) cos (Θ j )
fi→ j = δi j dA j dAi . (2.212b)
Aj Ai A j π r2

Here, the form factor fi→ j is the fraction of power leaving patch i that arrives at
patch j. Form factors depend solely on the geometry of the environment, i.e., the
size and the shape of the elements and their orientations relative to each other.
Therefore, the radiosity method is inherently view independent. It is normally
used as a preprocessing step to compute the diffuse light distribution over a scene
and is followed by a rendering step that produces the final image (or sequence of
images).

2.10.3 Photon Mapping

Light reflected off curved surfaces incident upon a Lambertian surface, or light
transmitted through a transparent object and incident upon a Lambertian surface,
are not accounted for in ray tracing and radiosity. These light paths create caus-
tics, patterns of light and dark, as for instance cast by the transparent objects in
Figure 2.60.
In ray tracing, where light paths are traced backwards, it is not possible to
determine in which direction to sample a light source, if a transparent object is
located along the line of sight between an intersection point and the light source.
The refraction caused by the transparent object would veer the ray off course and,
in all likelihood, it would miss the light source.
Thus, to properly account for caustics, rays should be traced from the light
source into the scene. However, as argued before, the probability that any of these
rays would hit the view plane is remote at best. A useful intermediate solution is
to begin by tracing rays/photons from the light source into the scene and deposit

i i

i i
 i i

i i

2.10. Application: Image Synthesis 101

Figure 2.60. Caustics cast onto the table by various objects.

energy whenever such rays hit a Lambertian surface. The data structure used to
store the energy of photons is called a photon map [549].
As light is properly refracted and reflected during this initial pass, the photon
map accurately encodes the caustics caused by transparent and specular objects.
During a second pass, conventional ray tracing may be used in a slightly mod-
ified form. Whenever a Lambertian surface is intersected, rather than continue as
normal, the photon map may be sampled instead. Variations are possible where
rather than storing and sampling energy in a photon map, angles of incidence are
recorded. In the second pass, rather than aiming rays directly at the light sources,
the values in the photon map are read to determine secondary ray directions. In
either solution, caustics can be rendered. In addition, photon mapping may be
extended to simulate the atmosphere as detailed in Section 2.11.

2.10.4 Rendering Inhomogeneous Media

The earth’s atmosphere typically has a varying index of refraction. This leads to
curved ray paths as discussed in Section 2.9.1. The density of air is determined by

i i

i i
 i i

i i

102 2. Physics of Light

sition
ent po
Appar
Ray
from
th e su
n
Earth

Figure 2.61. The varying density of the atmosphere causes the sun to be in a different
position than where it appears to be.

its temperature—the higher the temperature, the lower its density. The index of
refraction in turn depends on the density of the medium. Within a medium with a
smoothly varying refractive index, rays bend towards the area with a greater index
of refraction [112]. For the earth’s atmosphere, this has the implication that when
the sun is seen just above the horizon, it is in fact just below the horizon, as shown
in Figure 2.61.
In the atmosphere, the density of air is mediated by both pressure and tem-
perature, with both normally being highest near the earth’s surface. However,
local variations in temperature can bend rays in unexpected directions, yielding
effects such as mirages and fata morganas [783]. The inhomogeneous density of
air results in a spatially varying index of refraction.
To bend rays along their path, after advancing each ray by a small distance, its
direction l needs to be recomputed as a function of the local gradient of the refrac-
tive index ∇ n(r). This can be achieved by evaluating the ray equation (2.180).
If sunlight heats the ground, then close to the ground the index of refraction
changes more rapidly than higher above the ground. If an observer looks at a
distant object, then this may give rise to multiple paths that light rays may take
between the object and the observer. As an example, Figure 2.62 shows how
a curved path close to the ground plane may carry light between the roof and

Cool air

Viewpoint Direct path

ay
Curved r
Hot air
Ap
par
ent
pat
h

Figure 2.62. The roof of the building is visible along a direct path, as well as along a
curved path. The latter gives the impression of the building being reflected in the ground
plane.

i i

i i
 i i

i i

2.10. Application: Image Synthesis 103

Figure 2.63. Rendering of an inferior mirage. (Image courtesy of Diego Gutierrez, Adolfo
Munoz, Oscar Anson, and Francisco Seron; Advanced Computer Graphics Group (GIGA),
University of Zaragoza, Spain [411].)

the observer, whereas a straighter path farther from the ground also carries light
between the roof and the observer. The effect is that the building is visible both
upright and inverted. This is known as a mirage.
Dependent on atmospheric conditions, layers of colder and warmer air some-
times alternate, leading to different types of mirages. The configuration shown
in Figure 2.62 leads to an inferior mirage. A rendered example is shown in Fig-
ure 2.63. An inversion layer may cause the temperature at some altitude to be
higher than at ground level, and this gives rise to a superior mirage, as rendered
in Figure 2.64. Finally, if several such inversion layers are present, observers may
see a fata morgana, as shown in Figure 2.65. Further atmospherical phenomena,
including rainbows, halos, and scattering, are discussed in Section 2.11.

Figure 2.64. Rendering of a superior mirage. (Image courtesy of Diego Gutierrez, Adolfo
Munoz, Oscar Anson, and Francisco Seron; Advanced Computer Graphics Group (GIGA),
University of Zaragoza, Spain [411].)

i i

i i
 i i

i i

104 2. Physics of Light

Figure 2.65. Rendering of a fata morgana. (Image courtesy of Diego Gutierrez, Adolfo
Munoz, Oscar Anson, and Francisco Seron; Advanced Computer Graphics Group (GIGA),
University of Zaragoza, Spain [411].)

2.10.5 Optimizations
Many optimization techniques have been applied to make rendering algorithms
faster, including spatial subdivision techniques, which essentially sort the objects
in space allowing intersection points to be computed without testing each ray
against all objects [437]. Other strategies include distributed and parallel process-
ing, allowing multiple processing units to partake in the computations [161].
As sampling is a core technique in rendering algorithms, optimizing sampling
strategies has received a great deal of attention. In particular, from (2.207) we
have that each point in the scene contributes some amount of light to be reflected
in a direction of our choosing. Using pre-processing techniques, sampling may
be directed towards those parts of the scene that contribute most to the final re-
sult. Such techniques are collectively called importance sampling. An example is
Ostromoukhov’s technique, which is based on Penrose tiling [863].
At the same time, the BRDF kernel will weight these contributions. It is
therefore reasonable to only sample directions that will be weighted with a large
factor. Sampling techniques which take both the environment as well as the BRDF
into account to generate sampling directions should allow accurate results with
minimal loss of accuracy [655].

2.11 Application: Modeling the Atmosphere

Every day the Earth’s atmosphere produces beautiful displays of color, ranging
from blue skies, clouds, orange and red sunrises, and sunsets to rarer phenom-

i i

i i
 i i

i i

2.11. Application: Modeling the Atmosphere 105

Direction of Zenith
Sun
interest
Point of interest

θ θs
North
West γ

φ
φs
East
South

Figure 2.66. The geometry related to computing the position of the sun in the sky.

ena such as rainbows, halos, and the northern and southern lights. Each of these
phenomena can be explained [716, 783], and most have been modeled and ren-
dered. Examples include simulation of night scenes [1124], rendering of northern
and southern lights [59], and rendering of mirages and fata-morganas [410, 411].
In this section, we begin by briefly explaining several atmospheric phenomena
and then show how some of them may be incorporated into an image synthesis
algorithm.

2.11.1 The Sun

As the predominant light source, the interaction of the sun with the atmosphere
is the most important feature giving rise to sky coloring, as well as the sun’s
spectral distribution as seen at the surface of the Earth. The atmospheric con-
ditions as well as the length of the trajectory of the sun’s light through the at-
mosphere determine the range of colors that can be observed. To model atmo-
spheric phenomena, it is therefore important to know the position of the sun in
the sky, as well as the spectral distribution of the sun’s light before it reaches the
atmosphere.
The position of the sun is determined by the set of polar coordinates (θs ,φs ).
These, as well as several other angles used in the following section are depicted in
Figure 2.66. The position of the sun may be determined from solar time t which it-
self is given in terms of the Julian date J,11 the longitude (in radians), the standard
meridian MS for the time zone (in radians), and the standard time (in
decimal hours):

11 The Julian date is an integer in the range [1, 365] indicating the day of the year.

i i

i i
 i i

i i

106 2. Physics of Light

 
4 π (J − 80)
t = ts + 0.170 sin (2.213)
373
 
2 π (J − 8) 12 (MS − L)
− 0.129 sin + . (2.214)
355 π
The solar declination δ , also required to compute the position of the sun, is ap-
proximated by  
2 π (J − 81)
δ = 0.4093 sin . (2.215)
368
With the help of the site latitude l (in radians), the angle in the sky of the sun is
then given by (see also Figure 2.66)
π

π t 
θs = − sin−1 sin(l) sin(δ ) − cos(l) cos(δ ) cos , (2.216)
2 12
⎛
π t  ⎞
− cos(δ ) sin
⎜ ⎟
φs = tan−1 ⎝ 12
π t ⎠. (2.217)
cos(l) sin(δ ) − sin(l) cos(δ ) cos
12
The second characteristic of the sun required for accurate modeling of the
atmosphere is its spectral distribution. Outside the atmosphere, the spectral dis-
tribution of the solar irradiance is given in Figure 2.67 [403, 404]. At the Earth’s
Spectral irradiance (Wm -2 nm -1 )

2.5
Extraterrestrial solar spectral irradiance
Global tilt
2.0 Direct + circumsolar

1.5

1.0

0.5

0.0
0 500 1000 1500 2000 2500 3000 3500 4000
Wavelength (nm)

Figure 2.67. The extraterrestrial solar irradiance [403, 404], as well as two terrestrial
conditions, as specified by the American Society for Testing and Materials (ASTM) in
standard G173-03 [48].

i i

i i
 i i

i i

2.11. Application: Modeling the Atmosphere 107

surface, this distribution is altered due to interaction with the atmosphere [1037].
The American Society for Testing and Materials (ASTM) has defined two refer-
ence spectra using conditions representative of average conditions encountered in
mainland USA [48]; these are plotted in Figure 2.67. These terrestrial standards
make assumptions on the atmosphere, including the use of the 1976 U.S. Stan-
dard Atmosphere [1166] and a surface spectral reflectivity as produced by light
soil [486].

2.11.2 The Atmosphere’s Color

The most easily observed aspects of the atmosphere include the color of the sky
and clouds. The sky appears blue during the day and colors orange to red dur-
ing sunsets and sunrises. The range of colors the sky exhibits during the course
of a day are all explained by Rayleigh scattering of small particles in the upper
atmosphere.
As the upper atmosphere scatters light, the amount of light scattered is deter-
mined by the distance that light travels through the atmosphere. The longer the
pathlength, the more scattering can occur. During the day, when the sun is more
or less overhead, this pathlength is relatively short, and the sun therefore looks
pale yellow. At sunrise and sunset, the pathlength is much longer, and thus most
of the blue has been scattered out before the direct sunlight reaches the observer.

Figure 2.68. A sunset. The distance traveled by sunlight through the thin upper atmo-
sphere causes much of the light at blue wavelengths to be scattered out, leaving the red
part of the spectrum. This is an example of Rayleigh scattering; Clevedon, UK, August
2007.

i i

i i
 i i

i i

108 2. Physics of Light

Figure 2.69. The blue sky is caused by Rayleigh scattering in the thin upper atmosphere;
St. Malo, France, June 2005.

This colors the sun orange and red, as shown in Figure 2.68. For a model for
rendering twilight phenomena the reader is referred to a paper by Haber [413].
The color of the sky at mid-day is also due to Rayleigh scattering. While most
wavelengths travel through the atmosphere undisturbed, it is mostly the blue that
is scattered, as shown in Figure 2.69.
The atmosphere contains many particles that are relatively large. These parti-
cles give rise to Mie scattering , leading, for instance, to haze, a whitening of the
sky’s color. Finally, the scattering of light causes objects in the distance to appear
desaturated and shifted towards the blue. This phenomenon is known as aerial
perspective.
Haze may be described with a heuristic parameter called turbidity [760]. To
determine an appropriate value for turbidity on the basis of a description of the
atmosphere or the meteorological distance,12 Table 2.6 may be consulted (af-
ter [925]13 ).
With the position of the sun known, and a value for turbidity specified, it is
now possible to compute aerial perspective. When light is reflected off a distant
object, along its path towards the observer it may be attenuated. In addition, the
12 The meteorological distance is defined as the distance under daylight conditions at which the

apparent contrast between a black target and the horizon is just noticeable. It roughly corresponds to
the most distant feature that can be distinguished.
13 The following derivation to compute the attenuation along a given path through the sky is also

after Preetham, Shirley, and Smits [925].

i i

i i
 i i

i i

2.11. Application: Modeling the Atmosphere 109

Atmosphere Rm (in km) Turbidity T

pure air 256 1.0
very clear 55 2.0
clear 15 3.8
light haze 6 8.0
haze 3 24.0
thin fog 1 60.0

Table 2.6. Turbidity T as function of the type of atmosphere. Values are approximate.

atmosphere scatters light at all positions in space, and in particular along this
path, some light may be scattered into the direction of the observer. This is called
in-scattering. The light arriving at the observer is therefore a function of the in-
scattered light Lin and the attenuated light L0 :
L = τ L0 + Lin , (2.218)
where L is the amount of light arriving at the observer and τ is the extinction fac-
tor. Both the in-scattering term and the extinction factor depend on the scattering
medium and are, therefore, subject to both Rayleigh scattering for molecules and
Mie scattering for larger particles. The extinction factor (as well as Lin ) depends
on the scattering coefficients βm and βh for molecules and haze respectively:
  s   s 
τ (s) = exp − βm (h(x)) exp − βh (h(x)) . (2.219)
0 0

The function h(x) denotes the height of each point x along the path between 0 and
s and is given by h(x) = h0 + x cos(Θ). The extinction factor is thus an integral
over the path taken by the light through the atmosphere. As the scattering function
depends on the density of the particles, and therefore on the height h above the
Earth’s surface, the scattering function for both βm and βh is given by
β (h) = β 0 exp(−α h), (2.220)
where β 0 is the scattering coefficient at the Earth’s surface and α = 1.3 is the
exponential decay constant. The value of β 0 can be computed as a function of
angle θ (see Figure 2.66), or can be integrated over all angles. These expressions
for the Rayleigh scattering part are given by βm0 :
 2  
π 2 n2 − 1 6 + 3 pn  
βm0 (θ ) = cos 1 + cos2 (θ ) , (2.221)
2N λ 4 6 − 7 pn
 2  
8 π 3 n2 − 1 6 + 3 pn
βm =
0
. (2.222)
3N λ4 6 − 7 pn

i i

i i
 i i

i i

110 2. Physics of Light

η (log10)
1

-1 400 nm
450 nm
550 nm
650 nm
850 nm
-2
0 20 40 60 80 100 120 140 160 180
Scattering angle

Figure 2.70. Values for η given different wavelengths and scattering angles. These values
are valid for a value of 4 for Junge’s constant v.

Here, n = 1.0003 is the index of refraction of air, N = 2.54525 is the number

of molecules per unit volume, and pn = 0.035 is a de-polarization factor. The
scattering coefficients for haze are due to Mie scattering and are given by
 v−2
2π η (θ , λ )
βh0 (θ ) = 0.434 c , (2.223)
λ 2
 v−2
2π
βh0 = 0.434 c K(λ ), (2.224)
λ

where v is Junge’s exponent which takes a value of 4 for sky modeling. The
concentration factor c depends on the turbidity T as follows:

c = (0.6544 T − 0.6510) × 10−16 . (2.225)

Curves for η (θ , λ ) and K are plotted in Figures 2.70 and 2.71.

With the scattering coefficients for Rayleigh and Mie scattering defined, the
final expression of the extinction coefficient can be derived. As the height h varies
along the path as indicated above, the integrals in (2.219) can be solved, and this
equation can be rewritten as follows:

τ (s) = exp (−am (bm − um (s))) exp (−ah (bh − uh (s))) , (2.226)
u(s) = exp (−α (h0 + x cos(θ ))) , (2.227)

i i

i i
 i i

i i

2.11. Application: Modeling the Atmosphere 111

K
0.69

0.68

0.67

0.66

0.65
350 400 450 500 550 600 650 700 750 800
Wavelength

Figure 2.71. Values for K as function of wavelength λ .

β0
a=− , (2.228)
α cos(θ )
b = exp (−α h0 ) . (2.229)

This expression thus forms the first component of the computation of atmospheric
attenuation and can be plugged into (2.218). The second component involves the
computation of the light that is scattered into the path. At any given point along
the path, some light is scattered in, and this quantity thus needs to be integrated
over the length of the path.
We first compute the light scattered in at a specific point. This quantity de-
pends on the scattering functions βm0 and βh0 at sea level, as well as altitude h. We
can thus write expressions for the angular scattering functions in a manner similar
to (2.220):

β (ω , θ , φ , h(x)) = β 0 (ω , θ , φ ) u(x). (2.230)

The function β (ω , θ , φ , h(x)) can be seen as a phase function in that it describes

how much of the light coming from ω is scattered into the direction of interest
(θ , φ ). The total amount of light from ω scattering into this direction is then
Ls (ω ) β (ω , θ , φ , h(x)). The total amount of light scattered at position x is given by
the integral over all directions ω (repeated for both Rayleigh and Mie scattering

i i

i i
 i i

i i

112 2. Physics of Light

components):

Lin (x) = um (x) Ls (ω ) βm0 (ω , θ , φ ) d ω (2.231)
Ω

+ uh (x) Ls (ω ) βh0 (ω , θ , φ ) d ω ,
Ω

where we have applied (2.230) and taken u(x) out of the integrals. The total
amount of scattering along a path can be computed by integrating the above ex-
pression over this path, taking into account the attenuation τ that occurs at the
same time (this component can be taken from (2.226)):
 s
Lin = Lin (x) τ (s)dx. (2.232)
0

This equation can be solved numerically, or with suitable approximations it can

be simplified to allow analytical solutions, as shown by Preetham et al [925].
This then also allows (2.218) and hence the attenuation along a ray path due to
both molecular Rayleigh scattering as well as Mie scattering due to larger haze
particles to be computed.

2.11.3 Rainbows
During a rain shower, droplets of water are suspended in air. If a directional light
source, such as the sun, illuminates a volume with water droplets, each droplet

Figure 2.72. A rainbow; Millford Sound, New Zealand, 2001. (Photo courtesy of Timo
Kunkel (www.timo-kunkel.com).)

i i

i i
 i i

i i

2.11. Application: Modeling the Atmosphere 113

Secondary rainbow
42
51
Sun
Observer

Primary rainbow

Figure 2.73. Geometry associated with rainbows.

reflects and refracts light. The refraction causes dispersion of the light which de-
pends on wavelength. This gives rise to a rainbow (Figure 2.72). The rainbow is
due to a single internal reflection inside the water droplet. If a second reflection
occurs, this gives rise to a weaker secondary rainbow. The location of these rain-
bows with respect to the sun and the observer is diagrammed in Figure 2.73. The

Alexander’s dark band

Secondary rainbow Primary rainbow

Intensity (log10 )

0 -6

-7
-2
-8
-4
-9
120 130 140 150
-6

-8 440 nm
510 nm
650 nm
-10
0 20 40 60 80 100 120 140 160 180
Scattering angle

Figure 2.74. Intensity as function of scattering angle for water droplets of 1000 μ m
(std. dev. 120 μ m) suspended in air and lit by sunlight. These plots were computed for
perpendicular polarization. (Data generated using Philip Laven’s Mieplot program [654].)

i i

i i
 i i

i i

114 2. Physics of Light

Alexander’s dark band

Secondary rainbow Primary rainbow

2

Inensity (log 10 )
-7

-8
-2

-4
-9
-6 120 130 140 150

-8 440 nm
510 nm
650 nm
-10
0 20 40 60 80 100 120 140 160 180
Scattering angle

Figure 2.75. Intensity as function of scattering angle for water droplets of 1000μ m (std.
dev. 120μ m) suspended in air and lit by sunlight. These plots were computed for parallel
polarization. (Data generated using Philip Laven’s Mieplot program [654].)

region in between the primary and secondary rainbows is somewhat darker than
the surrounding sky and is known as Alexander’s dark band.
Droplets of a given size will scatter light over all angles, and this behav-
ior may be modeled with a phase function, which records for all outgoing an-
gles, how much light is scattered in that direction. Figures 2.74 and 2.75 show
the distribution of intensities generated by water droplets with a size distributed
around 1000 μ m with a log normal distribution having a standard deviation of
10 μ m. These distributions differ in their polarization and were both generated
for droplets suspended in air that was lit by sunlight. The plots show marked
peaks in the forward direction (no scattering, i.e., a scattering angle of 0 degrees),
as well as peaks for the primary and secondary rainbows. In addition, Alexander’s
dark band is visible.

2.11.4 Halos
Some clouds, such as cirrus clouds, are formed by ice particles rather than rain
droplets. These ice particles are usually assumed to be shaped like a hexagonal
slab with a varying height-to-width ratio. Crystals that are flat are called plates,
whereas crystals that are elongated are called pencils. Under wind-still conditions,

i i

i i
 i i

i i

2.11. Application: Modeling the Atmosphere 115

Plate Pencil

Figure 2.76. Ice crystals formed in cirrus clouds are usually assumed to be either plates
or pencils.

these crystals are suspended in air in approximately the orientation as shown in

Figure 2.76.
Sunlight penetrating cirrus clouds will interact with ice particles [387]. As
the normal laws of reflection and refraction apply when light enters these crystals,
light will be scattered into preferential directions determined by the geometry of
these crystals. These (multiple) scattering events lead to the formation of halos.
The halos occurring most often are located at an angle of 22◦ around the sun.
A weaker halo is sometimes seen at 46◦ . Further halo effects, such as sun-dogs
(parhely) and the upper tangent arc are also well known. Their position in relation
to the observer and the sun is diagrammed in Figure 2.77. A photograph of a halo
as well as a sun-dog is shown in Figure 2.78. The 46◦ halo is shown in Figure 2.79.
The scattering behavior of hexagonal ice particles can be modeled with a
phase function, i.e. a function that models for every direction how much light
is scattered in that direction [469, 691]. Phase functions indicate fractions of light
being scattered in specific directions and therefore integrate to one. Such phase
functions have been measured for halo-forming cirrus clouds, and are plotted in
Figure 2.80.
The assumption that all ice crystals are either plates or pencils is an oversim-
plification. In particular, the collection and analysis of samples of ice crystals has
revealed that ice crystals in halo-forming clouds are frequently hollow, and also

22 halo
46 halo
Parhely (sun-dogs)
Upper-tangent arc
Sun
Observer

Figure 2.77. Different types of halos and their position in relation to the observer and the
sun (after [377]).

i i

i i
 i i

i i

116 2. Physics of Light

Figure 2.78. Halo around the sun. Toward the lower left a second refraction is visible;
Southern Utah, 2001.

have more complicated shapes than simple hexagonal forms [996]. As a result,
the phase function shown in Figure 2.80, which is based on a simplified model,
suggests that the 22◦ and in particular the 46◦ are more common than they are in
practice. See also Sassen and Ulanowski for a discussion [997, 1161].

Figure 2.79. The 46◦ halo, photographed with a wide-angle lens; Wanaka, New Zealand,
2001. (Photo courtesy of Timo Kunkel (www.timo-kunkel.com).)

i i

i i
 i i

i i

2.11. Application: Modeling the Atmosphere 117

Phase (log 10 )
4
3
2 22 Halo
1 46 Halo
0
-1
-2
-3
0 30 60 90 120 150 180
Scattering angle

Figure 2.80. Phase function for hexagonal ice crystals (after [963]).

2.11.5 The Green Flash

At the last few seconds of sunset, sometimes the sun may be colored green, a
phenomenon known as the green flash. It results from a combination of Rayleigh
scattering and dispersion. While rarely seen in nature, predominantly because this
phenomenon only lasts for a brief moment, this effect can be reproduced with a
projector, a mirror, and a pan of water with a few drops of milk suspended [109].
With such a set of tools, it is possible to construct a dispersing prism made of
a material which exhibits Rayleigh scattering (water with some milk). In Fig-
ure 2.81, we show a similar demonstration, created by shining a beam of white
light through an ordinary dispersing prism, followed by passing the dispersed

Figure 2.81. On the left, white light is passed through a dispersing prism as well as a tank
of clear water. Normal dispersion is observed. On the right, a few drops of milk are added
to the water. The blue component of the dispersed light is scattered laterally, so that a red
disk with a green rim remains.

i i

i i
 i i

i i

118 2. Physics of Light

light through a container with milky water. The result is a simulation of a sunset
with added dispersion.

2.11.6 Rendering the Atmosphere

To render the atmosphere, a rendering algorithm such as the ones outlined in Sec-
tion 2.10 can be employed. However, without modification, these algorithms do
not model important components such as participating media, volumes of matter
interacting with light. Instead, conventional algorithms assume that light travels
through a medium without interacting with it, and only changes paths at surface
boundaries.
This is not sufficient to model the atmosphere. Instead, the radiative transfer
equation can be evaluated for every point along a ray path. Thus, tracing a ray
involves advancing the ray by a short distance, evaluating the radiative transfer
equation to determine the change in light due to the particles present in a small
volume around the point of interest. The ray is then advanced a short distance
again, and the process is repeated. At surface boundaries, light is reflected and
refracted as normal.
The radiative transfer equation takes into account how much light is emitted,
scattered, and absorbed at a point P on a ray aimed into direction Θ:

∂ L (P, Θ)
= α (P) Le (P, Θ) (2.233)
dP
+ α (P) Li (P, Θ)
+ σ (P) L(P, Θ)
+ α (P) L(P, Θ).

With α the absorption coefficient and σ the scattering coefficient, the four terms in
this equation model emission, in-scattering, out-scattering, and absorption. The
term in-scattering refers to light coming from any direction which is scattering
into the direction of the ray path. Out-scattering accounts for the light scattering
out of the ray’s path. The in-scattered light Li can come from any direction and is
therefore an integral over a full sphere Ω:

Li (P, Θ) = p(P, Θ, Θ ) L(P, Θ ) dΘ . (2.234)
Ω

Thus, the light coming from each direction is weighted by a phase function p
that models the scattering properties of the volume, i.e., the directionality of the
scattering. As desired, this function can be chosen to account for rainbows [963],
halos [371, 372, 377, 541, 542], as well as clouds. In addition, a specific form of

i i

i i
 i i

i i

2.12. Summary 119

Figure 2.82. Rendering of participating media. The left image shows distortions caused
by spatially varying temperature. The middle image shows a 3D turbulence field simulating
smoke. An anisotropic medium is modeled in the rightmost image. (Images courtesy of
Diego Gutierrez, Adolfo Munoz, Oscar Anson, and Francisco Seron; Advanced Computer
Graphics Group (GIGA), University of Zaragoza, Spain [411].)

this function was shown in Section 2.11.2 that accounts for the coloring of the
sky. Section 3.8.2 discusses an alternative phase function for rendering flames.
In the above, it is implied that each function is also dependent on wavelength.
The scattering is inelastic, i.e., scattering does not change a particle’s wavelength.
To account for elastic scattering, and thus model fluorescence and phosphores-
cence, an extra term can be added to the radiative transfer equation :
 
pλ (P, Θ , Θ)
αλi (P) f λi →λ (P) Lλi (P, Θ ) dΘ d λi . (2.235)
Ω λ 4π

Here, f is a probability distribution function that models the likelihood of a pho-

ton at wavelength λi being re-emitted at wavelength λ . Simulations of inelastic
scattering are shown in Figure 2.82.

2.12 Summary
Light is modeled as harmonic electromagnetic waves. The behavior of electric
and magnetic fields gives rise to the propagation of light through free space.
The electric and magnetic vectors are orthogonal to the direction of propagation,
which is indicated by the Poynting vector. The magnitude of the Poynting vector
is commensurate with the amount of energy carried by the wave.
While propagating, the electric and magnetic vectors normally rotate around
the Poynting vector, inscribing an ellipse. When the ellipse degenerates either
into a circle or a line, the light is said to be polarized.

i i

i i
 i i

i i

120 2. Physics of Light

In homogeneous isotropic materials, light propagates by the same principle,

albeit at a different velocity. If light is modeled as a superposition of harmonic
plane waves, then each of the constituent components oscillates with a given fre-
quency and associated wavelength. Visible light consists of a very narrow range
of wavelengths—roughly between 380 and 780 nm.
At surface boundaries, light is reflected and refracted. The angle of reflection
and refraction are given by Snell’s law, whereas the amount that is reflected and
refracted is modeled by the Fresnel equations.

2.13 Further Reading

There are many texts on the physics of light and optics. The classic texts are
Born and Wolf, Principles of Optics [112] as well as Hecht, Optics [447]. For the
preparation of this chapter, we also used Iskander, Electromagnetic Fields and
Waves [532] and the Handbook of Optics [73]. Other sources are [379, 428, 436,
736, 1016]. There are many books on computer graphics, of which we can list
only some [23, 271, 370, 895, 1043]. Several books on weather phenomena are
worth mentioning [387, 716, 783]. Williamson and Cummins, Light and Color
in Nature and Art [1247] discusses many of the topics of this chapter, including
modeling the atmosphere.

i i

i i
 i i

i i

Chapter 3
Chemistry of Matter

In the previous chapter, light was treated exclusively as a wave phenomenon. This
model is suitable for explaining reflection, refraction, and polarization. However,
the wave model alone is not able to explain all phenomena of light. It is therefore
necessary to extend the theory to include the notion of particles. This leads to
the concept of wave-particle duality: light has behaviors commensurate with both
wave models as well as particle models.
It was postulated by de Broglie that not only light, but all matter exhibits
both wave-like and particle-like behaviors. This includes protons, neutrons and
electrons that form atomic and molecular structures.
Thus, a model is needed to account for both wave and particle behaviors,
and such a model is afforded by Schrödinger’s equation. This equation can be
used to show that particles can only have specific energies, i.e., energy levels of
particles are generally quantized. The study of particles at the microscopic level
is therefore termed quantum mechanics [179, 683, 1109].
This chapter first briefly reviews classical physics which forms the basis of
quantum mechanics. Then, quantum mechanics and molecular orbital theory is
explained. These theories are necessary to understand how light interacts with
matter. In particular, we discuss the close interaction of energy states of elec-
trons orbiting atomic nuclei and light of particular wavelengths being emitted,
absorbed, or reflected.
The material presented in this chapter, in conjunction with electromagnetic
theory, also affords insight into the behavior of light when it interacts with matter
which itself has structure at the scale of single wavelengths.
121

i i

i i
 i i

i i

122 3. Chemistry of Matter

3.1 Classical Physics

As quantum mechanics relies heavily on classical physics, we start this chapter
with a brief overview of classical physics. We ignore magnetic interactions for
the moment to simplify the discussion.
A particle traveling through space has various inter-related attributes such as
mass m, (linear) velocity v, and momentum p:

p = m v. (3.1)

Along with linear momentum, a particle may possess angular momentum L if it

rotates around a point1 . If the point of rotation is a distance r away from the
particle, the angular momentum is defined as (see also Figure 3.1)

L = m v r. (3.2)

The momentum of a particular will change over time if a force is applied to it.
According to Newton’s second law of mechanics, a constant force F gives rise to
a linear change of momentum with time:

d(mv) d p
F= = . (3.3)
dt dt
For a particle with constant mass m, this law may be rewritten to yield

dv
F =m = ma, (3.4)
dt
where a is the acceleration of the particle. In addition to mechanical forces, parti-
cles may be subject to forces induced by charges, as predicted by Coulomb’s law
and given in Equation (2.1).
mv

L=mvr

Figure 3.1. Angular momentum.

1 In this chapter, the quantity L is reserved for angular momentum. However, this is also the com-

mon symbol used for luminance. In the remainder of this book, L is therefore taken to mean luminance.

i i

i i
 i i

i i

3.1. Classical Physics 123

The total energy E of a particle may be decomposed into a kinetic energy T ,

as well as a potential energy U:

E = T +U. (3.5)

Given the above, the kinetic energy is given by

mv2 p2
T= = . (3.6)
2 2m
Thus, the kinetic energy of a particle may be expressed in terms of its linear
momentum p. The potential energy of a charged particle in the electrostatic field
of another particle a distance r away is given by
Q1 Q2
U= . (3.7)
r
In the following, we will assume that, for a given system of particles, the total
energy remains constant. In that case, the Hamiltonian H of the system equals its
total energy:
H = E. (3.8)
If we assume that heavy particles, such as protons, are stationary, then the kinetic
energy is due to the electrons only. For instance, the Hamiltonian of a hydrogen
atom (which consists of one proton and one electron) becomes

H = T +U (3.9a)
p 2 e2
= − . (3.9b)
2m r
In this equation, e is the unit charge, which is positive for the proton and nega-
tive for the electron, and r is the distance between the proton and the electron.
For atoms with multiple protons and electrons, the Hamiltonian becomes more
complex as it will include a term for each pair of particles. As an example, the
Hamiltonian for a helium atom with two protons and two electrons would be

p21 p2 2e2 2e2 e2

H= + 2 − − + . (3.10)
2m 2m r1 r2 r12
For complex configurations, the Hamiltonian can not be evaluated directly, and
one would have to resort to approximations to estimate the total energy.
The total energy of a system is an indication of its stability. For instance,
if an electron with kinetic energy T approaches a proton, the system’s potential
energy decreases. However, the total energy of the system remains constant, and

i i

i i
 i i

i i

124 3. Chemistry of Matter

therefore the kinetic energy must increase. For the electron and the proton to
remain together, this excess kinetic energy must be lost. Otherwise, the kinetic
energy of the electron can become so large as to overcome the attractive force
of the proton. In that case, the electron would escape. Stability of a system
thus implies that the negative potential energy is greater than the positive kinetic
energy. A consequence is that the total energy of the system has to be negative
for the system to be stable.
While many phenomena may be adequately explained by classical physics, at
the microscopic level there are discrepancies between effects that can be measured
and those that are predicted by the theory. This has necessitated a new theory.

3.2 Quantum Mechanics

In 1905, Einstein postulated that energy and mass are related by his famous equa-
tion [279]:
E = mc2 . (3.11)

Here, E is the relativistic energy of a particle, and m is the relativistic mass of a

particle. The constant c is, as before, the speed of light in vacuum. One of the
implications of this equation is that the mass of a slow-moving particle is lower
than the mass of the same particle traveling at a higher speed.
For now, we will assume that a light wave is composed of photons2 . In that
case, the momentum of a photon traveling through a vacuum may be defined as

p = mc. (3.12)

Here, p is the relativistic momentum of the photon. The relativistic energy of the
photon is then given by
E = pc. (3.13)

For a given particle, such as an electron or a photon, Max Planck realized that
oscillations occur not at all frequencies, but only at specific frequencies. This
implies that the energy carried by a particle can only take on discrete values and
may be expressed in terms of frequency f :

E = nh f , n = 0, 1, 2, 3, . . . . (3.14)

2 This hints at the well-known wave-particle duality, which will be described in Section 3.2.1.

i i

i i
 i i

i i

3.2. Quantum Mechanics 125

Here, n is an integer value, and h is a constant of proportionality, known as

Planck’s constant. Its value is

h = 6.626176 × 10−34 J s. (3.15)

Thus, quantum theory derives its name from the quantization of energy levels.

3.2.1 Wave-Particle Duality

Light was explained as a wave phenomenon in the previous chapter. However,
several experiments demonstrate that the behavior of light can not be explained
solely by waves. Thus, the notion emerged that light is both. It loosely behaves as
a wave when traveling through free space and as a particle when it interacts with
matter.
This can be adequately demonstrated for light. The particles are called quanta,
except when the particles have frequencies in the visible range. Then, these parti-
cles are called photons.
Furthermore, the same wave-particle duality can be demonstrated for larger
particles such as electrons. If any particle can act as a wave, then associated with
it, there should be a wavelength. From (2.63), we have a relation between the
wavelength λ , the velocity of the wave v, and the frequency f . At the same time,
from (3.14) we have E = h f , and from (3.13) the energy of a photon is given as
E = pc. Combining these three equations, we get the de Broglie’s postulate:

h
λ= , (3.16)
p

where h is Planck’s constant. Thus, the wavelength of a wave/particle is related to

its momentum. The significance of this equation is that it directly links a quantity
associated with particles, namely their momentum, to a quantity λ associated with
waves.

3.2.2 Bohr’s Atom Model

In this section, we consider a very simple model of atoms, which we will refine
in later sections. Here, we assume that an atom consists of heavy particles such
as protons and neutrons forming a nucleus, with lighter particles orbiting around
the nucleus at some distance. This is known as Bohr’s model of the atom.
Protons have a positive charge, indicated with +e, whereas electrons have
a negative charge of −e. Neutrons have no charge. The number of protons in
a nucleus determines the atom’s atomic number, which is given the symbol Z.

i i

i i
 i i

i i

126 3. Chemistry of Matter

Normally, the number of protons and neutrons in a nucleus are identical. If these
numbers are different, then the atom is called an isotope.
The number of electrons surrounding the nucleus determines the charge of the
atom. If the number of electrons equals the number of protons in the nucleus, then
the total charge is zero. If the number of electrons and protons is not matched, the
charge is non-zero, and the atom is called an ion. If the charge is positive due to
a lack of electrons, the atom is called a cation. An excess of electrons causes the
atom to be negatively charged, and the atom is then called an anion.
We consider an atom with only a single proton and electron, i.e., a hydrogen
atom. Since the electron is in orbit around the proton, it neither gains nor loses
energy. If this were not the case, the electron would gravitate towards the proton
or spiral away from the proton.
The orbits of electrons are for now assumed to be circular. The angular mo-
mentum L of an electron can then be derived from de Broglie’s postulate. The
length of the orbit is related to its radius r by a factor of 2 π . If an electron is not
gaining or losing energy, then, after each orbit, it will return to the same position.
Ascribing a wave to this electron, the circumference should be an integral number
of wavelengths, i.e.,

2π r = nλ, n = 1, 2, 3, . . . . (3.17)

Substituting (3.16) then gives

nh nh
2π r = = , n = 1, 2, 3, . . . . (3.18)
p mv

Rearranging terms gives

nh
mvr = = L, n = 1, 2, 3, . . . . (3.19)
2π
Thus, the angular momentum is considered quantized and non-zero (because
n > 0). As an aside, the quantity 2hπ occurs frequently; it has been given its
own symbol h̄ = 2hπ . The expression for mvr allows us to solve for r, i.e., we can
compute the radius at which the electron is orbiting. However, the result depends
on its velocity, which is unknown.
To arrive at an expression for the radius, we proceed as follows. For a proton
with an electron in orbit, the Coulomb force is given by

e2
F= . (3.20)
r2

i i

i i
 i i

i i

3.2. Quantum Mechanics 127

From Newton’s second law, we know that this equals m a. The acceleration as a
result of the Coulomb force can be computed from the change in velocity after
the electron has rotated around the proton once. The magnitude of the velocity is
given by
2π r
v= , (3.21)
T
where T is the time required for one orbit. The change in magnitude Δv for a
single orbit is Δv = 2 π v. The acceleration is then simply a = Δv/T = 2 π v/T .
Using (3.21), this can be written as

v2
a= . (3.22)
r
We now use (3.20) and find the following relation:

e2 mv2
= . (3.23)
r2 r
To eliminate the unknown velocity, we use the expression for angular momentum
L to solve for v:
nh
v= . (3.24)
2π mr
Substituting into (3.23), we find

n2 h2
r= , n = 1, 2, 3, . . . . (3.25)
4 π 2 m e2
As a result of quantizing the angular momentum, we thus find that the distance
between the electron and the proton is quantized. Since all quantities are known,
the radius is a direct function of quantization level n. The resulting orbits are
called Bohr orbits.
Given these orbits, we can now compute the energy of the electron in terms of
its orbit. The total energy of a hydrogen atom is given by

mv2 e2
E = T +U = − (3.26a)
2 r
p 2 e2
= − . (3.26b)
2m r
Using the equations above, this may be restated as

e2 e2 e2
E= − =− . (3.27)
2r r 2r

i i

i i
 i i

i i

128 3. Chemistry of Matter

Substituting (3.25), the total energy associated with a hydrogen atom is given as

2 π 2 m e4 1
E =− , n = 1, 2, 3, . . . . (3.28)
h2 n2

The ground state of an atom occurs for n = 1. In this state, the energy is most
negative, and the atom is therefore most stable. For larger values of n, we speak of
excited states. As the energy for excited states is less negative than for the ground
state, these states are also less stable. The energy associated with the ground
state is the negative of the ionization potential, which is the energy required to
remove an electron. Thus, the quantization of the angular momentum has led to
the quantization of possible energy states of the atom.
Now, an electron may jump from one state to another, thereby changing the
total energy of the atom. If changing from state n1 to n2 , the corresponding change
in total energy is
 
2 π 2 m e4 1 1
ΔE = − − . (3.29)
h2 n22 n21

If this energy change is due to a jump to a lower excited state, i.e., n2 < n1 , then
this energy must have been dissipated in some manner, for instance by emission of
a photon. The energy of the photon must equal the change in energy of the atom.
We can now determine the frequency f of the photon, given that its energy equals
ΔE = h f . Solving for f yields an expression for the frequency of the photon:
 
2 π 2 m e2 1 1
f =− 2
− 2 . (3.30)
h3 n2 n1

With the frequency, as well as the medium through which the photon is traveling
known, the wavelength is also determined according to (2.63). As a result, the
change in energy state of an electron from a higher excited state to a lower excited
state has caused the emission of a photon with a wavelength determined by the
quantitative change in the atom’s energy. This change is quantized according to
the equations given above, and thus the color of the light emitted is quantized as
well.
Conversely, a photon may be absorbed by an electron. The extra energy im-
parted onto the electron is used to bring it to a higher excited state. This absorption
can only happen if the energy of the photon equals the amount of energy required
to make the electron jump between states of excitation. Thus, a photon of the right
wavelength may be absorbed, whereas photons of longer or shorter wavelengths
may pass through the atomic structure, or alternatively get deflected.

i i

i i
 i i

i i

3.2. Quantum Mechanics 129

Energy Frequency Color Wavelength

(eV) 10 14 (Hz) (nm)

5 12 250

Ultra-violet
10 300
4

8
Violet
3 400

Visible range
Blue
6 Green 500
Yellow
2 Orange 600
Red 700
4 800
1000
Infra-red

1
2

Figure 3.2. Energy diagram showing four different scales to represent energy.

The absorption of photons with specific wavelengths is thus governed by the

atomic structure of the matter with which the photon interacts. Different atoms
and molecules afford different possible states of excitation and will, therefore,
absorb photons with different wavelengths. Photons with other wavelengths are
not absorbed and cause reflected and transmitted light of different colors to appear.
As indicated above, the energy of excited states increases with increasing n,
i.e., becomes less negative. For E = 0, the electron is removed and the atom is
left in an ionized state. The electron is then said to be in free state or unbound
state. If more energy is added to the atom, if will be in the form of kinetic energy.
This energy may be added in any amount and is not limited to specific quantized
amounts.
Finally, the change of energy states can be conveniently represented in an en-
ergy diagram, as depicted in Figure 3.2. In addition to color, wavelength, and
frequency, a new unit of energy is introduced here, the electronvolt (eV; see also
Table D.2). This is a convenient unit to express the quantized energy of an elec-
tron, atom, or molecule. The conversion between wavelength λ and energy ΔE
(if specified in electron volts (eV )) is given by

λ ΔE = 1239.9. (3.31)

i i

i i
 i i

i i

130 3. Chemistry of Matter

3.2.3 The Time-Independent Schrödinger Equation

As particles are known to exhibit wave behavior, it may be assumed that a solu-
tion describing the wave behavior for light may be extended to include the wave
behavior of particles. If we associate a wave with the motion of an electron, its
solution in one dimension may be taken to be of the form
   
2π z 2π z
ψ = A sin + B cos . (3.32)
λ λ
This equation is of the same form as (2.59c), which we presented as a solution for
harmonic plane waves. However, we have removed the time-dependent compo-
nent. In addition, in this equation ψ is not a measure of energy, but the probability
of finding the electron at a particular position along the z-axis3 . We may substi-
tute (3.16) into this equation as follows:
   
2π pz 2π pz
ψ = A sin + B cos . (3.33)
h h
Under the assumption that the total energy of an electron remains constant, the
Hamiltonian for an electron is given by

p2
H =E = +U. (3.34)
2m
The equalities remain intact if all terms are multiplied by ψ . This yields the time-
independent Schrödinger equation:

H ψ = E ψ. (3.35)

In this formulation, the energy E of an electron is called the eigenvalue and ψ

is called the eigenfunction. The eigenfunctions correspond to orbitals. Thus, an
electron in orbital ψ has energy E. Equation (3.34) also yields

p2 ψ
= (E −U) ψ . (3.36)
2m
To derive an expression for the Hamiltonian operator, in this case, we begin by
differentiating ψ twice:
    
∂ 2ψ 4 π 2 p2 2π pz 2π pz
=− A sin + B cos (3.37a)
∂ z2 h2 h h
4 π2 2
=− p ψ. (3.37b)
h2
3 This expression will be extended to three dimensions later in this section.

i i

i i
 i i

i i

3.2. Quantum Mechanics 131

Rearranging terms and dividing left and right by 2 m yields

p2 ψ h2 ∂ 2 ψ
=− 2 , (3.38)
2m 8 π m ∂ z2
and, therefore,
h2 ∂ 2 ψ
− = (E −U) ψ . (3.39)
8 π 2 m ∂ z2
Thus, the Hamiltonian is now given in operator form as follows:

h2 ∂ 2 ψ
H =− +U. (3.40)
8 π 2 m ∂ z2
This operator may be extended to three dimensions by replacing the partial deriva-
tive in the z-direction with the Laplacian operator, as given in (A.19). The Hamil-
tonian operator is then formulated as

h2
H =− ∇2 +U. (3.41)
8π2 m

3.2.4 Quantum Numbers

The integer numbers n which characterize energy states of bound electrons are
called principal quantum numbers. The atomic structure may be further described
by additional quantum numbers.
As argued in Section 3.2.2, the angular momentum L is quantized. Empiri-
cally, it has been found that the square of the angular momentum can only have
specific values:

l(l + 1)h̄2 , l = 0, 1, 2, . . . , n − 1. (3.42)

The integer l is called the angular momentum quantum number, and its value,
describing the rotation of an electron around a nucleus, is strictly less than n.
This implies that for higher energy states, i.e., larger values of n, there are more
possible angular momentum states. The magnetic quantum number m can have
both positive and negative integral values between −l and l. It describes the
projection of the angular momentum of an electron onto an axis determined by an
external magnetic field [370]. The fourth and final quantum number is the spin
quantum number with values of either + 12 or − 12 ; it is a number describing the
rotation of an electron around itself.
An atom has a nucleus surrounded by an electron cloud. The electrons are
arranged in orbits (also called shells), with each orbit containing between zero

i i

i i
 i i

i i

132 3. Chemistry of Matter

l 0 1 2 3 4 5 ...
Symbol s p d f g h ...

Table 3.1. Symbols associated with values of the angular momentum quantum number l.

and two electrons. If there are two electrons in an orbit, then these will have
opposite spin.
By convention, orbital states are indicated with a value for n, followed by a
letter for l. The letter scheme is as indicated in Table 3.1. The first four letters
stand for sharp, principal, diffuse, and fundamental.The letters following these
four are arranged alphabetically [768]. Lowercase letters are used for single elec-
trons, whereas the total angular quantum number L of all the electrons in an atom,
taking values 0, 1, 2, 3, 4, 5, are given the uppercase letters S, P, D, F, G, H.
It is not possible for any two electrons in the same atom to have the same set
of quantum numbers. This is known as Pauli’s exclusion principle [750, 886].
Atoms are formed by filling orbitals in an order dictated by the lowest available
energy states.
The lowest energy orbit has principal quantum number 1 and has a single
atomic orbital, which is referred to as the s-orbital. If this orbital has one elec-
tron, then the electronic configuration corresponds to hydrogen (H), whereas the

n l m s Designation Number of electrons

1 0 0 + 12 ,− 12 1s 2
2 0 0 + 12 ,− 12 2s 2
2 1 -1 + 12 ,− 12
2 1 0 + 12 ,− 12 2p 6
2 1 1 + 12 ,− 12
3 0 0 + 12 ,− 12 3s 2
3 1 -1 + 12 ,− 12
3 1 0 + 12 ,− 12 3p 6
3 1 1 + 12 ,− 12
3 2 -2 + 12 ,− 12
3 2 -1 + 12 ,− 12
3 2 0 + 12 ,− 12 3d 10
3 2 1 + 12 ,− 12
3 2 2 + 12 ,− 12

Table 3.2. Quantum numbers associated with the lowest electron orbitals.

i i

i i
 i i

i i

3.2. Quantum Mechanics 133

Atom Configuration
Lithium Li 1s2 2s1 [He] 2s1
Beryllium Be 1s2 2s2 [He] 2s2
Boron B 1s2 2s2 2p1 [He] 2s2 2p1
Carbon C 1s2 2s2 2p2 [He] 2s2 2p2
Nitrogen N 1s2 2s2 2p3 [He] 2s2 2p3
Oxygen O 1s2 2s2 2p4 [He] 2s2 2p4
Fluorine F 1s2 2s2 2p5 [He] 2s2 2p5
Neon Ne 1s2 2s2 2p6 [He] 2s2 2p6

Table 3.3. Electron configurations for atoms with electrons in the first two shells.

two-electron case corresponds to helium (He). Their electronic configurations are

indicated with 1s1 and 1s2 , respectively. The notation names the principal quan-
tum number first (1), then the orbital (s), and the superscript indicates the number
of electrons in this orbital.
For a principal quantum number of 2, the next lowest shell contains one s-
orbital and three p-orbitals. All four orbitals have the same energy, and each can
hold up to two electrons. The p-orbitals are differentiated by a subscript: px , py ,
and pz . An electron in one of the 3p orbital indicates that n = 3 and l = 1. The
quantum numbers of several of the lowest electron orbitals are given in Table 3.2.
The electron configurations of atoms with electrons in the first two shells are given
in Table 3.3. As the first shell is always completely filled, an abbreviated notation
exists, listing the name of the atom rather than its corresponding electronic con-
figuration. The electronic configuration of Lithium is 1s2 2s1 and can therefore be
abbreviated to [He] 2s1 .
Atoms with principal quantum number 3 have three shells, with the first two
shells completely filled with electrons. The third shell has one s-orbital, three p-
orbitals, and five d-orbitals. Atoms with empty d-orbitals are listed in Table 3.4.
The d-orbitals are indicated with the following subscripts: dxy , dxz , dyz , dx2 −y2 ,
and dz2 . Each orbital takes at most two electrons, for a maximum of 10 electrons
in the d-orbitals. The atoms which have partially filled d-orbitals are called tran-
sition metals. Their electronic configuration is given in Table 3.5. Note that the
energy levels of the 4s orbital and the 3d orbitals are very similar, causing the 4s
orbital usually to be filled before the 3d orbitals.
The inert gases all have their s and p-orbitals fully filled. Their abbreviated
configurations are given in Table 3.6.
All atoms can be arranged according to their electronic states and chemical
behavior, such as in the periodic table of elements (Figure 3.3). The rows in this
table are called periods, whereas the columns are called groups.

i i

i i
 i i

i i

134 3. Chemistry of Matter

Atom Configuration

Sodium Na 1s2 2s2 2p6 3s1 [Ne] 3s1

Magnesium Mg 1s2 2s2 2p6 3s2 [Ne] 3s2
Aluminium Al 1s2 2s2 2p6 3s2 3p1 [Ne] 3s2 3p1
Silicon Si 1s2 2s2 2p6 3s2 3p2 [Ne] 3s2 3p2
Phosphorus P 1s2 2s2 2p6 3s2 3p3 [Ne] 3s2 3p3
Sulphur S 1s2 2s2 2p6 3s2 3p4 [Ne] 3s2 3p4
Chlorine Cl 1s2 2s2 2p6 3s2 3p5 [Ne] 3s2 3p5
Argon Ar 1s2 2s2 2p6 3s2 3p6 [Ne] 3s2 3p6

Table 3.4. Electron configurations for atoms with electrons in the first three shells (ex-
cluding d-orbitals).

Atom Configuration

Scandium Sc 1s2 2s2 2p6 3s2 3p6 3d1 4s2 [Ar] 3d1 4s2
Titanium Ti 1s2 2s2 2p6 3s2 3p6 3d2 4s2 [Ar] 3d2 4s2
Vanadium V 1s2 2s2 2p6 3s2 3p6 3d3 4s2 [Ar] 3d3 4s2
Chromium Cr 1s2 2s2 2p6 3s2 3p6 3d5 4s1 [Ar] 3d5 4s1
Manganese Mn 1s2 2s2 2p6 3s2 3p6 3d5 4s2 [Ar] 3d5 4s2
Iron Fe 1s2 2s2 2p6 3s2 3p6 3d6 4s2 [Ar] 3d6 4s2
Cobalt Co 1s2 2s2 2p6 3s2 3p6 3d7 4s2 [Ar] 3d7 4s2
Nickel Ni 1s2 2s2 2p6 3s2 3p6 3d8 4s2 [Ar] 3d8 4s2
Copper Cu 1s2 2s2 2p6 3s2 3p6 3d10 4s1 [Ar] 3d10 4s1
Zinc Zn 1s2 2s2 2p6 3s2 3p6 3d10 4s2 [Ar] 3d10 4s2

Table 3.5. Electron configurations for the 3d transition metals.

Atom Abbreviated configuration

Helium He 1s2
Neon Ne [He] 2s2 2p6
Argon Ar [Ne] 3s2 3p6
Krypton Kr [Ar] 3d10 4s2 4p6
Xenon Xe [Kr] 4d10 5s2 5p6

Table 3.6. Electron configurations of inert gases.

i i

i i
 i

i
i

i
3.2. Quantum Mechanics
Ia II a III b IV b Vb VI b VII b VIII Ib II b III a IV a Va VI a VII a O

1 H 2 He
1.0079 4.0026
Hydrogen Helium
Colorless Colorless
Legend: Color coding:

3 Li 4 Be 5 B 6 C 7 N 8 O 9 F 10 Ne
6.941 9.01218 Atomic number Symbol Metals 10.81 12.011 14.0067 15.9994 19.9984 20.179
Lithium Beryllium Atomic weight Transition Metals Boron Carbon Nitrogen Oxygen Fluorine Neon
Silvery white Lead gray Name Metalloids Black Black Colorless Pale blue Pale yellow Colorless
gray Color Non-Metals Colorless
Inert Gases
Figure 3.3. The periodic table of elements.

11 Na 12 Mg 13 Al 14 Si 15 P 16 S 17 Cl 18 Ar
22.98977 24.305 26.98154 28.086 30.97376 32.06 35.453 39.948
Sodium Magnesium Aluminium Silicon Phosphorus Sulfur Chlorine Argon
Silvery white Silvery white Silvery Dark gray, Colorless/red Lemon Yellowish Colorless
bluish tinge silvery white yellow green

19 K 20 Ca 21 Sc 22 Ti 23 V 24 Cr 25 Mn 26 Fe 27 Co 28 Ni 29 Cu 30 Zn 31 Ga 32 Ge 33 As 34 Se 35 Br 36 Kr
39.098 40.08 44.9559 47.90 50.9414 51.996 54.9380 55.847 58.9332 58.70 63.546 65.38 69.72 72.59 74.9216 78.96 79.904 83.80
Potassium Calcium Scandium Titanium Vanadium Chromium Manganese Iron Cobalt Nickel Copper Zinc Gallium Germanium Arsenic Selenium Bromine Krypton
Silvery white Silvery white Silvery white Silvery Gray Silvery Silvery Metallic, Metallic, Metallic, Copper, Bluish, pale Silvery white Grayish Metallic Metallic Red-brown, Colorless
metallic metallic metallic metallic grayish tinge grayish tinge silvery tinge metallic gray white gray gray metallic

37 Rb 38 Sr 39 Y 40 Zr 41 Nb 42 Mo 43 Tc 44 Ru 45 Rh 46 Pd 47 Ag 48 Cd 49 In 50 Sn 51 Sb 52 Te 53 I 54 Xe
85.4648 87.62 88.9059 91.22 92.9064 95.94 98.9062 101.07 102.9055 106.4 107.868 112.40 114.82 118.69 121.75 127.60 126.9045 131.30
Rubidium Strontium Yttrium Zirconium Niobium Molybdenum Technetium Ruthenium Rhodium Palladium Silver Cadmium Indium Tin Antimony Tellurium Iodine Xenon
Silvery white Silvery white Silvery white Silvery white Gray Gray Silvery gray Silvery white Silvery white Silvery white Silver Silvery gray Silvery Silvery Silvery Silvery Violet-dark Colorless
metallic metallic metallic metallic metallic metallic metallic gray gray gray gray gray

55 Cs 56 Ba 71 Lu 72 Hf 73 Ta 74 W 75 Re 76 Os 77 Ir 78 Pt 79 Au 80 Hg 81 Tl 82 Pb 83 Bi 84 Po 85 At 86 Rn
132.9054 137.34 174.97 178.49 180.9479 183.85 186.207 190.2 192.22 195.09 196.9665 200.59 204.37 207.2 208.9804 210 211 222
Caesium Barium Lutetium Hafnium Tantalum Tungsten Rhenium Osmium Iridium Platinum Gold Mercury Thallium Lead Bismuth Polonium Astatine Radon
Silvery gold Silvery white Silvery white Gray steel Gray blue Grayish Grayish Bluish gray Silvery white Grayish Gold Silvery white Silvery white Bluish Reddish Silvery Metallic Colorless
white white white white white

87 Fr 88 Ra 103 Lr 104 Rf 105 Db 106 Sg 107 Bh 108 Hs 109 Mt 110 Ds 111 Rg 112 Uub
223 226.0254 257 257 264 266 264 269 268 281 272 285
Francium Radium Lawrencium Rutherfordium Dubnium Seaborgium Bohrium Hassium Meitnerium Darmstadtium Roentgenium Ununbium
Metallic Metallic Unknown Unknown Unknown Unknown Unknown Unknown Unknown Unknown Unknown Unknown

57 La 58 Ce 59 Pr 60 Nd 61 Pm 62 Sm 63 Eu 64 Gd 65 Tb 66 Dy 67 Ho 68 Er 69 Tm 70
138.9055 140.12 140.9077 144.24 147 150.4 151.96 157.25 158.9254 162.50 164.9304 167.26 168.9342 Yb
Lanthanide
Lanthanum Cerium Praseodymium Neodymium Promethium Samarium Europium Gadolinium Terbium Dysprosium Holmium Erbium Thulium 173.04
series
Silvery white Silvery white Silvery white Silvery white Metallic Silvery white Silvery white Silvery white Silvery white Silvery white Silvery white Silvery white Silvery white Ytterbium
yellowish yellowish

89 Ac 90 Th 91 Pa 92 U 93 Np 94 Pu 95 Am 96 Cm 97 Bk 98 Cf 99 Es 100 Fm 101 Md 102 No

227 232.0381 231.0359 238.029 237.0482 239.11 241 244 249 252 253 254 256 254
Actinide
Actinium Thorium Protactinium Uranium Neptunium Plutonium Americium Curium Berkelium Californium Einsteinium Fermium Mendelevium Nobelium
series
Silvery Silvery white Silvery Metallic Silvery Silvery white Silvery white Silver Unknown Unknown Unknown Unknown Unknown Unknown
metallic gray metallic

135
i

i
i

i
 i i

i i

136 3. Chemistry of Matter

x
z
s

px py pz

d yz d xz d xy d x2 - y 2 d z2

Figure 3.4. A schematic of s-, p-, and d-orbitals. Note that this is a representation of a
probability density function, and that therefore the boundary is not sharp in reality, nor is
the shape accurate.

3.2.5 Orbital Shapes and Bonds

In Section 3.2.3, the time independent Schrödinger equation was presented. A
solution to this equation is a wave function ψ . Squaring and normalizing ψ yields
a function which indicates the electron cloud density. It is therefore a function of
the probability of finding an electron in any particular position around a nucleus.
For s orbitals, the electron cloud density is spherical, with the center coin-
ciding with the nucleus. There are three p-orbitals in the 2p shell, which are each
dumbbell-shaped and orthogonal to each other. The p-orbitals are aligned with the
x-, y-, and z-axes. The dxy -, dxz -, and dyz -orbitals have lobes that lie in between
the axes, whereas the dx2 -y2 - and dz2 -lobes are once more axis aligned [1128]. A
schematic of s-, p-, and d-orbitals is given in Figure 3.4.
When a pair of hydrogen atoms form a hydrogen molecule, their 1s orbitals
may overlap forming a bond between the two atoms. If the wave-functions asso-
ciated with both orbitals have the same sign, then we speak of bonding orbitals.
If the signs are opposite, the electrons in the two orbitals repel each other. This
is called anti-bonding . The combined 1s orbitals are termed σ -orbitals and their
associated bonds are called σ -bonds and σ ∗ -anti-bonds.

i i

i i
 i i

i i

3.2. Quantum Mechanics 137

σ-bonds

π-bonds

Figure 3.5. A schematic of bonding between s- and p-orbitals.

The three 2p-orbitals also form bonds, see Figure 3.5. In particular, the px -
orbitals of two atoms may directly overlap, forming a strong 2pσ -bond or 2pσ ∗ -
anti-bond. The two 2py - and 2pz -orbitals can also bond, forming weaker π -bonds
and π ∗ -anti-bonds, dependent on whether the wave-functions have identical or
opposite signs.
Molecules can bond more than once, i.e., double or triple bonding can occur.
For instance, oxygen (O2 ) produces a double bond—one 2pσ -bond due to the px -
orbital and a 2pπ -bond due to either the py - or pz -orbital. Nitrogen N2 , which has
a 2s2 2p3 configuration, forms a triple bond involving px -, py -, and pz -orbitals, as
shown in Figure 3.6.

double bond triple bond

Figure 3.6. A schematic of double and triple bonding.

Each bonding between atoms creates an energy level lower than the sum of
the energies of the orbitals associated in the bonding. Similarly, anti-bonding
involves a higher energy level than two unbonded orbitals. As such, we may
expect that molecules with orbitals when bonding with orbitals in other atoms,
absorb photons with different energies than single atoms. Thus, much of the
richness and variety of color in substances is due to the formation of bonds in
molecules, or to the presence of atoms and ions inside a crystal lattice consisting
of different types of atoms.

i i

i i
 i i

i i

138 3. Chemistry of Matter

Figure 3.7. The strong ionic bonding between Na+ and Cl − causes absorption at wave-
lengths outside the visible range. As a result, kitchen salt is colorless.

3.3 Atoms and Ions

In atoms, the energy states that electrons can attain are determined by the set of
allowable transitions between orbitals. In molecules, which consist of multiple
atoms, the electrons are involved in the bonding between individual atoms. The
strength of the bond is an indicator of the amount of energy required to free an
electron.
Thus, the ability of a molecule to selectively absorb light of specific wave-
lengths, is determined by the allowable energy states governed by the bonding
between atoms. For particularly strong bonds, only highly energetic photons can
be absorbed. Such photons are in the ultra-violet (UV) range of wavelengths, and
are therefore not visible. As a result, substances with particularly strong bonds
tend to appear colorless.
An example is the ionic bonding in salt (Figure 3.7), which consists of a
sodium atom and a chlorine atom. The sodium atom donates the electron in its
outermost shell to the chlorine atom. As a result, the two atoms become sodium
and chloride ions, Na+ and Cl− , each having a fully filled outer orbit, so that all
electrons are paired. Having all electrons paired creates a very stable bonding, so
that only highly energetic photons can be absorbed.
In contrast to ionic bonding, there is the situation where an electron in the
outer shell is shared equally by two atoms. Such covalent bonding can also pro-
duce very strong bonds between atoms, as evidenced by diamond. The strength

i i

i i
 i i

i i

3.3. Atoms and Ions 139

of the bonds in this cubic form of carbon, resulting from the pairing of electrons,
causes diamond to be colorless.

3.3.1 Color from 3d Transition Metals

In many inorganic compounds, the bonding is neither fully ionic, nor fully cova-
lent. An example is corundum, Al2 O3 , which has all its electrons paired off in
a configuration which is about 60% ionic and 40% covalent. The strength of the
bonding again causes the material in its pure form to be colorless (it is then known
as white sapphire).
However, if the material contains about 1% chromium sesquioxide, Cr2 O3 ,
the material becomes red and is then known as ruby. This color is a result of
unpaired electrons that can be excited to various states making transitions that
correspond to a release of radiation in the visible range, i.e., fluorescence.
It is the 3d-orbitals and, in particular their shape, that are involved in the
formation of color. In a single atom, each of the five different 3d-orbitals has
the same energy. However, the d-orbitals can be classified in two groups. These
are the dxy -, dxz -, and dyz -orbitals, which have lobes lying in between the axes,
and the dx2 −y2 - and dz2 -orbitals, which are axis aligned. The former group is
collectively called the t2g or t2 set, whereas the latter group is called the eg or e
set [1128].
If the atom is located in a crystal, then there are nearby electrons which re-
pel the electrons in the d-orbitals. This changes the shape of the d-orbitals by an
amount that depends on their orientation. The energy associated with some of
the d-orbitals increases more than others in the presence of a surrounding crys-
tal lattice. This phenomenon is called crystal field splitting or ligand field split-
ting [1128]. Of course, the shape of the surrounding crystal lattice, as well as the
distance to the transition metal, also play a role in which d-orbitals are raised in
energy, and by how much.
The energy levels associated with the d-orbitals in a transition metal lodged in
a crystal lattice determine which photons may be absorbed, and thereby determine
the absorption spectrum of the material.
If ruby is illuminated with white light, then electrons in Cr3+ ions are ex-
cited from their ground state E0 to two upper levels, E2 and E3 , leading to two
broad absorption spectra. The transition from E0 to E2 corresponds to a range of
wavelengths centered around 556 nm (yellow/green). The transition from E0 to
E3 corresponds to a range of wavelengths centered around 407 nm (violet). The
transition from E0 to E1 is not within the visible range of wavelengths. The light
that is not absorbed is predominantly red, i.e., the color of ruby [1128].

i i

i i
 i i

i i

140 3. Chemistry of Matter

A second example is given by emerald. The shape of the crystal lattice is the
same, albeit somewhat larger. The impurities are also formed by Cr3+ ions. This
results in an energy diagram which is essentially the same, although the upper
energy levels are closer to the ground state. This shifts the absorption spectrum
towards red, yielding the green color with a hint of blue typically seen in emeralds.
The impurities that give ruby and emerald their color (Cr2 O3 in this case),
are called transition metal impurities. The resulting colors are called allochro-
matic transition metal colors . Minerals may be colored with chromium, cobalt,
iron, manganese, nickel, samarium, and vanadium impurities, and, in general, the
transition metals listed in Table 3.5.
If the transition element is also the main compound, we speak of idiochro-
matic transition metal colors. Elements involved in idiochromatic transition met-
als are chromium, manganese, iron, cobalt, nickel, and copper. Green, for in-
stance, can be formed in several ways in crystals, including chrome green
(Cr2 O3 ), manganese (II) oxide (MnO), melantite (Fe2 O3 ), cobalt (II) oxide
(CoO), bunsenite (NiO), and malachite (Cu2 (CO3 )(OH)2 ). An example of a red
colorization is rhodochrosite (MnCO3 ), whereas blue can be formed for instance
by Thenard’s blue (Al2 CoO4 ), azurite (Cu3 (CO3 )2 (OH)2 ), and turquoise
(CuAl6 (PO4 )(OH)8 ) [815].

3.3.2 Charge Transfer Color

If multiple transition metals are present as impurities, it may become possible
for an electron to move from one transition metal ion to another and absorb a
photon in the process. As an example, corundum with 1% of titanium or iron
does not have color (titanium) or a very pale yellow color (iron). However, if both
impurities are present, inter-valence charge transfer can occur:

Fe2+ + Ti4+ → Fe3+ + Ti3+ (3.43)

Here, an electron has moved from the iron ion to the titanium ion. The difference
in energy between the two states is 2.11 eV. Thus, absorption of photons with en-
ergies of 2.11 eV is possible, resulting in a deep blue color. Corundum containing
both titanium and iron is known as blue sapphire.
The crystal structure of corundum is such that the above charge transfer occurs
for ions that are 2.65 Å apart.4 However, the crystal lattice also allows Ti and Fe
ions at a distance of 2.79 Å, resulting in a weaker second absorption which creates
a blue-green color. These two configurations occur at right angles, which means
4 The Ångstrom is a measure of distance, and equals one tenth of a nanometer.

i i

i i
 i i

i i

3.3. Atoms and Ions 141

that the blue and blue-green colors are seen dependent on the orientation of the
crystal. Blue sapphire thus exhibits dichroism.
Other examples of colors caused by charge transfer include many pigments,
such as Prussian blue, as well as yellow, brown, and black colors seen in rocks
and minerals that contain iron.

3.3.3 Metals and Semiconductors

Metals are characterized by high conductivity, which increases with temperature.
The opposite of conductivity is resistivity, which is therefore low in metals. Di-
electrics, on the other hand, have high resistivity, and thus do not conduct elec-
tricity very well. Materials that fall in between these extremes are called semicon-
ductors and are characterized by a decrease in conductivity when the temperature
is raised.
To explain the properties of metals and semiconductors, one may resort to
band theory, which can be seen as a special case of molecular orbital theory. To
begin, we consider a pair of hydrogen atoms approaching each other from infinity
(after [815]). Each atom has an electron in the 1s orbital. If these two electrons
have parallel spins, they will repulse more and more as they approach, creating a
higher energy. If the two electrons have opposite spin, then the energy will fall as
the two atoms approach. In this case, as the atoms move very close together, the
nuclei will start to repulse each other and will counteract the attraction between
the two opposite spinning electrons.
Energy

Antibonding
Bonding

>0

Normal bond length of H 2

<0
0 0.74
Distance (Å)

Figure 3.8. Morse curve showing energy as a function of distance. The inter-atomic
distance is normally measured in Ångstrom, which is one-tenth of a nanometer (see also
Figure 2.4).

i i

i i
 i i

i i

142 3. Chemistry of Matter

Figure 3.8 shows the energy of the ensemble of two hydrogen atoms as func-
tion of the distance between the two atoms. The top curve is for two electrons hav-
ing parallel spin, whereas the bottom curve is for electrons having opposite spin.
These types of plot are called Morse curves or potential energy diagrams. For
hydrogen with electrons in opposite spin, the distance between atoms is 0.74 Å.
The molecular hydrogen orbitals in one H2 molecule can accommodate the
same number of electrons in two orbitals as two atomic hydrogen orbitals in one
orbital each. Extending this approach by involving N hydrogen atoms, the number
of orbitals and energy levels increases by N. These energy levels will be very close
together (and depend on the inter-atoms distance between each pair of hydrogen
atoms), such that individual levels can not be distinguished. Thus, a large number
of hydrogen atoms in close proximity creates an energy band.
Typically, electrons in partially filled outer orbitals form bands, whereas fully
filled inner orbitals are not involved in bands unless their inter-atomic distance
becomes very small. Unfilled shells can form bands at much larger distances, as
shown in Figure 3.9.
For some materials, two or more of the outer orbitals, such as the 2s- and 2p-
orbitals of lithium, each broaden into two bands [781]. For metals, these bands
may overlap, causing interactions between electrons in different bands. This aids
conductivity in metals.
Energy

Empty outer
orbitals

Partly filled
valence orbitals

Filled inner
shells

0 d
Distance (Å)

Figure 3.9. Bands formed by clusters of atoms (after [815]).

i i

i i
 i i

i i

3.3. Atoms and Ions 143

The band structure of metals has the effect that light is strongly absorbed at
many different wavelengths. Refracted light usually travels into the metal for
less than a single wavelength (several hundred atoms), before it is completely
absorbed. This absorption creates a current through the metal near the surface,
which in turn produces new photons which appear as reflected light. As a conse-
quence, metals strongly reflect light.
Within the band structure, some transitions in energy levels are favored over
others. While the precise mechanism is currently poorly understood [815], the
limitations of state transitions give rise to a wavelength dependent reflectivity as
a function of the atomic structure of different metals. For instance, silver reflects
light over a broader range of wavelengths than gold, which reflects less light at
shorter wavelengths.
In non-conducting materials, bands formed by different orbitals may repel,
as in the case of the 2s- and 2p-electrons in diamond. This causes a gap to occur
between the 2s- and 2p-bands, which for diamond are well separated at the carbon
bond distance of 1.54 Å. The lower energy band is called the valence band and
the upper band the conduction band. As the conduction band is at a higher energy
level than the most energetic electrons in the material,5 the conduction band is
empty, and diamond therefore acts as an insulator [815].
The different atomic structures of materials thus give rise to different interac-
tions between bands, as well as different sizes of the energy gap between bands. If
the bands overlap in energy level, then we speak of metals. If there is a large gap
between bands, such as the 5.4 eV band gap of pure diamond, the material is an
insulator. If the gap is relatively small, then we speak of semiconductors—a class
of materials that plays a crucial role in the manufacture of transistors and chips.
The energy, and therefore the wavelength, of light that can be absorbed is
dependent on the size of the band gap. Light needs to have sufficient energy to
excite an electron such that it can jump from the valence band into the conduction
band. For diamond the band gap is so large that no visible light can be absorbed
and, hence, the diamond appears colorless.
For a medium band-gap semiconductor, such as the pigment cadmium yellow
(CdS), the band gap is 2.6 eV, and this is small enough to absorb some light in the
violet and blue range. The pigment vermillion (HgS) has a yet smaller band gap
of 2.0 eV, allowing the absorption of most visible wavelengths except red. Band
gaps below 1.77 eV allow absorption of all visible wavelengths, and materials
with small band gaps therefore appear black.
It is possible to add impurities to pure semiconductors. This involves replac-
ing a small amount of semiconductor material with a different substance. Such
5 The energy of the most energetic electrons in a material is called the Fermi energy or Fermi level.

i i

i i
 i i

i i

144 3. Chemistry of Matter

impurities are called activators or dopants. For instance, adding a small concen-
tration of nitrogen (N) to diamond will introduce extra electrons that form a new
energy level within the band gap, called a donor level. As a result, the band gap is
effectively split into two smaller gaps, and now light in the blue and violet range
can be absorbed. Thus, diamond containing some nitrogen impurities will have a
yellow color.
Similarly, impurities may be added that create an electron deficiency, and
therefore they may insert an acceptor level within the band gap. Wherever
an electron deficiency exists, we speak of a hole. An example is the addition of
boron which, when added to diamond, causes a blue coloration.

3.4 Molecules
Molecules consist of collections of atoms that form bonds. Collections of atoms in
a single molecule can be of a particular form and can occur in different molecules.
Such collections are called groups. Some groups form bonds with energy levels
enabling the absorption of photons within the visible range of wavelengths. In
that case, the group is said to be color-bearing.
Other groups may bring about shifts in intensity and/or wavelength. Groups
that increase or decrease intensity are referred to as hyperchromic and hypo-
chromic, respectively. Substances that aid in color shifts are called bathochromic
if the shift is towards the red end of the spectrum and hypsochromic if the shift is
towards the blue end.
In (organic) molecules, color-bearing groups of atoms are called chromo-
phores. Examples of chromophores include carbon-carbon double bonds as well
as nitrogen double bonds. The molecules containing such chromophores are
called chromogens, although they frequently have other groups as well that are
involved in color formation. These extra groups are called auxochromes [815,
1128]. The auxochromes shift as well as strengthen the color, i.e., they are both
bathochromic and usually hypochromic. Auxochromes include NH2 , NO2 , and
CH3 structures as well as NHR, and NR2 , where R is a more complex organic
group.
Several important classes of organic molecules exists, and these are briefly
discussed in the following sections.

3.4.1 Porphyrins
Carbon atoms often form double bonds, yielding an absorption spectrum that
peaks outside the visible range. However, chains of carbon atoms that alternate

i i

i i
 i i

i i

3.4. Molecules 145

Aldehyde Chain length λmax (in nm)

CH3 -(CH=CH)-CH3 1 220
CH3 -(CH=CH)2 -CH3 2 270
CH3 -(CH=CH)3 -CH3 3 312
CH3 -(CH=CH)4 -CH3 4 343
CH3 -(CH=CH)5 -CH3 5 370
CH3 -(CH=CH)6 -CH3 6 393
CH3 -(CH=CH)7 -CH3 7 415

Table 3.7. Peak absorption wavelength as a function of the number of conjugated double
bonds (after [1099]).

between single and double bonds create absorption spectra that are shifted toward
the visible range. Such chains are called conjugated double bonds. For instance,
a chain of six such double bonds:

CH3 − (CH = CH)6 − CH3 (3.44)

reaches the blue end of the spectrum. Molecules containing such a group therefore
appear yellow. Examples are α - and β carotene, which appear in carrots. The
number of conjugated double bonds determines where the peak in the absorption
spectrum lies. The π -electrons in such chains are mobile across the chain. Adding
further double bonds increases their mobility and, thereby, the number of resonant
configurations. This lowers the energy level of the lowest excited level and, thus,
shifts the absorption spectrum further into the visible range [767]. Table 3.7 lists
the peak absorption wavelength for several aldehydes.
Conjugated bonds also appear in more elaborate molecular structures, for in-
stance, in ones where bonds circle a single metal atom. Such structures are called
porphyrins, with chlorophyll being one of the most important naturally occur-
ring one, giving plants and algae their green color (Figure 1.1). The chlorophyll
molecule has a magnesium atom at its center.
Iron is the center of a different porphyrin, called heme, which in many species
is part of the hemoglobin molecule. This molecule is responsible for transporting
iron through the bloodstream. Heme is responsible for the color of blood. How-
ever, this color is derived from π to π ∗ transitions, rather than from the presence
of an iron atom.

3.4.2 Photochromic Compounds

Some groups of molecules change color under the influence of light. If such
a change is reversible, then the molecule is said to be photochromic. In par-

i i

i i
 i i

i i

146 3. Chemistry of Matter

ticular, some molecules turn colorless when illuminated, a process called called
bleaching.
The most important photochromic molecules are the ones that transduce light
in the visual system. For instance, the molecules present in the rods, called
rhodopsin, have a red-purple color, changing to colorless under the influence of
light. Under illumination, rhodopsin transforms via several intermediate forms,
to a range of bleached molecules called metarhodopsins. In the dark, these mole-
cules revert back to rhodopsin, through a range of intermediate forms.
The chromophore in rhodopsin is called 11-cis-retinal. It is connected to an
opsin through the amino acid lycene. The opsin acts as an auxochrome and causes
a bathochromic shift by raising the peak of the absorption spectrum of 11-cis-
retinal from 370 nm to 570 nm.
Under medium light intensities, the incidence of a photon causes the 11-cis-
retinal to change to the trans form, called all-trans-retinal. The reaction that ef-
fects this change also causes a chain of events that leads to the rod emitting a
signal, thereby mediating vision. The configuration of all-trans-retinal connected
to the opsin via lycene is called metarhodopsin.
In cones, the retinal is believed to be connected to different opsins, which
cause different bathochromic shifts. This creates different peak sensitivities in the
three cone types, thereby enabling color vision.
Under low light levels, the trans-retinal may become fully dislodged from
its opsin and will be transformed back to its cis form by means of enzymes lo-
cated in the eye. Under yet lower light conditions, the trans-retinal may leave
the eye altogether. It is then carried through the bloodstream to the liver, where
it is regenerated into the cis-form. This complex sequence of events under low
light conditions causes dark adaptation to progress slowly over a long period of
time [1128] (see Section 4.3.2).
Another application of a photochromic material is found in sunglasses that
change to a darker state under UV illumination. Hence, outside in the sun, they
will be darker than indoors.

3.4.3 Colorants
Color may be given to objects by means of colorants, which is the general term
used for substances that produce color. A distinction is normally made between
pigments and dyes, as seen in the definition of pigments offered by the Color
Pigments Manufacturers Association (CPMA) [678]:

Pigments are colored, black, white or fluorescent particulate organic and

inorganic solids which are usually insoluble in, and essentially physically

i i

i i
 i i

i i

3.4. Molecules 147

and chemically unaffected by, the vehicle or substrate in which they are in-
corporated. They alter appearance by selective absorption and/or by scatter-
ing of light.
Pigments are usually dispersed in vehicles or substrates for application,
as for instance in inks, paints, plastics or other polymeric materials. Pigments
retain a crystal or particulate structure throughout the coloration process.
As a result of the physical and chemical characteristics of pigments, pig-
ments and dyes differ in their application; when a dye is applied, it penetrates
the substrate in a soluble form after which it may or may not become insolu-
ble. When a pigment is used to color or opacify a substrate, the finely divided
insoluble solid remains throughout the coloration process.

If after application the substrate evaporates, leaving only the pigment particles
behind, we speak of an ink; an example is watercolor. In other cases, the substrate
may harden and form a protective coating, as for instance in oil paints.
Dyes, on the other hand, are usually dissolved in a solvent [1247]. The dye
molecules do not stay together, but mix throughout the solvent. These molecules
absorb light selectively, which gives the solution its color. In between dyes and
pigments is a third category of colorants, called lakes. These are essentially pig-
ments in which the color of the particles is created by dyeing. Before dyeing,
the particles, which are typically made of aluminium oxide, are colorless and
translucent.
Almost all colorants can be synthetically produced, with the exception of
chlorophyll, which is used to color, for instance, soap and chewing gum green.
For yellow, orange, and red colorants, a single absorption band in the blue-violet-
green region is required. A single absorption band in the red-orange-yellow region
creates blue and violet colors. However, to produce green, two absorption bands
in the red and violet parts of the spectrum are required, which presents a difficult
problem for designers of organic colorants [815]. Using a mixture of yellow and
blue colorants tends to produce less saturated colors than a single green colorant
and may cause a change in color when one of the two colorants fades faster than
the other.
There are many ways to specify a colorant. One is the color index, which
is produced by the Society of Dyers and Colourists and the American Associ-
ation of Textile Chemists and Colorists, and contains some 8000 dyes and pig-
ments [1067]. In addition, pigments and dyes can have many designations, given
by different associations and agencies. For instance color index C.I. 14700, which
was formerly used to color food, is also known as D & C Red No 4 by the U.S.
Food and Drug Administration, as well as Food Red 1 and Ponceau SX [815].
Some substances are neither dyes or pigments, such as substances used to color
glasses, glazes, and enamels.

i i

i i
 i i

i i

148 3. Chemistry of Matter

3.4.4 Dyes
As dye molecules are fully dissolved, the interaction of light with dyes is rela-
tively straightforward and follows Beer’s Law (see (2.191)). This means that the
transmittance of a material depends directly on its concentration of dye molecules.
Similarly, the amount of light absorbed depends on the thickness of the layer, as-
suming that the dye is applied in layers.
If a dye is constructed by mixing multiple types of dye, then the wavelength-
dependent transmittance of the mixtures can be inferred from the transmittance
functions Ti of each dye i separately by simply multiplying the transmittance func-
tions together for each wavelength:
n
Tmix (λ ) = ∏ Ti (λ ). (3.45)
i=0

Changing the concentration of a colorant may alter its perceived color. Several
mechanisms contribute to this, including scattering, which for instance causes
the white head on a glass of beer. A second mechanism is dye dichroism,6 as
illustrated in Figure 3.10. The two curves in this figure show absorption in the
same region, which is partly in the visible range. The low concentration absorbs a
small amount in the violet region, causing a yellow color. As the concentration is
increased, a more significant portion of the visible spectrum is absorbed, including
blue and green, yielding an orange to red color. This effect would not occur
Absorption

Visible range

Dye concentration:
High
Low

200 250 300 350 400 450 500 550 600 650
Wavelength (nm)

Figure 3.10. Dye dichroism as a result of varying concentrations of dye in a solution.

6 This phenomenon is to be distinguished from the orientation-sensitive dichroism discussed in

Section 3.7.

i i

i i
 i i

i i

3.4. Molecules 149

Figure 3.11. Undiluted yellow food coloring, containing E102, E110, and E124, is red
(left), whereas diluting this substance with water produces a yellow color (right).

if the absorption spectrum were entirely located in the visible region. Some yel-
low food colorings show this dependency of color on concentration, as shown in
Figure 3.11. Dichroic dyes find application in the construction of certain LCD
display devices [537].
A related phenomenon is outlined in Figure 3.12, where all absorption takes
place in the visible spectrum. At low concentrations, the dent around 600 nm
causes the dye to look yellow. Increasing the concentration makes this dent rela-
tively unimportant, as yellow wavelengths are now mostly absorbed. As a result,
the lack of absorption in the violet range causes the material to appear violet.
Assume that a surface is painted with a dye where the chromogens are dis-
solved at a concentration c, and that the layer has a thickness of h. The painted
layer can then be modeled using Beer’s law (Section 2.9.4). Dyes can thus be
characterized by their thickness as well as the concentration of their chromogens.
As dyes do not have a particulate structure, more complex phenomena, such as
scattering, do not need to be taken into account. This is in contrast to pigmented
paints, which are discussed in Section 3.4.6.

3.4.5 Bleaching
The chromogens of organic dyes can be split by chemical reactions. The active
ingredients involved in these reactions are usually either chlorine or hydrogen

i i

i i
 i i

i i

150 3. Chemistry of Matter

Absorption
Dye concentration:
High
Low

350 400 450 500 550 600 650 700 750 800
Wavelength (nm)

Figure 3.12. A change in concentration yields a different perceived color.

peroxide—substances used to bleach fabrics, paper, and hair. In each case, the
reaction breaks the π component of a double bond, leaving only a single σ -bond,
or it may even break both bonds.
As a result, the reaction causes a hypsochromic shift towards the blue end of
the spectrum and may move the peak of the absorption spectrum into the ultravi-
olet region. As the peak of absorption is then outside the visible region, the result
of chemical bleaching is a loss of color.
3.4.6 Pigments
As pigments are insoluble particles, their index of refraction is usually different
from the material they are suspended in. As a result, pigments can be used both to
give color to a material and to make the material more opaque. To maximize the
hiding power of a pigment, the difference between the index of refraction of the
pigment and its carrier should be as large as possible. This then creates particles
that mostly reflect and scatter light, rather than transmit it. Pigments in white
paints, for instance, usually have a refractive index greater than 2.0.
A paint usually contains a carrier, such as linseed oil in the case of artists’ oil
paints, with pigment particles suspended in it. A full account of light interaction
with a layer of paint would involve scattering, which could be modeled as Mie
scattering. However, this approach would be too complicated [629], and a simpler
theory is required to model how light is absorbed and scattered in a pigmented
paint. In this case, Beer’s law is not appropriate either, as this empirical law does
not assume that scattering occurs in the medium.
Instead, Kubelka-Munk theory is normally applied to describe light behavior
in a layer of paint [628]. Light traveling through such a layer has at every point

i i

i i
 i i

i i

3.4. Molecules 151

Paint dx
layer L L x

Substrate Rg

Figure 3.13. The geometry associated with Kubelka-Munk theory.

a probability of being scattered. The simplifying assumption is made that light

scatters only into the direction from which it came. Thus, for light penetrating
the layer in the forward direction, a scattering event would cause some fraction
of this light to be reflected back towards the surface of the layer. If light traveling
towards the surface undergoes a further scattering event, it will be scattered in a
direction away from the surface. Lateral scattering is not accounted for.
The geometry associated with Kubelka-Munk theory is shown in Figure 3.13.
Assume that the layer of paint has a uniform thickness of x. With S the scattering
component and K the absorption, in a thin sublayer of thickness dx, the light
traveling downward is attenuated by an amount equal to [412]

(K + S) L↓ dx. (3.46)

The attenuation induced by the same sublayer for light traveling upward is
similarly
(K + S) L↑ dx. (3.47)

Since we assume that the light that is scattered is always scattered in the opposite
direction, the attenuation in each direction is modified by −S L dx. Hence, the
total change for light traveling in either direction due to a sublayer of thickness
dx is

dL↓ = (K + S) L↓ dx − S L↑ dx, (3.48a)

−dL↑ = (K + S) L↑ dx − S L↓ dx. (3.48b)

This pair of equations can be solved as follows [412]. First, we rearrange terms
and substitute a = 1 + K/S:

dL↓
= a L↓ − L↑ , (3.49a)
S dx
−dL↑
= a L↑ − L↓ . (3.49b)
S dx

i i

i i
 i i

i i

152 3. Chemistry of Matter

Summing these equations and rearranging leads to

 2
L↓ dL↑ − L↑ dL↓ L↑ L↑
= − 2a + 1. (3.50)
L↓2 S dx L↓ L↓

Using the quotient rule and making the substitution r = L↑ /L↓ , we get
dr
= r2 − 2ar + 1. (3.51)
S dx
Rearranging and integrating yields
 
dr
=S dx. (3.52)
r2 − 2ar + 1
For a thickness of 0, the value of r will be equal to the reflectance of the un-
derlying material, Rg . For thicknesses larger than 0, the reflectance of the paint
plus its substrate is R. Hence, the integration limits of the left-hand side of the
√ equation are Rg and R, allowing us to carry out the integration. Using
above
b = a2 − 1, we have
 R  R 
dr 1 dr 1 dr
= − R (3.53a)
Rg r2 − 2ar + 1 2b Rg r − (a + b) 2b Rg r − (a − b)
 
1 (R − a − b)(Rg − a + b)
= ln , (3.53b)
2b (R − a + b)(Rg − a − b)

and therefore from (3.52)

= St, (3.53c)
where t is the thickness of the layer of paint. For a layer of paint with a hypo-
thetical thickness of infinity, the hiding of the substrate is complete. In this case,
the value of Rg can be set to any convenient value, such as Rg = 0. A further
rearrangement gives
(R − a − b)(−a + b)
lim = (R − a + b)(−a − b), (3.54a)
t→∞ exp (2Stb)
0 = (R − a + b)(−a − b). (3.54b)
Hence, for complete hiding, the reflectance of a paint is given by
1
R= √ (3.55a)
a + a2 − 1
1
=   . (3.55b)
K K 2
1+ + 1+ −1
S S

i i

i i
 i i

i i

3.4. Molecules 153

This last equation is the Kubelka-Munk equation, which expresses the reflectance
of a layer of paint given absorption and scattering coefficients, K and S, respec-
tively. Equivalent forms of this equation are
 
K K 2 K
R = 1+ − +2 , (3.56a)
S S S
K (1 − R)2
= . (3.56b)
S 2R
These equations are valid for pigmented layers of paint where the pigment is
dispersed homogeneously throughout the carrier and the pigment particles them-
selves all have equal density.
It is assumed here that both absorption and scattering coefficients are wave-
length-dependent functions. Their absolute units are unimportant, as only their
ratio is ever used. This simplifies measuring K and S values substantially. Con-
sider, for instance, that the spectral reflectance function R(λ ) of a paint is given.
It is then possible to simply set S(λ ) = 1, and compute K(λ ) using the Kubelka-
Munk equation (3.56b).
A paint containing several different pigments with absorption and scattering
coefficients Ki (λ ) and Si (λ ) yields absorption and scattering coefficients KM and
SM for the mixture as follows:
n
KM (λ ) = ∑ Ki (λ ) ci , (3.57a)
i=1
n
SM (λ ) = ∑ Si (λ ) ci . (3.57b)
i=1

Here, ci is the concentration of the ith pigment. Once the K and S functions
for a given pigment are known, the K and S values for paints containing this
pigment can be derived on the basis of the Kubelka-Munk equation and the linear
superposition of the known and unknown pigments.
One of the areas where the Kubelka-Munk theory is useful is in predicting the
difference between organic and inorganic paint mixtures. A mixture of white with
an inorganic pigment tends to appear more gray than a mixture of white with an
organic pigment at comparable concentrations. Haase and Meyer show an exam-
ple where a mixture of titanium dioxide (white) with cadmium red (inorganic) is
compared with a mixture of titanium dioxide with naphthol red (organic) [412].
Both reds appear similar when applied directly. However, when mixed with tita-
nium dioxide, the cadmium red looks more gray.
The Kubelka-Munk theory has several applications, including color match-
ing. In this case, color matching refers to the process of taking a given color and

i i

i i
 i i

i i

154 3. Chemistry of Matter

mixing a paint to match that color. It also has uses in computer graphics. The
discussion of both applications is deferred until Section 8.13, because the appli-
cations require knowledge of color spaces and color difference metrics, which are
not explained until Chapter 8.

3.4.7 Water and Ice

Water and ice consist of H2 O molecules. Water molecules can vibrate in three
different ways, as shown in Figure 3.14. These vibrations are designated v1 , v2 ,
and v3 , and they are associated with wavelengths in the infrared range (2755 nm,
6199 nm, and 2638 nm, or 0.45 eV, 0.20 eV, and 0.47 eV). These vibrations can
have harmonic overtones, i.e., linear combinations are possible, so that radiation
due to vibration is also possible at energies associated with, for example, v1 + v2 ,
2v2 + v3 , etc.
It is rare that any of these overtones reach the visible spectrum in isolated
water molecules. However, in both water and ice, H2 O molecules form hydro-
gen bonds between molecules causing stronger bonding and associated absorp-
tion nearer the visible spectrum. This typically leads to some absorption in the
red part of the visible spectrum [218]. As a result, clear water has a pale blue
color [1056], which can be observed, for instance, against a sandy sea floor as
shown in Figure 3.15.
It should be noted that the blue color of water is typically not due to the re-
flection of the blue sky, except perhaps at grazing angles. A polarizer can be used
to eliminate reflections, revealing that water is blue by itself.
Water and ice frequently contain contaminations, such as algae, which may
color the water blue-green to green. Water in glacier-fed lakes, for instance, con-
tains silt giving it its characteristic turquoise color, which is due to scattering
events (Figure 3.16) [1, 716]. As water is rarely ever pure, there is great vari-
ability of its color [108, 110, 905, 1055]. For uncontaminated water, absorption
spectra have been measured [917, 1068] and are shown in Figure 3.17.
When water stains other materials, such as paper, fabric or rock, the effect is
almost always that the material appears darker [39, 1156]. For dry porous materi-

v1 v2 v3
H 2O

Symmetrical stretch Symmetrical bend Anti-symmetrical bend

Figure 3.14. The three normal modes of vibration of water molecules (after [815].)

i i

i i
 i i

i i

3.4. Molecules 155

Figure 3.15. The pale blue color of water can be observed against sand, as seen here in
this aerial shot of a coast line; Grand Bahama, 2004.

Figure 3.16. The silt in this glacier-fed lake colors the water turquoise; Alberta, Canada,
July 2001. (Image courtesy of Greg Ward.)

i i

i i
 i i

i i

156 3. Chemistry of Matter

1.5

Absorption coefficient (m -1)

1.0

0.5

0.0
350 400 450 500 550 600 650 700 750
Wavelength (nm)

Figure 3.17. Absorption spectrum for clear water [917].

als, the index of refraction is high, for instance 2.0. The amount of light reflected
depends on the ratio of the two indices of refraction. Going from air to dry porous
material in our example, this ratio would be 1.0 / 2.0. Wet material on the other
hand is coated with a thin layer of water with an index of refraction of around
1.33. The above ratio is then reduced to 1.0/1.33, and therefore less light is re-
flected (Figure 3.18) [716].
For non-porous materials, water forms a thicker layer. Light penetrating this
layer is partially reflected by the underlying material. This reflected light under-

Figure 3.18. Water seeping down this rock changes the appearance of the material.

i i

i i
 i i

i i

3.4. Molecules 157

goes further reflections and refractions at the water-to-air boundary. Some total
internal reflection occurs, so that less light emerges [668]. A computational model
to synthesize imagery of wet materials was described by Jensen et al [548].

3.4.8 Glass
Glass is formed by melting and cooling. The cooling occurs sufficiently rapidly
so that crystallization does not occur. The material is said to be in a glassy or
vitreous state. Different types of glass are used for different applications. For
instance, mercury vapor lamp jackets are made of fused silica (SiO2 ). Ordinary
crown glass (soda-flint-lime) contains SiO2 as well as Na2 O and CaO and is used
for window panes and bottles. Ovenware is made from borosilicate crown and
contains SiO2 , B2 O3 , Na2 O, and Al2 O3 . Finally, crystal tableware is made out of
flint, containing SiO2 , PbO, and K2 O.
Typical crown glass has a green tinge, as can be observed by looking at a
window pane from the side, or, as shown in Figure 3.19, by observing the side
of a block of glass. This colorization is due to the use of iron in the manufac-
ture of glass. It is possible to minimize the amount of green ferrous iron (Fe2+ )
by means of chemical decolorizing or physical decolorizing. Adding some man-
ganese oxide (MnO2 ), also known as glassmaker’s soap, achieves both. First,
physical decolorizing occurs because manganese oxide has a purple color, offset-

Figure 3.19. The top of this block of glass shows a green tinge.

i i

i i
 i i

i i

158 3. Chemistry of Matter

ting the green of the ferrous iron. Second, chemical decolorizing is achieved by
the following reaction:

Mn4+ + 2Fe2+ → Mn2+ + 2Fe3+ . (3.58)

The resulting ferric iron (Fe3+ ) is colorless.

Color may be added to glass by adding small quantities of other materials.
The mechanisms by which color is formed include ligand field effects, such as
occurs when adding transition metals to the molten glass. Many allochromatic
colors can be created by adding a variety of irons, including Ti3+ , Mn2+ , Mn3+ ,
Fe2+ , Fe3+ , and Co.
By adding iron sulfates, yellow or brown charge-transfer colors may be cre-
ated, as is the case in the well-known beer-bottle brown. Adding gold or silver to
the molten glass will produce metal scattering. Gold produces a deep red or pur-
ple color (purple of Cassius), whereas silver creates a yellow color. The addition
of copper also produces red, albeit not as intense or deep as gold. Finally, adding
semiconductors to the glass, such as cadmium sulfide-selenide (CdS · CdSe) can
produce yellow, orange, red, and black [815].
The archetypical use of colored glass is in the stained-glass windows in
churches and cathedrals, as shown in Figure 3.20.

Figure 3.20. Colored glass in stained-glass windows; Rennes, France, June 2005.

i i

i i
 i i

i i

3.5. Sources of Radiation 159

3.5 Sources of Radiation

Atoms and molecules consist of nuclei with clouds of electrons circling around
the nuclei. By adding thermal energy (heat), molecules and atoms may be set into
vibration. Such vibrations and oscillations may cause energy to be released. It can
be shown that all atoms and molecules radiate energy, unless their temperature
is reduced to absolute zero (0K). Thus, the amount of radiation is dependent
on temperature, and for certain temperatures, the radiation may be in the visible
spectrum.
Thermal radiation is energy emitted by all objects as a result of motion by the
objects’ atoms. A body may absorb energy and radiate energy. If its temperature
remains constant, the amount of absorption and radiation is the same, and the
body is said to be at equilibrium. Absorption and emission may be modeled with
a wavelength-dependent absorption coefficient αλ and a wavelength-dependent
emission σλ .
The absorption coefficient is the unitless fraction of incident radiant energy
that is absorbed. The emission coefficient is the energy per unit time per unit area
that is emitted by a body at each given wavelength. Both coefficients depend on
the nature of the body, as well as wavelength.
If we were to heat a chamber, then its interior walls would absorb and emit
light. At each wavelength independently, the absorption and emission would be
equal if the temperature were to remain constant. This is formalized in Kirchhoff’s
Radiation Law:
αλ
= Meλ , (3.59)
σλ
where Meλ is known as the spectral radiant exitance,7 and is measured in W m−2 .
This function depends on temperature T and wavelength λ , but not on the ma-
terial of the cavity walls, nor on the cavity’s spatial configuration. This fact was
empirically observed as early as 1792 by ceramist Thomas Wedgwood who noted
that a kiln fired to a certain temperature would glow the same color as the pottery
inside [447].

3.5.1 Blackbody Radiation

If an object is a perfect absorber, i.e., all incident radiation is absorbed, then the
fraction αλ equals 1 for all wavelengths λ . Such an object appears black and is

7 This is a radiometric term and will be explained along with other radiometric terms in Chapter 6.

i i

i i
 i i

i i

160 3. Chemistry of Matter

Incident light

Stack of razorblades

Figure 3.21. The geometry of a stack of razor blades is such that it can absorb most light,
and thus simulate a blackbody.

called a blackbody radiator. Since no materials are perfect absorbers, it is not

possible to directly observe a perfectly black material.
However, it is possible to construct a cavity, as mentioned above, and create a
small hole in one of the walls. Nearly all of the energy entering the cavity from
outside is then absorbed after one or more reflections inside the cavity. Only a
tiny fraction of this light is reflected towards the hole and subsequently leaves the
cavity. The hole, therefore, appears nearly black and could be regarded as a close
approximation to a blackbody radiator.
As an aside, an alternative way to construct an object resembling a blackbody
radiator is by stacking a number of old-fashioned razor-blades. Although each
blade is a good reflector (αλ  1), the spatial configuration is such that light is
funneled into the grooves formed by the stack of blades and is, therefore, not able
to escape, as shown in Figure 3.21.
A stack of razor blades is shown in Figure 3.22, with the light source around
20 cm above the grooves [1247].8 The grooves absorb most of the light. At dif-
ferent angles, the grooves do not behave as a blackbody, as shown in Figure 3.23.
Returning to a cavity acting as a blackbody radiator, the walls absorb light
entering the cavity through a hole. If this cavity is heated, the atoms of the walls
vibrate more vigorously. As a result they each radiate more energy. The interac-
tion of atoms with neighboring atoms causes each to be capable of radiating over
a broad spectrum, rather than the narrow bandwidth spectra associated with the
excited states of individual electrons. Nonetheless, the broad spectrum generated
by a heated cavity shows a peak at a wavelength determined by the temperature
of the cavity. Wien’s displacement law provides a formula relating the peak wave-

8 This demonstration was given by Alain Fournier at his keynote speech for the 6th Eurographics

Workshop on Rendering, held in Dublin in 1995.

i i

i i
 i i

i i

3.5. Sources of Radiation 161

Figure 3.22. A stack of 30 razor blades approximating a blackbody radiator.

Figure 3.23. If the light is incident at a shallow angle, and the viewpoint is also chosen at
a shallow angle, the grooves may reflect rather than absorb.

i i

i i
 i i

i i

162 3. Chemistry of Matter

Figure 3.24. The stack of razor blades from Figure 3.22 is heated with a butane fueled
torch.

length λmax to the temperature of the cavity:

hc
λmax T = (3.60a)
4.965114kB
= 2.897791 × 10−3 , (3.60b)

where h is Planck’s constant, c is the velocity of light, and kB is Boltzmann’s

constant (see Appendix D). The right-hand side is given in (mK). This equation
shows that the peak wavelength decreases with increasing temperature, a fact that
can be empirically verified. For instance, iron begins to radiate dull red at around
700 ◦ C. At melting point (1500 ◦ C), the iron glows yellow. The temperature of
the sun, which is normally perceived as white, is around 5700 ◦ C. Even hotter
stars appear bluish-white [815]. An example of thermal radiation is shown in
Figure 3.24.
For a blackbody radiator, Planck’s radiation law provides a functional form
for the spectral density at each wavelength as function of the temperature [1262]:
   −1
8π hc hc
μeλ = exp − 1 . (3.61)
λ5 λ kB T

i i

i i
 i i

i i

3.5. Sources of Radiation 163

Iλ
Wienʼs displacement law
Planckʼs radiation law:
T = 9000 K
T = 8000 K
T = 7000 K
T = 6000 K
T = 5000 K
T = 4000 K

0 400 800 1200 1600 2000

λ (in nm)

Figure 3.25. The emission spectrum for blackbody radiators heated to different
temperatures.

The spectral radiant exitance Meλ as function of wavelength and temperature is:
c μeλ
Meλ = (3.62a)
4
   −1
2π hc2 hc
= exp − 1 . (3.62b)
λ5 λ kB T

This function is plotted for various values of T in Figure 3.25. The same figure
also shows in red the curve determined by Wien’s displacement law. Note that the
range of wavelengths shown is much larger than the visible range. Within the vis-
ible range, each of the curves is relatively flat, with the exception of very high and
very low temperatures. This means that for intermediate temperatures, blackbod-
ies radiate a broad spectrum, thereby appearing predominantly white with perhaps
a tinge of blue for higher temperatures and red for lower temperatures.
Planck’s radiation law for blackbody radiators gives a spectral distribution of
energy with its peak located at λmax . Wien’s displacement law predicts that this
peak shifts if the temperature is increased. Heating an object not only increases
the amount of light that is emitted, but also changes the apparent color. It is
therefore natural to refer to a given spectral distribution of a blackbody radiator
by its color temperature, which is given in degrees Kelvin.
The area under each of the curves in Figure 3.25 has special significance and
represents the total radiant exitance of the blackbody radiator. The radiant exi-
tance Me can be computed by integrating (3.62) over all wavelengths:
 ∞
Me = Meλ d λ = σ T 4 . (3.63)
0

i i

i i
 i i

i i

164 3. Chemistry of Matter

Here, σ is a constant of proportionality and is known as the Stefan-Boltzmann

constant :
2π 5 k 4
σ= = 5.67032 × 10−8 . (3.64)
15h3 c2
This constant is given in (W m−2 K−4 ). Rather than integrating over all wave-
lengths, the radiant exitance may be approximated by integrating over all wave-
lengths up to λmax :
 λmax
Meλ d λ ≈ 0.25σ T 4 . (3.65)
0
An example of a blackbody radiator is the sun, which radiates at approximately
5700 K, yielding a peak wavelength near 500 nm. Stars radiate between 5000
K and 11,000 K. Incandescent lamps have a temperature of around 3000 K. Hu-
man skin also radiates, but only weakly and at a temperature outside the visible
spectrum (around 10,000 nm) [1247].

3.5.2 Incandescence
Objects that are heated radiate energy. Some of the energy is radiated in the vis-
ible spectrum, dependent on the object’s temperature. Blackbodies are a special
case of radiating bodies, in that they absorb all incident light. In the more gen-
eral case, any radiation occurring as a result of heating objects is referred to as
incandescence.
Candles, fires, tungsten filaments, hot coal, and iron are all examples of in-
candescent light sources. The heat of a candle flame, for instance, melts the wax
which subsequently flows upward in the wick by means of capillary action and is
then vaporized. Such a flame emits yellow light near the top and blue light near
the bottom.
The vaporized wax is the fuel that can react with oxygen. The center of the
flame contains a limited amount of oxygen, which diffuses in from outside the
flame. The reaction in the center occurs at around 800 ◦ C to 1000 ◦ C [815]:

C17 H35 CO2 H + 11O2 → 9H2 0 + 5CO2 + 5CO + 8C + 9H2 . (3.66)

Near the outside of the flame where more oxygen is available, further reactions
occur at the higher temperature of between 1000 ◦ C to 1200 ◦ C, yielding yellow
light:

2CO + O2 → 2CO2 , (3.67a)

2C + 2O2 → 2CO2 , (3.67b)
2H2 + O2 → 2H2 0. (3.67c)

i i

i i
 i i

i i

3.5. Sources of Radiation 165

Figure 3.26. The flame of a lighter. The lower part of the flame contains a pre-mixed
combination of air and fuel, causing total combustion which leads to a blue flame. The
fuel in the top part of the flame is only partially combusted, leading to soot as well as a
yellow flame.

Thus, the above reactions produce energy in the form of light, as well as water and
carbon dioxide. A candle flame is classified as a diffusion flame, because oxygen
diffuses inward before it reacts with the fuel.
Near the bottom of the flame, a different situation arises, since before ignition
the fuel is mixed with oxygen. Such premixed flames appear blue and burn cleanly
since no soot particles are formed in the process.
Both premixed and diffusion flames occur in candle light, but they are gener-
ally easier to observe in the flame of a cigarette lighter, as shown in Figure 3.26.
The lower temperatures in the center of a flame due to a lack of oxygen in the
flame’s interior can be demonstrated by lowering a wire mesh into the flame. The
flame will burn underneath the mesh, but not above it. The mesh therefore allows
us to observe a cross-section of the flame, as shown in Figure 3.27. The dark
center is easily observed through the mesh.

i i

i i
 i i

i i

166 3. Chemistry of Matter

Figure 3.27. The emission of light from an open flame occurs near the surface of the
flame. The interior lacks sufficient oxygen to burn. That a flame is hollow is demonstrated
here by lowering a mesh into the flame of a candle, revealing a dark center. The flame
emits light underneath the mesh, but not above.

Incandescence is also the mechanism employed in tungsten filaments, as

shown in Figure 3.28. Here, an electric current is sent through a coil. The
coil’s resistance causes electric energy to be converted into heat, which causes
the tungsten to emit light. The temperature of filament light bulbs is around
2500 ◦ C. Incandescence is also found in old-fashioned vacuum tubes, as shown in
Figure 3.29.

Figure 3.28. An incandescent tungsten filament.

i i

i i
 i i

i i

3.5. Sources of Radiation 167

Figure 3.29. Incandescence in a vacuum tube.

Other forms of incandescent light are found in pyrotechnics, for instance when
magnesium powder is burned. This reaction produces very high temperatures,
leading to very bright white light, which may be colored by adding nitrates, chlo-
rates, or perchlorates. This reaction is also used in flares and tracer bullets. A final
example of incandescence is the photographic flashlight. The bulb of a flashlight
contains shredded zirconium metal and oxygen. On ignition, this produces molten
zirconium and zirconium oxide which radiates at a color temperature of around
4000 ◦ C [815].
Incandescent light is characterized by a lack of polarization. In addition, the
phase of each emitted electro-magnetic wave is random. Light is generally emit-
ted equally in all directions. The temperature of an incandescent light source de-
termines its color, similar to blackbody radiation. With temperature, the radiation
progresses from black to red, orange, yellow, white, and finally blue-white.

i i

i i
 i i

i i

168 3. Chemistry of Matter

Sodium

Neon

Xenon

400 440 500 550 600 650 700

Figure 3.30. Line spectra of sodium, neon, and xenon. Note that the re-
produced colors are approximate. (These spectra were generated using an online
applet provided by John Talbot, available from http://www.physik.rwth-aachen.de/ ˜
harm/aixphysik/atom/discharge/index1.html.)

3.5.3 Gas Excitation

Radiant energy may be produced by sending an electric current through a tube
filled with inert gas. This may be accomplished by two electrodes at either end of
the tube providing an electric field. Electrons accelerate through the tube and col-
lide with the gas molecules. The kinetic energy of the electrons is then absorbed
by the molecules, yielding electrically excited molecules, atoms, and ions, as well
as electrons. Of these, the charged particles (ions and electrons) are subjected to
the electric field and accelerate themselves, producing further collisions.
When excited molecules and atoms return to their original state, photons are
emitted. The number of allowable electric states is limited, and therefore line
spectra are produced, as shown in Figure 3.30. Such spectra are characterized by
narrow emission lines, superposed with a continuous broad spectrum. Joseph von
Fraunhofer mapped the wavelengths of a large number of spectral lines, which
were found to be associated with specific elements. The lines were assigned let-
ters and are collectively called Fraunhofer lines. A collection of these are listed in
Table 3.8. Fraunhofer lines are often used to characterize the dispersive properties
of dielectric materials (see for instance Section 12.3.6).

i i

i i
 i i

i i

3.5. Sources of Radiation 169

Designation Element Wavelength Designation Element Wavelength

T Fe 302.108 F H 486.134
P Ti+ 336.112 D2 Na 588.995
N Fe 358.121 D1 Na 589.592
L Fe 382.044 C H 656.281
K Ca+ 393.368 B O2 686.719
H Ca+ 396.847 A O2 759.370
G Ca 430.774 Z O2 822.696
G Fe 430.790

Table 3.8. A selection of Fraunhofer lines.

The spectrum emitted by a gas discharge lamp is primarily determined by the

inert gas used. Only inert gases are chosen since, otherwise, the gas may react
with the electrodes and break the lamp. Inert gases include neon, argon, helium,
krypton, and xenon (see Table 3.6).
Neon, as used in neon signs, gives a red color, as does neon with some argon
added. Argon by itself, or argon with added mercury, produces blue light. Yel-
low and green light may be created by using colored glass tubes. Less frequently

Ionized Na* (5.12 eV)

Energy (eV)

Wavelength (nm)

5 250

4D3/2 4F 5/2 300

4 5S 1/2
4P 1/2 4P 3/2
3D3/2
3 4S 1/2 400
Visible range

500

2 3P 1/2 3P 3/2 600

700
800
eV

V
5e

1000
2.03
2.0

0
3S 1/2

Figure 3.31. Energy levels associated with sodium, Na (after [815]).

i i

i i
 i i

i i

170 3. Chemistry of Matter

used are helium, krypton, and xenon because they are either not producing suf-
ficient intensity, or they are too costly to produce. These gases produce yellow,
pale lavender, and blue, respectively. Further colors may be produced by adding
phosphor powders or by using phosphorescent glass tubes.
Sodium vapor lamps operate on a similar principle, although sodium is a solid
metal at room temperature. To produce sodium vapor, neon is added which pro-
duces the initial discharge, as well as heat.9 The heat then vaporizes the sodium
(Na) which is then ionized due to collisions with electrons:

Na → Na+ + e− . (3.68)

Figure 3.32. Sodium street lights emit near-monochromatic light, making it difficult to
distinguish object reflectance; Zaragoza, Spain, November 2006.

9 A sodium vapor lamp emits pink light for the first couple of minutes after being switched on as a

result of the neon gas discharge.

i i

i i
 i i

i i

3.5. Sources of Radiation 171

The ion may later recombine with an electron and thereby return to a lower excited
energy level, before returning to a yet lower level and emitting a photon:

Na+ + e− → Na∗ , (3.69a)

∗
Na → Na + photon. (3.69b)

There are several different energy states the sodium may reach, as shown in Fig-
ure 3.31. Many of the energy states are outside the visible range and, therefore,
produce radiation that is not detected by the human eye. Around half of the light
emitted is due to a doublet of states at 2.03 eV and 2.05 eV, which both correspond
to yellow light (λ ≈ 589.6 and 589.0 nm). The strong yellow color makes sodium
vapor lamps less suitable for a range of applications, as it is difficult to determine
the color of objects under sodium lighting. Sodium lamps are predominantly used
in street lighting (Figure 3.32 and Section 9.3).
Mercury vapor lamps produce a bluish overall spectrum by a mechanism sim-
ilar to sodium vapor lamps. Their principle emission lines in the visible region
occur at around 405, 436, 546, and 578 nm. The color of mercury vapor lamps
can be improved by coating the glass envelope with a phosphor. Such phosphor-
coated mercury vapor lamps are in wide use today.

3.5.4 Fluorescence
In most cases, when a photon is absorbed by an atom, an electron jumps to a
higher state and immediately returns to its ground state, whereby a new photon is
emitted of the same wavelength as the photon that was absorbed. Some materials,
however, allow electrons to return to their ground state via one or more intermedi-
ate energy states. This creates state transitions during which photons of different
wavelengths are emitted. The wavelength of the emitted photon is determined by
the difference in energy levels of the two states involved in the transition. This
means that new photons are emitted at longer wavelengths, a phenomenon known
as Stokes’ law. The process of absorption and re-emission at a different wave-
length is called fluorescence.
As indicated in the preceding section, the inside of a mercury vapor lamp may
be coated with a phosphor, creating a fluorescent light source. An example of a
fluorescent light source is shown in Figure 3.33. The mercury vapor emits light,
predominantly in the ultra-violet region (253.7 nm), which excites the phosphor.
The phosphor in turn re-radiates light at a longer wavelength. The color of a
fluorescent tube is determined by the type of phosphor used. For instance, mag-
nesium tungstate creates an emission spectrum with a maximum emittance at 480
nm. Maximum emittance at longer wavelengths is possible with, for instance,

i i

i i
 i i

i i

172 3. Chemistry of Matter

Figure 3.33. A fluorescent light source.

calcium halophosphate (590 nm), strontium magnesium phosphate (620 nm), or

magnesium fluoro-germanate (660 nm) [1262].
Relative Spectral Power

90
80 F2
70 F7
F11
60
50
40
30
20
10
0
350 400 450 500 550 600 650 700 750 800
Wavelength (nm)

Figure 3.34. Relative spectral power distributions for CIE F2 , F7 and F11 fluorescent light
sources (data from [509]).

i i

i i
 i i

i i

3.5. Sources of Radiation 173

Figure 3.35. This set-up demonstrates frequency up-conversion. The small transparent
panel is coated with phosphors that emit at blue wavelengths when subjected to near-
infrared light. (Demonstration kindly provided by Janet Milliez, School of Optics, Uni-
versity of Central Florida.)

Fluorescent lamps can approximate daylight, although they are found to be

unsuitable for critical color-matching tasks, as their spectral power distribution
(shown for typical fluorescent lights in Figure 3.34) deviates too much from aver-
age daylight.
It is possible to construct phosphors that fluoresce in the visible range if ex-
cited by near-infrared electromagnetic radiation [866, 896]. In this case, the re-
emitted light is of a shorter wavelength than the incident light. This is effec-
tively the opposite of fluorescence, where re-emission takes place at longer wave-
lengths. The frequency of light being the reciprocal of its wavelength, this process
is termed frequency up-conversion. Recent advances in up-conversion materials
show promise for the construction of emissive display devices with a remark-
able brightness, paired with a wide color gamut which includes very saturated
colors [936, 937].
A demonstration of a small, yet very bright, vector display device is shown in
Figure 3.35. Here, near-infrared light is directed in a lissajous pattern toward a
phosphor-coated piece of plastic which is designed to emit blue light. Phosphors
emitting at red and green wavelengths are also possible, as shown in Figure 3.36.
The blue phosphor in this figure has a secondary emission at red wavelengths
and, therefore, appears somewhat purple. Nonetheless, this is a potentially suit-
able technology for constructing versatile display devices with low weight and
depth [936, 937].

i i

i i
 i i

i i

174 3. Chemistry of Matter

Figure 3.36. Red, green, and blue phosphors shown in bright and dim variants. The
substrate was opaque. (Demonstration kindly provided by Janet Milliez, School of Optics,
University of Central Florida.)

3.5.5 Phosphorescence
For some substances, electrons may be excited into a higher energy state, and
they may alternate between two meta-stable states before returning to the ground
state. These alternate states may be reached by further absorption of radiant en-
ergy. If the eventual return to the ground state emits a photon, it will be at a
later time than the initial absorption. Materials exhibiting this behavior are called
phosphorescent.
If the source of radiation that caused the transition between states is switched
off, there will be a delay before all atoms in the substance have returned to their
ground states. The decay time can be measured by noting the time necessary
before the emitted radiant power is reduced to a fraction of e−1 of its original
value.
Fluorescent materials generally have a very short delay time, on the order of
10−8 seconds. Phosphorescent materials have decay times that range from 10−3
seconds to many days [1262]. The decrease in energy is typically exponential.
However, when an electron absorbs enough energy to ionize the atom, the electron
will follow a power-law decay. The radiance L emitted by a phosphor as function
of time t may be modeled with an empirical formula [675]:

1
L=  2 , (3.70)
1
b √ +t
Lo b

i i

i i
 i i

i i

3.5. Sources of Radiation 175

Figure 3.37. The hands of this pocket watch are painted with a phosphorescent paint.
Before this photograph was taken, the watch was lit for around a minute by a D65 daylight
simulator. The photograph was taken in low light conditions to show the phosphorescent
behavior of the paint.

where L0 is the radiance emitted at time t = 0, and b is a material dependent

constant.
Examples of phosphorescent materials are the paint used to paint the hands
and dials of watches, as shown in Figure 3.37, and the light emitted by cathode
ray tubes (Figure 3.38).

Figure 3.38. Red, green, and blue dots of a phosphorescent cathode ray tube. At a dis-
tance, the human visual system fuses the separate dots and the result is perceived as a
uniform white.

i i

i i
 i i

i i

176 3. Chemistry of Matter

3.5.6 Luminescence
Fluorescence and phosphorescence discussed in Sections 3.5.4 and 3.5.5 are two
examples of a general principle, called luminescence. The term luminescence
refers to all production of light that is not caused by thermal radiation. This leaves
many forms of luminescence [815]:

Photoluminescence is luminescence induced by visible light or ultraviolet light.

Examples include day-glo paint, which contains organic dyes that fluoresce
causing an appearance brighter than white, and fluorescence brighteners
used in laundry detergent.

Radioluminescence in certain substances is caused by radiation of energetic par-

ticles such as alpha, beta, or gamma rays. For example, radioactive mate-
rials can be mixed with paint to create a self-luminous substance, used, for
instance, in dials of watches.

Electroluminescence occurs due to the application of an electric field or current.

Examples include gas discharge lamps as discussed in Section 3.5.3, as well
as the plasma technology outlined in Section 14.4.

Figure 3.39. Glow sticks emit light by chemiluminescence. The yellow, blue and green
glow sticks are enclosed in clear plastic holders, while the pink color occurs in part due to
the red plastic of the tube.

i i

i i
 i i

i i

3.5. Sources of Radiation 177

Triboluminescence is produced by mechanical force, such as grinding crystals. It

can be produced at home by crushing hard candy, in particular wintergreen
lifesavers.

Chemiluminescence is created by chemical energy, as for instance used in glow

sticks (Figure 3.39). Glow sticks can be frozen to halt the reaction that will
resume afterwards once the sticks reach room temperature.

Bioluminescence is a form of chemiluminescence produced by biological mecha-

nisms such as glow worms and fireflies. Also many (deep-)sea creatures ex-
hibit bioluminescence, including species of jellyfish and angler fish. Some
algae have a triboluminescent form, which makes them glow when dis-
turbed. This creates the “burning of the sea” effect.

Further distinctions between different forms of luminescence may be made [815],

which are beyond the scope of this book.

3.5.7 Light-Emitting Diodes

Semiconductors, as discussed in Section 3.3.3 can be doped with donors as well
as acceptors, and, in either case, this may change the color of the semiconductor.
In addition, donor impurities may increase the material’s conductivity at high
temperatures. Acceptors increase conductivity at all temperatures.
If a semiconductor such as silicon is doped with phosphorus, a donor, the
result is called n-type silicon because there exist negatively charged electrons in
the conduction band. Doping silicon with boron, an acceptor, results in positively
charged holes in the valence band, and such material is called p-type silicon. Plac-
ing two such materials side-by-side creates a p-n contact region (a p-n junction)
where all donors and acceptors are ionized. The Fermi level is now just above the
acceptor level in the p-region and just below the donor level in the n-region.
By applying a voltage to this configuration, the Fermi level will jump to a dif-
ferent level. Electrons in the conduction band will then be attracted to the holes in
the valence band, and their recombination will result in the emission of photons.
In silicon, the band-gap energy is at 1.1 eV, and this is sufficient to produce pho-
tons at infrared wavelengths. This process is called injection electroluminescence.
Other semiconductors exist with band gaps that correspond to visible light with
different colors, such as GaAs1−x Px (red), AlP (green) and ZnS (blue) [815].
Light sources based on this principle are called light-emitting diodes, or LEDs
for short [479], and are discussed further in Section 9.8. An example of a set of
red LEDs is shown in Figure 3.40.

i i

i i
 i i

i i

178 3. Chemistry of Matter

Figure 3.40. A set of red LEDs.

The principle used by LEDs can be reversed, such that light absorbed at a p-n
junction creates a voltage, a principle used by solar panels (Figure 3.41).

Figure 3.41. Solar panels use the same semiconductor technology as LEDs, albeit in
reverse by absorbing light and thus producing an electric current.

i i

i i
 i i

i i

3.5. Sources of Radiation 179

3.5.8 Organic Light-Emitting Diodes

Light-emitting devices can be constructed from semiconductor materials, as
discussed above. Recently, there has been considerable interest in light-emitting
devices built from organic molecules. Such devices are known as organic
light-emitting diodes (OLEDs) [1117]. This approach has the advantage of poten-
tially high efficiency and ease of manufacturing. In particular, there are develop-
ments towards modifying standard printing techniques such as screen
printing [96, 97, 663, 870] and ink jet printing [431] to produce OLED
displays [802].
OLEDs can be configured to emit light over the entire visible spectrum, for
instance by altering the driving voltage, or interference effects in multi-layer
OLEDs [385]. While high efficiency can currently be combined with a high
device stability for red and green emitters [79, 181, 852, 1148, 1224], stability,
efficiency, high color saturation, and a long lifetime are still problematic for blue
emitters [383, 1234].
Organic LED technology may be coupled with inorganic nanocrystals, for in-
stance made of cadmium-selenium (CdSe) [199]. These crystals afford narrower
emission spectra, i.e., more saturated colors, than conventional OLEDs which
typically emit light over a 50–100 nm range [1150].
OLEDs hold great promise for the future, in particular for the production of
efficient and bright display devices, as discussed in Section 14.6. An intermediate
solution is the use of hybrid TFT-OLED technology to overcome current device
limitations [1234].

3.5.9 Lasers
As we have seen in Section 3.5.4, an atom may absorb energy in the form of
a photon, causing an electron to jump to a higher energy level, or equivalently,
a different orbital. When the electron jumps back to its ground state, possibly
via one or more intermediate energy levels, new photons may be emitted, lead-
ing to fluorescence. If energy is added to the atom in the form of kinetic en-
ergy (heat), then electrons also jump to higher energy levels. Upon return to the
ground state, emission of photons in the visible range may also occur, leading to
incandescence.
We will assume that there is a population of N1 atoms at a given lower energy
level E1 , and a population of N2 atoms in a higher (excited) energy level E2 . In
a material with vast quantities of atoms, the relaxation to lower energy levels
occurs at an exponential rate, determined by the material itself. With t indicating
time, and γ2 the spontaneous decay rate, the size of the population of atoms at the

i i

i i
 i i

i i

180 3. Chemistry of Matter

higher energy level is given by [1045]:

N2 (t) = N2 (0) e−γ2 t (3.71a)

−t/τ2
= N2 (0) e . (3.71b)

The reciprocal of γ2 is given by τ2 = 1/γ2 and signifies the lifetime of the upper
level E2 . The above decay rate is independent of which lower energy levels are
reached. The decay rate from level E2 to level E1 can be modeled with a similar
equation, except that γ2 is replaced by a different constant γ21 . For convenience,
we may write this equation as a differential equation, which models the sponta-
neous emission rate:
dN2 (t)
= −γ21 N2 (t). (3.72)
dt
So far we have not discussed how atoms end up in a higher energy state,
other than that photons at the right wavelengths may be caught. Normally, this
happens irregularly dependent on the temperature of the material and the amount
of incident light. However, it is possible to force atoms into higher energy levels,
for instance, by flashing the material with a short pulse of light that is tuned to the
transition frequency of interest. This process is called pumping.
After such a short pulse is applied, the population of atoms in the higher en-
ergy state can be modeled by the following atomic rate equation:

dN2 (t)
= K n(t) N1 (t). (3.73)
dt
Here, n(t) is the photon density of the applied signal measured in number of
photons per unit volume, or equivalently as the electromagnetic energy density
divided by h̄ ω . This is the process of stimulated absorption, i.e., the application
of a short pulse of light forces atoms that are initially in the lower energy state E1
into the higher energy state E2 .
Similarly, the same pulse forces atoms that are initially in the higher energy
state E2 into the lower energy state E1 , thereby emitting photons. This process is
called stimulated emission and is modeled by the following atomic rate equation:

dN2 (t)
= −K n(t) N2 (t). (3.74)
dt
Note that the constant of proportionality K in the above two equation is identical,
but of opposite sign. This constant will be largest if the pulse of light that is
applied is of the same energy (frequency/wavelength) as the transition energy
between levels E1 and E2 of the material.

i i

i i
 i i

i i

3.5. Sources of Radiation 181

The total rate equation when pumping is applied to a volume of material is

then the superposition of the above three equations:

dN2 (t)
= K n(t) (N1 (t) − N2 (t)) − γ21 N2 (t). (3.75)
dt
Stimulated emission produces photons with energies that are able to force
stimulated absorption or emission in other atoms. As a result, stimulated emission
causes amplification of the initial pulse of light (hence the acronym laser, which
stands for light amplification by stimulated emission of radiation). Whether a
population of atoms, i.e., a volume of material, amplifies light or absorbs light
depends on the ratio of atoms in the lower and higher energy states, i.e., N2 /N1 .
In essence, if we can force a majority of the population of atoms into the higher
energy state, then the volume of material will act as an amplifier; otherwise, the
material will absorb energy.
This can also be seen from (3.75), where the difference between N1 and N2
is on the right-hand side of the equation. If this difference is positive, then the
material absorbs energy. If this difference is negative, i.e., when the material is in
a condition known as a population inversion, then more state transitions are from
E2 to E1 , leading to emission of additional photons which cause further stimulated
emissions. Under this condition, the initial pulse of light is amplified.
For a material that is in thermal equilibrium at temperature T , the ratio N2 /N1
adheres to a fundamental law of thermodynamics, known as Boltzmann’s
Principle:  
N2 E2 − E 1
= exp − . (3.76)
N1 kT
This ratio can also be expressed as a difference:
  
h̄ ω
N1 − N2 = N1 1 − exp − . (3.77)
kT

This equation shows that for a material in thermal equilibrium, the majority of the
population is in the lower energy level, and therefore light is absorbed. Thus, to
create laser light, a pumping mechanism may be used to move the majority of the
population into a non-equilibrium state.
When such a pumping process is applied, each atom may react to it by chang-
ing energy level. Since each atom responds in the same manner, the resulting
amplification is coherent in both phase and amplitude. Thus, the output of a laser
is an amplified reproduction of the input (pumping) signal—at least for a narrow
band around the wavelength associated with the transition energy E2 − E1 . How-
ever, some noise will be introduced due to spontaneous emission and absorption.

i i

i i
 i i

i i

182 3. Chemistry of Matter

Figure 3.42. Tunable laser photographed at various wavelengths. (Photographs created

with the expert help of Claudiu Cirloganu, School of Optics, University of Central Florida.)

The components of a practical laser include:

A laser medium, which is the collection of atoms, molecules, or ions that will
mediate amplification.

A pumping process,which will excite the laser medium into a higher energy level,
causing a population inversion.

Optical elements, which will mirror a beam of light to pass through the laser
medium one or more times. If the beam of light passes through once, we
speak of a laser amplifier, otherwise of a laser oscillator.

Lasers emit light at a narrow band of wavelengths, and therefore produce near
monochromatic light. In addition, as outlined above, this light is typically coher-
ent, i.e., in phase and amplitude, making laser light very different from most light
sources normally encountered. Most lasers are set to one specific wavelength, al-
though high-end lasers can be made to be tunable. For such lasers, it is possible to
dial in the wavelength at which they will emit monochromatic light (Figure 3.42).

3.6 Polarization in Dielectric Materials

In dielectric materials, electrons are bound to the atomic structure, i.e., conduction
of free electrons such as that which occurs in metals is absent. In the presence of
an electromagnetic field, electrons remain bound, but may be displaced. In such
cases, the electrons may be thought of as orbiting around a point that is some

i i

i i
 i i

i i

3.6. Polarization in Dielectric Materials 183

distance away from the nucleus. As a result, the negative charge of the electrons
is slightly offset with respect to the positive charge of the nucleus. An atom under
the influence of an electromagnetic field may thus form an electric dipole. The
atom is said to be polarized under these conditions.
Some dielectrics are known as polar substances, where the molecular struc-
ture is such that dipoles exist even in the absence of an external electromagnetic
field. This is for instance the case with water. However, for polar substances,
the orientation of each molecule is random, and therefore the net polarization of
the material in bulk is zero. When an electric field is applied, the forces acting
upon the molecules will align them with the field, and the material will become
polarized. This phenomenon is called orientational polarization.
Some materials consist of ions that are bound together. As an example,
sodium chloride NaCl consists of positive and negative ions. Application of an
external electric field will separate the ions and thus form electric dipoles. The
result is called ionic polarization.
A polarized atom may be modeled as a positive point charge +Q and a neg-
ative point charge −Q separated by a distance r. The dipole moment p is then
defined as
p = Q r. (3.78)
It would be difficult to model the interaction of an electromagnetic wave with a
material if the dipole moments of each individual atom or molecule need to be
considered. It is more convenient to compute the average behavior of the material
at the microscopic scale. The electric polarization P of a material is then given
by
P = N Q ra = N p, (3.79)
where N is the number of atoms per unit volume and ra is the average separa-
tion between positive and negative charges in atoms. It is now possible to model
the electric displacement D as an additive quantity rather than a multiplicative
quantity as shown in (2.5):
D = E + 4 π P. (3.80)
For suitably weak fields, the electric polarization can be taken to be proportional
to the magnitude of the electric field, i.e., P ∝ E. The constant of proportionality
is then called the dielectric susceptibility η :
P = η E. (3.81)
In this case, by comparing (2.5) and (3.81), we find a relation between the dielec-
tric susceptibility and the material constant ε :
ε = 1 + 4 π η. (3.82)

i i

i i
 i i

i i

184 3. Chemistry of Matter

3.6.1 The Lorentz-Lorenz Formula

Since an atom consists of a heavy nucleus surrounded by electrons orbiting at
specific distances, we can expect that the impact of an electromagnetic field on a
single atom yields somewhat different mathematics than the impact on a volume
of atoms or molecules. Thus, the effective field, E and H , is acting on an indi-
vidual atom, whereas the mean or observed field, E and H, is averaged over a
large number of atoms. It can be shown that the effective and mean fields can be
expressed in terms of each other by solving a Poisson equation over a uniformly
charged sphere, yielding [112]
4π
E = E + P. (3.83)
3
We assume that the electric dipole moment of an atom or molecule is proportional
to the strength of the electric field, and hence

p = α E . (3.84)

The constant α is called the mean polarizability. With N atoms per unit volume,
as given in (3.79), the electric polarization can be rewritten as

P = N p = N α E . (3.85)

Combining this with (3.83), we find

Nα
P= E. (3.86)
4π
1− Nα
3
We therefore have the following expression for the dielectric susceptibility:
Nα
η= . (3.87)
4π
1− Nα
3
Substitution of this result into (3.82) yields an expression for the permittivity ε :
8π
1+ Nα
ε= 3 . (3.88)
4π
1− Nα
3
√
With n = ε , the mean polarizability can be expressed in terms of the index of
refraction as follows:
3 n2 − 1
α= . (3.89)
4 π N n2 + 2

i i

i i
 i i

i i

3.6. Polarization in Dielectric Materials 185

In the context of refractivity, this relation is called the Lorentz-Lorenz formula,

and is named after two scientists who discovered this relation independently and
almost at the same time. When written in the following form,
ε −1
4π α N = 3 , (3.90)
ε +2
and used in the context of dielectric materials, this equation is also known as the
Clausius-Mossotti equation [422,829]. Finally, in the context of conductivity, this
equation is known as Maxwell’s formula [739]. The importance of these formulae
lies in the fact that they relate Maxwell’s equations with the atomic theory of
matter.

3.6.2 Dispersion
Newton’s famous experiment, in which light is refracted through a prism to reveal
a set of rainbow colors, cannot be explained with the machinery described so far.
In this experiment, the different wavelengths impinging upon a dispersing prism
are refracted by different amounts. This suggests that the index of refraction is
not a simple constant related to the permittivity and permeability of the material,
but that it is dependent on wavelength λ (or equivalently, angular frequency ω ).
It is indeed possible to derive a wavelength-dependent model for the index
of refraction to account for the wavelength-dependent nature of certain dielectric
materials.
We begin by assuming that electrons move with a velocity much smaller than
the speed of light. The Lorentz force acting on electrons, as given in (2.9), then
simplifies since the second term may be neglected. With a mass m, the position
of the electron, as a function of forces acted upon it, is modeled by the motion
equation:
∂ 2r
m 2 + q r = e E . (3.91)
∂t
If the electric field acting upon the electron is given by

E = E0 e−i ω t , (3.92)

then a solution in terms of r is given by [112]

e E
r=  . (3.93)
m ω02 − ω 2

Here, the quantity 

q
ω0 =
m

i i

i i
 i i

i i

186 3. Chemistry of Matter

is the resonance or absorption frequency. Thus, the angular frequency ω of the

incident field E causes the electron to oscillate with the same angular frequency.
The electron contributes a dipole moment p = e r to the total polarization P.
Under the assumption that the influence of the nuclei can be neglected, we there-
fore have for N atoms, each with a single electron with resonance frequency ω0 ,
a total polarization
e2 E
P = N er = N  2 . (3.94)
m ω0 − ω 2
However, if this equation is compared with (3.85), it can be seen that the mean
polarizability α is related to the angular frequency ω :

e2
Nα =N  . (3.95)
m ω02 − ω 2

We can now use the Lorentz-Lorenz formula (3.89) to relate the index of refrac-
tion to the angular frequency:

n2 − 1 4 π N e2
=  . (3.96)
n2 + 2 3 m ω02 − ω 2

It can be seen that the right-hand side of this equation tends to infinity for angular
frequencies ω approaching the resonance frequency ω0 . This is a consequence of
the formulation of the motion equation (3.91), which does not include a damping
term that models a resisting force. If this term were added, the position of an
electron would be modeled by

∂ 2r ∂r
m +g + q r = e E , (3.97)
∂ t2 ∂t
and the solution would be
e E
r=   . (3.98)
m ω02 − ω 2 − i ω g

As a result, the polarization P becomes a complex quantity, as does the index of

refraction n. The imaginary part of the index of refraction models the absorption
of the dielectric and is known as the extinction coefficient.
Materials normally have several resonance frequencies. Equation (3.96) may
therefore be extended as follows:
n2 − 1 4 π e2 Nk
n2 + 2
=
3m ∑ ω2 − ω2 . (3.99)
k k

i i

i i
 i i

i i

3.6. Polarization in Dielectric Materials 187

In this equation, Nk represents the number of electrons that correspond to res-

onance frequency ωk . While we omit the derivation (see [112]), this equation
may be rewritten for gasses as well as for liquids and solids. The derivation pro-
ceeds differently for gasses than for substances where n differs significantly from
unity. In either case, using the correspondence between angular frequency and
wavelength, ω /2π = c/λ , the above equation becomes
bk
n2 − 1 = a + ∑ . (3.100)
k λ 2 − λk2

The constants a and bk are given by

1
a=
c2 ∑ ρk λk2 , (3.101a)
k
1
bk = ρk λ 4 , (3.101b)
c2
e2 Nk
ρk = . (3.101c)
πm
The above equation is known as Sellmeier’s dispersion formula and the constants
a and bk are called Sellmeier’s constants. The refractive behavior of materials can
normally be modeled to a reasonable accuracy with a limited number of constants.
In many cases, taking into account around two resonant frequencies suffices—
especially in rendering applications.

Figure 3.43. Dispersion through a prism. The slit is visible towards the left. The dispersed
light is projected onto a sphere for the purpose of visualization.

i i

i i
 i i

i i

188 3. Chemistry of Matter

Figure 3.44. Dispersion through a prism. The light source is a Gretag Macbeth D65
daylight simulator, which produces light similar to that seen under an overcast sky.

Figure 3.45. A Gretag Macbeth D65 daylight simulator lights two sheets of polarizing
material from behind. Wedged between the two sheets are two pieces of crumpled cello-
phane.

i i

i i
 i i

i i

3.6. Polarization in Dielectric Materials 189

An example of dispersion is shown in Figure 3.43 and Figure 3.44. Here, a

D65 daylight simulator is used to shine light through a slit formed by two card-
board panels a short distance apart. The light is refracted and dispersed by a
35 mm dispersing prism. The dispersed light forms a rainbow pattern, which is
reflected by the diffuse white sphere.

Figure 3.46. Wavelength-dependent polarization shown by wedging pieces of crumpled

cellophane between polarizing sheets oriented at right angles.

Figure 3.47. Two polarizing sheets are aligned. The wedged pieces of cellophane produce
complementary colors to those shown in Figure 3.46.

i i

i i
 i i

i i

190 3. Chemistry of Matter

Another example, showing the relationship between polarization and wave-

length, is afforded by crumpled cellophane wedged between two linear polariz-
ers [447]. The setup of such an experiment is shown in Figure 3.45. Dependent on
its thickness, the cellophane produces different colors of the rainbow by rotating
the E field in a wavelength-dependent manner. Between crossed polarizing sheets,
the results are as shown in Figure 3.46. Complementary colors are produced if
the polarizing sheets are aligned, as shown in Figure 3.47.

3.7 Dichroism
There exist at least three phenomena which are each called dichroism. In this
section, we briefly describe each for the purpose clarifying the differences. In
each case, dichroism refers to the ability of a material or substance to attain two
different colors. However, the mechanism by which this two-coloredness occurs
is different.
First, dichroism can be related to polarization. In Section 2.3, we described
how birefringence may occur in certain minerals. In that case, the index of re-
fraction was dependent on the polarization of light. For certain materials, the
wavelength composition of transmitted light may depend on polarization as well.
Ruby crystals, for instance, have octahedral structures. If such crystals are lit by
polarized light, it is possible to change the color of the ruby from purple-red to
orange-red.
The absorption of polarized light by rods in certain species of fish has been
shown to differ by the angle of polarization [248]. Thus, at least some fish retinas
show dichroism of this form, particularly if the retina is unbleached.
A second effect occurs in other minerals, where the color of the transmitted
light is determined by the angle at which light strikes the stone. As such, the color
of transmitted light depends on the angle it makes with the optical axis of the ma-
terial. This effect is also known as dichroism, but it should be distinguished from
the case where the wavelength composition of transmitted light is dependent on
polarization. Iolite, a relatively common gem stone, exhibits this form of dichro-
ism, as shown in Figure 3.48. Some materials show different colors along three
or more optical axes, yielding an effect known as pleochroism.
Thirdly, dye dichroism refers to the different colors that dye solutions may
obtain as function of the concentration of the dye, as discussed in Section 3.4.3.
Finally, Dichroic filters and dichroic mirrors are based on interference effects.
Such devices selectively reflect part of the visible spectrum while transmitting the
remainder. They are built by coating a glass substrate with very thin layers of

i i

i i
 i i

i i

3.8. Application: Modeling of Fire and Flames 191

Figure 3.48. The stone in this pendant is made of iolite, a dichroic material. As a result,
viewing this stone from different angles reveals a different set of colors; in this case the
color varies between purple and yellowish-gray.

optical coatings. It is possible, for instance, to build a beam splitter by coating

several prisms with a dichroic optical material, and thus split light into red, green,
and blue components. Such beam splitters are used in some cameras that have
separate sensors for the three color channels. They are also commonly used in
LCD projectors (Section 14.12) as well as LCoS projectors (Section 14.14).

3.8 Application: Modeling of Fire and Flames

The chemical reactions occurring in a typical combustion process were discussed
in Section 3.5.2. Combustion can be classified into two groups, which are deto-
nations and deflagrations. The former are explosive reactions, whereas the latter
are slow reactions visible as flames and fire. During a deflagration the chemical
reactions cause an expansion of the fuel being burned, and create heat as well.
Both phenomena give rise to turbulence.
Modeling a flame can be achieved by finding a solution to the Navier-Stokes
equations for flow. Given a 3D grid with each grid point defining a velocity vec-
tor, each iteration simulates a time step, updating the velocity field, and thereby
the flow. These equations can be used for both fluid flow as well as gas flow.
There exist several variations of the Navier-Stokes equations, which model dif-

i i

i i
 i i

i i

192 3. Chemistry of Matter

ferent aspects of the fluid/gas. For flames, viscosity is unimportant, and therefore
the equations for incompressible flow are appropriate.
Considering a velocity field with each discrete position holding a velocity u,
then incompressibility can be enforced by ensuring that the divergence is zero:

∇·u = 0. (3.102)

Then, at time t the velocities are computed from the velocities of the previous
time step t − 1 using:

∇p
ut = − (ut−1 · ∇) ut−1 − + f, (3.103)
ρ

where p is the pressure, ρ is the density, and f models any external forces affecting
the flow. These quantities are defined at each grid point, and may or may not
themselves be updated for each time step. The above two equations together
are termed the Navier-Stokes equations for incompressible flow. Solutions to
this equation can be obtained for instance using semi-Lagrangian methods [307,
1079].
Unfortunately, the gas expansion discussed above cannot be taken into ac-
count with this model (it would require the more complex equations for com-
pressible flow). However, the Navier-Stokes equations for incompressible flow
can still be used, for instance by modeling the reaction zone where fuel is burnt as
a thin surface in 3D space, rather than a volume [741]. Recall from Section 3.5.2
that a deflagration requires oxygen which diffuses in from the surrounding air,
leaving a core within the flame where no reaction occurs. Thus, the reaction zone
is reasonably well modeled with an implicit surface.

3.8.1 Modeling a Flame

Nguyen et al. use a dynamic implicit surface to track a flame over time. Their
method is explained in some detail in this and the following subsection. They
use separate sets of equations to model the distribution of fuel and the flow of hot
gaseous products [835]. The updates to the two sets of Navier-Stokes equations
for each time step are coupled, however, enabling the conservation of mass and
momentum. To model turbulence, the Navier-Stokes equations are augmented
with a vorticity confinement term [307, 1084].
The implicit surface separates the gaseous fuel from the hot gaseous products
and the air around the flame. In a system where fuel is injected over a surface with
area A f , the fuel, which is premixed at first, will begin to heat up, and emit light.
In this stage, the color will be blue. Given a flame speed S indicating the rate with

i i

i i
 i i

i i

3.8. Application: Modeling of Fire and Flames 193

which fuel is burned (a fuel-dependent property [1155]), the surface area of the
blue core can be estimated to be AS using

v f A f = S AS , (3.104)

where v f is the speed of fuel injection. Thus, the size of the blue core depends on
the speed of injection in this model.
The reactions occurring in the flame, including the blue core, cause expansion
which is normally modeled as the ratio of densities ρ f /ρh , where ρ f is the density
of the fuel, and ρh is the density of the hot gaseous products. Across the dynamic
implicit surface modeling the thin flame, the expansion of the gas is modeled by
requiring conservation of mass and momentum, leading to the following equa-
tions:

ρh (Vh − D) = ρ f (Vh − D) , (3.105a)

2 2
ρh (Vh − D) + ph = ρ f (Vh − D) + p f . (3.105b)

Here, p f and ph are pressures, and V f and Vh are normal velocities of the fuel and
hot gaseous products. The speed of the implicit surface in its normal direction is
given by
D = V f − S. (3.106)
For the special case of solid fuels, the pressure will be zero; with the density and
normal velocity of the solid fuel given by ρs and Vs , we have
 
ρ f V f − D = ρs (Vs − D) , (3.107)

and therefore  
ρs
V f = Vs + − 1 S. (3.108)
ρf
Hence, the velocity of the gasified fuel in its normal direction at the boundary
between its solid and gaseous state is related to the velocity of the solid fuel plus
an expansion-related correction.
The dynamic implicit surface indicating the boundary of the blue core moves
at each time step with a velocity given by w:

w = u f + S n, (3.109)

where the velocity of the gaseous fuel is u f and S n is a term describing the veloc-
ity induced by the conversion of solid fuel to gas. At each grid point, the surface
normal n is determined by
∇φ
n= , (3.110)
|∇φ|

i i

i i
 i i

i i

194 3. Chemistry of Matter

where φ defines the region of the blue core. Its value is positive where fuel is
available, zero at the boundary, and negative elsewhere. The motion of the implicit
surface φ is then defined as
φt = −w · ∇ φ . (3.111)
This equation can be solved at each grid point, after which an occasional condi-
tioning step is inserted to keep the implicit surface φ well-behaved [835, 1027,
1146].
The implicit surface indicates the boundary between the fuel and hot gaseous
products. These two types of gas have their own velocity fields, u f and uh , which
are both independently updated with the Navier-Stokes equations above. As these
velocity fields are valid on opposite sides of the implicit surface, at this boundary
care must be taken to update the velocities appropriately. For instance, when
computing velocities uh for hot gaseous products near the implicit surface, values
on the other side of the boundary must be interpolated with the aid of the normal
velocity of the fuel V f , which is computed with

V f = u f · n. (3.112)

The corresponding normal velocity for the hot gaseous products can be computed
with  
ρf
Vh = V f +
G
−1 S (3.113)
ρh
The superscript G indicates that this is a ghost value, i.e., it is extrapolated for
a region where no hot gaseous fuel is present. The velocity of the hot gaseous
products on the opposite side of the implicit surface can now be computed with
 
uGh = Vh n + u f − u f · n n.
G
(3.114)

This mechanism allows the Navier-Stokes equations to be computed throughout

the volume, including the boundary of the flame.
The final component of the flame model involves external forces f which are
applied to the velocity field. These come in two forms. The first external force
is a dependency on temperature, which is defined as a scalar field T (discussed
below) and leads to buoyancy. The temperature-dependent force is defined as

fT = α (T − Tair ) z, (3.115)

where α is a user-defined constant, and z = (001)T is the up vector.

The second external force is termed vorticity confinement and constitutes a
correction to add turbulence and swirling, which is otherwise dampened as a re-
sult of the low resolution of the velocity field that is normally used in rendering

i i

i i
 i i

i i

3.8. Application: Modeling of Fire and Flames 195

applications [1084]. Vorticity confinement fC is defined as

fC = ε h (nω × ω ) . (3.116)

The amount of vorticity is determined by ε , with ε > 0, and h is the spacing

between the grid points. The vorticity ω and normalized vorticity gradients nω
are computed with

ω = ∇× u, (3.117a)
∇ |ω |
nω = . (3.117b)
| ∇ |ω ||
As fuel is converted from solid to gas, ignites, and moves through the blue
core out into the area where the flame is categorized as a diffusion flame (see
Section 3.5.2), the temperature follows a profile which rises at first and then cools
down again. To model this both over space and over time, it is useful to track fuel
particles during their life cycle, as follows:

Yt = − (u · ∇) Y − 1. (3.118)

This equation is solved once more with the aid of semi-Lagrangian fluid solvers.
With suitably chosen boundary conditions, the value 1−Y can be taken to indicate
the time elapsed since a particle left the blue core. The temperature T is then
linearly interpolated between the ignition temperature Tignition and the maximum
temperature Tmax , which occurs some time after leaving the blue core. In the
temperature fall-off region, with the help of a cooling constant cT , the temperature
can be computed with [835]
 
T − Tair 4
Tt = − (u · ∇) T − cT . (3.119)
Tmax − Tair
Finally, the density ρt at time step t is computed similarly, using

ρt = − (u · ∇) ρ . (3.120)

3.8.2 Rendering a Flame

While this procedure models the flame and the progression of its shape over time,
the rendering can be achieved by a ray-marching algorithm, visiting each grid el-
ement along its path. The temperature recorded at each grid point can be used
to compute the amount of blackbody radiation (see Section 3.5.1). The technique
for rendering a flame bears strong resemblance to the atmospheric rendering tech-
nique discussed in Section 2.11.6.

i i

i i
 i i

i i

196 3. Chemistry of Matter

Figure 3.49. A rendering of a campfire using the techniques discussed in this section.
(Image courtesy of Henrich Wann Jensen.)

However, in this case, it is appropriate to use a more general phase function

to account for scattering events. To model the spherical distribution of scattered
light, the Henyey-Greenstein phase function p can be used, which only depends
on incoming and outgoing directions Θ and Θ [460]:

1 − g2
p(Θ, Θ ) = 1.5
. (3.121)
4π (1 + g2 − 2g Θ · Θ )
This function is then used in (2.234), which models in-scattering. Accounting for
in-scattering, out-scattering, as well as light emitted along each ray segment, we

i i

i i
 i i

i i

3.9. Further Reading 197

use (2.233). However, light emission for modeling flames is governed by Planck’s
radiation law, so that Leλ equals the spectral radiant exitance Meλ given in (3.62).
An example result obtained with this combined modeling and rendering tech-
nique is shown in Figure 3.49, showing a campfire where a cylindrical log emits
fuel which then catches fire.

3.9 Further Reading

A more elaborate discussion on the sources of color can be found in Nassau [815]
and Tilley [1128]. Wyszecki and Stiles is a must-have reference work, contain-
ing a vast amount of data [1262]. Several books on quantum mechanics were
consulted for the preparation of this chapter [179, 683, 1109]. Williamson and
Cummins also covers some of the material in this chapter [1247].

i i

i i
 i i

i i

Chapter 4
Human Vision

In the preceding chapters, light and its interaction with matter was discussed. One
particularly important interaction occurs when light enters the human eye. Light
falling on the retina triggers a most remarkable sequence of events. In this and the
following chapter, we will focus on this chain of events, insofar the current state of
knowledge allows us. It should be noted that much of the neurological underpin-
nings of color still forms an active area of research, and knowledge of higher-level
color processing in the human brain therefore remains relatively sparse [496]. In
addition, the more we learn about the functional organization of the visual cortex,
the more complex the human visual system turns out to be. Nonetheless, we con-
sider this topic the other half of essential background information, necessary to
understand and effectively use theories of color.
In this chapter, we primarily deal with the substrate that forms the human
visual system: the eye, the retina, and neural circuitry in the brain. Much of
the information collected in this chapter stems from studies of Old-World mon-
keys [227, 1032]. It has been shown that their photoreceptors have the same spec-
tral tuning as human photoreceptors [78, 1009, 1010], as well as a similar retinal
layout [972]. In addition, cell types and circuits found in the macaque visual
system are very similar to those in the human retina [224–226,604,606,608,892].
Both anatomy and histology use terms and descriptions for structures relative
to the orientation of the body. Figure 4.1 shows the planes of reference used to
describe where structures are. These planes are [329]:

The midsagittal plane, which divides the head into a left and right half. Objects
located closer to this plane are called medial and those that are further away
are called lateral.

The coronal plane divides the head into an anterior and posterior region.
199

i i

i i
 i i

i i

200 4. Human Vision

Posterior

Anterior

Lateral
Medial

Horizontal plane Superior

Inferior

Median or
midsagittal plane Coronal plane

Figure 4.1. Terminology used to indicate anatomic planes of reference.

The horizontal plane divides the head into a superior and inferior region. These
regions are also called cranial and caudal.

One further term of reference is deep, indicating distance from the surface of the
body.

4.1 Osteology of the Skull

The skull consists of an upper part, called cranium or neurocranium, and a lower
part, called facial skeleton or viscerocranium. The cranium itself consists of the
cranial vault and the cranial base.
The skull contains four cavities, each with a different function:

The cranial cavity houses the brain.

The nasal cavity serves in respiration and olfaction.
The orbits house the eyes.
The oral cavity forms the start of the gastrointestinal tract.

i i

i i
 i i

i i

4.2. Anatomy of the Eye 201

The two orbital cavities lie between the cranium and the facial skeleton, and
are separated by the nasal cavity. They serve as sockets to hold the eyes as well as
the adnexa, and they transmit nerves and blood vessels that supply the face around
the orbit [329].
The shape of each orbit is roughly a quadrilateral pyramid with the apex form-
ing the optic canal (which provides passage for the optic nerve and the ophthalmic
artery between the orbits and the cranium). Their walls are approximately trian-
gular, except for the medial wall which is oblong. The medial walls are parallel
to the midsagittal plane, whereas the lateral walls are angled at 45 degrees to this
plane. The height, width, and depth of the orbits are around 40 mm each, with a
volume of 30 ml.

4.2 Anatomy of the Eye

The eyes are used for photoreception and for communication of the resulting ac-
tion potentials to the brain. A diagram of a horizontal cross-section is shown in
Figure 4.2. The eyes are located in the anterior portion of the orbit, closer to the
lateral wall than the medial wall, and closer to the roof than the floor of the orbit.
The shape of the eye is approximately that of two intersecting spheres, the
smaller one forming the cornea and the larger one the sclera. The cornea has
a radius of 7.8 mm and forms one-sixth of the circumference of the eye. The
sclera accounts for the rest of the circumference, forming a sphere with a radius
of 11.5 mm. The axial length of the globe is on average 24 mm.

Pupil Cornea
Trabecular Anterior chamber
meshwork Iris
Posterior chamber
Canal of Limbus
Schlemm
{

Zonular apparatus
Lens Rectus muscle,
Ciliary body
tendon & belly

Retina
Sclera Optic disk
Fovea Vitreous
{

Choroid
{

Corneo-scleral coat
Retinal Uveal tract
{

pigment Neural layer

Optic nerve epithelium Contents

Figure 4.2. Diagram of a horizontal cross-section of the human eye.

i i

i i
 i i

i i

202 4. Human Vision

The eye has a layered structure, with three layers (or tunics) present: the outer
corneoscleral envelope, the uveal tract, and the inner neural layer. The sclera and
the cornea together form the corneoscleral coat, which provides a tough outer
shell supporting and protecting the inner layers as well as the contents of the eye.
Extraocular muscles attach to this layer to mediate eye movement. The uveal
tract, or uvea, consists of the iris, ciliary body, and the choroid and contains two
openings: the pupil anteriorly and the optic nerve canal posteriorly. The innermost
layer consists of the retina which is located in the image plane of the eye. It has
two layers, the inner neurosensory retina and the retinal pigment epithelium. The
retina is described in greater detail in Section 4.3.
The transparent contents of these three layers are the aqueous humor, the lens,
and the vitreous body. Light passes through these parts as well as through the
cornea before it reaches the retina. In the following, we describe each of the
components of the eye that are involved in the projection of images on the retina.

4.2.1 Eye Movements

The ability of eyes to move in their sockets is important for two reasons. First,
the visual acuity is highest in the fovea. Thus, eye movement is required to orient
the eyes such that features of interest are projected onto the fovea. This type of
eye movement is called gaze shifting. Second, the human visual system should be
able to track both stationary as well as moving objects (gaze holding), even when
the head of the observer itself is moving.
The eye can move about 45 degrees in all directions, although in practice
the eye rarely moves more than 20 degrees. Objects further in the periphery
are tracked with a combination of eye and head movements [166]. If the an-
gle between the optical axes of the two eyes does not change as a result of eye
movement, we speak of version movement. If the angle between the two opti-
cal axes changes, the eye movement is called vergence movement (or disjunctive
movement) .
Version movements are involved in bringing different parts of the visual field
onto the fovea. These movements, called saccades, are very fast, lasting around
100 ms. During this time, angular velocities can reach 700 deg/s [969]. Normal
vision is largely suppressed during these movements.
Vergence movements are typically slower, with a maximum velocity of 200
deg/s [285,286]. The predominant stimulus for vergence movements is the differ-
ent projected location of objects on the fovea (called disparity).
During fixation, an observer’s eyes are not completely stationary, but are sub-
ject to very small movements. Such miniature eye movements have three compo-
nents, namely drift, tremor, and microsaccades [166]. For a given fixation point,

i i

i i
 i i

i i

4.2. Anatomy of the Eye 203

1.0
p(Θ) 0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
-30 -20 -10 0 10 20 30
Θ

Figure 4.3. The normalized probability of the eye deviating from the fixation point as a
function of deviation angle (after [75, 321]).

the probability of the eye deviating from a fixation point as a result of miniature
eye movements can be modeled as the following function of angle Θ (plotted in
Figure 4.3) [75]:

p(Θ) = 0.62 e−(0.167 Θ) + 0.38 e−(0.05 Θ) .

2 2
(4.1)

These miniature movements are important for normal vision. If they are artifi-
cially suppressed, the visual image may break down or disappear altogether [256].

4.2.2 The Cornea

The cornea has a diameter of 10.6 mm vertically, and 11.7 mm horizontally, al-
though if viewed from behind its circumference is circular. The central diameter
is 7.8 mm with the remainder of the cornea being less curved. Its thickness varies
between 0.52 mm in the center to 0.67 mm in the periphery.
The refractive index of the cornea (n = 1.376) and its associated tear film are
responsible for most of the focusing of light that occurs in the eye. This index of
refraction changes little with age.
The cornea consists of five layers: the corneal epithelium, the anterior limit-
ing lamina (Bowman’s layer) , the corneal stroma, the posterior limiting lamina
(Desçemet’s membrane), and the corneal endothelium. Each of these layers are
transparent, with factors such as regularity of the intracellular and extracellular
structures, smoothness of the epithelium and its absence of blood vessels con-
tributing to its transparency.

i i

i i
 i i

i i

204 4. Human Vision

The cornea consists of fibrils which are large compared with the wavelength of
light. This was established by measuring the birefringence of the fibrils [236,755],
suggesting that scattering in the cornea should lead to a highly opaque material.
However, these fibrils are uniform in shape and are arranged in parallel rows.
The diffracted light thus cancels out, leaving the undiffracted light which passes
through the cornea unhindered [755].
For medical purposes, some work has focused on modeling the cornea [70,
418].

4.2.3 The Iris

The iris is suspended in the coronal plane anterior to the lens, surrounded by the
aqueous humor which fills the anterior and posterior chambers. The free edge of
the iris is called the pupillary margin.
The iris has a circumference of 37 mm and a 12 mm diameter. Its pupillary
margin is located more anteriorly than the root. The pupil varies between 1 to
8 mm in diameter, dependent on the state of contraction of two muscles present
in the iris: the sphincter pupillae muscle and the dilator pupillae muscle. The
anterior surface of the iris has a thickened ring called the collarette, which divides
the iris into an inner pupil zone and an outer ciliary zone.
There are four layers in the iris: the anterior border layer, the stroma, the
dilator pupillae, and the posterior pigment epithelium. The anterior border layer
has gaps allowing the flow of fluid between the aqueous humor in the anterior
chamber and the stroma during contraction and dilation. These gaps are called
crypts. Heavily pigmented clumps of melanocytes also appear, which are known
as naevi. The stroma also contains pigments, but it is also the main source of scat-
tering, causing the variety of eye colors seen in mankind. The sphincter muscle,
responsible for contraction of the pupil, lies free in this layer. The dilator pupil-
lae forms the third layer and mediates dilation of the pupil. The fourth layer is
the posterior pigment epithelium, which is always heavily pigmented and appears
black.
The complexity of different layers, crypts, naevi, muscles, and other folds,
furrows, and radial streaks creates an intricate pattern which is unique for every
person. Artificially recreating such patterns, for instance, for the creation of ocu-
lar prosthetics, requires a skillful combination of both art and science. Ocularists
usually paint an iris button with up to thirty to seventy layers of paint, alternated
with varnish to provide depth (Figure 4.4). A non-photorealistic rendering ap-
proach was taken by Lefohn et al., which was inspired by the work of ocularists
(Figure 4.5) [667]. Finally, the uniqueness of each iris makes them suitable for
use in identification purposes [234].

i i

i i
 i i

i i

4.2. Anatomy of the Eye 205

Figure 4.4. An ocular prosthesis. The iris was created by alternating dozens of layers of
paint with clear coats. Compare with the green eye of Figure 2.46, which served as the
model for this prosthesis. (Prosthesis courtesy of Richard Caruso, Eye Prosthetics of Utah,
Inc.)

Figure 4.5. Non-photorealistic rendering of an eye. This rendering is matched to the

prosthesis of Figure 4.4 and the eye of Figure 2.46. (Aaron Lefohn, Richard Caruso,
Erik Reinhard, Brian Budge, and Peter Shirley, “An Ocularist’s Approach to Human Iris
Synthesis,” IEEE Computer Graphics and Applications, 23(6), c 2003 IEEE.)

4.2.4 The Pupil

The pupil controls the amount of light entering the eye. Its size can be adjusted
at a rate of 4 Hz [699] and depends on the amount of ambient illumination, as
plotted in Figure 4.6. There are many other factors that influence pupil diameter,

i i

i i
 i i

i i

206 4. Human Vision

pupil diameter (mm)

6

-4 -2 0 2 4
log luminance (cd/m 2)

Figure 4.6. Entrance pupil diameter as a function of ambient illumination for young
adult observers (after [305]). The photometric unit of candela per square meter is formally
introduced in Chapter 6.

including age, accommodation, drugs, and emotion [166]. The range of pupillary
constriction affords a reduction in retinal illumination by at most a factor of seven
(0.82 log units [118]) to ten [1262].

4.2.5 The Lens

Posterior to the iris and the pupil lies the lens [236]. It is suspended between the
aqueous humor and the vitreous body and connects to lens zonules (suspensory
ligament), which in turn connect to the ciliary body. The latter contains the ciliary
muscle, which together with the elasticity of the lens itself is capable of changing
the refractive power of the lens, a process called accommodation. This process
allows the eye to focus on objects both near and far. The ability to accommodate
diminishes with age, which is why many older people need reading glasses.
The transparency of the lens is mediated by its shape, internal structure, and
biochemistry. It receives nourishment from the aqueous humor and the vitreous
body, thereby allowing the lens to be avascular. The curvature of the lens anteri-
orly has a radius ranging from 8 to 14 mm, whereas the curvature posteriorly has
a radius between 4.5 and 7.5 mm. The lens has a diameter of around 10 mm and
an axial length of 4 mm, although the lens continues to grow at a rate of 0.0023
mm per year.
The lens consists of three layers: the capsule, the lens epithelium, and the lens
fibers. The capsule, which envelopes the lens, is elastic and shapes the lens to be
more rounded when not under tension by the ciliary muscle.

i i

i i
 i i

i i

4.2. Anatomy of the Eye 207

1.0

Diopter D = 1/f
0.0

-1.0

-2.0

400 500 600 700

Figure 4.7. Chromatic aberration in the human eye, measured in diopters D, as function
of wavelength λ (after [80]).

The curvature of the lens is not the same in horizontal and vertical direc-
tions, leading to astigmatism (the phenomenon that horizontal and vertical lines
have different focal lengths). A consequence of astigmatism is that, at the retina,
horizontal edges will consistently show colored fringes, chromatic aberration,
whereas vertical edges will show differently colored fringes [477]. Both Young
and Helmholtz have demonstrated chromatic aberration in their own eyes [164,
165], and this effect was later measured by Wald and Griffin [1191] and Bed-
ford and Wyszecki [80]. The amount of chromatic aberration in the human eye is
measured in diopters, i.e., the reciprocal of focal length, as shown in Figure 4.7.
This effect is believed to be corrected with later neural mechanisms and possi-
bly also by waveguiding due to the packing of photoreceptors, as discussed in
Section 4.2.7.
Other than accommodation, a second function of the lens is to prevent optical
radiation between 295 and 400 nm from reaching the retina [344]. This filter-
ing is required to protect the retina from damage by ultra-violet (UV) radiation.
With age, the lens yellows and also becomes more opaque, which increases the
amount of UV filtering that occurs. The yellowing is due to photochemical re-
actions rather than biological aging [254, 255, 344, 1223] and causes changes in
color perception with age. The decrease in light transmittance is in part due to
disturbances in the regular packing of crystallins, which in turn induces increased
scattering [329], thus yielding a more opaque lens. This is known as a cataract
(see also Section 5.11).
Finally, the lens is a source of fluorescence (Section 3.5.4), in that incident
UV radiation is absorbed by the lens and re-emitted in the visible range. This
phenomenon increases with age. For a healthy lens of a 30-year old, the ratio
between incident (sky) light and the amount of produced fluorescence is around

i i

i i
 i i

i i

208 4. Human Vision

0.002 [1222]. This ratio increases with age to around 0.017 (noticeable) for a
60-year old and 0.121 for an 80-year old. At these ratios, the occurrence of fluo-
rescence is seen as veiling glare.
Dependent on the wavelength of the incident UV light, the range and peak
wavelength of the fluorescence varies. For instance, incident light at 360 nm
produces fluorescence with a range between 380 nm to over 500 nm, with a peak
at 440 nm [672,1302]. The wavelength of 360 nm is near the absorption maximum
of the lens.
In the aging lens, the UV absorption band additionally broadens into the blue,
suggesting that human vision may be aided by employing filters that block UV
and short wavelengths in the visible range. This would reduce both veiling glare
as well as fluorescence [657, 1297].

4.2.6 The Ocular Media

In addition to accommodation, the ciliary body is responsible for the production
of the aqueous humor, the fluid that fills the anterior and posterior chambers. Fluid
secreted by the ciliary body first enters the posterior chamber, which communi-
cates with the anterior chamber through the pupil. The flow continues towards
the trabecular meshwork and the canal of Schlemm (Figure 4.2). Replenishment
of the aqueous humor occurs at a rate of 4 μ l per minute. The anterior chamber
holds around 250 μ l of fluid.
The functions of the aqueous humor are to nourish the cornea and the lens,
drain toxic metabolic waste, and maintain pressure on the cornea and lens. The
latter is achieved by balancing the production and drainage of the fluid and is
necessary for the normal functioning of the eye.
The vitreous cavity makes up about two-thirds of the volume of the eye. The
contents are called the vitreous body or vitreous and weighs 3.9 grams. It consists
of a transparent viscoelastic gel with an index of refraction of 1.337. The vitreous
body is largely acellular.

Medium n
Cornea 1.376
Aqueous humor 1.336
Lens (center) 1.406
Lens (outer layers) 1.386
Vitreous body 1.337

Table 4.1. Indices of refraction for the different ocular media.

i i

i i
 i i

i i

4.2. Anatomy of the Eye 209

Image formation on the retina is thus mediated by the cornea, the aqueous
humor, the lens, and the vitreous body. Their respective refractive indices are
summarized in Table 4.1 [447]. Together, they are able to focus light on the
retina. However, the ocular media is not a perfect transmitter. Light losses occur
due to Fresnel reflection, absorption, and scattering [1262]. Fresnel reflection is
generally less than 3 to 4 percent [166]. The optical system is optimized for use in
open air. For instance, in underwater environments, the index of refraction from
water to cornea is different, and vision appears blurred as a result [208].
Further, the index of refraction of the ocular media is dependent on wave-
length, causing edges projected on the retina to show colored fringes, known as
chromatic dispersion [50].
Aberrations are strongest at the first refractive surface, the anterior corneal
surface. The aberrations occurring here only slightly increase with age and are
largely corrected by subsequent ocular surfaces in young individuals. However,
the corrective power of later refractive boundaries in the ocular media decreases
with age, so that the total aberration at the retina increases with age [43]. For
instance, the continued growth of the lens causes aberration of the lens to increase
with age [367].
Absorption of short wavelengths occurs at the cornea. This may disrupt sur-
face tissue, leading to photokeratitis (snow blindness or welder’s flash) [166].
The damage threshold is lowest (0.4 J/cm2 ) for light at 270 nm [904]. Further
absorption at short wavelengths occurs in the center of the lens [772]. This ab-
sorption increases with age. Finally, anterior to the outer segments of the cones
in the fovea lies a layer of yellow pigment (xanthophyll), which absorbs light at
short wavelengths [1065]. The average transmittance of the human eye is plotted
in Figure 4.8.

1.0
Normalized Transmittance

0.8

0.6

0.4

0.2

0.0
400 600 800 1000 1200 1400
λ

Figure 4.8. Wavelength-dependent transmittance of the ocular media (after [351]).

i i

i i
 i i

i i

210 4. Human Vision

1.4
p(Θ) 1.2

1.0

0.8

0.6

0.4

0.2

0.0
-10 -8 -6 -4 -2 0 2 4 6 8 10
Θ

Figure 4.9. Point spread function for ocular light scatter (after [75]).

Some light is scattered in the eye. This means that if the eye is exposed to
a point light source, a small area of the retina is illuminated, rather than a single
point. Ocular light scatter is also known as veiling glare. The intensity fall-off
as a function of radial angle (in arc minutes) can be modeled by a point spread
function p(Θ) [75]:
⎧
⎨1.13 e−5.52 Θ2 + 1.42
, |Θ| ≤ 7 ,
p(Θ) = 9.85 + Θ3.3 (4.2)
⎩9.35 × 10−6 (7.8 + Θ)−2.5 , |Θ| > 7 .

This function is plotted in Figure 4.9, and following Baxter et al. [75], it consists
of data collected for visual angles less than 5 arc minutes [615] superposed with
data for angles larger than 9 arc minutes [336, 402, 1185].

4.2.7 The Retina and the Retinal Pigment Epithelium

The (neural) retina is a neurosensory layer able to transduce light into signals and
transmit those to the brain. It is largely a loose layer, attached only at its anterior
termination and at the margins of the optic nerve head [329]. It has a surface area
of 1250 mm2 and varies in thickness between 100 μ m and 230 μ m.
The retina consists of various regions, as indicated in Figure 4.10. While
different terminology is in use by clinicians and anatomists, here we will use the
anatomic terms:

The area centralis is circular with a diameter of 5 to 6 mm. It lies between the
superior and inferior temporal arteries.

i i

i i
 i i

i i

4.2. Anatomy of the Eye 211

Clinical terms Anatomic terms

Central posterior retina Area centralis
Macula Fovea
Fovea Foveola
Optic disk Optic disk

Peripheral retina with blood vessels

Figure 4.10. The retina consists of various regions of interest (after Forrester et al. [329]).

The fovea is 1.5 mm in diameter and lies 3 mm lateral to the optic disk. It contains
a layer of xanthophyll pigment which gives it a yellowish color (hence the
clinical term macula lutea).

The foveola is a 0.35 mm-diameter zone which consists of a depression sur-

rounded by a rim. This area is avascular. The visual acuity of the retina
(resolving power) is highest in this region.

The optic disk contains no retinal layer or photoreceptors. It is known as the blind
spot. It lies 3 mm medial to the center of the fovea.

The peripheral retina is the remainder of the retina.

The histology of the neurosensory retina is further discussed in Section 4.3.

The retina is backed by the retinal pigment epithelium, which provides ad-
hesion and nourishment to the retina. It is also heavily pigmented with melanin,
which aids in the absorption of light. This reduces light scatter in the eye and thus
improves image resolution.
A peculiarity of the retina is that its BRDF is not Lambertian, which implies
that its ability to absorb light depends on the direction of incidence. As a con-
sequence of the positioning of the retina with respect to the pupil, light reaching
the periphery of the retina will be incident at an angle, where it is less efficiently
absorbed than in the center of the retina.
Thus, light entering the pupil through the center and reaching a given position
on the retina will be absorbed better than light reaching the same location after

i i

i i
 i i

i i

212 4. Human Vision

passing through the pupil nearer the pupillary margin. This is known as the Stiles-
Crawford effect of the first kind [167, 789, 1088].
At the same time, the absorption shows a wavelength dependence with angle
of incidence, causing changes in hue with eccentricity of retinal location. This
is known as the Stiles-Crawford effect of the second kind [34, 1089, 1262]. One
of the explanations for this effect is that light incident upon the retina at an angle
passes through matter over a longer distance before being absorbed. As a result,
more attenuation has occurred, hence explaining color difference with angle of
incidence [85]. In addition, backscattering off the choroid may account for some
of this effect [85].
Waveguiding is also implicated in the explanation of the Stiles-Crawford ef-
fects [40, 1183]. The size of the photoreceptors and their dense packing creates
an effect similar to fiber-optics (more on waveguiding in Section 4.7), creating di-
rectional sensitivity. The waveguiding appears to reduce the effect of aberrations
in the ocular media [1183].

4.3 The Retina

When light enters the eye, it passes through the ocular media, and then passes
through several layers of cells before being absorbed by the photoreceptors. The
photoreceptors are thus at the very back of the retina [803], in close contact with
the retinal pigment epithelium, which provides the photoreceptors with retinal
(vitamin A). This molecule is essential in photo-transduction.

Fovea

Photoreceptors
Periphery Neural layers

Figure 4.11. Neural cell layers are pushed aside in the fovea to allow more light to reach
the photoreceptors.

i i

i i
 i i

i i

4.3. The Retina 213

Fovea Optic disk

mm 0.3

0.0
Temporal Nasal

Figure 4.12. The thickness of the retinal layers varies with position (after [329]).

The layers of cells preceding the retina are varied in thickness, and, in partic-
ular in the fovea, light passes through a much thinner layer before reaching the
photoreceptors, as shown in Figure 4.11. A plot of the variation in thickness of
the retina is given in Figure 4.12.
The neurosensory retinal layer transduces light into neural signals that are then
transmitted to the brain for further processing. To this end, the retina consists of
several types of cells, including photoreceptors, bipolar, ganglion, amacrine, and
Incident light direction

Ganglion cells
C Amacrine cells
Bipolar cells
D Horizontal cells
Photoreceptors

F
Rods Cones

A Optic nerve fiber layer E Outer plexiform layer

B Ganglion cell layer F Outer nuclear layer
C Inner plexiform layer G Inner and outer segments
D Inner nuclear layer of rods and cones

Figure 4.13. Connectivity of neural cells in the retina (after [329]). In this figure, light
comes from the top, passes through the layers of ganglion, amacrine, bipolar, and half cells
before reaching the light sensitive photoreceptors.

i i

i i
 i i

i i

214 4. Human Vision

half cells [262]. Within each class, further sub-classifications are possible based
on morphology or response type. Dependent on species, there are between one
and four types of horizontal cells, 11 types of bipolar cells, between 22 and 40
types of amacrine cells, and around 20 types of ganglion cells [227, 609, 723, 749,
1177, 1217]. Although the predominant cell types are all involved in neural pro-
cessing of visual signals, the retina also contains glial cells, vascular endothelium,
pericytes, and microglia.
The photoreceptors, bipolars, and ganglion cells provide vertical connectiv-
ity [915], whereas the horizontal and amacrine cells provide lateral connectivity.
A schematic of the cell connectivity of neural cells in the retina is given in Fig-
ure 4.13. These cells are organized into three cellular layers, indicated in the
figure with blue backgrounds, and two synaptic layers where cells connect. The
latter are called the inner and outer plexiform layers [262]. Each of the cell
types occurring in the retina is discussed in greater detail in the following sec-
tions. Visual processing typically propagates from the outer layers towards the
inner retina [935].
Cells involved in the transmission of the signal to the brain can be stimulated
by a pattern of light of a certain size, shape, color, or movement. The pattern
of light that optimally stimulates a given cell, is called the receptive field of the
cell [430]. Photoreceptors, for instance, respond to light directly over them. Their
receptive field is very narrow. Other cells in the visual pathways have much more
complex receptive fields.

4.3.1 Photoreceptors
Photoreceptors exist in two varieties, namely rods and cones [1015]. The rods
are active in low light conditions and mediate the perception of contrast, bright-
ness, and motion. Such light conditions are called scotopic. Cones function in
bright light (photopic light levels) and mediate color vision. Scotopic and pho-
topic lighting conditions partially overlap. The range of light levels where both
rods and cones are active is called the mesopic range.
There are approximately 115 million rods and 6.5 million cones in the retina.
The density of rods and cones varies across the retina. The periphery consists
predominantly of rods, having a density of around 30 000 rods per mm2 . The
cone density is highest in the fovea with around 150 000 cones per mm2 . In the
fovea there are no rods [1015]. The rod and cone density across the retina is
plotted in Figure 4.14.
Rods and cones form the beginning of a sequence of processing stages called
pathways. Pathways are sequences of cells that are interconnected. For instance,
photoreceptors are connected to bipolar cells as well as to half cells. Bipolars are

i i

i i
 i i

i i

4.3. The Retina 215

Fovea Blind spot

+ 10 5 /mm 2
1.6

1.2

0.8 Cones
Rods
0.4

0.0
-60 0 60
Angular extent

Figure 4.14. The density of rods and cones in the retina (after [901]).

also interconnected by a layer of amacrine cells before connecting to the ganglion

cells. The ganglion cells in turn carry the signal away from the eye and into the
lateral geniculate nucleus (LGN), a small area in the brain that in turn connects to
the visual cortex.
The interconnections at each stage along the visual pathways are complex
and varied. In addition to feed-forward processing, all stages and neural layers
have feedback loops. There also exists feedback from the brain back into the
retina [347, 482].
A full understanding of the visual processing occurring in the retina remains
beyond our abilities. Nonetheless, more is known about the retina than most other
centers of processing in the brain. In the remainder of this chapter, we give a
brief overview of some of the peculiarities that are known about the human visual
system, as this is the substrate of human (color) vision.

4.3.2 Rods
Under favorable circumstances the human visual system can detect a single pho-
ton [446, 1018, 1205], a feature attributed to the sensitivity of the rods, and the
pooling occurring in the rod pathway. Rods have a peak density in a ring 5 mm
from the center of the fovea [862] (see also Figure 4.14). Rods are packed in a
hexagonal pattern, separating cones from each other.
Much is known about the molecular steps involved in light transduction in
both rods and cones [289,928]. A simplified overview is given here. Rods contain
a visual pigment called rhodopsin, which has a peak spectral sensitivity at λmax =
496 nm (see Figure 4.15). Whenever a rhodopsin molecule absorbs a photon, a
chemical reaction occurs that bleaches the molecule, leaving it in a state that does

i i

i i
 i i

i i

216 4. Human Vision

Log relative sensitivity

2

-2

-4

400 500 600 700

Wavelength (nm)

Figure 4.15. The log relative sensitivity of the rods.

not allow it to absorb another photon. This reduces the sensitivity of the rod. After
some period of time, the process is reversed, and the rhodopsin is regenerated.
The regeneration process is relatively slow, although not as slow as the process
of dark adaptation. Dark adaptation can be measured psychophysically as the time
it takes to regain sensitivity after entering a dark room. While dark adaptation may
take at least 30 minutes, rhodopsin regeneration requires around 5 minutes. Some
of the reasons for the duration of dark adaptation are outlined in Section 3.4.2.
The rate of regeneration is related to the amount of rhodopsin present. If
p ∈ [0, 1] is the fraction of unbleached rhodopsin, the rate of change d p/dt can be
described by [118]:
dp 1− p
= , (4.3)
dt 400
where 400 is a time constant. A solution to this differential equation is given by
 
−t
p = 1 − exp . (4.4)
400
This behavior was indeed observed [987]. The bleaching process follows a similar
rate of change [118]:
dp −Lv p
= . (4.5)
dt 400 L0
Here, Lv is the intensity of the light causing the bleaching, and L0 is the intensity
that bleaches 1/400 of the rhodopsin per second. The solution to this differential
equation is  
−Lv t
p = exp . (4.6)
400 L0

i i

i i
 i i

i i

4.3. The Retina 217

The combined effect of bleaching and generation is then given by

 
dp 1 Lv p
= 1− p− . (4.7)
dt 400 L0

As bleaching and regeneration occur concurrently, exposure to a constant amount

of light will cause the amount of rhodopsin to converge to a steady state. Once
such a steady state is reached, the rate of change by necessity becomes zero. As a
result, (4.7) gives a relation between the fraction of unbleached rhodopsin p and
the steady state illumination Lv :
Lv
1− p = . (4.8)
Lv + L0
The value of L0 is known as the half-bleach constant. This equation shows that
for very small values of Lv  L0 , the relation between Lv and p is nearly linear.
For very large values of Lv , the fraction of unbleached rhodopsin asymptotically
approaches zero. Centered around L0 , the range of intensities for which p is
between 0.01 and 0.99 is approximately 2 log units (i.e., L p=0.01 /L p=0.99 ≈ 100).

4.3.3 Cones
The bleaching and regeneration processes known to occur in rhodopsin also occur
in the opsins (visual pigments) associated with cones. The rate of change of opsins
p as function of intensity Lv is given by [118]
 
dp 1 Lv p
= 1− p− . (4.9)
dt 120 L0

As only the constant 120 is different from (4.7), the steady-state behavior for
opsins is the same as for rhodopsin and is given by (4.8).
The cones come in three types which have peak sensitivities to different wave-
lengths. These three types can be broadly classified as sensitive to long, medium,
and short wavelengths (abbreviated as L, M, and S cones), as shown in Fig-
ure 4.16.1 The peak sensitivities λmax lie around 440 nm, 545 nm, and 565 nm,
respectively [1036]. With three different cone types, human color vision is said
to be trichromatic.
Figure 4.16 also shows that each cone type is sensitive to a wide range of
wavelengths, albeit in different relative proportions. The signal produced by a
photoreceptor is, therefore, proportional to the mix of wavelengths incident upon
1 Note that the curves shown here are not corrected for scattering and absorption of the ocular

media.

i i

i i
 i i

i i

218 4. Human Vision

1.0 S cone
M cone

Normalized sensitivity
L cone

0.5

0.0
400 500 600 700
λ

Figure 4.16. The normalized sensitivities of the three cone types peak at different wave-
lengths (after [1092]).

the receptor, weighted by the receptor’s sensitivity curve. As a result, a single

cone is not able to distinguish color. The perception of color arises from the
comparison of signals produced by different cone pigments later in the processing
pipeline.
The cone output is thus a weighted sum over all incident wavelengths, with
the sensitivity curves shown in Figure 4.16 as the weighting function. This gives
rise to metamerism, i.e., the fact that two different distributions of wavelengths are
perceived to have the same color. Light of different spectral compositions may in-
duce the same perception of color if they lead to an identical triplet of absorptions

0
-1
log S (λ max /λ)

-2
-3
-4
-5
-6
-7
-8
0.7 0.8 0.9 1.0 1.1 1.2 1.3
λ max /λ

Figure 4.17. The log spectral sensitivity template plotted against normalized wavelength
λmax /λ .

i i

i i
 i i

i i

4.3. The Retina 219

by the three different photopigments. If spectra lead to different absorptions, the

light will generally be perceived as different.
In addition to the wavelength of the peak sensitivities, the shape of the ab-
sorbance spectra of the cone pigments bears mention [643, 644, 732]. It has
been found that the shape of the curves becomes identical if plotted on a log-
linear scale, where instead of plotting the log sensitivity against wavelength, the
log sensitivity is plotted against the normalized wavelength, i.e., λmax /λ (Fig-
ure 4.17) [644].
Using a curve fitting procedure, the following expression is obtained, which
models the spectral sensitivity S of a cone pigment as a function of λmax /λ :
     
λmax λmax
S = exp 70 0.880 − (4.10)
λ λ
  
λmax
+ exp 28.5 0.924 −
λ
   −1
λmax
+ exp −14.1 1.104 − + 0.655 .
λ

On a linear scale, the curves for the three peak sensitivities mentioned above are
shown in Figure 4.18. These curves compare well with those of Figure 4.16,
which were obtained through measurement [1092]. They also fit data obtained
through psychophysical experiments on human observers, as well as electro-
physiological experiments on single photoreceptors in monkey, squirrel, and hu-
man retinas [644].

1.0
S cone
S(λ)

M cone
0.8 L cone

0.6

0.4

0.2

0.0
400 500 600 700
λ

Figure 4.18. The spectral sensitivity template plotted against wavelength λ for peak sen-
sitivities of λmax = 440, λmax = 545, and λmax = 565.

i i

i i
 i i

i i

220 4. Human Vision

logLv -0.82 0 1 2 3 4 5 6
A 7.18 6.56 5.18 3.09 1.87 1.31 1.09 1.00

Table 4.2. Pupillary area as a function of luminance log Lv (after [656]).

Upon stimulation, photoreceptors produce a graded potential which in turn

triggers a response in subsequent cell layers (the bipolar cells and amacrine cells).
The dynamic range of the output is limited, as in all neurons [355], and it can
be modeled for brief flashes of light with either a modified Michaelis-Menten
function VMM , or an exponential function Vexp :
j Lvn
VMM (Lv ) = , (4.11)
Lv + σ n
n


Vexp (Lv ) = j 1 − 2−Lv /σ .
n n
(4.12)

Here, Lv is the luminance (photometrically weighted radiance, see Chapter 6),

j is the maximum photocurrent, σ is the semi-saturation constant, and n is an
exponent which normally has a value somewhat below 1 [77, 483, 1169]. In par-
ticular, Boynton and Whitten estimate the exponent to be 0.7 [117].
The semi-saturation constant determines which value of Lv is mapped to half
the maximum output. The photoreceptor output is thus modeled with a function
that is linear for small values and, for large values, asymptotically reaches the
maximum output j. However, photoreceptors, and in particular cones, are able to
adjust their gain, i.e., σ [355].
Equation (4.11) models the non-linearity of the cone response under the as-
sumptions that the pupil is fixed and no bleaching occurs. Taking these effects
into account, the cone response may be modeled by [118]:
j (Lv Ap)n
VMM (Lv ) = , (4.13)
Lv Apn + σ n
where p is the fraction of unbleached opsins derived from (4.8) and A is the rela-
tive pupil area, which is given as function of Lv in Table 4.2. This equation applies
to a steady-state situation where Lv is kept constant.
A recent and more complete computational model of cone response, including
photo-transduction, calcium feedback, and the transduction occurring in the cone
inner segment is described by van Hateren (see Section 4.8) [434].

4.3.4 The Photoreceptor Mosaic

The S-cones are sparse in comparison with the L- and M-cones, and become even
sparser in the fovea. They are absent in a central retinal disk (the foveola) with

i i

i i
 i i

i i

4.3. The Retina 221

Figure 4.19. Photoreceptor mosaic. (From Mark D Fairchild, Color Appearance Models,
2nd edition, Wiley, Chichester, UK, 2005.)

a diameter of 100 μ m (0.34 deg) [220]. This means that even in humans with
normal color vision, the fovea is tritanopic (loss of S-cone function) for small
and brief targets [610, 1248].
With the peak response of L- and M-cones being only 30 nm apart, they ex-
hibit large amounts of overlap in spectral sensitivity, thus contributing more or
less evenly to a luminance signal. The S-cones, on the other hand, have a very dif-
ferent peak sensitivity and are a poor representation for a luminance signal [861].
The central part of the fovea, therefore, appears to be geared towards maintaining
high spatial acuity rather than high contrast color vision [352].
An alternative explanation for the absence of S-cones in the center of the fovea
is possible, however. For instance, it can be argued that this absence of S-cones
in the fovea helps to counteract the effects of Rayleigh scattering in the ocular
media [1036].
The S-cones in the periphery are packed at regular distances in non-random
order [1245]. The packing of L- and M-cones is essentially random. A schematic
of the packing of cones in the retina is given in Figure 4.19.
The ratio between cone types in the human retina varies considerably between
individuals [785, 977]. For instance, the measured ratio of L- and M-cones was
3.79:1 for one individual and 1.15:1 for another [1245]. This difference has only
minor influence on the perception of color [121, 911], which may be due to a

i i

i i
 i i

i i

222 4. Human Vision

combination of various factors. First, the statistics of natural scenes are such that
high frequency, high contrast signals are rare. Further, the scatter occurring in
the ocular media may help to prevent aliasing. Lastly, post-receptoral process-
ing may compensate for the idiosyncrasies of the photoreceptor mosaic [824].
Nonetheless, some variability in photopic spectral sensitivity remains between
healthy individuals, which is largely attributed to the random arrangement of L-
and M-cones and the variability in ratio between these cone types [227, 670].
The ratio of the combined population of L- and M-cones to S-cones is around
100:1 [1205]. If we assume that the L/M ratio is 2:1, then the ratio between all
three cone types is given by L:M:S = 0.66 : 0.33 : 0.01.
Finally, the packing density of cones in the fovea varies significantly be-
tween individuals. Measurements have shown individuals with as few as 98,200
cones/mm2 in the fovea up to around 324,100 cones/mm2 [219]. Outside the
fovea, the variance in receptor density is much lower. It can be argued that the
cone density outside the fovea is just enough to keep the photon flux per cone
constant, given that the retinal illumination falls off with eccentricity. Between
φ = 1◦ and φ = 20◦ degrees of eccentricity, the cone density d(φ ) can be mod-
eled by [1157]
 
φ −2/3
d(φ ) = 50, 000 . (4.14)
300
Finally, the packing density is slightly higher towards the horizontal meridian as
well as towards the nasal side [219].

4.3.5 Bipolar Cells

Bipolar cells take their primary input from the photoreceptors. They are differ-
entiated into different types according to their response to the neurotransmitter
glutamate received from the photoreceptors. For instance, they may be tuned to
faster or slower fluctuations in the visual signal [609]. The neuronal chemistry of
bipolar cells (and all other cells in the brain) determines their function [457].
Some bipolars activate ON pathways, i.e., they detect light against a darker
background. Other bipolars respond to dark spots against a light background
and start an OFF pathway. ON-bipolars and OFF-bipolars work in parallel and
may take their input from the same cones. The parallel processing of ON and
OFF signals are continued later in the visual pathway; bipolar ON and OFF cells
connect to separate ON and OFF ganglion cells.
Having separate ON and OFF pathways implies that the sensation of light and
dark is mediated by separate processes. On the basis of psychophysical experi-
ments this was already postulated over a century ago by Hering [461].

i i

i i
 i i

i i

4.3. The Retina 223

In the fovea, L- and M-cones each connect to a single midget bipolar cell. As
a consequence, midget bipolar cells carry chromatic information. However, the
fovea also contains diffuse bipolar cells that connect to different cone types, and
thus carry achromatic information [1217]. Diffuse bipolar cells have a center-
surround organization [223] (see Section 4.3.6).
The S-cones form the start of a separate ON pathway. Separate blue-cone
bipolar cells connecting to blue cones have been identified [612], which mediate
the S-cone ON pathway. In addition, each S-cone connects to a single midget
bipolar cell [597, 598], which potentially forms the substrate of a blue OFF sig-
nal [227].

4.3.6 Horizontal Cells

In addition to taking input from photoreceptors, bipolar cells are interconnected
laterally by horizontal cells. These cells have very wide receptive fields, i.e., they
connect to a large number of bipolars and photoreceptors. At the same time,
the plasma membranes of horizontal cells interconnect, thereby widening their
receptive field even further.
Horizontal cells connect to all cones within their dendritic field (i.e., receptive
field). As different cone types are likely to be available within a horizontal cell’s
receptive field, this implies that horizontal cells are not necessarily chromatically
selective [1217]. However, two dominant horizontal cell types have been found,
H1 and H2, whereby the former takes input preferentially (but not exclusively)
from L- and M-cones, whereas the latter preferentially connects to S-cones [223].
Horizontal cells provide an inhibitory opponent signal to the bipolar cells,
either through direct connection, or through a proposed feedback mechanism in-
volving photoreceptors [609]. If the area surrounding the receptive field of a
bipolar cell is different from the bipolar’s receptive field itself, the signal is am-
plified further due to the effect of the horizontal cells. Thus, the receptive field of
a bipolar cell is indirectly modified by activity of neighboring photoreceptors and
bipolar cells.
The response to light of bipolars due to their connectivity with photoreceptors
is called the center, whereas the modulation by a wider receptive field due to hori-
zontal cell activity, is called the surround. Typically, an ON-bipolar is augmented
by an OFF-surround, and an OFF-bipolar has an ON-surround—hence the term
opponent signal.
The activity of horizontal cells and, therefore, indirectly photoreceptors and
bipolar cells, is modulated by further feedback from signals originating in the in-
ner plexiform layer. This allows photoreceptors to neuronally adapt to light, in
addition to the non-linear response mechanisms outlined in Section 4.3.3. This

i i

i i
 i i

i i

224 4. Human Vision

feedback mechanism also affords the possibility of color coding the response of
bipolar cells [609]. As a result, the notion that retinal processing of signals is
strictly feed-forward is false, as even the photoreceptors appear to receive feed-
back from later stages of processing.

4.3.7 Ganglion Cells

Ganglion cells form the connectivity between the retina and the rest of the brain.
Their input, taken from both bipolar cells and amacrine cells, is relayed to the
lateral geniculate nucleus through the ganglion’s axons which form the optic nerve
bundle.
The predominant ganglion cell types in the human retina are ON-center and
OFF-center [430]. An ON-center ganglion cell is activated when a spot of light
falls on the center of its receptive field. The opposite happens when light falls onto
the periphery of its receptive field. For OFF-center ganglion cells, the activation
pattern is reversed with activation occurring if light falls on the periphery; illumi-
nation of the center of its receptive field causing inhibition. Thus, the receptive
fields of these ganglion types are concentric, with center-surround organization
(ON-center, OFF-surround and vice versa) [632].
In the fovea, the receptive fields of ganglion cells are much smaller than in
the rest of the retina. Here, midget ganglion cells connect to midget bipolar cells
in a one-to-one correspondence. Each red and green cone in the fovea (but not
blue cone) connects to one ON and one OFF midget bipolar cell, and these are
in turn connected to an ON and OFF ganglion cell, thereby continuing the ON
and OFF pathways. This allows cones in the fovea to transmit dark-on-light and,
separately, light-on-dark information at very high spatial resolution [609].
Ganglion cells respond to colored stimuli in one of two ways: achromatically
or in color-opponent fashion. A particular ganglion cell type, the parasol ganglion
cell, is the substrate for carrying achromatic information, i.e., information about
the sum of the L and M channels.
The classic color opponent pattern arises when a ganglion cell is stimulated by
long wavelength light (red) and inhibited by short wavelength light (green) [1190].
Such red/green color opponency is counter-balanced by the existence of a
green/red color opponent ganglion cell type.2 The information about the differ-
ence between L- and M-cone excitations is carried by midget ganglion cells [223].
However, it should be noted that this color opponency only arises in the fovea
where the size of the center is reduced to a single cone type. In the periphery, the
2 Note that although we call this opponent process “red/green,” this process corresponds in practice

to a color axis that goes from pinkish-red to cyan [209], see Section 4.6.

i i

i i
 i i

i i

4.3. The Retina 225

receptive field size gradually increases with eccentricity, and here midget ganglion
cells receive input from L- and M-cones for both center and surround. This is
consistent with psychophysical evidence that red/green sensitivity declines with
eccentricity [223, 801].
Small-bistratified ganglion cells form a further cell type that carry
blue-ON/yellow-OFF signals [221].3 They are not spatially opponent, in that
their center and surrounds have nearly the same size. The ON-signal is derived
from S-cones (through the blue-cone bipolars), whereas the surround is derived
from L- and M-cones (by means of diffuse bipolar cells). The bistratifed ganglion
cells thus carry an S-(L+M) signal.
Several other ganglion cell types with complex receptive fields exist in the
retina, for instance, ones that are selective for direction and motion [60, 62].

4.3.8 Ganglion Cell Classifications

A further classification of ganglion cells is given by Gouras [382], who subdivides
these cells into tonic and phasic types, which exhibit different responses to light
stimuli. Tonic cells respond to light in a steady manner. They have very small
receptive fields and are color-specific. These cells project to the parvocellular
layer of the Lateral Geniculate Nucleus (LGN; see Section 4.4) and are therefore
part of the parvocellular pathway. Their optimal response is for high-contrast
signals with high spatial resolution.
Tonic cells can have red, green, or blue centers (both ON and OFF), cor-
responding to the three cone types. Their surround is also spectrally selective,
although drawn from a different cone type than the center. Blue OFF-center types
are extremely rare [788].
The phasic ganglion cells have larger receptive fields and respond transiently
to either the onset or the offset of a stimulus. For small spots, the transient dis-
charge is followed by a smaller sustained discharge. Large spots of light only
produce the initial transient. The transient behavior is caused by the inhibiting
surround response, which is delayed with respect to the center response.
Phasic cells are generally achromatic and project to the magnocellular layer
of the LGN. Changes in luminance are detected, but color changes are not [221].
Phasic cells are more sensitive to low contrast signals covering larger areas [826].
There are further color-opponent ganglion cell types. For instance, cells exist
with a color-opponent center, but with no surround. Such cells exist for blue-
yellow color opponency [221].
3 Note that although this dimension is referred to as blue/yellow, it encodes colors that are perceived

as violet/lime [257] (see Section 4.6).

i i

i i
 i i

i i

226 4. Human Vision

Ganglion cells are tuned to spatial frequency, i.e., their sensitivity to spatially
varying patterns is highest for certain spatial frequencies of the pattern. The fre-
quency response of ganglion cells can be characterized by a grating composed of
a sine wave oriented in some direction. The contrast of the sine wave is measured
as the difference between the lightest and darkest parts of the wave. By intro-
ducing such a grating into the receptive field of the ganglion cell, and reducing
the contrast until the cell just barely responds, the threshold of visibility can be
deduced [826].
By repeating this process for sine waves of different spatial frequencies, a
threshold curve can be created. Such curves, known as contrast-sensitivity func-
tions (CSFs), show a peak at a given frequency [284]. This is the frequency to
which the ganglion cell is most responsive.
In addition to peaked-contrast sensitivity, ganglion cells may be characterized by
the sensitivity to motion within their receptive fields (called hyper-acuity). The
smallest detectable displacements are frequently much smaller than might be ex-
pected given the size of the center receptive field, or the peak contrast sensitiv-
ity [1031].
A useful feature of both contrast sensitivity and hyper-acuity measures is that
they apply to ganglion cells as well as human psychophysical performance. This
leads to the conclusion that organisms cannot detect visual stimuli that are not
detected by ganglion cells [826].
Ganglion receptive fields can be reasonably well modeled by the subtraction
of two Gaussian profiles. The center is modeled by a positive Gaussian function
with a smaller radius and larger amplitude than the surround. The Gaussian profile
modeling the surround is subtracted from the center profile. This is the so-called
Difference of Gaussians or DoG model.

4.3.9 Amacrine Cells

Amacrine cells provide lateral connectivity between bipolar cells and ganglion
cells, as well as vertical connections between ON and OFF visual pathways. In
addition, the rod and cone pathways are interconnected through amacrine cells,
as discussed in Section 4.3.10.
There is a large number of different amacrine cell types which appear to be
involved in adaptation processes that correct for changes in the environment. For
instance, the spectral composition of daylight during a sunrise and a sunset is very
different from mid-day. To feed the brain with an optimal signal during changing
light conditions requires a light-driven modulatory system that is perhaps medi-
ated by amacrine cells.

i i

i i
 i i

i i

4.3. The Retina 227

As an example, feedback is provided to the horizontal cells. For instance, a

specific amacrine cell releases dopamine when stimulated with a flashing light.
This neurotransmitter in turn decouples horizontal cells, thereby effectively re-
ducing their receptive field size [609].
Another example is given by the somatostatin-immunoreactive amacrine cells
[961, 988] which have cell bodies and dendrites largely located in the lower part
of the retina. Their axons project into the upper-half of the retina. Under normal
viewing conditions light from the sky is projected predominantly onto the lower-
half of the retina, whereas reflections off the ground are projected onto the upper-
half of the retina. Somatostatin-immunoreactive amacrine cells may therefore be
involved in adaptation to the tremendous variation in light received from different
parts of the visual field [1217].

4.3.10 The Rod Visual Pathway

Rods project to just one type of bipolar cell—the rod bipolar. Large numbers of
rods project to a single rod bipolar cell. These bipolars do not directly connect to
ganglion cells, but synapse into amacrine cells instead [605, 607], thus diverging
the signal. There are two types of amacrine cells dominant in the rod visual
pathway, with designated names AII and A17.
The AII amacrines link the rod and cone pathways, such that the rod signals
can also use the cone bipolars to reach the ganglion cells. This type of amacrine
is mostly post-synaptic to rod bipolar cells and synapses into ganglion cells. They
are small-field, which means that their lateral spread is limited. Nonetheless, AII
amacrines are also directly connected to each other, much like horizontal cells are
interconnected. This widens their area of influence. In bright light conditions,
when the narrow-field cone pathways need to maintain their spatial resolution, a
different type of amacrine cell releases a neurotransmitter that uncouples the AII
amacrines, thereby narrowing their receptive field and minimizing the interference
of the rod pathway.
On the other hand, the A17 amacrine cells are wide-field and typically re-
ceive input from approximately 1000 bipolar cells, making them very sensitive to
light. The A17 does not synapse into other amacrine or ganglion cells. Its large
receptive field appears to be used to integrate the signal, helping to set sensitivity
levels [825].
The significant convergence of rods into bipolars and the divergence of the
signal through amacrine cells is thought to increase light sensitivity of the rod
pathway. It achieves this very high sensitivity at the cost of spatial resolution,
which is reduced by the pooling of signals.

i i

i i
 i i

i i

228 4. Human Vision

Finally, in the mesopic range, rod vision has a (relatively small) effect on color
vision [599, 1077, 1078].

4.4 The Lateral Geniculate Nucleus

The signal leaving the retina projects to several parts of the brain, including the
superior colliculus and the lateral geniculate nucleus.4 Of interest to color vision
are the pathways that go through the lateral geniculate nuclei (LGN). These are
located on either side of the thalamus and relay the signal from the eye to the
primary visual cortex.
The optic nerve leaving each eye is split, with the nasal part crossing over
(Figure 4.20). This means that each LGN receives input from both eyes. The left
LGN is responsible for processing the visual input of the left half of each eye,
and the right LGN processes the right half of the visual field. The part where the
nerve bundles cross over is called the optic chiasm.
There exist three pathways that project from the retina into each LGN and
from there into the cortex. These pathways are collectively known as the retino-
geniculo-cortical pathway. Each LGN contains several layers, and each layer
receives input from one eye only. In addition, each layer is classified by cell
type and receives input from predominantly one ganglion cell type. Thus, signals
that are processed separately in the retina remain separate in the LGN
(Figure 4.21). There are three types of layers, with the third mediating output
from the blue cones being a relatively recent discovery [157, 257, 458].

Fibers from the

temporal retina
(nasal visual field)
do not cross

Fibers from the

nasal retina (temporal Optic chiasm
visual field) cross to
the contralateral side

Figure 4.20. The optic chiasm.

4 The term geniculate refers to its structure and means knee-shaped.

i i

i i
 i i

i i

4.4. The Lateral Geniculate Nucleus 229

Diagram of the left Lateral Geniculate Nucleus

Parvocellular (6)
6
Parvocellular (5) Striate cortex V1
5
4Cβ / 6
Parvocellular (4)
Right eye 4
CO blobs / 4a
Parvocellular (3)
3 4Cα / 6

Magnocellular (2)
Left eye
2
Ipsilateral (same side) connections Magnocellular (1)
Contralateral (opposite side) connections 1
Koniocellular layers

Figure 4.21. A diagram of the LGN showing both its main inputs and main outputs, which
project to the striate cortex.

The three types of layers are [257]5 :

Magnocellular layers. These layers receive input from parasol ganglion cells and
are therefore responsible for carrying the achromatic L+M signal [227,
660]. These cells have a spatial center-surround configuration, where ON-
center cells receive excitatory input from both L- and M-cones in the center
and inhibitory input from both L- and M-cones in the surround. OFF-center
cells have opposite polarity [352].

Parvocellular layers. Receiving input from midget ganglion cells, these layers
are responsible for carrying the L-M signal, i.e., they mediate red-green
color opponency [673,891]. Wiesel and Hubel first proposed the possibility
of neural circuitry whereby the center of the receptive field was excited by
one cone type, and the surround by another [1244]. Evidence for such
organization has been confirmed, for instance, in the red-green opponent
ganglion cells [661] and the LGN [746, 946, 947]. However, it should be
noted that the dendritic field sizes of midget ganglion cells in the parafovea
increase in size, and as a result both their centers and surrounds take input
5 Note that the following description of these three layers in the LGN is simplified and not necessar-

ily universally accepted. However, for the purpose of this book, the current description is sufficiently
complete.

i i

i i
 i i

i i

230 4. Human Vision

from both L- and M-cones. As a result, red-green color opponency is not

maintained in the parafovea [222, 223].
The absence of S-cone input to parvocellular layers was shown by com-
paring the spectral cone response data [1091] with a Principal Components
Analysis of measurements of cells in this layer [976] using data obtained
by de Valois and colleagues [1172].

Koniocellular layers. These layers are also sometimes called intercalated [322,
406], although the preferred naming is koniocellular [564]. There are three
pairs of layers of this type. The dorsal most pair transmits low-acuity infor-
mation to the primary visual cortex (also known as area V1). The middle
pair projects the signal that originates in bistratified ganglion cells and car-
ries the blue yellow S-(L+M) signal to the cytochrome oxydase blobs of
the primary visual cortex [257, 458]. The ventral-most pair is related to
the superior colliculus. In addition to these dominant pathways, konio-
cellular neurons project to parts of the extrastriate cortex, bypassing V1
altogether [458, 973, 1052, 1286].

4.5 The Visual Cortex

The visual cortex is the primary location for processing visual signals in the
brain. It contains a large number of separate functional units that are intercon-
nected through both feed-forward and feedback connections. These cortical areas
include:

V1. This is the principle part of the visual cortex that receives input from the
LGN. In an older classification scheme, this area is also known as Brod-
mann area 17 [138, 710]. It is also known as the primary visual cortex or
the striate cortex.

V2. Area V2, after staining with cytochrome oxydase, reveals thin stripes, thick
stripes, and pale stripes, each receiving input from different areas in V1.

V3. This is an area thought to be involved in the integration of signals from dif-
ferent pathways.

V4. This area is thought to play an important role in the processing of color in-
formation.

i i

i i
 i i

i i

4.5. The Visual Cortex 231

Projections dominated by parvo signals

Projections dominated by magno signals
Mixed projections MT

V1

V2 V3

V4
Parvo Magno
pathway pathway

Figure 4.22. The main feed-forward projections between several cortical areas (af-
ter [592]).

MT. Area MT is also known as area V5, or the middle temporal area. Cells in
this area tend to be strongly sensitive to motion. It predominantly receives
input from the geniculate magno system.

Many other areas exist, but most is known about areas V1, V2, and MT [563].
The main feed-forward projections between these areas are diagrammed in Fig-
ure 4.22. Color-selective cells have been found in several areas, including V1,
V2, V3, and in particular V4 [206,624,705,1113,1289,1290]. However, it should
be noted that the localization of V4 as a center for color processing is still un-
der debate, as some studies suggest a different organization [282, 407, 1139].
In the following we briefly discuss some of the cortical areas relevant to color
processing.

4.5.1 Area V1
Area V1 is the main entry point to the visual cortex for signals from the lateral
geniculate cortex. It consists of six layers, with some layers divided into sub-
lamina [138, 710]. In addition, several layers, including layers two and three,
have columnar structures called blobs or patches. The areas in between are called
inter-blobs or inter-patches. The separation between patches and inter-patches is
based on a histochemical technique called cytochrome oxydase (CO) that differ-
entiates cells based on metabolism, revealing structure in layers two and three of
the primary visual cortex.
One of the challenges still remaining is to determine what the function might
be of cells located in different layers. This problem may be interpreted as un-

i i

i i
 i i

i i

232 4. Human Vision

raveling the functional architecture, i.e., determining if there is a tendency for

neighboring cells to share one or more receptive-field characteristics [493].
It is thought that the columnar structure relates to functional structure, with
proposals existing for several different types of function, including ocular dom-
inance, orientation, spatial frequency, and color. For instance, many cells in the
primary visual cortex are orientation-sensitive. Such orientation-selective cells re-
spond best to a receptive field that is elongated, rather than circularly symmetric,
and that is oriented in a particular direction. It is found that this type of selectivity
is prominent in layer 4B, but exists in other layers as well [408, 438, 964].
It was long thought that there exists a segregation in functional pathways from
the retina all the way to the primary visual cortex and beyond, separating the
processing of color, form, and motion/stereo [287, 694]. In this tripartite model,
area V1 transforms the three streams which are then projected to area V2 [692–
694].
The optic radiations originating in the lateral geniculate nucleus synapse into
different layers and into the patches of layers two and three, as diagrammed in the
left part of Figure 4.23. Here, it is seen that different layers in the LGN project
to different parts of area V1. This appears to provide evidence that the different
pathways originating in the retina remain functionally segregated in V1.
However, the intra-cortical circuitry, i.e., the interconnectivity within area V1
after signals from the LGN have been projected, suggests that the pathways do

LGN inputs First intracortical Second intracortical

synapse synapse
1 1
2 2
Primary visual cortex
(V1, striate cortex)

3 3
4A 4A
4B 4B
4Cα 4Cα
4Cβ 4Cβ
5 5
6 6

Konio Parvo Magno Cytochrome oxydase (CO) patch

Parvo Magno +
Parvo

Figure 4.23. Area V1 (primary visual cortex) and its intracortical circuitry (after [1051]).

i i

i i
 i i

i i

4.5. The Visual Cortex 233

not remain separate, but are combined instead [148, 636, 1051, 1265, 1281]. A
summary of intra-cortical connectivity is shown in Figure 4.23. This finding was
corroborated by studies that revealed that orientation-selective cells can be tuned
to color-selective as well [209, 491, 669, 674, 1127, 1181, 1189].
A further complicating factor is that area V1 has both feed-forward and feed-
back connections from many other functional areas in the brain [1008]:

• Feed-forward projections from other areas to V1 originate in: pulvinar,

LGN, claustrum, nucleus paracentralis, raphe system, locus coeruleus, and
the nucleus basalis of Meynert [6, 104, 260, 384, 459, 635, 844, 889, 957].

• Feedback projections from other areas to V1 originate in: V2, V3, V4, MT,
MST, frontal eye field (FEF), LIP, and the inferotemporal cortex [69, 889,
970, 1041, 1108, 1164, 1165].

• Area V1 projects forward to V2, V3, MT, MST, and FEF [115, 323, 693,
709, 753, 1041, 1164, 1165].

• Area V1 extends backward projections to SC, LGN, pulvinar, and pons

[333, 334, 384, 409, 709].

For the purpose of modeling the micro-circuitry of area V1, as well as understand-
ing the function of V1, these feedback mechanisms are likely to play roles that
are at least as important as the geniculate input [747].
While the color processing in the LGN occurs along dark-light, red-green, and
yellow-blue axes, this is not so in area V1. LGN projections are recombined in
both linear and non-linear ways, resulting in color sensitivity that is more sharply
tuned to specific ranges of color, not necessarily aligned with the color directions
seen in the LGN [209, 419, 1175]. In particular, clusters of cells have been iden-
tified that respond to purple [1189]. Psychophysical studies have confirmed the
likelihood of a color transform from the LGN to the striate cortex [1174, 1176].
There is a possibility that these non-linear recombinations underlie the human
perception of color categories. It is also possible that separate color transfor-
mations in V1 contribute to the perception of color boundaries and colored re-
gions [1030]. A proposed three-stage color model, accounting for phenomena
measured through neurophysiological experiment as well as psychophysics, is
discussed in Section 4.6.
In summary, many cells in area V1 respond to multiple stimuli, such as color,
position, orientation, and direction. The notion that pathways for the processing
of color, shape, and the like remain separate is unlikely to be true. In addition,
area V1 is one of many areas where color processing takes place, thus moving

i i

i i
 i i

i i

234 4. Human Vision

gradually from the color-opponent mechanisms found in the LGN to mechanisms

that are more closely aligned with color perception.

4.5.2 Area V2
Area V2 is characterized by thick, thin, and pale stripes that are revealed under
cytochrome oxydase staining. There is some degree of functional segregation
along these histochemically defined zones, although many cells are found that are
selective for multiple stimulus attributes [592]. The thick and pale stripes have
a higher degree of orientation selectivity than the thin stripes [726]. The thin
stripes have a tendency towards color selectivity and are found to be sensitive to
wavelength composition [800]. Nonetheless, the precise nature of functional seg-
regation in area V2 is still under dispute [1051] and significant cross-talk between
the processing of various stimulus attributes exists.
The pathway between V1 and V2 is currently known to be bipartite [1050],
with patches in V1 connecting to thin stripes and inter-patches connecting to both
thick and pale stripes in V2 [1051]. These projections are well segregated [490].
There is also significant feedback from V2 to V1 [971], although the organi-
zation of these feedback projections is only beginning to emerge [38, 726, 1044,
1251]. The effect of removing these feedback connections produces only subtle
changes in the responses of V1 neurons [511, 993]. It is possible that these feed-
back projections alter the spatial extent of surround inhibitory circuits [1033],
increase the saliency of stimuli [510], or possibly aid in disambiguation [854].

4.5.3 Area V3
Current techniques have not revealed anatomically or functionally distinct com-
partments within area V3 [592], although a columnar organization has been re-
ported [4]. In addition, little is known about the physiology of neurons in this
area [308]. V3 may play a role in the integration of different stimulus attributes,
aiding in visual perception. It has also been suggested that this area is involved in
the analysis of three-dimensional form [4].
The selectivity of neurons in V3 for different stimulus attributes is similar to
those in V2, except that a higher sensitivity to direction was found (40% of the
cells in V3 versus 20% of the cells in V2 [592]). Orientation selectivity is found
in around 85% of the cells, selectivity for size is found in approximately 25%,
and color sensitivity in approximately 50% of the cells in both V2 and V3 [592].
There is no correlation between selectivity for different attributes, suggesting that
different attributes are not processed along separate pathways in V3 (or in V2).

i i

i i
 i i

i i

4.5. The Visual Cortex 235

4.5.4 Area V4
Area V4 has been the subject of significant controversy. It has been implicated
as a color center on the basis of a rare condition called achromatopsia. This is a
deficiency that causes the world to appear colorless and is also known as cerebral
color blindness. This deficiency has been related to lesions in a particular part of
the brain, known as V4. As a result, this area may be involved in computations re-
lating to the perception of color [71, 230, 887, 1289]. It should be noted, however,
that this viewpoint is challenged by several studies [282, 407, 1139] and, as such,
the question of whether V4 can be deemed a color center remains open [392].

4.5.5 Area MT
Area MT is known to be involved in the perception of motion. It responds more
strongly to moving versus stationary stimuli. Direction-selective populations of
neurons have been found in both humans [498,1138] and macaque monkeys [135,
754, 991]. This area is predominantly sensitive to achromatic stimuli.

4.5.6 Retinotopic Maps

The optics of the eye project an image onto the retina, where neighboring pho-
toreceptors image neighboring portions of the retinal image. The signal that is
transmitted towards the visual cortex generally maintains this topography. In the
LGN, the spatial localization of image features is preserved, and the mapping
of the retinal image to neurons in this area is therefore topographic. The term
retinotopic is reserved for this mapping.
In the primary visual cortex, this mapping is preserved, although the fovea
is represented in a relatively large portion of area V1 and beyond. Although
neighboring areas in the retinal image are still represented by neighboring clus-
ters of neurons, in V1 the mapping has undergone a log-polar transformation.
This means that along one dimension, the log distance from the fovea is mapped,
and the other dimension represents the polar angle with respect to the (arbitrarily
chosen) horizontal axis of the retinal image (Figure 4.24). Retinotopic mapping
exists in area V1 as well as areas beyond V1, although further away from V1 the
precision of this organization declines [393, 394, 432, 677].
The logarithmic relationship between retinal distance to the fovea and distance
measured along the surface of the primary visual cortex results in a large area of
the cortex being devoted to processing signals stemming from the fovea. This is
known as cortical magnification [268, 1017].

i i

i i
 i i

i i

236 4. Human Vision

Retina V1
y r

Fovea x θ

Figure 4.24. The retinal image is laid out in the cortex according to a log-polar mapping
(after [392]).

4.5.7 Summary
As the simple model of segregated pathways through the modules in the visual
cortex is being refined, it is too early to attribute specific functionality to separate
regions in the visual cortex. The exact processing that takes place is yet to be
unraveled, and it is therefore not appropriate to assign a substrate to the higher
levels of functioning of the human visual system. This is the case for the primary
visual cortex, as well as for processing elsewhere in the visual cortex.
Dependent on the techniques used to measure functional aspects of the vi-
sual cortex, it is possible to arrive at different conclusions. For instance, there are
those that argue in favor of segregated pathways for color, motion, form, etc. Oth-
ers have argued in favor of a distributed processing of features, where functional
areas in the visual cortex are sensitive to combinations of motion, form, direction,
orientation, color, and the like.
On perceptual grounds, it is also not possible to arrive at a single conclusion.
While some perceptual attributes appear to be dissociated from others, this is not
always strictly true [1202]. For instance, the apparent velocity of an object is
related to its contrast and color [159], and high spatial and temporal frequencies
appear desaturated [908].
One difficulty for determining the functional architecture of the visual cor-
tex is due to the limitations of the currently available methodologies. Most use
point methods, i.e., record responses from single neurons by means of single-
cell recordings or tracer micro-injections. Given the number of neurons present,
and their vast interconnection network, the problems involved are obvious [494].
Nonetheless, there is promise for the future, in that new techniques are becoming
available that allow large populations of cells to be imaged at cell resolution [846].
In the following chapter, we therefore turn our attention to an altogether dif-
ferent strategy to learn about the (color) processing that takes place in the human
visual system, namely psychophysics. Here, the human visual system is treated

i i

i i
 i i

i i

4.6. A Multi-Stage Color Model 237

as a black box, which reveals some of its functionality through measuring task
response.

4.6 A Multi-Stage Color Model

The wavelengths making up the visible spectrum can be arranged on a circle. The
0◦ , 90◦ , 180◦ and 270◦ angles are chosen to coincide with stimuli that isolate
responses from the two types of opponent neuron in the LGN [209], as shown in
Table 4.3.

Angle LGN axis Perceptual color

0◦ L-M “pinkish-red”
90◦ S-(L+M) “violet”
180◦ M-L “cyan”
270◦ -S+(L+M) “greenish-yellow”

Table 4.3. Major axes in MBDKL color space.

The relative strength of each color is adjusted such that all colors appear
equally bright. In particular, the contrasts along the L-M axis are 8% for the
L-cones and 16% for the M-cones, whereas the S-cone contrast is 83% [1175].
Contrasts for intermediate angles are interpolated. The resulting colors are called
iso-luminant. This configuration is known as the MBDKL color space, named
after Macleod, Boynton, Derrington, Krauskopf, and Lennie [250, 722]. Compu-
tational details of this color space are given in Section 8.6.3.
Although its axes are commonly referred to as red-green and yellow-blue, the
actual colors they encode are as given above. The axes of the color space used by
the LGN therefore do not coincide with the perceptual dimensions of red, green,
blue, and yellow, which are rotated in the MBDKL color space. It was postulated
by Russell and Karen de Valois that the red, green, yellow, and blue perceptual
color systems are formed from the LGN output by adding or subtracting the S
signal from the L and M opponent cells [1171].
The multi-stage color model proposed by Russell and Karen de Valois is
composed of four separate stages. They will be discussed in the following sec-
tions [1171]:

1. Three cone types with three to five pigments;

2. Cone opponency;

i i

i i
 i i

i i

238 4. Human Vision

3. Perceptual opponency;

4. Color selectivity in complex cells.

4.6.1 First Stage

The three-cone model is currently widely accepted, each with a single cone pig-
ment. However, it is possible that there exist two variants of the L-cone pigment,
with a peak sensitivity spaced a few nanometers apart, and possibly a similar pair
of M-cone pigments [823]. The ratio of L:M:S cones is taken to be 10:5:1. The
S-cones are assumed to be regularly spaced, whereas the L- and M-cones are
clustered together.

4.6.2 Second Stage

The second stage is postulated to occur in the retina and is involved with the
formation of spectrally opponent and non-opponent signals. The horizontal cells
in this model sum their inputs, which are taken from several cones in a spatial
region, calculating a L+M+S signal. At the bipolar cell level, the midget bipolars
compute L - (L+M+S), M - (L+M+S), and S - (L+M+S) signals. This notation
indicates that the center is formed by either L, M, or S, and the spatial surround
is (L+M+S). For brevity, these signals are denoted Mo , Lo and So .
In addition, the opposite signals are also computed: -L + (L+M+S), -M +
(L+M+S), and -S + (L+M+S), assuming that the spatial surround is formed by
indiscriminately summing over all L-, M-, and S-cones within a local neighbor-
hood.6 These signals are the start of the parvocellular pathway. These signals are
also written as -Lo , -Mo , and -So .
The center and the surround have more or less the same strength. Taking into
account the cone ratios, for Lo , the model thus weights 16L for the center against
a surround of 10L + 5M + S:

Lo = 16L - (10L + 5M + S) (4.15a)

= 6L − 5M − S. (4.15b)

In the discrete model, diffuse bipolars are weighted as 6L - 5M. In either case, the
result is close to a conventional “red-green” color opponent cell. The same is true
for Mo and So , which end up being weighted as 11M - 10L - S and 15S - 10L -
5M, respectively.
6 An alternative model is also presented, whereby the receptive fields are modeled as L-M, -L+M,

M-L, and -M+L. This model assumes that the surround consists of only L and M inputs, and is there-
fore termed “discrete.”

i i

i i
 i i

i i

4.6. A Multi-Stage Color Model 239

The diffuse bipolars appear to receive input from a small number of L- and
M-cones, as well as a surround input derived from horizontal cells. The signals
computed by diffuse bipolar cells is therefore designated L+M - (L+M+S) and
-L-M + (L+M+S), forming the start of the magnocellular pathway.

4.6.3 Third Stage

In the third stage, perceptual opponency is modeled. The output of the diffuse
bipolars, through parasol ganglion cells and the magnocellular layer of the LGN,
is assumed to continue separately and to form the photopic luminosity function
(see Section 6.1).
The six parvocellular LGN types, Lo , Mo , So , -Lo , -Mo , and -So undergo
further transformations as the third stage of the multi-stage color model. LGN
cells respond to both luminance and color variations, but with different receptive
fields [1170]. For instance, an Lo cell has an excitatory center and inhibitory
surround with respect to an increase in luminance. The reason for this behavior is
that both L- and M-cones respond similarly to changes in luminance, thus driving
the center and surround in opposite directions.
However, at the same time, this cell has a uniformly excitatory response to an
equiluminant change in color if the shift is towards longer wavelengths. Here, the
L- and M-cones are driven in opposite directions, yielding a synergistic response
at an Lo cell. Cells at the second stage therefore confound luminance and chro-
matic information. Similar arguments can be made for the five other cell types in
the LGN. An -Mo , for instance, has an inhibitory center and excitatory surround
for luminance changes, but a uniform excitatory response to a chromatic change
towards longer wavelengths.
It is, however, possible to separate color and luminance information by recom-
bining the outputs of second stage cells in specific ways. For instance, summing
Lo and -Mo at the third stage results in cancellation of sensitivity to luminance,
yielding only a response to chromatic changes. Summing Lo and -Mo , on the
other hand, leads to canceling of chromatic responsivity and extracts the lumi-
nance. Hence, Lo - Mo and Mo - Lo form chromatic signals, sensitive to changes
from red to green, and from green to red. Similarly, Lo + Mo and Mo + Lo are
spatially opponent luminance signals.
In this model, the third stage is dominated by Lo and Mo signals, contributing
to both the perceptual red-green and yellow-blue systems. It is thus postulated,
for instance, that the M-cones are the main contributors to the blue signal. The
So signal is then used to modulate this dominant L/M pathway to retrieve the per-
ceptual red, green, yellow, and blue color axes. In particular, these color systems

i i

i i
 i i

i i

240 4. Human Vision

Signal 1 Signal 2 Sum Result

L o - Mo +So L o - Mo + S o Red
Lo - Mo -So Lo - Mo - So Yellow
Mo - Lo +So M o - Lo + S o Blue
Mo - Lo -So Mo - Lo - So Green
Lo + Mo +So L o + M o + So Light
-Lo - Mo -So -Lo - Mo - So Dark

Table 4.4. Recombination of Lo , Mo , and So signals.

can be computed by adding or subtracting So from the Lo - Mo and Mo - Lo cells.

The four possible combinations are listed in Table 4.4
It is assumed that the contribution from So is given different weights for the
four color systems. To match hue scaling experiments [1174], the So is given a
weight of 1.0 for the perceptual red system, 2.0 for green, 2.0 for blue, and 3.5 for
the yellow system. The contributions of the three cone types to each of these four
perceptual color axes is given in Table 4.5.

Perceptual color axes Stage-3 weights Cone contributions

Red 10Lo - 5Mo + So 100L - 110M + 10S
Green 10Lo - 5Mo + 2So 90L - 115M + 25S
Yellow -10Lo + 5Mo + 3.5So -145L + 87.5M + 57.5S
Blue -10Lo + 5Mo + 2So -130L + 95M + 35S

Table 4.5. Cone contributions to each of the four perceptual color axes. The weights given
to So are 2 for all four axes in the original model, but have been adjusted here to account
for more recent hue-scaling results [1174].

4.6.4 Fourth Stage

It has been shown that most simple cells produce a half-wave rectified output. The
fourth stage produces such an output. Complex cells, located beyond V1 can then
be modeled by summing together two half-wave rectified simple-cell responses.

4.6.5 Neurophysiological Experiments

Using the MBDKL color space as a basis, neurophysiological and psychophys-
ical experiments have been carried out that support this view [209, 1174–1176].
The color responses measured in parvocellular LGN cells of macaque monkeys
are shown in Figure 4.25. The LGN responses as function of color angle are well

i i

i i
 i i

i i

4.6. A Multi-Stage Color Model 241

Spikes/s (normalized)
100
+S - (L+M)
+L - M
50 -L + M

-50

-100
0 90 180 270 360
MBDKL Color Angle

Figure 4.25. Sinusoidal fit to color responses of various LGN cells (after [1175]).

modeled by sinusoidals, indicating that the input transmitted from the photore-
ceptors is linearly recombined in the LGN.
The optimal tuning of LM opponent cells is along the 0◦ -180◦ axis in MBDKL
space. The M center cells (+M-L and -L+M) on average have their maximum
response at 180◦ and 0◦ . The L center cells (+L-M and -L+M) peak slightly off-
axis at approximately 345◦ and 165◦ . The +S-(L+M) cells peak at 90%. The LGN
cells, in general, respond strongly to colors along the 0◦ -180◦ and 90◦ -270◦ axes
and show little response to other colors, as shown in Figure 4.26.
This figure also shows the color tuning of a set of cortical cells. Cortical
cells are tuned to a variety of different colors, albeit with dips around the 0◦ -180◦
and 90◦ -270◦ axes [1175]. Other studies have found a similarly wide variety of
tunings in V1 [669, 1127], although other experiments have shown cell types that
respond discretely to red, green, yellow, and blue [1147, 1180].

30% LGN cells

Cortical cells

20%

10%

0%
0 90 120 150 180 210 240
MBDKL Color Angle

Figure 4.26. The tuning distribution of 100 LGN cells and 314 striate cortex cells. Oppo-
site phases are combined, i.e., 0◦ and 180◦ are both mapped to 180◦ (after [1175]).

i i

i i
 i i

i i

242 4. Human Vision

60

Cells
LGN cells
50
Cortical cells
40
30
20
10
0
0.1 1 10 100
n

Figure 4.27. The distribution of exponents for cells in the LGN and in the striate cortex.
The median values for LGN and V1 cells are 1.08 and 1.90 (after [1175]).

While the LGN cells respond linearly to their inputs, V 1 cells exhibit a range
of non-linearities. Such non-linearities can be modeled with an exponentiated
sine function in the MBDKL space:

R = A sinn (C − φ ) . (4.16)

Here, the response R is modeled as a function of gain A, the color axis C under
investigation, and phase φ . The exponent n models the degree of non-linearity.
For a group of LGN and V1 cells, the distribution of exponents that best fits the
individual cells’ responses are plotted (Figure 4.27). The median value found for
V1 cells is 1.90, indicating that nearly one half of the cells exhibit significant
non-linear behavior.

1.0
sin (x)

n=1
n

0.8 n = 1.9
n = 10
0.6

0.4

0.2

0.0
0 40 80 120 160
x

Figure 4.28. A sine function exponentiated with different values of n. As can be seen
here, larger exponents cause the function to be more peaked, i.e., tuned to a narrower set
of wavelengths.

i i

i i
 i i

i i

4.6. A Multi-Stage Color Model 243

The interpretation of an exponent n that is larger than 1 is that such responses

are more sharply tuned, i.e., the responses are limited to a smaller range of wave-
lengths. This can be seen in Figure 4.28 which plots sine functions exponentiated
with different values of n.

4.6.6 Psychophysical Experiments

A further experiment, this time psychophysical, compared responses for large
stimuli (subtending 2◦ of visual angle) against small stimuli (0.1◦ of visual angle)
both presented to the foveal region [1176]. As S-cones are absent in the fovea,
the small stimuli activate only L- and M-cones, whereas the large stimuli activate
all three cone types. The result of the 2◦ color-scaling experiment is reproduced
in Figure 4.29.
Once more, the peaks in the graphs for red, green, blue, and yellow color
regions do not coincide with the LGN cell peaks, but are rotated. Second, the
shape of the graphs is non-sinusoidal, but the regions are better modeled with an
exponentiated sine curve (4.16). The exponents found for red, green, blue, and
yellow are, respectively, 1, 1.7, 1.8, and 2.9.
For small stimuli presented to the fovea, the S-cones will contribute little or
nothing to the perception of color [1176]. As a result, stimuli around the 90◦
and 270◦ angles are frequently not noticed. The color scaling results for small
stimuli are reproduced in Figure 4.30. Here, stimuli in the 90◦ to 270◦ range were
perceived almost equally as green and blue. The range between 270◦ and 450◦
was seen as red with a small yellow component added in.
The absence of input from the S-cones has therefore resulted in only two re-
gions with more or less uniform color appearance. The contribution of the S
system is summarized in Table 4.6. It can be seen that the S-cone signals make
Spikes/s (normalized)

100

80 Blue Green Yellow Red

60

40

20

0
0 90 180 270 360 450
MBDKL Color Angle

Figure 4.29. Color-scaling results for 2◦ stimuli (after [1176]).

i i

i i
 i i

i i

244 4. Human Vision

Spikes/s (normalized)
100
Blue
80 Green
Yellow
60 Red

40

20

0
0 90 180 270 360 450
MBDKL Color Angle

Figure 4.30. Color scaling results for 0.1◦ stimuli. Near-zero responses, as well as data
around 90◦ and 270◦ , are not shown (after [1176]).

MBDKL angle Perceptual color S-cone contribution

90◦ - 180◦ Green Decrease
90◦ - 180◦ Blue Increase
180◦ - 270◦ Green Increase
180◦ - 270◦ Blue Decrease
0◦ - 90◦ Red Slight increase
270◦ - 360◦ Red Slight decrease
0◦ - 90◦ Yellow Decrease
270◦ - 360◦ Yellow Increase

Table 4.6. S-cone contribution to color perception.

significant contributions to the perception of both green and blue. This is consis-
tent with the possibility that in V1, S signals are added to the +M-L cells to form
blue and subtracted from the -M+L cells to form green. Similarly, these findings
are in accord with a model which adds S-cone signals to +L-M signals to form
red and subtracts them from -L+M signals to create yellow.
Moreover, the responses in the absence of S-cone input are not peaked. This
suggests that the sharpening of responses in V1 is due to the contribution the
S-cone pathway.

4.6.7 Discussion
The magnocellular pathway is traditionally seen as predominantly responsible
for carrying the luminance signal. In the multi-stage model, this pathway car-
ries essentially luminance information akin to the V (λ ) curve (introduced in Sec-
tion 6.1). However, there also exists evidence that the parvocellular pathway may
carry achromatic information which is multiplexed with color information [1170,

i i

i i
 i i

i i

4.7. Alternative Theory of Color Vision 245

1173]. The multi-stage color model is able to model this and then separate the
chromatic and achromatic information at the cortical stage.
In this model, the L- and M-cones provide the dominant input to both the red-
green and yellow-blue systems. The S-cones modulate this dominant pathway in
the third stage to construct perceptual color axes. It suggests that the dominant
input to the blue system is formed by the M-cones, not the S-cones. This feature
is in accord with measurements of tritanopes (see Section 5.11) who are missing
or have non-functional S-cones. They typically see all wavelengths below 570
nm as blue or partially blue. However, S-cones do not respond to wavelengths
above around 520 nm. The gap between 520 and 570 nm where blue is detected
cannot be explained by conventional color models, but can be explained with the
multi-stage model.
Finally, the rotation to perceptual color axes is in accord with both hue-cancel-
lation experiments [515], as well as cross-cultural color-naming research [87,
449].

4.7 Alternative Theory of Color Vision

This chapter so far has focused on generally accepted theories of color vision.
However, recently an intriguing alternative theory has emerged which can be
viewed as complementary to known facts about human vision. This theory deals
with the transduction of photons into electrical signals in the photoreceptors.
It can be argued that there are several unusual aspects in current theories and
measurements. First, the absence of blue cones in the fovea does not appear to
impede color perception in the foveal region. Second, the seemingly random dis-
tribution of L- and M-cones does not appear to have a marked effect on color
vision. This is odd, if it is assumed that each of the three cone types is character-
ized by the presence of a single type of photopigment.
Moreover, recent studies have for the first time been able to probe individual
cones by correcting the light scatter in the ocular media [476]. By stimulating
individual cones with monochromatic light at wavelengths of 600, 550, and 500
nm, the expectation is that dependent on the cone type, the corresponding sensa-
tion of red, green, or blue will be elicited. However, this did not turn out to be
the case, and it was concluded that the diversity in cone response could not be
completely explained by L- and M-cone excitation. Instead, single cones appear
to be capable of producing more than one sensation.
A further counter-intuitive property of color vision that cannot be explained
by the three-receptor model was first reported by Tyndall [1158] and discussed

i i

i i
 i i

i i

246 4. Human Vision

by Mollon and Estévez [786]. It was observed that wavelength discrimination at

around 455 nm improves when a white desaturating light is added. For a three-
cone mechanism, any wavelength discrimination must be reduced if white light is
added to the stimulus, and therefore Tyndall’s discovery is not well-explained by
current theories of color vision.
There are several other anomalies in current models of color vision [771].
Rather than listing them all, we mention one further remarkable observation. It
is possible to mix red and green light to form yellow light. It is also possible to
construct monochromatic yellow light. These two yellows are perceived the same,
i.e., they are metameric (see Section 7.2.1). However, when a vertical yellow bar
is moved horizontally across the visual field, the monochromatic yellow remains
yellow, whereas the yellow color mixture is broken up into a red leading edge, a
green trailing edge, and a yellow middle region. This phenomenon was noted first
by Ives in 1918 [539].
In a three-cone model, the explanation of this result is problematic, even if
different time constants are assumed for the different cone types. In particular, it
would be expected that the moving color mixture can be broken up into red and
green edges. However, it would be expected that the monochromatic yellow bar
would also break up into a red leading edge and a green trailing edge, on the basis
that the quantal catches for each cone type should be identical for both stimuli.
This is contrary to the experimental result.
Although the three-cone model is able to explain many visual phenomena and
is a useful starting point for most applications in engineering-related disciplines,
there are sufficiently many observations which cannot be explained by this model.
There is therefore room for alternate hypotheses which may explain the same
phenomena, but in addition explain some of the cases where the current model
breaks down.
One such model considers the shape of the photoreceptors. In particular, the
cone shape of the receptors that mediate color vision may be important. In Sec-
tion 2.5, we showed that a tube with a different index of refraction than its sur-
rounding area may conduct light as a result of total internal reflection. If the
cross-section of the tube is made very small, of the order of a single wavelength
of light, the tube becomes a waveguide that is characterized by the waveguide
parameter V [1064]:

πd
V= n21 − n22 . (4.17)
λ
Here, d is the diameter of the tube, n1 and n2 are the refractive indices of the tube
and its surround, and λ is the wavelength for which the waveguide parameter is
computed. The waveguide parameter V can be used to determine if the tube is

i i

i i
 i i

i i

4.8. Application: Modeling a Human Retina 247

wide enough for geometric optics to be a valid model, or if waveguide effects

play a role. The latter would be the case for smaller values of V . For very small
values of V , light cannot be confined to the interior of the tube, and light will
radiate away.
As V depends on both the diameter of the tube, as well as the wavelength, it
is clear that given a specific tube diameter, longer wavelengths cannot be propa-
gated through the tube, whereas short wavelengths can. Replacing the tube with a
tapered cylinder, at the wide entrance many different wavelengths can be propa-
gated, whereas moving towards the thin end of the taper, only shorter and shorter
wavelengths can be propagated.
If the wavelength composition along the length of the tapered cylinder could
be read out, then the waveguiding principle can be used to construct a spectrom-
eter. It turns out that the size and shape of the cones in the retina are such that
waveguiding could occur. It is therefore possible that cones act as waveguides,
irrespective of which photopigment, or combination of photopigments, is present
in each.
This model of color vision, however, is currently not widely accepted. Al-
though it appears to explain several phenomena that current models of human
color vision cannot explain, significant further research would be required to de-
termine if cones in the human retina use waveguiding to mediate color vision. So
far, the model is commensurate with existing literature. However, direct verifica-
tion of this model has yet to occur.

4.8 Application: Modeling a Human Retina

All displays and all rendering algorithms produce stimuli that are ultimately in-
tended to be viewed by humans. A relatively complete model of low-level human
vision may therefore help in the assessment of how images are actually perceived.
Such a model may, for instance, help assess the retinal image quality of images
displayed on new hardware before the hardware is built [243]. For rendering al-
gorithms, such a model may help assess where shortcuts in the rendering pipeline
may be achieved.
Deering presents a model taking into account the eye optics, the photoreceptor
mosaic, as well as the transduction of photons by the cones (the rods are excluded,
so that the model is suitable for photopic light conditions only) [243]. As part of
the simulation, which covers much of the material discussed in this chapter, a
new algorithm for synthesizing artificial retinas is developed, placing and pack-
ing cones according to some simple rules. A new retina is started with just seven

i i

i i
 i i

i i

248 4. Human Vision

cones; a center cone surrounded by six hexagonally packed cones. In each fol-
lowing iteration, a ring of new cones is placed outside this cluster. Based on the
location where cones are created, they have a larger or smaller anatomical target
radius. The cone centers are then migrated inward.
This migration is governed by two rules, the first being adherence to each
cone’s target radius and the second a tendency to move inward as far as possible.
For a pair of cones with positions pi and target radii ri , the normalized distance d
is given by
|pi − p j |
d(i, j) = . (4.18)
ri + r j
Cones i and j are defined as neighbors if d(i, j) ≥ 1.5.
The cone positions are then iteratively updated using the cone force equation:
pi − p j
pi = pi + k1 n + k2 r + k3 ∑ s(d(i, j))
|pi − p j |
. (4.19)
j|d(i, j)≥1.5

Here, n is a vector pointing from pi to the center of the retina, and r is a random
vector such that |r| ∈ [0, 1]. The factors k1 , k2 , and k3 are constant, and s is an
interpolation function given by
⎧
⎪
⎨0, d > 1,
s(d) = 3.2 (1 − d), 0.75 < d ≤ 1, (4.20)
⎪
⎩
0.266 (3.75 − d), 0 ≤ d ≤ 0.75.
These migration steps are executed between 25 and 41 times to let the cones settle
into a pattern. The upper bound of migration iterations leads to a retinal mosaic
that is mixed regular and irregular in a manner similar to real retinas. With this
parameterization, each cone has on average 6.25 neighbors, which is a number
also found in human retinas. Outside the fovea, fewer migration iterations are
carried out to simulate that there the retina is less regularly packed.
Once the current ring of cones has finished migrating, the next iteration starts
by surrounding the current cluster with a new ring of cones, which in turn will
migrate inwards.
It would be possible to model the response of each cone with any of a num-
ber of models. A recent model that predicts a wide range of measurements was
presented by van Hateren [434]. It is based on state-of-the-art knowledge of the
various stages involved in photo-transduction. For a full account of this model,
the reader is encouraged to track down the reference [434], which is freely avail-
able on the Journal of Vision website. As the details of the model are beyond the
scope of this book, we only give an overview of this model in the remainder of
this section.

i i

i i
 i i

i i

4.8. Application: Modeling a Human Retina 249

The basic building blocks are temporal low-pass filters, static (non-) linear
filters, and subtractive and divisive feedback loops. Each of these building blocks
forms a plausible model of some aspect of cone functionality, whether cellular or
molecular. A low-pass filter transforms an input x(t) to an output y(t) with
 ∞
y(t) = h(t − s) x(s) ds, (4.21a)
−∞
⎧  
⎨ 1 exp −t , t ≥ 0
h(t) = τ τ (4.21b)
⎩
0 t <0

This equation can also be written in differential form:

dy(t)
τ = x(t) − y(t). (4.22)
dt
In both forms, τ is a time constant.
In Figure 4.31, the cascading of processing functions is indicated with a con-
trol flow diagram. The four main blocks are the photo-transduction stage, a cal-
cium feedback loop, processing occurring in the inner segment, and finally the

Phototransduction Calcium feedback

*
I E β X Ios
τR τE cβ+kβE* 1/β / τX Xnx

1/α C
1+(aCC)nc τC

Inner segment Horizontal cell feedback

Ios Vis Vs It Vb
/ τm - It(Vs) τ1 aIτ2 To bipolar cells

gi Vh
τis aisVisγ aIτh

To horizontal cells
τ Low pass filter with time constant τ

1/β Static filter

/ Divisive feedback
- Subtractive feedback

Figure 4.31. A model of cone response (after [434]).

i i

i i
 i i

i i

250 4. Human Vision

horizontal cell feedback. While the constants and functions indicated in this flow
chart are detailed in van Hateren’s paper [434], it is included here to show the sort
of complexity that is required to model only the very first stage of human vision,
the process of photoreception in a single cone.
Compare this complexity with the simple sigmoidal response functions
(Michaelis-Menten or Naka-Rushton) discussed in Section 4.3.3. These mod-
els treat photoreceptors as black boxes with an input and an output, whereas the
present model incorporates most of the separate molecular and cellular compo-
nents that define cone functionality. In addition, it takes temporal aspects into
account (except slow processes such as bleaching).
No matter which model is preferred, it is now known that most of the sensi-
tivity regulation in the outer retina of primates is already present in the horizontal
cells and can be measured there [1057]. This is necessarily so, and is consistent
with the model discussed here.
It makes the model useful for applications in visual computing. For instance,
it has been shown to be capable of encoding high dynamic range video [435].

4.9 Further Reading

Hubel’s book Eye, Brain, and Vision can bow be found online at http://neuro.
med.harvard.edu/site/dh/ [496]. Another excellent online resource is Webvision
at http://webvision.med.utah.edu. Classical books on human vision are those by
Cornsweet [208], Wandell [1205], and Palmer [868]. Lots of information on the
retina can be found in Dowling [262], whereas the anatomy and histology of the
eye are described in much detail by Forrester et al. [329]. The collection of papers
edited by Gegenfurtner and Sharpe gives deep insight into early visual processing,
including the molecular structure that mediates vision [353].

i i

i i
 i i

i i

Chapter 5
Perception

So far, we have described components of the human visual system (HVS) and
some of what is known about their interconnectivity. Long before anything was
known about the components of the human visual system—the substrate that me-
diates vision—inferences about how humans make sense of the world were based
on psychophysical experiments.
In a typical psychophysical experiment, a participant is shown a stimulus and
is given a task. By varying aspects of the stimulus, the task performance can be
related to the variation in the stimulus. For instance, it is possible to ask a par-
ticipant if a patch of light can be distinguished from its background. By varying
the intensity of the patch between trials, the threshold at which the patch becomes
visible can be determined.
In a sense, psychophysical experiments treat the human visual system as a
black box. The input is the stimulus, and the output is the task response given by
the observer. With suitably designed psychophysical experiments, many aspects
of human vision can be inferred. These types of experiments can be designed
to ask questions at a much higher level than neurophysiology currently is able to
address. In particular, many experiments exist that answer questions related to
human visual perception.
Although many different definitions of human visual perception exist, we will
refer to human perception as the process of acquiring and interpreting sensory
information. Human visual perception is then the process of acquiring and in-
terpreting visual information. Thus, perception is related to relatively low-level
aspects of human vision.
High-level processing of (visual) information is the domain of cognition and
involves the intelligent processing of information, as well as memory, reasoning,
251

i i

i i
 i i

i i

252 5. Perception

attention, thought, and emotion. In general, the higher-level processes are much
more difficult to understand through psychophysical means than the lower-level
perception attributes of human vision. As such, most of this chapter will focus on
perception, rather than cognition.
One of the cornerstones of color perception is the triple of trichromatic, color-
opponent, and dual-process theories. After giving general definitions and dis-
cussing some of the problems faced by the human visual system, we therefore
discuss these early in this chapter.
The trichromatic, opponent-color and dual process theories are able to explain
several aspects of human vision. However, there are many more peculiarities
of vision that they do not explain. Thus, dependent on the application, further
refinements and additions to these models will have to be considered. Rather than
present a comprehensive catalogue of visual illusions, we show a more-or-less
representative selection. It will be clear that human vision has many complexities
that may have to be taken into account when developing visual applications. In
addition, visual illusions play an important role in understanding the HVS [272].
After all, knowing when visual processing breaks down allows us to learn about
the operation of the HVS.
In the remainder of this chapter, we then discuss some of the mechanisms of
early vision that have been uncovered that help to explain some of these visual
illusions. These mechanisms, which include various forms of adaptation, sen-
sitivity, and constancy, can be used as building blocks in engineering solutions.
As an example, models of lightness and brightness perception have been used
as part of tone-reproduction operators (which are the topic of Chapter 17), and
retinex theory has been used in several different applications. Chromatic adapta-
tion (Chapter 10) is important in white-balancing applications in photography.
We conclude this chapter with a discussion of higher-level processing involv-
ing color naming, as well as color deficiencies and tests for them.

5.1 Lightness, Brightness, and Related Definitions

Before we are able to discuss human vision in greater detail, it is helpful to in-
troduce several definitions related to human visual perception. The definitions we
have chosen here, are following Wyszecki and Stiles [1262]:

Light is that aspect of radiant energy of which a human observer is aware through
the visual sensations that arise from the stimulation of the retina of the eye
by radiant energy.

i i

i i
 i i

i i

5.1. Lightness, Brightness, and Related Definitions 253

Color is that aspect of visual perception by which an observer may distinguish

differences between two structure-free fields of view of the same size and
shape, such as may be caused by differences in the spectral decomposition
of the radiant energy concerned in the observation.

Hue is the attribute of the perception of a color denoted by blue, green, yellow,
red, purple, and so on.

Unique hues are hues that cannot be further described by the use of the hue name
other than its own. There are four unique hues, each of which shows no
perceptual similarity to any of the others. They are red, green, yellow, and
blue.

Brightness is the attribute of visual sensation according to which a given visual

stimulus appears to be more or less intense, or, according to which the area
in which the visual stimulus is presented appears to emit more or less light.
Brightness ranges from “dim” to “bright.”

Lightness is the attribute of visual sensation according to which the area in which
the visual stimulus is presented appears to emit more or less light in propor-
tion to that emitted by a similarly illuminated area perceived as the “white”
stimulus. It is sometimes referred to as relative brightness. Lightness
ranges from “dark” to “light”.

Chromaticness is the attribute of visual sensation according to which the (per-

ceived) color of an area appears to be more or less chromatic. It is also
known as colorfulness.

Chroma is the attribute of visual sensation that permits a judgment to be made

of the degree to which a chromatic stimulus differs from an achromatic
stimulus of the same brightness.

Saturation is the attribute of visual sensation that permits a judgment to be made

of the degree to which a chromatic stimulus differs from an achromatic
stimulus regardless of their brightness.

In addition to these basic terms, it should be noted that the above definition of
color refers to a perceived quantity. It is therefore also sometimes referred to as
perceived color.

i i

i i
 i i

i i

254 5. Perception

Further specifications of color are as given below [1262]:

Related colors are perceived to belong to an area seen in relation to other colors.

Unrelated colors are perceived to belong to an area seen in isolation from other
colors.

Chromatic colors are perceived as possessing hue.

Achromatic colors are perceived as devoid of hue.

The four unique hues have specific properties, in that unique red and unique
green are perceived to be neither yellow or blue. Similarly, unique yellow and
unique blue are perceived to be neither red or green. The hueness of a color stim-
ulus is sometimes described as a combination of two unique hues. For instance,
orange can be described as reddish-yellow or yellowish-red [1262].
The terms brightness and lightness are often confused with each other, and
with luminance and illuminance (which are photometric terms introduced in Sec-
tion 6.2). Wyszecki and Stiles’ definition given above will be the one that we use
as well, but it should be noted that in some fields these terms are used to describe
other quantities. For instance, brightness is sometimes defined as perceived lumi-
nance and lightness as perceived reflectance [13].1 In those cases, the term surface
lightness could be used [140]. Finally, brightness is also known as luminosity.

5.2 Reflectance and Illumination

The human visual system solves a remarkably difficult problem, namely to make
sense of the environment. The carrier of information is light, which falls on the
retina and is then processed. In Section 2.10, the process of image synthesis is
outlined. This process is a computational model of how light interacts with the
environment to ultimately form an image on either a camera sensor or the retina.
In particular, the rendering equation (2.207) provides a relatively complete
model of light interaction. For a given point on a surface, this equation computes
how much light is reflected into a direction of choice, given the light incident
upon this point from all surface points in the scene. Thus, a single evaluation of
this equation requires the full geometry of the scene as well as all light incident
upon a single point, arriving from all other points.
While this is a computationally expensive equation, the human visual system
has evolved to evaluate the inverse of this equation. Thus, starting with light
1 A lively debate on lightness and brightness perception is available online, at www.psych.ndsu.

nodak.edu/mccourt/CV/Hyperlinkedpapers/YorkCD/.

i i

i i
 i i

i i

5.2. Reflectance and Illumination 255

Light spectrum
A Reflectance spectrum
Spectrum reaching retina

Wavelength

Figure 5.1. The light spectrum and the surface reflectance together determine the spectrum
of light reaching our eyes (after [868]).

falling on the retina, an impression of the distribution of light in the environment

is derived. In addition, the placement and shape of objects in the environment is
deduced.
One of the tasks the human visual system appears to carry out is the compu-
tation of surface reflectance, i.e., the illumination and surface-reflectance com-
ponents are disentangled. We note, however, that this is not a universally held
opinion. There are still many uncertainties and open questions related to what the
human visual system perceives, and how.
However, at some point in the chain of processing, the human visual system
infers 3D shape from 2D retinal input. This is an under-constrained problem.
First, the environment has three dimensions, whereas the retinal input has only
two. Second, Equation (2.207) is not invertible.
An example of the complexities involved are shown in Figure 5.1. Here, the
same surface (middle column) is illuminated with three different light sources (left
column). The resulting spectral composition falling on the retina (right column)
is different in each case. Yet, under suitable conditions, the human visual system
is able to infer the reflectance spectrum of the surface to a reasonable level of
accuracy. This allows the illuminant to be discounted and objects in environments
to be distinguished. Ignoring the illumination is known as color constancy [42,
120] and is discussed in more detail in Section 5.9.
Any under-constrained (computational) system requires additional informa-
tion to allow a solution to be approximated. This additional information has to

i i

i i
 i i

i i

256 5. Perception

come from something other than the retinal input. As the human visual system
has evolved to view natural scenes, i.e., environments as commonly encountered
in daily life, it is reasonable to believe that the HVS has come to rely on statistical
regularities available in such environments. There may be many assumptions that
the HVS makes about its environment. Some of these are known and others have
yet to be discovered. In any case, these assumptions serve the role of additional
information that helps disentangle illumination and geometry.
The key insight, therefore, is that human visual processing is beautifully
matched to commonly encountered environments. On the other hand, if some
of these known or unknown assumptions are broken, the HVS may not be able to
infer its environment correctly. In those cases, a visual illusion may result.

5.3 Models of Color Processing

The perception of color is mediated by many different processes occurring along
the visual pathways. Before any deep insights were gained into the inner workings
of the human visual system, several models were developed to account for some
of the peculiarities of human color vision. These include trichromatic theory,
postulating three types of light transducer, and opponent-process theory, which
better predicts certain forms of color blindness. Both theories are now held to
be true, and are united into the dual-process theory. Each of these are discussed
in the following sections. Note that these theories are in many ways the percep-
tual equivalents of the processing in the early visual system, as inferred from the
neurophysiology discussed in the preceding chapter.

5.3.1 Trichromatic Theory

Trichromatic theory was introduced by George Palmer in 1777 [867] and inde-
pendently proposed by Sir Thomas Young in 1802 [1283]. It was extended in
1855 by James Clerk Maxwell [756] and then by Hermann von Helmholtz [453]
in 1867. It is now known as the Young-Helmholtz trichromatic theory.
In particular, Helmholtz determined that new colors can be formed by mix-
tures of other colors. His color-matching experiments involved a test color and
a set of light sources that can be blended in various amounts. A participant in a
color-matching experiment can adjust the strength of each light source in the set
to visually match the test color. Helmholtz found that the minimum number of
light sources required to match any given test color was three. Thus, trichromatic
theory predicts that any perceivable color can be matched with a mixture of three
primary colors.

i i

i i
 i i

i i

5.3. Models of Color Processing 257

Hence, it was surmised that the human visual system has three different mech-
anisms to detect combinations of wavelengths. It was confirmed many years later
that the human visual system indeed has three types of photoreceptors that are
active under photopic conditions, as outlined in Section 4.3.3.

5.3.2 Opponent-Process Theory

While the trichromatic theory for color vision successfully predicts the outcome
of color-matching experiments, it fails to give a full account of human color vi-
sion. In particular, it does not account for the specific anomalies related to color
blindness. In color blind individuals, pairs of colors are lost simultaneously. In
addition, these pairs are always either red and green, or yellow and blue.
Second, the subjective experience of color is not explained by the trichromatic
theory. Although the color mixing of three colors to match a test color may ex-
plain why purple appears as a mixture of red and blue, the subjective experience
of a mixture of red and green is yellow, rather than the predicted reddish-green.
To account for these unexplained phenomena, the opponent-process theory
was proposed [462]; it was later validated by Jameson and Hurvich [516]. Here,
primary colors are recombined into opponent processes that pair colors in a mu-
tually exclusive manner. In particular, reds and greens are paired, and so are
yellows and blues. The red-green channel encodes all shades of red and green in
one dimension, with the reds encoded as positive values and the greens as negative
values.
Thus, the red-green channel progresses from highly saturated greens to desat-
urated greens, to gray, to desaturated reds, and finally to saturated reds. A second
channel encodes yellows against blues in a similar manner. The third channel is
non-opponent and encodes achromatic responses, from black to white.
A consequence of this encoding is that some pairs of colors are perceived as a
mixture of two colors, whereas others are not. For example, the subjective experi-
ences of a mixture of red and blue is purple, and that of red and yellow is orange.
On the other hand, mixtures along each of the dimensions defined by the color-
opponent theory are not perceived together. Thus, reddish-green and yellowish-
blue are not among our subjective color experiences (see also Section 4.3.7).
Color-opponent theory successfully predicts that disorders affecting one of
the neural encoding mechanisms causes colors to be lost in pairs. This is borne
out in practice, with color blindness affecting either the red-green channel, or the
yellow-blue channel (see Section 5.11).
Finally, each of the two opponent channels have a point that is perceived as
neutral grey. The neutral point in the red-green channel is perceived as neither
red or green. A similar neutral point lies between yellow and blue. Prolonged

i i

i i
 i i

i i

258 5. Perception

Chromatic valence
1.0
Achromatic
Blue-yellow
Yellow Red-green
Red
0.5
Red

0.0

Blue
-0.5 Green

400 500 600 700

Wavelength (nm)

Figure 5.2. The spectral response of the red-green and yellow-blue opponent processes.
Also shown is the spectral response of the achromatic channel (after [516]).

exposure to colored stimuli causes adaptation to occur, resulting in a shift that

places the neutral gray at a different value. Hence, opponent processing may
account for simultaneous color contrast as well as color after-images.
In addition, each channel responds to one of two colors in a mutually exclusive
manner. This affords an opportunity to devise a psychophysical experiment that
determines the spectral response function of the color-opponent mechanisms.
A monochromatic test stimulus of a given wavelength may be presented to an
observer, who is given control over red, green, yellow, and blue light sources that
can be mixed with the test stimulus. The aim of the experiment is to adjust one
of the red and green, and one of the yellow and blue light sources, to yield a light
mixture that is perceived as achromatic. By recording the amount of red, green,
yellow, and blue that needs to be added to the test stimulus for each wavelength,
the spectral distribution of each opponent process is found, as shown in Figure 5.2.
The positive and negative designations are appointed arbitrarily.

5.3.3 Dual-Process Theory

While the trichromatic theory predicts the results of color-matching experiments,
and the opponent-process theory predicts the pairwise loss of color in color blind-
ness and the subjective experiences of mixtures of colors, neither theory gives a
full account of all color phenomena. In particular, the phenomena explained by
one theory are not explained by the other, and vice versa. As such, there is value
in both theories, and in practice they can be unified into a single model of human
visual processing.

i i

i i
 i i

i i

5.4. Visual Illusions 259

Such a unified approach is afforded by the dual-process theory. The precise,

testable form of this theory was developed by Leo Hurvich and Dorothea Jameson
in 1957 [516]; they suggested two sequential stages of color processing. The first
step includes a trichromatic stage, in which light entering the eye is converted into
three values determined by the sensitivity of the photoreceptors in the eye. The
second step is an opponent-process stage which encodes the three outputs from
the previous step and thus forms color opponency.
Although the trichromatic, opponent-color, and dual-process theories were de-
veloped on the basis of psychophysical results, they are confirmed by the neuro-
physiological results presented in the preceding chapter. Computational imple-
mentations of these models are discussed in Chapter 8. Although they constitute
a useful set of color models, they do not form a complete model of human color
vision. There are several levels of visual processing that influence the perception
of color, as discussed in the remainder of this chapter.

5.4 Visual Illusions

When the human visual system is faced with a stimulus that is in some way
anomalous, for instance, because it is designed to deviate from natural images
in a specific way, then it may misinterpret the stimulus. For instance, a static dis-
play may be perceived as in motion. The specific combination of color, contrast,
luminance, and geometry, or a subset of these, are required to create a visual illu-
sion. This serves to show that human visual processing is inherently non-linear,
and that different attributes of a stimulus interact in fascinating and unexpected
ways.
Arguably, one of the tasks of the human visual system is to make sense of
the world. When we take a white object into a sunlit environment, we are able
to determine that the object is white and, at the same time, establish that it is
in a bright environment. Taking that object subsequently into a dark room, we
are still able to detect that the object is white and that it is now placed in a dim
environment. In effect, we are sensitive to relative luminances, but at the same
time are able to detect absolute values. A different way of looking at this is to say
that the human visual system is able to separate reflectance from illumination.
If the human visual system acted as a linear light meter, we would never be
able to accomplish this. Taking the object into the dark room would simply mean
that the measured luminance values would be lower. Thus, if the human visual
system does not act as a simple photometer, then it becomes interesting to ask
what it is that the human visual system computes.

i i

i i
 i i

i i

260 5. Perception

Figure 5.3. The elephants in both panels are identical, yet they appear to be darker on the
left and lighter on the right.

For the perception of lightness, it is now accepted that the retinal input is split
into separate components that are processed more or less independently [602].
However, there are two proposals as to how such a decomposition may occur,
namely into frameworks and into layers [366].
For frameworks, the retinal image is thought to be decomposed into contigu-
ous regions of illumination. Within each region, the highest intensity serves as an
anchor (see Section 5.8.5), and all other values within a region are relative to this
anchor.
The alternative theory presumes that the retinal image is split into overlapping
layers that are each treated separately. The image is then assumed to consist of
an illumination pattern that is projected onto a pattern of surface reflectances. Of
course, this would enable the human visual system to reverse engineer reflectance
from illumination by a process known as inverse optics [742]. There are many
real-world examples where reflectance and illumination are perceived simultane-
ously. An example is a white house reflected in a black car [366]. The intensity
of the reflected pattern is gray; yet, neither the house nor the car appear gray.
Instead, the image is perceptually split into two separate layers.
Another example where an image is split into separate layers is shown in Fig-
ure 5.3 [35]. Here both images show an elephant with identical pixels. Only the
surround is changed, namely lightened in the left image and darkened in the right
image. Both images, however, still appear to be composed of a cloudy pattern
with a dark elephant on the left and a light elephant on the right. The result of this
apparent layered processing is that the human visual system has not perceived the
two elephants as identical.
Thus, the accuracy with which the human visual system is able to disentan-
gle reflectance from illumination is not perfect, and this is one source of visual

i i

i i
 i i

i i

5.4. Visual Illusions 261

illusions. Visual illusions can be seen as systematic errors made by the visual sys-
tem, and they constitute a “signature of the software used by the brain” [365,366].
Many of the errors made by the human visual system can be explained by a model
of frameworks [13, 363]. Likewise, a layers model can also be expanded to in-
clude an account for errors [35, 366]. It is currently not known which model is
more plausible, or even if the two models can be unified.
In any case, visual illusions are important to understand aspects of human
visual processing. In addition, some of them have been directly applied in some
applications. For instance, counter-shading is a well-known technique used by
artists to increase the apparent contrast of images and painting (see Section 17.9).
The Cornsweet-Craik-O’Brien illusion, introduced in Section 5.4.4, has been used
to assess tone-reproduction operators (Section 17.11.4).

5.4.1 The Ponzo Illusion

The Ponzo illusion can be created by placing two identical bars over parallel lines
that recede into the distance [916]. Although the two bars are of identical width,
they appear to be of different size, as shown in Figure 5.4. One explanation for
this effect is that the top bar appears to be further away from the viewer due to
the presence of converging lines which are often associated with distance [134].
As objects that are further away also tend to be smaller, the human visual system
interprets the top bar as being both further away and larger.
This is an example of size constancy, which is the ability of the human visual
system to perceive objects to be of the same size, independent of their distance,
and therefore the size of the projection of the object onto the retina (Figure 5.5).
Different forms of constancy play an important role in human vision. Varia-
tions of it occur in many different image attributes, such as lightness constancy,
color constancy, and contrast constancy, which are discussed in the following
section, as well as in Section 5.8 (lightness constancy) and Section 5.9 (color
constancy).

Figure 5.4. The Ponzo illusion. Two identical colored horizontal bars appear to be of
different length.

i i

i i
 i i

i i

262 5. Perception

Figure 5.5. An example of size constancy. The menhirs in this photograph are all ap-
proximately the same size; Parque Escultórico de la Torre de Hécules, A Coruña, Spain,
August 2005.

5.4.2 Simultaneous Contrast and Contrast Constancy

A uniform gray patch will appear darker or lighter depending on the environment
it is in. We have shown in Section 5.4 that the human visual system does not
act as a linear light meter, but rather it must balance the detection of relative val-
ues, which arguably is required for effective object recognition, with the detection
of absolute values, which is necessary to determine the illumination of an
environment.
While the human visual system gets it right in most naturally occurring en-
vironments, in degenerate cases such as presented in Figure 5.6, the stimulus is
simple enough to see that this balance between absolute and relative values is ten-
uous. Here, the two patches are identical, but their surrounds are not. As a result,
the two patches appear somewhat different.
The traditional explanation of simultaneous contrast is that the edges between
patch and surround are detected by antagonistic center-surround receptive fields
(Section 4.3.4), after which edge information is propagated into the interior of the
patch [207, 396, 869]. The propagation of values from an edge to an interior of a
patch is called filling-in and is discussed further in Section 5.8.3.
The simultaneous-contrast illusion demonstrates that we cannot simply take
a colored patch and derive from it the visual percept that it generates. Instead,
we have to take the viewing environment into account. The sub-domain in color

i i

i i
 i i

i i

5.4. Visual Illusions 263

Figure 5.6. The simultaneous-contrast illusion. The two gray patches appear somewhat
different, but they are in fact identical.

science that deals with this is called color-appearance modeling, and is discussed
in detail in Chapter 11.
Figure 5.7 shows two checkerboard patterns each surrounded by a larger
checkerboard pattern with either a higher or a lower contrast. Although the
two center panels have the same contrast, the one on the right appears to have
higher contrast. This phenomenon is known as contrast constancy or contrast
contrast [180, 1069]. Compare this illusion to the one shown in Figure 5.3.

Figure 5.7. Contrast constancy. The center part of both checkerboard patterns has the
same contrast (after [180]).

5.4.3 The Café-Wall Illusion

The café-wall illusion was named after the place where it was first observed and,
consequently, described: on the wall of a café in Bristol [388,389]. In this illusion,
tiles in different rows are offset by some amount. Between the tiles, there is mortar
of a color that is in between the dark and the light tiles. The presence of the mortar
is essential, as having no lines between the tiles would not create an illusion. The
café-wall illusion produces lines that are perceived as curvy, whereas the lines are
actually perfectly straight, as shown in Figure 5.8.

i i

i i
 i i

i i

264 5. Perception

Figure 5.8. The café wall at the bottom of St. Michael’s Hill in Bristol, UK. The mortar
between the tiles forms straight lines that are perceived to be curved; March 2005.

5.4.4 The Cornsweet-Craik-O’Brien Illusion

The Cornsweet-Craik-O’Brien illusion [208, 842, 1038] is shown in Figure 5.9.
This illusion consists of two identical gray patches separated by two non-linear
ramps of opposite polarity. A sharp gradient is present between the two ramps,
resulting in a luminance profile.
The shape of the ramps is chosen such that they are interpreted by the human
visual system as constant luminance. The C1-continuity between the ramps and
the constant sections ensures that the ramps cannot be differentiated from the
constant-luminance sections. It may be argued that low level center-surround
processes, such as those found in the retina [262, 632, 742, 868], are the cause for
this failure to detect the ramps. These processes are tuned to sharp discontinuities
rather than slow gradients. The discontinuity between the left half and the right
half of this illusion is therefore the only feature that is detected by low level visual
processing. Also, note that center-surround processing of an actual step function
would produce a similar signal.
This visual illusion is implicitly exploited by many artists who apply counter-
shading to their drawings and paintings. A computational simulation for the pur-

Figure 5.9. The Cornsweet-Craik-O’Brien illusion. The left and right quarters of this
image have the same value, whereas the middle half contains two smooth ramps, separated
by a sharp jump. This configuration is perceived as a step function, with the left half and
right half of unequal brightness (after [208]).

i i

i i
 i i

i i

5.4. Visual Illusions 265

pose of contrast enhancement is outlined in Section 17.9. In addition, this model

was used as a stimulus in a validation study to assess the performance of tone-
reproduction operators, as discussed in Section 17.11.4.

5.4.5 The Hermann Grid and the Scintillating Grid

The Hermann grid, shown in Figure 5.10 (left) consists of a regular array of black
blocks which are separated by a white background [130, 463, 1076, 1145]. In the
periphery of vision, the crossings of the white lines will exhibit dark patches.
The classic explanation of this effect is that center-surround mechanisms have
much smaller receptive fields in the fovea than in the periphery. Thus, if the
stimulus is viewed from an appropriate distance, the black patches only occur in
the periphery of vision, where the surround overlaps with both the white bands
and the black squares. When focused on a crossing of white bands, the center-
surround mechanisms in the fovea are small enough to fit entirely within the white
bands and, thereforeb do not overlap with the black blocks. The differential output
of the center and surrounds in the fovea and the periphery are therefore different.
While this explanation may seem reasonable, recently modifications to the
Hermann grid have made it possible to make the illusion disappear. As shown in
Figure 5.10 (right), by replacing the straight edges of the black blocks with wavy
patterns, the Hermann illusion vanishes. As the width between the blocks and
the size of the blocks is not altered, the classic explanation for the Hermann grid
cannot be maintained [354].
A stronger version of the Hermann grid illusion is known as the scintillation
illusion [1011, 1012, 1178]. Figure 5.11 shows the scintillation illusion in two

Figure 5.10. The Hermann grid (left), showing phantom dark blobs on the intersections
of the white bands. The classical explanation of the Hermann grid can be invalidated by
creating curved sides to each block, which makes the illusion disappear (right).

i i

i i
 i i

i i

266 5. Perception

Figure 5.11. By placing white dots at the cross sections of a Hermann grid, the scintillation
effect can be induced. The dark blobs take on the same color as the blocks.

versions. Note that the black boxes give rise to black illusory patches, whereas
the green boxes give rise to green illusory patches.

5.4.6 The Munker-White Illusion

The Munker-White illusion consists of horizontal black and white stripes. If, in
one vertical area, all the black stripes are replaced with stripes of a different color,
and, in another vertical area, all the white stripes are replaced by the same color,
then these two colored areas appear to be of different intensity, as shown in Fig-
ure 5.12 [1241]. This illusion shows that the apparent lightness of a surface can
be affected by the spatial configuration in which it appears.
In comparison with the simultaneous-contrast illusion discussed in Section
5.4.2, the perceived lightness values are opposite. Here, in a predominantly light
surround, the colored bars appear lighter rather than darker. In a predominantly
dark surround, the illusion is that the colored bars are darker. As a consequence, a
simple center-surround explanation given for simultaneous contrast does not ap-
ply to this illusion, as it would predict the wrong perception of these patches [37].

Figure 5.12. The Munker-White illusion.

i i

i i
 i i

i i

5.4. Visual Illusions 267

This has resulted in several possible explanations of the Munker-White illu-

sion. In the first, it is argued that a more complex center-surround configuration
can still give rise to this illusion [102, 325, 799]. Second, the illusion may be the
result of grouping and segmentation. It is typically assumed in such models that
T-junctions are detected by the human visual system, which are then interpreted
in a specific way.
For instance, T-junctions may be thought of as providing information regard-
ing occlusions of objects. In a T-shaped edge, the surface above the horizontal bar
occludes the objects separated by the stem of the T-shape. Thus, T-shaped edges
allow surfaces to be ranked in order of “belongingness”: the objects either side
of the stem of the T belong together more than they belong to the surface above
the horizontal bar [363]. In White’s illusion, the colored bars therefore belong
more to the bars to which they are superposed than to the ones to which they are
adjacent [1129, 1288].
The anchoring theory proposed by Alan Gilchrist also provides an explanation
of White’s illusion [37,363]. This theory is discussed further in Section 5.8.5. An
extensive review of the Munker-White illusion is presented by Anderson [37].

Figure 5.13. Neon color-spreading examples. On the top left, an achromatic version is
shown, which is also known as neon brightness spreading. The top-right figure is due to
Varin [1179], whereas the bottom-left illusion was developed by van Tuijl [1151]. The
bottom right illusion is from [127].

i i

i i
 i i

i i

268 5. Perception

5.4.7 Neon Color Spreading

The neon color-spreading illusion typically occurs when an incomplete black line
pattern is completed with lines of a different color. The colored parts of the figure
then dilute into the white surrounding areas, as seen in Figure 5.13. The color
of the red bars in this figure spreads into its neighboring surrounding area, but is
importantly halted by an illusory contour.
This illusion was first discovered by Varin [1179] and later independently
rediscovered by van Tuijl [1151]. The bottom-left figure is a variation of the
Ehrenstein figure, showing that a very minimal configuration can still result in the
neon color spreading effect [277, 278].
Discussion of this illusion can be found in various places [593, 942, 943].
See also Bressan et al. for an overview of conditions under which the neon color
spreading effect is likely to occur [127].

5.4.8 Color After-Effects

Staring at a field of a particular color for a period of time, and then focusing on
a white background, may yield an after image of opposite color. As the human
visual system adapts to prevailing conditions, for instance, the prolonged presence
of a particular color, once that color is removed, the HVS remains adapted to
this color for a while longer before re-adapting to the new situation. During this
period, an after-effect may be visible.
Interestingly, some after-effects last for more than a few seconds. In particular,
contingent after-effects may endure for several days. Such after-effects depend on
two stimulus variables. For instance, in the case of the McCollough effect, the
variables are color and orientation [762].
The McCollough effect can be experienced by observing Figure 5.14. The
two chromatic gratings are the adapting stimuli, whereas the achromatic grating
at the bottom is the test stimulus. Stare at the two chromatic gratings for at least
10 minutes, alternating between the two every 15 seconds. Move your eyes slowly
around the white circles. This procedure will allow your eyes to adapt to the two
colored gratings.
Afterward, look at the achromatic gratings. The horizontal bars should have
a greenish tinge, whereas the vertical bars should be colored magenta. These are
thus the complements of the colors in the adapting stimuli.
The induction protocol that must be followed to create the McCollough ef-
fect includes a uniformly oriented pattern of a particular color, which may be
alternated with orthogonally oriented patterns of a complementary color. After

i i

i i
 i i

i i

5.4. Visual Illusions 269

Figure 5.14. The McCollough effect.

induction, observing achromatic versions of these patterns, a faint tinge of a com-

plementary color can be seen [139].
It should be noted that the McCollough effect can be extremely long lasting,
making it special among visual illusions. The 10-minute adaptation time rec-
ommended above was found to be sufficient to make the after effect last for 24
hours [557]. It is therefore known an a contingent after-effect.
As discussed in Section 4.2.5, as a result of astigmatism, the optics of the
eye produces colored fringes whereby the color depends on the orientation of
the edges. One theory proposes that the McCollough effect is a manifestation
of the neural circuitry that corrects for some of the artifacts, namely chromatic
aberration, introduced by the eye’s optics [139]. Overviews of the McCollough
effect are published by Stromeyer [1100] and McCollough [763].

i i

i i
 i i

i i

270 5. Perception

5.5 Adaptation and Sensitivity

When observers move through an environment, they get accustomed to the par-
ticular lighting conditions of the environment. According to Goethe, “Prisoners
who have been long confined in darkness acquire so great susceptibility of the
retina, that even in the dark (probably a darkness very slightly illumined) they can
still distinguish objects.” [373]. Under normal conditions, changes in lighting by
factors of up to 1000 go largely unnoticed [63]. Much larger changes in lighting
are very noticeable. For instance, entering a cinema during the day leads to tem-
porary blindness; sight is slowly regained after some time. Leaving the cinema
and entering daylight may be a painfully bright experience as well.
In both cases, the human visual system adapts, so that after some period of
time, the vision system becomes accustomed to the new lighting conditions. The
process of getting accustomed to prevailing (lighting) conditions is called adap-
tation. It regulates the visual sensitivity to the current environment in order to
maximize the ability for higher-level processes to extract information from the
environment.
Within the concept of adaption, two further distinctions need to be made.
First, steady-state adaptation refers to the performance of the human visual sys-
tem when fully adapted. Thus, visual attributes can be measured with respect to
the lighting conditions when the human visual system is at equilibrium.
On the other hand, temporal adaptation refers to the performance of the hu-
man visual system during the adaptation process. Thus, visual performance is
measured as a function of the time that has elapsed since the onset of adaptation.
There are many attributes to which the human visual system adapts. The
two specific examples given above are dark adaptation and light adaptation for
entering and leaving the cinema, respectively. There are many other examples of
situations where the human visual system adapts, including chromatic adaptation
which allows the human visual system to adapt to lighting of a particular color.
An example in which this process causes unexpected behavior are the color after-
effects discussed in Section 5.4.8.
Light adaptation, dark adaptation, and chromatic adaptation are important ef-
fects for the study of color and are discussed in the following sections. We also
discuss sensitivity, as sensitivity and adaptation are closely related.

5.5.1 Dark Adaptation

The human visual system functions over a very large range of illumination. While
the range of light that can be processed and interpreted simultaneously is limited,

i i

i i
 i i

i i

5.5. Adaptation and Sensitivity 271

Log intensity
8

7
Cones
Rod-cone break
6

5
Rods
4

2
0 5 10 15 20 25 30
Time in the dark (min)

Figure 5.15. The time course of dark adaptation (after [443]).

over the course of time the human visual system adapts to a much larger range
of illumination. These processes are not instantaneous. The time course of adap-
tation for the transition between a light environment and a dark environment is
called dark adaptation.
There are several mechanisms at play during dark adaptation, including di-
lation of the pupil, photochemical activity in the retina, and neural processes in
the remainder of the visual system. While the pupil rapidly adjusts to new light-
ing conditions, the purpose of constriction is to improve the optical resolution
of the stimulus falling on the retina, rather than a mechanism to stabilize image
luminance [150, 151, 1238].
Dark adaptation can be measured by letting an observer gaze at a pre-adapting
light for approximately five minutes before placing the observer in a dark room
and increasing the intensity of a test stimulus until it is detected by the observer
[53]. The intensity of this threshold stimulus can be plotted as a function of time,
as shown in Figure 5.15. The adaptation curve followed by individual observers
varies somewhat, but tends to lie within the light gray region in this figure.
The curve consists of two separate regions that are marked by a break. During
the first few minutes of dark adaptation, the cones gain in sensitivity and dominate
the adaptation curve. The threshold level at the end of the cone adaptation curve
is called the cone threshold. The rod system is slower to adapt, and only begins
to dominate the adaptation process after the cones are at equilibrium. The curve
followed for the rods ends with a maximum sensitivity after around 40 minutes.
The threshold reached at that point is called the rod threshold and has a value of
around 10−5 cd/m2 .

i i

i i
 i i

i i

272 5. Perception

If the test stimuli used during dark adaptation are colored, then before the rod-
cone break is reached, these stimuli appear colored. After the rod-cone break, they
appear colorless.
Dark adaptation is affected by several factors, including the intensity of the
pre-adapting light, the exposure duration, the size and location of the spot on the
retina used to measure adapting behavior, the wavelength of the threshold light,
and rhodopsin regeneration [208, 444, 445, 899].

5.5.2 Light Adaptation

When a fully dark-adapted individual is exposed to light, the human visual sys-
tem adapts to the new and brighter lighting conditions. This process is called light
adaptation. It can be measured with increment threshold experiments, whereby
an observer is exposed to test stimuli presented on a background with a given in-
tensity. The test stimulus is then increased until it is detected. The results of such
an experiment are plotted in Figure 5.16. On the horizontal axis the background
intensity is plotted, and the vertical axis plots the threshold intensity of the test
stimulus. For this reason, curves as shown in this plot are called threshold-versus-
intensity (or TVI) curves .
In this particular plot, the curve consists of two branches; the lower branch
is caused by the rod system and the upper branch is due to the cone system.
TVI curves as a function of the adapting luminance LA can be modeled for cones

-3
Log threshold intensity

-4

-5

Cones
-6

-7
Rods

-8
-9 -7 -5 -3 -4
Log intensity

Figure 5.16. The time course of light adaptation (after [236]).

i i

i i
 i i

i i

5.5. Adaptation and Sensitivity 273

Log ΔI
Saturation
2

0 Weber’s
Law
-1

-2
Dark
Light
-3
Square Root
Law
-4
-8 -4 -2 0 2 4
Log I

Figure 5.17. A schematic threshold-versus-intensity curve for a single mechanism. The

background intensity I is plotted against the difference threshold ΔI. Four different sections
can be discerned (after [16]).

by [310]
⎧
⎪
⎨−0.72 if log(LA ) ≤ −2.6,
log(TVI(LA )) = log(LA ) − 1.255 if log(LA ) ≥ 1.9,
⎪
⎩
(0.249 log(LA ) + 0.65)2.7 − 0.72 otherwise.
(5.1)

For rods, the TVI curve can be approximated by [310]

⎧
⎪
⎨−2.86 if log(LA ) ≤ −3.94,
log(TVI(LA )) = log(LA ) − 0.395 if log(LA ) ≥ −1.44,
⎪
⎩ 2.18
(0.405 log(LA ) + 1.6) − 2.86 otherwise.
(5.2)

Under suitable experimental conditions, it is possible to isolate either the rod or

the cone system, and then a single curve is obtained [16, 63]. Such a curve, plot-
ting the difference threshold ΔI against the background intensity I, is shown in
Figure 5.17.
This curve shows four different sections. At the bottom left, a linear section
characterizes the human visual system’s behavior at the dark-light level. Increas-
ing the background intensity does not initially bring about an increase in detection

i i

i i
 i i

i i

274 5. Perception

threshold, because the sensitivity at these levels is determined by neural noise (the
dark light).
The second part of the curve behaves according to the square root law, also
known as the de Vries-Rose law [979, 1188]:
ΔI
√ = k. (5.3)
I
Here, light levels are such that quantal fluctuations determine the threshold of
vision. Thus, to determine a test stimulus, it must be sufficiently different from
the background to exceed these fluctuations. Under such conditions, an ideal
light detector would record a threshold that varies according to the square root of
the background intensity. In a log-log plot, as shown in Figure 5.17, this would
yield a curve segment with a slope of 0.5. The rod pathway typically has a slope
of 0.6 [417]. The shape of the transition to the next curve segment is determined
by several factors, including the size of the test spot and the duration of the test
stimulus.
A large part of the threshold-versus-intensity curve has a slope of around 1.
This means that the intensity threshold ΔI is proportional to the background in-
tensity I [114, 1225, 1226]. This relation is known as Weber’s law [100, 900]:
ΔI
= k. (5.4)
I
The constant k is the Weber constant or the Weber fraction. The intensity thresh-
old ΔI is also known as the just noticeable difference (JND) . To account for neural
noise, the dark light may be modeled as a small constant I0 that is added to the
background intensity:
ΔI
= k. (5.5)
I + I0
The last part of the curve in Figure 5.17 steepens, indicating that at those back-
ground levels, very large intensity differences become undetectable. Thus, the
system saturates. For the rod system, this occurs at around 6 cd/m2 for a natural
pupil in daylight [63].
The cone system does not saturate, but shows a constant difference threshold
over a very large range of background intensities [133].

5.5.3 Supra-Threshold Sensitivity

Weber’s law, as discussed in the preceding section, relates stimulus detection at
threshold to the magnitude of the background. It is therefore valid to detect incre-
ments, but not directly applicable to supra-threshold stimuli, i.e., stimuli that are

i i

i i
 i i

i i

5.5. Adaptation and Sensitivity 275

well above the detection threshold. To yield a relation between the magnitude of
the stimulus and the strength of the evoked sensation, it is tempting to integrate
Weber’s law, after noting that the Weber fraction can be interpreted as a constant
times the differential change in perception d ψ , i.e., k = k d ψ . Restating Weber’s
law and integrating yields
ΔI
dψ = , (5.6a)
k I
ψ = k ln(I) +C. (5.6b)

For convenience, we set the arbitrary constant of integration C to −k ln(I0 ),

where I0 is a threshold stimulus. The result is Fechner’s law and relates the per-
ceived intensity ψ of a patch to its radiometric value I:
 
 I
ψ = k ln . (5.7)
I0
According to this equation, the magnitude of the sensation has a logarithmic re-
lation to the size of the stimulus. This law makes the erroneous assumption that
threshold discriminations, which are a measure of uncertainty and variability in
sensory discrimination, are integrable to give a relation between stimulus and
sensation.
Thus, currently Fechner’s law has largely been abandoned in favor of a dif-
ferent law that holds for many modalities, including brightness, lightness, smell,
taste, temperature, and many others. It goes by the name of Stevens’ power law
and relates stimulus I to subjective magnitude ψ as follows [1086]:

ψ = k(I − I0 )n , (5.8)

where the exponent n varies according to which of the many sensations is mea-
sured. For example, the exponent is 0.33 for brightness,2 1.2 for lightness, and
3.5 for electric shock.

5.5.4 Contrast Sensitivity

The threshold of visibility depends on many factors. As suggested by Weber’s
law, the visibility of a test stimulus depends on the intensity of the background.
In addition, the ability to discern stimuli from their background depends on the
spatial frequency of the stimuli. A convenient way to adjust the spatial frequency
of a stimulus is by employing a grating, which is a pattern that alternates regularly
over space.
2 Assuming a 5◦ target in the dark.

i i

i i
 i i

i i

276 5. Perception

Figure 5.18. Campbell-Robson contrast-sensitivity chart [152] (Figure courtesy of Izumi

Ohzawa, Osaka University with permission).

A sine function is an example of a grating. Such a function can be created

at many spatial frequencies, as well as different amplitudes. It is thus possible to
construct a 2D function where the spatial frequency is increased along the x-axis
and the amplitude is varied along the y-axis:
1 + sin (exp(x/xmax ))
I(x, y) = log(y/ymax ) , (5.9)
2

500
-1
(Modulation threshold)

0.0009 tr
200 0.009 tr
0.09 tr
100 0.9 tr
9 tr
50 90 tr

20
20

1
0.5 1 2 5 10 20 50
Spatial frequency (cycles/deg)

Figure 5.19. Contrast sensitivity for different levels of retinal illumination. (Data from
van Ness [827]; figure redrawn from [646].)

i i

i i
 i i

i i

5.5. Adaptation and Sensitivity 277

where xmax and ymax are the maximum display dimensions. The resulting image
is shown in Figure 5.18, which is called a Campbell-Robson contrast-sensitivity
chart. In this chart, it is possible to trace from left to right where the sine grating
transitions into a uniform gray. In the uniform gray area there is still contrast,
but it falls below the threshold of visibility. The trace forms an inverted U-shape,
which delineates the contrast sensitivity as a function of spatial frequency.
Contrast sensitivity is, in addition, dependent on the amount of light falling
on the retina. This amount itself is modulated by both the environment as well
as the pupil size. For this reason, the preferred measure for retinal illuminance is
the troland, which is luminance multiplied by the pupillary area.3 The contrast
sensitivity for different levels of retinal illuminance are plotted in Figure 5.19.
Finally, the sensitivity to contrast may be altered in individuals with certain
diseases such as multiple sclerosis and disorders such as cataracts and ambly-
opia [41].

5.5.5 Chromatic Adaptation

The TVI curves explained in Section 5.5.2 are applicable to both the rod and
the cone systems. As there are three types of cones with peak sensitivities at
different wavelengths, it is reasonable to assume that each cone type follows the
TVI curve independently. This means that if an environment has a strong color
cast, usually because the illuminant is colored, each of the three cone types will
adapt separately and, simultaneously, have a different difference threshold ΔI.
The different levels of adaptation of the three cone types cause the human
visual system to be adapted to the dominant color of the environment. This mech-
anism is called chromatic adaptation. As normally the dominant color of the en-
vironment is caused by the color of the light source, the human visual system thus
becomes adapted to the prevailing lighting conditions. This property of the HVS
is known as discounting the illuminant. The independence of the three receptor
types thus creates an automatic white balancing of the HVS.
Figure 5.20 shows an example of chromatic adaptation. By staring at the
cross in the panel at the top for a few minutes, the retina becomes adapted to this
bipartite field. The image on the bottom is processed to have a similar color cast
in both the left- and right-half to that of the bipartite field. By observing the image
after adaptation to the top panel, the image will look normal, because the HVS
will discount the color casts present in the image.
If the spectral sensitivities of the three receptors at one state of adaptation are
given by l¯1 (λ ), m̄1 (λ ), and s̄1 (λ ), and by l¯2 (λ ), m̄2 (λ ), and s̄2 (λ ) for a different
3 Luminance is a photometric term, explained in Section 6.2.7.

i i

i i
 i i

i i

278 5. Perception

state of adaptation, then these spectral sensitivities are related as follows:

l¯1 (λ ) = kl l¯2 (λ ), (5.10a)

m̄1 (λ ) = km m̄2 (λ ), (5.10b)
s̄1 (λ ) = ks s̄2 (λ ). (5.10c)

These relations are known as the von Kries coefficient law [622]. The coefficients
kl , km , and ks are inversely related to the relative strengths of activation [544]. Von
Kries chromatic adaptation is further discussed in Chapter 10.

Figure 5.20. An example of chromatic adaptation. Stare for 20 seconds at the small cross
in the top panel. Then focus on the cross in the bottom panel. The image should look
normal; Rathaus Konstanz, Germany, June 2005.

i i

i i
 i i

i i

5.6. Visual Acuity 279

5.5.6 Spatial Summation

The human visual system has the ability to sum light falling on neighboring parts
of the retina. This implies that there is a relation between the threshold intensity
ΔI and the size of the stimulus A. It has been found that if the stimulus area is
doubled, the value of the threshold is halved:
I An = k, (5.11)
where the exponent n describes the degree of summation. If summation occurs
fully, then n equals 1. Otherwise, n lies between 0 and 1. This relation is known
as Ricco’s law [236, 329, 570].

5.6 Visual Acuity

The photoreceptor mosaic is a discrete array of light receptors. This means that
the spatial resolution of detail that can be processed by the HVS is limited. The
spatial resolving power of the human visual system is referred to as visual acuity.
There are several factors limiting visual acuity, including the photoreceptor
density and diffraction and abberations in the eye [1054]. In addition, there are
several other factors affecting visual acuity, which include illumination, contrast,
color, and the location of the retina under consideration.
Based on cone spacing alone, the resolving power of the retina is around 60
cycles per degree of visual angle [219]. This is higher than can be achieved with
clinical techniques to infer visual resolving power, as no correction for the optics
of the eye and neural processing was applied.

5.6.1 Factors Determining Visual Acuity

For very young children, visual acuity dramatically depends on age [492]. Fig-
ure 5.21 shows a computer simulation of how visual acuity develops over the first
few months of life. The object is assumed to be 24 inches away. Note that both
visual acuity and color vision are not well developed at birth.
There are at least seven factors that have an impact on visual acuity for adult
humans in addition to the hard limits imposed by the photoreceptor array. These
factors are refractive errors in the optics of the eye, the size of the pupil, the
amount of illumination, the time the target was exposed, the area of the retina
involved in target detection, the state of adaptation of the eye, and the presence of
eye movement.
When a point light is present in the visual field, a thin beam of light is pro-
jected onto the retina. The optics of the eye will spread the beam somewhat.

i i

i i
 i i

i i

280 5. Perception

Figure 5.21. An image processed to simulate human vision after 0, 4, and 8 weeks of
development (top row), and 3 and 6 months of development followed by adult vision (bot-
tom row). The simulation is from tinyeyes.com. Statue by James Tandi; photographed at
Native Visions Art Gallery, Orlando, FL.

The distribution of light on the retina due to a point source is characterized by a

point-spread function (PSF).
Errors in the optics of the eye cause light to not be sharply focused on the
retina; these errors create some amount of defocus. The PSF will overlap with a
larger portion of the retina, leading to a blurred view of the world. Some causes
of blur are myopia (short-sightedness) and hyperopia (long-sightedness). A my-
opic eye causes the light to be sharply focused somewhere in front of the retina,
whereas a hyperopic eye focuses the image behind the retina.
The pupil determines how much light enters the eye. A larger pupil causes
more light to enter. More light will allow more photons to stimulate the retina
and also reduces diffraction effects. However, abberation will increase. Reduc-
ing the pupil size, on the other hand, will reduce abberation effects, but increase

i i

i i
 i i

i i

5.6. Visual Acuity 281

Visual acuity
Max cone
1.6
resolution

1.2 Cones

0.8

Max rod
0.4
resolution
Rods
0.0
-4 -2 0 2 4
Log I

Figure 5.22. The visual acuity changes with illumination, shown here separately for the
rods and the cones (after [962].)

diffraction [51]. Thus, the size of the pupil has a direct effect on the point-spread
function of the projected light.
The amount of background illumination has a direct impact on visual acuity.
For recognition tasks, the relationship between the two is given in Figure 5.22.
The resolving power is not uniformly distributed over the retina, but decreases
with distance to the fovea. The relationship is shown in Figure 5.23.
Under photopic lighting conditions, the best visual acuity is achieved for test
stimuli with the same intensity as the intensity to which the eye is adapted. Under
scotopic conditions, the visual acuity is much lower and is determined by the AII
amacrine cell [606, 1218].

1.0
Snellen fraction

0.7
0.5

0.3

0.2

0.1
0.07
0 5 10 15 20 25 30
Eccentricity (degrees)

Figure 5.23. Visual acuity as function of retinal position, measured as the eccentricity
from the fovea (after [1239]).

i i

i i
 i i

i i

282 5. Perception

5.6.2 Visual Acuity Measurements

The visual acuity differs between individuals and tends to degrade with age. It
may be measured in individuals in different ways. The first is target detection
where the location of a target needs to be determined. Examples are Landolt rings
(Figure 5.24 (left)) and the illiterate E (Figure 5.24(middle)).

Figure 5.24. Landolt rings (left) and illiterate Es (middle) are used for target detection,
whereas Snellen charts (right) are useful for target recognition.

Target recognition tasks are also frequently used to measure an individual’s

visual acuity. The Snellen chart is a good example (Figure 5.24 (right)).
Finally, target localization is a separate technique whereby a small displace-
ment of an object, such as a discontinuity in a contour, can be detected (Fig-
ure 5.25). If visual acuity is measured in this way, it is called Vernier acuity.

Figure 5.25. Vernier acuity can be measured by having observers locate small displace-
ments. In both panels, the displacements are increased from left to right.

5.7 Simultaneous Contrast

One of the tasks of the human visual system is to detect objects in its environ-
ment. To perform this task, the human visual system uses cues such as brightness,
color, texture, shape, motion, and size. The difference in these cues, coming from
objects and their immediate surroundings, allows for the detection of these ob-
jects. For example, people with normal vision are easily able to detect a red apple
against green foliage, by using differences in the color cue coming from the apple
and its surround.4
4 A relation between the need to distinguish certain tropical fruits and the evolution of primate

vision has been proposed. An overview of arguments for and against this hypothesis is given by

i i

i i
 i i

i i

5.7. Simultaneous Contrast 283

Figure 5.26. An example of simultaneous contrast. The central circle on the left appears
larger than the one on the right, even though they are the same size (after [868]).

The human visual system enhances such differences in cues, allowing for eas-
ier detection of objects. Take the identical inner circles in Figure 5.26, for ex-
ample. In this artificial scenario, the circle on the left is surrounded by smaller
circles, and the one on the right is surrounded by larger circles. The human visual
system enhances the difference between the sizes of the central circles and their
surrounds, making the circle on the left appear larger and the one on the right
appear smaller. As a result, the two central circles appear to have different sizes.
The enhancement of the difference between cues from the object and its surround
is termed simultaneous contrast.
Similarly, a gray region appears brighter when viewed on a darker surround,
and darker when viewed on a brighter surround [171]. This type of simultaneous
contrast is called simultaneous lightness contrast. Simultaneous color contrast
refers to the perceived enhancement of differences in the color cue. Both these
types of simultaneous contrast are discussed in the following sections.

5.7.1 Simultaneous Lightness Contrast

An example of simultaneous lightness contrast is shown in Figure 5.27, where
the ring of uniform gray appears to be composed of four different shades of gray.
The ring, or stimulus, appears darker if its surround is lighter, and lighter if its
surround is darker. If the ring were cut out and removed from its surround, it
would appear a uniform gray.
Mach bands form another well-known effect of simultaneous lightness con-
trast. These are named after their discoverer, Ernst Mach. Mach bands are the
perception of bands along C1-discontinuities in images. These bands accentuate
the difference across such edges, as illustrated in Figure 5.28. There is a percep-
tion of bands on either side of the edge. The band on the left of the edge appears
brighter than the region on the left, and the band on the right appears darker than
Regan et al [944].

i i

i i
 i i

i i

284 5. Perception

Figure 5.27. A gray ring of uniform lightness appears to have four sections of different
lightness. The perceived lightness of each section is influenced by the lightness of the
surround (after [868]).

the region on the right. The graph shows the difference between the perceived and
the actual lightness across the edge.

1.0
Intensity
0.8 Perceived intensity

0.6

0.4

0.2

0.0
0.0 0.5 1.0

Figure 5.28. Bands are perceived on either side of the edge shown above. The presence
of these bands enhances the physical difference across the edge.

5.7.2 Simultaneous Color Contrast

Similar to color after-effects, which involve the perception of complementary
hues over time, simultaneous color contrast involves the perception of comple-
mentary hues across space. Simultaneous color contrast is also known as induced

i i

i i
 i i

i i

5.7. Simultaneous Contrast 285

Figure 5.29. The uniformly colored ring is perceived to have different colors, determined
by the background color behind the ring (after [868]).

color. As with simultaneous lightness contrast, induced color may be demon-

strated by placing a uniform color against different backgrounds, so that the color
appears different. Figure 5.29 demonstrates this effect. The perceived color of the
uniform gray ring is determined by the background behind it. In the case of a blue
background, the ring appears to have the hue of the complementary hue (yellow),
while against the yellow background, the ring appears blue.
The magnitude of the perceived contrast effect may be measured by using a
nulling method [614]. In this method, the observer is asked to fixate on a white
disk. If its surround is set to green, for example, the white disk appears to be
colored by the opposite hue, which is red in this case. The perceived red color of
the disk may be neutralized by adding a little of the opposite color (green) to the
disk. The amount of green light required to make the disk appear white again is
the measure of the contrast effect.
When this experiment is repeated for different colors, the observer is required
to keep a standard white in mind. This may be avoided by modulating the color
of the surround from one color to its complementary hue. In this version, the
observer simply has to keep the appearance of the disk steady as the color of the
surround is modulated from say, red to green. Figure 5.30 illustrates the result of
such a measurement.

i i

i i
 i i

i i

286 5. Perception

Red Inducing color

‘Nulling’ color
Induced color

Green
Wavelength (nm)

Figure 5.30. Nulling technique for measuring the effect of simultaneous color contrast
(after [614]).

5.8 Lightness Constancy

The color theories described in the preceding sections are valid descriptions for
isolated patches of color. The context in which these colors appear are therefore
ignored. Nevertheless, this context plays an important role in the perception of
color. Our color perception is affected by factors such as the geometry of the
environment and its lighting conditions.
An example is shown in Figure 5.31. Here, the colors of the left and right half
of the light house are different when seen in isolation. When observed in context,
it is clear that the surface reflectance of the light house remains constant, but the il-
lumination changes. The complexities involved in disentangling reflectance from
illumination were outlined in Section 5.2. Figure 5.31 shows that the HVS effort-
lessly separates surface reflectance from illumination. This particular feature is
called constancy. It enables the perception of objects in scenes, independent of
illumination.
Formally, constancy may be defined as the ability to perceive the reflectance
properties of surfaces despite changes in lighting and other viewing conditions
[868]. It is a feature of the human visual system that applies to several different
modalities. In particular, lightness constancy applies to achromatic stimuli and
the perception of black and white. An example is given in Figure 5.32. The
perceived difference in the two checks marked A and B are due to some of the
inferences the human visual system makes.
First, local contrast suggests that check A is darker than average, whereas
check B is interpreted as lighter than average. As this assumption is based on
local contrast only, the presence of the shadow does not alter this logic. Second,
high frequency edges are typically interpreted as belonging to textures, whereas
low frequency content may be due to shadows, and is therefore ignored.

i i

i i
 i i

i i

5.8. Lightness Constancy 287

Figure 5.31. The left and right half of the light house each stimulate the photoreceptors
differently, as demonstrated by the insets. However, the wall of the light house is still
perceived as one continuous surface. The difference in illumination is discounted by the
human visual system, as long as the context is present; Light house at Ponce’s Inlet, FL,
June 2005.

These inferences normally help in reconstructing 3D information from retinal

input, and are the cause of lightness constancy. Assume, for instance, that the
human visual system operated as a linear light meter. In that case, we would not
be able to distinguish a white surface in a dark environment from a black surface
in a very light environment [13]. However, humans can normally accomplish
this task, which is a desirable property enabling us to make inferences about the
environment. There exist several theories to explain different forms of constancy,
as discussed in the following sections.

5.8.1 Adaptational Theories

One of the mechanisms that partially aid color and lightness constancy is adapta-
tion. If it is assumed that the lighting in an environment is constant, then adapta-
tion to this lighting will subdue the active perception of it. Therefore, the lighting
in the scene is automatically discarded by retinal mechanisms [517, 622].

i i

i i
 i i

i i

288 5. Perception

Figure 5.32. The check shadow illusion. Although the two squares marked A
and B have the same luminance, they are perceived differently due to the pres-
ence of a shadow (left). If the illusion of a shadow is removed, then the
two checks appear the same (right). (Images courtesy of Edward H. Adelson,
(http://web.mit.edu/persci/people/adelson/checkershadow illusion.html).)

Though adaptation may aid constancy, it cannot give a full account of color
and lightness constancy. Adaptation theories are only valid to explain global
changes in lighting conditions; they do not address local conditions such as the
change in lighting due to shadows in the same environment. In particular, the
simultaneous effects in Figure 5.31 and 5.32 are not well explained by adaptation
theories.

5.8.2 Relational Theories

It is a well-known fact that the human visual system is largely sensitive to differ-
ences, contrasts, and changes over time. Sensitivity to absolute values is far less
dominant than sensitivity to relative values. As relative values can be inferred
from the retinal input only, no prior knowledge is required, nor are absolute val-
ues important. In addition, the ratio of intensity values received by the retina
remains invariant under changing lighting conditions. This invariance is the ba-
sis for relational theories, which propose that relative values convey information
regarding surface reflectance [331].
This notion was first introduced by Hering and formulated in a testable form
by Wallach [1196], who suggests that perceived lightness is determined by ratios
of light entering the eyes. In Wallach’s experiment, a white screen receives light
from two different projectors, such that the light from one of the projectors creates
a disk on the surface, and the second projector displays a ring around that disk (see
Figure 5.33). The disk and the ring are displayed at different intensities.
This set-up is then duplicated, albeit with the ring displayed at a different in-
tensity than the first ring. The intensity of the disk is controllable by the observer.

i i

i i
 i i

i i

5.8. Lightness Constancy 289

Figure 5.33. The experimental setup of Wallach’s experiment. Subjects were asked to
modify the light of the circle on the right so that it was perceived to have the same re-
flectance as the circle on the left (after [1196]).

Participants in this experiment are then asked to adjust the intensity of the second
disk such that its reflectance is perceived to be identical to the first disk.
If participants are able to perceive absolute values, then the values of the two
disks should be matched. If participants are only able to perceive relative values,
then the ratios of the values of the disk and ring should be matched. Wallach’s
experiment revealed that participants tended to match ratios rather than absolute
values. This gives weight to the hypothesis that ratios of light entering the eye are
important in the perception of lightness.

5.8.3 Filling-In
Given that relative contrast is important to human vision, the question arises at
which spatial scale such computations may occur. In particular, such computa-
tions could be carried out on either a local or global scale. Visual phenomena,
such as the Cornsweet-Craik-O’Brien illusion imply that this computation may
be local (see Section 5.4.4).
In this illusion, the left and right quarters of the image have identical values.
Nevertheless, these parts of the image are perceived to be different due to the
presence of an edge. This edge is easily perceived, but the gradual ramps directly
adjoining the edge contain no C1 -discontinuities, and they are therefore not no-
ticed. As a result, the change in value due to the edge is propagated along the
surface; hence, the perceived intensity difference between left and right sides of
this illusion.
The propagation of edge information to the interior of surfaces is called filling-
in [280, 395, 616, 1197]. This process is under-constrained, since many different

i i

i i
 i i

i i

290 5. Perception

ways to reconstruct a scene from edge information are possible. It is thought that
the human visual system makes specific assumptions about the world to enable
a reasonable reconstruction. For instance, it may be assumed that the luminance
distribution between sharp discontinuities is more or less uniform. This leads to
a computational model of filling-in which employs a (reaction-) diffusion scheme
to propagate edge information into the interior of uniform regions [830].
A different useful assumption is that the real world exhibits the well-known
1/ f image statistic, which is discussed further in Section 15.1 [52, 55, 58, 82, 83,
144, 172, 259, 313, 314, 651, 878, 879, 983, 985, 986, 1001]. Given that low spatial
frequencies are attenuated in the edge signal, but are not completely absent, a
signal may be reconstructed by boosting low frequencies according to this 1/ f
statistic [228].

5.8.4 Retinex Theory

The filling-in procedure described in the preceding section provides a model for
the global perception of intensities derived from edge information. The con-
trast ratio produced by the edge is then a cue for the relative lightness level at
a global scale.
An alternative model for the propagation of locally computed contrast ratios
is afforded by the retinex model [647]. In smooth regions, local contrast is small
enough to fall below the detection threshold, and thus smooth gradients are dis-
counted. Contrast ratios are then propagated over large regions by multiplication
of all the ratios that are above threshold.
This model may be tested by the construction of an appropriate psychophys-
ical test. To this end, Land and McCann constructed 2D displays of uniform
rectangular patches. These displays are called mini-worlds of Mondriaan. These
displays are lit with a light source creating a spatially slowly varying illumination.
In this set-up, regions of Mondriaans with low surface reflectance could possibly
reflect more light than regions with higher reflectance. Figure 5.34 shows a sim-
plified Mondriaan along with the computations that Land and McCann suggested
may be taking place in the HVS.
Through these informal experiments, they were able to show how humans
are capable of perceiving relative reflectances of surfaces, despite slowly vary-
ing scene lighting and regardless of the absolute values of light in the scene.
They proposed the retinex theory, which suggests that the HVS has a densely
connected structure of systems. This structure computes products of ratios as de-
scribed above. Retinex theory is used in several applications, including shadow
removal from images, which is discussed in Section 5.12.

i i

i i
 i i

i i

5.8. Lightness Constancy 291

Illumination
Luminance
32 40
A
80% Reflectance
48
60
24
B
40% Reflectance
32
80
16
B
20% Reflectance
20 100
Luminance edge calculation (A to C):
(48/24) (32/16) = 4/1

Figure 5.34. Retinex theory proposes an explanation for how humans achieve color con-
stancy despite gradual changes in incident light falling on surfaces (after [647]).

5.8.5 Anchoring
While human vision is predominantly driven by relative values, the absolute im-
pression of how light or dark an environment, or part thereof is, also plays an
important role. The Retinex theory offers no model for how humans are able to
perceive absolute values. To compute absolute values given the relative values
from the Retinex theory, several heuristics may be applied. Finding the absolute
values from relative values is termed the scaling problem.
An example of a heuristic that may be involved includes the anchoring heuris-
tic, which maps the surface with the highest reflectance to white [679]. Evidence
for this heuristic can be demonstrated by showing a range of gray values between
black and dark gray. The lightest of these will be perceived as white, whereas the
black will be perceived as gray [158]. A related example is shown in Figure 5.35,
where in the left image, a piece of paper is shown that appears white. When suc-
cessively whiter sheets are added, the original piece of paper will appear less and
less white. This is known as the Gelb effect and is an example of a situation where
lightness constancy breaks.
In addition to a tendency for the highest luminance to appear white, there is
a tendency for the largest surface area to appear white [363]. If the largest area
also has the highest luminance, this becomes a stable anchor which is mapped
to white. However, when the largest area is dark, the highest luminance will be
perceived as self-luminous.
In all likelihood, a combination of both heuristics is at play when anchoring
white to a surface in an environment. In addition, the number of different sur-

i i

i i
 i i

i i

292 5. Perception

Figure 5.35. The paper in the left image appears less and less white when whiter paper is
added to the scene.

faces and patches in a region affects the perception of white. This is known as
articulation. A higher number will facilitate local anchoring (see Figure 5.36).
Finally, insulation is a measure of how different surfaces or patches in a scene are
grouped together [363]. The patches surrounding the test square in Figure 5.36,
for instance, may be considered a local framework, and the rest of the page may
be considered a global framework.
In general, frameworks can be viewed as regions of common illumination. In
addition, proximity is an important grouping factor. The latter essentially means
that nearby regions are more likely to be in the same framework than distant
regions.
Another anchoring rule may be that the average luminance in an environment
is mapped to middle gray [456]. Similar rules are used frequently in photography,
as well as in tone reproduction [950] (see Chapter 17). However, experimental ev-
idence suggests that anchoring to highest and lowest luminance as well as largest
area provides a better model of human visual perception [362, 679]. A compu-
tational model of lightness constancy was implemented for the purpose of tone
reproduction, as discussed in Section 17.8.

Figure 5.36. The patches in the centers of all four groups have the same reflectance. The
right pair of patches look somewhat different because of their surrounds. If the surrounds
are made of a larger number of patches, the effect of lightness difference is enhanced (left
pair), even though the average intensity of light reaching the eyes from the surrounds in
both groups is the same (after [13]).

i i

i i
 i i

i i

5.8. Lightness Constancy 293

Once an anchor is found, all other surface reflectances can then be derived
from the lightest surface by relative Retinex-like computations [647].

5.8.6 Shadow and Reflectance Edges

Images contain edges, usually at different spatial frequencies. The sharpness of an
edge may be a cue to the visual system as to what type of edge it is. In particular,
edge sharpness may help in determining if the edge signals a change in surface
reflectance or in illumination. The Retinex theory, however, does not distinguish
between different types of edges.
To determine whether it matters that an edge is of a particular type, an elegant
experiment may be devised where participants match a test patch against a set
of reference patches [364]. The test patch is placed in the shade, whereas the
reference patches are directly illuminated by a light source. Thus, there exists a
shadow edge between the test and reference patches, as shown in Figure 5.37.
Observers participated in the matching experiment twice. During one set of
the trials, a mask was placed in front of the observer such that it was not possible to
determine the nature of the edge. In the other set of trials, the mask was removed,
and the observers were made aware that the edge was a shadow edge.

Observer’s view
without mask
Bird’s eye view of
experimental set-up

Display

Shadow
Occluder
Observer’s view
with mask

Removable
mask

Observer

Figure 5.37. The set-up of Gilchrist’s experiment to determine the importance of shadow-
edge identification for constancy. Subjects were asked to select from the squares on the
right, the one that best matched the reflectance of the square on the left. They performed
the experiment twice. Once without the mask (top right), so they knew the edge in the
center was a shadow edge, and once with the mask (bottom right), so they did not know
what type of edge it was (after [364]).

i i

i i
 i i

i i

294 5. Perception

Equal ratio match Equal luminance match

Without mask
3
2
1

3 With mask
2
1

Munsell value

Figure 5.38. The results from Gilchrist’s experiment. When subjects were aware that
the edge in the middle was a shadow edge, they were able to ignore it and select a square
such that its reflectance was almost identical to the original patch. When this information
was not available, they chose squares that had the same amount of light reflected off their
surfaces as the original square. Thus, lightness constancy failed when the subjects were
unaware that an edge was a shadow edge (after [364]).

When the subjects knew that the edge was due to a change in incident light,
they chose squares from the candidate squares that were very close to the actual
reflectance of the original square. However, when they were not aware that the
edge was a shadow edge, they chose squares that reflected the same intensity of
light as the original square, as shown in Figure 5.38.
These results imply that the human visual system is able to ignore shadow
edges, but not reflectance edges. This, in turn, implies that we may encode light-
ing information separately from reflectance information by computing local con-
trasts separately in light maps and reflectance maps.

5.8.7 Distinguishing between Shadow and Reflectance Edges

Given the fact that knowledge of the type of edges present in an environment
mediates lightness perception, we may wonder how the human visual system de-
termines that an edge is of a particular type. There exist several heuristics that
may be used to discriminate between the two kinds of edges, of which three are
discussed here
In the absence of any other distinguishing information, the HVS may decide
on the nature of an edge based on how sharp it is. Fuzzy edges are typically due to
changes in lighting rather than reflectance. This is the reason why the sharp edges
in Wallach’s experiment discussed earlier were interpreted as reflectance edges.

i i

i i
 i i

i i

5.9. Color Constancy 295

Sometimes, a sharp edge may be perceived as a luminance edge. This is due

to another heuristic employed by the HVS. If depth information available to the
HVS shows that two regions are not coplanar, the edge dividing them, even if it is
sharp, is interpreted as a shadow edge. Such edges explain the different amounts
of light falling on the different planes of an object, due to its shape and position.
The Retinex theory implies that the HVS keeps track of ratios across edges.
If these edges are caused by changes in reflectance, then the maximum ratio is
obtained if the edge is adjoined by a black and a white surface. A white surface
typically does not reflect more than 90 percent of the incident light, whereas black
reflects no less than about 10 percent of the incident light [868]. Therefore, if a
ratio across an edge is more than 9:1, then the edge is most likely caused by
changes in lighting.

5.9 Color Constancy

The chromatic equivalent of lightness constancy is called color constancy. Color
constancy is not studied in as much detail as lightness constancy, and conse-
quently, less is known about how color constancy is achieved by humans. To
achieve color constancy, on the basis of the photoreceptor input, reflectance and
illuminance need to be disentangled. With three types of photoreceptor, there are
six unknowns corresponding to the reflectance and illuminance for each of the
three response functions.
As with lightness constancy, it is assumed that chromatic adaptation plays a
partial role in discounting effects such as the chromatic content of the lighting
in the scene [1227, 1300]. However, the higher-dimensionality of the problem
affords additional assumptions to be made about the environment and, thus, aids
in the interpretation of the environment.

5.9.1 Color Constraints

The environment around us typically contains certain unvarying features which
make it easier to use heuristics to infer information from the scene. One of these
is consistency of light within a single scene. Usually, scenes are lit by a specific
spectrum of light from a single type of light source. This implies that the intensity
of the light may vary across a scene, but not its spectrum [440]. A recent proposal
exploits this by assuming that the illumination can be modeled by blackbody ra-
diation, using Planck’s law to solve the color constancy problem [583].
Another important constraint is that there is only a limited number of differ-
ent types of light spectra in typical scenes. The spectrum of light in an outdoor

i i

i i
 i i

i i

296 5. Perception

environment is different depending on whether there are clouds, what time of day
it is, etc. However, these changes are relatively minor. For indoor environments,
a similar argument can be made. Here, there are several different kinds of light
sources, but their variation is not significant.
Finally, only a limited number of different surface reflectances are typically
encountered [233]. Thus, the limited variability in lighting and surface reflectance
may be exploited by the human visual system to achieve color constancy.

5.9.2 Shadow and Reflectance Edges

Color is a useful attribute for discriminating between shadow and reflectance
edges [868]. An interesting difference to note between shadow and reflectance
edges is their change in spectrum across the edges. Across a shadow edge, there is
typically a sharp change in intensity, but not in the shape of the spectrum. Across
a reflectance edge, on the other hand, there is typically a change in the shape of
the spectrum (see Figure 5.39).
Thus, the absence or presence of changes in spectral composition across an
edge may be exploited to determine the edge’s type [982]. Given knowledge of
the intensity at two different wavelengths at either side of the edge, the type of
edge may be inferred. The two known wavelengths on either side define both
slopes as well as spectral cross-points.
Figure 5.40 shows the four possible configurations. The diagram at the top left
illustrates the case in which the measured intensity of wavelength 1 has increased
across the edge while intensity of wavelength 2 has decreased. Since there is
Photons

Photons

Shadow edge Reflectance edge

Region 2
Region 1

Region 2

Region 1

400 700 400 700

Wavelength (nm) Wavelength (nm)

Figure 5.39. Possible changes in spectra across shadow and reflectance edges (af-
ter [868]).

i i

i i
 i i

i i

5.9. Color Constancy 297

Photons

Photons
Region 2
Region 1

Region 2
Region 1

400 700 400 700

Wavelength (nm) Wavelength (nm)
Photons

Region 2 Photons Region 1

Region 1 Region 2

400 700 400 700

Wavelength (nm) Wavelength (nm)

Figure 5.40. The four different cases in the edge-classification algorithm proposed by
Rubin and Richards [982].

incomplete information about the spectrum, it is approximated with a line joining

the two known measurements. In such a situation, the slopes intersect, indicating
that the spectra change not only in intensity across the edge, but also in shape.
This spectral crosspoint indicates the presence of a reflectance edge.
If the shape of the spectrum has in fact changed across the edge, then the
slopes of the spectra on either side of the edge may also be of different signs,
as shown in the diagram on the top right. Opposite slope signs also indicate the
presence of reflectance edges.
Thus, if either (or both) of these features are present across an edge, as in
all cases except for the diagram at the bottom right, the edge is classified as a
reflectance edge. If neither are present, as in the diagram at the bottom right, then
the shape of the spectra is assumed to be similar enough on either side of the edge
to classify it as a shadow edge.

5.9.3 Applications of Color Constancy

Algorithms modeling color constancy attempt to retrieve the perceived surface
color of an object projected onto the photoreceptors or camera sensor, indepen-

i i

i i
 i i

i i

298 5. Perception

dent of the illuminant [317, 319, 474]. Recent evaluations of collections of com-
putational color-constancy algorithms were presented by Barnard et al [66] and
Hordley and Finlayson [487]. Such algorithms are important in situations where
the illumination is not controlled, for instance, in remote-sensing applications,
guiding robots, as well as color reproduction [1203, 1204]. It also finds practical
application in matching paints [474]. Finally, Retinex theory has been adapted to
solve the color constancy problem [738, 965].

5.10 Category-Based Processing

Humans tend to classify colors into various categories. Each of these categories
is accorded a name, and all the colors in that category, even though they generate
different response patterns in the retina, may be called that name. Names of some
commonly used categories include red, green, blue, and yellow. Objects such as
leaves, grass, and unripe fruit are all called green, even though they have different
reflectance spectra. This categorization may be viewed as dividing the three-
dimensional color space into several different subsets and giving each subset a
different name.
Given our ability to categorize color, several questions arise. When we ob-
serve a color, why do we immediately classify it as a specific category and give it
a label? Why have we divided the space of colors into various regions with labels?
Is it because we were taught these labels in our childhood and, therefore, learned
to classify all colors according to these labels? Or is there some physiological
reason for these specific categories? Various experiments have been conducted to
study the nature of color categorization, which indicate that the physiology of the
human visual system determines to a large extent how we categorize color. We
will discuss these in the remainder of this section.

5.10.1 Color Naming

Before any cross-cultural research had been performed, color naming was thought
to be an example of cultural relativism, which is the theory that a person’s culture
influences his or her thoughts and experiences. It implies that two people from
different cultures think about the world differently. In linguistics, this hypothesis
is named the Sapir-Whorf hypothesis after two linguists who made it famous,
Edward Sapir and Benjamin Lee Whorf.
In terms of color naming, cultural relativism implies that a name for a particu-
lar kind of color emerges in a language as its need develops in the society speaking
that language. One example of this theory is that the Inuit, whose surroundings

i i

i i
 i i

i i

5.10. Category-Based Processing 299

are predominantly covered with snow, and who thus have a greater need to dis-
tinguish between different kinds of snow, have more than a dozen names for it.
Cultural relativists further believed that this was the only factor influencing the
naming of colors in different languages, and that human physiology played no
part in determining it. Hence, there would be no correlation in the division of
color space into various categories across different languages.
The first cross-cultural research to study the nature of color categorization was
performed by Brent Berlin and Paul Kay in 1969 [87]. This work was prompted
by their intuitive experience in several languages of three unrelated stocks. They
felt that color words translated too easily among various pairs of unrelated lan-
guages for the relativity thesis to be valid.
Basic color terms are names of colors with the following properties. First,
they are single lexical terms such as red, green, or pink; not light-green or
greenish-blue. Second, they refer to the color of objects, rather than objects them-
selves. Therefore, colors such as bronze and gold are not valid color terms. Third,
the colors are applicable to a wide range of objects, so that colors like blonde are
not valid. Fourth, the colors are in frequent use.
Berlin and Kay use an array of Munsell color chips (see Section 8.9.1) to
determine the references of basic color terms of a language. For each basic color
term, participants in the experiment are asked to specify the best example of the
color, as well as the region of chips on the array that can be called by the same
name as the basic color term. Figure 5.41 shows a copy of the Munsell Chart that
was shown to the participants.
Berlin and Kay experimentally examined 20 languages directly and 78 more
through literature reviews. They discovered two interesting trends. First, there

Figure 5.41. An illustration of the Munsell Chart that was shown to partici-
pants in the work by Berlin and Kay [87]. (See also the World Color Survey,
http://www.icsi.berkeley.edu/wcs/.)

i i

i i
 i i

i i

300 5. Perception

Type Basic terms White Black Red Green Yellow Blue Brown Pink Purple Orange Gray
1 2 • • ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
2 3 • • • ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
3 4 • • • • ◦ ◦ ◦ ◦ ◦ ◦ ◦
4 4 • • • ◦ • ◦ ◦ ◦ ◦ ◦ ◦
5 5 • • • • • ◦ ◦ ◦ ◦ ◦ ◦
6 6 • • • • • • ◦ ◦ ◦ ◦ ◦
7 7 • • • • • • • ◦ ◦ ◦ ◦
8 8 • • • • • • • • ◦ ◦ ◦
9 8 • • • • • • • ◦ • ◦ ◦
10 8 • • • • • • • ◦ ◦ • ◦
11 8 • • • • • • • ◦ ◦ ◦ •
12 9 • • • • • • • • • ◦ ◦
13 9 • • • • • • • • ◦ • ◦
14 9 • • • • • • • • ◦ ◦ •
15 9 • • • • • • • ◦ • • ◦
16 9 • • • • • • • ◦ • ◦ •
17 9 • • • • • • • ◦ ◦ • •
18 10 • • • • • • • • • • ◦
19 10 • • • • • • • • • ◦ •
20 10 • • • • • • • • • ◦ •
21 10 • • • • • • • • ◦ • •
22 11 • • • • • • • • • • •

Table 5.1. The 22 possible subsets of the 11 basic color terms.

appear to be just 11 universal basic color terms.5 This small number is due to the
fact that color categories are shared by languages. The 11 basic colors are: white,
black, red, green, yellow, blue, brown, purple, pink, orange, and gray.
Second, there are only 22 different subsets of these basic color terms from a
total of 216 − 1 possibilities. If a language does not have all of these basic color
terms, then it has a specific subset of these color terms. For example, if a language
has only two basic color terms, then these color terms are the equivalent of the
English color names white and black. Table 5.1 shows the possible subsets that
are found across all these languages. Based on this result, it is hypothesized that
as languages develop, they acquire the basic color terms in a specific order that
can be termed evolutionary.
These results indicate that there are physiological reasons for the very specific
evolution of languages in terms of color. It is worth noting that colors important
in Hering’s opponent theory (red, green, blue, yellow, white, and black) are also
5 This number was increased to 16 by later studies [584].

i i

i i
 i i

i i

5.10. Category-Based Processing 301

some of the first colors to be recognized in languages as separate categories. On

the basis of these results, the Sapir-Whorf hypothesis has been severely under-
mined. There is now concrete evidence for the existence of semantic universals
in the domain of color vocabulary.
Research on color naming continues to this day [391, 945], and the reader is
encouraged to visit the World Color Survey at www.icsi.berkeley.edu/wcs/.

5.10.2 Focal Colors and Prototypes

In the course of their research, Berlin and Kay also discovered the existence of
best examples, or focal colors, in color categories. To find out which basic color
terms best describe any particular color, they use two different approaches. In one
approach, they obtain information about the boundary of color terms by showing
participants various colors and asking them to point out those colors that lie on the
boundary of a specified color term. In the second approach, they show colors and
ask participants to identify the colors that best describe a specified color term.
Participants invariably give faster and more consistent responses in the second
approach. This implies that color categories are structured around focal colors,
rather than around colors at the boundaries of color terms.
This idea received more support from several experiments conducted by
Eleanor Rosche and colleagues [450, 978]. She performed a number of experi-
ments with members of the Dani tribe in New Guinea as her participants. The
Dani have just two basic color terms in their language. In one experiment, the
Dani are shown different colors and later asked to recognize them from a set of
colors. The Dani point out focal colors faster and more reliably than those lying
at the boundaries of color categories. In a similar experiment, the Dani’s learn-
ing ability is tested in terms of colors. They learn focal colors more quicker and
more accurately than colors at the boundaries of categories. These and other re-
sults form mounting evidence that focal colors play a crucial role in our internal
representation of color categories.
In light of these discoveries, Kay and McDaniel [584] propose a model for
color naming which is based on fuzzy set theory. They argue that since people
are able to find and remember focal colors more easily than boundary colors,
focal colors are better members of their color categories. Thus, each category
can be seen as a set, such that membership of a particular color is defined by a
number ranging from 0 to 1, where 1 indicates that the color is a focal color, and
0 indicates that the color does not belong to that specific color category. A color
may be a member in more than one category. Bluish-green for instance, would be
a member of both blue and green categories. The membership of such a color in
these categories is represented by fractions between 0 and 1.

i i

i i
 i i

i i

302 5. Perception

This model predicts that it is easier to reliably detect and remember focal
colors. A focal color clearly belongs to one category, in which its membership
value is high. For colors on the boundary, belonging to more than one category,
a decision has to be made regarding which category they belong to. This could
possibly make the task of detection and recognition less reliable.

5.11 Color Anomalies

Individuals with normal color vision are called trichromats, referring to the triplet
of cone types present in a normal healthy retina. About 8% of all males, how-
ever, suffer from some color vision deficiency [128]. Common visual anomalies
affecting color vision fall in two dominant classes. The first class is anomalous
trichromacy and is characterized by the presence of one or more photopigments
that have a peak sensitivity different from normal opsins [30–33].
There are three different types of anomalous trichromats. In a protanomalous
trichromat, the absorption spectrum of the L-Cone (erythrolabe) is displaced. In a
deuteranomalous trichromat, the absorption spectrum of the M-Cone (chlorolabe)
is displaced. Finally, in a tritanomalous trichromat, the absorption spectrum of
the S-Cone (cyanolabe) is displaced. These displacements of cone photopigments
from their optimal positions result in deficient color discrimination [1019].
Apart from displaced absorption spectra, anomalous trichromats may have
weaker absorption spectra, leading to an anomaly called color weakness. Such
individuals require a more intense stimulus to perceive the same color.
The second class of anomalies is called dichromacy. Dichromatic individuals
perceive the world as mixtures of only two basic colors. This anomaly is also
related to the photopigments of the cones in the eye. A protanope is missing
erythrolabe (L-cones are affected), a deuteranope is missing chlorolabe (affecting
the M-cones), and a tritanope is missing cyanolabe (affecting the S-cones). An
overview is given in Table 5.2.
The terms protan, deutan, and tritan refer to the affected photo-pigment.
Protans and deutans tend to confuse reds and greens. These defects are usu-

Red-green defects Yellow-blue defects

Photopigment status Protan defect Deutan defect Tritan defect
Dichromat Protanope Deuteranope Tritanope
Anomalous trichromat Protanomalous trichromat Deuteranomalous trichromat Tritanomalous trichromat

Table 5.2. A table listing the different types of dichromacies and anomalous trichromacies.

i i

i i
 i i

i i

5.11. Color Anomalies 303

ally, but not always, inherited. Tritans tend to confuse blues and yellows, a defect
that is almost always acquired. Figure 5.42 and Figure 5.43 show images as they
appear to a normal trichromat, as well as how they might appear to the different
types of dichromats. These images were created with Vischeck, a freely available
program to treat images so that they appear as they would to people with different
color anomalies. Computer simulation of color deficiencies is further discussed
by Brettel et al [129].
Achromotopsias are rare anomalies in which the affected person behaves
largely as a monochromat [3, 910]. As the name implies, monochromats see the
world as shades of a single color. The most common form of achromatopsia
is rod monochromacy, in which individuals have no cones. Rod monochromats
do not show the Purkinje effect (see Section 6.1), have achromatic vision, low
visual acuity, high sensitivity to bright lights, nystagmus (involuntary eye move-
ment), and macular dystrophy. It is a color deficiency that occurs in about 1 in
30,000 [236, 329].

Figure 5.42. An image as it would be perceived by normal trichromats (top left),

protanopes (top right), deuteranopes (bottom left), and tritanopes (bottom right); Insel
Mainau, Germany, August 2005. (Images created with the freely available Vischeck pro-
gram (www.vischeck.com)).

i i

i i
 i i

i i

304 5. Perception

Figure 5.43. An image as it would be perceived by normal trichromats (top left),

protanopes (top right), deuteranopes (bottom left), and tritanopes (bottom right); Zaragoza,
Spain, November 2006. (Images created using Vischeck (www.vischeck.com).)

Another form of achromatopsia is cone monochromacy, which occurs in about

1 in one-hundred million humans. Here, the discrimination of colored lights is
lost if their luminances are equalized. Visual acuity is unaffected. Although the
name suggests that cone monochromats have only one type of functioning cone,
experiments have shown that it is more likely of cortical origin [316, 329, 361].
Further, chromatopsias are color anomalies that may not be termed color de-
fects, as they do not result in decreased ability to discriminate colors. They simply
present a distortion of color vision, much like looking through a colored filter. As
an example, such anomalies may be caused temporarily due to cataract extrac-
tion. A particular kind of cataract, known as a nuclear cataract, behaves like a
yellow filter, thereby absorbing blue light. Since the individual with the cataract
has grown accustomed to having less blue light incident on the retina, he or she
becomes extremely sensitive to blue stimuli over time. When the cataract is finally
removed and the individual is exposed to a “bluer” spectrum, he or she usually
reports a blue cast on the whole scene. As might be expected, this perception
fades away with time.

5.11.1 Spectral Sensitivity

There are various ways in which the differences in color perception by people with
color anomalies may be visualized and compared with the perception of normal
trichromats.

i i

i i
 i i

i i

5.11. Color Anomalies 305

Luminous efficiency
1.0
Trichromacy
Deuteranopia
Protanopia

0.5

0.1

440 480 520 560 600 640 680

Wavelength (nm)

Figure 5.44. Achromatic response function for normal trichromats, protanopes, and
deuteranopes (after [516]).

One such mechanism was introduced in Section 5.3.3 and quantifies an ob-
server’s spectral sensitivity through a nulling method [516]. When the experiment
is performed by an observer with a color anomaly, the different spectral sensitiv-
ity of this observer may be compared with that of a normal observer, resulting
in the achromatic response functions shown in Figure 5.44. The protanope’s re-
sponse function is significantly different from the trichromat’s and deuteranope’s
response function. Figure 5.45 shows the theoretical chromatic and achromatic
responses of a protanope and a deuteranope. In this case, there is no perception
of reds and greens, leading to only a single chromatic response.

1.0
Relative visual responses

0.0

Deuteranopia
Protanopia
White
Yellow
Blue
-1.0
400 500 600 700
Wavelength (nm)

Figure 5.45. Spectral response for dichromats with red-green defect (after [516]).

i i

i i
 i i

i i

306 5. Perception

Δλ (nm)
Trichromacy
14 Deuteranopia
Protanopia

10

450 500 550 600

Wavelength (nm)

Figure 5.46. Wavelength-discrimination functions for protanopia and deuteranopia [903].

Δλ (nm)

Trichromacy
Tritanopia

450 500 550 600 650

Wavelength (nm)

Figure 5.47. Wavelength discrimination functions for tritanopia [1254].

5.11.2 Wavelength Discrimination

To visualize the differences between normal vision and dichromatic vision, the
wavelength-discrimination functions for individuals may be plotted, as shown in
Figure 5.46 and Figure 5.47. Protanopes and deuteranopes respond like mono-
chromats for wavelengths greater than about 545 nm. In contrast, tritanopes dis-
criminate well for wavelengths above 545 nm. However, they behave as mono-
chromats for wavelengths centered around 490 nm.

5.11.3 Pseudoisochromatic Plate Tests

A variety of color vision tests have been developed to differentiate between nor-
mal trichromats and those with various color anomalies. Among these, the sim-
plest to perform and the most easily available is the pseudoisochromatic plate test,
also known as the Ishihara test [531] . This test consists of a booklet of different

i i

i i
 i i

i i

5.11. Color Anomalies 307

plates. Each plate is drawn such that its pattern is only detectable by observers
with normal color vision.
The Ishihara test set exists in two variations, one consisting of 24 plates and
one consisting of 34 plates. If administering an Ishihara test, it is recommended
to use plates from the larger set, as the 24-plate set contains relatively few reliable
plates. Both sets contain two groups of plates. The first group is suitable for
numerates, containing numerals painted in a different color from the background.
The second group is suitable for innumerates, contains snaking patterns, but is
rarely used because the results are much more difficult to assess, and these plates
are less reliable.
The plates containing numerals can be classified into four colorimetrically
different groups:
Transformation plates cause anomalous color observers to give different answers
than those with normal vision. Plates 2–9 are in this group (of the 34-plate
set).
Disappearing digit plates, also known as vanishing plates contain numerals that
can only be observed by those with normal color vision. For color-deficient
viewers, these plates appear as a random dot pattern. These are Plates 10–
17.
Hidden digit plates are designed such that anomalous observers will report a nu-
meral, whereas normal observers do not see any pattern. Plates 18–21 are
in this group.
Qualitative plates are designed to assess the severity of color blindness and sep-
arate protan from deutan color perception. These are Plates 22–25.
Of these four types, only the transformation and disappearing plates are reliable
and should be used to test for color blindness. Figure 5.48 shows examples of all
four plate types.
It should be noted that the Vischeck simulation applied to the hidden-digit
plate does not reveal the number that was intended. Hidden-digit plates are typi-
cally not very reliable and, in addition, the Vischeck algorithm may not be opti-
mally matched to this particular stimulus. It is included for completeness.
A typical color plate test procedure begins by showing the observer Plate 1
(see Figure 5.49), which is a demonstration plate. If the observer does not give
the correct answer for this plate, then the remainder of the test should not be
administered.
To interpret the results, the false positives and false negatives should be taken
into consideration. The false-positive rate varies significantly with age, and to a

i i

i i
 i i

i i

308 5. Perception

Figure 5.48. Ishihara test plates from the 24-set to determine color blindness. From left
to right: Ishihara plate, same plate as seen by a protanope, a deuteranope, and a tritanope.
Top to bottom: a transformation plate, a disappearing plate, a hidden-digit plate, and a
qualitative plate. Note that due to color reproduction issues as well as viewing conditions,
these reproductions are not suitable to determine color blindness. (Test plates from [531].
Color anomalies simulated using Vischeck (www.vischeck.com).)

lesser extent with education [638]. The false-negative rate depends on the type of
anomalous vision, but tends to be low. The question then is where to place the
cut-off, i.e., how many errors before we can decide that someone is color normal?
The answer depends on many factors. For a male population, a good balance
between false positives and false negatives is achieved by placing the cut-off be-
tween four and six errors, provided Plates 2–17 are used (in particular excluding
the hidden-digit plates).
For females, the occurrence of anomalous color vision is lower, meaning the
cut-off point would have to be chosen differently, and even then the interpretation

i i

i i
 i i

i i

5.12. Application: Shadow Removal from Images 309

Figure 5.49. Demonstration test. This numeral should be visible by both color blind
observers as well as color normal observers. (Test plate from [531].)

of the results would be more complicated. In addition, as color deficiencies in fe-

males may have a somewhat different nature, the Ishihara test is not recommended
as a stand-alone test, but should always be augmented with other tests.
If the hidden-digit plates are included, then the interpretation of the results
also becomes much more difficult. The results obtained with this group of plates
varies significantly with age and with degree of anomaly.
In summary, the Ishihara test is suitable for quick and convenient screening of
individuals. For any one suspected to have anomalous color vision, further tests
should always be carried out, as failing the Ishihara test does not automatically
mean that the observer is color deficient. In addition, the Ishihara test is not
suitable for determining the degree of color deficiency.

5.12 Application: Shadow Removal from Images

Retinex theory, introduced in Section 5.8.4, is the basis for several applications
(see [332] and [340] for implementations). For instance, it has been used success-
fully for dynamic range reduction [553, 775], the topic of Chapter 17. Retinex
theory has also been used in algorithms addressing the issues of color correc-
tion [64, 553, 737], gamut mapping [759], and image enhancement [774].
In this section, we will discuss an algorithm that employs Retinex theory to
remove shadows from images [320]. Shadow removal in an image is useful for
a variety of tasks. Several tasks, such as image segmentation and object tracking
and recognition, might perform better on an image from which the effects of scene
lighting have been removed. Often these algorithms are fooled by the presence
of shadow edges which they mistakenly respond to as reflectance edges. If shad-
ows are removed first, then image data will conform more closely to assumptions
made by these algorithms. Removing lighting effects from an image may also be
considered as the first step in scene relighting, a task in which images of a scene
are re-rendered with modified lighting.

i i

i i
 i i

i i

310 5. Perception

Retinex theory explains how slowly varying lighting may be discounted by

the human visual system and may, therefore, be used to remove some lighting
effects from an image. The algorithm presented in this section adapts this theory
so that it can be used to remove the effects of sharp variations in lighting as well.
Location of shadow edges needs to be specified as input into the algorithm. This
information may be marked by hand, or may be computed automatically in some
cases, by using existing techniques [317].
In the following, we first discuss an algorithm for Retinex theory (Section
5.12.1) and then detail the changes that are made to this algorithm to enable the
removal of shadows (Section 5.12.2).

5.12.1 Retinex Algorithm

Given an image as input, the general concept of Retinex theory, as discussed in
Section 5.8.4, may be implemented in several different ways [340, 553, 647–649,
737]. Land’s path-based algorithm is one such approach that lends itself easily to
adaptation, so that shadows may be removed [647].
Land proposed that a region’s reflectance, relative to the region in the scene
with the highest reflectance, may be obtained by computing the product of ratios
along a path starting from the region with the highest reflectance and ending at
the region whose relative reflectance is being computed. For instance, if the path
starts from pixel p1 , goes through pixels p2 , p3 , and p4 , and ends at p5 , then the
relative reflectance Θ1,5 at p5 with respect to p1 may be computed in the following
manner:
p5 p2 p3 p4 p5
Θ1,5 = = . (5.12)
p1 p1 p2 p3 p4
If products of ratios are computed in log space, ratios become subtractions
and products become summations. Therefore, the relative reflectance becomes
 
Θi, j = exp ∑ log pk+1 − log pk , (5.13)
k∈P,k< j

where Θi, j is the reflectance of pixel j with respect to pixel i and P is the path
from pixel i to pixel j. This computation requires knowledge about the location
of the pixel with the highest reflectance. Since the value at a pixel is affected
by its reflectance, as well as the intensity of incident light, it cannot be used to
determine the pixel with the highest reflectance. Such a pixel may only be found
with an exhaustive evaluation of all possible sequential products, comparing the
results and selecting the lowest value. This value corresponds to the ratio of pixels
with the highest and lowest reflectance in the scene.

i i

i i
 i i

i i

5.12. Application: Shadow Removal from Images 311

Land suggested that an approximation to the above procedure may be obtained

without an exhaustive search by computing products of ratios along several ran-
dom paths ending at the pixel whose relative reflectance is being computed. Each
path is assumed to start from a pixel with a 100% reflectance. Since this is ob-
viously incorrect for most randomly selected starting pixels, often the product of
ratios becomes greater than 1 somewhere along the path. This occurs every time
the starting pixel’s reflectance is lower than that of the current pixel on the path.
In such cases, the ratio computation is set to 1 and started afresh from the current
pixel. This pixel is now considered the new starting pixel of the path.
If several such paths are computed for each pixel, they produce different rel-
ative reflectances, since the starting pixel for each reflectance computation is dif-
ferent. However, averaging these relative reflectances gives a reasonable estimate
of the final relative reflectance of each pixel:
1 N
ρj = ∑ Θi, j ,
N i=1
(5.14)

where ρ j is the final relative reflectance of pixel j and N is the number of paths
used in the computation.
The ratios computed between different adjacent pixels may give some indi-
cation of whether the pixels are on opposite sides of an edge or belong to the
same uniform region. If the pixels have similar values, then the ratio will be close
to 1. In log space, this is equivalent to a value close to 0. Log ratios significantly
different from 0 will mean that an edge exists between the pixels. To eliminate
the effects of slowly varying lighting, a threshold t may be applied to the ratios.
Those ratios that are close to 0 will be set to 0, as shown below:

Θi, j = ∑ T (d pk ) , (5.15a)
k∈P,k< j

d pk = logpk+1 − logpk , (5.15b)


0 if |d pk | < t,
T (d pk ) = (5.15c)
d pk otherwise.

5.12.2 Shadow Edge Removal

So far, the algorithm for Retinex will only discount small changes in illumination.
If shadow edges are to be removed using the same technique, a higher value for
the threshold t has to be set. However, a higher value for t not only discounts
shadow edges, but some reflectance edges as well, giving rise to erroneous rela-
tive reflectances. Therefore, some mechanism is required to distinguish between
shadow and reflectance edges, so that the algorithm can react differently to them.

i i

i i
 i i

i i

312 5. Perception

A classification of edges can be given as input to the Retinex algorithm. This

can be done by providing a hand-marked binary image S as input, in which the
shadow edges are marked white and the rest of the image is black. Alternatively,
the binary image may be computed automatically in some cases, using existing
techniques [317]. If a shadow edge is encountered, then the ratio at that point can
be set to 0 in log space. The function T now becomes

0 if |d pk | < t or Sk = 1,
T (d pk ) = (5.16)
d pk otherwise.

By incorporating this change into the Retinex algorithm, sharp changes in

lighting will also be discounted, producing a representation of the scene approxi-
mately devoid of lighting effects [320].

5.13 Application: Graphical Design

Understanding the different forms of color deficiency is helpful in graphical de-
sign tasks, including the design of logos, brochures, and web pages. In this sec-
tion, we briefly list some of the design rules that may help to create safe color
combinations, i.e., designs that can be distinguished by both color normal and
color-deficient observers.6
There are certain color combinations that are preferred for the color of text
printed on a background and there are others that should be avoided, for various
reasons. In particular, black letters on a white background are always good. Black
letters on other backgrounds should mostly be avoided. In particular, black on red
or green backgrounds will be seen by some as black on black. On the other hand,
black on a yellow background is typically a good combination.
White lettering on a black background is also always acceptable. However,
other colors on a black background should be avoided. Blue tends to be a better
color than others, but it should not be used on a black background for fine detail,
since visual acuity for blue colors is worse in all types of color vision.
Mixed backgrounds should be avoided, as masking effects will make text dif-
ficult to read for anyone, but especially for color-deficient viewers.
Saturation also plays a role in graphical design for color-deficient viewers. In
general, more saturated colors will be better. However, for green dichromats, red
will be seen as yellow, and may therefore largely vanish against a white back-
ground. To accommodate green dichromats, it would be best to ensure that the
6 See www.firelilly.com/opinions/color.html.

i i

i i
 i i

i i

5.13. Application: Graphical Design 313

red is displayed against a background that would yield good contrast with yellow
as well. Making the text large and bold will also be beneficial.
A saturated red right next to a saturated blue is a special case. The wavelengths
of these colors are at opposite ends of the visible spectrum, and this causes the
focal length for lenses (including the ones in human eyes) to be distinctly different
for these colors. This may leave the impression that the text is not stationary, but
moves around with eye movements. This is due to continuous refocusing of the
eye for this color combination, even though the eye-object distance is identical.
This effect occurs for color-normal observers.

Figure 5.50. From left to right, top to bottom: input image, simulation of a deuteranope,
daltonized input image, and a simulation of how a deuteranope may perceive the daltonized
input image; Insel Mainau, Germany, August 2005. (Simulations created using Vischeck
(www.vischeck.com/daltonize).)

i i

i i
 i i

i i

314 5. Perception

While the above ought to help with choosing text and backgrounds, it is not
so easy to adjust images so that they can be viewable for all web users. However,
it is possible to adjust images so that color-deficient observers find it easier to
recognize objects in them. One approach is to increase the red and green contrast
in the image, as many color deficient viewers have residual red-green sensitivity.
Second, red-green variations can be mapped to other image attributes, such as the
yellow-blue channel, or luminance. The combination of these image-processing
techniques is called daltonization [261]. An example of such a simulation is
shown in Figure 5.50.

5.14 Application: Telling Humans

and Computers Apart
Visual illusions afford the possibility to detect the difference between humans and
computers, for instance in Internet-based applications where a webpage should
be visible to humans, but not to, say, malicious web crawlers. A commonly used
technique is image-based, whereby the viewer is presented an image of some
text and is asked to type in the text they observe [18, 19]. Geometric distortions
and random patterns can be added to the image to make it more difficult for a
computer to decipher the text. However, such systems can be defeated by clever
object-recognition algorithms [790].
Gavin Brelstaff and Francesca Chessa have argued that it may be possible
to use human visual illusions, which may be designed such that the complicated
human visual machinery is required to decipher a message encoded within the
display [125]. Given the huge complexity of the human visual system, and the
fact that much of it has not yet been explored, it may be very difficult for compu-
tational models to defeat such approaches.

5.15 Further Reading

Margaret Livingstone’s book is a good source linking human vision with art [695],
as is Semir Zeki’s book [1290]. Good general resources for the material discussed
in this chapter are Davson [236], Palmer [868], and Marr [742]. Further reading
on lightness perception can be found online at

• http://www-psych.rutgers.edu/˜alan/theory3/ [363]

i i

i i
 i i

i i

5.15. Further Reading 315

Interesting web sites with further optical illusions can be found at:

• http://www.echalk.co.uk/amusements/OpticalIllusions/illusions.htm

• http://www.michaelbach.de/ot/

• http://www.ritsumei.ac.jp/ akitaoka/index-e.html

• http://www.viperlib.org/

Color deficiencies are described in Donald McIntyre’s Colour Blindness: Causes

and Effects [766].
Many of the plots in this book were derived from data publicly available
on the Internet. In particular, we refer to the Munsell Color Science Labora-
tory (http://www.cis.rit.edu/mcsl/online/) and the CVRL color and vision database
(http://www.cvrl.org), which comes courtesy of the Institute of Ophthalmology,
University College London.

i i

i i
 i i

i i

Part II
Color Models

i i

i i
 i i

i i

Chapter 6
Radiometry and
Photometry

Photons are the carriers of optical information. Light, emitted by display devices,
light sources, or the sun, reaches our eyes, usually after reflection off objects in
a scene. After transduction, a signal is transmitted to the brain which is inter-
preted, and dependent on the circumstances, this may provoke a response from
the observer.
Since light is the carrier of visual information, a fundamental activity related
to light is its measurement. Light measurement exists in two forms. The first,
radiometry, is the science of measuring all optical radiation whose wavelengths lie
between approximately 10 nm and 105 nm (see Figure 2.4). This region includes
ultraviolet, visible, and infrared radiation. We only consider optical radiation, i.e.,
radiation that obeys the principles of optics.
Whereas radiometry considers many more wavelengths than those to which
the human visual system is sensitive, the second form of light measurement,
photometry, is concerned only with radiation in the visible range. The quan-
tities derived in radiometry and photometry are closely related: photometry is
essentially radiometry weighted by the sensitivity of the human eye. As such, ra-
diometry measures physical quantities and photometry measures psycho-physical
quantities.
In this chapter, we first revisit the topic of human wavelength sensitivity, en-
abling us to refine our explanation of radiometry and photometry. We then define
several important quantities relevant to both disciplines. We introduce computa-
tional models that attempt to compute perceptual quantities from their physical
counterparts. Thus, this chapter affords an insight into radiation measurements
319

i i

i i
 i i

i i

320 6. Radiometry and Photometry

that are important to understand the color measurements discussed in the next
chapter.

6.1 The Sensitivity of the Human Eye

The human eye is only sensitive to electromagnetic radiation between approx-
imately 380 to 780 nm. Sensitivity peaks at λ = 555 nm and tails off toward
the minimum and maximum wavelengths. Visual sensitivity varies slightly be-
tween individuals, and depends upon factors such as the illumination of the view-
ing environment, the extent of the viewing field, and the observer’s adaptation
level [183].
Under controlled viewing conditions, the sensitivity of the human eye can be
measured with sufficient accuracy. Such measurements were performed in the
early 1900s [358, 360]. Here, the field of view was fixed at 2◦ (approximately the
angular extent of a thumb nail when viewed at arm’s length) and the luminance of
the viewing field was set to at least 8 cd/mm2 , i.e., the lighting conditions are in
the photopic range. The accepted international values are based on five different
sets of measurements [1199] and form the basis of the CIE 1924 2◦ photopic
luminous efficiency curve V (λ ) (see Figure 6.1).
V (λ ) was modified by the CIE in 1988 to better account for human visual
sensitivity below 460 nm. This modified curve is denoted by VM (λ ), and is con-
sidered supplemental to V (λ ) rather than a replacement (Figure 6.1) [189].
For scotopic vision a separate curve was standardized by the CIE in 1951 [182,
212, 1194] and is called the CIE scotopic luminous efficiency curve,
denoted by V  (λ ). This curve is plotted in Figure 6.2 in both linear and loga-
rithmic scales. The sensitivity curves are tabulated in Appendix E.

0
of the human eye (log10)
Luminous efficiency

V(λ)
-1 VM(λ)
-2
-3
-4
-5
-6
-7
400 500 600 700 800
Wavelength (nm)

Figure 6.1. V (λ ) and VM (λ ).

i i

i i
 i i

i i

6.1. The Sensitivity of the Human Eye 321

Luminous efficiency (log10)

1.0
Luminous efficiency

V(λ) V(λ)
-1 V'(λ)
0.8 V'(λ)
-2
0.6 -3

0.4 -4
-5
0.2
-6
0.0 -7
400 500 600 700 800 400 500 600 700 800
Wavelength (nm) Wavelength (nm)

Figure 6.2. Photopic V (λ ) and scotopic V  (λ ) luminous efficiency curves.

As mentioned in the introduction of this chapter, photometric quantities can

be obtained by spectrally weighting their radiometric counterparts. The weighting
function is given by the V (λ ) curve.
Under photopic lighting conditions the L-, M-, and S-cones mediate vision.
Each cone type has a different spectral sensitivity and population (see Sections
4.3.3 and 4.3.4). Therefore, V (λ ) approximates the weighted sum of the sen-
sitivities of each type and takes into account their relative populations [302].
Conversely, as only rods are active under scotopic lighting conditions (see Sec-
tion 4.3.2), V  (λ ) models the sensitivity of the rods to light at different wave-
lengths.
Under mesopic lighting conditions, the relative contributions of rods and
cones vary with illumination level. As a result, a single curve would not rep-
resent human visual sensitivity appropriately; a family of curves would be a more
likely solution [567,611,1199]. Nevertheless, no curve has officially been recom-
mended by the CIE.
As shown in Figure 6.2, the shapes of the luminous efficiency curves for pho-
topic and scotopic vision are similar. However, the peak of the scotopic curve is
shifted by approximately 50 nm toward the shorter wavelengths (i.e., toward the
blue). Thus, a change in illumination causes a shift in the relative intensity of col-
ors. For instance, given a red and a blue object with identical surface reflectance
under ordinary illumination, the red one will appear darker than the blue object as
the illumination is reduced to the scotopic level. This phenomenon is called the
Purkinje effect or the Purkinje shift [930, 1199].

i i

i i
 i i

i i

322 6. Radiometry and Photometry

6.2 Radiometric and Photometric Quantities

Radiometric and photometric quantities are used to measure the amount of radi-
ation incident on, exitant from, or passing through a surface or point in space.
All radiometric and photometric quantities are wavelength dependent, and most
of them have positional and directional dependencies as well. Therefore, when
referring to a quantity Q that depends on all these factors it may be written as
Q (x, y, θ , φ , λ ), where (x, y) denotes the position, (θ , φ ) denotes the direction,
and λ denotes the wavelength. However, we drop some or all of these parameters
when this does not lead to confusion.
We assume that the radiometric quantities we begin with are spectrally un-
weighted (or have an equal weight applied to all wavelengths). Then, any given
photometric quantity Qv (λ ) may be derived from its radiometric counterpart
Qe (λ ) by
 780
Qv = Km Qe ( λ ) V ( λ ) d λ , (6.1)
380
where the e subscript (for energy) is used for radiometric quantities and the v sub-
script (for visual) is used for photometric quantities. These may also be dropped
when the meaning is clear. The result of integration is normalized by a constant
Km = 683 lm/W to obtain the photometric units at an absolute scale. The unit
lm/W (lumens per watt) affords a conversion from the fundamental radiometric
unit watt to the photometric unit lumen.
A similar computation can be carried out for values in the scotopic range.

Here, V (λ ) is replaced with V (λ ) and Km is replaced with Km = 1700 lm/W
resulting in
 780
Qv = Km Qe (λ ) V  (λ ) d λ . (6.2)
380
To calculate the value of a photometric quantity for only a single wavelength λ0 ,
the computation reduces to

Qv (λ0 ) = Km Qe (λ0 ) V (λ0 ). (6.3)

The most commonly used radiometric and photometric quantities are listed in
Table 6.1; they are explained in the following sections.

6.2.1 Radiant and Luminous Energy

To measure the total energy in a beam of radiation, a quantity called radiant en-
ergy is used. Radiant energy is denoted by Qe and measured in joules (J).
Similarly, to measure the total energy in a beam of radiation as perceived by
humans luminous energy is used. Thus, luminous energy is zero outside the

i i

i i
 i i

i i

6.2. Radiometric and Photometric Quantities 323

Radiometric quantity Unit Photometric quantity Unit

Radiant energy (Qe ) J (joule) Luminous energy (Qv ) lm · s
Radiant flux (Pe ) J · s−1 = W (watt) Luminous flux (Pv or Fv ) lm (lumen)
Radiant exitance (Me ) W · m−2 Luminous exitance (Mv ) lm · m−2 = lx (lux)
Irradiance (Ee ) W · m−2 Illuminance (Ev ) lm · m−2 = lx (lux)
Radiant intensity (Ie ) W · sr−1 Luminous intensity (Iv ) lm · sr−1 = cd (candela)
Radiance (Le ) W · m−2 · sr−1 Luminance (Lv ) lm · m−2 · sr−1 = cd · m−2 (nit)

Table 6.1. Radiometric and photometric quantities.

380 – 780 nm range. Luminous energy may be computed from radiant energy
using (6.1). It is denoted by Qv and measured in lumen seconds (lm s). The lumen
is the base unit of photometry and is explained in the next section. Radiant and
luminous energy are used in applications where the aim is to measure the total
dosage of radiation emitted from a source.

6.2.2 Radiant and Luminous Flux

As radiation may be variable over time, an important quantity is the rate of change
of radiation. This is called radiant flux. The symbol used for radiant flux is Pe and
it is defined as the time derivative of radiant energy:

dQe
Pe = . (6.4)
dt
With Qe expressed in joules and t in seconds, radiant flux Pe is given in watts
(1 W = 1 J/s). The photometric counterpart of radiant flux is called luminous
flux:
dQv
Pv = . (6.5)
dt
The same quantity may be derived from radiant flux:
 780
Pv = Km Pe (λ ) V (λ ) d λ . (6.6)
380

The unit of luminous flux is the lumen (lm), which is the base unit of photometry.
Luminous flux is a measure of the capacity of radiant flux to invoke visual sensa-
tions and is equal to 1/683 watt of radiant power at a frequency of 540 × 1012 Hz,
which corresponds to a wavelength of about 555 nm.

i i

i i
 i i

i i

324 6. Radiometry and Photometry

6.2.3 Radiant and Luminous Intensity

Radiant flux emitted from an ideal point source is not constrained as a function
of direction. As an emitter may radiate more in some directions than others, a
further unit of measurement is used. Radiant intensity may be used to quantify
the amount of flux emitted in a given direction. It is defined as the amount of
radiant energy emitted per second into a differential solid angle (solid angle is
defined in Appendix B.6). Therefore the radiant intensity Ie is given by

dPe
Ie = (6.7a)
dω
d 2 Qe
= . (6.7b)
dt d ω
The unit of radiant intensity is watts per steradian (W/sr). If the radiant intensity
of a point source is the same in all directions, the source is said to be uniform
or isotropic. For a uniform source with radiant intensity Ie , the total radiant flux
equals
Pe = ω Ie = 4π I. (6.8)
For instance, if Ie = 1 W/sr then Pe ≈ 12.56 W. A related quantity to radiant
intensity is the average radiant intensity which is defined as the ratio of the total
flux emitted by a source to the solid angle of emittance which is usually 4π (i.e.,
the entire sphere surrounding the source).
The luminous intensity is the photometric counterpart of radiant intensity and
may be computed from luminous flux:

dPv
Iv = (6.9a)
dω
d 2 Qv
= , (6.9b)
dt d ω
or may be derived from radiant intensity:
 780
Iv = Km Ie (λ ) V (λ ) d λ . (6.10)
380

The unit of luminous intensity is lumens per steradian (lm/sr) which is normally
called candela (cd). The candela is defined as the amount of luminous intensity
in a given direction of a source that emits monochromatic radiation of frequency
540 × 1012 Hz and that has a radiant intensity in that direction of 1/683 W sr−1 .
Luminous intensity may be computed for any source (i.e., both point and ex-
tended), but in general it is only meaningful for point sources [839]. For extended

i i

i i
 i i

i i

6.2. Radiometric and Photometric Quantities 325

sources, radiance and luminance are appropriate. They are defined in the follow-
ing sections.
The average luminous intensity, which is the photometric equivalent of av-
erage radiant intensity, is also referred to as the mean spherical candle power
(MSCP) [1199].

6.2.4 Irradiance and Illuminance

When optical radiation reaches a surface, that surface is irradiated or illuminated.
The radiometric irradiance and the photometric illuminance are measures of how
intensely irradiated or illuminated a point on a surface is.
The irradiance Ee is defined as the ratio of the radiant flux striking a differen-
tial surface area dA (i.e., radiant energy per second per surface area):

dPe
Ee = . (6.11)
dA
As irradiance does not have directional dependency, to compute irradiance at a
point, the radiation arriving from all incident directions must be included. For
a differential surface area of an opaque surface, this requires integrating over a
hemisphere covering the area of interest. Irradiance is measured in watts per
square meter (W/m2 ).
The illuminance Ev at a point is defined as the luminous flux per differential
area:
dPv
Ev = . (6.12)
dA
It may also be derived directly from irradiance Ee :
 780
Ev = Km Ee (λ ) V (λ ) d λ . (6.13)
380

The unit of illuminance is lumens per square meter (lm/m2 ), which is also called
lux.
Irradiance and illuminance are measures of the density of incident radiation.
To define the surface density of exitant radiation leaving a point, the quantities
called radiant exitance and luminous exitance are used. They have the same units
as irradiance and illuminance.
In the following sections we demonstrate the computation of illuminance with
two examples. In the first example, the surface lit by a light source is assumed
to be differential. Therefore, the intensity of the light source can be regarded as
the same for all directions spanning the surface. In the second example, a larger
surface is considered where this assumption cannot be made.

i i

i i
 i i

i i

326 6. Radiometry and Photometry

Iv
N
θ d θ ⊥
d

dA

Figure 6.3. A differential surface area lit by a point light with intensity Iv .

6.2.5 Illuminance on a Differential Surface Area

The configuration under consideration is depicted in Figure 6.3. As the light
strikes a differential surface area, we can assume that all points contained in this
area are exposed to the same luminous intensity. The flux dPv emitted from the
light source that reaches the surface element dA is then
dA cos (θ )
dPv = Iv , (6.14)
d2
where Iv is the luminous intensity of the source in the direction of the surface, θ
is the angle of incidence, and d is the distance between the source and the surface.
Note that dA cos (θ )/d 2 is the solid angle spanned by the surface with respect to
the source and is obtained by applying both the cosine law and the inverse square
law (Sections 2.9.2 and 2.9.3). The illuminance is then given by
dPv
Ev = (6.15a)
dA
cos (θ )
= Iv . (6.15b)
d2
This equation may be written in a different form which will help the computation
of illuminance over large surface areas. Since d equals d ⊥ / cos (θ ), where d ⊥ is
the distance between the surface and the point source along the surface’s normal,
substituting d in (6.15b) gives

cos3 (θ )
Ev = Iv . (6.16)
(d ⊥ )2
This equation is known as the cosine-cubed law. In this formulation, each point
on a surface lit by a uniform light source receives illumination that varies with θ ,
since d ⊥ is the same for all points on a large flat surface. For non-uniform light
sources, the illumination changes with θ and Iθ (the intensity of the source in the
given direction θ ).

i i

i i
 i i

i i

6.2. Radiometric and Photometric Quantities 327

I
θ2

⊥ θ1
d

S S

Figure 6.4. A large surface is illuminated by a point source.

6.2.6 Illuminance on Large Surfaces

The equations of the previous section can be extended to compute the average
illumination of any surface, for instance by computing the illuminance at each
point and then averaging the results. Consider the surface and the point light in
Figure 6.4.
For simplicity, we assume that the surface is disk-shaped and has a radius R.
We also assume that a point source of luminous intensity Iv is located directly
above the center of the disc. Under this configuration, the cosine-cubed law sug-
gests that all points lying on a ring with a given radius will have the same illumi-
nance (Figure 6.5). The area of a ring of radius r is equal to 2π r dr, where dr is
the differential increase in radius across the ring. The total illuminance over this

I
θ θ
⊥
d

R
r

dr
2π r

Figure 6.5. Top: Each differential area on a ring has the same angle of incidence θ and,
therefore, has the same illuminance. Bottom: The area calculation of the ring is shown.

i i

i i
 i i

i i

328 6. Radiometry and Photometry

ring is then
cos3 (θ )
Pv,r = Iv 2π r dr. (6.17)
(d ⊥ )2
Integrating over all rings that form the disk yields the total illuminance over the
surface:  R
cos3 (θ )
Pv = Iv 2π r dr. (6.18)
0 (d ⊥ )2
Here, both r and θ are changing together, making the integration problematic.
This can be avoided by rewriting r in terms of θ as r = d ⊥ tan (θ ). Then dr =
d ⊥ sec2 (θ ) d θ and the upper limit of the integral becomes tan−1 (R/d ⊥ ):
 tan−1 (R/d ⊥ )
Pv = 2π Iv sin (θ ) d θ (6.19a)
0 
d⊥
= 2π Iv 1 −  . (6.19b)
R2 + (d ⊥ )2

The average illumination E¯v is calculated by dividing Pv by the disc’s surface area:
 
2 Iv d⊥
Ēv = 2 1 −  . (6.20)
R R2 + (d ⊥ )2

It can be verified that as R becomes small compared to d ⊥ (i.e., R/d ⊥ → 0), this
equation reduces to Ēv = Iv /(d ⊥ )2 .

6.2.7 Radiance and Luminance

The quantities discussed so far were either measured per second (flux), over dif-
ferential surface areas (irradiance), or over differential angles (intensity). An in-
teresting unit of measurement arises when energy is measured per second, per unit
surface area, and per solid angle all at the same time. Such quantities are called
radiance (radiometric) and luminance (photometric) and allow the measurement
of energy arriving at a surface from a particular direction integrated over a period
of time. The same units are used to indicate flux emitted by a surface element into
a particular direction, as well as flux along the path of a beam.
The symbol used for radiance is Le and its value is computed as

d 2 Pe
Le = , (6.21)
d ω dA cos (θ )
where θ is the angle of incidence or exitance (see also Figure 6.6). The unit of
radiance is watts per steradian per square meter (W sr−1 m−2 ).

i i

i i
 i i

i i

6.2. Radiometric and Photometric Quantities 329

Le (x,y,θ,φ,λ)

dω

N
θ

dA cos (θ)

φ
dA

Figure 6.6. Radiance Le emitted from a point (x, y) on a differential surface dA in the
direction of (θ , φ ).

The radiance Le emitted from a differential area is directly related to its radiant
intensity Ie :
dIe
Le = , (6.22)
dA cos (θ )
where θ is the exitant angle.
Luminance Lv is the photometric counterpart of radiance. It may be computed
from luminous flux Pv as
d 2 Pv
Lv = . (6.23)
d ω dA cos (θ )
Alternatively, luminance may be derived from radiance:
 780
Lv = Km Le (λ ) V (λ ) d λ . (6.24)
380

Luminance is measured in candela per square meter (cd m−2 ). The nit is also
sometimes used instead of candela per square meter, although this term is now
obsolete.
It is easy to see that a camera records spectrally weighted radiance. Light
arriving from many directions is focused by the lens, measured over a given ex-
posure time, and for each pixel individually recorded over a small surface area.
Thus, the charges recorded in the sensor, between the time that the shutter is
opened and closed, represents radiance values. The firmware in the camera then
processes the data, so that we can not be certain that the values retrieved from the
camera still correspond to radiances (although most cameras now export a raw
format, designed to bypass much of the in-camera processing).

i i

i i
 i i

i i

330 6. Radiometry and Photometry

In the human eye a similar process takes place. Here the refractive indices of
the various ocular media focus light, which is then transduced by the photorecep-
tors. The human equivalent of exposure time is not so easily defined. Photore-
ceptors transmit signals at the same time as they receive photons. The amount
of signal is modulated by the amount of light, bleaching, and by neural feedback
mechanisms. Nevertheless, this combination of processes causes the signal to
be commensurate with the radiance reaching each receptor. Subsequent visual
processing stages then cause the signal to be photometrically weighted, so that
luminance values result.

6.2.8 Invariance of Radiance

If we assume that energy propagates between an emitter and a surface through a
medium that does not absorb any energy, then this medium is called lossless. In
such media, it is possible to measure the emitted radiance from a source by only
measuring the incident radiance on a detector.
In Figure 6.7, two surface elements are placed in arbitrary orientations; the
distance between them is also arbitrary. We assume that the surfaces are small
enough such that the radiance emitted and received by any one point on either
surface is the same for all other points on the same surface.

dS

NS
θS
d

LS
NR LR
θR

dR

Figure 6.7. A differential surface dR is radiated by a differential source dS located at a

distance d from the surface. In a lossless medium, the radiance LS emitted from the source
is equal to the radiance LR received by the surface.

i i

i i
 i i

i i

6.2. Radiometric and Photometric Quantities 331

If the radiance emitted from the surface with surface area dS is equal to LS ,
then the total flux dPS that is emitted in the direction of dR equals

dPS = LS cos (θS ) dS d ωR , (6.25)

where d ωR is the solid angle subtended by dR at the center of dS. The differential
solid angle d ωR can be rewritten as follows:

cos (θR ) dR
d ωR = , (6.26)
d2
where d is the distance between the center of the surfaces. Substituting d ωR into
the previous equation we obtain

LS cos (θS ) cos (θR ) dS dR

dPS = . (6.27)
d2
If the radiance received by any point on the surface with area dR is equal to LR ,
then the total flux received from the direction of dS is

dPR = LR cos (θR ) dR d ωS . (6.28)

This expression can be rewritten as

LR cos (θR ) cos (θS ) dR dS

dPR = . (6.29)
d2
In a lossless medium, the flux emitted from dS in the direction of dR must be
equal to the flux received by dR from the direction of dS (see Figure 6.8):

LS cos (θR ) cos (θS ) dR dS LR cos (θR ) cos (θS ) dR dS

= , (6.30)
d2 d2
and hence
LS = LR . (6.31)
This result expresses the invariance of radiance. Since invariance of radiance
is independent of the points chosen on dS and dR (their centers were arbitrarily
chosen), it holds between any pair of points on both surfaces. Furthermore, the
equation is also independent of d, which means that for any pair of points along a
ray connecting dS to dR, the radiance is the same. The significance of this result
lies in the fact that by measuring the incident radiance on a receiver, such as a
detector, the radiance emitted from a source can be inferred.
So far we have assumed that the intervening medium is lossless, and that
there is no optical system between the source and the receiver. However, most

i i

i i
 i i

i i

332 6. Radiometry and Photometry

dS

NS

NR

dR

Figure 6.8. The flux emitted by dS in the direction of dR is equal to the flux received by
dR from the direction of dS.

measurement instruments have lenses to refract the incident radiation towards a

detector.
At a differential surface element dA of the lens, a beam of radiation is refracted
according to Snell’s law, given in (2.102). The differential flux on either side of
the surface of the lens is given by
dP1 = L1 dA cos (θ1 ) d θ1 d φ , (6.32a)
dP2 = L2 dA cos (θ2 ) d θ2 d φ . (6.32b)
In a lossless medium, both quantities will be identical (dP1 = dP2 ). Combining
this result with (2.102), we find
L1 L2
2
= 2. (6.33)
n1 n2
The quantity L/n2i is called the basic radiance, and it is always preserved in a
lossless medium [839]. Once light exits the lens and returns to the medium with
refractive index n1 , the radiance also returns to its old value.
In practice, radiance is not fully invariant due to energy losses during trans-
mission, refraction, and reflection. The fraction of initial radiance LS that is suc-
cessfully propagated to a receiver LR is called the propagance τ of the entire path
of radiation:
LS
τ= . (6.34)
LR

i i

i i
 i i

i i

6.2. Radiometric and Photometric Quantities 333

The value of propagance of an optical path is due to all attenuation caused by

absorption, scattering, reflection, and refraction.

6.2.9 Irradiance from Radiance

Given a source and a receiver that face each other (Figure 6.7), it is possible
to compute the irradiance on a receiver as function of the radiance emitted by
the source. The flux dPR received by dR from dS is given by (6.29). Since the
irradiance is equal to the incident flux per differential surface area, the irradiance
ER on dR is equal to dPR /dR, and hence

LS cos (θR ) cos (θS ) dS

ER = , (6.35)
d2
where we have applied (6.31). In this equation, the factor cos (θS ) dS/d 2 is the
solid angle subtended by dS, and hence (6.35) may be rewritten as follows:

ER = LS cos (θR ) d ωS . (6.36)

An interesting result is obtained if both the source dS and the receiver dR lie
on the interior surface of a sphere. In that case, we have,

θR = θS = θ , (6.37a)
d = 2r cos (θ ), (6.37b)

where r is the radius of the sphere as shown in Figure 6.9. Substituting these
into (6.35), the irradiance on dR is given by
LS dS
ER = . (6.38)
4 r2

dS
θR = θS = θ
d = 2r cos (θ)
θS r
d
θR O
r
dR

Figure 6.9. The two differential patches dS and dR lie on a sphere. In this case, the
irradiance of dR due to emission from dS is independent of the distance d and both of the
angles θS and θR .

i i

i i
 i i

i i

334 6. Radiometry and Photometry

Note that this result is independent of both θ and d. Therefore, all the points
on a sphere are equally irradiated from any other point on the sphere. This is
a particularly convenient configuration which may be exploited to measure both
power as well as irradiance (or their photometric equivalents) using a device called
the integrating sphere (see Section 6.8.2).

6.2.10 Iso-Radiant Surfaces

As discussed in Section 2.9.2, the radiation emitted from any point of an iso-
radiant surface decreases with the cosine of the angle between the surface normal
and the direction of emission. In other words, for an iso-radiant surface the radiant
intensity of each point (i.e., differential surface element) along the direction θ is
equal to
dI = I0 cos(θ ), (6.39)
where I0 is the radiant intensity in the direction normal to the surface. For such
surfaces, the radiance emitted from any point in the direction θ is given by

dI I0
L= = . (6.40)
dA cos (θ ) dA

As such, radiance (and similarly luminance) emitted from an iso-radiant surface

is the same irrespective of the outgoing direction. For such surfaces, the flux dP
emitted per differential area may be conveniently related to the radiance of the
surface:
 π /2  2π
dP = L dA cos (θ ) d ω (6.41a)
0 0
 π /2  2π
= L dA cos (θ ) sin (θ ) d φ d θ (6.41b)
0 0
 π /2
= 2π L dA cos (θ ) sin (θ ) d θ (6.41c)
0
= π L dA. (6.41d)

The total flux emitted from the entire surface is then computed as

P= dP (6.42a)
A

= π L dA (6.42b)
A
= π L A. (6.42c)

i i

i i
 i i

i i

6.2. Radiometric and Photometric Quantities 335

Receiver patch dR
N
I

d θ ⊥
d
θ
dS

Source
r

Figure 6.10. The differential surface patch dR is illuminated by a flat circular extended
source. A differential ring on the source surface is shown, and the radiance L emitted from
a differential area dS of this ring is shown to strike dR.

6.2.11 Error Associated with the Point-Source Assumption

While computations with point sources are convenient, in practice, light sources
are extended. In Section 2.9.3, we gave a rule of thumb stating that if the distance
to the source is at least ten times the size of the largest dimension of the source,
this source can be approximated as a point source. The error associated with this
rule is analyzed here.
We begin by considering an extended circular source of constant radiance L
illuminating a differential surface area dR which is d ⊥ units apart from the source
(see Figure 6.10). The source and receiver are facing each other, i.e., their surface
normals point in opposite directions.
The amount of flux received by dR from each differential element dS of the
source is equal to
L cos2 (θ ) dR dS
d2 P = . (6.43)
d2
As θ is the same for all points on a ring, we can rewrite the total flux received
from each ring as
L cos2 (θ ) dR
d 2 Pring = 2π r dr , (6.44)
d2
and, therefore, the flux received by the receiver dR from the extended source is
equal to
 R
L cos2 (θ ) dR
dP = 2π r dr . (6.45)
0 d2
This form is difficult to integrate since both d and θ are changing together for
each differential band. Making similar substitutions as outlined in Section 6.2.6

i i

i i
 i i

i i

336 6. Radiometry and Photometry

θΝ
(x 0, y 0)
L (x 0, y0 , θ, φ)

θΝ

(x 1, y 1) L (x 1, y1 , θ, φ)

Figure 6.11. The angle θN may change based on the surface curvature whereas θ is fixed.

yields
 tan−1 (R/d ⊥ )
P= 2π L sin (θ ) cos (θ ) d θ dR (6.46a)
0
 
R2
= π L dR , (6.46b)
R2 + (d ⊥ )2
and the irradiance E is equal to
 
dP R2
E= = πL . (6.47)
dR R2 + (d ⊥ )2
In the general case, the radiant intensity of an extended source in the direction
of (θ , φ ) is defined as

I (θ , φ ) = L (θ , φ ) cos (θN ) dS, (6.48)
S
!
where S indicates that integral is computed over the entire surface. Also note
that θN is the angle between the surface normal and the direction of emission.
The angle θN is not necessarily equal to θ as illustrated in Figure 6.11.
For an extended circular source (Figure 6.10), the intensity in the direction of
the surface normal is given by
I = π R2 L. (6.49)
By substituting this result into (6.47) we obtain the irradiance Ea on the differen-
tial surface dR in terms of the intensity of the source:
I
E= . (6.50)
R2 + (d ⊥ )2
If the source is very small (R → 0) or if the distance from the source is very large
(d ⊥ → ∞) (e.g., starlight), the approximate irradiance Ea is given by
I
Ea ≈ . (6.51)
(d ⊥ )2

i i

i i
 i i

i i

6.3. The Efficacy of Optical Radiation 337

Relative irradiance (log10)

E’
0 E’a

-1

-2

-3
-2 -1 0 1 2
log (d ⊥/ R)

Figure 6.12. The relative irradiance on a differential surface due to an extended source.
Both exact values E  and the point-source approximation Ea are shown.

Therefore, the error associated with treating an extended source as a point source
stems from ignoring the R2 in the denominator. We can see this by dividing both
E and Ea by R2 so that

I
E = 2
, (6.52a)
1 + (d ⊥ /R)
I
Ea ≈ 2
. (6.52b)
⊥
(d /R)

If we plot both E  and Ea for different ratios of d ⊥ and R (Figure 6.12), we see
that the two curves almost overlap when the distance between the source and the
receiver is at least ten times the radius of the source. At this distance, the error
drops below 1%, and it reduces further for longer distances.

6.3 The Efficacy of Optical Radiation

As evidenced by the V (λ ) curve, the human visual system has a different sensitiv-
ity to each wavelength. Wavelengths close to the peak sensitivity (555 nm) invoke
a greater visual sensation than other wavelengths. Thus, radiation composed of
wavelengths around the HVS’ peak sensitivity have a higher luminous efficiency
than other radiation.
The ability to invoke a visual sensation can be expressed in terms of efficacy
as well as efficiency. The efficacy of radiation is the ratio of its luminous power
to the radiant (or electrical) power that produced it. Thus, efficacy is measured

i i

i i
 i i

i i

338 6. Radiometry and Photometry

in lumens per watt (lm/W). On the other hand, efficiency is a unitless term that
expresses the ratio of the efficacy at a particular wavelength to the efficacy at
555 nm [1090].
Two types of efficacies are distinguished: the efficacy of electromagnetic ra-
diation and the efficacy of lighting systems. The former is defined as the ratio of
the luminous power to the radiant power in a beam of radiation, whereas the lat-
ter is defined as the ratio of the luminous power produced to the electrical power
consumed in a lighting system. These concepts are discussed in the following
sections.

6.3.1 Luminous Efficacy of Radiation

The luminous efficacy of radiation Kr is defined as the quotient of the luminous
flux and the radiant flux in a beam of electromagnetic radiation:
Pv
Kr = . (6.53)
Pe
It is measured in lm/W. When a beam of radiation is composed of a single wave-
length λ0 , its efficacy Kr (λ0 ) is equal to the value of the luminous sensitivity
curve for that wavelength times a normalization constant. For instance, for pho-
topic vision the luminous efficacy at a single wavelength λ0 is
683 Pe (λ0 ) V (λ0 )
Kr (λ0 ) = = 683 V (λ0 ) lm/W (6.54)
Pe (λ0 )
The luminous efficacy curve for photopic vision may be obtained by plotting
Kr (λ ) for all wavelengths in the visible spectrum. A similar curve Kr (λ ) can
be created from V  (λ ) for scotopic conditions. Both curves are plotted in Fig-
ure 6.13. Whereas V (λ ) and V  (λ ) give relative information, Kr (λ ) and Kr (λ )
provide absolute values of sensitivity.
For polychromatic light, the luminous efficacy may be computed by
! 780
683 Pv (λ ) V (λ ) dλ
Kr = !∞
380
. (6.55)
0 Pe (λ ) d λ
The luminous efficacies of ultraviolet and infrared radiation, as well as of other
types of radiation outside the visible spectrum, are zero. The maximum possible
luminous efficacy for photopic vision is equal to 683 lm/W, and it is only attain-
able if all the radiation is concentrated at 555 nm. As an aside, the efficiency of
such radiation is unity. If radiation is uniformly distributed at all wavelengths in
the visible range and zero everywhere else, its efficiency is around 0.4 [1199].
The luminous efficacies of commonly encountered types of radiation are given in
Table 6.2.

i i

i i
 i i

i i

6.3. The Efficacy of Optical Radiation 339

1800
Efficacy (lm/W) 1600
K (λ)
1400 K’ (λ)
1200
1000
800
600
400
200
0
400 500 600 700 800
Wavelength (nm)

Figure 6.13. Luminous efficacy curves for photopic (K(λ )) and scotopic (K  (λ )) vision.

Source Luminous efficacy (lm/W)

Monochromatic light (555 nm) 683
White light, constant over visible spectrum 220
Blue-sky light 125–140
Direct-beam sunlight, midday 90–120
Direct-beam sunlight, extraterrestrial 99.3
Direct-beam sunlight, sunrise/sunset 50–90
Overcast sky 103–115
Uncoiled tungsten wire at its melting point 53
Radiation from typical tungsten filament lamp 15
Phosphorescence of cool white fluorescent lamp 348

Table 6.2. Luminous efficacies of commonly encountered light sources (from [761]).

6.3.2 Luminous Efficacy of Lighting Systems

The efficacy of electromagnetic radiation does not give insight into how much
power needs to be supplied to a lighting system for it to create that radiation.
For instance, a monochromatic laser beam emitting at 555 nm has high efficiency,
even if a large amount of electrical power is consumed in the operation of such
a laser. Thus, a different measure is required to compute the overall efficacies of
lighting systems.
The luminous efficacy of lighting systems Ks is defined as the ratio of the
total luminous flux Pv produced by a lighting system (i.e., lamp, laser, etc.) and

i i

i i
 i i

i i

340 6. Radiometry and Photometry

Source Lighting system efficacy (lm/W)

Standard incandescent filament lamp, 2700K 17
Tungsten-halogen (linear), 2950K 20
Fluorescent T12 4ft, 3500K 82
Mercury vapor, 6800K 45

Table 6.3. Luminous efficacies of lighting system for commonly used light sources.

the total electrical power Pel supplied to the system:

Pv
Ks = . (6.56)
Pel
The efficacies of commonly used light sources are given in Table 6.3.

6.4 Luminance, Brightness, and Contrast

While luminance is a psycho-physical quantity relating radiance to the spectral
sensitivity of the human visual system, it is not always a good measure of how
the stimulus that caused this quantity of luminance is perceived. The perception
of a certain amount of luminance, whether emitted by a source, or reflected off
a surface, depends on environmental and physiological conditions, and the per-
ception of the same absolute amount of luminance may thus vary with viewing
conditions. For example, the headlights of a car appear brighter during the night
than during the day, even though they are the same lights (see Figure 6.14).
To describe the perception of how much light is emitted, brightness, as defined
in Section 5.1, can be used. Although brightness depends on external factors,

Figure 6.14. The car lights are on in both images, but they appear much brighter in the
dark environment. (This image was published in Erik Reinhard, Greg Ward, Sumanta
Pattanaik, and Paul Debevec, “High Dynamic Range Imaging, Acquisition, Display, and
Image-Based Lighting”, c Elsevier 2006.)

i i

i i
 i i

i i

6.4. Luminance, Brightness, and Contrast 341

such as the appearance of surrounding areas, in many cases it is proportional to

luminance. Under appropriate conditions, the relation between brightness B and
luminance L can be approximated with Stevens’ power law (Section 5.5.3) [1262]:

B = a L1/3 − B0 . (6.57)
Here, a and B0 are constants depending on the viewing conditions. These condi-
tions include the state of adaptation of the observer as well as the background and
spatial structure of the viewed object [865].
A fundamental quantity which is closely related to luminance and brightness
is contrast. Essentially, two types of contrast may be identified: physical contrast
and perceived contrast.
Physical contrast, also known as luminance contrast, is an objective term used
to quantify luminance differences between two achromatic patches.1 It may be
defined in several ways depending on the spatial configuration of patches. For in-
stance, the contrast C between a foreground patch superimposed on a background
patch is characterized by " "
" L f − Lb "
C=" " ", (6.58)
Lb "
where L f and Lb are the luminances of the foreground and background patches. If
two patches L1 and L2 reside side by side forming a bi-partite field, their contrast
may be computed by
|L1 − L2 |
C= . (6.59)
max (L1 , L2 )
Finally, for periodic patterns, such as sinusoidal gratings, the contrast may be
computed by
Lmax − Lmin
C= , (6.60)
Lmax + Lmin
where Lmax and Lmin are the maximum and the minimum luminances in the grat-
ing. This type of contrast is also called Michelson contrast or modulation .
For pairs of diffuse surfaces, the luminance contrast depends solely on their
reflectances [941]. Thus, even if the illumination on these surfaces is altered, the
luminance will change proportionally, so that the luminance contrast will remain
the same.
The second type of contrast is perceived contrast, also known as brightness
contrast. It is the apparent difference between the brightnesses of two achromatic
patches. Since brightness does not only depend on luminance, brightness con-
trast does not strictly depend on luminance contrast either. In general, brightness
1 Chromatic patches might also induce color contrast; this is discussed in Chapter 5.

i i

i i
 i i

i i

342 6. Radiometry and Photometry

contrast increases with luminance, and this phenomenon is called the Stevens
effect [302]. For instance, a black-and-white photograph appears less rich in con-
trast if viewed indoors compared to viewing it outdoors. Outdoors, blacks appear
more black and whites appear more white. The luminance contrast is the same in
both cases, unless the photograph is held at angles that cause highlights.

6.5 Optical Detectors

The various radiometric and photometric terms explained so far provide quanti-
ties that indicate how much light is emitted, transmitted, or received. We now
discuss the devices that are used to measure light and how these quantities may
be measured in practice.
The basic function of an optical detector is to measure optical radiation. When
a photon strikes the surface of a detector, the detector undergoes a change of state.
A photon striking a material may free one of its electrons if the energy of the
photon is above a certain threshold, a process called photo-electricity. The energy
of the released electron is equal to the energy imparted on it by the photon minus
the energy it required to depart from its orbit. Thus, by measuring the energy of a
released electron, one can find the energy of the incident photon.
Most optical detectors work on this principle, although they vary in construc-
tion. Three main types of optical detectors may be distinguished: photo-emissive,
photo-conductive, and photo-voltaic detectors. Each detector has a relationship
between the incident radiation, and the output signal produced. This relationship
is characterized by a detector’s responsivity.

6.5.1 Responsivity
The output signal of an optical detector may be expressed in several ways, such
as a change in the voltage, current, or conductivity. If we denote the output signal
by S and the incident flux by P, the flux responsivity RP of a detector is given by
S (λ , λ0 )
RP ( λ , λ 0 ) = , (6.61)
P (λ )

where λ0 is the wavelength to which the detector is most sensitive. Thus, the
responsivity may change as a function of position, direction, and wavelength.
Furthermore, the responsivity may change as a function of incident radiation. A
detector is said to have a linear response if its flux responsivity does not change
with the amount of incident radiation, assuming all other factors are kept constant.
Linearity is a desired property of optical detectors and often satisfied within the
wavelength range of the detector [839].

i i

i i
 i i

i i

6.5. Optical Detectors 343

Relative response
100

80
Selenium
60 Silicon
Standard observer
40 Tungsten (2856 K)

20

0
400 500 600 700 800 900 1000 1100
Wavelength (nm)

Figure 6.15. The spectral response of various materials used in detectors, as well as the
responsivity of the human eye (after [1090]).

In Figure 6.15, the responsivities of two commonly used detector materials,

selenium and silicon, are plotted together with the sensitivity of the human eye
(i.e., V (λ )). As can be seen, there are significant differences between the re-
sponsivities of different materials and the human eye. In practice, no material
has precisely the same responsivity as the human eye. Therefore, to make photo-
metric measurements with optical detectors, specially designed light filters called
photopic correction filters are used to match the responsivity of detectors to the
sensitivity of the human eye.

6.5.2 Photo-Emissive Detectors

Photo-emissive detectors are composed of a vacuum tube and an open circuit
through which no current passes in the absence of light (see Figure 6.16, left).
The vacuum tube comprises a positively charged anode and a negatively charged
cathode. The cathode contains a suitable material that releases electrons when
exposed to radiation. A diagram of a vacuum tube is shown in Figure 6.16 (right).
The envelope of the vacuum tube is made of glass which lets light in. The en-
velope is attached with earthed-metal guard rings to prevent leakage of electrons
over the glass surface. In some cases the tube is filled with a noble gas, such as
argon, to enable a higher current between anode and cathode [1199].
Usually, the cathode contains alkali metals which release their single valence
band electrons easily when exposed to light. Therefore, when light strikes the
cathode it releases some of its electrons. These free electrons are collected at the

i i

i i
 i i

i i

344 6. Radiometry and Photometry

Anode
Resistor
+
Anode Vacuum tube
Cathode -

Guard rings
Light rays
Electron flow Cathode

Figure 6.16. Photo-detection by photo-emission. On the left, a photo-emissive circuit is

shown. The figure on the right presents a diagram of a vacuum tube.

anode due to its positive voltage, resulting in a continuous flow of current. By

measuring this current, the flux incident upon the detector can be inferred.
Vacuum tubes by themselves are rarely used in practice. However, a variant
called the photo-multiplier tube (PMT) is used in applications requiring high sen-
sitivity [941]. The difference between a PMT and a classical vacuum tube is
that there are several other electrodes (called dynodes) between the anode and the
cathode. When an electron released from the cathode hits the first dynode, more
electrons are released. Each subsequent dynode further amplifies the signal, until
the anode collects a large number of electrons, yielding a sensitive measuring de-
vice. One disadvantage of PMTs is that they require stable and very high voltages
to operate.

6.5.3 Photo-Conductive Detectors

Photo-conductive detectors are built from semiconductor materials such as sili-
con. Normally semiconductors have high resistance and, therefore, low conduc-
tivity. In a circuit, as shown in Figure 6.17, no current passes in the absence of
light.
However, when exposed to light, the photo-electric effect causes some of the
valence electrons in the semiconductor to be freed. With an increasing number
of free electrons, the conductivity of the material increases (i.e., the resistance
decreases), which results in a higher electrical current flowing through the circuit.
Thus, by measuring the change in the current, one can measure the amount of
light falling on the detector.
Historically, the most common materials used in photo-conductive detectors
are cadmium sulfide (CdS), cadmium selenide (CdSe), and lead sulfide (PbS),
whereas currently silicon (Si) is widely used [1024].

i i

i i
 i i

i i

6.6. Light Standards 345

Glass window Resistor

+
Photoconductive
material -
Light rays
Electron flow

Figure 6.17. Photo-detection by photo-conduction.

6.6 Light Standards

The efforts to quantify light and define certain light sources as standards dates
back to the early 1800s. First, flames of certain candles were used as standards,
but these were not reliable to the accuracy needed by photometric measurements.
Then, more controlled flame lamps such as the Carcel lamp, the pentane lamp,
and the Hefner lamp replaced the standard candles during the nineteenth cen-
tury [1199]. Although they are more reliable and accurately reproducible than
candles, flame lamps also do not afford the desired accuracy. After their inven-
tion, incandescent filament lamps were used as standards. However, they also
lack the desired accuracy and stability.
Proposals were made from time to time to use the radiation emitted from a
heated metal of a specified surface area to be used as standard. Different types of
metals were tested,2 and platinum at its melting temperature was found to be the
most appropriate. Even platinum was not perfectly reliable due to the difficulties
of preventing contamination of pure platinum. Later this standard was revamped,
and a full radiator3 immersed in pure molten platinum at its melting temperature
was used as the primary standard.
Although this latest standard served reasonably well in photometric measure-
ments, it has difficulties in construction. It has been a goal for a long time to relate
the unit of luminous intensity, the candela, to the optical watt [242]. This desire
was realized in 1979 when a new definition of candela was adopted.
The candela is now defined as the amount of luminous intensity in a given
direction of a source that emits monochromatic radiation of frequency 540 × 1012
Hz and that has a radiant intensity in that direction of 1/683 W/sr. This definition
is been adopted as the primary standard of light.

2 Suggested radiators were an iron vessel containing boiling zinc, a coil of platinum wire heated to

the temperature of a hydrogen flame, and burning magnesium [1199].

3 Although it is not possible to obtain a full (blackbody) radiator in nature, very close approxima-

tions can be made with a thin hollow cylinder with absorbing walls.

i i

i i
 i i

i i

346 6. Radiometry and Photometry

Of course, all of the standards described above can be realized under labora-
tory conditions. However they are not practical for everyday use. For practical
measurements, more portable and readily usable light sources are created by com-
parison with primary standards. These standards are called secondary standards
(or sub-standards).
Secondary standards should approximate the primary standards to the ade-
quate degree of accuracy. In most cases, the hierarchy is deepened one more level
to produce what is called working standards by comparing against secondary stan-
dards. The accuracy of working standards should be adequate enough to be used
in ordinary photometric work [1199].

6.7 Detector Standards

Light standards discussed in the previous section are used to carry out photomet-
ric measurements by comparing test lamps against them (see also Section 6.9.1).
However, today, such comparisons are rarely made, mainly because of the prac-
tical difficulties involved and low reproducibility of results. Instead, calibrated
optical detectors are used for both radiometric and photometric measurements.
The primary standard that underpins the calibrated detectors is the electrical
substitution radiometer4 (ESR) that operates at cryogenic temperatures (about
5K = −272◦ C). The ESR works on the principle of matching the tempera-
ture change caused by incident radiant flux (radiant power) with the temperature
change due to applied electrical power. When an equilibrium is achieved, the radi-
ant flux is computed from the equivalent electrical power. A simplified illustration
of the ESR is shown in Figure 6.18.
When the shutter is open, the radiant flux enters the receiving cone and in-
creases the temperature of the heat sink whose initial temperature is T0 . Then, the
shutter is closed and the temperature of the heat sink is maintained by increas-
ing the electrical current through the heater with resistance R. If we denote the
current that is just sufficient to maintain the temperature of the heat sink by i, the
electrical power P consumed at the heater and hence the optical power Pe will be
equal to
Pe = P = i2 R. (6.62)
Due to its low operating temperature, very accurate measurements can be per-
formed with a cryogenic ESR. However, the main purpose of an ESR is to cali-
brate secondary standards or transfer standard detectors which in turn are used to
calibrate working standards [877].
4 Also known as absolute cryogenic radiometer or electrically calibrated radiometer.

i i

i i
 i i

i i

6.8. Measurement of Optical Radiation 347

Po
w
+

er
su
pp
-

ly
i
Shutter

H
Radiant flux

ea
te
r (R
)
ne

Th
co

er
g
in

m
iv

al
ce

so e

lin
en tur
Re

k
) s ra
r
(T pe
m
Te

Heat sink
(temperature = T 0 )

Figure 6.18. An illustration of the ESR, simplified to bring out its important components.
(This illustration is based on the drawing in NIST Technical Note 1421 [877].)

6.8 Measurement of Optical Radiation

We discuss some of the equipment that may be used to measure various radio-
metric and photometric quantities. These include the integrating sphere, goniora-
diometers, photometers, and spectroradiometers. We begin with the measurement
equation, which lies at the heart of the operation of these measurement devices.

6.8.1 The Measurement Equation

The measurement equation relates the exitant radiation from a surface to the out-
put signal of a radiometric instrument. All radiometric quantities can be measured
by proper manipulation of the measurement equation.
The element of radiant flux incident on the detector surface at the differential
area dA centered around point (x, y), and within a solid angle d ω centered around
direction (θ , φ ), is given by (Figure 6.19)

dP (x, y, θ , φ , λ ) = L (x, y, θ , φ , λ ) dA cos (θ ) d ω d λ . (6.63)

If the responsivity of the detector is RP , each element of radiant flux produces in

the detector a differential output signal dS equal to

dS (x, y, θ , φ , λ , λ0 ) = RP (x, y, θ , φ , λ , λ0 ) dP (x, y, θ , φ , λ ). (6.64)

i i

i i
 i i

i i

348 6. Radiometry and Photometry

(θ,φ)

LS
NS L
θ
dw

Detector area = A
Responsivity = R
dA

Figure 6.19. The detector with responsivity R and total surface area A is configured to
measure the radiance emitted from the light source. For simplicity, the optical system that
typically sits in front of the detector is not shown.

Here, λ0 is the wavelength at which the detector is designed to operate. There

is a range of wavelengths Δλ around λ0 for which the detector has a significant
non-zero response. The total signal S produced by the detector is equal to the sum
of all differential signals dS, yielding the measurement equation:
  
S (A, ω , Δλ , λ0 ) = dS (x, y, θ , φ , λ , λ0 ) d ω dA d λ (6.65a)
Δλ A ω
= RP (x, y, θ , φ , λ , λ0 ) L cos (θ ) d ω dA d λ , (6.65b)
Δλ A ω

where ω is the solid angle subtended by the source being measured at point (x, y).
As an aside, it should be noted that ω may change based on the position (x, y) on
the detector surface, and, therefore, ω is a function of position (x, y).
Equation (6.65b) is valid for the configuration given in Figure 6.19. In this
form, we are measuring the incident radiance on the detector rather than the emit-
ted radiance from the source. In a lossless medium, these would be equal due to
the invariance of radiance. In practice, however, the radiance may be attenuated
by the intervening medium and due to losses during refraction by the optical sys-
tem. As discussed in Section 6.2.8, the incident radiance L is related to the source
radiance LS via the propagance τ of the optical path:

L (x, y, θ , φ , λ ) = LS (x, y, θ , φ , λ ) τ (x, y, θ , φ , λ ). (6.66)

i i

i i
 i i

i i

6.8. Measurement of Optical Radiation 349

Consequently, the final form of the measurement equation becomes

  
S (A, w, Δλ , λ0 ) = RP LS τ cos (θ ) d ω dA d λ . (6.67)
Δλ A ω

In this equation, LS is the radiance to be measured, and S is the output signal of

the measurement instrument. If the instrument is calibrated, then S equals LS .
However, if the device is uncalibrated, a second measurement should be made
whereby the light source is replaced with a standard lamp with known radiance.
For a standard lamp, the measurement equation is
  
Ss (A, w, Δλ , λ0 ) = RP LSs τ s cos (θ ) d ω dA d λ . (6.68)
Δλ A ωs

Here, LSs is the known radiance of the standard lamp. The known quantities are
the results of the two measurements S and Ss and the radiance LSs of the standard
lamp. To solve for LS using (6.67) and (6.68), we assume that the values of LS and
LSs are constant over the integration intervals A, w, ws , and Δλ . In practice, the
aperture size of the measuring device is about 1◦ , and the response is sufficiently
flat over Δλ , making this assumption valid [839]. As a result, LS and LSs may be
taken out of their integrals in (6.67) and (6.68).
In addition, we will assume that τ equals τ s and ω equals ω s . The integrals in
these equations are then identical, so that combining these two equations yields a
solution for LS :  s
LS
LS = S . (6.69)
Ss

6.8.2 The Integrating Sphere

An apparatus commonly used for radiant and luminous power as well as irra-
diance and illuminance measurements is the integrating sphere, also known as
Ulbricht sphere or averaging sphere. The integrating sphere is a hollow sphere
with uniform diffusely reflecting inner walls (Figure 6.20). Light entering the en-
trance port or diverging from a lamp inside the sphere is diffusely reflected several
times until it strikes the detector. The detector is protected from direct light with
a screen, which ensures that radiation is reflected at least once before impinging
upon the detector.
Of the two integrating spheres shown in Figure 6.20, the sphere on the left is
appropriate to measure the radiant power of the light source inside, whereas the
sphere on the right is designed to measure the irradiance at the entrance port.
The usefulness of the integrating sphere lies in the fact that illumination due to
reflected light at any point on the surface of the sphere is the same independent of

i i

i i
 i i

i i

350 6. Radiometry and Photometry

Entrance port

P
P: power
S S S: screen
D: detector
D D

Figure 6.20. The left integrating sphere is designed to measure the radiant power of a
light source whereas the right sphere is designed to measure the irradiance at the entrance
port. In both cases, a screen is used to prevent direct light from reaching the detector.

the position of the light source [1105]. Furthermore, the amount of illumination
bears a direct relation to the power of the light source, as can be derived as follows.
The flux received by a differential surface dR from another differential surface
dA with luminance L can be deduced from (6.38) under the condition that both
surfaces lie on a sphere (see Figure 6.9):
dR
dP = L dA. (6.70)
4r2
The luminance reflected by a surface Lo is equal to the incident luminance Li
times the bi-directional reflectance distribution function (BRDF) fr , which is de-
termined by the reflective properties of the coating of the interior of the integrating
sphere:
Lo (x, y, Θi , Θo ) = fr (x, y, Θi , Θo ) Li (x, y, Θi ). (6.71)
The luminance outgoing into a given direction at a surface element is the result of
all incident luminance that is being reflected into the outgoing direction. Within
an integrating sphere, the integral over all incoming directions is equivalent to an
integral over all surface elements dA:

Lo = fr Li dA. (6.72)
A

The coating of the interior of the integrating sphere is diffuse, and therefore the
BRDF fr can be approximated with a constant, the reflectance factor ρ . The
integrating sphere enables this integral to be evaluated with the help of (6.41d):

dP
Lo = ρ (6.73a)
A π
ρP
= . (6.73b)
π

i i

i i
 i i

i i

6.8. Measurement of Optical Radiation 351

Substitution into (6.38) yields the flux dPi received by differential area dR as a
function of the flux P emitted by the iso-radiant interior surface of the sphere:
dPi ρP
= . (6.74)
dR 4 π r2
This expression is valid for all flux occurring from one reflection inside the sphere.
As light can reflect multiple times before being absorbed or detected by the sensor,
this expression can be refined for multiple reflections:
dPi ρP
= . (6.75)
dR 4 π r2
Finding the illuminance due to reflected light is a matter of tracing light rays
for a few bounces. First, light emitted by the light source will reflect off the
sphere’s walls. We call this reflected luminance Lo,1 . The total flux received by
any area dR due to the first reflection is equal to

dR
dPi,1 = Lo,1 dA. (6.76)
4r2
A

The value of Lo,1 may be different for different points on the sphere. For instance,
the points closer to the light source will have higher luminance. However, the
!
aggregate value Lo,1 dA is related to the total flux of the source via
A
  
ρ dP ρ P
Lo,1 dA = ρ Li,0 dA = = , (6.77)
π π
A A A

where Li,0 is the incident luminance due to direct lighting. Therefore the flux
received by dR due to the first reflection is equal to
dR ρ P
dPi,1 = , (6.78)
4r2 π
and, consequently, the luminance received due to the first reflection equals
dPi,1 ρP
Li,1 = = 2 2. (6.79)
π dR 4π r
At this point, the luminance emitted from any point on the sphere will be equal to
Lo,2 = ρ Li,1 . (6.80)
Similar to the computation just performed, one may compute the flux received by
dR due to the first and the second reflections:

dR
dPi,1,2 = 2 (Lo,1 + Lo,2 ) dA, (6.81)
4r
A

i i

i i
 i i

i i

352 6. Radiometry and Photometry

which can be rewritten as

dR ρ P dR ρ 2 P
dPi,1,2 = + 2 . (6.82)
4r2 π 4r π
After an infinite number of reflections, the flux received by dR can be written as
dR P
dP = dPi,1,2,... = (ρ + ρ 2 + . . . ). (6.83)
4π r 2
Since 0 ≤ ρ < 1, this expression can be further simplified:
dR P ρ
dP = . (6.84)
4π r 2 1 − ρ
And, finally, the irradiance (or the illuminance) on dR is equal to
dP P ρ
E= = , (6.85)
dR 4π r2 1 − ρ
which is independent of the location of the light source and directly proportional
to its power. Thus, by measuring the irradiance or illuminance due to reflected
light (i.e., by means of a detector), it is possible to compute the total flux of
the light source. The first use of an integrating sphere in this manner (i.e., as a
photometer) was made by Ulbricht [1162].
If used for irradiance measurements, rather than the total flux of the source,
the irradiance at the entrance port may be computed from the incident flux and
the area of the entrance port.
So far, we assumed a perfect sphere with perfectly diffusing walls. Unfor-
tunately, the presence of the light source, the screen, the entrance port, and the
detector cause deviations from perfect behavior. This problem may be mitigated
by constructing the sphere as large as possible. For flux measurements, it is not
uncommon to use integrating spheres with a diameter of 5 m [841]. The distance
of the screen to the light source is also important. It is found that optimum results
are achieved with a screen distance of 0.4 r [1121].
Integrating spheres used for irradiance measurements are much smaller, often
in the order of a few centimeters. The main reason to use an integrating sphere
for measuring irradiance is to enable an accurate cosine response. In the absence
of an integrating sphere, radiation arriving from wide angles tends to reflect more
of the detector surface, causing an inaccurate measurement.
The integrating sphere has proven to be useful for purposes other than flux
or irradiance measurements. For instance, if the detector is removed, radiance
emitted from the exit port will be uniform in all directions. Therefore, it can serve
as a uniform diffuser which enables its use for accurate camera calibration.

i i

i i
 i i

i i

6.8. Measurement of Optical Radiation 353

6.8.3 Measurement of Radiant and Luminous Flux

Since radiant and luminous flux are measures of the total radiation emitted from
a source in all directions, the preferred technique to measure these quantities is
by use of an integrating sphere. Dependent on the geometry and the emission
characteristics, the source may be placed at the entrance port or inside the sphere.
For instance, if the emission is limited to a hemispherical solid angle, such as
most LEDs and spot lights, the source is placed at the entrance port facing toward
the center of the sphere. For sources emitting over wider solid angles, including
incandescent and fluorescent lamps, the source must be placed inside the sphere.
The total flux of a test light source may be measured with one of the following
four techniques:

1. the direct measurement method;

2. the substitution method;

3. the simultaneous method;

4. the auxiliary lamp method.

In the direct measurement method, the detector is replaced with a spectroradiome-

ter to obtain the spectral radiant flux. This method has several benefits, in that not
only the total power of the lamp is obtained but also important color properties
such as chromaticity coordinates (Section 7.5.1), the source’s correlated color
temperature (Section 9.2.3), and its color-rendering index (Section 9.3).
In the other three methods, a standard lamp and the test lamp are measured
alternatively. In this case, the ratio of the photo-detector signals is assumed to be
directly proportional to the ratio of the power of each lamp:
 
ST
PT = PS , (6.86)
SS

where S denotes the output signal of the photo-detector and the subscripts S and
T identify the standard and the test lamps.
In the substitution method, the lamps are placed inside the sphere one by one,
preferably in the center of the sphere [1199]. In the simultaneous method, both
lamps are in the sphere together (symmetrically along the vertical axis), although
only one lamp is lit for each measurement. An additional screen is placed between
the lamps to prevent the opaque parts of the lamps absorbing direct light from each
other. In the auxiliary lamp method, a third lamp is placed inside the sphere to
control the self-absorption of light emitted by each lamp itself.

i i

i i
 i i

i i

354 6. Radiometry and Photometry

6.8.4 Measurement of Radiant and Luminous Intensity

Radiant and luminous intensity are measures of the directional flux distribution of
a light source. Therefore, to measure these quantities a means of taking an ade-
quate number of measurements from various angles around the source is
necessary.
An instrument that is used for this purpose is called the goniophotometer (and
similarly gonioradiometer for radiometric measurements). A goniophotometer
allows rotation of an optical detector over 360◦ around a fixed light source both
in vertical and horizontal directions. The frequency of measurements depends
on the directional characteristics of the light source. For instance, if the light
source is fairly uniform, 36 measurements each separated by 10◦ in the horizon-
tal plane may be sufficient [1199]. More measurements are likely to be needed
around the vertical plane, as most sources manifest a greater variation along their
vertical axis.
The movement of the detector follows a spherical path such that the distance
between the measured source and the photo-detector is unaltered between mea-
surements. Also, the light source should be entirely within the field of view of the
detector. All additional areas that are in the field of view must be blackened so
that only the source radiation is measured.

6.8.5 Measurement of Irradiance and Illuminance

Irradiance and illuminance represent the surface density of the incident electro-
magnetic radiation. In most cases, radiation arrives from a hemispherical solid
angle surrounding the surface to be measured.
The chief practical consideration for measuring irradiance is to account for the
cosine law (Section 2.9.2). According to the cosine law, a surface can only receive
radiation proportional to its area projected in the direction of propagation. If the
angle between the surface normal and the incident light rays is θ , the projected
area of the surface in the direction of propagation is equal to

A = A cos(θ ). (6.87)

The difficulty to satisfy the cosine law arises from the fact that as the light rays
arrive at the surface from grazing angles, they also tend to reflect more off the
surface (see Section 2.5.2).5 Without accounting for this effect, irradiance and
illuminance values will be underestimated.
5 This is a different way of saying that near grazing-angles paints tend to be more reflective than a

Lambertian surface.

i i

i i
 i i

i i

6.8. Measurement of Optical Radiation 355

Screen Extruded detector Screen

surface

Figure 6.21. An extruded detector surface. Paths of three of the light rays are depicted
with dashed lines. The rays incident from narrow angles are received by the top of the
detector, while the rays incident from wide angles are received by the edges. Screening
prevents contribution of light from very wide angles.

The solution to this problem is called cosine correction and can be achieved
in several ways. Most detectors have highly diffuse covers to minimize the re-
flectance of the detector. In addition, the detector surface may be extruded to
enable light entrance through the edges of the detector as well. Such a detector is
depicted in Figure 6.21. In this case, light arriving from the horizontal direction
(θ = 90◦ ) should be blocked with proper screening. A second alternative is to use
an integrating sphere to average out light arriving from a very wide solid angle.

6.8.6 Measurement of Radiance and Luminance

Radiance and luminance are best measured with a spectroradiometer, which can
also reveal the wavelength distribution of electromagnetic radiation. Luminance
may be computed from radiance with built-in circuitry.
D
iff
ra

Depolarizer
ct

Aperture
io
n
gr
at

Incoming radiation
in
g

Detector array
380 nm 800 nm

Figure 6.22. Incident radiation passing through the aperture is depolarized and diffracted
into its constituent components. Each component impinges on the detector cell sensitive to
that wavelength.

i i

i i
 i i

i i

356 6. Radiometry and Photometry

A spectroradiometer is composed of an optical system that admits radiation

from a very narrow solid angle and a small surface area. The admitted radiation
is then diffracted into its constituent wavelengths by use of a diffraction grating
as shown in Figure 6.22 (see Section 2.7.3). Each wavelength is directed to a cell
on the detector array which is most sensitive to that wavelength. The number of
cells on the detector array may be less than the total number of wavelengths (e.g.,
a cell for every four wavelengths). This then becomes the wavelength resolution
of the instrument.

6.9 Visual Photometry

Photometric measurements may also be carried out using visual comparison per-
formed by human observers. Using human observers in photometric measure-
ments was predominantly the method of choice until the 1940s. Such methods,
however, suffer from several drawbacks. For instance, results obtained from one
set of measurements can deviate from the results obtained from another, even
for the same group of observers. Thus, beside the inconvenience of an elaborate
set-up, visual photometry has the disadvantages of poor reproducibility and low
precision [185].
A further disadvantage is that humans can only compare lights in a subjective
sense. For instance, one can distinguish the brighter of the two sources, but one
cannot quantitatively say how much brighter that source is. In general, visual
photometry is impractical for daily use. However, a faster and more convenient
method for measuring color matching functions is outlined in Section 7.7 that
largely overcomes the aforementioned disadvantages.

6.9.1 Visual Photometry using Brightness Matches

The fundamental technique used in visual photometry is matching the brightness
of two surfaces illuminated by two light sources. Usually, one of the sources is
taken as a standard with known characteristics, and the other is a test source for
which the measurement is made.
The surfaces lit by these light sources are presented to the observer by means
of an instrument called the photometer head (see Figure 6.23). The observer can
adjust the amount of light reflected by the surfaces by moving the light sources
closer to or farther from the photometer head, or by placing neutral density filters
with different densities over the light sources.
In this system, observers are asked to match the brightness of the two surfaces
in the photometer head illuminated by the two light sources. Several carefully

i i

i i
 i i

i i

6.9. Visual Photometry 357

ds dt

Bench

Screens Head Screens

Figure 6.23. A photometer based on the inverse square law.

placed screens are used to block stray light so that light enters the photometer
head only from the light sources. One of the sources is a known standard whose
intensity is previously established, and the other source is a test source whose
intensity is to be measured. The observer is able to adjust the brightness of the
illuminated surfaces by sliding the lights towards and away from the photometer
head, effectively using the inverse-square law (Section 2.9.3)
The brightness match is obtained when the illumination of both surfaces inside
the photometer head appear the same:

Is ρs It ρt
= 2 , (6.88)
ds2 dt

where Is and It are the intensities of the standard and the test sources, ρs and ρt
are the reflectances of the illuminated surfaces, and ds and dt are the distances of
the sources from the surfaces. The measured term It may then be computed with
 2
ρs dt
It = Is , (6.89)
ρt ds

or, if the surfaces have the same reflectance,

 2
dt
It = Is . (6.90)
ds

Although the photometer in this example is designed to measure luminous

intensity, other photometric quantities may be measured similarly with different
configurations of the instrument. For instance, since luminance is intensity per
unit area, luminance at any point of the light source may be measured by admitting
light from a smaller portion of the source.
It should be noted that there are numerous instruments and measurement tech-
niques used in visual photometry. An excellent reference is Walsh [1199].

i i

i i
 i i

i i

358 6. Radiometry and Photometry

6.9.2 Heterochromatic Photometry

It has been implicitly assumed in the previous discussion that the compared
sources are the same or very similar in color. In principle, when observers are
making brightness matches, this should be accomplished under specific test con-
ditions, including the use of identical or at least very similar spectral distributions
of test and reference sources.
However, there are numerous situations where lights with different spectral
distributions need to be compared. Even for two metameric light sources, their
spectral distribution is unlikely to be identical. Therefore when making photomet-
ric measurements, techniques to mitigate the effects of color differences should
be considered. Visual photometry carried out with stimuli of different spectral
composition is called heterochromatic photometry.
The first solution is comparing lights of similar colors. Even though the spec-
tral compositions of the lights may not be identical, if the differences are subtle
these might not be perceived by the eye and may not harm the measurement. For
instance, with a difference of 100◦ K in color temperature, measurements can be
performed with an accuracy of about 1 percent [1199] (color temperature is dis-
cussed in Sections 3.5.1 and 9.2). When the color differences are larger, the gap
in color between the sources may be divided into steps, and at each step an inter-
mediate auxiliary lamp may be used. Then each lamp is compared with the next
one, forming a link between the original sources that are compared.
Another technique is called the compensation method or the mixture method.
In this method, light emitted from one of the sources is mixed with light from an
auxiliary source of a different color forming a mixture which is similar in color
to the other source. For instance, by adding a small amount of blue light to light
emitted from a standard source of low color temperature, its color temperature
may be increased to a level where it becomes comparable to the test source. As
long as the intensity of the added blue light is known, the intensity of the test
source can be calculated.
A fundamental but more involved method splits the radiation emitted from
each light into its spectral components that are monochromatic and compares each
component separately to find the intensity of each. These intensities are then
summed to find the total intensity of the test source. This process may be carried
out with spectrophotometers. In this method, measurements should be made with
care, since errors introduced for each component will also add up in the final
result.
It is possible to use color filters to match the spectral distributions of the com-
pared lights. If a specially designed color filter is placed in front of one of the
lamps, the spectral distribution of the light passing through this filter will match

i i

i i
 i i

i i

6.10. Application: Measuring Materials 359

that of the other light with which the comparison is made. Then, classical meth-
ods of homochromatic photometry may be used to measure the intensity of the
transmitted light. Of course the transmission characteristics of the filter should
be known (or measured with some other method) to adjust the measured intensity
accordingly.
In the case of large chromatic differences between source and test patches, it
is still possible to carry out brightness matches with good accuracy, for instance,
with the flicker photometer method. An instrument called the flicker photometer
is used for the purpose of presenting two lights on a surface in alternation at a
minimum frequency such that when the brightness of the two lights match, the
perception of flicker will disappear [1199]. The method relies on the fact that the
temporal resolving power of the human eye is lower for colored stimuli than for
achromatic stimuli. This means that there exist frequencies for which chromatic
differences between the alternating patches is not perceived. By adjusting the
intensity of one of the patches, a brightness match occurs if the perceived achro-
matic differences disappear as well. Typical frequencies are between 10 Hz and
40 Hz [1262].
While a heterochromatic flicker-photometry experiment can be constructed
with a physical device whereby synchronized shutters are opened and closed in
opposite phase, an accurate experiment carried out on a display device is compli-
cated. Computers and graphics cards are getting ever faster, but are rarely able
to guarantee a fixed and stable frame-rate. Drift and fluctuations are likely to
occur, and they affect the accuracy of the measurement [591]. A general prob-
lem with heterochromatic flicker photometry is that the frequencies required to
measure brightness matches may induce photosensitive epileptic seizures in some
individuals, making the technique usable only in controlled laboratory experi-
ments [530].
Brightness matching, in the case of large chromatic differences between
source and test patches, moves into the realm of color matching, which is dis-
cussed in the following chapter on colorimetry. There, we will also introduce
a recent technique to improve the accuracy of visual brightness matching that
can be performed on computer displays and show how it can be used to derive
observer-specific iso-luminant color maps (Section 7.7.1).

6.10 Application: Measuring Materials

Measurements of object reflectance and light emission are important in many dif-
ferent fields, for instance in lighting design. In computer graphics, accurate image

i i

i i
 i i

i i

360 6. Radiometry and Photometry

synthesis can only be achieved if modeling, rendering, and display are all based
on accurate data and measurements. Even the best renderer is still a garbage-in-
garbage-out system, meaning that if the modeling is not accurate, the resulting
image will not be accurate.
Modeling of an environment requires the acquisition of geometry, the emis-
sion spectra of light sources, as well as the measurement of materials. In this
section, we consider the problem of measuring materials. Opaque spatially ho-
mogeneous materials can be modeled with bi-directional reflectance distribution
functions (BRDFs), which were introduced in Section 2.9.7. Such functions re-
late the amount of light reflected in a direction of interest to the amount of light
incident upon a point on a surface. The relation is given as a ratio and may be
defined as a function of wavelength.
With incident and exitant direction defined as two-dimensional polar coordi-
nates, a BRDF is a five-dimensional function. For isotropic BRDFs the dimen-
sionality is reduced to four, as rotation of the surface around the surface normal
at the point of interest does not alter the ratio of reflected light. If the BRDF is, in
addition, measured with tristimulus values rather than a full spectral set of sam-
ples, the wavelength dependence is removed, and the dimensionality is reduced
to three dimensions.
This means that to measure an isotropic BRDF for the purpose of using it in a
renderer requires a material sample to be measured along three dimensions. The
traditional way to measure a BRDF is with the use of a gonioreflectometer.6 A
flat sample containing the material to be measured is placed in a fixed position.
A calibrated point light source can move along one direction over the hemisphere
placed over the sample material. The detector can independently move over the
entire hemisphere. Together, they span the three degrees of freedom required to
sample all possible combinations of incident light and measurement angles, as
shown in Figure 6.24 [744]. Anisotropic BRDFs can be measured by giving the
position of the light source a second degree of freedom.
Every measurement, thus, provides a data point for one angle of incidence
paired with one angle of reflectance. A similar result can be obtained by plac-
ing the detector in a fixed position and rotating the sample over two degrees of
freedom. The light source is then still moved over the same path as before.
In either set-up, the number of samples needs to be large to get a dense sam-
pling of the BRDF. Such a dense sampling is required in the absence of further
knowledge about the sort of material being measured. This makes the acquisition
of a full BRDF long-winded, even if the gantry maneuvering the detector and the
6 See the Cornell Light Measurement Laboratory (http://www.graphics.cornell.edu/research/

measure/) and the Nist reference reflectometer at the Starr facility (http://physics.nist.gov/).

i i

i i
 i i

i i

6.10. Application: Measuring Materials 361

Detector

Light source

Sample

Figure 6.24. In a gonioreflectometer, the light source and detector have one and two
degrees of freedom to move over the hemisphere placed over the sample to be measured.
This configuration enables the acquisition of isotropic BRDFs.

light source is under computer control. If specific materials are being measured,
then it may be possible to measure only a subset of all angles of incidence and
reflectance.
A different approach to make the acquisition procedure tractable is to capture
many samples at once, usually by employing a digital camera. Two of the de-
grees of freedom in the measurement device are then replaced by the two image
dimensions. For flat samples this may be achieved by placing a hemispherical
mirror over the sample and photographing it through a fish-eye lens [1212]. A
perhaps simpler hardware set-up may be achieved by requiring the sample to be
curved. Spherical [744,745,751,752] or cylindrical samples [701] are simplest to
capture, since a single photograph will capture many different orientations of the
surface normal with respect to the light source. The requirement of simple shapes
exists because this allows an analytic model to infer the surface orientation at each
pixel.
More complex material samples can also be used, as long as they are convex
to avoid inter-reflections within the sample material. In that case, the camera
may be augmented with a laser scanner device to capture the 3D geometry of
the sample [744]. From the geometry, the surface normals at each pixel can be
inferred; these are, in turn, used to relate pixel intensities to angles of incidence
and angles of reflectance.
For concave objects one may expect significant inter-reflections that will ren-
der the material measurement inaccurate. One would have to separate the pixel
values into direct and indirect components. While to our knowledge this has not
been achieved in the context of material measurements, separation of direct and
indirect components is possible by placing an opaque surface with small regu-
larly spaced holes (an occlusion mask) in front of the light source. After taking

i i

i i
 i i

i i

362 6. Radiometry and Photometry

multiple captures with the occlusion mask laterally displaced by small amounts,
image post-processing techniques can be employed to infer the direct and indirect
illumination on surfaces separately [818].
BRDFs can reflect almost all incident light. If, in addition, the BRDF is highly
specular, then the amount reflected in the reflected direction is only limited by the
intensity of the light source. As a result, the imaging system should be able to
capture a potentially very large dynamic range. The capture of high dynamic
range images is the topic of Chapter 12.
The number of materials that can be represented by a BRDF, although large,
is still limited as it assumes that all surface points have identical reflectance prop-
erties. It is possible to extend the measurement of BRDFs to account for spatially
varying reflectance. Such spatially varying reflectance functions are called bi-
directional texture functions (BTFs). They can be measured with the aid of a
robot arm, a light source, and a digital camera using flat samples [231, 232].7
In the case of BRDFs, and especially BTFs, the number of samples required to
represent a material is very large. It is therefore desirable to fit this quantity of data
to analytic BRDF models. This can be achieved in several different ways. Some
of the resulting models are discussed in Section 2.9.7. For further information,
we refer the interested reader to the relevant references [231, 582, 601, 637, 639,
640, 671, 744, 764, 998, 1013, 1212, 1240, 1285].

6.11 Further Reading

Radiometry and photometry are explained in Wyszecki and Stiles’ Color Science:
Concepts and Methods, Quantitative Data and Formulae [1262]. There are sev-
eral dedicated books, including Introduction to Radiometry and Photometry by
McCluney [761], Optical Radiation Measurements by Grum and Becherer [397],
Handbook of Applied Photometry by DeCusatis [242], and Photometry and Ra-
diometry for Engineers by Stimson [1090].

7 A collection of BTFs and BRDFs obtained with this technique is available from the Columbia-

Utrecht reflectance and texture database at http://www.cs.columbia.edu/CAVE/software/curet/index.

php.

i i

i i
 i i

i i

Chapter 7
Colorimetry

Color is a purely psychological phenomenon; that is to say, without a human

observer there could be no color. Though perceptual in nature, that does not mean
that color cannot be well measured, quantified, and communicated. Colorimetry
is the science of color measurement and description. To measure color, we must
take into account the physical properties of the environment including material
and light sources, as well as the physiological properties of the human visual
system.
Since color is ultimately a human response, the science of color measurement
should begin with human observations. As we learned in Section 4.3.1, the hu-
man visual system contains two types of photodetectors; rods and cones. The
three types of cones are primarily responsible for color vision. The simplest form
of color measurement might be to directly record the cone responses for any given
stimulus. While this might create a direct physiological measurement, the poten-
tially invasive procedure, not to mention the variation between any two individ-
uals, probably outweighs its usefulness. Another method for color measurement
that correlates with the human cone responses is necessary, namely colorimetry.
This chapter outlines the experimental procedures, as well as statistical mod-
els, used to measure and specify colors. We will start with an explanation of
Grassmann’s laws of additive color matching. Using these laws, we will build
a visual color-matching experiment, as well as a visual colorimeter. Historical
color-matching experiments and their results will be discussed. These results have
resulted in several standardized “observers,” which can be thought of as statistical
approximations to actual human observers. Color measurements can be specified
through integration of these standard observers, as discussed in Section 7.5. Fi-
nally, the practical applications of colorimetry and the standard observers will be
discussed.
363

i i

i i
 i i

i i

364 7. Colorimetry

It is very important to stress that although colorimetry is based upon the hu-
man visual system, it is not a technique for describing how colors actually appear
to an observer. It is purely a method for accurately quantifying color measure-
ments and describing when two colors “match” in appearance. The ultimate vi-
sual response to any stimulus requires much more information about the viewing
conditions, including (but not limited to) size, shape, and structure of the stimulus
itself, color of surround, state of adaptation, and observer experience. The sci-
ence of describing what a stimulus actually looks like is called color appearance
modeling and is described in detail in Chapter 11.

7.1 Grassmann’s Laws

Before we can begin any discussion of colorimetry, or visual color matching, it
is important to make a few assumptions. Wyszecki and Stiles sum up the experi-
mental laws of color matching as those of trichromatic generalization [1262]. The
trichromatic generalization states that since humans have three fundamental cone
types, any color stimulus can be matched completely with an additive mixture
of three appropriately modulated fixed primary-color sources. We see this gen-
eralization in practice daily, as most television and desktop monitors utilize only
three primaries (typically red, green, and blue) to recreate most of the colors of
the world. In practice, there are limitations to the total number of colors that can
reproduced with a given set of primaries, known as the reproducible gamut of the
display device. More details on display devices can be found in Chapter 14.
The trichromatic generalization allows us to make color matches between any
given stimulus and an additive mixture of other color stimuli. These additive
color matches follow the general rules of algebra. The mathematical rules of
color matching were described by Grassmann in 1853 and are known as Grass-
mann’s laws of additive color matching [386]. These laws, like typical algebraic
expressions, can be explicitly broken down into four rules [1262]:

Symmetry Law. If color stimulus A matches color stimulus B then B matches A.

Although this may seem obvious, it is included to show the similarities
between Grassmann’s laws, and generalized algebra.

Transitive Law. If color stimulus A matches B, and color stimulus B matches C,

then color stimulus A matches C. Again, this should seem obvious.

Proportionality Law. If color stimulus A matches B, then α A matches α B, where

α is any positive scaling factor. It is important to stress, that for the propor-
tionality law to hold, only the radiant power of color stimulus can change;

i i

i i
 i i

i i

7.1. Grassmann’s Laws 365

Radiant Power
A+B
A
B

400 700
Wavelength (nm)

Figure 7.1. Additive color mixing of the spectral power distributions of two color stimuli
A and B.

the overall spectral power distribution cannot change. This can be impor-
tant when designing color-matching experiments. For instance, if the three
primaries are tungsten light bulbs, the radiant power cannot be changed
by decreasing or increasing the current to the light source as this will also
change the fundamental spectral power distribution of the lights.
Additivity Law. The additivity law is the most important law for generalized color
matching and forms the basis for colorimetry as a whole. Essentially, if
there are four color stimuli A, B, C, and D, and if two of the color matches
are true, i.e., A matches B and C matches D, A + C matches B + D, then it
follows that A + D matches B + C.

When we talk about additive color mixing, in these particular examples, we

are essentially talking about adding the spectral power distributions of radiant
light. That is to say, we are assuming there are no interactions between wave-
lengths, and we simply add the power of two stimuli or sources at any given
wavelength interval. An example of this is shown in Figure 7.1.
It is important to understand that additive color mixing is generally only valid
for adding light itself. It should not be assumed that material properties such
as transmittance and reflectance will behave in a linear manner. Wyszecki and
Stiles also stress that there are three fundamental considerations that are ignored
regarding Grassmann’s laws [1261]:
1. The dependence on the observational conditions on a match. If two stimuli
match under one given viewing condition, there is no reason to assume that

i i

i i
 i i

i i

366 7. Colorimetry

the match will continue to hold if one of the stimuli is placed in a different
condition.

2. Observer adaptation state. Although the match will generally hold if the
observer’s adaptation state changes while viewing both stimuli, an effect
known as persistence of color matches, the match may break down when
viewing the stimuli independently under two different adaptation states.
More details on chromatic adaptation are given in Chapter 10.

3. The dependence on the match for a given observer. If two stimuli (with
disparate spectral power distributions) match for one person, there is no
reason to assume that the match will hold for another person. The vari-
ability between color matches among the general population is surprisingly
large. When we add in the color deficient population, it should be obvious
to the casual observer that what they consider to be a color match will not
necessarily match for all (or any!) other people.

Grassmann’s laws provide the framework for developing a robust method for
measuring and communicating colors. Remember that an ideal method for mea-
suring colors would be to directly measure the human cone responses for any
given stimulus. Unfortunately this is not a feasible method for specifying color.
If we had the spectral sensitivities of the cone photoreceptors, as described in
Section 4.3.3, it would be possible to estimate the response through mathematical
integration of the color stimulus and the cones. It is only within the last several
decades that we have had a solid understanding of what the cone responses actu-
ally are. So, historically, we have not been able to use the cones as a method for
describing color measurements.
Using Grassmann’s laws of additive color mixing, researchers in the early
1920s were able to estimate the cone responses in a series of color-matching ex-
periments. Section 7.2 describes those historical experiments and how they were
used to generate computational systems of colorimetry.

7.2 Visual Color Matching

If we knew the cone responses, it would be easy to specify color matches in
terms of integrated responses to radiant power stimuli. Thus, for each of the
three cone types, the radiant power of the stimulus is multiplied on a wavelength-
by-wavelength basis with the cone spectral sensitivity, and then the response is
integrated across all wavelengths. This concept is illustrated in (7.1), where Φ (λ )

i i

i i
 i i

i i

7.2. Visual Color Matching 367

is the spectral power distribution of the stimulus, and L (λ ), M (λ ), and S (λ ) are

the relative cone spectral sensitivities:

L1 = Φ (λ )L (λ ) d λ , (7.1a)
λ

M1 = Φ (λ )M (λ ) d λ , (7.1b)
λ

S1 = Φ (λ )S (λ ) d λ . (7.1c)
λ

From these equations, we can state that if the relative integrated cone responses
L1 , M1 , and S1 for one stimulus equal those for a second, L2 , M2 , and S2 , then
the two color stimuli must, by definition, match.
The integrated LMS cone responses can be thought of as tristimulus values.
We can now state that two color stimuli match when their integrated tristimulus
values are identical. Knowing the cone spectral-power distributions would allow
us to perform these tristimulus calculations to specify color measurements. Un-
fortunately, it is only recently that we know how to obtain cone sensitivities. His-
torically, they have been very difficult to measure. If the cone sensitivities were
idealized, such as those shown in Figure 7.2, then they would be easy to mea-
sure. We could just use a monochrometer to scan the visible wavelength range
and record the relative response to each wavelength. The real cone spectral sen-
sitivities, as discussed in Section 4.3.3 and shown again in Figure 7.3 have a high
degree of overlap. This overlap means that it is impossible to isolate the visual
response to a single cone excitation and impossible to experimentally measure the
individual cone sensitivities. Other techniques for estimating the cone responses
are necessary.
At this point, one might question why our eyes have evolved in such a way that
our photodetectors are so highly overlapping. From an information-processing
point of view, this suggests that the LMS cone signals are highly correlated and
not data-efficient. This is true, although processing at the ganglion level does
partially de-correlate the cone signals into opponent colors before transmission to
the brain. The high degree of cone overlap actually provides the human visual
system with a very high sensitivity to wavelength differences by examining the
relative LMS signals with each other.
Imagine two monochromatic lights with the same radiant power at 540 nm and
560 nm. If these lights are viewed using the idealized cone spectral sensitivities
shown in Figure 7.2, only the middle sensitive cone will be excited. The integrated

i i

i i
 i i

i i

368 7. Colorimetry

1.0

sensitivity
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2 L
M
0.1 S
0.0
400 450 500 550 600 650 700
wavelength (nm)

Figure 7.2. Idealized cone spectral sensitivities. The cones evenly sample the visible
wavelengths and have minimal overlap.

response to both the lights will be the same, and the observer will not be able to
distinguish the color difference between the two light sources.
Now consider the same two light sources, and the real cone spectral sensitiv-
ities as shown in Figure 7.3. All three cones will respond to the 540 nm light,
while the L and M cones will respond to the 560 nm light. The relative LMS
responses to the two lights will be very different, however. By comparing the

1.0
sensitivity

0.9 L
M
0.8 S
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
400 450 500 550 600 650 700
wavelength (nm)

Figure 7.3. Actual cone spectral sensitivities. Notice the high degree of overlap (correla-
tion) between the three cone types. (From [1091].)

i i

i i
 i i

i i

7.2. Visual Color Matching 369

relative responses, it is then possible for the visual system to distinguish the two
color stimuli as being different. This behavior is discussed in further detail in
Section 7.2.1.

7.2.1 Univariance and Metamerism

An important consideration for colorimetry, and generating color matches in gen-
eral, should be immediately obvious when viewing (7.1). The act of integrating
the cone responses suggests that our visual system does not act as a wavelength
detector. Rather, our photoreceptors act as approximate linear integrators. That
means that two color stimuli that have distinctly different spectral power distribu-
tions can elicit identical cone responses. In the example given in the previous sec-
tion, we saw that two different light sources can cause identical cone responses.
This is known as the principle of univariance.
Another example of this can be seen in Figure 7.4. Two monochromatic light
sources with the same radiant power, at approximately 420 nm and 475 nm have
the same integrated response from our short wavelength S cones. The relative
responses of the L and M cones is different, so our visual system is still able to
detect the difference between these two light sources in this example.
The principle of univariance, and the fact that our cones act as photo-
integrators, allows two color stimuli that have very different spectral-wavelength
properties to appear to match to a human observer. This is known as metamerism
and allows for a wide range of color-reproduction devices. Essentially, we can

1.0
sensitivity

0.9 L
M
0.8
S
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
400 450 500 550 600 650 700
wavelength (nm)

Figure 7.4. Demonstration of the principle of univariance. Two monochromatic light

sources at different wavelengths integrate to the same S response.

i i

i i
 i i

i i

370 7. Colorimetry

1.0

sensitivity
0.9 L
M
0.8 S
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
400 450 500 550 600 650 700
wavelength (nm)

Figure 7.5. Iconic example of metamerism. The two color stimuli, represented by the solid
and dashed black lines, integrate to the same LMS signals and appear to match, despite
have markedly different spectral power distributions.

now conceive of creating a color match between any given spectral power distri-
bution by generating a differing spectral power distribution that integrates to the
same LMS cone responses. An iconic example of this is shown in Figure 7.5.

7.2.2 A Device for Visual Color Matching

Using this principle of metamerism, it is possible to conceive of a device that
allows us to create color matches for arbitrary color stimuli using a small number
of known color primaries. The simplest concept may be to think of using your
computer and a display device, such as an LCD panel. We know that we can
adjust the red, green, and blue digital counts sent to the LCD to generate arbitrary
colors.
Suppose you have a favorite color flower on your desk that you want your
friend who is in a different country to see. You could, conceptually, open up your
favorite graphics software and create a simple color patch that matches the color
of the flower. You could then send your friend your computer and LCD screen
and have them examine the color patch that you have created. They will then
know what color the flower on you desk is. This is somewhat inconvenient, but in
essence you have created what is known as a visual colorimeter; a device used to
generate color measurements (of the flower) visually.
If your friend has the identical computer and LCD panel, it may be possible
to simply send them the graphics file that contains the correct red, green, and blue

i i

i i
 i i

i i

7.2. Visual Color Matching 371

values so that they can open that file on their own system. If the manufacturer of
the computer and LCD has excellent quality control, then it is possible that your
friend will see the same color that matches the flower on your desk. Of course,
because color perception depends on many factors other than just the display it-
self, such as whether the computer is viewed in the dark or in a brightly lit office,
there is no way to guarantee that your friend is seeing the same color.
Now we can imagine a device specifically designed to generate these types
of visual color matches. It could be as simple as a box with three light sources
and a screen. We can generate color matches by adjusting the relative amounts of
those three primaries and then send those coordinates to anyone else in the world
who has the same device. If the device is viewed in identical conditions, then
we can be relatively satisfied that when they dial in the same relative amounts
of the primaries, they will get a color match. With this concept, it is possible
to measure the color of any stimulus in the world simply by dialing in a color
match. This form of color measurement does not need to have any knowledge of
the actual cone spectral sensitivities, as long as we create the color matches with
a consistent set of primaries.
There are obvious practical limitations to this type of color measurement. For
one, it is quite inconvenient to have to create visual matches for all measurement
needs. There is also the problem with the difference between matches created by
different people. What matches for one person will quite possibly not match for
another person.
If we instead perform this type of visual color matching for a wide variety
of colored stimuli and a wide variety of observers, it is possible to generate an
“average” color-matching data set. Experimentally, this can be performed using a
set-up similar to that shown in Figure 7.6.
An observer looks at an opening in a wall and sees a circular patch with two
colors shown in a bipartite field. His task is to adjust the colors of one of the fields
such that it matches the other. Behind the wall are three light-source primaries on
one side and a single monochromatic light source on the other, all pointing at
a white screen. The observer can adjust the relative amounts of the three pri-
maries (without altering the spectral power distribution) to create a match for the
monochromatic light. He can then repeat this procedure across all wavelengths in
the visible spectrum.
In practice, as shown in Figure 7.6, it is necessary to also have adjustable
primaries in the reference field. This is because monochromatic light appears
very saturated to humans, and it is difficult to match with three broad primaries.
Adding light from the broad primaries to the monochromatic light will desaturate
the appearance. In essence, using Grassmann’s laws as described in Section 7.1,

i i

i i
 i i

i i

372 7. Colorimetry

Test Field Reference Field

Figure 7.6. A generic color-matching experiment. The observer adjusts the relative pri-
maries of the test field to create a match to the monochromatic illumination in the reference
field.

this is equivalent to subtracting light from the test field. Algebraically, a color
match can then be defined using (7.2), where R, G, and B are the test primaries, r,
g, and b are arbitrary units or amounts of those primaries, and λ is the monochro-
matic test light:

R(r) + B(b) + G(g) = λ − R(r2 ) − B(b2 ) − G(g2 ). (7.2)

This is exactly the type of experiments that were performed by two scien-
tists in the early 1920s. Wright performed this experiment with ten observers
in 1928–1929 using monochromatic primaries in the test field [1253]. Indepen-
dently, Guild performed a similar experiment with seven observers, using broad-
band primaries in 1931 [405]. Both experiments used similar set-ups, with the
reference and test bipartite field subtending approximately two degrees (2◦ ) and
viewed in darkness. The two-degree field is the approximate size of the fovea in
the retina, as discussed in Section 4.2.7. These experiments generated two sets
of spectral tristimulus values that are essentially the relative amounts of each of
the RGB primaries needed to match any given wavelength. In (7.2), these are the
amounts of r, g, and b needed at each wavelength.

i i

i i
 i i

i i

7.3. Color-Matching Functions 373

0.5

sensitivity
r (λ)
g (λ)
0.4
b (λ)

0.3

0.2

0.1

0.0

−0.1
400 450 500 550 600 650 700
wavelength (nm)

Figure 7.7. The spectral tristimulus values from the Wright and Guild experiments, aver-
aged and transformed to the primaries at 435.8, 546.1, and 700 nm. These spectral tristim-
ulus values are also referred to as the CIE r̄ (λ ), ḡ (λ ), b̄ (λ ) color-matching functions.

7.3 Color-Matching Functions

In 1931, the colorimetry committee of the CIE examined the two data sets from
the Wright and Guild data. They first transformed the data from the two experi-
ments into a common set of primaries. Basically, this involves describing the pri-
maries from one experiment in terms of spectral tristimulus values of the second
experiment. Both data sets were converted to a set of primaries corresponding to
435.8 nm, 546.1 nm, and 700.0 nm. More details of this conversion can be found
in Berns [93] and Wyszecki [1262].
Once in this common primary space, the CIE found the two data sets in close
enough agreement, and so they averaged the results of the Wright and Guild ex-
periments. This agreement verified the general assumptions used to generate the
color-matching data, as well as the linear manipulations allowed by Grassmann’s
laws. The spectral tristimulus values for the three monochromatic primaries de-
scribed above are shown in Figure 7.7.
The tristimulus values needed to create a match for any given wavelength can
be both positive and negative. Since it is impossible to have negative amounts of a
primary, this indicates that it is necessary to add that amount of the primary to the
reference field, as shown in the generic color-matching experiment of Figure 7.6.
Another way to think about the negative tristimulus values, is that the particular
monochromatic illumination is outside the gamut of the primaries used in the
matching experiment.

i i

i i
 i i

i i

374 7. Colorimetry

The spectral tristimulus values indicate the amount of each of the three pri-
maries (hence, tristimulus) necessary to generate a color match, for any given
spectral wavelength. For instance, at 440 nm, we would need approximately
−0.0026 units of R, 0.0015 units of G and 0.3123 units of B. These curves
can also be thought of as color-matching functions. We can can determine the
amount of each primary necessary to create a color match by integrating the
spectral power distribution of the color stimuli with the color-matching functions.
These integrated values can then be thought of as tristimulus values. (Note, we no
longer need to refer to them as spectral.) The generic form of this calculation is:


R= Φ (λ ) · r̄ (λ ) d λ , (7.3a)
λ

G= Φ (λ ) · ḡ (λ ) d λ , (7.3b)
λ

B= Φ (λ ) · b̄ (λ ) d λ . (7.3c)
λ

In these equations, R, G, and B are the tristimulus values, Φ is the color-

stimuli spectral power distribution, and r̄ (λ ), ḡ (λ ), b̄ (λ ) are the CIE rgb color-
matching functions shown in Figure 7.7. Using these tristimulus values to mea-
sure color, it is now possible to determine if two stimuli match. If their RGB
values are the same, then the colors match. We can eliminate the need for visual
observation by calculating these relationships directly. The generalized relation-
ship for a color match is shown in (7.4). It is important to stress that all three
tristimulus values must be the same, in order to guarantee a color match.

 
Φ1 (λ ) · r̄ (λ ) d λ = Φ2 (λ ) · r̄ (λ ) d λ , (7.4a)
λ λ
 
Φ1 (λ ) · ḡ (λ ) d λ = Φ2 (λ ) · ḡ (λ ) d λ , (7.4b)
λ λ
 
Φ1 (λ ) · b̄ (λ ) d λ = Φ2 (λ ) · b̄ (λ ) d λ . (7.4c)
λ λ

We now have a method for calculating when colors match, or we can measure
a color, based upon the spectral power distribution of the stimuli. It should be
possible to communicate the RGB tristimulus values to our friend and have them

i i

i i
 i i

i i

7.4. CIE 1931 and 1964 Standard Observers 375

generate a color match using their three monochromatic primaries at 435.8, 546.1,
and 700 nm. If they do not have access to those primaries, they can calculate the
appropriate linear transform (3 × 3 linear matrix) necessary to generate a match
with the primaries that they do have. One such transformation, to an imaginary
primary set, is discussed in Section 7.4.

7.4 CIE 1931 and 1964 Standard Observers

Using Grassmann’s laws of additive color mixing, as well as the transformation of
primaries, the CIE was able to calculate a set of color-matching functions based
upon the 17 observers from the Wright and Guild experiments. These color-
matching functions, r̄ (λ ), ḡ (λ ), and b̄ (λ ), were based upon a specific set of
real primaries: monochromatic lights at 435.8, 546.1, and 700 nm.
These primaries were chosen because the first two were spectrum lines from
a mercury lamp, while the 700 nm light was chosen so as to isolate the L cone
as much as possible, using the knowledge of the visual system available at the
time [93]. In 1931 numerical integration techniques were difficult and expensive
to perform, and integration with both positive and negative numbers (as necessary
using the CIE RGB color-matching functions) were even more difficult. Addi-
tionally, the CIE desired to combine the system of colorimetry with the 1924 CIE
system of photometry, as discussed in Chapter 6. Having a fourth integrating
function, CIE V (λ ) would require more computation than what was desired.
The CIE set out to define a transform to a new set of primaries with two
specific goals: having an all positive response and having one of the primaries,
or color-matching functions, be the photopic luminance response function, CIE
V (λ ). More details on the derivation of this transformation can be found in Fair-
man [304].
This linear transformation from RGB to an all positive set of color-matching
functions required a set of imaginary primaries. That is to say, in order to suc-
cessfully reproduce all colors in the visible spectrum, we need to use lights that
cannot be physically realized. The color-matching functions settled upon were
referred to as x̄ (λ ), ȳ (λ ),and z̄ (λ ), and the integrated tristimulus values are CIE
XYZ so as to not be confused by realizable RGB tristimulus. Using imaginary
primaries means that it is impossible to create a device to actually recreate the
color matches, as we could using the three monochromatic primaries described in
Section 7.3. However, this just means that we use this new system of colorimetry
strictly in a computational manner. The linear transformation from CIE RGB to
XYZ is shown in Equation 7.5:

i i

i i
 i i

i i

376 7. Colorimetry

1.8

sensitivity
x (λ)
1.6
y (λ)
1.4 z (λ)
1.2
1.0
0.8
0.6
0.4
0.2
0.0
400 450 500 550 600 650 700
wavelength (nm)

Figure 7.8. The CIE 1931 x̄ (λ ), ȳ (λ ), and z̄ (λ ) color-matching functions.

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
X 0.4900 0.3100 0.2000 R
1 ⎣ 0.17697
⎣ Y ⎦= 0.81240 0.01063 ⎦ · ⎣ G ⎦ . (7.5)
0.17697
Z 0.0000 0.0100 0.9900 B

1
The scaling factor 0.17697 shown in this equation was used to normalize the
functions to the same units as CIE V (λ ). Applying (7.5) to the r̄ (λ ), ḡ (λ ), and
b̄ (λ ) color-matching functions gives us the x̄ (λ ), ȳ (λ ), and z̄ (λ ) color-matching
functions:
The very first thing to notice about these color-matching functions is that,
indeed, they are all positive. It should also be immediately obvious that these
functions do not look like the human cone spectral sensitivities, although we can
use them to calculate color matches in much the same way as we could with
cones. The ȳ (λ ) color-matching function should look almost identical to the
1924 photometric standard observer. These color-matching functions have come
to be known as the CIE 1931 standard observer, or similarly the CIE 2◦ -observer;
they are meant to represent the color-matching results of the average human pop-
ulation [186].
The term 2◦ -observer comes from the fact that both Wright and Guild used
a bipartite field that subtended approximately two degrees of visual angle, and
as such all the color matches were made using just the fovea. As we learned
in Section 4.3.1, this meant that the color matches were performed using only the
cones while eliminating any rod contribution. This was vitally important, because

i i

i i
 i i

i i

7.4. CIE 1931 and 1964 Standard Observers 377

Grassmann’s laws of additive color mixing can break down when there are both
rod and cone contributions to color mixing.
What this means, is that the CIE 1931 standard observer is designed to predict
color matches of very small stimuli (think a thumbnail at arm’s length). Another
important consideration is that the average human response used for color mea-
surements was taken from a pool of 17 observers, over 75 years ago! Despite the
age, and apparent limitations of the 1931 standard observer, it is still used with
great success today and is still the foundation of modern colorimetry.
Not willing to rest on their laurels with the introduction of the 1931 stan-
dard observer, the CIE continued to encourage active experimentation to both
validate the existing color-matching functions and to test the use of them for
larger color patches. Through observations lead by Stiles at the National Physics
Laboratory in the UK, with ten observers using both 2◦ and 10◦ , it was deter-
mined that the 1931 standard observer was appropriate for the small-field color
matches [1262].
There were discrepancies between the small field results and the large field re-
sults, and so further testing was required. Stiles and Burch proceeded to
measure an additional 49 observers using the 10◦ -bipartite field [1087]. The
Stiles and Burch data was collected at a relatively high luminance level, in an
attempt to minimize the contribution of the rods to the color matches. Some
computational techniques were also utilized to eliminate any remaining rod
contribution [93].
Around the same time, Speranskaya measured the color-matching functions
for 27 observers, also using a 10◦ viewing field [1074]. This data was measured
at a lower luminance level, and so the contributions of the rods were thought to
be higher. The CIE computationally removed the rod contribution and combined
these two data sets. A transformation to imaginary primaries, very similar to that
performed to generate the 1931 standard observer, followed. What resulted were
the 1964 CIE supplementary standard observer [186, 1262].
Like the 1931 standard observer, the 1964 standard observer is also commonly
referred to as the 10◦ -standard observer. The color-matching functions for the
10◦ observer are expressed as: x̄10 (λ ), ȳ10 (λ ), and z̄10 (λ ). These functions are
shown in Figure 7.9. For comparisons to the 1931 standard observer, both sets of
color-matching functions are shown in Figure 7.10.
The CIE recommends the use of the 1964 color-matching functions for all
color stimuli that are larger than 4◦ of visual angle. It is important to emphasize,
as shown in Figure 7.10, that the ȳ10 (λ ) color-matching function does not equal
the photometric observer (ȳ (λ ) or V (λ )), and that calculations using the 1964
standard observer do not directly translate into luminance measurements.

i i

i i
 i i

i i

378 7. Colorimetry

2.5

sensitivity
x10(λ)
y10 (λ)
2.0
z10(λ)

1.5

1.0

0.5

0.0
400 450 500 550 600 650 700
wavelength (nm)

Figure 7.9. The CIE 1964 x̄10 (λ ), ȳ10 (λ ),and z̄10 (λ ) color-matching functions.

2.5
sensitivity

x (λ)
y (λ)
z (λ)
2.0
x10(λ)
y10 (λ)
z10(λ)
1.5

1.0

0.5

0.0
400 450 500 550 600 650 700
wavelength (nm)

Figure 7.10. The CIE 1931 x̄ (λ ), ȳ (λ ), and z̄ (λ ) and the CIE 1964 x̄10 (λ ), ȳ10 (λ ), and
z̄10 (λ ) color-matching functions.

7.5 Calculating Tristimulus Values

and Chromaticities
We now have two sets of color-matching functions available with which we can
calculate color matches. The generic form of these calculations is:

i i

i i
 i i

i i

7.5. Calculating Tristimulus Values and Chromaticities 379


X= x̄ (λ ) · Φ (λ ) d λ , (7.6a)
λ

Y= ȳ (λ ) · Φ (λ ) d λ , (7.6b)
λ

Z= z̄ (λ ) · Φ (λ ) d λ . (7.6c)
λ

From these calculations, we can measure the “color” of any given stimulus,
Φ (λ ), and determine if two stimuli match. By definition, two stimuli will match
if they have identical tristimulus values. This match is considered to be valid for
the average population. It is also important to stress that this calculation of a
tristimulus value does not give us any insight into what the color actually looks
like, nor does it guarantee a match if any aspect of the viewing conditions change:

XYZ1 = XYZ2 . (7.7)

In practice, we often want to make calculations for real objects, rather than
generic radiant spectral power distributions. Typically, these objects are reflect-
ing objects. In this case, the color stimulus, Φ (λ ), is calculated by multiplying
the spectral reflectance or reflectance factor of an object with the spectral power
distribution of the light source. In most cases, we use relative reflectance factors
that have been normalized between 0–1 by dividing by the reflectance of a perfect
reflecting diffuser. Likewise, as will be discussed in Section 9.1, we typically use
the normalized spectral power distribution of the light sources, which is defined
to be either 1.0 or 100 at 560 nm. The normalized reflectance power and spectral
power distribution of the light source are then multiplied by the chosen standard
observers and integrated to calculate tristimulus values. This relationship is shown
in Figure 7.11.
The CIE color-matching functions are defined from 360 to 800 nm, in 1 nm
intervals. In practice, light-source and reflectance data are not available in this
range, or in that fine a wavelength interval. Typically, tristimulus values are cal-
culated using a more limited range of data; often from 380–720 nm or 400–700
nm at 5 or 10 nm increments. We also typically measure what are known as rela-
tive tristimulus values, which are normalized to a Y of 100.0 for the light source
itself. The relative discrete tristimulus-value calculation for reflecting objects, as
seen in Figure 7.11 is shown in (7.8). Similar calculations can also be performed
for transmitting colors.

i i

i i
 i i

i i

380 7. Colorimetry

Relative power
1.2

1.0
S(λ)
0.8

0.6
Multiply
Reflectance factor

0.6

R(λ)
0.4

0.2

0.0
Multiply
2.5
Tristimulus values

2.0 x(λ)
X = ∫ S(λ) R(λ) x(λ)
y(λ)
λ
1.5 z(λ)
Y = ∫ S(λ) R(λ) y(λ)
1.0 λ

Z = ∫ S(λ) R(λ) z(λ)

0.5 λ

0.0
400 450 500 550 600 650 700
wavelength λ (nm)
Integrate

Figure 7.11. Tristimulus values X, Y , and Z for real objects are calculated by multiplying
the light source, S (λ ), with the reflectance factor, R (λ ), and with the color-matching
functions, x̄ (λ ), ȳ (λ ), and z̄ (λ ) and then integrating.

These calculations are

X = k ∑ x̄ (λ ) S (λ ) R (λ ) Δλ , (7.8a)
λ
Y = k ∑ ȳ (λ ) S (λ ) R (λ ) Δλ , (7.8b)
λ

i i

i i
 i i

i i

7.5. Calculating Tristimulus Values and Chromaticities 381

Z = k ∑ z̄ (λ ) S (λ ) R (λ ) Δλ (7.8c)
λ
100
k= (7.8d)
∑ ȳ (λ ) S (λ ) Δλ .
λ

In these equations, we see that the normalization constant k is designed to

assure that calculations for a perfect reflecting diffuser will always result in a lu-
minance factor of Y = 100. The luminance factor can be defined as the ratio of
the luminance of an object to that of a perfect reflecting diffuser, when viewed
under the same geometry and lighting conditions. When we calculate relative
tristimulus values, as we have above, the CIE Y tristimulus value always corre-
sponds to the luminance factor. In (7.8), Δλ is the sampled wavelength interval,
typically 5, 10, or 20 nm. In practice, when normalizing the tristimulus values,
the sampling interval gets removed from the calculations due to its presence in
the k calculation.
Suppose we do not want to measure reflecting objects, but rather actual light
sources or self illuminating colors (such as a CRT display) that have physically
meaningful units. To do this, we must revisit the photometric luminous efficiency
function, V (λ ). Recall that a lumen is a unit of luminous flux defined to be 1/683
of a watt at 555 nm, the peak of V (λ ). Thus, there are 683 lm/W at 555 nm, and
there is a relationship between lumens and the weights of the V (λ ) photometric
observer.
When measuring self-illuminating materials, we typically measure either the
spectral radiance (Wm−2 sr−1 ) or irradiance (Wm−2 ). Using (7.9), we can calcu-
late either luminance units (cd m−2 ) or illuminance units (lux, or lm m−2 )1 . In
these equations, we will use Φ (λ ) to be either radiance or irradiance measure-
ments. It is important, for these calculations, to include the wavelength-sampling
interval, Δλ , as there is no normalization constant k:
X = 683 ∑ x̄ (λ ) Φ (λ ) Δλ , (7.9a)
λ
Y = 683 ∑ ȳ (λ ) Φ (λ ) Δλ , (7.9b)
λ
Z = 683 ∑ z̄ (λ ) Φ (λ ) Δλ . (7.9c)
λ

7.5.1 Chromaticity Coordinates

The CIE standard observers allow us to calculate consistent color measurements
and also determine when two colors match each other. Although these coordinates
1 See appendix D for an overview of SI units as well as derived units.

i i

i i
 i i

i i

382 7. Colorimetry

Figure 7.12. The visible spectrum displayed in CIE XYZ tristimulus space, seen from
different vantage points. This can also be considered a three-dimensional plot of the 2◦
color-matching functions. It’s projection onto a 2D plane produces the more familiar chro-
maticity diagrams.

do not represent color appearances, we can use them to represent a stimulus’ lo-
cation in a three-dimensional color space. Each of the axes in this space represent
the imaginary X, Y, and Z primaries, and a color stimulus’ location in the space
is its integrated tristimulus values. We can plot the location of the visible spec-
trum to get an idea of the behavior and shape of the CIE XYZ space. Figure 7.12
shows the spectral colors, which are also the color-matching functions of the 1931
standard observer, plotted in XYZ tristimulus space.
Although we cannot ascertain the appearance of any color based upon its tris-
timulus values, we show the spectral lines in Figure 7.12 in color. While not en-
tirely appropriate and done mostly for illustrative purposes, the monochromatic
colors of the spectrum do not drastically change as a result of viewing conditions
and chromatic adaptation. In essence, we are assuming that the colors drawn are
those of the individual wavelengths when viewed in isolation, otherwise known
as unrelated colors.
It is difficult to visualize the shape of the spectral colors in the three-dimen-
sional space. It is also fairly difficult to determine where any given color lies in
the space. Often we are interested in knowing and specifying the general region
of space a color occupies, but in an easy to understand manner. We know, by
definition, that for the 1931 standard observer, the Y tristimulus value is a measure
of luminance, which is directly related to our perception of lightness.
The other two tristimulus values do not have as easily interpretable meanings.
By performing a projection from three to two dimensions, we can generate a space
that approximates chromatic information: information that is independent of lu-
minance. This two-dimensional projection is called a chromaticity diagram and is
obtained by performing a type of perspective transformation that normalizes the
tristimulus values and removes luminance information.

i i

i i
 i i

i i

7.5. Calculating Tristimulus Values and Chromaticities 383

We can think of chromaticity coordinates as ratios of tristimulus values that

do not contain any magnitude information. This normalization is shown in (7.10).
Chromaticity coordinates are generally expressed using lowercase letters, while
we use uppercase letters for tristimulus values.

X
x= , (7.10a)
X +Y + Z
Y
y= , (7.10b)
X +Y + Z
Z
z= , (7.10c)
X +Y + Z
x + y + z = 1. (7.10d)

It should be obvious that the projection of three-dimensions of information

down to two always results in the loss of information. For chromaticity coor-
dinates, the ratio normalization essentially throws away the luminance informa-
tion. It should also be obvious that because the chromaticity coordinates always
sum to 1, if we know two of them we can easily determine the third. On their
own, chromaticity coordinates do not possess enough information to adequately
describe a color measurement or a color match. Traditionally, we can use two
coordinates (usually x and y) along with one of the tristimulus values to fully de-
scribe a color. Since the Y tristimulus value describes luminance, you will often
see colors described with their Y xy values. We can calculate tristimulus values
back from chromaticities using (7.11).
x
X = Y, (7.11a)
y
z
Z = Y. (7.11b)
y

Plotting the spectral colors in the chromaticity diagram results in the familiar
horseshoe-shaped curve. These spectral locations are referred to as the spectrum
locus, and they represent the boundary of physically realizable colors. The spec-
trum locus for the 1931 standard observer is shown in Figure 7.13. Once again,
in this figure we plot the spectral wavelengths in color. While this is somewhat
appropriate for the monochromatic lights (again, when viewed in isolation) it is
not appropriate to draw colors inside the spectral locus.
As we have emphasized in this chapter, colorimetry is not designed to predict
how colors appear, but rather just provide a means for measuring and describ-
ing color matches. Chromaticity diagrams have an even more limited description

i i

i i
 i i

i i

384 7. Colorimetry

Figure 7.13. The visible spectrum displayed in the CIE xy chromaticity diagram.

of color, as they lack any luminance information. Depending on viewing condi-

tions, luminance levels, and state of adaptation, it is possible to make almost any
location in the chromaticity diagram appear either black or white (or any other
color!). This is evidenced in Figure 7.14, where the chromaticities of several CIE
standard illuminants are plotted in a chromaticity diagram. Despite the variety of
chromaticity locations, real light sources at these same coordinates would gener-
ally appear white.
The “purple” line plotted at the bottom of the horseshoe-shaped spectrum lo-
cus in Figure 7.13 does not exist as a pure spectral color, but rather as a combina-
tion between the short and long wavelength stimuli. Appropriately enough, this
is referred to as the purple line or purple boundary.
As chromaticities are simply normalized tristimulus values, it is possible to
specify the range of colors that can be generated by two or more primaries, when
these primaries are represented in a chromaticity diagram. For instance, for a two-
primary system, we draw a straight line in the chromaticity diagram connecting
the two primaries together. The system can then reproduce any of the chromatic-
ities on that line, by changing the relative luminances of the two primaries.
For a three-primary system, we draw lines in the chromaticity diagram that
connect the three primaries together, with the range of reproducible chromaticity
coordinates lying inside the triangle bounded by those lines. This is illustrated in

i i

i i
 i i

i i

7.5. Calculating Tristimulus Values and Chromaticities 385

Figure 7.14. The location of various CIE standard illuminants in the 1931 xy chromaticity
diagram.

Figure 7.15. The chromaticity boundaries of the CIE RGB primaries at 435.8, 546.1, and
700 nm (solid) and a typical HDTV(dashed).

i i

i i
 i i

i i

386 7. Colorimetry

Figure 7.15 for the monochromatic primaries used to specify the CIE RGB color-
matching functions, as well as the primaries specified for a typical high-definition
television [536]. It is tempting, and erroneous, to refer to these triangles as the
gamut of the display system. This is very misleading, as the gamut of reproducible
colors must at the very least include the third dimension of luminance to have
any meaning. We can say that these triangles represent the range of all possible
chromaticities that we can create through additive mixing of the three primaries,
following Grassmann’s laws.
Another word of warning: Chromaticity diagrams, while sometimes useful,
should be used with diligence and care. They do not represent meaningful colors
and should not be used to specify what a color looks like. They do provide insight
into additive color mixing, which is why they are often still used in the display
industry. There is no reason to believe that equal distances, or areas for that matter,
should represent equal perceptual distances. In fact, we should be surprised to find
out if they did. The CIE color-matching functions that they are based upon are
not meant to be human cone sensitivities (though we can consider them to be
approximate linear transformations from cones).
Nevertheless, there has been considerable research on creating chromaticity
diagrams that are more nearly uniformly spaced. In 1976, the CIE recommended
using the u v Uniform Chromaticity Scale (UCS) chromaticity diagram [186].
The general goal of this diagram was to be more uniformly spaced, such that
distances and areas held a closer relationship to actual perceptions. It should be
noted that just like the xy chromaticity diagram, the u v diagram does not contain
any luminance information. Thus, any distances that are meant to be perceptually
meaningful do not include the luminance relationships between the colors. This
may be acceptable when dealing with primaries or light sources that have similar
luminances, but it is not acceptable when examining those with widely disparate
luminances. The calculation of u v from CIE XYZ tristimulus values is given by

4X
u = , (7.12a)
X + 15Y + 3Z
9Y
v = . (7.12b)
X + 15Y + 3Z

The u v -coordinates can also be computed from xy-chromaticity coordinates:

4x
u = , (7.13a)
−2x + 12y + 3
9y
v = , (7.13b)
−2x + 12y + 3

i i

i i
 i i

i i

7.6. Practical Applications of Colorimetry 387

Figure 7.16. The 1976 u v Uniform Chromaticity Scale Diagram.

with the inverse transform given by

9u
x= , (7.14a)
6u − 16v + 12
4v
y=  . (7.14b)
6u − 16v + 12
The corresponding u v chromaticity diagram is shown in Figure 7.16.

7.6 Practical Applications of Colorimetry

Colorimetry, as we have learned throughout this chapter, provides a means to
accurately specify color measurements and determine if two colors match. Al-
though not meant to describe how colors appear (for that we need a full color
appearance model) colorimetry does provide a framework from which to build
an accurate color-reproduction system. Indeed, colorimetry forms the basis of all
modern color-management systems and now exists in some form or another in all
desktop operating systems.
What colorimetry gives us, in essence, is a method for describing color as
device-independent. That is to say, we can measure and specify colors in a way

i i

i i
 i i

i i

388 7. Colorimetry

that is not tied to any specific display device or specific set of primaries. Most
readers have probably encountered the situation where they create or examine an
image on one computer display and then look at the image on a second computer
display, only to find out that the image looks markedly different. Despite the
fact that the digital file contains the exact same red, green, and blue values, each
computer interprets these data differently. The images are assumed to be device-
dependent, essentially measurements tied to the specific device on which they
were created. Imagine if other measurements in the world were based in such a
manner, for instance a unit of length based upon each individual foot size!
By removing the display device from the picture, we can specify color in a
meaningful way, based upon the imaginary but well-defined XYZ primaries. This
is similar to basing the measurement of length on an imaginary, but well-defined,
golden foot. We can then transform our colorimetric units into any other unit, or
any other display device, much like we can convert the foot to the meter, fathom,
or furlong. This is the basis of modern color management.
So how can one take advantage of the CIE system of colorimetry? When creat-
ing digital images, we can create them directly in tristimulus space, though due to
the imaginary nature of the space this may prove difficult. For computer-graphics
renderings that rely on physics, there are several ways to directly take advantage
of colorimetry. If, in the rendering process, all calculations are performed in the
spectral domain, including light source propagation as well as material and light
source interactions, then it is easy to just integrate the resulting spectral radiance
values of the scene with the CIE x̄ (λ ), ȳ (λ ), and z̄ (λ ) color-matching functions,
and integrate as shown in (7.6). We will then have perceptually accurate measure-
ments of the rendered image.
But how do we display colorimetric data? For that, it is necessary to apply
another transformation of primaries.

7.6.1 Transforming Primaries

Recall from Section 7.3, that the original color-matching data from Wright and
Guild were transformed to a common set of primaries before being averaged to-
gether. This transformation utilized Grassmann’s laws to express one set of spec-
tral tristimulus values in terms of another. This was accomplished using a linear
transform, which for a three-primary display is represented by a 3 × 3 matrix
transform. For additive color systems that obey Grassmann’s laws, the calcula-
tion of this transform is very straightforward.
Desktop display systems, such as CRTs, behave as additive systems and, as
such, calculating the transformation from XYZ tristimulus values into the native

i i

i i
 i i

i i

7.6. Practical Applications of Colorimetry 389

1.0

Relative Power
0.9 Red
Green
0.8 Blue
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
400 450 500 550 600 650 700
wavelength (nm)

Figure 7.17. The relative spectral power distribution of a representative CRT display.

primaries of the display is relatively easy. This is known as display character-

ization; more details can be found in Section 14.18, as well as in the works by
Berns [93] and Day [237]. LCDs are more complicated, as they have a very bright
black level, and also have different spectral characteristics dependent on intensity.
The following discussion is therefore only approximate for LCD displays.
It is important to stress that Grassmann’s laws assume a completely linear
system, which most displays are inherently not. When characterizing a display,
it is very important to first remove the nonlinear components, often referred to
as gamma, before applying the colorimetric transformation. We will ignore the
nonlinear component in this section, though further details are presented in Chap-
ter 14.
A typical CRT display has three additive primaries; red, green, and blue. An
example of such a display is shown in Figure 7.17. Due to the phosphors of a
typical CRT display, there are relatively broadband blue and green primaries and
a relatively sharp red primary. If we would like to figure out the transformation
from CIE XYZ tristimulus values into the units of the display device, we need
to first measure the XYZ values of the three primaries of the CRT. We can then
generate a transformation that will take us from display RGB values into XYZ
tristimulus values. This is accomplished using (7.15):
⎡ ⎤ ⎡ ⎤⎡ ⎤
X Xred Xgreen Xblue R
⎣Y ⎦ = ⎣Yred Ygreen Yblue ⎦ ⎣G⎦ . (7.15)
Z Zred Zgreen Zblue B

i i

i i
 i i

i i

390 7. Colorimetry

In (7.15), the Xred , Yred , and Zred terms represent the colorimetric measurements
of the CRT red primary, while the same holds for green and blue. These are
measured by driving the display with the maximum RGB digital counts, typically
(255, 0, 0) for the red primary on a 8-bit system.
Examining the form of the matrix in (7.15), we can immediately see that the
tristimulus calculation of any color on the display is literally just an additive mix-
ture of the three primaries, e.g., Xcolor = Xred + Xgreen + Xblue . By inverting the
3 × 3 matrix, it is easy to calculate the RGB values necessary to display a given
tristimulus value:
⎡ ⎤ ⎡ ⎤−1 ⎡ ⎤
R Xred Xgreen Xblue X
⎣G⎦ = ⎣Yred Ygreen Yblue ⎦ ⎣Y ⎦ . (7.16)
B Zred Zgreen Zblue Z
So in our computer rendering example, if we have rendered the spectral ra-
diance values, and then calculated the tristimulus values, we can apply (7.16) to
determine the RGB values to display on our CRT (again, temporarily ignoring the
nonlinear behavior). To display the same colors on a different device, it is not nec-
essary to re-render the scene, but rather just calculate a different transformation
matrix.
If it is not possible to directly measure the output display device, there are
assumptions that can be made. For instance, the primary set for HDTV (ITU-
R BT.709/2) is well defined [536]. Not coincidentally, this same primary set,
dubbed sRGB, was chosen to be the “standard” RGB for color management as
well as for the Internet [524]. If the primaries of the output display are unknown,
it is possible to assume that they are consistent with the sRGB primaries. The
transformation to sRGB from CIE XYZ tristimulus values is given by
⎡ ⎤ ⎡ ⎤⎡ ⎤
sR 3.2410 −1.5374 −0.4986 X
⎣sG⎦ = ⎣−0.9692 1.8760 0.0416⎦ ⎣Y ⎦ . (7.17)
sB 0.0556 −0.2040 1.0570 Z

When applying any transformation of primaries, it is important to watch the scale

of the input and output equations. For instance, the matrix shown in (7.17) expects
the input XYZ tristimulus values to be in the range of 0–1, rather than the more
common range of 0–100 for relative colorimetry. The sRGB/ITU-R BT.709/2
primary set has a more limited gamut than the imaginary primaries of XYZ, as
suggested by the dashed triangle in Figure 7.15. It is quite possible that transfor-
mations from CIE XYZ will result in sRGB values that are negative, or above 1.0.
In this case, these colors represent objects that are outside the sRGB gamut and
need to be mapped back into the 0–1 range. The most common form of gamut

i i

i i
 i i

i i

7.6. Practical Applications of Colorimetry 391

mapping for this case is simple clipping, though more complicated algorithms are
in wide use (see Section 15.4).

7.6.2 Transformation of Color-Matching Functions

In the previous sections, we have described a practical application of using CIE
colorimetry to measure and display accurate spectral computer graphics render-
ings. While ideal, it is not always feasible to use full spectral calculations for
rendering purposes. Suppose, rather, we wanted to use a traditional three-channel
rendering package, but still allow for accurate specifications of color. In this case,
it would be possible to specify all light sources and material properties in the scene
in terms of CIE XYZ tristimulus values. These can be calculated by measuring
the spectral reflectance and transmittance properties of the materials and then in-
tegrating into CIE XYZ using (7.8). The radiant spectral power distribution of the
lights can also be calculated, using (7.9), and the rendering package can take care
of the interactions between materials and sources. More details on sampling the
spectral curves for computer rendering purposes can be found in Hall [415].
These techniques still require a final transformation in RGB value before
the images can be displayed, through the transformation of primaries discussed
above. One practical alternative is to generate a new set of color-matching func-
tions, based upon the primaries of the desired output display. If the desired output,
for instance, is a CRT then it would be possible to calculate an appropriate set of
color-matching functions that would then correctly generate the appropriate inte-
grated tristimulus values that could be directly displayed.
An example of the color-matching functions generated for the sRGB and
ITU-R BT.709/2 primaries are shown in Figure 7.18. Notice how there are con-
siderable negative lobes in these color-matching functions, suggesting that there
are a significant number of spectral colors that can not be correctly generated if
our rendering and storage system does not allow for both positive and negative
numbers. If integrating a full spectrum is not possible, one possible way to sam-
ple the wavelength spectrum is to look where the sRGB color-matching functions
peak and sample the color spectrum at those three locations: 445 nm, 540 nm, and
605 nm. While not ideal, sampling at those wavelengths should produce colors
reasonably close to those desired.
Using these techniques, it is also possible to generate a set of color-matching
functions, or spectral sensitivities such as those in Figure 7.18 for a digital capture
device such as a camera that are a direct representative of the output primaries.
Often it is not possible to generate sensitivities that have both positive and neg-
ative values, and so often we generate all positive sensitivities that are a linear

i i

i i
 i i

i i

392 7. Colorimetry

2.0

Sensitivity
r sRGB
1.5
gsRGB
bsRGB
1.0

0.5

0.0

−0.5

−1.0
400 450 500 550 600 650 700
Wavelength (nm)

Figure 7.18. The color-matching functions corresponding to the sRGB primaries.

transform from the desired primaries. The special case, where the spectral sen-
sitivities of the capture device are a direct linear transform from CIE XYZ tris-
timulus values, means that the device satisfies the Luther-Ives condition and is
colorimetric [538, 715].

7.6.3 XYZ to Spectrum Conversion

For some applications, including rendering if a mixture of spectral and tristimu-

lus data is to be used within the same application [416], it may be necessary to
convert a tristimulus value XYZ to a plausible spectral representation Φ (λ ). This
amounts to inverting (7.6), yielding infinitely many solutions, which are each
metamers of each other.
To solve this problem, additional constraints must be imposed. The simplest
approach is to introduce three linearly independent basis functions f1 (λ ), f2 (λ ),
and f3 (λ ). The derived spectrum is then a linear combination of these basis
functions:
3
Φ (λ ) = ∑ x j f j (λ ) , (7.18)
j=1

where x j are three coefficients. The values for these coefficients can be found by
substituting this equation into (7.6), giving

i i

i i
 i i

i i

7.6. Practical Applications of Colorimetry 393

 3
X= x̄ (λ ) ∑ x j f j (λ ) d λ , (7.19a)
j=1
λ
 3
Y= ȳ (λ ) ∑ x j f j (λ ) d λ , (7.19b)
j=1
λ
 3
Z= z̄ (λ ) ∑ x j f j (λ ) d λ . (7.19c)
j=1
λ

Making the following substitutions:

c0 = X, (7.20a)
c1 = Y, (7.20b)
c2 = Z, (7.20c)
c̄0 = x̄, (7.20d)
c̄1 = ȳ, (7.20e)
c̄2 = z̄, (7.20f)

and rearranging terms in (7.19), we have

3 
ci = ∑ c̄i (λ ) f j (λ ) d λ x j . (7.21)
j=1
λ

With the color-matching functions c̄i (λ ), the basis functions f j (λ ), and the XYZ
tristimulus value ci known, the coefficients x j can be computed. Plugging these
coefficients into (7.18) yields the desired spectral representation of the input XYZ
tristimulus value ci .
However, the choice of basis functions f j (λ ) is important, as it will have
a profound impact on the shape of the resulting spectrum. For computational
simplicity, a set of delta functions could be chosen [369]:

1 for λ = λi ,
f j (λ ) = δ (λ − λ j ) = (7.22)
0 for λ = λi .

The coefficients x j can then be derived from (7.21), which simplifies to

3
ci = ∑ c̄i (λ j ) x j . (7.23)
j=1

To achieve a numerically stable solution, the values for λ j can be chosen to co-
incide with the peaks in the color-matching functions c̄i , which are λ1 = 590 nm,

i i

i i
 i i

i i

394 7. Colorimetry

λ2 = 560 nm, and λ3 = 440 nm. Of course, with this approach the resulting
spectrum will only have three non-zero values. This is far removed from typical
spectra, which tend to be much smoother.
Alternatively, the following box functions could be used as basis functions
[414, 415]:

1 for λi < λ < λi+1 ,
f j (λ ) = (7.24)
0 otherwise.

The boundaries of the intervals are given by λ1 = 400 nm, λ2 = 500 nm, λ3 = 600
nm, and λ4 = 700 nm. This approach yields spectra which vary abruptly at the
interval boundaries.
For use in rendering, it is desirable to create spectra which are relatively
smooth, especially if they represent reflectance functions of materials [1063]. To
enforce a smoothness constraint, the first three Fourier functions can be chosen as
basis functions [265, 1204]:

f1 (λ ) = 1, (7.25a)
 
λ − λmin
f2 (λ ) = cos 2π , (7.25b)
λmax − λmin
 
λ − λmin
f3 (λ ) = sin 2π . (7.25c)
λmax − λmin

This approach leads to smooth spectra, but there is a significant probability that
parts of the spectra are negative, and they can therefore not be used. This tends to
happen most for highly saturated input colors. Although the negative parts of any
spectrum could be reset to zero, this would introduce undesirable errors.
An arguably better approach would be to parameterize the basis functions to
make them adaptable to some characteristic of the tristimulus value that is to be
converted to a spectrum. In particular, it would be desirable to create smooth
spectra for desaturated colors, whereas saturated colors would yield more spiked
spectra. As an example, Gaussian basis functions may be parameterized as fol-
lows [1106]:
  
2 (λ − λ j ) 2
f j (λ ) = exp − ln 2 . (7.26)
wj

Here, λ j indicates the center position of the jth Gaussian, and w j is its width. The
values λ j are λ1 = 680 nm, λ2 = 550 nm, and λ3 = 420 nm.

i i

i i
 i i

i i

7.6. Practical Applications of Colorimetry 395

For a tristimulus value given in, for instance, the sRGB color space, a rough
indication of the degree of saturation is obtained by the following pair of values:
|sR − sG|
s1 = , (7.27a)
sR + sG
|sB − sG|
s2 = . (7.27b)
sB + sG
We can now compute the widths of the Gaussian basis functions by linearly in-
terpolating between the user-defined minimum and maximum Gaussian widths
wmin = 10 nm, and wmax = 150 nm:

w1 = s1 wmin + (1 − s1 ) wmax , (7.28a)

w2 = s2 wmin + (1 − s2 ) wmax , (7.28b)
w3 = min(w1 , w2 ). (7.28c)

This approach will produce smooth spectra for desaturated colors and progres-
sively more peaked spectra for saturated colors. As there are infinitely many
solutions to the conversion between XYZ tristimulus values and spectral repre-
sentations, this appears to be a reasonable solution, in keeping with many spectra
found in nature.

7.6.4 Beginnings of Color Management

We have now seen how it is possible to measure and specify color in a device-
independent manner. That is to say, we can describe color in a way that is not tied
to a specific input or output display device. This specification lies at the heart of
all modern color-management systems. The goal of a color-management system
is to have a color, or image, specified in such a way that we can reproduce it
with accurate results on any other device. This allows us to see an image on our
computer monitor, print the image, or send the image to our friends across the
country, and still have all the images match in appearance.
We can do this by calculating the transformation of primaries from one de-
vice to all the others, as described earlier in this section. This, of course, assumes
all the devices are linear and additive, which is most definitely not the case for
printers. The other technique is to calculate the colorimetric characterization for
each device, both on the input side and the output side, from their own respective
primaries into CIE XYZ tristimulus space. From this space, which is no longer
tied to any particular device, we can then transform back out into any other de-
vice. The general flowchart for such a color management system is shown in
Figure 7.19.

i i

i i
 i i

i i

396 7. Colorimetry

Input Device
Characterization

Device
Independent CIE
XYZ

Output Device
Characterization

Figure 7.19. Using device-independent colorimetry as a basic color management process.

While such techniques do lie at the heart of most modern color-management

systems, Figure 7.19 represents an over-simplification. Recall that colorimetry
is only designed to measure colors and determine if two colors match, but not
specify what two colors look like. This color match may break down when certain
aspects of the viewing conditions change, such as overall amounts or colors of
the lighting. In these situations, a colorimetric match between input and output
devices does not guarantee an appearance match. In fact, more often than not,
an appearance mis-match is to be expected. As an example, displaying the same
image on two different displays will yield a different percept even if the primaries
are identical, but only the white points do not match.
In order to obtain more accurate color reproduction across a wide variety of
media, it is necessary to employ chromatic-adaptation models or a full color ap-
pearance model, as discussed in Chapter 11.

i i

i i
 i i

i i

7.7. Application: Iso-Luminant Color Maps 397

7.7 Application: Iso-Luminant Color Maps

The field of visualization is concerned with the mapping of data to pixel values
in such a way that meaningful conclusions about the data can be drawn. There
are several forms of visualization, including medical visualization, scientific vi-
sualization, and information visualization. The origins of the data are therefore
varied. However, in general, the data have many different attributes that need to
be mapped to pixel values. This means that high-dimensional attribute spaces are
mapped to three-dimensional color spaces.
It is frequently desirable to visualize several attributes of the data simultane-
ously. There are many different ways to accomplish this, but we focus on the
sub-problem of mapping data to color values. As luminance plays an extremely
important role in human vision, and in particular in the perception of image struc-
ture and surface shape [868], it requires careful consideration when designing
visualization algorithms.
For instance, it may be possible to map one attribute of the data to a range
of luminance values, and a second attribute of the data to chromaticity values,
leading to a bivariate color map. In such an approach it would be crucial to design
a color map in which luminance variations are absent, i.e., an iso-luminant color
map. In the case of univariate color maps where only one attribute is mapped to a
color scale, iso-luminant color maps are also desirable, as they help avoid errors in
interpretation as a result of perceptual effects such as simultaneous contrast [1215]
(see Section 5.7). Similar advantages apply to the use of univariate color maps in
combination with surface shading [326].
The design of an iso-luminant color map requires at least three problems to be
solved. First, the display device used for the visualization is often uncalibrated,
with unknown primaries, white point, and gamma function. Variations between
display devices can be considerable [370]. Second, the lighting conditions in the
viewing environment are typically unknown and uncontrolled [251,1216], leading
to reflections off the display and an unknown state of adaptation of the viewer.
Third, the spectral sensitivity in humans may vary between individuals, and may
even change with age, for instance as the result of the build-up of yellow pigments
in the ocular media [1262].
To overcome these problems, it is possible to design an iso-luminant color
map specifically for a particular combination of monitor, display environment,
and observer. This can be achieved by modifying an existing color-matching
experiment, such that it can be conveniently and quickly executed by any ob-
server [591]. Unlike most research on the design and application of color maps in
visualization [86, 958, 959, 966, 1215], this technique does not assume calibrated
displays and/or viewing environments.

i i

i i
 i i

i i

398 7. Colorimetry

7.7.1 Color Matching

The construction of an iso-luminant color map requires matching a set of colors to

a single achromatic luminance value. This is similar to the problem of measuring
the luminous efficiency function V (λ ), which is normally achieved by means of a
color-matching experiment. However, in our case, the color-matching experiment
takes place in the device-dependent color space of the monitor that is to be used
for visualization.
Color-matching experiments either present stimuli in spatial juxtaposition, or
exploit differences in temporal resolution between chromatic and achromatic per-
ception [118, 1143, 1262]. Temporal approaches such as flicker photometry show
stimuli that rapidly alternate between achromatic and chromatic values, whereby
the observer can adjust the achromatic signal until the image appears stable. Al-
though accurate results can be achieved, the flicker frequency should be around
15 Hz to 20 Hz, which is in a range that is difficult to reproduce accurately on typ-
ical display devices. In addition, this frequency may induce epileptic seizures in
some individuals. This makes flicker photometry unattractive for use as a general
technique to be applied outside carefully controlled laboratory conditions.
An alternative is to display two patches side by side in a bipartite field, and
the observer can adjust the intensity of one of the patches to match the apparent
brightness of the other patch. If both patches are chromatic, but of different wave-
lengths, the resulting technique is called heterochromatic brightness matching.
This task is rather difficult, making it problematic to obtain accurate results.
In particular, this method suffers from a phenomenon called additivity failure,
described as follows [118]. First, two chromatic patches are individually matched
to an identical white field by adjusting the intensity of the chromatic patches. If
these two chromatic patches are then super-posed, and the white field is doubled
in intensity, then the combined colors should match the new white field.
However, in practice, the combined field will appear dimmer than the doubled
white field. In addition, saturated colors will appear to glow with a brightness
that is not in proportion to their actual luminance. The latter is known as the
Helmholtz-Kohlrausch effect [302, 1262] (see also Section 11.2.2).
Related to direct heterochromatic brightness matching is the minimally dis-
tinct border technique. Here, the two fields are precisely juxtaposed, i.e., the
fields are not separated in space. One of the fields can now be adjusted until the
border between them is minimally distinct to the observer. The results of this
method tend to be similar to heterochromatic brightness matching, although it
solves the problem of additivity failure, and therefore does not suffer from the
Helmholtz-Kohlrausch effect [118, 1262]. It is therefore possible to regard the

i i

i i
 i i

i i

7.7. Application: Iso-Luminant Color Maps 399

minimally distinct border technique as a luminance matching task, making it an

interesting technique for the construction of iso-luminant color maps.
However, finding the minimally distinct border between two patches is still
not a simple task. Especially if the patches are very different in spectral com-
position, the border between them will remain visible. As a result, the matches
obtained with this method may show relatively large variance.

7.7.2 Color Matching Using Face Stimuli

The lack of accuracy of the minimally distinct border technique is a problem if
it is to be used in visualization tasks. A luminance-matching task with higher
accuracy may be constructed by considering human faces [591].
Humans are generally very good at distinguishing and recognizing faces, pos-
sibly because the brain contains dedicated circuitry for face recognition [94, 281,
350]. However, at the same time, face recognition is much more difficult if the
main lighting is not coming from above [934], which was confirmed in an fMRI
study [356]. In addition, incorrect luminance levels are more easily detected in
faces than in other objects [1101]. Given that luminance mediates face percep-
tion in thresholded face images [160], this can be exploited by constructing a test
pattern as shown in Figure 7.20.
This pattern was constructed by thresholding a photograph of a human face
which was lit by a strong directional light source. The pattern on the left in this

Figure 7.20. The two-tone pattern on the left is easily recognized as a human face, whereas
the image on the right is not. (Gordon Kindlmann, Erik Reinhard, and Sarah Creem, “Face-
Based Luminance Matching for Perceptual Color Map Generation,” Proceedings of IEEE
Visualization, pp 309–406,  c 2002 IEEE.)

i i

i i
 i i

i i

400 7. Colorimetry

Figure 7.21. Example of varying the colors in face stimuli. Note the different cross-
over points for the three colors, where the perception of a face flips from the left image
to the right image. (Gordon Kindlmann, Erik Reinhard, and Sarah Creem, “Face-Based
Luminance Matching for Perceptual Color Map Generation,” Proceedings of IEEE Visual-
ization, pp. 309–406, c 2002 IEEE.)

i i

i i
 i i

i i

7.7. Application: Iso-Luminant Color Maps 401

Figure 7.22. The two-tone pattern used in the minimally distinct border experiment.
(Gordon Kindlmann, Erik Reinhard, and Sarah Creem, “Face-Based Luminance Matching
for Perceptual Color Map Generation,” Proceedings of IEEE Visualization, pp. 309–406,
c 2002 IEEE.)

figure is easily recognized as a human face, whereas inverting this pattern breaks
human face perception.
In this stimulus, it is possible to replace the black with a desired shade of
gray. The white is replaced with a color, and the observer is given control over
the intensity value of the color. By appropriately changing the intensity value of
the color, the stimulus that was initially seen as a face will disappear, and the
pattern on the other side will then be recognized as a face. The range of intensity

0.70 0.09
Mean lightness

Standard deviation

0.65 0.08 Minimally distinc border

Face-based matching
0.60 0.07
0.55 0.06
0.50 0.05
0.45 0.04
0.40 0.03
0.35 0.02
0.30 0.01
0.25 0.00
red yellow green cyan blue magenta red yellow green cyan blue magenta

Figure 7.23. Mean and standard deviation for both the minimally distinct border and
the face-based luminance matching techniques. (Gordon Kindlmann, Erik Reinhard, and
Sarah Creem, “Face-Based Luminance Matching for Perceptual Color Map Generation,”
Proceedings of IEEE Visualization, pp. 309–406, 
c 2002 IEEE.)

i i

i i
 i i

i i

402 7. Colorimetry

Figure 7.24. Color maps generated using face-based luminance matching. (Gordon Kindl-
mann, Erik Reinhard, and Sarah Creem, “Face-Based Luminance Matching for Perceptual
Color Map Generation,” Proceedings of IEEE Visualization, pp. 309–406, c 2002 IEEE.)

values where both the left and right halves of the stimulus are perceived to be
ambiguous tends to be small, suggesting a high level of accuracy. An example of
a set of stimuli is shown in Figure 7.21.
In a user study, the face-based luminance-matching technique was compared
against a conventional minimally distinct border technique, using a stimulus that
can not be recognized as a human face, but has otherwise the same boundary
length and a similar irregular shape (as shown in Figure 7.22). The results, shown
in Figure 7.23, reveal that the two techniques produce the same mean values, but
that the face-based technique has a significantly higher accuracy.
Color maps generated with this technique are shown in Figure 7.24. This
figure shows that different observers generate somewhat different iso-luminant
color maps. The variances between different observers have been measured using
face-based luminance matching. Results for seven participants, each for normal
color vision, are shown in Figure 7.25. Aside from showing differences between
observers, this figure also shows that the standard rainbow color map is far from
iso-luminant.
It is also shown in Figure 7.24 that monotonically increasing color maps can
be constructed with this technique. This is achieved by choosing increasing lu-
minance values from left to right along the color scale. Thus, different colors
are matched to increasing luminance values, yielding the monotonically increas-
ing luminance map shown in this figure. Note that face-based techniques have
also been developed to test if a given color map is monotonically increasing or
not [974].

i i

i i
 i i

i i

7.8. Further Reading 403

Mean Lightness 1.0

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0.0
red yellow green cyan blue magenta

Figure 7.25. The perception of lightness varies between observers, as shown here for
18 different colors and seven participants. (Gordon Kindlmann, Erik Reinhard, and Sarah
Creem, “Face-Based Luminance Matching for Perceptual Color Map Generation,” Pro-
ceedings of IEEE Visualization, pp. 309–406,  c 2002 IEEE.)

In summary, exerting luminance control over color maps used in visualization

tasks is desirable. Taking the display, the viewing environment, as well as differ-
ences between observers into account, a simple perceptual test can be constructed
to create iso-luminant color maps optimized for one specific viewing condition.
This will help produce good visualizations and minimize errors in interpretation.
The software for this technique is available online at http://www.cs.utah.edu/∼gk/
facelumin.html.

7.8 Further Reading

A recent book specifically on colorimetry is Colorimetry: Fundamentals and Ap-
plications by Ohta and Robertson [849]. More general texts are Hunt’s book, The
Reproduction of Colour [509] and, of course, Wyszecki and Stiles, Color Science:
Concepts and Methods, Quantitative Data and Formulae [1262].

i i

i i
 i i

i i

Chapter 8
Color Spaces

As shown in Chapter 7, due to the trichromatic nature of the human visual sys-
tem, color may be represented with a color vector consisting of three numbers,
or tristimulus values. The CIE XYZ color space is derived from spectral power
distributions by means of integration with three color-matching functions. These
color-matching functions can be considered approximate linear transformations
of the human cone sensitivities. For the XYZ color space, the Y channel was
derived to be identical with the photopic luminance response function.
Closely tied to all color-matching functions, and their corresponding tristim-
ulus values, are the primaries or light sources that can be used to re-derive the
functions in a color-matching experiment. One of the main goals in creating the
the CIE XYZ color space was to have tristimulus values that are all positive. In
order to do that, it was necessary to select an imaginary set of primaries. Essen-
tially, this means that it is not possible to construct a set of only three positive
light sources that, when combined, are able to reproduce all colors in the visible
spectrum.
In general practice, the XYZ color space may not be suitable for performing
image calculations or encoding data, as there are a large number of possible XYZ
values that do not correspond to any physical color. This can lead to inefficient
use of available storage bits and generally requires a higher processing bit-depth
to preserve visual integrity. It is not possible to create a color display device that
corresponds directly to CIE XYZ; it is possible to create a capture device that
has sensor responsivities that are close to the CIE XYZ color-matching functions,
though due to cost of both hardware and image processing, this is not very com-
mon. Therefore, since the corresponding primaries are imaginary, a transcoding
405

i i

i i
 i i

i i

406 8. Color Spaces

into a specific device space is necessary. Depending on the desired task, the use
of a different color space may be more appropriate. Some of the reasons for the
development of different color spaces are:

Physical realizability. Camera and display manufacturers build sensors and dis-
plays that have specific responsivities (cameras) and primaries (displays).
The output of any given camera is often implicitly encoded in a color space
that corresponds to the camera’s hardware, firmware, and software. For in-
stance, most digital sensors capture data linearly with respect to amounts
of light, but this data is gamma corrected by compressing the signal using
a power function less than 1.0. This gamma-corrected data may or may
not still correspond to the physical RGB sensors used to capture the image
by the camera. The signal that is sent to a display device such as a cath-
ode ray tube, which has its own device-dependent color space defined by
its phosphors, may not be in the same color space as the cathode ray tube;
it must first be transformed or color reproduction will be impaired. In a
closed system, it is possible to directly capture images using a camera that
is specifically tied to an output device, removing the need for transforming
the colors between capture and display.

Efficient encoding. Some color spaces were developed specifically to enable ef-
ficient encoding and transmission of the data. This encoding and trans-
mission may rely on certain behaviors of the human visual system, such
as our decrease in contrast sensitivity for chromatic signals compared to
achromatic signals. In particular, the color encodings used by the different
television signals (NTSC, PAL, SECAM) exploit this behavior of the visual
system in their underlying color models.

Perceptual uniformity. For many applications, we would like to measure the per-
ceived differences between pairs of colors. In most linear color spaces,
such as CIE XYZ, a simple Euclidean distance does not correspond to the
perceptual difference. However, color spaces can be designed to be percep-
tually uniform, such that the Euclidean distance in those color spaces is a
good measure of perceptual difference.

Intuitive color specification. Many color spaces are designed to be closely tied
to output display primaries, such as the red, green, and blue primaries of a
typical CRT display. However, the mapping from such values encoded in
these spaces to the appearance of a color can be very non-intuitive. Some
color spaces are designed to provide a more intuitive meaning to allow for
ease of specifying the desired color. These spaces may encode the data in

i i

i i
 i i

i i

407

ways that more closely match our perception of color appearance, using
terms such as hue, saturation, and lightness (HSL).
Color spaces are often characterized by their bounding volumes, or the range
of colors that they are capable of representing. This is often referred to as the
gamut of colors that a space can represent. This is only a valid assumption if
the values that color vectors can store are limited, for example all positive with
a range between 0 and 1. For instance, many (but not all) color spaces do not
allow negative values—typically because the encoded values correspond to the
light energy coming from a specific set of output primaries, and light cannot be
negative. We generally assume that only output devices have color gamuts, while
capture devices have a dynamic range or limitations in their spectral sensitivity.
This is often an area of great philosophical and semantic debate and discussion in
the color-imaging community.
The conversion of colors represented in one gamut to another, often smaller,
gamut is a frequently occurring problem in color reproduction. This field of study
is known as gamut mapping. There are many ways to remap colors from a larger
to a smaller gamut, with some of the possible approaches outlined in Section 15.4.
As display devices and color printers get more sophisticated, the opposite problem
is also becoming an active area of research, often called gamut expansion. How
we move colors that have been encoded in a limited range to a device capable of
displaying a much larger range has both similar and unique problems to traditional
gamut mapping.
The CIE XYZ color space is still actively used for converting between differ-
ent color spaces, as many color spaces are actually defined by means of a trans-
form to and from XYZ. We can think of CIE XYZ as a device-independent method
for representing colors, and any given output device can be described by its rela-
tionship to XYZ. It is important to stress that when a color space that corresponds
to physical primaries has a well-documented and specified relationship with CIE
XYZ, then this device can also be thought of as device-independent. Although the
space may be tied to physically realizable primaries, it is still possible to specify
any color match using that space (allowing for both positive and negative encoded
values), or also to derive a unique set of color-matching functions that correspond
to those primaries.
For simple, linear and additive trichromatic display devices the transformation
to and from CIE XYZ can usually be given by a 3 × 3 matrix. This transforma-
tion relies on Grassmann’s Laws of additive color mixing. Often, color spaces are
defined by the linear relationship to CIE XYZ, with additional non-linear process-
ing. In most cases, such non-linear processing is designed to minimize perceptual
errors when storing in a limited bit-depth without ample color precision, or to di-

i i

i i
 i i

i i

408 8. Color Spaces

R G B White
x 0.6400 0.3000 0.1500 0.3127
y 0.3300 0.6000 0.0600 0.3290

Table 8.1. The xy chromaticity coordinates for the primaries and white point specified by
ITU-R BT.709. The sRGB standard also uses these primaries and white point.

rectly be displayed on devices that have a non-linear relationship between signal

and energy emitted.
To construct such a linear 3 × 3 matrix transforming a specific additive display
device to CIE XYZ, the three primaries as well as the white point of the device (or
color space) need to be known. These are typically described as xy chromaticity
coordinates, though they can also be specified as CIE XYZ tristimulus values.
The xy chromaticity coordinates of the primaries represent the colors obtained
for the color vectors (1, 0, 0), (0, 1, 0), and (0, 0, 1). Basically, they represent
chromaticity of the maximum value of any given channel when displayed alone.
Recall from Chapter 7 that chromaticity coordinates are two-dimensional projec-
tions of a three-dimensional color space, and that they have been normalized to
remove the luminance component from the specification. Without the luminance
component, it is impossible to know the relative power or energy of the three pri-
maries, and so we cannot know what color will be shown when all three primaries
are displayed at their maximum. Thus, we need to include a fourth description of
the device; the display or color space white point. The white point of a color space
is the color obtained by the maximum of all three primaries, typically expressed
as the color vector (1, 1, 1). As an example, the ITU-R BT.709 specifies a color
space with primaries and white point as specified in Table 8.1. The white point is
effectively D65.
Assuming that the primaries represent some form of red, green, and blue, the
xy chromaticity coordinates can be extended to compute chromaticity triplets by
setting z = 1 − x − y. This leads to chromaticity triplets (xR , yR , zR ), (xG , yG , zG ),
and (xB , yB , zB ), as well as a triplet (xW , yW , zW ) for the white point. If we know
the maximum luminance of the white point, we can compute its tristimulus value
(XW ,YW , ZW ). Using these values, we solve the following set of equations for the
luminance ratio scalers SR , SG , and SB :

XW = xR SR + xG SG + xB SB , (8.1a)
YW = yR SR + yG SG + yB SB , (8.1b)
ZW = zR SR + zG SG + zB SB . (8.1c)

i i

i i
 i i

i i

409

The conversion between RGB and XYZ is then given by

⎡ ⎤ ⎡ ⎤⎡ ⎤
X xR SR xG SG xB SB R
⎣Y ⎦ = ⎣yR SR yG SG yB SB ⎦ ⎣G⎦ . (8.2)
Z zR SR zG SG zB SB B
More often than not, we do not have the maximum luminance of the white
point provided to us, rather just the chromaticity coordinates. Recall from Chap-
ter 7 that for a linear, additive display device we can determine the relationship
between CIE XYZ and the device RGB using the following equation:
⎡ ⎤ ⎡ ⎤⎡ ⎤
X XRmax XGmax XBmax R
⎣Y ⎦ = ⎣YRmax YGmax YBmax ⎦ ⎣G⎦ , (8.3)
Z ZRmax ZGmax ZBmax B
where XY ZRmax represents the tristimulus values of the primary when displayed at
its maximum power. When we do not have the full tristimulus values of the pri-
maries, we can solve for the relative tristimulus ratios in order to convert between
RGB and XYZ. We can expand Equation (8.3) to be a product of the chromaticity
coordinates along with the luminance component:

⎡  ⎤ ⎡     ⎤
xw xr xg xb
⎡ ⎤⎡ ⎤
⎢ yw Yw ⎥ ⎢ yr yb ⎥ 0 0
⎢ ⎥ ⎢ yg ⎥ Yr R
⎢ ⎥ ⎢ ⎥⎣ 0 ⎦ ⎣G⎦ .
⎢ Yw ⎥ = ⎢ 1   1   1 ⎥ 0 Yg (8.4)
⎣ zw ⎦ ⎣ zr zg zb ⎦ 0 0 Yb B
Yw
yw yr yg yb

This expansion still requires the knowledge of both the luminance of the
white, Yw , as well as the luminance of the red, green, and blue channels Yr ,Yg ,Yb .
Since we do not have this information, we can calculate the luminance ratios Y R
that would be necessary to obtain the chromaticity of the given white point, with
the chromaticities of the red, green, and blue channel we are given. First, we as-
sume that the maximum luminance Yw occurs when R = G = B = 1, and that it
has a luminance ratio of 1.0. Equation (8.4) can then be reduced to
⎡ ⎤ ⎡     ⎤
xw xr xg xb
⎡ ⎤
⎢ yw ⎥ ⎢ yr y y ⎥ R
b ⎥ Yr
⎢ ⎥ ⎢ g
⎢ 1 ⎥=⎢ 1 ⎥ ⎣ R⎦
⎢ ⎥ ⎢   1   1 ⎥ Yg . (8.5)
⎣ zw ⎦ ⎣ zr zg zb ⎦ YbR
yw yr yg yb
R , by inverting the
We can solve for the luminance ratios, represented by Yrgb
individual chromaticity matrix, as shown in Equation (8.6a). We then use these

i i

i i
 i i

i i

410 8. Color Spaces

luminance ratios to solve for the 3 × 3 matrix from Equation (8.3) that will trans-
form from device RGB into XYZ. This technique is also useful when you want
to force a set of primaries with given chromaticities into having a specific white
point chromaticity:
⎡     ⎤−1 ⎡ ⎤
⎡ R⎤ xr xg xb xw
Yr ⎢ yr yg yb ⎥ ⎢ yw ⎥
⎢ R⎥ ⎢ ⎥ ⎢ ⎥
⎣Yg ⎦ = ⎢⎢ 1  1
⎥ ⎢ ⎥
 1 ⎥ ⎢ 1 ⎥ ; (8.6a)
R ⎣ zr zg zb ⎦ ⎣ zw ⎦
Yb
yr yg yb yw
⎡     ⎤
xr xg xb
⎡ ⎤ ⎡ R ⎤
⎢ yr yb ⎥
XRmax XGmax XBmax ⎢ yg ⎥ Yr 0 0
⎣YRmax YGmax YBmax ⎦ = ⎢
⎢ 1  1
⎥⎣
 1 ⎥ 0 Yg
R 0 ⎦.
ZRmax ZGmax ZBmax ⎣ zr zg zb ⎦ 0 0 YbR
yr yg yb (8.6b)

The conversion from XYZ to RGB can be computed by inverting the 3 × 3

matrix calculated above. For any matrix constructed in this manner, the luminance
of any given color can by computed by evaluating

Y = yR SR + yG SG + yB SB . (8.7)

Alternatively, the luminance can be calculated by adding up the luminance of

each of the individual components, Y = YR + YG + YB . As discussed in previous
chapters, the Y channel in XYZ color space is identical to the photopic luminance
response function and since light is additive, we can calculate the luminance of
any given display by simply adding the luminance of each of the color channels.
If an image is presented in an RGB color space for which the primaries or the
white point are unknown, then it may be beneficial to assume that the image is en-
coded in a standard color space. Most standardized color spaces have well-defined
linear matrix transformations to and from XYZ. Many standard color matrices are
based upon actual primaries of output devices, or are designed to provide max-
imum precision to a desired range of colors. One common choice of encoding
(and broadcasting) color spaces is that of HDTV, whose primaries are specified
by ITU-R BT.709 (see Table 8.1). These primaries are also known as the Rec.
709 primaries; they were designed to be representative of the typical CRT dis-
play used for broadcast video and computer monitors [923]. It is also important
to note that the primaries for ITU-R BT.709 are identical to those of the sRGB
color space, which was also designed to be representative of the typical computer
display connected to the Internet [524]. The transformation between the ITU-R

i i

i i
 i i

i i

8.1. RGB Color Spaces 411

BT.709/sRGB primaries and the CIE XYZ tristimulus values is

⎡ ⎤ ⎡ ⎤⎡ ⎤
X 0.4124 0.3576 0.1805 R
⎣Y ⎦ = ⎣ 0.2126 0.7152 0.0722⎦ ⎣G⎦ ; (8.8a)
Z 0.0193 0.1192 0.9505 B
⎡ ⎤ ⎡ ⎤⎡ ⎤
R 3.2405 −1.5371 −0.4985 X
⎣G⎦ = ⎣−0.9693 1.8706 0.0416⎦ ⎣Y ⎦ . (8.8b)
B 0.0556 −0.2040 1.0572 Z
Again, we can calculate the luminance of any given color in this color space by
summing the values of the individual RGB channels:

Y = 0.2126 R + 0.7152 G + 0.0722 B. (8.9)

It is important to note that we can also calculate the white point of any given
display by calculating the XYZ tristimulus values of the maximum RGB values:
⎡ ⎤ ⎡ ⎤⎡ ⎤
XW 0.4124 0.3576 0.1805 Rmax
⎣YW ⎦ = ⎣0.2126 0.7152 0.0722⎦ ⎣Gmax ⎦ . (8.10)
ZW 0.0193 0.1192 0.9505 Bmax
The ITU-R BT.709 primaries have a defined white point equivalent to that of CIE
D65:
⎡ 709 ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
XW 0.4124 0.3576 0.1805 100 95.05
⎣Y 709 ⎦ = ⎣0.2126 0.7152 0.0722⎦ ⎣100⎦ = ⎣100.00⎦ . (8.11)
W
ZW709 0.0193 0.1192 0.9505 100 108.90
This tristimulus value is equivalent to the chromaticity values of the white point
listed in Table 8.1.
In this chapter, many of the more common color spaces used in various indus-
tries are discussed. These are generally grouped into device-specific and device-
independent color spaces. While it is tempting to group the color spaces that are
defined by a specific (and real) primary set as being device-dependent, if they
have a well-defined transform to and from a device-independent space such as
CIE XYZ, then these spaces are indeed device-independent. To facilitate imple-
mentation and ease of communication between the wide variety of color spaces,
we include transformation matrices where possible.

8.1 RGB Color Spaces

The phosphors used in cathode ray tubes and plasma devices, as well as the il-
luminants and colored filters used in LCD and DLP displays (see Chapter 14),

i i

i i
 i i

i i

412 8. Color Spaces

implicitly define the primaries and native white points of display devices. Typical
additive display devices use a red, a green, and a blue primary. However, different
display devices can have very disparate primaries, even if they are all described
as being RGB. As a result, each display device has a different range of colors,
or color gamut, that it is able to display. Nonetheless, these color spaces are all
generically referred to as RGB color spaces.
It is important to remember that, as a result, there is no such thing as the RGB
color space. Every device has its own color space, and these classes of RGB
color spaces are therefore called device-dependent; if we send the same RGB
values to each of the displays we will see different colors depending on their
choice of primaries. This is in contrast to the XYZ color space, which is well
defined, based on a set of imaginary primaries, and is an example of a device-
independent color space. The practical implication is that an image can appear
very different depending on the display, unless care is taken to process the data
prior to output. The process of measuring how a display handles color is called
device-characterization and is necessary to achieve color matches between differ-
ent devices (see Section 14.18).
To enable some consistency in how images appear on a display, the images
can be converted to a standardized set of primaries, with a standardized white
point. This standardized set can be specified by the target display, e.g., a display
requesting that all images be described using the ITU-R BT.709 primaries, which
then performs an internal conversion from this space into the primaries of the
display, or this conversion can be done using the model of modern color manage-
ment. This model is achieved by converting the image into a device-independent
space such as the XYZ or CIELAB color space and, subsequently, to the RGB
color space of the target display. All too often, in general practice, such correc-
tion does not happen and images appear different on different displays. This may
be because the color space of the image is unknown or the primaries and white
point of the display are unknown.
Some standardization has occurred in recent years, for instance, through the
use of the sRGB color space. Some cameras are able to output images in this
color space, and some display devices are sRGB compliant. This means that in
an imaging pipeline that employs the sRGB color space throughout, conversion
between different device-dependent RGB color spaces is unnecessary. We can
think of using the sRGB color space in this manner as a means for a default form
of color management. If all devices work in sRGB, then all devices theoretically
should match each other. This makes the sRGB color space suitable for such
consumer applications as shooting a digital image and displaying this image on a
CRT, or placing it on a website and having it appear somewhat consistent wher-

i i

i i
 i i

i i

8.1. RGB Color Spaces 413

ever it is viewed. If the output device is not assuming an sRGB input, however,
then the colors may mismatch.
The sRGB color space, like many other RGB color spaces, is defined by a
3 × 3 matrix as well as a non-linear luminance encoding. A maximum luminance
of 80 cd/m2 is assumed, as this was representative of a typical display when the
sRGB space was created. The matrix transforms between linearized sRGB and
XYZ is the same as that specified by ITU-R BT.709, as given in (8.8). The non-
linear encoding is given by

1.055 R1/2.4 − 0.055 R > 0.0031308,
RsRGB = (8.12a)
12.92 R R ≤ 0.0031308;


1.055 G1/2.4 − 0.055 G > 0.0031308,
GsRGB = (8.12b)
12.92 G G ≤ 0.0031308;

1.055 B1/2.4 − 0.055 B > 0.0031308,
BsRGB = (8.12c)
12.92 B B ≤ 0.0031308.

The non-linear encoding, which is also known as gamma encoding, and which
is therefore frequently confused with gamma correction, is required to minimize
quantization errors in digital applications. Since sRGB is an 8-bit quantized color
space, it is beneficial to have a non-linear relationship between the color values
and the luminance values that they represent. By compressing the data with a
power function less than 1.0, we use more bits of precision for the darker colors
where color differences are more perceptible. This helps keep quantization errors
below the visible threshold where possible. It is important to note that although
gamma correction and gamma encoding are different entities, the ultimate behav-
ior is very similar. Gamma correction is necessary when an output display device
responds non-linearly with regard to the input signal. Typical CRTs have a non-
linear relationship between input voltage and output luminance, often described
by a power function greater than 1.0 (generally around 2.4). Gamma correction
is necessary to compress the signal prior to sending it to a display, in order to get
a linear, or close to linear, response out. Since sRGB was defined to be represen-
tative of a typical CRT display, the gamma encoding also serves the purpose of
gamma correction and the non-linearly encoded data can be sent directly to the
output device.
For very small values in the sRGB space, the encoding is linear, as this pro-
duces better behavior for near-black values. When we combine the linear compo-
nent with the exponent of 1/2.4, we get a behavior that is very similar to having

i i

i i
 i i

i i

414 8. Color Spaces

just an exponent of 1/2.2, although it deviates for the dark colors. In practice, this
short linear ramp is often abandoned for simplicity of implementation, and a sin-
gle exponent of 1/2.2 is used. Note that this simplified formulation is a deviation
from the sRGB specification and is therefore not recommended.
The non-linear encoding given for the sRGB space can be parameterized as in
Equation (8.13) [882]:

(1 + f ) Rγ − f t < R ≤ 1,
Rnonlinear = (8.13a)
sR 0 ≤ R ≤ t;

(1 + f ) Gγ − f t < G ≤ 1,
Gnonlinear = (8.13b)
sG 0 ≤ G ≤ t;

(1 + f ) Bγ − f t < B ≤ 1,
Bnonlinear = (8.13c)
sB 0 ≤ B ≤ t.
Together with specifications for the primaries and the white point, the parameters
f , s, and t specify a class of RGB color spaces that are used in various industries.
The value of s determines the slope of the linear segment, and f is a small off-
set. The value of t determines where the linear slope changes into the non-linear
encoding.
Table 8.2 lists the primaries and white points of a collection of commonly
encountered device RGB spaces. The associated conversion matrices and non-
linearities are given in Table 8.3. The two-dimensional gamuts spanned by each
color space are shown in Figure 8.1, as projections in the CIE u v chromaticity
diagrams. The projected gamut for the HDTV color space is identical to the sRGB
standard, and it is therefore not shown again. It is important to stress that device
gamuts are limited by the range of colors that a real output device can create; how-
ever the encoding color space inherently has no gamut boundaries. Only when we
impose a limitation on the range of values that can be encoded, such as [0, 1] or
[0, 255] do color spaces themselves have gamuts. It is also important to point
out that the two-dimensional gamut projection into a chromaticity diagram is in-
herently simplifying gamut descriptions. Since we need at least three dimensions
to fully specify color appearance (see Chapter 11) color gamuts should also be
at least three-dimensional. Since the non-linear encoding of HDTV and sRGB is
different, it is possible for these spaces to have different gamuts when represented
in a three-dimensional space, such as CIELAB. For additive display devices, how-
ever, the triangle formed by the primaries in a chromaticity diagram is often used
in industrial applications to represent the ultimate gamut of the device.
The Adobe RGB color space was formerly known as SMPTE-240M, but was
renamed after SMPTE’s gamut was reduced. It has a larger gamut than sRGB, as

i i

i i
 i i

i i

8.1. RGB Color Spaces 415

shown in the chromaticity diagrams of Figure 8.1. This color space was developed
with the printing industry in mind, as there were many colors that could be printed
that could not be represented by a smaller gamut color space. It is important to
note that when a limited precision is used, e.g. 8-bits for each color channel,
then an extended gamut inherently means there will be less bits for all colors and
quantization may occur. Thus, it is important to choose an encoding color space
that is best representative of the most likely or most important colors, in order to
not “waste bits” where they are not needed. Many digital cameras today provide
an option to output images in either the Adobe RGB color, as well as sRGB.
High definition television (HDTV) and sRGB standards specify identical pri-
maries, but they differ in their definition of viewing conditions. This difference
is represented by the non-linear encoding. The sRGB space has a linear segment
and a power function of 1/2.4, while the ITU-R BT.709 has a linear segment and
power function of 1/2.2. These linear segments make the effective non-linearity
approximately 1/2.2 for sRGB and 1/2 for HDTV. Thus, if improper assumptions
are made between these color spaces, color mismatches can occur. The American
National Television Systems Committee (NTSC) standard was created in 1953
and has been used as the color space for TV in North America. As phosphors

Color space Primaries White point

R G B (Illuminant)
sRGB/ITU-R BT.709 x 0.6400 0.3000 0.1500 D65 0.3127
y 0.3300 0.6000 0.0600 0.3290
Adobe RGB (1998) x 0.6400 0.2100 0.1500 D65 0.3127
y 0.3300 0.7100 0.0600 0.3290
HDTV (HD-CIF) x 0.6400 0.3000 0.1500 D65 0.3127
y 0.3300 0.6000 0.0600 0.3290
NTSC (1953) x 0.6700 0.2100 0.1400 C 0.3101
y 0.3300 0.7100 0.0800 0.3161
PAL/SECAM x 0.6400 0.2900 0.1500 D65 0.3127
y 0.3300 0.6000 0.0600 0.3290
SMPTE-C x 0.6300 0.3100 0.1550 D65 0.3127
y 0.3400 0.5950 0.0700 0.3290
Wide Gamut x 0.7347 0.1152 0.1566 D50 0.3457
y 0.2653 0.8264 0.0177 0.3584

Table 8.2. Chromaticity coordinates for the primaries and white points of several com-
monly encountered RGB color spaces.

i i

i i
 i i

i i

416 8. Color Spaces

Color space XYZ to RGB matrix RGB to XYZ matrix Non-linear transform
⎡ ⎤ ⎡ ⎤ γ = 1/2.4 ≈ 0.42
3.2405 −1.5371 −0.4985 0.4124 0.3576 0.1805
⎣−0.9693 f = 0.055
sRGB 1.8760 0.0416⎦ ⎣ 0.2126 0.7152 0.0722⎦
s = 12.92
0.0556 −0.2040 1.0572 0.0193 0.1192 0.9505
t = 0.0031308

⎡ ⎤ ⎡ ⎤ γ = 1
51 ≈ 1
2.2
2.0414 −0.5649 −0.3447 0.5767 0.1856 0.1882 2 256

Adobe RGB ⎣−0.9693 1.8760 0.0416⎦ ⎣ 0.2974 0.6273 0.0753⎦ f = N.A.

(1998) 0.0134 −0.1184 1.0154 0.0270 0.0707 0.9911 s = N.A.
t = N.A.
⎡ ⎤ ⎡ ⎤ γ = 0.45
3.2405 −1.5371 −0.4985 0.4124 0.3576 0.1805
f = 0.099
HDTV (HD-CIF) ⎣−0.9693 1.8760 0.0416⎦ ⎣ 0.2126 0.7152 0.0722⎦
s = 4.5
0.0556 −0.2040 1.0572 0.0193 0.1192 0.9505
t = 0.018
⎡ ⎤ ⎡ ⎤ γ = 0.45
1.9100 −0.5325 −0.2882 0.6069 0.1735 0.2003
f = 0.099
NTSC (1953)/ ⎣−0.9847 1.9992 −0.0283⎦ ⎣ 0.2989 0.5866 0.1145⎦
s = 4.5
ITU-R BT.601-4 0.0583 −0.1184 0.8976 0.0000 0.0661 1.1162
t = 0.018
⎡ ⎤ ⎡ ⎤ γ = 0.45
3.0629 −1.3932 −0.4758 0.4306 0.3415 0.1783
f = 0.099
PAL/SECAM ⎣−0.9693 1.8760 0.0416⎦ ⎣ 0.2220 0.7066 0.0713⎦
s = 4.5
0.0679 −0.2289 1.0694 0.0202 0.1296 0.9391
t = 0.018
⎡ ⎤ ⎡ ⎤ γ = 0.45
3.5054 −1.7395 −0.5440 0.3936 0.3652 0.1916
f = 0.099
SMPTE-C ⎣−1.0691 1.9778 0.0352⎦ ⎣ 0.2124 0.7010 0.0865⎦
s = 4.5
0.0563 −0.1970 1.0502 0.0187 0.1119 0.9582
t = 0.018
⎡ ⎤ ⎡ ⎤ γ = N.A.
1.4625 −0.1845 −0.2734 0.7164 0.1010 0.1468
f = N.A.
Wide Gamut ⎣−0.5228 1.4479 0.0681⎦ ⎣ 0.2587 0.7247 0.0166⎦
s = N.A.
0.0346 −0.0958 1.2875 0.0000 0.0512 0.7740
t = N.A.

Table 8.3. Transformations for standard RGB color spaces (after [882]).

and other display technologies now allow much more saturated colors, it has been
deprecated and replaced with SMPTE-C to match phosphors in current display
devices, which are more efficient and brighter. Phase alternation line (PAL) and
Séquentiel Couleur à Mémoire (SECAM) are the standards used for television in
Europe. These spaces are discussed further in Section 8.4.
Finally, the Wide gamut color space is shown for comparison [882]. Its pri-
maries are monochromatic light sources with wavelengths of 450, 525, and 700
nm. By moving the primaries closer to the spectrum locus, a larger triangle of rep-
resentable colors is formed. In the limit, the primaries become monochromatic,
as in the case of this color space.

i i

i i
 i i

i i

8.1. RGB Color Spaces 417

0.7 0.7 0.7

v’ Adobe (1998) Gamut v’ sRGB Gamut v’ SMPTE-C Gamut
0.6 0.6 0.6
0.5 0.5 0.5
0.4 0.4 0.4
0.3 0.3 0.3
0.2 0.2 0.2
0.1 0.1 0.1
0.0 0.0 0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
u’ u’ u’

0.7 0.7 0.7

v’ NTSC Gamut v’ PAL/SECAM Gamut v’ Wide Gamut
0.6 0.6 0.6
0.5 0.5 0.5
0.4 0.4 0.4
0.3 0.3 0.3
0.2 0.2 0.2
0.1 0.1 0.1
0.0 0.0 0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
u’ u’ u’

Figure 8.1. CIE (u ,v ) chromaticity diagrams showing the color gamuts for various color
spaces.

The shift of primaries toward the spectrum locus yield primaries that, when
plotted against wavelength, are more peaked. This is where the name sharpen-
ing stems from. Hence, color spaces formed by employing primaries with such
peaked response functions are called sharpened color spaces. The associated
widening of the gamut is beneficial in several applications, including white bal-
ancing, color appearance modeling (Chapter 11), and rendering (Section 8.12).
Transforms for sharpened color spaces are introduced in Section 8.6.2.
Finally, to demonstrate that conversion between color spaces is necessary for
correct color reproduction, Figure 8.2 shows what happens when an image’s pri-
maries are misinterpreted. In this figure, we have converted an image between
different RGB color spaces. This has the same effect as displaying an image
on different monitors without adjusting for the monitors’ primaries. The error
created with this procedure is what is commonly seen when color conversion is
omitted. Note that even if care is taken to transform image data between color
spaces, it is also important to assure that white points are correctly mapped to-

i i

i i
 i i

i i

418 8. Color Spaces

Figure 8.2. A set of images obtained by converting to different RGB color spaces. These
images demonstrate the types of error that can be expected if the primaries of a display
device are not taken into account. From left to right and top to bottom: the original image
(assumed to be in sRGB space) and the image converted to Adobe RGB, NTSC, PAL, and
SMPTE-C spaces; Zaragoza, Spain, November 2006.

gether. For instance, when moving from a space with a white point of D65 to
one with a white point of D50, we cannot just move into CIE XYZ and then back
out into the new RGB color space. We must first perform a chromatic-adaptation
transform between the white points, as discussed in Chapter 10.

8.2 Printers
Printers are one technology that can be used to produce hardcopy images. We
use the term hardcopy for printed images because once an image is created it
cannot be changed, unlike the softcopy display device. Typically a printed image

i i

i i
 i i

i i

8.2. Printers 419

is formed by depositing specified amounts of colorants onto paper or some other

substrate. If we assume that that three different inks are used to create a color,
the amount of each colorant used to form a colored dot can be expressed as a
triplet of percentages. This triplet forms another type of color cube, known as
the colorant space. An example of colorant spaces are the well-known CMY and
CMYK spaces, which are discussed in Section 8.2.1.
To encapsulate the range of colors that any given set of colorants can produce,
it is generally necessary to first print those colorants onto paper and then measure
the printed pages. These measurements, when expressed in a device-independent
space such as CIE XYZ or CIELAB, form a three-dimensional volume essentially
bounded by a deformed color cube. This irregular shape is then the gamut of
the printing device. Unlike softcopy additive display devices, subtractive color
mixing of printed colorants does not behave in a simple, linearly additive manner.
The gamut boundary reflects this non-linearity, as the size and shape depends on
the colorants used, the size of each dot, the opacity of the inks, the amount each
dot overlaps with previously deposited dots, the order in which different colors
are applied, and the paper used for printing.
It is often desirable to define for each printer a printer model which is a pre-
dictive model of the printer’s gamut as well as the necessary amount of colorants
necessary to generate a desired output color, given the above printer-dependent
features. The Kubelka-Munk equations, given in Section 3.4.6, can for instance
be used to map a printer’s colorant space to a linear set of equations that can be
used for predictive purposes. An alternative additive predictive model utilizes the
Demichel and Neugebauer equations, which are discussed in Section 8.2.2.

8.2.1 CMY and CMYK Colorant Spaces

The CMY and CMYK colorant spaces are two common spaces utilized by print-
ers. The ink deposited on paper reduces the amount of light that is reflected from
the surface, and so this type of color creation is often called a subtractive color sys-
tem. For example, the CMY color space uses cyan, magenta, and yellow primaries
to generate colors. We can think of, as a loose approximation, cyan colorants as
“subtracting” red light, magenta colorants as subtracting green light, and yellow
colorants as subtracting blue light. Color values encoded in these type of spaces
specify how much ink, toner, or dye needs be deposited for each point in the im-
age to generate the desired output color. Often, a fourth black ink is used, which
relaxes the constraints posed on the other three inks to produce neutral black. The
extra black ink should be considered a practical addition for color reproduction
using subtractive colorants. If black is used, colors can be specified in the CMYK
color space, where the K indicates black. Many modern printers now use a large

i i

i i
 i i

i i

420 8. Color Spaces

number of colorant primaries, in order to achieve accurate color reproduction, as

well as an expanded gamut. It is not uncommon to see printers with six, ten, or
even 12 different colorants or primaries. With so many primaries to choose from,
it is generally up to the device manufacturers to determine the combinations of
primaries, halftone patterns, and ink coverage necessary to generate any desired
output color. For simplicity sake, we will focus on the more traditional CMY and
CMYK printers as the methods for handling increasing numbers of primaries are
similar.
The CMY and CMYK spaces are inherently device-dependent, as the actual
tristimulus values of the output depend strongly on the types of colorants used,
the technology of the printer itself, the properties of the halftone pattern used, the
surface characteristics of the paper, and in many cases on the age of the print as
well. In most cases, printing itself is inherently a binary process, as there is either
colorant on a page or not. A wide range of colors can be produced by spatially
varying the amount and location of colorants on the printed page. Furthermore,
there is no reason to suspect that the transfer function between specified CMYK
values and the amount of ink produced on paper is linear or uniform. In gen-
eral, all of the variables that can affect the color produced by printers are not
standardized, and thus differ on a per-printer type. Color management using a
device-independent color space is one technique that has been used with success
to achieve accurate and reliable colors across a wide variety of printing devices.
That said, a common conversion that needs to be carried out is between a
given RGB color space and CMY(K). As both classes of color space can be highly
device dependent, it is not possible to define a single transform that will provide
acceptable results for all applications. A more appropriate way to perform such a
task would be to first convert RGB values into a device-independent space, such
as XYZ tristimulus values, and from there construct a conversion to CMY(K)
through a printer model or an exhaustive series of measurements. This conversion
is known as device characterization and forms the basis of most modern color-
management systems.
For non-critical color reproduction, computationally cheap approximations
are sometimes used. Common use of such conversions is one of the possible
reasons that printed output rarely matches the softcopy output. The simple (and
generally incorrect) conversion between RGB and CMY, recalling that a subtrac-
tive system removes light from a scene, is as follows:

C = 1 − R, (8.14a)
M = 1 − G, (8.14b)
Y = 1 − B; (8.14c)

i i

i i
 i i

i i

8.2. Printers 421

and

R = 1 −C, (8.15a)
G = 1 − M, (8.15b)
B = 1 −Y. (8.15c)

If a black ink is used, we can calculate a simple approximation using what is

often called gray component replacement (GCR) or undercolor removal (UCR).
Traditionally, we would say that UCR reduces the total amount of colorant cov-
erage necessary to achieve a neutral color, whereas GCR replaces the CMY pri-
maries with the black primary. This may be done to remove any color cast from
the CMY primaries and achieve a more neutral gray, as a monetary savings if
black colorant is cheaper, or to minimize the color inconstancy of the printed
page. Kang argues that GCR is a more sophisticated type of algorithm, and that
UCR can actually be considered a subset of GCR [575]. Again, using a very sim-
plified algorithm, one such conversion between CMY and CMYK using a type of
GCR is

K = min (C, M,Y ) , (8.16a)


C = C − K, (8.16b)

M = M − K, (8.16c)

Y = Y − K. (8.16d)

The process therefore removes some of the cyan, magenta, and yellow ink, and
replaces it with black. The most that can be removed is the minimum of the three
C, M, and Y inks, so that neutral colors are preserved. The new values for C, M,
and Y are as they were before, but with the value of K subtracted. Note that it
would be possible to chose any value for K as long as

0 ≤ K ≤ min (C, M,Y ) . (8.17)

The reverse conversion is given by

C = C + K, (8.18a)

M = M + K, (8.18b)

Y = Y + K. (8.18c)

We must again stress that the CMY(K) values obtained in this manner are directly
related to the primaries and white point of the original (perhaps unknown) RGB
color space. These primaries will no doubt be very different from the CMY(K)

i i

i i
 i i

i i

422 8. Color Spaces

primaries of whichever printing device will ultimately be used, and so the results
of this type of RGB to CMY(K) conversion can only be considered approximate
at best.
The color gamut of the CMY(K) spaces are typically said to be much smaller
than the gamuts of RGB devices. This means that there are ranges of colors that
can be displayed on a monitor that can not be reproduced in print. Section 15.4
discusses gamut mapping, a process that maps a larger gamut to a smaller one. It
is important to stress that not all printers have a smaller gamut than all additive
display devices, especially for colors closely associated with the printer primaries,
such as bright yellows or dark cyan colors. The ultimate gamut of a printer device
is heavily tied to the materials used to create the printed colors. For example, the
same printer may produce a gamut much smaller than a typical LCD display when
printing on plain or matte paper, but may have a gamut much larger than the LCD
when printing on a high-quality glossy paper. In some instances, gamut-mapping
algorithms may be best served by both reducing a larger gamut to a smaller one in
some areas, and mapping a smaller to a larger (gamut expansion) in others. This
just suggests that overarching statements such as “displays have larger gamuts
than printers” should be met with a high degree of skepticism.
More recently, printers are using more than four inks to achieve a higher de-
gree of color fidelity. For instance, some desktop printers have added two extra
inks for light cyan and light magenta. This resulting device-dependent colorant
space is known as CcMmYK, in which the c and m indicate light cyan and light
magenta, respectively. This may be done to reduce the visibility of the half-toning
process, and to create a more uniform printed surface. Other multiple ink config-
urations, such as the addition of green, red, orange, blue, or even clear inks, are
certainly possible as well. These inks can be added to increase the overall gamut
of the printer, increase the color constancy of the printer, or to increase overall
image quality (perhaps by reducing the size of the halftone dots).

8.2.2 The Demichel-Neugebauer Printer Model

The Demichel-Neugebauer model can be used to infer the gamut boundaries of
printed colorants on paper. It is therefore an empirical mapping between the
device-colorant space and the achievable colors on paper. Typically, we con-
sider printers and colorants to be a form of subtractive color mixing. Subtractive
color mixing is generally a non-linear process, and more advanced mathematics
are necessary to model and predict these non-linearities. If we consider that the
binary process (either colorant or no colorant) of printers produces a series of
dots that are then integrated by our visual system, then we can consider this inte-

i i

i i
 i i

i i

8.2. Printers 423

gration to be an additive process. While the mixing of the colorants themselves

are most likely non-linear, the mixing of the colors that the eye perceives can be
modeled as an additive mixture of the colorant’s spectral reflectance distributions.
Essentially this is what the Demichel-Neugebauer equations, often referred to as
just the Neugebauer equations, do. On a simplified level, the Demichel equa-
tions form a probabilistic model predicting how much of any given primary forms
the basis of a color, and the Neugebauer equations model the additive color mix-
ing of the primaries. The Neugebauer equations are often used for three-ink and
four-ink processes, but can be extended for more inks. These equations are most
appropriate to predict colors resulting from halftone printing.
In halftone printing, each of the colored inks is applied sequentially, with
dots of ink either placed next to one another or partially overlapping. If three
inks are utilized, then the paper may be covered by one ink, two inks, three inks,
or may be left blank. There are eight combinations of overlap, assuming that
the order in which inks are applied, is fixed. The eight colors associated with
these combinations of overlap are called the Neugebauer primaries. We indicate
these primaries with XW YW ZW , X1Y1 Z1 , X2Y2 Z2 , X3Y3 Z3 , X12Y12 Z12 , X13Y13 Z13 ,
X23Y23 Z23 , and X123Y123 Z123 .
The tristimulus value of a color on paper is then a weighted combination of
these eight primaries, given by the Neugebauer equations [828]:

X = aW XW + a1 X1 + a2 X2 + a3 X3 + a12 X12 + a13 X13 + a23 X23 + a123 X123 ,

(8.19a)
Y = aW YW + a1 Y1 + a2 Y2 + a3 Y3 + a12 Y12 + a13 Y13 + a23 Y23 + a123 Y123 ,
(8.19b)
Z = aW ZW + a1 Z1 + a2 Z2 + a3 Z3 + a12 Z12 + a13 Z13 + a23 Z23 + a123 Z123 .
(8.19c)

Note that the Neugebauer equations are traditionally calculated as a weighted

sum of spectral reflectances rather than tristimulus values, though due to the addi-
tivity of light and Grassmann’s laws we can use either. The Neugebauer primaries
consist of the printer primaries themselves, as well as the combinations of these
primaries, which form secondaries. For a CMY printer, the secondaries are red
(M + Y ), green (C + Y ), and blue (M + C), while black is formed by adding all
three primaries together and white is formed by just the paper itself. The relative
weights for each of the Neugebauer primaries are calculated using the percentage
coverage for each printer primary as a halftone dot percentage, given by c1 , c2 ,
and c3 . The relationship between the dot-area percentages and the weight factors

i i

i i
 i i

i i

424 8. Color Spaces

aW , a1 , . . . are given by the Demichel equations [246, 247]:

aW = (1 − c1 )(1 − c2 )(1 − c3 ), (8.20a)

a1 = c1 (1 − c2 )(1 − c3 ), (8.20b)
a2 = (1 − c1 ) c2 (1 − c3 ), (8.20c)
a3 = (1 − c1 )(1 − c2 ) c3 , (8.20d)
a12 = c1 c2 (1 − c3 ), (8.20e)
a13 = c1 (1 − c2 ) c3 , (8.20f)
a23 = (1 − c1 ) c2 c3 , (8.20g)
a123 = c1 c2 c3 . (8.20h)

We should consider the weights calculated by the Demichel equations to be prob-

abilities of any given primary, based upon the percentage of dot-area coverage.
For instance if any given area has a high percentage of cyan and yellow colorant,
then the probability of that area being blue increases. The combined Demichel-
Neugebauer model can thus be used to predict the XYZ tristimulus value for each
of the eight Neugebauer primaries that are generated in a three-ink process. The
equations can be extended for more than three inks, though the number of Neuge-
bauer primaries increases exponentially with the number of printer inks used. In
order to properly utilize the Demichel-Neugebauer equations, it is necessary to
know the tristimulus values, or spectral reflectance of each of the Neugebauer
primaries. Typically this is done through careful printing and measurement. The
Neugebauer equations also have a fundamental assumption of additivity of the
light, though in practice absorption and scattering of the inks, as well as optical
diffusion of light through the inks and paper affect the additivity.
To test the additivity of any given printer a set of uniform color ramps can
be generated and measured. If the system behaves in a relatively linear man-
ner, and we know how much of each of the colorant will be applied, we can use
the Demichel-Neugebauer equations to predict the output color on paper. Fur-
ther information on printer color modeling can be found in the texts of Yule and
Kang [575, 1287]. Wyble gives an excellent overview of more recent models
of modeling color printers using the Demichel-Neugebauer equations along with
other techniques [1258].
The opposite problem, namely that of color separation, is concerned with find-
ing how much colorant to apply to reproduce a given color specified in XYZ color
space. The nature of this problem is revealed by substituting the Demichel equa-
tions into the Neugebauer equations, yielding a polynomial in the dot percent-
ages. Solving these equations for c1 , c2 , and c3 leads to an expression mapping

i i

i i
 i i

i i

8.2. Printers 425

an XYZ color to dot percentages for each of the three colorants in terms of the
eight Neugebauer primaries (which need to be measured for these equations to be
useful).
As the solution is non-linear, an analytical inversion of the printer model can
only be achieved if simplifying assumptions are made. Otherwise, regression
techniques can be used to express each of the three dot percentages in terms of
X, Y , and Z. First, we will assume that the color of paper results in a tristimulus
value of (XW ,YW , ZW ) = (1, 1, 1). Second, we assume that the densities of inks
are additive, i.e., X12 = X1 X2 and so on for all other ink combinations [912–914].
This allows the Neugebauer equations to be factored as follows [1287]:

X = (1 − c1 + c1 X1 ) (1 − c2 + c2 X2 ) (1 − c3 + c3 X3 ) , (8.21a)
Y = (1 − c1 + c1 Y1 ) (1 − c2 + c2 Y2 ) (1 − c3 + c3 Y3 ) , (8.21b)
Z = (1 − c1 + c1 Z1 ) (1 − c2 + c2 Z2 ) (1 − c3 + c3 Z3 ) . (8.21c)

A simplification is obtained by assuming that X2 = X3 = Y3 = 1, so that we have

X = (1 − c1 + c1 X1 ) , (8.22a)
Y = (1 − c1 + c1 Y1 ) (1 − c2 + c2 Y2 ) , (8.22b)
Z = (1 − c1 + c1 Z1 ) (1 − c2 + c2 Z2 ) (1 − c3 + c3 Z3 ) . (8.22c)

Although these equations can be inverted, a further simplifying assumption will

make inversion more straightforward. In particular, it may be assumed that X1 =
Y2 = Z3 = 0, so that we get

1 − c1 = X, (8.23a)
Y
1 − c2 = , (8.23b)
Y1 + (1 −Y1 ) X
Z
1 − c3 = . (8.23c)
(Z1 + (1 − Z1 ) X) (Z2 + (1 − Z2 )(1 − c2 ))
The three-ink model may be extended to four colors, for use with CMYK
printer models [426]. While the procedure is straightforward [577], the equations
become much more difficult to invert so that numerical techniques are required.
Real printing processes often deviate from the Demichel-Neugebauer model.
The above equations therefore have to be used with care. The Neugebauer printer
model can suffer from accuracy problems when the general assumptions are sim-
plified beyond physical reality.
Various extentions have been proposed [540, 907], including the Clapper-
Yule model for multiple inks [196, 197, 1287] and the spectral Neugebauer equa-
tions [1182]. The Clapper-Yule model itself was extended to model printing with

i i

i i
 i i

i i

426 8. Color Spaces

embedded metallic patterns [465]. Another class of improved models is known

as the cellular or localized Neugebauer equations [725, 864], where color space
is subdivided into disjunct regions that are treated independently.
An extension that accounts for dot gain may also be incorporated into the
Neugebauer equations. When a droplet of ink is deposited onto paper, it spreads
out. This is called the physical dot gain. Light scattering inside the paper will
also cause dots to absorb more light than the area coverage would predict, a phe-
nomenon called optical dot gain. The desired coverage ci can be recovered using
the following empirical relation:

1 − ci = (1 − c̄i )γi , (8.24)

where c̄i is the specified dot percentage, and γi is an ink-specific parameter that
can be derived from experimental data [1093]. Again, it is important to stress
that this non-linear exponent is only an empirical approximation to better predict
printed results and should not be taken as a model of a real physical process.
A further extension takes into account trapping. This is a phenomenon
whereby the amount of ink that sticks to the page is less than the amount of ink
deposited. Trapping thus reduces the dot percentage of coverage by a fraction tip ,
where i indicates ink i sticking directly to the paper. For overprinting, the fraction
tipj models the trapping of ink i printed on top of ink j. The Neugebauer equations
(given only for X) then become [1093]:

X = (1 − t1p c1 )(1 − t2p c2 )(1 − t3p c3 ) XW (8.25a)

+ t1p c1 (1 − t12p p
c2 )(1 − t13 c3 ) X1 (8.25b)
+ (1 − t1p c1 )t12 p p
c2 (1 − t13 c3 ) X2 (8.25c)
p p p
+ (1 − t1 c1 )(1 − t2 c2 )t3 c3 X3 (8.25d)
+ t1p c1 t12
p p
c2 (1 − t123 c3 ) X12 (8.25e)
+ t1p c1 (1 − t12p p
c2 )t13 c3 X13 (8.25f)
p p p
+ (1 − t1 c1 )t2 c2 t23 c3 X23 (8.25g)
+ t1p c1 t12
p p
c2 t123 c3 X123 . (8.25h)

The Neugebauer equations assume that the primaries of the printing inks are
known, or that they can be measured. This may not always be practical to do. In
that case, it may be possible to compute an approximation of the primaries using
a more complicated non-linear model such as the Kubelka-Munk theory used to
calculate the absorption and scattering of translucent and opaque pigments [1093]
(see Section 3.4.6).

i i

i i
 i i

i i

8.3. Luminance-Chrominance Color Spaces 427

8.3 Luminance-Chrominance Color Spaces

RGB color spaces are generally characterized by three physical primaries, or three
primaries that are well defined in a device-independent manner. For any three-
channel color space, we can say that tristimulus values that determine the amount
of each of the three primaries are necessary to generate the final color. Although
these spaces are useful for specifying and encoding color for direct transmission
and encoding to real devices, it is often convenient to encode color into a lu-
minance channel and two chrominance channels. The luminance channel then
represents how light or dark the color is, and should (by definition) relate directly
to the Y channel of the CIE XYZ color space. The two chrominance channels
together then determine the chromatic content of any given color.
One utility of such color representation is the use of the chrominance channels
to generate chromaticity diagrams. Recall from Chapter 7 that a chromaticity dia-
gram is a two-dimensional projection of a three-dimensional space. Chromaticity
diagrams can be useful for plotting the general location of chromatic primaries in
an additive display device, or for determining complementary colors of a given
stimulus. Common example of chromaticity diagrams are formed by plotting the
x against the y channel of the Yxy color space, or the more perceptually uniform
chromaticity diagram created by the u and v channels of the CIE Yu v space.
Although it is tempting to use chromaticity diagrams to plot the gamut of additive
display devices, such as CRTs and LCDs, it should be readily evident that the two-
dimensional projection immediately eliminates a large amount of information. In
most luminance-chrominance spaces, this information loss is the actual luminance
information, which means we cannot determine the actual color appearance of any
given point in a chromaticity diagram. Representing device gamuts, which inher-
ently need to have at least three color dimensions in a chromaticity space, can be
misleading, at best, and completely inaccurate, at worst.
Another reason for using luminance-chrominance color spaces is that the hu-
man visual system naturally utilizes such a representation for transmission of sig-
nals between the retina and the visual cortex. The visual system utilizes a type
of color opponent space, with a white-black channel, a reddish-greenish channel,
and a bluish-yellowish channel. Due to the nature of generating these channels
from the LMS cone signals, the visual system’s spatial sensitivity (acuity and con-
trast sensitivity) is generally higher for variations in luminance than for variations
in chrominance. As all cone types contribute to the formation of a luminance sig-
nal in the human visual system, its spatial resolution is generally determined by a
single cone receptor. However, color opponent signals require at least two cones
(four if the random distribution of L and M cones is taken into account [785]). As

i i

i i
 i i

i i

428 8. Color Spaces

Figure 8.3. The images on the left have all three channels subsampled, whereas the images
on the right have only their chrominance channels subsampled. From top to bottom the
subsampling is by factors of 2, 4, 8, 16, and 32; Long Ashton, Bristol, UK.

i i

i i
 i i

i i

8.3. Luminance-Chrominance Color Spaces 429

a result, color opponent signals have a spatial resolution corresponding to about

four cone diameters [505].
It is therefore possible to use a lower spatial resolution for encoding and trans-
mitting the chrominance channels than for the luminance channel, without greatly
affecting how such encoded images are perceived. As luminance and chromi-
nance are not encoded separately in RGB color spaces, the same efficient encod-
ing could not be achieved in RGB.
An example of the effect of lowering the spatial resolution of the chromi-
nance channels is shown in Figure 8.3. The images on the left in this figure have
the spatial resolution of all three channels progressively lowered. On the right,
only the chrominance channels are subsampled by corresponding amounts. The
effect of lowering the spatial resolution is far more notable when all channels are
subsampled.
This feature is exploited by the transmission encodings used in television sig-
nals to save bandwidth, an idea attributed to A. V. Loughren [507]. Television
encodings are further discussed in Section 8.4. The ability to spatially subsample
chrominance channels with minimal perceptual artifacts is also taken advantage
of in most modern image-compression algorithms, such as JPEG.

8.3.1 CIE XYZ and Yxy

The CIE XYZ color space was introduced in Chapter 7. Due to the device-
independent nature of the CIE XYZ space, along with its basis with the human
visual system, CIE XYZ is often used to define many other color spaces. One
important consideration in the derivation of the CIE XYZ space was the ability
to specify all visible colors with positive values for X, Y , and Z. Due to the high
degree of overlap between the cone responses, this is not possible using only three
physically realizable primaries. As a result, the CIE XYZ space is defined using
imaginary primaries.
XYZ tristimulus values can be plotted on a chromaticity diagram by comput-
ing chromaticity coordinates, as shown in Section 7.5.1. The combination of the
XYZ luminance channel Y and two chromaticity coordinates x and y can be re-
garded as a color space. As chromaticity coordinates are normalized, they carry
no information about luminance. In the Yxy color space, luminance is therefore
separated from chrominance, though we cannot say that the xy chromaticities
represent red-green and yellow-blue opponent color representations.
In chromaticity diagrams, the spatial difference between pairs of points is a
poor indicator of perceptual difference between pairs of colors. An alternative
way of describing this is that the xy chromaticity diagram is not perceptually uni-
form. This has given rise to several other color spaces, which can be regarded as

i i

i i
 i i

i i

430 8. Color Spaces

improvements over Yxy in this regard. Examples are the CIE Yuv, CIE u v uni-
form chromaticity scale (UCS) diagrams, and also the CIE L∗ a∗ b∗ (CIELAB) and
CIE L∗ u∗ v∗ (CIELUV) color spaces. It should be noted that although the CIELAB
color space was developed as a perceptually uniform luminance-chrominance
space, it does not have a defined chromaticity diagram that goes along with it.

8.3.2 CIE Yuv (1960) and CIE Yu v

In the CIE Yuv color space, the luminance channel is the same as in the CIE
XYZ and CIE Yxy color spaces (see Section 7.5.1). However, the chromaticity
coordinates are defined differently, to achieve a more uniform chromaticity plane:

2x
u= , (8.26a)
6y − x + 1.5
3y
v= . (8.26b)
6y − x + 1.5

A further refinement was introduced by the CIE Yu v color space, achieving yet
greater uniformity in its chromaticity coordinates:

2x
u = , (8.27a)
6y − x + 1.5
4.5y
v = . (8.27b)
6y − x + 1.5

Note that the only difference between Yuv and Yu v is between the v and v chan-
nels, whereby v = 1.5 v. Finally, as both these color spaces are derived from CIE
XYZ, they can also be considered to be well-defined device-independent color
spaces. It should also be noted that the Yu v space generally has superseded the
Yuv space, although for historical compatibility that space is still used for calcu-
lated correlated color temperatures.

8.4 Television and Video

Broadcasting involves several stages, starting with the capture of video or film,
followed by various stages of processing on the captured and encoded signal.
This processed signal is then transmitted either by air, by satellite, by cable, or by
a combination of all of these methods. Once received, the signal is decoded and
displayed on a television.

i i

i i
 i i

i i

8.4. Television and Video 431

To ensure that this series of events leads to an image that is reasonably close
to the image captured at the start of this sequence, standard definitions for each of
the processing steps are required. Such standards pertain to the color spaces used
at various stages, the encoding of the signal, sample rates, the number of pixels in
the image, and the order in which they are scanned.
Unfortunately, there is no single standard, and different broadcasting systems
are currently in use. For instance, for scanning North America and Japan use the
480i29.97 system, whereas most of the rest of the world uses the 576i25 system.1
The notation refers to the number of active scanlines (480 or 576), the i refers to
interlaced video, followed by the frame rate. Note that the North American system
has in total 525 scanlines. However, only 480 of them appear on screen. Aside
from interlaced video, progressive scan systems are also in use. For instance, high
definition television (HDTV) includes standards for both; 720p60 and 1080i30.
The transmission of signals requires the conversion of a captured RGB signal
into a space that is convenient for transmission. Typically, a non-linear transfer
function is applied first, leading to an R G B signal. The general form of this
non-linearity is given in (8.13). Following Charles Poynton, in this section primed
quantities denote non-linear signals [923].
The luminance-chrominance color spaces are suitable, as they allow the
chrominance channels to be compressed, saving transmission bandwidth by ex-
ploiting the human visual system’s relatively poor color acuity. Converting from
a linear RGB space to one of these spaces yields luminance and chrominance
signals. However, if the input to the transform is deliberately non-linear, as in
R G B , the conversion does not result in luminance and chrominance, but in a
non-linear luminance and chrominance representation. These are called luma and
chroma. Note that the non-linear luma representation is a rough approximation to
the color appearance description of lightness, though the chroma bears little to no
resemblance to the color appearance attribute of chroma. These appearance terms
are defined later in Chapter 11.
An image encoded with chroma subsampling is shown in Figure 8.3. The
sampling is usually described by a three-fold ratio, e.g., x:y:z which represents
the horizontal and vertical subsampling. The first number represents the hori-
zontal sampling reference of luminance with respect to a sampling rate of 3.375
MHz; thus a 4:4:4 represents a luminance sampling rate of 13.5 MHz. The second
number refers to the chrominance horizontal factor with respect to the first num-
ber. The third number is either the same as the second or it is zero in which case

1 We use the notation introduced by Charles Poynton in Digital Video and HDTV: Algorithms and

Interfaces [923].

i i

i i
 i i

i i

432 8. Color Spaces

4:4:4 4:2:2 4:1:1 4:2:0

Luminance

Color 1

Color 2

Figure 8.4. Four instances of chroma subsampling: 4:4:4, 4:2:2, 4:1:1, and 4:2:0.

the vertical color resolution is subsampled at a factor of 2. Examples of chroma

subsampling are 4:4:4, 4:2:2, 4:1:1, and 4:2:0, as illustrated in Figure 8.4.
The encoding of signals is standardized by the NTSC standard for North
America and PAL for Europe. These encodings turn RGB into a composite signal
that is used in analog broadcast. On the other hand, newer video equipment uses
component video, denoted Y  PB PR for analog components or Y CBCR for digital
components. Older equipment may use composite video. High definition televi-
sion always uses either analog or digital component video.
The several systems in use today are identified by the number of active scan-
lines and characterized by the chromaticity coordinates of the primaries, the non-
linear transfer function, and the coefficients used to compute luma and chroma.
Each of these parts is governed by different standards, as summarized in Table 8.4.
With both standard definition television (SDTV; 480i and 576i systems) and
HDTV (720p, 1080p, and 1080i systems) currently in use, the current situation
is that there are two encoding systems for luma and chroma. This means, for
instance, that there exist two different forms for YCBCR , dependent on which
system is used. The same is true for its analog equivalent Y PB PR . The accurate
interchange of signals across different systems is therefore problematic.

System Primaries Transfer Luma

480i SMPTE RP 145 Rec. 709 Rec. 601
576i EBU Tech. 3213 Rec. 709 Rec. 601
720p Rec. 709 Rec. 709 Rec. 709
1080p Rec. 709 Rec. 709 Rec. 709
1080i Rec. 709 Rec. 709 Rec. 709

Table 8.4. Primaries, non-linear transfer function, and luma encoding for the different
television systems currently in use (after [923]).

i i

i i
 i i

i i

8.4. Television and Video 433

8.4.1 EBU Y U V
Starting with linear RGB values, the EBU standard, which is used in European
PAL and SECAM color encodings, assumes that the D65 white point is used,
as well as the primaries listed under PAL/SECAM in Table 8.2. The associated
transformation matrix, as well as the parameters for the non-linear transfer func-
tion are listed under the PAL/SECAM entry in Table 8.3. The R G B signal is
then converted to Y U V transmission primaries. Here, the chromatic channels
are formed by subtracting luma from blue and red, respectively [975]:

U  = 0.492111(B −Y  ), (8.28a)
  
V = 0.877283(R −Y ). (8.28b)

This leads to the following transformation matrix:

⎡ ⎤ ⎡ ⎤⎡ ⎤
Y 0.299 0.587 0.114 R
⎣U  ⎦ = ⎣−0.147141 −0.288869 0.436010⎦ ⎣G ⎦ . (8.29)
V 0.614975 −0.514965 −0.100010 B

The inverse is given by

⎡ ⎤ ⎡ ⎤⎡ ⎤
R 1.000000 0.000000 1.139883 Y 
⎣G ⎦ = ⎣ 1.000000 −0.394642 −0.580622⎦ ⎣U  ⎦ . (8.30)
B 1.000000 2.032062 0.000000 V 

8.4.2 NTSC (1953) Y I Q

The color coding and transmission primaries for the American National Tele-
vision Systems Committee (NTSC) standard are described in this section. The
transmission primaries use a luminance-chrominance encoding, which was intro-
duced in the early 1950s [698].
The color coding is no longer in wide use; it has been superseded by the
SMPTE-C standard which is discussed in the following section. The RGB pri-
maries for NTSC are listed in Table 8.2 under NTSC (1953). The corresponding
conversion matrix and non-linear transfer function are listed in Table 8.3 as NTSC
(1953). The NTSC standard uses Y I Q transmission primaries, where the letters
I and Q stand for in-phase quadrature. The transmission primaries are constructed
by differencing B and R from Y as follows:

I  = −0.27(B −Y  ) + 0.74(R −Y  ), (8.31a)

    
Q = 0.41(B −Y ) + 0.48(R −Y ). (8.31b)

i i

i i
 i i

i i

434 8. Color Spaces

Figure 8.5. Visualization of the Y I Q color channels. The original image is shown in the
top-left panel. The Y channel is shown in the top right and the bottom two images show
the Y +I channels (left) and Y +Q channels (right); Zaragoza, Spain, November 2006.

i i

i i
 i i

i i

8.4. Television and Video 435

Y = 1.00 Y = 0.89 Y = 0.70 Y = 0.59 Y = 0.41 Y = 0.30 Y = 0.11

I = 0.00 I = -0.31 I = -0.21 I = -0.52 I = 0.52 I = 0.21 I = 0.31
Q = 0.00 Q = 0.32 Q = -0.60 Q = -0.28 Q = 0.28 Q = 0.60 Q = -0.32

White Yellow Cyan Green Magenta Red Blue

Y = 1.00 Y = 0.89 Y = 0.70 Y = 0.59 Y = 0.41 Y = 0.30 Y = 0.11

Figure 8.6. Standard NTSC color bars along with their corresponding luminance bars.

The I  is the orange-blue axis while Q is the purple-green axis (see Figure 8.5
for a visualization). The reason for this choice is to ensure backward compatibil-
ity (the Y  channel alone could be used by black and white displays and receivers)
and to utilize peculiarities in human perception to transmit the most visually pleas-
ing signal at the lowest bandwidth cost. Human visual perception is much more
sensitive to changes in luminance detail than to changes in color detail. Thus,
the majority of bandwidth is allocated to encoding luminance information and the
rest is used to encode both chrominance information and audio data. The color
coding is given by
⎡ ⎤ ⎡ ⎤⎡ ⎤
Y 0.299 0.587 0.114 R
⎣ I  ⎦ = ⎣ 0.596 −0.275 −0.321⎦ ⎣G ⎦ , (8.32)
Q 0.212 −0.523 0.311 B
and its inverse transform is
⎡ ⎤ ⎡ ⎤⎡ ⎤
R 1.000 0.956 0.621 Y 
⎣G ⎦ = ⎣ 1.000 −0.272 −0.647⎦ ⎣ I  ⎦ . (8.33)
B 1.000 −1.107 1.704 Q
The standard NTSC color bars are shown alongside their luminance bars in
Figure 8.6.
For practical purposes, the transform between EBU and NTSC transmission
primaries is possible because their RGB primaries are relatively similar. The
transform is given by
    
I −0.547 0.843 U 
= ; (8.34a)
Q 0.831 0.547 V 
    
U −0.547 0.843 I 
= . (8.34b)
V 0.831 0.547 Q

i i

i i
 i i

i i

436 8. Color Spaces

Note that the reverse transform uses the same matrix. The conversion between
NTSC and EBU primaries and back is given by
⎡ ⎤ ⎡ ⎤⎡ ⎤
RNTSC 0.6984 0.2388 0.0319 REBU
⎣GNTSC ⎦ = ⎣ 0.0193 1.0727 0.0596⎦ ⎣GEBU ⎦ ; (8.35a)
BNTSC 0.0169 0.0525 0.8459 BEBU
⎡ ⎤ ⎡ ⎤⎡ ⎤
REBU 1.4425 −0.3173 −0.0769 RNTSC
⎣GEBU ⎦ = ⎣−0.0275 0.9350 0.0670⎦ ⎣GNTSC ⎦ . (8.35b)
BEBU −0.0272 −0.0518 1.1081 BNTSC

8.4.3 SMPTE-C RGB

Television in North America currently uses the SMPTE-C standard. Its RGB
primaries, transformation matrices, and transfer functions are listed in Tables 8.2
and 8.3. The conversion matrix to Y I Q is the same as for the NTSC (1953)
standard discussed in the preceding section. As a result, the conversion between
EBU and SMPTE-C transmission components is given in (8.34). The conversion
between SMPTE-C and EBU RGB signals is as follows:
⎡ ⎤ ⎡ ⎤⎡ ⎤
RSMPTE-C 1.1123 −0.1024 −0.0099 REBU
⎣GSMPTE-C ⎦ = ⎣−0.0205 1.0370 −0.0165⎦ ⎣GEBU ⎦ ; (8.36a)
BSMPTE-C 0.0017 0.0161 0.9822 BEBU
⎡ ⎤ ⎡ ⎤⎡ ⎤
REBU 0.9007 0.0888 0.0105 RSMPTE-C
⎣GEBU ⎦ = ⎣ 0.0178 0.9658 0.0164⎦ ⎣GSMPTE-C ⎦ . (8.36b)
BEBU −0.0019 −0.0160 1.0178 BSMPTE-C

8.4.4 ITU-R BT.601 Y PB PR (SDTV)

The ITU-R BT.601 standard defines the transform from R G B to transmission
primaries. It does not define the primaries, white point, or transfer function of the
R G B space. This standard is used for digital coding of SDTV signals, and the
conversion is as follows:
⎡ ⎤ ⎡ ⎤⎡ ⎤
Y 0.299 0.587 0.114 R
⎣PB ⎦ = ⎣−0.169 −0.331 0.500⎦ ⎣G ⎦ ; (8.37a)
PR 0.500 −0.419 −0.089 B
⎡ ⎤ ⎡ ⎤⎡ ⎤
R 1.000 0.000 1.403 Y 
⎣G ⎦ = ⎣ 1.000 −0.344 −0.714⎦ ⎣PB ⎦ . (8.37b)
B 1.000 1.772 0.000 PR

i i

i i
 i i

i i

8.4. Television and Video 437

8.4.5 ITU-R BT.709 Y CB CR (HDTV)

The ITU-R recommendation for HDTV signals uses the primaries listed under
HDTV in Table 8.2. Its associated matrices and non-linear transfer function are
given in Table 8.3. The transmission coefficients are computed by
⎡ ⎤ ⎡ ⎤⎡ ⎤
Y 0.2215 0.7154 0.0721 R
⎣CB ⎦ = ⎣−0.1145 −0.3855 0.5000⎦ ⎣G ⎦ ; (8.38a)
CR 0.5016 −0.4556 −0.0459 B
⎡ ⎤ ⎡ ⎤⎡  ⎤
R 1.0000 0.0000 1.5701 Y
⎣G ⎦ = ⎣ 1.0000 −0.1870 −0.4664⎦ ⎣CB ⎦ . (8.38b)
B 1.0000 1.8556 0.000 CR

Further, the conversion between EBU and Rec.709 RGB signals is

⎡ ⎤ ⎡ ⎤⎡ ⎤
R709 1.0440 −0.0440 0.0000 REBU
⎣G709 ⎦ = ⎣ 0.0000 1.0000 0.0000⎦ ⎣GEBU ⎦ ; (8.39a)
B709 0.0000 −0.0119 1.0119 BEBU
⎡ ⎤ ⎡ ⎤⎡ ⎤
REBU 0.9578 0.0422 0.0000 R709
⎣GEBU ⎦ = ⎣ 0.0000 1.0000 0.0000⎦ ⎣G709 ⎦ . (8.39b)
BEBU 0.0000 0.0118 0.9882 B709

The conversion between NTSC and Rec. 709 RGB is

⎡ ⎤ ⎡ ⎤⎡ ⎤
R709 1.5073 −0.3725 −0.0832 RNTSC
⎣G709 ⎦ = ⎣−0.0275 0.9350 0.0670⎦ ⎣GNTSC ⎦ ; (8.40a)
B709 −0.0272 −0.0401 1.1677 BNTSC
⎡ ⎤ ⎡ ⎤⎡ ⎤
RNTSC 0.6698 0.2678 0.0323 R709
⎣GNTSC ⎦ = ⎣ 0.0185 1.0742 −0.0603⎦ ⎣G709 ⎦ , (8.40b)
BNTSC 0.0162 0.0432 0.8551 B709

and finally between SMPTE-C and Rec.709 RGB,

⎡ ⎤ ⎡ ⎤⎡ ⎤
R709 0.9395 0.0502 0.0103 RSMPTE-C
⎣G709 ⎦ = ⎣ 0.0178 0.9658 0.0164⎦ ⎣GSMPTE-C ⎦ ; (8.41a)
B709 −0.0016 −0.0044 1.0060 BSMPTE-C
⎡ ⎤ ⎡ ⎤⎡ ⎤
RSMPTE-C 1.0654 −0.0554 −0.0010 R709
⎣GSMPTE-C ⎦ = ⎣−0.0196 1.0364 −0.0167⎦ ⎣G709 ⎦ . (8.41b)
BSMPTE-C 0.0016 0.0044 0.9940 B709

i i

i i
 i i

i i

438 8. Color Spaces

8.4.6 SECAM Y DB DR
The transmission encoding of SECAM is similar to the YIQ color space. Its
conversion from RGB is given by
⎡ ⎤ ⎡ ⎤⎡ ⎤
Y 0.299 0.587 0.114 R
⎣DB ⎦ = ⎣−0.450 −0.883 1.333⎦ ⎣G ⎦ ; (8.42a)
DR −1.333 1.116 0.217 B
⎡ ⎤ ⎡ ⎤⎡  ⎤
R 1.0000 0.0001 −0.5259 Y
⎣G ⎦ = ⎣ 1.0000 −0.1291 0.2679⎦ ⎣DB ⎦ . (8.42b)
B 1.0000 0.6647 −0.0001 DR

This color space is similar to the YUV color space, and the chromatic channels
DB and DR may therefore also be directly derived as follows:

DB = 3.059U, (8.43a)
DR = −1.169V. (8.43b)

The luma channels in both spaces are identical.

8.4.7 Kodak’s Photo Y CC

Kodak’s Photo Y CC color space is used in the PhotoCD system. This system is
intended to enable the display of imagery on both television, video, and computer
monitors. It uses the RGB primaries and non-linear transfer function as ITU-R
BT.709, which is listed under HDTV in Tables 8.2 and 8.3. Luma Y  and the two
chroma channels are encoded by
⎡ ⎤ ⎡ ⎤⎡ ⎤
Y 0.299 0.587 0.114 R
⎣C1 ⎦ = ⎣−0.299 −0.587 0.886⎦ ⎣G ⎦ . (8.44)
C2 0.701 −0.587 −0.114 B

The Kodak Y CC color space includes an explicit quantization into an 8-bit for-
mat. The Y CC channels are quantized using
⎡8  ⎤ ⎡ 255Y 
⎤
Y
⎣8C1 ⎦ = ⎢ 1.402
⎣111.40C
⎥
⎦. (8.45)
8C 1 + 156
2 135.64C + 137
2

To enable display of encoded imagery on various display devices, the inverse

transform from Y CC to RGB depends on the target display device. For display

i i

i i
 i i

i i

8.5. Hue-Saturation-Lightness Spaces 439

by computers on a cathode ray tube, the conversion is

⎡ ⎤ ⎡ ⎤
Y 1.3584 8Y 
⎣C1 ⎦ = ⎣2.2179 (8C1 − 156)⎦ . (8.46a)
C2 1.8215 (8C2 − 137)

Non-linear R G B is then reconstructed:

⎡ ⎤ ⎡ ⎤
R Y  +C2
1
⎣G ⎦ = ⎣Y  − 0.194C1 − 0.509C2 ⎦ . (8.47)
k
B Y  +C2

For computer monitors, the value of k equals 1. The range of values of R G B is

between 0 and 346. To bring this range to the more conventional range of [0, 255],
a look-up table can be used. For display of YCC encoded images on televisions,
we set k = 353.2. The resulting values are non-linear, which is fine if the display
has a gamma of 2.2. However, for different displays, further gamma correction
may have to be carried out.

8.5 Hue-Saturation-Lightness Spaces

As discussed in the introduction to this chapter, intuitive selection of colors re-
quires the design and utilization of an appropriate color space. Ideally, for such a
space, each axis should carry an easy to understand set of information. The hue,
saturation, lightness (HSL) spaces exist in several variations, where the lightness
channel is replaced with either intensity (HSI) or value channels (HSV). A variant
with a chroma channel is also known (HCI).
In reality, these spaces should carry along an asterisk or quotation marks to
differentiate them from spaces that are more closely tied to the actual color ap-
pearance attributes. They can be considered crude approximations of more visu-
ally precise color spaces. The definition of these color spaces are given in terms
of an abstract RGB color space that is assumed to be normalized. The RGB
color cube is rotated such that the diagonal between (0, 0, 0) and (1, 1, 1) forms
the lightness axis. Note that this inherently suggests that “lightness” is a linear
combination of R + G + B, without regard to the particular RGB color space used.
Perpendicular to this axis is a circular plane defining color. The distance of a point
in this plane to the lightness axis defines “saturation.” The angle between a given
defined point and the color of interest defines “hue,” and thus ranges between 0◦
and 360◦ .
The color spaces in this section share all their disadvantages with using an
assumed RGB. As the color spaces are defined with respect to RGB, HSL-type

i i

i i
 i i

i i

440 8. Color Spaces

color spaces are generally device-dependent. If these spaces are defined using a
specific set of RGB primaries, they could be considered to be device-independent,
though in practice this is rarely done. Despite the suggestive names given to the
axes, the hue, saturation, and lightness do not correlate well with the perceptual
attributes of hue, saturation, and lightness as defined in Chapter 11. In particular,
the lightness axis (or value, or brightness, or whatever other name is given to
this channel) is not linear in any perceived quantity. To compute lightness, one
should consider computing XYZ tristimulus values first and then converting to an
appropriate perceptually uniform color space, such as CIE L∗ a∗ b∗ .
Further, as the computation for hue is usually split into 60◦ segments, one may
expect discontinuities in color space at these hue angles. Computations in these
color spaces are complicated by the discontinuity that arises at the 0◦ / 360◦ hue
angle [922].
In summary, the HSL-related color spaces discussed in the following sections
could be used for selecting device-dependent colors, for instance in drawing pro-
grams. Nonetheless, it should be noted that the meaning of the axes does not
correlate well with perceptual attributes of the same names, and care must be
taken not to confuse these spaces with more advanced color appearance spaces.

8.5.1 HSI
Replacing the lightness axis with an intensity axis, which is computed somewhat
differently, we arrive at the HSI space. The minimum of the three R, G, and B
values is used as an intermediary variable:

vmin = min (R, G, B) . (8.48)

We also compute an intermediary parameter h that will be used to compute hue H:

⎛ ⎞
R−G
+ (R − B)
1 ⎜ ⎟
h= cos−1 ⎝  2 ⎠. (8.49)
360 (R − G)2 + (R − B)(G − B)

The color space is then defined as

R+G+B
I= , (8.50a)
3
vmin
S = 1− , (8.50b)
I
(8.50c)

i i

i i
 i i

i i

8.5. Hue-Saturation-Lightness Spaces 441

⎧
⎪ B G
⎪
⎪1−h if > ∧ S > 0,
⎪
⎨ I I
H= B G (8.50d)
⎪h if ≤ ∧ S > 0,
⎪
⎪ I I
⎪
⎩
undefined if S = 0.
The hue angle H is undefined for colors with S = 0. This stems from the fact that
in this case the color is monochromatic, and defining a value for hue would not
be meaningful.
The conversion from HSI to RGB begins by converting H to degrees:
H  = 360 H. (8.51)
An RGB triplet is then formed as follows:
⎧⎡
⎤
⎪ S cos(H  )
⎪⎢ 3
⎪
1
1 + cos(60−H  ) ⎥
⎪
⎪ ⎢ 1 − (B + R) ⎥
⎪
⎪
⎪ ⎣ ⎦ if 0 < H  ≤ 120,
⎪
⎪
⎪
⎪ 3 (1 − S)
1
⎪
⎪
⎪
⎪ ⎡ ⎤
⎡ ⎤ ⎪ ⎪
⎪ 1
(1 −
R ⎪
⎨⎢
3 S)
⎥
⎣G⎦ = ⎢ 1 1 + S cos(H  −120) ⎥ if 120 < H  ≤ 240, (8.52)
⎪ ⎣3 cos(180−H ) ⎦
B ⎪
⎪
⎪
⎪ 1 − (R + G)
⎪
⎪
⎪⎡
⎪ ⎤
⎪
⎪ 1 − (G + B)
⎪
⎪
⎪
⎪ ⎢ ⎥
⎪
⎪ ⎢ 1
(1 − S) ⎥ 
⎪
⎪
⎪ ⎣
3 ⎦ if 240 < H ≤ 360.
⎩ 1 1 + S cos(H  −240)
3 cos(300−H  )

It should be noted that some of the R, G, and B values depend on the prior com-
putation of other R, G, and B values in the same triplet. This implies that these
values must be computed sequentially, as they are interdependent.

8.5.2 HSV
The conversion between RGB and HSV begins by finding the minimum and max-
imum values of the RGB triplet:
vmin = min (R, G, B) , (8.53a)
vmax = max (R, G, B) . (8.53b)
Saturation S and value V are defined in terms of these values:
vmax − vmin
S= , (8.54a)
vmax
V = vmax . (8.54b)

i i

i i
 i i

i i

442 8. Color Spaces

Of course, when R = G = B we have a gray value, and thereby a color for which
the saturation is 0 since vmax = vmin . In that case, hue H is undefined. Otherwise,
hue is computed as an angle ranging between 0◦ and 360◦ :
⎧
⎪ G−B
⎪
⎪ 60 if R = vmax ,
⎪
⎪ v − vmin
⎪
⎪ 
max
⎪
⎨ 
B−R
H = 60 2 + if G = vmax , (8.55)
⎪
⎪ vmax − vmin
⎪
⎪  
⎪
⎪ R−G
⎪
⎪
⎩60 4 + if B = vmax .
vmax − vmin

If the hue angle is less than 0 after this computation, the angle should be brought
within range by adding 360◦ . The inverse transformation, from HSV back to
RGB, begins by dividing the hue angle by 60:
H
H = . (8.56)
60
The fractional part of H  is given by
) *
f = H − H . (8.57)

Intermediate variables a, b, and c are then computed as follows:

a = V (1 − S), (8.58a)
b = V (1 − S f ), (8.58b)
c = V (1 − S (1 − f )) . (8.58c)

Figure 8.7. Hue angles (right) computed for the image on the left; Zaragoza, Spain,
November 2006.

i i

i i
 i i

i i

8.5. Hue-Saturation-Lightness Spaces 443

Figure 8.8. Cardinal hue angles. From top to bottom, left to right: original image, fol-
lowed by images with hue set to 0◦ , 60◦ , 120◦ , 180◦ , 240◦ , and 300◦ ; Zaragoza, Spain,
November 2006.

i i

i i
 i i

i i

444 8. Color Spaces

The integer part of H  determines which of the four quadrants the hue angle
H lies in. It is used to select which values to assign to R, G, and B:
⎧+ ,T
⎪
⎪ V c a if H   = 0,
⎪
⎪
⎪
⎪ + ,T
⎪
⎪
⎪
⎪ b V a if H   = 1,
⎪
⎡ ⎤ ⎪ ⎪+
⎪
,T
R ⎨ a V c if H   = 2,
⎣G⎦ = + ,T (8.59a)
⎪
⎪ H   = 3,
B ⎪
⎪ a b V if
⎪
⎪ + ,
⎪
⎪ c a V T if H   = 4,
⎪
⎪
⎪
⎪
⎪
⎪ + ,
⎩ V a b T if H   = 5.

An example of the hue angles present in an image is shown in Figure 8.7

(right), which was created by converting an RGB image to HSV, then setting the
saturation and value components to 1 before converting back to RGB.
The hue angles are arranged in a hexcone, with cardinal hue angles located at
every 60 degrees between 0◦ and 360◦ . The cardinal hue angles are visualized in
Figure 8.8, where the saturation and value components are retained, but the hue
of each pixel is set to the six cardinal angles.

8.6 HVS Derived Color Spaces

As discussed in Chapter 4, the early part of the human visual system carries out
various color transforms, starting with the transduction of light at the photorecep-
tor level, and followed by the recombination into color opponent spaces at the
ganglion level. In this section, we discuss various color spaces that are based on
our knowledge of the human visual system. Color opponent spaces are presented
in Section 8.7.

8.6.1 LMS Cone Excitation Space

The response of the cones in the retina are modeled by the LMS color space,
with L, M and S standing for long, medium, and short wavelengths. Although
the CIE XYZ color matching functions are closely related to a linear transform of
the LMS cone signals, they are not exact. Schanda gives an excellent overview
between the relationship of the cone fundamentals to CIE XYZ color-matching
functions [1002]. Nevertheless, approximations can be used to convert between
CIE XYZ and LMS cone signals. One such approximation is the Hunt-Pointer-
Estevez cone fundamentals, used in many color appearance models. The matrices

i i

i i
 i i

i i

8.6. HVS Derived Color Spaces 445

von Kries
Relative response 1.0
L
0.8 M
S
0.6

0.4

0.2

0.0
400 450 500 550 600 650 700
Wavelength (nm)

Figure 8.9. Relative response functions for the LMS cone-excitation space (after [318]).

that convert between XYZ and LMS are given by

⎡ ⎤ ⎡ ⎤⎡ ⎤
L 0.3897 0.6890 −0.0787 X
⎣M ⎦ = ⎣−0.2298 1.1834 0.0464⎦ ⎣Y ⎦ ; (8.60a)
S 0.0000 0.0000 1.0000 Z

⎡ ⎤ ⎡ ⎤⎡ ⎤
X 1.9102 −1.1121 0.2019 L
⎣Y ⎦ = ⎣ 0.3710 0.6291 0.0000⎦ ⎣M ⎦ . (8.60b)
Z 0.0000 0.0000 1.0000 S

As the LMS cone space represents the response of the cones in the human visual
system, it is a useful starting place for computational models of human vision.
It is also a component in the CIECAM02 and iCAM color appearance models
(see Chapter 11). The relative response as a function of wavelength is plotted in
Figure 8.9.

8.6.2 Sharpened Color Spaces

An alternative cone-response domain has been developed for use specifically

in chromatic-adaptation transforms and color-constancy experiments, known as
sharpened cone responses [581, 685] (see Figure 8.10 and 8.11). Two such ex-
amples are the cone signals used in the Bradford chromatic adaptation transform
and, more recently, the CIECAT02 chromatic-adaptation transform. More details

i i

i i
 i i

i i

446 8. Color Spaces

Bradford
1.0

Relative response
R
0.8 G
B
0.6

0.4

0.2

0.0

-0.2
400 450 500 550 600 650 700
Wavelength (nm)

Figure 8.10. Relative response functions for the Bradford chromatic-adaptation transform
(after [318]).

CAT02
Relative response

1.0
R
0.8 G
B

0.6

0.4

0.2

0.0

-0.2
360 410 460 510 560 610 660 710 760
Wavelength (nm)

Figure 8.11. Relative response functions for the CAT02 chromatic-adaptation transform.

on these spaces can be found in Section 10.4.5:

⎡ ⎤
0.8951 0.2664 −0.1614
MBradford = ⎣−0.7502 1.7135 0.0367⎦ ; (8.61a)
0.0389 −0.0685 1.0296
⎡ ⎤
0.9870 −0.1471 0.1600
−1
MBradford = ⎣ 0.4323 0.5184 0.0493⎦ ; (8.61b)
−0.0085 0.0400 0.9685

i i

i i
 i i

i i

8.6. HVS Derived Color Spaces 447

⎡ ⎤
0.7328 0.4296 −0.1624
MCAT02 = ⎣−0.7036 1.6975 0.0061⎦ ; (8.62a)
0.0030 0.0136 0.9834
⎡ ⎤
1.0961 −0.2789 0.1827
−1
MCAT02 = ⎣ 0.4544 0.4735 0.0721⎦ . (8.62b)
−0.0096 −0.0057 1.0153

8.6.3 MBDKL Color Space

The multi-stage color model proposed by de Valois et al. (see Section 4.6) is based
on a color space developed to conveniently represent color signals as they are
processed by early human vision. This color space was developed by MacLeod,
Boynton, Derrington, Krauskopf, and Lennie [250, 722] and named accordingly
the MBDKL space.
The three axes in this color space are chosen to coincide with human visual
responses occurring in early human vision. In particular, the first axis corresponds
to a luminance signal that is derived from the LMS color space (see Section 8.6.1)
by summing the L and M channels. Here, it is assumed that the S cones do not
contribute to the luminance signal.
The second and third channels are chromatic, with the second channel modu-
lating color such that the excitation of S cones is unaffected. This may be termed
the constant S axis. Along this dimension, the L and M components co-vary such
that their sum remains constant. The third channel is chosen such that chromatic
changes to the signal do not change the excitation of L cones or the M cones.
Hence, this is the constant L + M axis. Changes along this channel will only af-
fect the excitation of the S cones. The point where the three axes intersect is the
white point.
This three-dimensional color space can be unfolded to lie in a plane for con-
venience of representation. This is achieved by converting to polar coordinates.
The two chromatic channels are represented by azimuth φ ∈ [0, 360), whereas lu-
minance is represented by elevation θ ∈ [−90, 90]. A shift from white towards
pinkish-red along the constant S axis is indicated with an azimuthal angle of 0◦
(see also Table 4.3 on page 237). The opposite direction, which realizes a shift
from white towards cyan, is indicated with an angle of 180◦ . Similarly, a shift to-
wards violet along the L + M axis is at an angle of 90◦ . The angle of 270◦ marks
a shift toward greenish-yellow.
The elevation angle encodes luminance changes. An increase in luminance
is given by positive elevation angles, whereas decreases are marked by negative
elevation angles.

i i

i i
 i i

i i

448 8. Color Spaces

The azimuthal angle φ is computed from LMS signals as follows:

 
−1 L−M
φ = tan . (8.63)
S − (L + M)
The luminance signal is computed by summing the contribution from L and M
cones. However, the elevation angle is defined for changes in luminance, i.e.,
contrasts. For a pair of LMS signals, indicated with subscripts 1 and 2, the eleva-
tion angle is given by

θ = sin−1 (L1 + M1 − L2 − M2 ) . (8.64)

For any cell in the human visual system that linearly combines cone signals, there
will be an azimuth and elevation pair (φ0 , θ0 ) defining a null plane. The cell will
not respond to any colors lying in this plane. Modulating colors along a direction
perpendicular to this plane will maximally excite this cell. For a given color
encoded by (φ , θ ), the response R of a cell in spikes per second is

R = K (sin(θ ) cos(θ0 ) − cos(θ ) sin(θ0 )) sin (φ − φ0 ) (8.65a)

= K sin (θ − θ0 ) sin (φ − φ0 ) . (8.65b)

Here, K is a suitably chosen scale factor. Plots showing R as function of az-

imuthal angle are shown in Section 4.6, where the MBDKL color space was used
to analyze properties of cells in the early human visual system.
The unfolding of a color space into a 2D plane essentially returns us to plots
that are in a sense equivalent to figures plotting variables directly against wave-
length. However, the significance of the MBDKL color space is that the azimuthal
angle does not directly correspond to wavelength, but instead is encoded such that
signals eliciting maximal responses in LGN cells are at angles of 0◦ , 90◦ , 180◦ ,
and 270◦ , respectively.

8.7 Color Opponent Spaces

Color opponent spaces are characterized by a channel representing an achromatic
signal, as well as two channels encoding color opponency. The two chromatic
channels generally represent an approximate red-green opponency and yellow-
blue opponency. In this section, we discuss the CIE 1976 L∗ a∗ b∗ , also called
CIELAB, and the CIE 1976 L∗ u∗ v∗ color spaces CIELUV. These spaces also in-
clude a non-linear compression to achieve better perceptual uniformity throughout
the color space. This perceptual uniformity, therefore, enables color differences

i i

i i
 i i

i i

8.7. Color Opponent Spaces 449

between pairs of color to be assessed by simply computing their Euclidean dis-

tance. Improved color difference metrics are discussed in Section 8.8. We also
discuss a color space IPT that was designed to minimize some of the perceptual
errors of CIELAB, specifically with regards to lines of constant hue. The last color
space discussed in this section is the Lαβ space, which was derived by applying
principal components analysis (PCA) to a representative set of natural images,
encoded first in LMS cone-response space. This space represents a generic or-
thogonal type of color representation, where each channel is decorrelated from
the others.
Both the CIE 1976 L∗ a∗ b∗ and the CIE 1976 L∗ u∗ v∗ color spaces were
standardized by the CIE when they recognized the need for a uniform color space
in which color differences could be calculated. Two competing proposals were
evaluated, and at the time both were deemed equally valid. As such, they were
both formally recommended by the CIE [93, 1097]. Over time the CIE 1976
L∗ a∗ b∗ space has been recognized to be more accurate, especially in the
printing, imaging, and materials industries. In part this is due to the more accu-
rate chromatic-adaptation type model inherent in the calculation (CIELAB uses a
normalization of XYZ, while CIELUV uses a subtraction of XYZ). Although it
has fallen out of favor, the CIE 1976 L∗ u∗ v∗ space has found use in the display
and television industries, mostly due to its accompanying chromaticity diagram
[1097].

8.7.1 CIE 1976 L∗ a∗ b∗

The input into the CIE 1976 L∗ a∗ b∗ model (CIELAB) are the stimulus (X,Y, Z)
tristimulus values, as well as the tristimulus values of a diffuse white reflecting
surface that is lit by a known illuminant, (Xn ,Yn , Zn ). The CIELAB equations
calculate values representing the lightness of a color L∗ , as well as two opponent
chromatic channels [577]:
⎡ ⎤
⎡ ∗⎤ ⎡ ⎤ f (X/Xn )
L 0 116 0 −16 ⎢
f (Y /Yn ) ⎥
⎥
⎣a∗ ⎦ = ⎣500 −500 0 0⎦ ⎢ (8.66)
∗
⎣ f (Z/Zn ) ⎦
b 0 200 −200 0
1

The function f is defined as

⎧√
⎨3r for r > 0.008856,
f (r) = 16 (8.67)
⎩7.787 r + for r ≤ 0.008856.
116

i i

i i
 i i

i i

450 8. Color Spaces

The representation of lightness L∗ is normalized between 0 and 100 for black

and white. The non-linear compression function is defined by a linear component
near black and a cube-root function above that. The lightness function is consid-
ered to be a perceptually uniform scale between black and white and is a function
only of input luminance, Y and Yn . The chromatic channels encode a value be-
tween reddish-magenta and green for a∗ and a value between yellow and blue for
b∗ . These channels typically fall in the -128 to 128 range, though they are not
explicitly constrained. All three of these channels can be considered to encode a
measure of the human visual response to color stimuli. We can consider CIELAB
to be a rudimentary color appearance space, as we will see in Chapter 11.
As can be seen from (8.66), the values of a∗ and b∗ depend on the relative
luminance Y . This means that increasing or decreasing the luminance of the color
will have an impact on the values of a∗ and b∗ . Although this is perceptually
accurate, this does indicate that it is not possible to use the two chromatic channels
as a chromaticity diagram.
The inverse transform between CIE L∗ a∗ b∗ and CIE XYZ can be computed
directly2 :
⎧ 
⎪
⎪ L∗ a∗ 16 3
⎨ + + if L∗ > 7.9996,
X = Xn 116  500 116  (8.68a)
⎪
⎪ 1 L∗ a∗
⎩ + if L∗ ≤ 7.9996,
7.787 116 500
⎧ 3
⎪
⎪
∗
⎨ L + 16 if L∗ > 7.9996,
Y = Yn 116 116 (8.68b)
⎪
⎪ 1 L∗
⎩ if L∗ ≤ 7.9996,
7.787 116
⎧ 
⎪
⎪ L∗ b∗ 16 3
⎨ − + if L∗ > 7.9996,
Z = Zn 116  200 116 (8.68c)
⎪
⎪ 1 L∗ b∗
⎩ − if L∗ ≤ 7.9996.
7.787 116 200

Within the CIE L∗ a∗ b∗ color space, which is approximately perceptually lin-

ear, the difference between two stimuli may be quantified with the following color
difference formula:
+ ,1/2
∗
ΔEab = (ΔL∗ )2 + (Δa∗ )2 + (Δb∗ )2 . (8.69)

The letter E stands for difference in sensation (in German, Empfindung) [560].
2 This inverse transform is approximate, but widely used in practice. It would only be inaccurate

for colors that are both very dark and saturated. Such colors are rare in nature.

i i

i i
 i i

i i

8.7. Color Opponent Spaces 451

Figure 8.12. An example of the use of color-difference metrics. Here the ΔEab ∗ metric is

applied to a pair of images with the top-left image encoded in a lossless file format (PPM),
whereas the top-right image was stored as a jpeg file. The bottom image shows the per
∗ color difference; Wick, UK, October 2006.
pixel ΔEab

An example of the use of color-difference metrics is given in Figure 8.12,

where a pair of images is shown with one image subjected to lossy jpeg com-
pression. The difference image encodes ΔEab ∗ as a set of gray levels, with lighter

shades indicating larger errors. This image shows qualitatively where in the im-
∗
age the largest errors occur. For quantitative evaluation of errors, the actual ΔEab
values should be examined.
It should be noted that this color-difference metric is only approximately lin-
ear with human visual perception. The reason is that the laboratory conditions
under which this difference metric was tested, namely by observing flat reflection
samples on a uniform background, has little bearing on typical imaging applica-
tions. Hence, this is a case where laboratory conditions do not extend to the real
world. Also, it should be stressed that the CIELAB color-difference equations
were not designed for predicting very large color differences. Further develop-
ments have led to improved color difference metrics, and these are outlined in
Section 8.8.

i i

i i
 i i

i i

452 8. Color Spaces

b* yellowness
Cab
hab

-a* greenness a* redness

-b* blueness

Figure 8.13. Hue and chroma is an alternative to specify color.

8.7.2 L∗ C∗ hab Color Space

Conversion from CIE L∗ a∗ b∗ to L∗ C∗ hab is given by:
∗
Lab = L∗ , (8.70a)

∗
Cab = a∗ 2 + b∗ 2 , (8.70b)
 ∗
−1 b
hab = tan . (8.70c)
a∗

Essentially the transform from CIELAB to L∗ C∗ hab space can be thought of as

a rectangular to polar conversion and should be considered just an alternate rep-
resentation. The C∗ value defines the chroma of a stimulus, whereas hab denotes
a hue angle. This transformation may be more intuitive to understand for users
wishing to describe a color, and it is used in the CMC and CIE ΔE ∗ color differ-
ence metrics, which are presented in Sections 8.8.1 and 8.8.2. Note that the hue
angle hab should be between 0◦ and 360◦ . If the result is not within this range,
then 360 should be added or subtracted to the angle. For a fixed lightness, colors
can be visualized along red-green and yellow-blue axes, or as hue-chroma pairs,
as shown in Figure 8.13.
The CIELAB color difference can be specified in L∗ C∗ hab color space by first
calculating a rectangular hue difference, ΔHab∗ :

+ ,1/2
∗
ΔHab = (ΔE ∗ )2 − (ΔLab
∗ 2 ∗ 2
) − (ΔCab ) , (8.71a)
+ ,1/2
∗
ΔEab = (ΔL∗ )2 + (ΔCab
∗ 2 ∗ 2
) + (ΔHab ) . (8.71b)

This general form of CIELAB is used in color-difference metrics described in

Section 8.8, where this basic formulation is expanded to include different weights
for the three terms on the right-hand side. These different weights create ellipses
of tolerances in L∗ C∗ hab space. Rather than using a∗ and b∗ to approximate visual

i i

i i
 i i

i i

8.7. Color Opponent Spaces 453

b* b*

Δb*

ΔHab ΔCab
Δa*
a* a*

Figure 8.14. Tolerances in (a∗ ,b∗ ) space yield axis-aligned error metrics, whereas in
CIELAB L∗ C∗ hab space, the error metric is aligned with the principal axis of the ellipsoids.

tolerances, the use of weighted values for lightness, hue, and chroma yields closer
approximations to these elliptical areas of equal tolerance [93]. This is visualized
in Figure 8.14. Thus, color-difference metrics using well-chosen weights on light-
ness, hue, and chroma differences are considered to give more accurate results
than the color-difference formulae associated with the CIELAB and CIELUV
color spaces.
The inverse transform from L∗ C∗ hab to CIE L∗ a∗ b∗ is given by

L∗ = L∗ , (8.72a)
∗ ∗
a = C cos (hab ) , (8.72b)
∗ ∗
b = C sin (hab ) . (8.72c)

A similar conversion is possible from CIE L∗ u∗ v∗ to L∗ C∗ huv . The equations

are given in (8.70), except a∗ is replaced with u∗ and b∗ is replaced with v∗ .

8.7.3 CIE 1976 L∗ u∗ v∗

Similar to the CIELAB model, the inputs to the CIELUV calculations are a stim-
ulus (X,Y, Z) as well as the tristimulus values of white reflecting surface that is
lit by a known illuminant, (Xn ,Yn , Zn ). The conversion from CIE 1931 tristimulus
values to CIE 1976 L∗ u∗ v∗ (abbreviated as CIELUV) is then given by
⎧  1/3
⎪
⎪ Y Y
⎨ 116 − 16 > 0.008856,
∗ Yn Yn
L = (8.73a)
⎪
⎪ Y Y
⎩903.3 ≤ 0.008856;
Yn Yn
u∗ = 13 L∗ (u − un ), (8.73b)
v∗ = 13 L∗ (v − vn ). (8.73c)

i i

i i
 i i

i i

454 8. Color Spaces

The primed quantities in the above equations are computed from (X,Y, Z) as
follows:
4X 4 Xn
u = , un = ; (8.74a)
X + 15Y + 3 Z Xn + 15Yn + 3 Zn
9Y 9Yn
v = , vn = . (8.74b)
X + 15Y + 3 Z Xn + 15Yn + 3 Zn
The inverse transform from CIE L∗ u∗ v∗ to CIE XYZ begins by computing the
luminance channel Y ,  ∗ 
Yn L + 16 3
Y= (8.75)
100 116
and three intermediary parameters,
u∗
a= + un , (8.76a)
13 L∗
v∗
b= + vn , (8.76b)
13 L∗
c = 3Y (5b − 3). (8.76c)
The Z coordinate is then
(a − 4) c − 15abY
Z= , (8.77)
12b
and the X coordinate is
c 
X =− + 3Z . (8.78)
b
The transformation to CIE L∗ u∗ v∗ creates a more or less uniform color space,
such that equal distances anywhere within this space encode equal perceived color
differences. It is therefore possible to measure the difference between two stimuli
(L1∗ , u∗1 , v∗1 ) and (L2∗ , u∗2 , v∗2 ) by encoding them in CIELUV space, and applying the
following color difference formula:
+ ,1/2
∗
ΔEuv = (ΔL∗ )2 + (Δu∗ )2 + (Δv∗ )2 , (8.79)

where ΔL∗ = L1∗ − L2∗ , etc.

In addition, u and v may be plotted on separate axes to form a chromaticity
diagram. Equal distances in this diagram represent approximately equal percep-
tual differences. For this reason, where possible we show CIE (u ,v ) chromaticity
diagrams rather than perceptually non-uniform CIE (x,y) chromaticity diagrams.
The reason for the existence of both CIELUV and CIELAB color spaces is
largely historical. Both color spaces are in use today, with CIELUV more com-
mon in the television and display industries and CIELAB in the printing and ma-
terials industries [1097].

i i

i i
 i i

i i

8.7. Color Opponent Spaces 455

8.7.4 IPT Color Space

With more psychophysical data becoming available, including Hung and Berns
[500] and Ebner and Fairchild [273], the limitations of CIELAB and CIELUV
have become apparent. In particular, the uniformity of hue can be improved.
A simple transform that better models the linearity of constant hue lines, while
predicting other appearance attributes as well as or better than CIELAB and
CIECAM97s (the latter is discussed in Chapter 11), is afforded by the IPT color
space [274]. The abbreviation stands for image-processing transform, while the
letters also carry meaning to indicate the three channels of this color space. The
I channel represents intensity, whereas the P and T channels encode color op-
ponency. The P channel, for protan, encodes an approximate red-green oppo-
nency, whereas the T channel, for tritan, encodes an approximate yellow-blue
opponency.
The IPT color space is defined, as usual, as a transform from CIE XYZ. The
conversion between XYZ and IPT begins by a matrix multiplication, taking the
XYZ tristimulus values, specified according to the CIEXYZ 1931 2◦ observer
model, to a variant of the LMS cone excitation space, which is close to the Hunt-
Pointer-Estevez cone primaries (which are discussed in Section 10.4.1):
⎡ ⎤ ⎡ ⎤⎡ ⎤
L 0.4002 0.7075 −0.0807 X
⎣M ⎦ = ⎣−0.2280 1.1500 0.0612⎦ ⎣Y ⎦ . (8.80)
S 0.0000 0.0000 0.9184 Z
Note that the transformation into cone signals has assumed a normalization into
CIE D65, rather than an equal energy stimulus. More details on this normalization
can be found in Chapter 10. The LMS cone signals are then subjected to a non-
linearity:

 L0.43 if L ≥ 0,
L = 0.43 (8.81a)
− (−L) if L < 0;

 M 0.43 if M ≥ 0,
M = (8.81b)
− (−M)0.43 if M < 0;

 S0.43 if S ≥ 0,
S = 0.43 (8.81c)
− (−S) if S < 0.

Although real cone signals would never have negative components, the linear
matrix transform used in IPT can generate negative signals for real colors. This
is why care must be taken to maintain the proper sign before applying the non-
linear compression. The IPT opponent space is then reached by a further matrix

i i

i i
 i i

i i

456 8. Color Spaces

transform:
⎡ ⎤ ⎡ ⎤⎡  ⎤
I 0.4000 0.4000 0.2000 L
⎣ P ⎦ = ⎣ 4.4550 −4.8510 0.3960⎦ ⎣M  ⎦ , (8.82)
T 0.8056 0.3572 −1.1628 S
where I ∈ [0, 1] and P, T ∈ [−1, 1] under the assumption that the input XYZ values
were normalized. If the IPT axes are scaled by (100, 150, 150), they become
roughly equivalent to those found in CIELAB. The inverse transform begins by
transforming IPT to non-linear cone excitation space:
⎡ ⎤ ⎡ ⎤⎡ ⎤
L 1.8502 −1.1383 0.2384 I
⎣M  ⎦ = ⎣ 0.3668 0.6439 −0.0107⎦ ⎣ P ⎦ . (8.83)
S 0.0000 0.0000 1.0889 T
Linearization is then achieved by

L 1/0.43 if L ≥ 0,
L= (8.84a)
− (−L )1/0.43 if L < 0;

M  1/0.43 if M  ≥ 0,
M= (8.84b)
− (−M  )1/0.43 if M  < 0;

S 1/0.43 if S ≥ 0,
S= (8.84c)
− (−S )1/0.43 if S < 0,
and the transform to CIE XYZ is given by

⎡ ⎤ ⎡ ⎤⎡ ⎤
X 1.0000 0.0976 0.2052 L
⎣Y ⎦ = ⎣ 1.0000 −1.1139 0.1332⎦ ⎣M ⎦ . (8.85)
Z 1.0000 0.0326 −0.6769 S

The IPT color space bears similarities with CIELAB, although with different co-
efficients. It was designed specifically to have the strengths of CIELAB, but to
avoid the hue changes that could occur when compressing chroma along lines of
constant hue. It is a suitable color space for many color-imaging applications,
such as gamut mapping.

8.7.5 Lαβ Color Space

An interesting exercise is to determine what color space would emerge if an en-
semble of spectral images were converted to an LMS cone space and, subse-
quently, subjected to principal components analysis (PCA) [984]. The general

i i

i i
 i i

i i

8.7. Color Opponent Spaces 457

idea behind this approach is to determine if there is any link between the encod-
ing of signals in the human visual system and natural image statistics. A set of
natural images is used here, so that any averages computed are representative for
a canonical natural scene. The photoreceptor output is simulated by converting
the spectral images to the LMS color space.
The principal components analysis rotates the data so that the first princi-
pal component captures most of the variance. This can be understood by as-
suming that the data forms, on average, an elliptical point cloud. PCA rotates
this ellipse so that its major axis coincides with the axis defined by the first
principal component. Its second most important axis is aligned with the sec-
ond principal component. The remaining axis is aligned with the third principal
component. The principal components form an orthogonal coordinate system
and, thus, form a new color space. In this color space, the data is maximally
decorrelated.
By analyzing the resulting color space, it was found that this color space
closely resembles a color opponent space. The 3 × 3 matrix that transforms be-
tween LMS and the PCA-derived color space consists of elements that, aside from
a weight factor, recombine LMS signals in close-to-integer multiples. In addition,
it was found that the orientation of the axes is such that they can be assigned mean-
ing. The first principal component represents luminance, whereas the second and
third principal components represent yellow-blue and red-green color opponent
axes, respectively. For this reason, this color space is named Lαβ .
Additionally, it was found that the point cloud obtained becomes more sym-
metrical and well behaved if the color space is derived from logarithmic LMS
values.
The color opponency is shown in Figure 8.15, where the image is decom-
posed into its separate channels. The image representing the α channel has the β
channel reset to 0 and vice versa. We have retained the luminance variation here
for the purpose of visualization. The image showing the luminance channel only
was created by setting both the α and β channels to zero. The transform between
LMS and Lαβ is given by

⎡ ⎤
1
⎡ ⎤ ⎢ √3 0 0
⎥⎡ ⎤⎡ ⎤
L ⎢ ⎥ 1 1 1 log L
⎣α ⎦ = ⎢
1 ⎥⎣
⎢ 0 √ 0 ⎥ 1 1 −2⎦ ⎣log M ⎦ ; (8.86a)
⎢ 6 ⎥
β ⎣ 1 ⎦ 1 −1 0 log S
0 0 √
2

i i

i i
 i i

i i

458 8. Color Spaces

Figure 8.15. The top-left image is decomposed into the L channel of the Lαβ color space,
as well as L + α and L + β channels in the bottom-left and bottom-right images; Rochester,
NY, November 2004.

⎡√ ⎤
3
⎡ ⎤ ⎡ ⎤⎢ 0 0 ⎥⎡ ⎤
log L 1 1 1 ⎢ 3 √ ⎥ L
⎣log M ⎦ = ⎣1 ⎢ 6 ⎥⎣ ⎦
1 −1⎦ ⎢ 0 0 ⎥ α . (8.86b)
log S 1 −2 0 ⎣
⎢ 6 √ ⎥ ⎦ β
2
0 0
2

i i

i i
 i i

i i

8.8. Color Difference Metrics 459

The diagonal matrix contains weight factors that scale the relative contribution of
each of the three axes.
The theoretical importance of this color space lies in the fact that decorrelation
of natural images, represented in LMS color space, yields a color opponent space
which is very close to the one thought to be computed by the human visual system.
In other words, for natural images, the human visual system appears to decorrelate
the signal it transmits to the brain.
In practice, this color space can be used whenever natural images need to be
processed. Although PCA only decorrelates the data, in practice, the axis tend to
be close to independent. This means that complicated 3D color transforms on im-
age data can be replaced by three simpler 1D transforms. An example application
is discussed in Section 8.10.

8.8 Color Difference Metrics

Both the CIE L∗ a∗ b∗ and CIE L∗ u∗ v∗ are intended to be perceptually uniform
color spaces, such that the Euclidean distance between two colors is a good esti-
mation of their perceived difference. However, over time it has become clear that
neither space is adequately uniform, and therefore the need for improved color
difference metrics arose.
The Colour Measurement Committee of the Society of Dyers and Colourists
developed the CMC color difference metric in 1984. The CIE also refined its
color difference metrics, resulting in CIE 1994 ΔE∗ and later CIEDE2000.

8.8.1 CMC (l:c)

The CMC color difference metric [198, 765] is an elaboration of the CIE L∗ a∗ b∗
color difference formula. It uses a finer resolution for desaturated colors to ac-
count for the fact that small errors in the white balance of images are easily no-
ticed [507]. The reason human vision is sensitive to small differences for near-
achromatic signals is that a true achromatic stimulus silences the red-green and
yellow-blue pathways.
CMC is derived from the CIE LCH color space, for which a transform was
given in Section 8.7.2. Measurement of color differences is assumed to occur
under D65 lighting, using the 10◦ standard observer function.
This color difference metric is parameterized by the ratio l : c, where l denotes
a scale factor for lightness, and c is a scale factor for chroma. To determine
the perceptability of the difference between two colors, the ratio is set to 1, i.e.,

i i

i i
 i i

i i

460 8. Color Spaces

l = c = 1. In some applications, the ratio 2 : 1 is used. The parameters l and c

appear as variables in the formulation below.
To compute a CMC color difference, we begin by calculating two intermedi-
ary variables, both derived from CIE LCH values:

(C̄∗ )4
f= , (8.87a)
(C̄∗ )4 + 1900


0.36 + 0.4 |cos(35 + H̄)| if H̄ ≤ 164◦ ∨ H̄ > 345◦ ,
t= (8.87b)
0.56 + 0.2 |cos(168 + H̄)| if 164◦ < H̄ ≤ 345◦ .

The values of SL , SC , and SH define an ellipsoid in LCH color space, as follows:

⎧ ∗
⎨ 0.040975 L for L̄∗ > 16,
SL = 1 + 0.01765 L̄∗ (8.88a)
⎩0.511 for L̄∗ ≤ 16,

0.0638 C̄∗
SC = + 0.638, (8.88b)
1 + 0.0131 C̄∗
SH = SC ( f t + 1 − f ). (8.88c)

In the above equations, L̄∗ = 0.5(L1∗ + L2∗ ), and C̄ and H̄ are computed similarly.
The color difference ΔECMC is then computed by
L1∗ − L2∗
ΔLCMC = , (8.89a)
l SL
C1∗ −C2∗
ΔCCMC = , (8.89b)
c SC
H1∗ − H2∗
ΔHCMC = , (8.89c)
SH

ΔECMC = ΔLCMC
2 + ΔCCMC
2 + ΔHCMC
2 . (8.89d)

This color difference metric quantifies the perceived difference between two col-
ors, indicated with L1C1 H1 and L2C2 H2 .

8.8.2 CIE 1994 (ΔL∗ , ΔCab

∗ , ΔH ∗ )
ab
The CIE 1994 color difference formula [29, 90, 1249] is considered a further de-
velopment of the CIE L∗ a∗ b∗ color difference formula. It has in common with

i i

i i
 i i

i i

8.8. Color Difference Metrics 461

the CMC color difference metric that the relative contributions of lightness, hue,
and chroma are divided by functions of chroma to account for the relatively small
color differences that humans can distinguish for near-achromatic colors [507]. In
addition, the CIE 1994 color difference metric also specifies a set of experimental
conditions under which this formula is valid.
The reference conditions include D65 illumination of two homogeneous
patches of color placed side by side in direct contact, and each subtending at
least 4◦ of visual field. The illuminance is rated at 1000 lux. The observer should
have normal color vision.
Like CMC, the CIE 1994 color difference formula derives from the CIE LCH
color space and is also parameterized by a set of weights. They are kL , kC , and
kH , which are normally each set to unity. However, in the textile industry, a value
of 2 is used for kL . First, values for SL , SC , and SH are computed:

SL = 1, (8.90a)

SC = 1 + 0.045 C1∗ C2∗ , (8.90b)

SH = 1 + 0.015 C1∗ C2∗ . (8.90c)

These values are used to compute differences along each of the dimension L, C,
and H:
L1∗ − L2∗
ΔL94 = , (8.91a)
kL SL
C∗ −C2∗
ΔC94 = 1 , (8.91b)
kC SC
H ∗ − H2∗
ΔH94 = 1 . (8.91c)
kH SH
∗ (abbreviated as CIE94) is then given by
The color difference ΔE94

∗
ΔE94 = ΔL94 2 + ΔC 2 + ΔH 2 .
94 94 (8.92)

Finally, if the weights kL , kC , or kH differ from unity, then their values should be
included in the notation, i.e., CIE94 (2 : 1 : 1) or ΔE94∗ (2 : 1 : 1).

The interpretation of values obtained with this metric is that a ΔE94 ∗ value of 1

corresponds to the detection threshold of human vision. A value of 2 will be

discernible for patches of color that are next to each other. Values larger than 5
will be easy to see in side-by-side image comparisons.
While this color difference metric is preferred over ΔEab ∗ , it still does not corre-

late well with perceived color difference for saturated blue colors and near neutral
colors [273, 500, 666, 714].

i i

i i
 i i

i i

462 8. Color Spaces

8.8.3 CIEDE2000 Color Difference Metric

The CIEDE2000 color difference metric derives directly from the CIE L∗ a∗ b∗
color space. It was developed to improve the predictive power of the CIE94 metric
in the saturated blue and near-neutral regions. The near-neutral color prediction is
improved by scaling the a∗ axis differently. The hue and chroma differences are
weighted differently to aid in the prediction of blue color differences.
Given two colors specified in this space, we carry out the following calcula-
∗ :
tions for each color. First, for both samples, we compute Cab


∗
Cab = (a∗ )2 + (b∗ )2 , (8.93)

∗ values, which we denote with C̄ ∗ .

as well as the arithmetic mean of the two Cab ab
A second intermediary value g is computed:

⎛-
.  ∗ 7 ⎞
. C̄
g = 0.5 ⎝1 − /  7ab ⎠. (8.94)
∗
C̄ab + 257

With this value, we compute L , a , b , C , and h for each of the two colors.

L  = L∗ , (8.95a)
 ∗
a = (1 + g) a , (8.95b)
b = b ∗ , (8.95c)

C = (a )2 + (b )2 , (8.95d)
 
 180 −1 b
h = tan . (8.95e)
π a

Note that h and values derived from it are specified in degrees rather than in
radians. The arithmetic means of the pairs of L , C , and h values are also com-
puted. These are denoted by L̄ , C̄ , and h̄ . Checks should be made to ensure that
hue angles remain positive. This also has implications for the computation of the
arithmetic mean of the pair of h values. Here, it is important to ensure that the
hue difference remains below 180◦ . If a larger hue difference is found, then 360◦
is subtracted from the larger of the two hue angles, and the arithmetic mean is
recomputed. Before computing SL , SC , and SH , several intermediary values are

i i

i i
 i i

i i

8.8. Color Difference Metrics 463

calculated:

(C̄ )7
RC = 2 , (8.96a)
(C̄ )7 + 257
   2 
h̄ − 275
RT = −RC sin 60 exp − , (8.96b)
25
   
T = 1 − 0.17 cos h̄ − 30 + 0.24 cos 2h̄ (8.96c)
   
+ 0.32 cos 3h̄ + 6 − 0.20 cos 4h̄ − 63 , (8.96d)

followed by

0.015 (L̄ − 50)

2
SL = 1 +  , (8.97a)
2
20 + (L̄ − 50)
SC = 1 + 0.045 C̄ , (8.97b)

SH = 1 + 0.015 C̄ T. (8.97c)

The CIEDE2000 color difference metric, denoted with ΔECIE00 , is computed by

L1∗ − L2∗
ΔLCIE00 = , (8.98a)
kL SL
C1∗ −C2∗
ΔCCIE00 = , (8.98b)
kC SC

 
h1 − h2   
2 sin C1 C2
2
ΔHCIE00 = , (8.98c)
k H SH

ΔECIE00 = ΔLCIE00
2 + ΔCCIE00
2 + ΔHCIE00
2 + RT ΔCCIE00 ΔHCIE00 . (8.98d)

An informal comparison between the color difference metrics discussed in this

section is given in Figure 8.16. The top-right image was created by applying
a sharpening operator, which has increased contrast in the mid-range of values
at the cost of a loss of detail in the dark and light regions. The color difference
metrics typically show where these changes have occurred, although the CIE ΔEab ∗

and ΔEuv∗ predict a much larger visible difference than the other metrics.

i i

i i
 i i

i i

464 8. Color Spaces

Figure 8.16. The top-right image is a sharpened version of the top-left image. They
are compared in the remaining images using the color difference metrics discussed in this
section; Castle Combe, UK, September 2006.

i i

i i
 i i

i i

8.9. Color Order Systems 465

8.9 Color Order Systems

Color specification can be achieved in three fundamentally different ways. First,
colors can be specified as an additive mixture. Given a set of primaries, we can
then specify how much of each primary is added together to form the color. Ex-
amples are the RGB color spaces, as well as HLS-related spaces. Second, colors
may be specified in systems derived from color-matching experiments. These in-
clude XYZ and its derived color spaces such as CIELAB and CIELUV. It should
be noted that both the XYZ system, and its derivatives CIELAB and CIELUV also
rely on additive mixtures of light, in the color-matching experiments themselves.
The third class of color specification methods is based directly on visual per-
ception. Here, the aim is to produce color scales which are perceptually uniform.
A color order system is defined as a conceptual system of organized color percep-
tion [93]. Such systems are separate from their physical incarnations, known as
emphcolor order system exemplifications. For instance, the well-known Munsell
color order system is published as an atlas of painted chips.
In color order systems, the objective is to subdivide color space into percep-
tually equal spacings. This can typically only be achieved under relatively con-
trolled conditions. This includes the viewing environment, its geometry, as well as
its illumination. The spectral distribution of the illuminant, as well as its strength,
are defined. In a physical implementation of a color order system, the samples
should remain constant over time, i.e., fading, yellowing, and changes with tem-
perature and/or humidity should be avoided. If a color atlas is used to specify
color, then in order to minimize the effects of metamerism, it is desirable that the
sample to be specified is made of the same material as the paint chips in the atlas.
There are several color order systems with associated exemplifications avail-
able, including the Munsell Color System, the Natural Color System, the OSA
Uniform Color Scales System, as well as several others. These are briefly intro-
duced in the following sections.

8.9.1 Munsell Color System

Developed as a teaching aid for art students, the Munsell color system divides
color space into equally spaced segments along the dimensions of value (light-
ness), hue, and chroma, which are then called Munsell value (V), Munsell hue
(H), and Munsell chroma (C) [805, 808].
Rather than have the more conventional four primary hues of red, green, yel-
low, and blue, Munsell defined five principal hues by adding purple. This was
because the general gap between the neighboring hues of red and blue was con-
sidered to be perceptually too large. The hue of any given stimulus can thus be

i i

i i
 i i

i i

466 8. Color Spaces

designated by a single letter if it lies on one of the primaries, or a combination of

two primaries: red (R), yellow-red (YR), yellow (Y), green-yellow (GY), green
(G), blue-green (BG), blue (B), purple-blue (PB), purple (P), and red-purple (RP).
These colors follow the same progression as a rainbow and can be placed on a cir-
cle. Each of the ten hues and hue combinations can be further subdivided into
ten smaller steps, indicated with a number preceding the letter designation. For
instance, purple ranges from 1P to 10P.
Munsell values range between black (0) and white (10). In addition, achro-
matic colors can be indicated with the prefix N, i.e., N0 . . . N10. Hue and value
together form a cylindrical shape if hue values are located on an axis orthogonal
to the hue circle.
For each hue/value pair, the value of chroma represents how much the sample
deviates from achromatic. The distance from the hue axis determines the chroma
of a sample. The highest chroma value available is not uniformly distributed in
color space, but instead depends on both hue and value. This system therefore
leads to an asymmetrical shape in 3D space, known as the Munsell color solid.
In the Munsell notation system, a color is specified as H V/C. For instance, a
bluish-green color can be written as 10BG 6/4. The Munsell color atlas [806] is
now available as the Munsell book of color [807]

8.9.2 Natural Color System

Hering’s hypotheses on color opponency and the eye’s physiology leads to the
general theory of six primary colors, which are black, white, red, green, yel-
low, and blue [461]. These ideas were later transformed into a color order sys-
tem [554], as well as a color atlas [470]. After many years of improvements and
refinements, this resulted in the Natural Color System (NCS) and the Swedish
Standard Color Atlas [423–425, 1053, 1111].
In this system, the elementary colors are arranged in a hexagon, such that
possible combinations are all allowable. Recall that colors on opposite ends of a
color opponent channel cannot be perceived simultaneously (Section 5.3.2)
Beside the hexagon which encodes NCS hue, the two remaining perceptual
quantities in NCS are NCS blackness (or whiteness), and NCS chromaticness.

8.9.3 OSA Uniform Color Scales System

Both the Munsell color system and NCS are cylindrical in the sense that hues are
arranged in either a circle or a hexagon with the hue axis orthogonal, and Munsell
chroma or NCS chromaticness defined as the distance to the hue axis. This means
that the difference between two neighboring hues depends on the value for chroma

i i

i i
 i i

i i

8.10. Application: Color Transfer between Images 467

or chromaticness. For a small change in hue, the distance between two colors
of low chroma is much smaller than for colors with higher chroma. This just
suggests that steps of hue in these systems cannot be considered perceptually
uniform, though for any given value and chroma the steps should be close to
uniform.
The Optical Society of America (OSA) Uniform Color Scales System (OSA-
UCS) was designed to remedy this problem [718, 719, 836, 837]. It accomplished
this by structuring the color space according to a crystal lattice, shaped as a cuboc-
tahedron. The complexity of this system, however, requires a software implemen-
tation for it to be useful [93].

8.10 Application: Color Transfer between Images

In general, algorithms in computer graphics and computer vision may benefit
from the appropriate choice of color space. In this chapter we have shown that the
most common of all color spaces, namely RGB, is generally the least well defined.
There is no one RGB color space, and many “standards” have emerged. Each of
them have different primaries, a different white point, and a different non-linear
encoding.
On the other hand, the CIE XYZ color space is well defined, and it is capa-
ble of encompassing the full range of colors visible by humans in an all positive
manner. However, because the XYZ space also contains information that is well
outside the visible range, it is not the most efficient means of encoding and trans-
mitting information. If encoding and calculations are to be performed in the CIE
XYZ space, then it is important to utilize a higher bit-depth representation in order
to minimize quantization artifacts.
While many applications and algorithms continue to be defined in some RGB
color space, we argue that in most cases an appropriate choice of color space
can bring very significant benefits. As an example, in this section we outline an
algorithm that crucially hinges upon the appropriate choice of color space. The
task is the transfer of the color palette of one image to another. The question then
is how to capture the mood of an image and apply it to a second image.
The most obvious approach is to measure certain statistical properties of one
image and then modify a second image so that it obtains the same statistical prop-
erties. As color is represented with tristimulus values, and we have one such value
per pixel, the problem is essentially a complicated three-dimensional problem. If
the three color channels could be decorrelated first, the problem could be split into
three one-dimensional sub-problems, which would arguably be easier to solve.

i i

i i
 i i

i i

468 8. Color Spaces

As we are predominantly interested in images of environments we would en-

counter in the real world, i.e., natural images, we require a color space for which
the three color channels are decorrelated for natural images. Natural images have
various statistical regularities that can be exploited in image processing. The most
well-known is the 1/ f shape of the amplitude spectrum to which natural images
by and large adhere.3
A second, and very useful, statistic emerges when the correlation between
color channels is considered. Converting an ensemble of spectral images of nat-
ural scenes to LMS color space, we essentially end up with an image represented
in the same space as the photoreceptor’s output.
A standard technique to decorrelate data is called principal components anal-
ysis (PCA) [579,696]. This technique rotates the data such that the data is aligned
with the axes as well as possible. The first principle component is then the axis
which accounts for the largest variance in the data, whereas each subsequent axis
accounts for less and less variance. Applying PCA to an ensemble of images rep-
resented in LMS color space therefore rotates the data into a new color space in
which the three color axes are maximally decorrelated.
The surprising result that emerges is that when PCA is applied in this manner
to natural image ensembles, a new color space is found which strongly resembles
a color opponent space similar to the opponent space computed by the retinal
ganglion cells [984]. It is therefore apparent that the visual system decorrelates
the visual signal before sending it to the brain. In practice, the three color channels
are close to independent. This color space is the L α β color space discussed in
Section 8.7.5.
For 2000 random samples drawn from each of the images shown in Fig-
ure 8.17, their distribution is plotted in Figure 8.18. The point clouds form more
or less diagonal lines in RGB space when pairs of channels are plotted against
each other, showing that the three color channels in RGB space are almost com-
pletely correlated for these images. This is not the case for the same pixels plotted
in L α β space.
The practical implication for this is that by converting an image to the L α β
color space, the three channels are decorrelated and, in practice, can often be
treated as independent. The three-dimensional color transfer problem can there-
fore be recast into three independent one-dimensional problems [949].

3 Natural image statistics are normally collected over large sets of images, or image ensembles.

On average, the 1/ f statistic holds to a high degree of accuracy. However, for individual images this
statistic (and all other natural image statistics) may deviate from the ensemble average by an arbitrary
amount.

i i

i i
 i i

i i

8.10. Application: Color Transfer between Images 469

Figure 8.17. Examples images used to demonstrate the correlation between channels.
The first two images are reasonable examples of natural images, whereas the third image
is an example of an image taken in a built-up area. Built environments tend to have some-
what different natural image statistics compared with natural scenes [1295,1296]; Top left:
Mainau, Germany, July 2005. Top right: Dagstuhl, Germany, May 2006. Bottom: A
Coruña, Spain, August 2005.

To transfer the color palette of one image to another, both images are thus
first converted to the L α β space. Within this space, along each of the axes a
suitable statistic needs to be transferred. It turns out that a very simple matching
of means and standard deviations is frequently sufficient to create a plausible
result.
Naming the image from which its color palette is gleaned, the source image,
and the image to which this palette is applied, the target image, we use the s and
t subscripts to indicate values in source and target images. We first subtract the
mean pixel value in each of the axes:

Ls (x, y) = Ls (x, y) − L̄s , (8.99a)

αs (x, y) = αs (x, y) − ᾱs , (8.99b)
βs (x, y) = βs (x, y) − β̄s , (8.99c)

i i

i i
 i i

i i

470 8. Color Spaces

-2 1.6

Second channel
Red - Green l-α
-4 Red - Blue 1.2 l-β
Green - Blue α-β
-6
0.8
-8
0.4
-10
0.0
-12
-10 -9 -8 -7 -6 -5 -4 -3 -2 -16 -14 -12 -10 -8 -6 -4 -2 0 2
1.6
Second channel

0 Red - Green l-α

Red - Blue 1.2 l-β
-2 Green - Blue α-β
0.8
-4
0.4
-6
0.0
-8
-6 -5 -4 -3 -2 -1 0 1 -10 -8 -6 -4 -2 0 2
0 0.8
Second channel

Red - Green l-α

-1 Red - Blue 0.6 l-β
Green - Blue α-β
0.4
-2
0.2
-3
0.0
-4
-3.5 -3 -2.5 -2 -1.5 -1 -0.5 0 -6 -5 -4 -3 -2 -1 0 1
First channel First channel

Figure 8.18. Random samples plotted in RGB color space (left) and L α β color space
(right). The top to bottom order of the plots is the same as the order of the images in
Figure 8.17.

Lt (x, y) = Lt (x, y) − L̄t , (8.99d)

αt (x, y) = αt (x, y) − ᾱt , (8.99e)
βt (x, y) = βt (x, y) − β̄t . (8.99f)

β β
Second, the six standard deviations σsL , σsα , σs , σtL , σtα , and σt are computed.
It is now possible to make the means and standard deviations in the three target

i i

i i
 i i

i i

8.10. Application: Color Transfer between Images 471

Figure 8.19. Color transfer between images. The left and middle images serve as input,
and the right image was produced by matching means and standard deviations.

channels the same as those in the source channels:

σsL 
Lt (x, y) = L (x, y) + L̄s , (8.100a)
σtL t
σα
αt (x, y) = sα αt (x, y) + ᾱs , (8.100b)
σt
β
σs
βt (x, y) = β
βt (x, y) + β̄s . (8.100c)
σt

After conversion back to RGB, we obtain a displayable image. Examples of this

approach are shown in Figure 8.19.
Matching means and standard deviations implies that the point cloud in each
of the three dimensions is essentially Gaussian. This may be too simple a model.
It may therefore be possible to match higher-order moments such as skew and
kurtosis as well. Of course, the more high-order moments are taken into consid-
eration, the better the shape of the point cloud in the source and target images are
matched.
Alternatively, it would also be possible to match the histograms that can be
computed along each axis. Histogram matching begins by computing the his-
tograms of both the source and target channels. Next, the cumulative histograms
are computed. Cumulative histograms are monotonically non-decreasing func-
tions. The source histogram is then inverted. Pixels values in a particular channel
in the target image are then converted by first looking up their corresponding fre-

i i

i i
 i i

i i

472 8. Color Spaces

Figure 8.20. The top two images served as input to the color transfer algorithm. The
bottom-left image shows a result whereby means and standard deviations are matched,
whereas the bottom-right image shows a result obtained with histogram matching. Top-left
image: Turtle Mound, New Smyrna, FL, November 2004. Other images: Lake Konstanz,
Germany, July 2005.

quency in the cumulative target histogram. This frequency is then mapped to a

different pixel value by applying the inverted cumulative source histogram.
An example of this approach is shown in Figure 8.20. Although matching
of means and standard deviations will provide plausible results for many im-
age pairs, the example shown in this figure is not one of them. However, the
histogram-matching technique produces a result that is much closer to what one
might intuitively expect the result to look like.
There are several application areas for this simple technique. First, this
method provides an aesthetically pleasing way to enhance images. Second, it
has proven a useful method to create night-time images out of photographs taken
by daylight.
Third, the technique is important in augmented reality applications where real
environments are merged with computer-generated data. To make the computer-
generated elements mesh with the real scene, a fast algorithm is required to ac-
count for mismatches in the choice of surface materials as well as differences in
the lighting. While the color transfer algorithm cannot account for missing or
wrongly placed shadows or other light interactions between the real and virtual
components, the computer-generated elements can be adjusted to have the same

i i

i i
 i i

i i

8.10. Application: Color Transfer between Images 473

Figure 8.21. The University of Central Florida logo superimposed on a background [951].
The color transfer algorithm was applied to the image on the right and makes the logo blend
in with the background better; UCF Campus, Orlando, FL, June 2004.

Figure 8.22. The color transfer algorithm benefits from the appropriate choice of color
space. The top and left-most images are the input, whereas the bottom-middle image was
computed in RGB space. The bottom-right image was computed in the decorrelated L α β
space. Both transfers were computed using histogram matching; Nicosia, Cyprus, June
2006.

i i

i i
 i i

i i

474 8. Color Spaces

color statistics, which makes them fit in the scene better [951]. An example is
shown in Figure 8.21.
It is interesting to compare the results obtained in L α β color space with im-
ages created in different color spaces. In particular, when the results are created in
an RGB color space, it is expected that due to correlations between the axes, the
results will not be as good. That this is indeed the case is shown in Figure 8.22.
As mentioned above, the color transfer algorithm implicitly assumes that the
source and target images adhere to natural image statistics. Although image en-
sembles tend to show remarkable statistical regularities, individual images may
have different statistical properties. It is therefore not guaranteed that this algo-
rithm will work on any pair of images.
In particular, if the composition of the source and target images is very differ-
ent, the algorithm may fail. In general, transformations whereby the source image
has a fairly small color palette (such as night images, sunsets, and everything with
a limited number of different colors) tend to work well.
If the source and/or the target image is not a natural image in the statistical
sense, then the L α β color space is not an appropriate choice of color space. In
that case, it may be possible to explicitly decorrelate the data in the source and
target images separately, i.e., run PCA on both images [551]. It is then possible to
construct a linear transform between the source and target image. This transform
thus defines an image-specific color space, which may then be used to equalize
the three means and standard deviations.
This section has shown that the appropriate choice of color space is sometimes
crucial to achieve a particular goal. Most applications in computer vision and
computer graphics will benefit from the appropriate choice of color space. Color
should be an important consideration in the design of visual algorithms.

8.11 Application: Color-to-Gray Conversion

Conversion of color images to gray is often necessary, for instance in printing
applications. Starting with an image defined in a given three-dimensional color
space, grayscale conversion essentially constitutes a dimensionality reduction
from three dimensions to one. Given the salience of luminance in human visual
perception, a first approach would be to convert the color image to a color space
that has a luminance channel. It is then possible to linearly remap the luminance
values to the desired grayscale.
As an example, Figure 8.23 shows an image and its grayscale equivalent. The
grayscale image was created by converting the color image to Lαβ space (see
Section 8.7.5) and retaining only the luminance channel.

i i

i i
 i i

i i

8.11. Application: Color-to-Gray Conversion 475

Figure 8.23. Conversion from color to grayscale using the luminance channel of the Lαβ
color space; Castle Combe, UK, September 2006.

The use of the Lαβ color space is motivated by the fact that, for natural image
ensembles, this color space is obtained by employing PCA, as outlined in the
preceding section. As the luminance channel is the first principal component, this
particular grayscale conversion will, on average, retain the most information.
Although an appropriate choice on average, for a specific image this conver-
sion may yield a distinct loss of information. For instance, if the image contains
large variations in chromatic content, without exhibiting large variations in its
luminance channel, then the Lαβ color space will yield unconvincing results.
In such cases, we may apply PCA to each image individually, yielding the
best possible color space achievable with linear transformations only. By taking
only the first principal component and removing the second and third principal
components, a grayscale image is obtained.
While this approach may produce somewhat better results for certain images,
this is largely dependent on the distribution of points in color space. Principal
components analysis is only able to rotate the point cloud such that the direction
of greatest variance becomes axis-aligned. By projecting the data onto the first
principal component, we are not guaranteed a plausible grayscale conversion. For
instance, consider a dumbbell-shaped distribution of points such as in images of
green foliage with red fruit. The greens will be clustered in one sub-volume of
color space, and the reds will be clustered elsewhere. In the best case, PCA will
rotate the data such that all the greens are mapped to very light colors, and all the
reds are mapped to very dark colors. The mid-range of the grayscale will end up
being under-utilized.
A better approach would be to preserve the color difference between pairs of
pixels as much as possible. Here, it is important to start with a perceptually uni-
form color space, so that color differences have perceptual meaning. For instance,
the CIE94 color difference metric could be used. Assuming that a tristimulus
value c is mapped to grayscale by a function T (c), and the color difference met-

i i

i i
 i i

i i

476 8. Color Spaces

ric is written down as ci − c j , then the color difference between pairs of pixels
should remain constant before and after the mapping [938]:
ci − c j T (ci ) − T (c j )
= . (8.101)
Crange Trange
A measure of the error introduced by the mapping T is given by
 
ci − c j T (ci ) − T (c j ) 2
ε =∑ ∑
2
− . (8.102)
i j=i+1 Crange Trange

Here, Crange and Trange are the maximum color differences in the color and
grayscale images. Restricting the mapping T to linear transformations in the
CIELAB color space, the general form of T is

T (c) = (g · c, 0, 0). (8.103)

As two of the three dimensions are mapped to 0, the function T may be replaced
by the scalar function g · c, where g is the vector in color space that determines
the axis to which all color values should be projected to obtain a grayscale image.
Its direction will be determined shortly. The color difference metric for T is now
the absolute value of the difference of two gray values. The error function thus
simplifies to
 
ci − c j |g · (ci − c j ) | 2
ε =∑ ∑
2
− . (8.104)
i j=i+1 Crange Trange
A solution to the grayscale conversion is then found by minimizing this error
function, which is achieved with standard minimization techniques. For instance,
the Fletcher-Reeves conjugate gradient method can be used. The optimization
procedure can be seeded with an initial solution of g = (1, 0, 0). After conver-
gence, the vector g for which the error is minimized then determines the axis onto
which all colors are projected, leading to the desired grayscale image.
The power of this approach, however, is still limited by the fact that all pixel
values are projected to a single dimension. Many different colors may project to
the same gray value. A possible alternative approach can be derived by realizing
that different colors should not be projected to the same gray value if they are
neighboring. However, pixels separated from each other by differently colored
image regions may successfully be mapped to the same gray value. Such an ap-
proach would help overcome the problem of having a limited number of different
gray values to which all pixels must be mapped.
One such approach, called Color2Gray, begins by converting the image to the
CIE L∗ a∗ b∗ space [378]. For a pair of pixels, indicated with subscripts i and j, the

i i

i i
 i i

i i

8.11. Application: Color-to-Gray Conversion 477

conventional color-difference metric computes the Euclidean distance between

the two points. However, for the purpose of conversion to grayscale, it is better to
split this color-difference metric into two metrics, one measuring the luminance
difference and a second measuring the chromatic difference:

ΔLi∗j = Li∗ − L∗j , (8.105a)

∗ ∗ ∗
ΔC i j = (Δai j , Δbi j ) (8.105b)
= (a∗i − a∗j , b∗i − b∗j ). (8.105c)

This separation allows iso-luminant colors with different chromaticities to be dis-

tinguished and, subsequently, to be mapped to different gray values. In addition,
some user control is desirable, since, depending on the application, some colors
should be preferentially mapped to lighter gray values than others. Such creative
control would be difficult to automate.
For each pair of pixels, a desired difference in gray value δi j can now be
computed. As each pair of pixels generates such a desired difference, there may
not exist a grayscale image that satisfies each of these differences. Hence, these
target differences are input to an optimization problem, with objective function:

∑ ((gi − g j ) − δi j )2 . (8.106)
i, j

Here, gi is the gray value of pixel i which is found by minimizing this function.
It is initialized by setting g = Li . Once again, conjugate gradient methods are
suitable for finding a minimum.
The target differences δi j are computed under user control. Just as the a∗
and b∗ channels span a chromatic plane, the color differences in these channels,
Δa∗ and Δb∗ , span a chromatic color-difference plane. To determine which color
differences are mapped to increases in gray value, the user may specify an angle
Θ in this plane. The associate unit length vector is denoted with vΘ .
A second user parameter α steers a non-linearity cα (x) that is applied to the
chromatic color difference ΔCi j . This compressive function assigns relatively
more importance to small color differences than large ones:

x
cα (x) = α tanh . (8.107)
α
The target color difference for a pair of pixels (i, j) is then
⎧
⎪
⎨ΔLi j if |ΔLi∗j | > cα ( ΔCi∗j ),
∗
δi j = cα ( ΔCi j ) if ΔCi∗j · vΘ ≥ 0, (8.108)
⎪
⎩
cα ( − ΔCi∗j ) otherwise.

i i

i i
 i i

i i

478 8. Color Spaces

Figure 8.24. The red of this anturium stands out clearly in the color image. However,
conversion to gray using the luminance channel of Lαβ space makes the flower nearly
disappear against the background. The bottom-left image shows the result of applying
PCA directly to the color image. The bottom-right image shows the result obtained by
applying the spatial Color2Gray algorithm [378].

An example result is compared with a standard gray conversion in Figure 8.24.

The image shown here has a dumbbell-shaped distribution along the red-green
axis, which results in the red flower disappearing in the gray conversion for the
standard luminance-channel approach. Applying PCA directly to this image does
not improve the results notably, as shown in the same figure. However, the spa-
tially variant technique described above, is able to assign a different range of gray
levels to the red flower, making it distinctly visible in the gray-level image.
In summary, a good transformation to grayscale constitutes a dimensionality
reduction from three to one dimensions, whereby some perceptually significant
attribute is preserved. In particular, chromatic differences between pairs of iso-
luminant colors should be expressed in differences in gray level [939], especially
if these colors are located in spatial proximity.

8.12 Application: Rendering

The rendering algorithms described in Section 2.10 quietly assumed that all light-
ing and materials are spectral, i.e., defined per wavelength. For some applica-
tions, a spectral renderer will be invaluable. As an example, it would be difficult

i i

i i
 i i

i i

8.12. Application: Rendering 479

to render a material with a reflectance that is sharply peaked at a small number of

wavelengths when lit by, for example, a fluorescent light source, which itself has
a sharply peaked emission spectrum.
However, most applications do not require a fully spectral renderer. As the
human visual system under photopic conditions processes light in three different
channels, it is often not required to render images using more than three chan-
nels. Thus, the emission spectra of lights and reflectance spectra of materials are
represented by three values each. Each shading calculation is then carried out
independently for the three components.
The choice of color space to use for rendering images is usually not consid-
ered. Lights and materials are frequently represented with an RGB triplet without
concern for exactly which of the many RGB color spaces were intended. In the
best case, one of a few standard color spaces is used, such as sRGB. Such a choice
is motivated by the fact that the images produced by the renderer are already in
the same color space that is used by conventional display devices.
However, the sRGB color space has a rather narrow color gamut (see Sec-
tion 15.4). It is found that within a rendering application, the use of a wider gamut
produces images that are closer to a full spectral reference image [1208]. Wider
gamuts are available in the form of sharpened color spaces (see Section 8.6.2).
Given a spectral light source L(λ ) and a material ρ (λ ), a CIE XYZ tristimulus
value can be computed by integrating using standard color matching functions
x̄(λ ), ȳ(λ ), and z̄(λ ):

X= L(λ ) ρ (λ ) x̄(λ ) d λ , (8.109a)
λ
Y= L(λ ) ρ (λ ) ȳ(λ ) d λ , (8.109b)
λ
Z= L(λ ) ρ (λ ) z̄(λ ) d λ . (8.109c)
λ

The resulting XYZ values represent the material as seen under illuminant L. The
dominant illuminant available within the environment should be chosen for L.
This is possible, even if there are many light sources in an environment, as most
of these light sources typically have the same spectral emission spectra.
Conversion to a sharpened color space then involves multiplication by a 3 × 3
matrix:
⎡ ⎤ ⎡ ⎤
R X
⎣G ⎦ = Msharp ⎣Y ⎦ . (8.110)
B Z

i i

i i
 i i

i i

480 8. Color Spaces

A suitable conversion to a sharpened color space is given by the following matrix:

⎡ ⎤
1.2694 −0.0988 −0.1706
Msharp = ⎣−0.8364 1.8006 0.0357⎦ ; (8.111a)
0.0297 −0.0315 1.0018
⎡ ⎤
0.8156 0.0472 0.1372
−1
Msharp = ⎣ 0.3791 0.5769 0.0440⎦ . (8.111b)
−0.0123 0.0167 0.9955

After rendering in this color space, the resulting pixels need to be converted to
the space in which the display device operates. This is achieved by converting
each pixel first to XYZ and then to the desired display space. Assuming this is
the sRGB color space, the conversion is then
⎡ ⎤ ⎡ ⎤
Rd R
⎣Gd ⎦ = MsRGB M −1 ⎣G ⎦ . (8.112)
sharp
Bd B

Comparing renderings made in the sRGB and sharpened color spaces with
a direct spectral rendering, it is found that the difference between the reference
and the sharpened images remained at or below detectable levels. Measured us-
ing the CIE ΔE94∗ color difference metric (see Section 8.8.2), the average differ-

ence is 4.6 for the 98th percentile (with 5.0 visible in side-by-side image com-
parisons) [1208]. The sRGB color space does not perform as well, producing
differences with the reference image that are more easily perceived at 25.4.
Aside from the choice of color space, color-accurate rendering will require
white balancing, as spectral measurements of materials are typically made under
different illumination than they will be rendered. Second, the dominant illumi-
nant used in the rendering is usually different from the conditions under which
the resulting image is viewed. Both are sources of error for which correction is
needed. This topic is discussed in more detail in Section 10.7, where rendering is
revisited.

8.13 Application: Rendering and Color-

Matching Paints
Most image-synthesis applications render directly into an RGB space; usually
the space spanned by the primaries of the targeted output device. However, this
space is inadequate for some applications. One of them is simulating and ren-
dering paint. As shown in Section 3.4.6, pigmented paints cause scattering and

i i

i i
 i i

i i

8.13. Application: Rendering and Color-Matching Paints 481

absorption, which can only be properly accounted for by employing a sufficiently

accurate simulation. Using either RGB or CMY directly is not sufficient.
However, the Kubelka-Munk theory provides an alternative that may be em-
ployed in the computer simulation of paints. It is possible to derive the K(λ ) and
S(λ ) functions by measuring the reflectance function R(λ ) for pigments in their
masstone states (i.e., a pure pigment) as well as in tinted states (i.e., mixed with
white in a known concentration). Using the techniques described in Section 3.4.6,
reflectance functions for paints containing the same pigments but in different con-
centrations can be derived. Such spectral reflectance functions can be converted
to XYZ by integrating with the appropriate color-matching functions. The XYZ
triplets can in turn be converted to the color space used for rendering—for exam-
ple, the sharpened RGB space discussed in Section 8.12.
Thus, by adding absorption and scattering functions for different pigments to
derive a new spectral absorption and scattering function for the desired mixture,
a spectral reflectance function can be computed, which is then converted to the
preferred RGB rendering space. This approach leads to more accurate results
than by adding RGB triplets directly to simulate a paint mixture. Hence, the use
of Kubelka-Munk theory is recommended for simulating paint mixtures within
rendering applications. It can also be used in drawing programs to simulate, for
instance, airbrushing [412].
Finally, an important application lies in the conversion from RGB triplets to a
recipe for mixing pigments in given concentrations to mix a paint that is as close
as possible to the color defined by the RGB tristimulus value. This is the elec-
tronic equivalent of a procedure called color matching that should not be confused
with the color-matching experiments discussed in Chapter 7. Color matching for
paints refers to the matching of a paint mixture to a given sample of paint.
We assume that a database of spectral reflectance functions is available for a
set of pigments. As Kubelka-Munk theory can predict the spectral reflectance of
any mixture of these pigments, the conversion from a given spectral reflectance
function to a mixture of paints can be posed as an optimization problem.
One approach to achieve this, called spectrophotometric matching, is to con-
vert an RGB triplet to a spectral reflectance function, and then searching for a
concentration for each pigment, such that when mixed, the resulting paint matches
the target spectral reflectance function as well as possible [27, 81, 629, 874, 1198].
Algorithms to convert RGB triplets to spectral functions are discussed in Sec-
tion 7.6.3.
Alternatively, color matching may proceed in tristimulus space and is then
called tristimulus matching [27, 28, 345, 412, 629, 743]. The advantage of this
approach is that only the tristimulus values of the sample need to be known for a

i i

i i
 i i

i i

482 8. Color Spaces

paint to be matched. However, the number of different pigments that can be used
in the mixture is limited to three [412].
The approach described here follows Haase and Meyer’s technique and is
based on minimizing the color difference between a trial match and the paint
sample [412]. The CIE L∗ a∗ b∗ ΔE color difference, explained in Section 8.7.1, is
used for this application.
As the aim is to find concentrations for each pigment, a mixture where one
or more of the concentrations is negative, is not allowed. To avoid setting up an
optimization problem where negative concentrations are a possible outcome, the
problem is restated to find the square of concentrations for each pigment, labeled
qi = c2i for the ith pigment.
Initially, a random XYZ tristimulus value is guessed, and the ΔE color differ-
ence with the target sample is computed. To enable a gradient descent method to
be used, the gradient of ΔE 2 with respect to each of the squared concentrations qi
is computed:
∗ 2 ∗ 2
∂ ΔEab ∂ ΔEab
= 2 qi . (8.113)
∂ qi ∂ ci
By the chain rule, we have
∗ 2
∂ ΔEab ∗ 2
∂ ΔEab ∗ 2 ∗ 2
∂ X ∂ ΔEab ∂ Y ∂ ΔEab ∂Z
= + + . (8.114)
∂ ci ∂ X ∂ ci ∂ Y ∂ ci ∂ Z ∂ ci
The right-hand side of this equation can be expanded by expressing each of the
partials as follows:
∗ 2
∂ ΔEab 1000Δa∗
= , (8.115a)
∂X 1/3
3X 2/3 X0
∗ 2
∂ ΔEab −1000Δa∗ + 232ΔL∗ + 400Δb∗
= , (8.115b)
∂Y 1/3
3Y 2/3 X0
∗ 2
∂ ΔEab −400Δb∗
= . (8.115c)
∂Z 1/3
3Z 2/3 Z0
The remaining partials are expanded as follows:
∂X ∂ R(λ )
= k ∑ L(λ )x̄(λ ) , (8.116a)
∂ ci λ
∂ ci
∂Y ∂ R(λ )
= k ∑ L(λ )ȳ(λ ) , (8.116b)
∂ ci λ
∂ ci
∂Z ∂ R(λ )
= k ∑ L(λ )z̄(λ ) . (8.116c)
∂ ci λ
∂ ci

i i

i i
 i i

i i

8.13. Application: Rendering and Color-Matching Paints 483

Here, x̄(λ ), ȳ(λ ), and z̄(λ ) are standard color-matching functions, L(λ ) is the
emission spectrum of the light source lighting the sample, and k is a normalization
constant given by 100/ ∑λ L(λ )ȳ(λ ). The partial ∂ R(λ )/∂ ci is given by
⎛ ⎞
KM (λ )
⎜ +1 ⎟
∂ R(λ ) SM (λ )Ki (λ ) − KM (λ )Si (λ ) ⎜⎜ SM (λ ) ⎟
⎟.
= ⎜ 1−   ⎟
∂ ci S M (λ )
2
⎝ KM (λ ) 2 KM (λ ) ⎠
+2
SM (λ ) SM ( λ )
(8.117)
In this equation, KM and SM are the absorption and scattering functions for the
current trial, and Ki and Si are the absorption and scattering functions for the ith
pigment.
With the current match trial a gradient vector is computed; this is then added
to the current squares of the concentrations for each pigment qi . This leads to
new values for the concentrations ci , which in turn are used to compute a new
match trial. This process iterates until convergence is reached, i.e., all gradients
are below a specified threshold. The concentrations found for each pigment can
then be used to mix a paint that will look similar to the target XYZ tristimulus
value.
The match is valid under the chosen illuminant L. However, it may be benefi-
cial to minimize the effects of metamerism if the paint samples are to be viewed
under different illuminants. For instance, if it is assumed that a match is required
under CIE illuminant C, but metamerism needs to be minimized under CIE illumi-
nant A with the initial match being four times as important, then the minimization
function could be modified to

ΔE 2 = 4ΔEC2 + ΔEA2 . (8.118)

The matching algorithm outlined here has important benefits over standard tris-
timulus matching approaches that produce a match in XYZ space. By using the
uniform CIE L∗ a∗ b∗ color space, the match is found by minimizing the ΔE color-
difference metric that corresponds to how humans perceive color differences. Sec-
ond, an arbitrary number of pigments can be used in the match.
It would be possible to improve this method, though, as the steepest descent
method is not guaranteed to find a global minimum. Second, the method does not
take into account the fact that it may be both costly and inconvenient to mix as
many pigments as there are reflectance functions in the database used to compute
the match. As a result, it is likely that the color match involves small concentra-
tions of many pigments. It would be beneficial to devise a method that would find
the smallest number of pigments that would create a match.

i i

i i
 i i

i i

484 8. Color Spaces

8.14 Application: Classification of Edges

Image understanding refers to the process of obtaining high-level information
about the scene that is captured in images and videos. Answers to questions such
as What are the different objects in the scene? What actions are they performing?
Are there humans in the scene? are all provided by this process. The main purpose
of image understanding is to enable the user to make optimal decisions and take
appropriate actions based on information from the scene. Examples of systems
that use image understanding are surveillance systems and robots.
Image understanding often relies on lower-level information such as segmen-
tation, motion estimation, and object tracking. Algorithms for these tasks often
produce incorrect results due to the presence of shadows in the scene [217]. This
is because pixel values of the same object across edges created by cast shadows
are often very different and are, therefore, incorrectly recognized as belonging to
different objects of the scene.
Clearly, removal of shadows from the input images and videos would improve
the performance of these algorithms. Shadow removal may become easier if edges
cast by shadows (henceforth called shadow edges) are somehow distinguishable
from other edges, such as those cast by differences in reflectance properties of
adjacent surfaces in the scene (also called reflectance edges).
Color is an important feature that may be exploited to distinguish between
shadow and reflectance edges. Specifically, the choice of color space to represent
data has been shown to be important [585]. The obvious difference across shadow
edges is that regions under direct light appear brighter than the corresponding re-
gions in shadows. The regions in shadow get light from their surroundings [337].
This light is referred to as ambient light. The difference across shadow edges
is apparent only in the brightness of objects across shadow edges, if the spec-
tral composition of ambient light is the same as that of the direct light. Such a
phenomenon can occur in cases when the gray world assumption holds, that is,
the average reflectance of objects in the scene is gray [513]. In this case, the re-
flectance of objects will not alter the spectral composition of light reflecting off
of objects. Furthermore, there should be no participating media in the scene, as
this could again change the spectral composition of ambient light.
When the above assumptions hold, and the only difference across shadow
edges is in luminance, then an appropriate choice of color space can be employed
to differentiate between shadow edges and reflectance edges [217, 357, 990]. Any
color space that separately encodes luminance and chromatic information may be
used. Edges that appear only in the luminance channel and not in color channels
are more likely to be shadow edges, whereas edges appearing in both luminance

i i

i i
 i i

i i

8.14. Application: Classification of Edges 485

Figure 8.25. The image on the top left was captured in a scene where the gray world
assumption was assumed to hold approximately. The image on the right shows the infor-
mation captured by the luminance channel of the Yxy color space. The bottom two images
show the information encoded by the color channels (x and y, respectively). Notice how
most of the information across the shadow edge is contained in the luminance channel,
whereas all channels encode the change across reflectance edges.

and color channels are probably reflectance edges. Figure 8.25 illustrates this
difference between shadow and reflectance edges in scenes where the spectral
component of ambient light is the same as that of direct light.
It is hypothesized that the difference across shadow edges in sunlit scenes
will be in luminance as well as color, as the spectral component of ambient light
is possibly different from that of direct light for such scenes. Sunlight, which is
the direct light source, has a distinctly yellow spectrum, while the rest of the sky
appears blue due to Rayleigh scattering (as discussed in Section 2.8.1). Ambient
light therefore tends to have a bluish color. Figure 8.26 demonstrates this change
across shadow edges in outdoor scenes. There appears to be a change in color
as well as luminance across the edge. Choosing color spaces that separately en-
code luminance and color information is therefore not helpful in distinguishing
between shadow and reflectance edges in outdoor scenes.
Given that the changes in color across shadow edges are possibly in blue and
yellow (as the direct light is predominantly yellow and the ambient light is blue),
color opponent spaces may be better choices for the purpose of edge classifica-
tion than other color spaces, since these separately encode blue-yellow and red-
green information. Shadow edges will appear in the luminance and blue-yellow

i i

i i
 i i

i i

486 8. Color Spaces

Figure 8.26. Colors appear to have a bluer shade in shadows and a yellower shade under
direct light. The left image shows a color under direct sunlight, and the right image shows
the same color in shadow. Images were derived from the photograph shown in Figure 8.27.

channels, but not in the red-green channel, while all three channels will tend to
encode the difference across reflectance edges. The two edge types will therefore
often be distinguishable in outdoor scenes. The invariance to shadows in the red-
green channel was shown to have an evolutionary advantage in monkey foraging
behavior [1144].
The performance of 11 color spaces were tested for the purpose of shadow-
edge identification [585]. These include RGB, XYZ, Yxy, normalized rgb, LMS,
Luv, Lab, Lαβ , HSV, AC1 C2 , and the linear version of Lαβ (where the log of
LMS values is not taken before conversion to Lαβ ). When an image is converted
into a color space, one channel in the color space might encode the difference
across reflectance edges only, while another channel might encode the difference
across both shadow and reflectance edges. In that case, classification of edges
in the image may be done by first converting the image to the color space and
then using these two channels to distinguish between the two types of edges. The
most suitable color space for distinguishing between the edge types will have one
channel that does not distinguish between the edge types and one channel that
best discriminates between the edge types when compared with all the channels
in all the color spaces in the experiment.
Data for hundreds of edges of both types were collected and encoded by each
of the 11 color spaces. Edge data was generated by photographing a variety of dif-
fuse paint samples and then cropping these photographs to obtain uniform patches
of each paint sample under direct light as well as in shadow. Figure 8.27 shows
one of the photographs and the corresponding patches obtained from it. An edge
may be represented by a pair of such patches. For a shadow edge, one patch will
contain a color in direct light, and the second patch will contain the same color in
shadow. For a reflectance edge, the patches will contain different colors in direct
light; the patches might also contain different colors in shadow.
To compute the response of a color space to a shadow or reflectance edge,
the two patches representing the edge are converted into that color space. The
values of each patch are then averaged to obtain a single value representing color,

i i

i i
 i i

i i

8.14. Application: Classification of Edges 487

Figure 8.27. The image at the top shows a set of colored paint samples. The patches at the
bottom were cropped from this photograph. The top row shows patches from the shadow
region, and the bottom row shows corresponding patches from the lit region.

followed by an assessment of how different these values are for the two patches. If
a channel shows large differences in the average values of the patches representing
reflectance edges and small differences for patches representing shadow edges,
then that channel is a good classifier.
The discrimination of each channel is computed using receiver operating char-
acteristic (ROC) curves [773]. A ROC curve is a graphical representation of the
accuracy of a binary classifier. They are created by thresholding each channel’s re-
sponse to reflectance and shadow edges. The true positive fraction is the fraction
of correctly identified reflectance edges against the total number of reflectance
edges, and the true negative fraction is the same fraction for shadow edges. The
true positive fraction is plotted against the false positive fraction to give the ROC
curve. The area under the ROC curve (AUC) is the most commonly used quan-
titative index describing the curve. This area has been empirically shown to be

i i

i i
 i i

i i

488 8. Color Spaces

1.0 1.0 1.0

RGB LMS CIE XYZ
0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

R L X
0.2 G 0.2 M 0.2 Y
B S Z
0.0 0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

1.0 1.0 1.0

Yxy CIE Luv HSV
0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

Y L H
0.2 x 0.2 u 0.2 S
y v V
0.0 0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

1.0 1.0 1.0

Normalized RGB AC1 C2 CIE Lab
0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

r A L
0.2 g 0.2 C1 0.2 a
b C2 b
0.0 0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

1.0 1.0
Lαβ Linear Lαβ
0.8 0.8

0.6 0.6

0.4 0.4
L L
0.2 α 0.2 α
β β
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Figure 8.28. ROC curves for each channel in each of the 11 color spaces included in the
experiment. (Erum Arif Khan and Erik Reinhard, “Evaluation of Color Spaces for Edge
Classification in Outdoor Scenes,” IEEE International Conference on Image Processing,
Genova, Italy, pp 952–955, c 2005 IEEE.)

i i

i i
 i i

i i

8.14. Application: Classification of Edges 489

Channel AUC (outdoors) AUC (indoors)

rgb - r 0.435 0.598
rgb - g 0.477 0.573
rgb - b 0.649 0.567
XYZ - X 0.362 0.553
XYZ - Y 0.397 0.547
XYZ - Z 0.590 0.521
LMS - L 0.384 0.554
LMS - M 0.412 0.545
LMS - S 0.590 0.521
Yxy - Y 0.397 0.547
Yxy - x 0.795 0.902
Yxy - y 0.798 0.881
Luv - L 0.397 0.547
Luv - u 0.588 0.643
Luv - v 0.647 0.612
HSV - H 0.798 0.847
HSV - S 0.707 0.760
HSV - V 0.372 0.549
CIELAB - L 0.394 0.569
CIELAB - a 0.826 0.848
CIELAB - b 0.762 0.778
Lαβ - L 0.422 0.532
Lαβ - α 0.798 0.910
Lαβ - β 0.818 0.895
linear Lαβ - L 0.408 0.530
linear Lαβ - α 0.682 0.626
linear Lαβ - β 0.635 0.613
AC1 C2 - A 0.388 0.509
AC1 C2 - C1 0.627 0.653
AC1 C2 - C2 0.679 0.601

Table 8.5. Channels and their AUCs for outdoor and indoor scenes.

equivalent to the Wilcoxon statistic and computes the probability of correct clas-
sification [421]. An area of 1 signifies perfect classification while an area of 0.5
means that the classifier performs at chance.
A suitable color space should have one channel that has an AUC close to
0.5, and another channel with an AUC close to 1. The most suitable color space
is one for which the discriminating channel’s AUC is closest to 1. The ROC

i i

i i
 i i

i i

490 8. Color Spaces

curves computed for each channel for sunlit scenes are shown in Figure 8.28 and
the AUCs for each channel (for both indoor and outdoor scenes) are shown in
Table 8.5. Note that the ROC curve for normalized rgb shows the absence of any
channel that does not discriminate between the two edge types. It is therefore
removed from further consideration.
Four of the color spaces under consideration show color opponency. These
color spaces are CIELAB, Lαβ , linear Lαβ and AC1 C2 . While CIELAB is a
perceptually uniform space, and Lαβ is perceptually uniform to a first approxi-
mation, the latter two are linear color spaces. The performance of the perceptually
uniform color spaces is significantly better than their linear counterparts.
For indoor scenes however, where there is only a change in luminance across
shadow edges, and all that is required of a color space is separate encoding of
luminance and chromatic information, other color spaces tend to perform as well
as color opponent spaces. For outdoor scenes, color opponent spaces perform
well. The two perceptually uniform color opponent spaces prove to be the most
suitable choices. CIELAB and Lαβ may be used to get the best discrimination
between shadow and reflectance edges, followed by Yxy and HSV color spaces.
From this experiment, it is shown that color is an important cue in the classi-
fication of edges for sunlit scenes even when no assumptions are made about the
gray world and the effects of participating media. Experimental results show that
opponent color spaces are a good choice for the purpose of edge classification.

8.15 Further Reading

RGB color spaces are discussed in detail by Pascale [882], whereas transforms
between different spaces are presented in [327]. A full account of standard and
high definition television is presented by Charles Poynton [923]. The CIE colori-
metric systems are discussed in various places, including Hunt’s book on color
reproduction [509] and Wyszecki and Stiles [1262]. Rolf Kuehni has written an
entire text on the historical use of color spaces [631]. Bruce Lindbloom maintains
a website with tools and equations dealing with different color spaces [685].

i i

i i
 i i

i i

Chapter 9
Illuminants

The real world offers a wide range of illumination, which changes based on the
atmospheric conditions of the moment, the time of the day, and the season of the
year. Similarly, the range of illumination from man-made light sources is vast,
ranging from the yellowish light of a tungsten filament to the bluish-white of a
fluorescent light source.
The apparent color of an object is determined by the spectral power distribu-
tion of the light that illuminates the object and the spectral reflectance properties
of the object itself. Therefore, the wide range of illumination under which an
object may be viewed affects the perception of its color. This poses a significant
problem, especially in industrial colorimetry, where the main concern is to match
the color of a product to a given standard. The match achieved under one light
source may no longer hold when the product is displayed under a different light
source.
To alleviate this problem, the Commission Internationale de l’Éclairage (CIE)
defines a set of specific spectral power distributions called CIE Standard illumi-
nants. Using CIE illuminants in colorimetric computations ensures consistency
across measurements. It also simplifies the color match problem and makes it
independent of individual differences between light sources.

9.1 CIE Standard Illuminants and Sources

When we speak of an illuminant, we refer to a (relative) spectral power distribu-
tion, which may or may not be realizable by a light source. On the other hand,
the term source refers to a physical emitter of light such as the sun and the sky, or
a tungsten light bulb. As such, the definition of an illuminant precedes the con-
491

i i

i i
 i i

i i

492 9. Illuminants

struction of a light source that is aimed to represent that illuminant. The rationale
is that new developments in light-source technology are likely to produce better
light sources that more accurately represent any desired illuminant [194].
CIE illuminants represent general types of light sources commonly encoun-
tered in real life. For example, incandescent and fluorescent light sources and
different phases of daylight are all represented by CIE illuminants.

9.1.1 CIE Standard Illuminant A

CIE Standard Illuminant A represents typical domestic tungsten-filament light-
ing. Its spectral radiant power distribution is determined by light emitted from
a Planckian radiator (i.e., a blackbody or full radiator) operated at an abso-
lute temperature of 2856 K according to the International Practical Temperature
Scale [142].
The spectral radiant exitance Me,λ (T ) of a Planckian radiator at temperature
T (in Kelvin K) is approximated by
+
c ,−1
Me,λ (T ) = c1 λ −5 exp W m−3 ,
2
−1 (9.1)
λT

where the radiation constants c1 and c2 are equal to

c1 = 3.74150 × 10−16 Wm2 , (9.2a)

−2
c2 = 1.4388 × 10 mK. (9.2b)

The spectral power distribution of CIE illuminant A is then computed by insert-

ing T = 2856 K into (9.1). However, in most colorimetric measurements, only
relative spectral power distributions are necessary. As such, the spectral power
distribution SA (λ ) of CIE illuminant A (as is the case for other illuminants) is
typically normalized to 100 at a wavelength of 560 nm:

Me,λ (2856)
SA (λ ) = 100 . (9.3)
Me,560 (2856)

The spectral power distribution of CIE illuminant A is shown in Figure 9.1.

Correlated color temperature, tristimulus values, and chromaticity coordinates
of illuminant A are listed in Table 9.1 along with other CIE illuminants. The other
CIE illuminants listed in this table will be explained in the subsequent sections.
Spectral data for all CIE illuminants are given in the accompanying DVD-ROM.

i i

i i
 i i

i i

9.1. CIE Standard Illuminants and Sources 493

300
Relative Radiant Power
250

200

150

100

50

0
300 400 500 600 700 800 900
Wavelength (λ)

Figure 9.1. Relative spectral radiant power distribution of CIE standard illuminant A.

In practice, a light source approximating CIE illuminant A can be built by

using a gas-filled coiled tungsten filament lamp operating at a correlated color
temperature of 2856 K [1262]. Correlated color temperature as well as other
types of color temperature are explained in Section 9.2. For now, it suffices to say
that the correlated color temperature of a light source equals the temperature of
a Planckian radiator whose chromaticity coordinate is closest to that of the light
source.

Illuminant CCT (K) Tristimulus values Chromaticity coordinates

X Y Z x y u’ v’
A 2856 109.85 100.00 35.58 0.45 0.41 0.26 0.52
B 4874 99.10 100.00 85.32 0.35 0.35 0.21 0.49
C 6774 98.07 100.00 118.22 0.31 0.32 0.20 0.46
D50 5003 96.42 100.00 82.51 0.35 0.36 0.21 0.49
D55 5503 95.68 100.00 92.14 0.33 0.35 0.20 0.48
D65 6504 95.04 100.00 108.88 0.31 0.33 0.20 0.47
D75 7504 94.97 100.00 122.61 0.30 0.32 0.19 0.46
E 5400 100.00 100.00 100.00 0.33 0.33 0.21 0.47
F2 4230 99.20 100.00 67.40 0.37 0.38 0.22 0.50
F7 6500 95.05 100.00 108.72 0.31 0.33 0.20 0.47
F11 4000 100.96 100.00 64.37 0.38 0.38 0.23 0.50

Table 9.1. Correlated color temperature (CCT), tristimulus values, CIE 1931 xy and CIE
1976 u v chromaticity coordinates of commonly used CIE illuminants. Spectral data of all
CIE illuminants are given in the accompanying DVD-ROM.

i i

i i
 i i

i i

494 9. Illuminants

9.1.2 CIE Standard Illuminant D65

CIE standard illuminant D65 represents average daylight with a correlated color
temperature of approximately 6504 K. The reason to use a correlated color tem-
perature of 6504 K, instead of the more rounded value of 6500 K, is a result of
the slight adjustment of the radiation constant c2 in Planck’s formula (see (9.1)).
Originally, c2 = 1.4380 10−2 mK was accepted, but the value of c2 is changed
to 1.4388 10−2 mK in the International Practical Temperature Scale 1968. The
1.4380 ≈ 6504.
difference of 4 K follows from 6500 1.4388
Illuminant D65 is based on numerous spectro-radiometric measurements taken
in the United States, Canada, and Great Britain [559]. Although the spectral dis-
tribution of daylight on the earth’s surface is known to change as a function of
season, time of day, geographic location, weather conditions, air pollution, etc.,
the CIE recommends the use of D65 pending the availability of additional infor-
mation [193]. The relative spectral power distribution of CIE illuminant D65 is
shown in Figure 9.2 together with other CIE D-series illuminants. These illumi-
nants are explained in the next section.
As a result of the rather jagged spectral distribution of the CIE standard illu-
minant D65 , it has proven difficult to build a light source that faithfully represents
this illuminant. Consequently, at present the CIE has not recommended any stan-
dard light source to realize illuminant D65 . Despite this, daylight simulators suit-
able for psychophysics, can be constructed by placing a suitable filter in front of a

140
Relative Radiant Power

D50
120 D55
D65
D75
100

80

60

40

20

0
300 400 500 600 700 800 900
Wavelength (λ)

Figure 9.2. Relative spectral radiant power distributions of CIE D illuminants.

i i

i i
 i i

i i

9.1. CIE Standard Illuminants and Sources 495

Figure 9.3. The Gretag Macbeth D65 daylight simulator consists of a halogen light source
with a blue filter, shown here.

halogen light source. An example is the Gretag MacBeth D65 daylight simulator,
shown in Figures 2.50 and 3.45. The filter is shown in Figure 9.3.
The CIE Colorimetry Committee has developed a method that can be used
to assess the usefulness of existing light sources in representing CIE D illumi-
nants [184, 192]. The method is too elaborate to reproduce here in full. Nonethe-
less, the main steps of the method are:

1. Measure the spectral power distribution of the light source under consid-
eration by means of spectro-radiometry from λ = 300 to λ = 700 at 5-nm
intervals.

2. Compute the CIE 1976 (u10 , v10 ) chromaticity coordinates of the light
source from the measured spectral data. These coordinates must fall within
a circle of radius 0.015 centered at the chromaticity point of D65 on the
same diagram. Otherwise the tested source does not qualify as a daylight
simulator.

3. Even if the test in the previous step succeeds, the quality of the daylight
simulator needs to be determined. For this purpose two tests are performed:

(a) Compute the mean index of metamerism MIvis (see Section 9.4) of
five pairs of samples which are metamers under CIE illuminant D65 .
The spectral radiance factors of these samples are provided by the
CIE [1262]. These computations should use CIE 1964 (X10 ,Y10 , Z10 )

i i

i i
 i i

i i

496 9. Illuminants

CIEL∗ a∗ b∗ CIE L∗ u∗ v∗ Category

< 0.25 < 0.32 A
0.25 to 0.50 0.32 to 0.65 B
0.50 to 1.00 0.65 to 1.30 C
1.00 to 2.00 1.30 to 2.60 D
> 2.00 > 2.60 E

Table 9.2. Categories of daylight simulators (from [1262]). Using the method explained
in the text, MIvis and MIuv are computed for the test source. Depending on the color-
difference formula, use either the CIE L∗ a∗ b∗ or CIE L∗ u∗ v∗ values to find the corre-
sponding category for the MIvis and MIuv indices. The result is a two-letter grade, such as
BD or AC, indicating the quality of the daylight simulator in the visible and ultra-violet
regions of the spectrum.

tristimulus values. The resulting value indicates how good the meta-
merism of these five sample pairs are under the tested daylight
simulator.
(b) Compute the mean index of metamerism MIuv of three pairs of sam-
ples, where one member of each pair is non-fluorescent and the other
member is fluorescent. The spectral values of these pairs are provided
by the CIE [1262]. These pairs also yield identical tristimulus val-
ues under CIE illuminants D65 . All computations should once more
use CIE 1964 (X10 ,Y10 , Z10 ) tristimulus values. The resulting value
indicates how good the daylight simulator represents the standard il-
luminant D65 in the ultra-violet region of the spectrum.

4. Use MIvis and MIuv as an index to Table 9.2 to find the category under
which the daylight simulator falls. A category of AA indicates that the day-
light simulator has excellent performance both in the visible and ultraviolet
region of the spectrum.

9.1.3 Other CIE Illuminants D

For standardization purposes, the CIE recommends use of illuminant D65 for all
calculations that involve daylight. However, the CIE also defines a set of supple-
mentary illuminants called D50 , D55 , and D75 which may be used according to
the task at hand. These illuminants have correlated color temperatures of 5003 K,
5503 K, and 7504 K; therefore, their color ranges from yellowish-white to bluish-
white.
Different disciplines have adopted different types of CIE daylight illuminants.
For instance, D65 is commonly used by the paint, plastic, and textile industries,

i i

i i
 i i

i i

9.1. CIE Standard Illuminants and Sources 497

while the graphics art and photography communities favor D50 [93]. The relative
spectral radiant power distributions of these illuminants, together with the CIE
standard illuminant D65 , are shown in Figure 9.2.

9.1.4 CIE Method of Calculating D-Illuminants

The measurements by Judd et al. [559] revealed a simple relationship between the
correlated color temperature of a CIE D-series illuminant and its relative spectral
distribution. Denoting the correlated color temperature by Tc , the spectral radiant
distribution S(λ ) of a daylight illuminant can be calculated as follows:

1. Use Tc to find the first chromaticity coordinate xD . If Tc is between 4000 K

and 7000 K, then
−4.6070 × 109 2.9678 × 106 0.09911 × 103
xD = + + + 0.244063. (9.4)
Tc3 Tc2 Tc
If Tc is between 7000 K and 25000 K, then
−2.0064 × 109 1.9018 × 106 0.24748 × 103
xD = + + + 0.237040. (9.5)
Tc3 Tc2 Tc

2. Use xD to compute the second chromaticity coordinate yD :

yD = −3.000xD
2
+ 2.870xD − 0.275. (9.6)

3. Use xD and yD to compute the factors M1 and M2 :

−1.3515 − 1.7703xD + 5.9114yD
M1 = , (9.7)
0.0241 + 0.2562xD − 0.7341yD
0.0300 − 31.4424xD + 30.0717yD
M2 = . (9.8)
0.0241 + 0.2562xD − 0.7341yD

4. Using M1 , M2 , and the spectral distributions S0 (λ ), S1 (λ ), and S2 (λ ) (given

in Table F.1), compute the relative spectral distribution S (λ ):
S (λ ) = S0 (λ ) + M1 S1 (λ ) + M2 S2 (λ ). (9.9)

5. Normalize S (λ ) to 100 at wavelength 560 nm:

S (λ )
S(λ ) = 100 . (9.10)
S (560)

i i

i i
 i i

i i

498 9. Illuminants

9.1.5 CIE F Illuminants

Fluorescent lighting is commonly used owing to its high electrical efficacy and
its similarity to natural daylight. Therefore, the CIE originally defined 12 types
of fluorescent illuminants denoted by F1 to F12 that represent typical fluores-
cent lamps used in practice [186]. F-series illuminants are categorized into three
groups based on the emission spectra of the light sources they represent:
• F1–F6: standard fluorescent lamps;

• F7–F9: broadband fluorescent lamps;

• F10–F12: three-narrow-band fluorescent lamps.

From each of these groups, the CIE emphasizes one particular illuminant for use
in colorimetric work. These are the F2 cool white single phosphor lamp; the
F7, a daylight simulator; and the F11, which is a three-narrow-band fluorescent
lamp [194]. Of these three illuminants, the cool white F2 is the most similar
to typical fluorescent lights found in homes, offices, stores, and factories [735].
The spectral power distributions of all CIE F-series illuminants are given in the
accompanying DVD-ROM.

Group Illuminant Commercial Chromaticity Coordinates CCT (K) CRI

type x y
Standard F1 DAY 0.3131 0.3371 6430 76
F2 CWFa 0.3721 0.3751 4230 64
F3 WHITE 0.4091 0.3941 3450 57
F4 WWFb 0.4402 0.4031 2940 51
F5 DAY 0.3138 0.3452 6350 72
F6 LWFc 0.3779 0.3882 4152 59
Broad-band F7 D65 0.3129 0.3292 6500 90
F8 D50 0.3458 0.3586 5000 95
F9 CWXd 0.3741 0.3727 4150 90
Three-narrow-band F10 TL85 0.3458 0.3588 5000 81
F11 TL84 0.3805 0.3769 4000 83
F12 TL83 0.4370 0.4042 3000 83
a Cool white fluorescent
b Warm white fluorescent
c Lite white fluorescent
d Cool white deluxe
Table 9.3. CIE F illuminants, their commercial equivalents, and several important prop-
erties. CCT: Correlated color temperature, CRI: Color-rendering index (see Section 9.3);
(from [960]).

i i

i i
 i i

i i

9.1. CIE Standard Illuminants and Sources 499

400
Relative Radiant Power
350 D 65
F2
300 F7

250

200

150

100

50

0
300 400 500 600 700 800 900
Wavelength (λ)

Figure 9.4. CIE illuminants F2 and F7 and the CIE standard illuminant D65 .

Although these illuminants were derived from actual fluorescent lamps, their
commercial types were not disclosed. Recognizing that this may limit their us-
ability, this issue has been addressed by work of other researchers [960]. Table 9.3
lists F-series illuminants and their commercial types given by Rich [960].
In this table, several abbreviations are used to denote the commercial type of
each illuminant. For instance CWF, LWF, and WWF represent cool white, lite
white, and warm white fluorescent lamps, respectively, based on their correlated
color temperatures. CWX represents a cool white deluxe lamp which is con-
structed by supplementing a single phosphor lamp, such as CWF, with a second
phosphor to enhance the under-represented parts of the emission spectrum [941].
Figure 9.4 shows the relative spectral radiant power of F2 and F7 along with
the CIE standard illuminant D65 . F11 is shown in Figure 9.5. For this illuminant,
normalization to 100 at 560 nm exaggerates its entire spectral distribution, since
there is very little power at that wavelength. Therefore, for such illuminants,
different normalization strategies, such as normalizing by the entire area under
the spectral curve, are also employed [960].
Recognizing the need to catch up with the rapidly changing fluorescent in-
dustry the CIE released 15 new fluorescent illuminants in 2004 [194]. These new
illuminants are denoted by symbols from FL3.1 to FL3.15 and are classified into
several groups as follows:
• FL3.1 − 3.3: standard halo-phosphate lamps;
• FL3.4 − 3.6: deluxe type lamps;

i i

i i
 i i

i i

500 9. Illuminants

3000

Relative Radiant Power

D65
2500 F11

2000

1500

1000

500

0
300 400 500 600 700 800 900
Wavelength (λ)
Figure 9.5. CIE illuminant F11 and the CIE standard illuminant D65 . Note that since F11
has very little power at λ = 560, normalization to 100 at this wavelength exaggerates its
entire spectral power distribution.

• FL3.7 − 3.11: three band lamps;

• FL3.12 − 3.14: multi-band lamps;
• FL3.15: daylight simulator.
Relative Radiant Power

400

350 D65
FL3.15
300

250

200

150

100

50

0
300 400 500 600 700 800 900
Wavelength (λ)
Figure 9.6. CIE illuminant FL3.15 (daylight simulator) shown with CIE standard illumi-
nant D65 .

i i

i i
 i i

i i

9.1. CIE Standard Illuminants and Sources 501

Spectral radiant distributions of these illuminants are given in the accompanying

DVD-ROM, and a comparison of FL3.15 with the CIE standard illuminant D65 is
shown in Figure 9.6.

9.1.6 Other Illuminants

An important theoretical illuminant is one whereby at each wavelength the same
amount of light is emitted. Typically, this amount is assumed to be unity. This
leads to a flat emission spectrum, which is called an equal-energy stimulus. It is
also often referred to as an equal energy spectrum on the wavelength basis [1261].
Table 9.1 lists the correlated color temperature and chromaticity coordinates for
the equal-energy stimulus under illuminant E.
The CIE defined two other illuminants, namely B and C, together with illu-
minant A, in 1931. Illuminant B is intended to represent direct sunlight with a
correlated color temperature of 4874 K, while illuminant C is intended to model
average daylight with a correlated color temperature of 6774 K.
Both of these illuminants can be realized by using special filters to steer the
distribution of light emitted from a CIE standard source A [558]. Illuminant B can
be realized by using a double layered colorless optical glass filter where each layer
is 1 cm thick and contains one of the solutions B1 and B2 shown in Table 9.4.
Relative Radiant Power

300
A
250 B
C
D65
200

150

100

50

0
300 400 500 600 700 800 900
Wavelength (λ)

Figure 9.7. Relative spectral radiant power distributions of CIE standard illuminants A,
D65 , and deprecated CIE standard illuminants B and C. Note the discrepancy of B and C
from D65 especially in the ultra-violet region.

i i

i i
 i i

i i

502 9. Illuminants

B1 Compound Formula Amount

Copper Sulphate CuSO4 . 5H2 O 2.452 g
Mannite C6 H8 (OH)6 2.452 g
Pyridine C 5 H5 N 30.0 ml
Distilled water to make 1000.0 ml
B2
Cobalt Ammonium Sulphate CoSO4 . (NH4 )2 SO4 . 6H2 O 21.71 g
Copper Sulphate CuSO4 . 5H2 O 16.11 g
Sulphuric Acid density 1.835 g.ml−1 10.0 ml
Distilled water to make 1000.0 ml

Table 9.4. Solutions B1 and B2 used for producing filters to realize illuminant B from
illuminant A.

C1 Compound Formula Amount

Copper Sulphate CuSO4 . 5H2 O 3.412 g
Mannite C6 H8 (OH)6 3.412 g
Pyridine C5 H5 N 30.0 ml
Distilled water to make 1000.0 ml
C2
Cobalt Ammonium Sulphate CoSO4 . (NH4 )2 SO4 . 6H2 O 30.58 g
Copper Sulphate CuSO4 . 5H2 O 22.52 g
Sulphuric Acid density 1.835 g.ml−1 10.0 ml
Distilled water to make 1000.0 ml

Table 9.5. Solutions C1 and C2 used for producing filters to realize illuminant C from
illuminant A.

Illuminant C can be realized similarly using the solutions C1 and C2 given in

Table 9.5. These filters are commonly called Davis-Gibson filters [235]. The
spectral distributions of illuminants B and C are shown in Figure 9.7.
However, both of these illuminants inadequately represent their intended spec-
tral distributions. The chief deficiency is in the ultra-violet range of the spectrum,
which plays an important role in the colorimetry of fluorescent materials. As
such, they are dropped from the list of standards by the CIE, and illuminant D65
is recommended as a replacement. Currently illuminants B and C are not recom-
mended for use in any application, but are included here for completeness.

i i

i i
 i i

i i

9.2. Color Temperature 503

Approximate image color of a blackbody with changing temperature

1000

3000

5000

7000

9000

11000

13000

25000
Temperature in Kelvins

Figure 9.8. Various colors of a blackbody at different absolute temperatures. Note that
these colors only approximate the true colors due to limitations in the printing process.

9.2 Color Temperature

When a blackbody is heated so that its temperature rises, its color changes from
reddish to bluish (Section 3.5.1). This effect is shown in Figure 9.8. This figure
shows that the color temperature of a blackbody is equal to its absolute temper-
ature in Kelvin. Additionally, it is clear from this figure that the scale is non-
linear with human visual perception, in that a very large part of the scale is taken
up by blue colors. A related measure, known as the reciprocal megakelvin or
remek,1 is computed as the reciprocal color temperature. It’s unit is 106 K−1 , and
is approximately perceptually linear [561, 927, 1262].
Most light sources and illuminants are not perfect blackbodies. Furthermore
for non-incandescent sources, such as fluorescent lamps, the color they emit is not
directly related to their temperature. Their color is determined by the composition
of the gas mixture used in their tubes (see Section 3.5.4). Therefore, color tem-
perature is used exclusively for blackbodies, and different terms are introduced
for other types of sources. These are explained in the following sections.

9.2.1 Distribution Temperature

The distribution temperature is a property of a light source whose relative spec-

tral distribution is similar to that of a blackbody radiator. The temperature of the
blackbody radiator for which this match occurs is then called the distribution tem-
perature of the light source. This implies that both the light source and the black-
body at that particular temperature emit light of the same chromaticity [1262].

i i

i i
 i i

i i

504 9. Illuminants

Temperature CIE 1931

T (in K) x y
1500 0.5857 0.3931
2000 0.5267 0.4133
2500 0.4770 0.4137
2856 0.4475 0.4074
4000 0.3802 0.3767
5000 0.3444 0.3513
5500 0.3314 0.3405
6500 0.3118 0.3224
7500 0.2978 0.3083
10000 0.2762 0.2840

Table 9.6. Chromaticity coordinates of a blackbody radiator at various color temperatures.

9.2.2 Color Temperature of a Source

The color temperature of a source is the temperature (in Kelvin) of a blackbody
radiator which emits light of the same (or almost the same) chromaticity as the
light source. Note that the relative spectral power distributions of the source and
the blackbody radiator need not be identical.

y 0.9

0.8

0.7

0.6
Planckian locus
0.5
4000 2500
5500 1500
0.4
7500
0.3 2000
2856
10000 5000
0.2 6500
0.1

0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
x

Figure 9.9. The Planckian locus sampled at various absolute temperatures.

1 This quantity was formerly known as the microreciprocal degree or mired [1262].

i i

i i
 i i

i i

9.2. Color Temperature 505

To illustrate the concept of color temperature, let us consider the chromaticity

coordinates of a blackbody radiator at various temperatures . These coordinates
may be computed by integrating the spectral distribution of a blackbody radiator
by the CIE 1931 color-matching functions. A set of such chromaticity coordi-
nates is computed in the interval 1500 K–10000 K and tabulated in Table 9.6. By
plotting these coordinates on the CIE 1931 chromaticity diagram, as shown in
Figure 9.9, we find the Planckian locus, the line connecting subsequent color
temperatures.
It is only appropriate to speak of a color temperature for a light source if its
chromaticity coordinates lie on the Planckian locus. In general, only incandes-
cent sources (e.g., a tungsten filament light bulb) satisfy this requirement. An
example is the CIE standard source A which has a color temperature of 2856 K.
For other light sources, the related attribute correlated color temperature may be
determined, as discussed next.

9.2.3 Correlated Color Temperature

The term correlated color temperature is introduced when the chromaticity co-
ordinates of a source are not identical to that of a full blackbody radiator (i.e.,
when the chromaticity coordinates of a light source do not lie on the Planckian
locus). In that case, the correlated color temperature denotes the absolute temper-
ature of a blackbody radiator whose perceived color is closest to that of the light
source [1262].

v’ 0.7
Planckian locus
0.6 Iso-temperature lines

0.5

0.4

0.3

0.2

0.1

0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
u’

Figure 9.10. Iso-temperature lines shown on the CIE 1976 u v chromaticity diagram.

i i

i i
 i i

i i

506 9. Illuminants

v’ 0.54

0.52

0.50

0.48

0.46

Planckian locus
0.44 Iso-temperature lines

0.18 0.20 0.22 0.24 0.26 0.28

u’

Figure 9.11. A close-up view of iso-temperature lines shown on the CIE 1976 u v chro-
maticity diagram.

This temperature may be determined graphically using a chromaticity dia-

gram. However, in this case, a uniform chromaticity diagram such as the CIE
1976 u v Uniform Chromaticity Scale (UCS) needs to be employed. In UCS,
there is a direct correlation between the distance between the chromaticity points
and their perceived difference, a property missing in CIE 1931 xy chromaticity
diagrams.

y
Planckian locus
0.8
Iso-temperature lines

0.6

0.4

0.2

0.0
0.0 0.2 0.4 0.6 0.8
x

Figure 9.12. Iso-temperature lines transformed back to the CIE 1931 xy chromaticity
diagram.

i i

i i
 i i

i i

9.2. Color Temperature 507

y
Planckian locus
0.45 Iso-temperature lines

0.4

0.35

0.3

0.30 0.34 0.38 0.42 0.46 0.50

x

Figure 9.13. A close-up view of iso-temperature lines on the CIE 1931 chromaticity
diagram. Note that intersections are not orthogonal as a result of the perceptual non-
uniformity of this diagram.

The method of computing the correlated color temperature is illustrated in

Figure 9.10. This figure shows the same data as Figure 9.9 except that it is trans-
formed into the CIE u v chromaticity diagram. The forward and backward trans-
formations are given by Equations (7.13) and (7.14).
In Figure 9.10, the blue line shows the Planckian locus and the green lines
show the so-called iso-temperature lines. Iso-temperature lines intersect the

y Daylight locus
Planckian locus D50
0.36
Iso-temperature lines
D55 5000
0.34
D65
6000
0.32
D75
7000

0.30 8000
9000

0.28 0.30 0.32 0.34 0.36

x

Figure 9.14. The daylight locus together with the Planckian locus sampled at various
absolute temperatures. With the aid of iso-temperature lines, one may obtain the correlated
color temperatures for daylight illuminants.

i i

i i
 i i

i i

508 9. Illuminants

Planckian locus perpendicularly, and the temperature at the intersection gives the
correlated color temperature for all stimuli whose chromaticity points lie on that
iso-temperature line. A close-up diagram is shown in Figure 9.11.
This diagram can be converted back into the more familiar CIE 1931 chro-
maticity diagram using (7.14) (Figure 9.12; a close-up is shown in Figure 9.13).
Note that iso-temperature lines are no longer orthogonal to the Planckian locus as
a result of the non-uniform nature of the CIE xy chromaticity diagram.
As an example, consider finding the correlated color temperature of CIE day-
light illuminants. In Figure 9.14, the chromaticity coordinates of CIE illuminants
D50 , D55 , D65 , and D75 are shown on a line, which is also called the daylight lo-
cus. Since these coordinates do not lie on the Planckian locus, we can only speak
of correlated color temperature for these illuminants. Correlated color tempera-
tures may be found by tracing the iso-temperature lines to their intersection point
with the Planckian locus.

9.3 Color-Rendering Index

Together with their reflectance properties, one of the key parameters that deter-
mines how objects appear to us is the color of the illuminating light source. The
color-rendering index (CRI) of a light source thus is intended to quantify the ef-
fect the light source has on the appearance of objects in comparison with their
appearance under a reference illuminant [1262].
Before explaining how to compute the color-rendering index, a practical ex-
ample may illustrate why it is important. During the day people normally do not
struggle to pick out their cars in a parking lot since the color of each car can eas-
ily be distinguished. However, at night, this simple task becomes significantly
more difficult if the parking lot is illuminated with orange-colored lamps such as
sodium vapor lamps (see also Figure 3.32). This is because the color-rendering
index of these lamps is much lower than the CRI of daylight.
The CIE recommends the following procedure for determining the CRI of a
test light source:
1. Select a reference illuminant. The reference illuminant appropriate for test
sources with correlated color temperatures below 5000 K is a Planckian
radiator and for 5000 K and above, one of the CIE D-series illuminants.
Additionally, the reference illuminant needs to be of the same or nearly the
same chromaticity as the test source [190]. A practical threshold for the
chromaticity difference is ΔCT = 5.4 10−3 , where ΔC is computed as
0 10.5
ΔC = (uk − ur )2 + (vk − vr )2 . (9.11)

i i

i i
 i i

i i

9.3. Color-Rendering Index 509

The CIE samples under illuminant D 65

Sample 1 2 3 4 9 10 11

5 6 7 8 12 13 14

The CIE samples under illuminant A

Sample 1 2 3 4 9 10 11

5 6 7 8 12 13 14

Figure 9.15. The appearance of the CIE samples under two different lighting conditions.
These samples are specified by their radiance factors and are used to compute the color-
rendering index of light sources.

Here (uk , vk ) and (ur , vr ) are the chromaticity coordinates of the test and the
reference source, with respect to CIE 1960 UCS.

2. Obtain the spectral-radiance factors of standard samples provided by the

CIE for computation of the color-rendering index. These samples are in-
tended to span a large range of colors commonly encountered in practice.
The CIE defines 14 samples; the first eight samples are moderately satu-
rated, similar in lightness, and cover the hue circle [190]. In contrast, the
last six samples vary widely in saturation and lightness. The appearance
of these samples under daylight and incandescent light is illustrated in Fig-
ure 9.15. Their Munsell notations are given in Table 9.7. The full spectral
data are provided on the accompanying DVD-ROM.

As will be further detailed in the following discussion, the CRI obtained

when any one of these samples is used, is called the CIE special color-
rendering index . On the other hand, the mean CRI of the first eight samples
gives the CIE general color-rendering index.

3. Determine the CIE 1931 XYZ tristimulus values and xy chromaticity coor-
dinates of each sample under the test source and the reference illuminant,
yielding (Xk,iYk,i Zk,i ) and (Xr,i Xr,i Xr,i ) for the ith sample.

i i

i i
 i i

i i

510 9. Illuminants

Sample No. Approximate Munsell notation Color appearance under daylight

1 7.5 R 6/4 Light grayish-red
2 5 Y 6/4 Dark grayish-yellow
3 5 GY 6/8 Strong yellow-green
4 2.5 G 6/6 Moderate yellowish-green
5 10 BG 6/4 Light bluish-green
6 5 PB 6/8 Light blue
7 2.5 P 6/8 Light violet
8 10 P 6/8 Light reddish-purple
9 4.5 R 4/13 Strong red
10 5 Y 8/10 Strong yellow
11 4.5 G 5/8 Strong green
12 3 PB 3/11 Strong blue
13 5 YR 8/4 Light yellowish-pink (human complexion)
14 5 GY 4/4 Moderate olive green (leaf green)

Table 9.7. Test color samples used for computation of the CIE color-rendering index of
light sources.

4. Transform the tristimulus values into the CIE 1960 UCS uv chromaticity
coordinates. This can be accomplished by either of the following equations:

4X 6Y
u= v= ; (9.12)
X + 15Y + 3Z X + 15Y + 3Z
4x 6y
u= v= . (9.13)
−2x + 12y + 3 −2x + 12y + 3

5. Account for the chromatic-adaptation difference (see Chapter 10) that would
occur between the test source and the reference illuminant. This results in
the following adapted-chromaticity coordinates:

cr dr
10.872 + 0.404 ck,i − 4 dk,i
ck dk
uk,i = , (9.14)
cr dr
16.518 + 1.481 ck,i − dk,i
ck dk
5.520
vk,i = , (9.15)
cr dr
16.518 + 1.481 ck,i − dk,i
ck dk

where subscripts k and r identify terms associated with the test source and
the reference illuminant, respectively, and i denotes the sample number.

i i

i i
 i i

i i

9.3. Color-Rendering Index 511

The terms c and d are computed by

1
c = (4 − u − 10v), (9.16)
v
1
d = (1.708v + 0.404 − 1.481u). (9.17)
v

6. Transform the adaptation-corrected uk,i , vk,i values into the CIE 1964 uni-
form color space.2 This is accomplished by

Wr,i∗ = 25(Yr,i )1/3 − 17 ∗

Wk,i = 25(Yk,i )1/3 − 17, (9.18)
∗
Ur,i = 13Wr,i∗ (ur,i − ur ) ∗
Uk,i ∗
= 13Wk,i (uk,i − uk ), (9.19)
Vr,i∗ = 13Wr,i∗ (vr,i − vr ) ∗
Vk,i ∗ 
= 13Wk,i (vk,i − vk ). (9.20)

Here Yr,i and Yk,i must be normalized such that Yr = Yk = 100.

7. Finally, the resultant color shift is computed by the CIE 1964 color differ-
ence formula:

ΔEi = (Ur,i ∗ −U ∗ )2 + (V ∗ −V ∗ )2 + (W ∗ −W ∗ )2 (9.21)
k,i r,i k,i r,i k,i

= (ΔUi∗ )2 + (ΔVi∗ )2 + (ΔWi∗ )2 . (9.22)

ΔEi denotes the color shift that occurs for the ith sample when the test
source is replaced by the reference illuminant. Once ΔEi is obtained the
CIE special color-rendering index, Ri is computed by

Ri = 100 − 4.6(ΔE)i . (9.23)

This indicates that the maximum special CRI equals 100, and this condition
only occurs if the color difference is zero for that particular sample when
the test source is replaced by the reference illuminant. The constant 4.6
ensures that a standard warm white fluorescent lamp attains a CRI of 50
when compared against an incandescent reference [190].

8. The CIE general color-rendering index is computed as an average of the

first eight special color-rendering indices:

1 8
Ra = ∑ Ri .
8 i=1
(9.24)

2 Although the CIE 1964 uniform color space and color-difference formula has been replaced by

the CIE 1976 uniform color space and color-difference formula, they have been retained for the time
being for computing color-rendering indices [190].

i i

i i
 i i

i i

512 9. Illuminants

The color-rendering index is an important metric especially in the lighting

industry. Usually lamps with high color-rendering indices are preferred to those
with low color-rendering indices.
The CIE procedure used to determine the color-rendering index has been sub-
jected to various criticisms [1142]. One of the objections to the CIE method is
that it uses an obsolete color space and an obsolete color-difference formula. Al-
though improvements have been made in both of these areas, the recommendation
has not been updated. Furthermore, the improvements in the study of chromatic-
adaptation transforms are also not reflected in the recommended procedure.
In addition, it has been argued that even if its out-of-date components are
replaced with the more improved versions, the procedure would still be inad-
equate [1142]. In this respect, the main criticism is directed at the possibility
to choose more than one reference source and the insufficiency of the test sam-
ples [1142].

9.4 CIE Metamerism Index

If two object-color stimuli of different relative spectral radiant power distributions
yield the same tristimulus values according to a given colorimetric observer and
under a reference illuminant, these stimuli are said to be metameric to each other,
or simply called metamers.
Denoting the spectral reflectance of objects as β1 (λ ) and β2 (λ ) and the spec-
tral concentration of the illuminant as S(λ ), the necessary conditions for meta-
merism are
 
β1 (λ )S(λ )r̄(λ ) = β2 (λ )S(λ )r̄(λ ), (9.25)
λ λ
 
β1 (λ )S(λ )ḡ(λ ) = β2 (λ )S(λ )ḡ(λ ), (9.26)
λ λ
 
β1 (λ )S(λ )b̄(λ ) = β2 (λ )S(λ )b̄(λ ), (9.27)
λ λ

where r̄(λ ), ḡ(λ ), and b̄(λ ) are color-matching functions.

Metamers are perceived to be of the same color under a reference illuminant,
S(λ ). However, these equations indicate that if either the illuminant or the color-
matching functions are changed, these equalities may no longer hold; i.e., pairs of
colors may no longer be metamers. Also, it is intuitively clear that the more dif-
ferent β1 (λ ) is from β2 (λ ), the more likely that this equality is distorted, should

i i

i i
 i i

i i

9.4. CIE Metamerism Index 513

either the source or the observer change [1262]. The degree of this mismatch, as
defined by the CIE when either the illuminant or the observer is changed, is called
the CIE metamerism index.
Since the metamerism index indicates how much the perceived color of two
samples will change when the lighting is changed, it is of particular importance
for the paint and the textile industries. For instance, it is important to ensure that
all parts of a car exterior look the same regardless of changes in illumination.
The CIE defines two types of metamerism index. The first index is used to
quantify the degree of mismatch when the illuminant is changed for a constant
observer. This is called the CIE special metamerism index: change in illuminant .
The second index is used when the illuminant is fixed and the observer is allowed
to change. This type of index is called the CIE special metamerism index: change
in observer [188].
The CIE recommends the following procedure to compute the CIE special
metamerism index of two samples:

1. Compute the tristimulus values Xm,i ,Ym,i , Zm,i , (i = 1, 2) and (m = r), of the
two samples under the reference illuminant (the preferred reference illumi-
nant is D65 ):

Xm,i = k ∑ βi (λ )Sm (λ )x̄(λ )Δλ , (9.28)

λ

Ym,i = k ∑ βi (λ )Sm (λ )ȳ(λ )Δλ , (9.29)

λ

Zm,i = k ∑ βi (λ )Sm (λ )z̄(λ )Δλ , (9.30)

λ

where Sm=r (λ ) is the spectral power distribution (SPD) of the reference il-
luminant, the βi are the spectral reflectance of the samples, and x̄(λ ), ȳ(λ ),
and z̄(λ ) are the color-matching functions (CIE 1931 or 1964); Δλ is the
sampling resolution of the spectrum. The normalization factor k is calcu-
lated as
100
k= . (9.31)
∑ Sm (λ )ȳ(λ )Δλ
λ

Normalization ensures that the Y -component of a perfect reflecting diffuser

is equal to 100. Since the samples are metameric under the reference illu-

i i

i i
 i i

i i

514 9. Illuminants

minant, their tristimulus values must be equal:

Xm=r,1 = Xm=r,2 , (9.32)

Ym=r,1 = Ym=r,2 , (9.33)
Zm=r,1 = Zm=r,2 . (9.34)

2. Compute the tristimulus values Xm,i ,Ym,i , Zm,i , (i = 1, 2) and (m = t), of

the two samples under the test illuminant using the procedure in the previ-
ous step. This yields two different sets of tristimulus values, namely one for
each sample. Here, the choice of test illuminant depends on the target appli-
cation. For instance, if the metamerism needs to be sustained under typical
fluorescent lighting the CIE F2 illuminant may be chosen [136, 137, 1262].

3. The CIE metamerism index Milm is equal to the CIELAB color differ-
ence ΔEab ∗ between the tristimulus values X
m=t,1 , Ym=t,1 , Zm=t,1 and Xm=t,2 ,
Ym=t,2 , Zm=t,2 (see Section 8.7.1):
∗
Milm = ΔEab . (9.35)

If a different color-difference formula, including the ones introduced in

Section 8.8, is used this should be noted [194].

The CIE metamerism index for a change in observer can be computed similarly.
The sole difference is that the illuminant is kept constant while the observer (i.e.,
its associated color-matching functions) is changed.
The CIE metamerism index is particularly important for the paint and textile
industries where all products of a given type should match a standard. In this case,
a lower degree of metamerism is desired [558].

9.5 Dominant Wavelength

The CIE tristimulus values of a color stimulus, or its related chromaticity coor-
dinates, provide numerical information that can be used in colorimetric computa-
tions. However, they provide little information to the inexperienced with regard to
what that color stimulus actually looks like. It is possible to infer that a stimulus
with tristimulus values X > Y > Z may look more reddish than bluish; it requires
significant experience to determine what the actual color is.
Measures of dominant wavelength and purity are used in this context to de-
scribe colors in a more intuitive manner. To understand the appearance of a color

i i

i i
 i i

i i

9.5. Dominant Wavelength 515

stimulus from these two concepts, one should still know the progression of colors
from the short to the long wavelengths in the light spectrum.
The dominant wavelength of a color stimulus is the monochromatic wave-
length of the spectrum such that when light of this wavelength is mixed with
an achromatic stimulus, the mixture matches the original stimulus in color. It
is customary to use either CIE standard illuminant D65 or A as the achromatic
stimulus.3
The dominant wavelength is denoted by λd and can be computed with the
aid of a CIE chromaticity diagram. Recall, that in this diagram, all colors with
chromaticity coordinates lying on a straight line can be produced by mixing, in
appropriate proportions, the two colors with chromaticity coordinates defining
the end points of this line. Therefore, to find the dominant wavelength of a color
stimulus, we can draw a straight line starting from the chromaticity coordinate
(xw , yw ) of an achromatic stimulus (e.g, D65 ) and passing through the chromatic-
ity coordinate (x, y) of the stimulus in question. If we continue extending this
line until it intersects with the spectrum locus, the wavelength corresponding to
this intersection point gives the dominant wavelength of the color stimulus. This
process is illustrated in Figure 9.16.

y 0.9

0.8 520
Spectrum locus
0.7
560
0.6
500 λd (xb,yb)
0.5 580

0.4 S1 (x,y) 600

0.3 D65 (xw,yw)
830
0.2
480
0.1 Purple line
360
0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
x

Figure 9.16. The dominant wavelength λd of stimulus S1 is computed by finding the

intersection with the spectrum locus of the straight line drawn from the chromaticity point
of an achromatic stimulus (D65 ) toward S1 .

3 Although illuminant A has a yellowish hue, it still qualifies as achromatic under some circum-

stances due to the chromatic-adaptation mechanism of the human visual system [558].

i i

i i
 i i

i i

516 9. Illuminants

Unfortunately, the spectrum locus is not represented with an analytical equa-

tion, so numerical or graphic solutions are required. One alternative is to first
graphically determine the approximate position of the dominant wavelength by
using the method described above. For instance, from Figure 9.16, we can see
that the dominant wavelength of S1 is between 560 nm and 580 nm. Then we
can test the chromaticity coordinates of all monochromatic wavelengths in this
interval such that the slope of the line drawn from S1 to these points is closest to
the slope of the line drawn from D65 to S1 . This procedure can be stated as an
optimization problem, where the goal is to minimize the difference in these two
slopes. The wavelength of the monochromatic stimulus that minimizes the slope
then represents the dominant wavelength.
The dominant wavelength is useful to describe the approximate hue of a color
stimulus. For instance, if one knows that the light of wavelengths between roughly
570 nm and 650 nm has an orange hue, a light source with its dominant wavelength
falling in this interval will also tend to have an orange hue. The dominant wave-
length is commonly used by LED manufacturers to specify the characteristics of
light sources [1094].
Not all color stimuli have a dominant wavelength. For some stimuli, the
straight line may intersect with the purple line rather than the spectrum locus;

y 0.9

0.8 520 λc or -λd

Spectrum locus
0.7
560
0.6
500
0.5 580

0.4 600
D65 (xw, yw)
0.3
830
0.2 S2 (x,y)
480 Purple line
0.1
360 P (xb, yb)
0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
x
Figure 9.17. The complementary wavelength λc (or −λd ) of stimulus S2 is computed by
finding the intersection with the spectrum locus of the straight line drawn from S2 toward
the achromatic stimulus (D65 ). The intersection of this line with the purple line is shown
by P and used for computing excitation or colorimetric purity. All data is shown on the
CIE 1931 chromaticity diagram.

i i

i i
 i i

i i

9.6. Excitation Purity 517

an example is shown in Figure 9.17. It is evident from the figure that this happens
when the chromaticity coordinates of a stimulus lie in the triangle whose apex is
the achromatic stimulus.
For such stimuli, we speak of complementary wavelength rather than dom-
inant wavelength. The complementary wavelength λc (or −λd ) indicates the
monochromatic wavelength such that when the stimulus is mixed with a color
of this wavelength, the mixture matches a reference achromatic stimulus in color.
The value λc can be computed similarly to λd .

9.6 Excitation Purity

Although the dominant and complementary wavelengths give cues about the hue
of a color stimulus, they do not convey information about how saturated a color
stimulus is. That is, all color stimuli that lie on the same line as S1 (see Fig-
ure 9.16) will have the same dominant wavelength, but they may differ signifi-
cantly in saturation. For instance, we regard the stimuli close to the chromaticity
coordinate of the achromatic stimulus as being desaturated. On the other hand,
the stimuli close to the spectrum locus is considered be highly saturated.
Thus the saturation of a color stimulus is expressed by a different term called
the excitation purity. It indicates how far a stimulus is from an achromatic stimu-
lus on the CIE chromaticity diagram. The excitation purity is defined as a ratio of
distances: ⎧ x−x
⎪
⎨
w
if x − xw ≥ y − yw ,
xb − xw
pe = y − y (9.36)
⎪
⎩ w
if x − xw < y − yw ,
yb − yw
Note that of the two forms on the right-hand side, the one with the largest numer-
ator is chosen, as it results in the smallest precision error [194]. The excitation
purity is only a loose measure of the saturation of a color stimulus [1262]. For
more accurate estimations, more elaborate models must be used (see Chapter 11).

9.7 Colorimetric Purity

A related concept to excitation purity is the colorimetric purity, denoted by pc .
The colorimetric purity of a color stimulus expresses how much white light (i.e.,
achromatic light), with a wavelength equal to the dominant wavelength of the
stimulus considered, must be added to the monochromatic stimulus so that their
mixture matches the color of the stimulus in question. As such, the colorimetric

i i

i i
 i i

i i

518 9. Illuminants

purity pc can be computed by

Ld
pc = , (9.37)
Ld + Ln
where Ld and Ln represent the luminances of the monochromatic and achromatic
stimuli, respectively. By using Grassmann’s laws this equation can be rewritten
as
Ld
pc = , (9.38)
Ls
where Ls is the luminance of the stimulus for which the purity is computed, and
clearly Ls = Ld + Ln . Colorimetric purity may also be derived from the excitation
purity, and vice versa:
⎧y x−x
⎪
⎨
b w
if x − xw ≥ y − yw ,
y x b − xw
pc = y y − y (9.39)
⎪
⎩ b w
if x − xw < y − yw ,
y y b − yw

or equivalently,
yb
pc = pe . (9.40)
y
The coordinates (x, y), (xb , yb ), and (xw , yw ) have the same meanings as those
used for excitation purity with one crucial exception. For stimuli that possess
a complementary wavelength, (xb , yb ) represents the chromaticity coordinates of
the complementary wavelength λc , rather than the chromaticity coordinates on
the purple line (see Figure 9.17).
In all, it is important to remember that the quantities discussed in this section
are defined on a non-uniform CIE chromaticity diagram. This fact in general, may
limit their use and should be kept in mind whenever a color is expressed by using
these quantities [509].

9.8 Application: Modeling Light-Emitting Diodes

The technology for the construction of light-emitting diodes (LEDs) is rapidly
advancing to the point that they have become useful in various novel applica-
tions, including car lights, illumination, and as backlights in LCD displays (see
Section 14.2.1).
Important characteristics of LEDs are their luminous efficacy of radiation and
their color-rendering index. In practical LED designs, there is a trade-off between
these two characteristics: higher luminous efficacies can be achieved at the cost of

i i

i i
 i i

i i

9.8. Application: Modeling Light-Emitting Diodes 519

Dichromatic white LED Trichromatic white LED Tretrachromatic white LED

Figure 9.18. LED-based lightsources configured as dichromatic, trichromatic, and tetra-

chromatic sources (after [1014]).

a lower color-rendering index, and vice-versa [1014]. Nonetheless, the luminous

efficacy follows Haitz’s law, which states that the luminous output of LEDs is
doubling every 18–24 months [1083]4 .
White LEDs can be fabricated in several different ways. A main distinction
between different types of LEDs is whether they contain phosphors or not, i.e.,
LEDs or phosphor-converted LEDS (pc-LEDs). Each of these types can be further
classified into dichromatic, trichromatic, and tetrachromatic sources, as illustrated
in Figures 9.18 and 9.19. The spectral power distribution functions for some of
these types are shown in Figure 9.20.

Dichromatic white: Trichromatic white: Tretrachromatic white:

blue LED plus UV LED plus red and blue LEDs
yellow phosphor tri-phosphor plus cyan and green phosphors

Figure 9.19. pc-LED lightsources configured as dichromatic, trichromatic, and tetrachro-

matic sources (after [1014]).

Dichromatic LEDs tend to have good luminous efficacy, whereas tetrachro-

matic LEDs have a better color-rendering index. Trichromatic LEDs are in be-
tween, with a good luminous efficacy as well as a good color-rendering index.
In pc-LEDs, the conversion from short-wavelength photons to long-wavelength
photons involves an energy loss, reducing their efficiency by between 10% and
30% [1014]. As tetrachromatic LEDs and pc-LEDs have four emission peaks,
spread over the visible wavelength, their emission spectra are most uniform, lead-
ing to the highest color-rendering index of the types discussed here.
The spectral power distribution of an LED has two, three, or four peaks, as-
sociated with the specific semiconductors or phosphors used. The shape of each

4 This law is analogous to Moore’s law, which predicts that the number of transistors on chips

doubles every 18–24 months.

i i

i i
 i i

i i

520 9. Illuminants

Blue LED, yellow phosphor Blue LED, red LED, yellow phosphor
1.0

Relative luminous
0.8
intensity 0.6
0.4
0.2
0.0
UV LED, RGB phosphors RGB LED
1.0
Relative luminous

0.8
intensity

0.6
0.4
0.2
0.0
400 450 500 550 600 650 700 750 400 450 500 550 600 650 700 750

Figure 9.20. Spectral power distribution functions for some LED types. (after [1059]).

peak can be approximated with the superposition of two Gaussian curves [848]:
   5
1 (λ − λ0 )2 2 (λ − λ0 )2
Si (λ ) = exp − + exp − , (9.41)
3 Δλ 3 Δλ

where λ0 is the center of the peak, and Δλ is the width of the peak. The spectral
power distribution of an LED is then the weighted sum of the individual peaks:

S(λ ) = ∑ wi Si (λ ), (9.42)
i

where the weights wi adjust the relative magnitude of the peaks. The widths of
the peaks, given by Δλ , are between approximately 20 and 30 nm. Typical peak
wavelengths for a tetrachromatic LED are λ1 = 450 nm, λ2 = 510 nm, λ3 = 560
nm, and λ4 = 620 nm [1014].

9.9 Application: Estimating the Illuminant

in an Image
Given an image, it is often desirable to know the spectral composition of the
light source in the scene [1133]. For instance, object-segmentation algorithms
can be based on color differences. If the nature of the illuminant is known, then

i i

i i
 i i

i i

9.9. Application: Estimating the Illuminant in an Image 521

such segmentation algorithms can be constructed in a more robust manner, as

one can account for changes due to the illuminant as well as to highlights [56,
57, 132, 442, 595, 1236]. In addition, digital cameras typically employ corrective
algorithms that make use of an estimate of the scene illuminant [153, 495, 1034,
1256]. A further application lies in image database retrieval, where knowledge of
the illuminant will allow a better match between a source image and the collection
of images in the database [266, 338, 1003, 1110].
To estimate the color of the illuminant on the basis of a single image, it is
often convenient to make additional assumptions on the nature of the illuminants’
spectra that are to be recovered. Given that the number of different illuminants
encountered in average environments is limited to daylight spectra, as well as dif-
ferent types of incandescent and fluorescent artificial lights, this is not unreason-
able. In fact, it allows the problem to be simplified from estimating the illuminant
to classifying the illuminant to be one of a small set of spectra [1133, 1134].
Many different computational algorithms for separating the color of a surface
and the color of the illuminant have been proposed, and they can be classified into
different groups [1133], including those based on general linear models [141,474,
662, 728, 1204], algorithms based on an analysis of diffuse inter-reflection and
the appearance of highlights [339, 441, 596, 664], Bayesian methods [119, 994,
1301], multi-view methods [1149, 1298, 1299], and algorithms based on spectral
images [1132, 1135, 1137].
The problem of illuminant estimation, or classification, emerges in human
vision as well, where it is related to color constancy (Section 5.9). Humans are
able to perceive the color of surfaces in a variety of everyday situations. The
spectrum of the illuminant can vary without the perception of surface color being
greatly affected [120,123], which requires either implicit or explicit knowledge of
the illuminant. Of course there are also viewing conditions where color constancy
breaks down [454]. We will assume that under normal viewing conditions, color
constancy holds sufficiently well that an estimation of the illuminant is possible.
Of the different computational techniques mentioned above, some are more
plausible to be part of human vision than others. In particular, there is some evi-
dence that the human visual system may use specular highlights to infer the color
of the illuminant [1271]. As an example of how the color of the dominant illumi-
nant may be estimated, we therefore focus our example on one of the algorithms
that is are based on an analysis of highlights.
The nature of highlights visible on an object depends on its material. Many
materials have a high oil or water content which causes specular reflections as a
result of their constant index of refraction over the visible spectrum [578]. In ad-
dition, such materials may exhibit diffuse reflection. This includes, for instance,

i i

i i
 i i

i i

522 9. Illuminants

human skin. As a result, the specular highlight has the same color as the illu-
minant, whereas off-peak angles show the color of the surface. This behavior
is captured in the dichromatic reflection model, which essentially models the bi-
directional reflectance distribution function (BRDF) as an additive mixture of a
specular and diffuse component [596, 1028, 1136].
It is now possible to analyze image areas that are known to contain highlights,
for instance the nose in human faces [1098]. The pixels away from the highlight
represent the diffuse reflection component; these are the darker pixels. They will
lie on an approximately straight line in RGB color space. The lighter pixels will
form a separate cluster, lying on a differently oriented straight line. Separating
the two clusters can be achieved using a variation of principal components analy-
sis [1098].
Alternatively, one may view each surface point as casting a vote for its illumi-
nant [664]. Using the same dichromatic reflection model, whereby each point on
the surface reflects partly its diffuse body color and partly the specular illuminant
color, we can plot pairs of surface points in a CIE chromaticity diagram. Each
pair forms a line segment which, when extended in the appropriate direction, will
include the chromaticity point of the illuminant. In other words, colors from the
same surface will have different purities, and they will all tend to lie on the same
line. Of course, when a surface does not have a strong highlight, then the change
in purity between different points on the surface can be small and, therefore, prone
to error.
Repeating this procedure for a set of points drawn from a different surface,
we find a second line in the chromaticity diagram. The intersection point with the
first line will then yield an estimation of the chromaticity coordinate of the illu-
minant [664]. If there are more surfaces in a scene, then more lines can be drawn,
which theoretically will all pass through the same intersection point. However,
noise in the measurements, as well as diffuse inter-reflection between objects,
may cause these lines to not pass exactly through the same point, thereby intro-
ducing inaccuracies. In that case, each surface in the scene creates a different line
which can be seen as a vote for a particular illuminant color. The actual color of
the illuminant is then estimated to be identical to the majority vote [664].
If it is known beforehand that the illuminant is a blackbody radiator, then the
above algorithm can be simplified to the analysis of a single highlight-containing
surface. Its line on the chromaticity diagram will intersect the Planckian locus
(Figure 9.9).
An example image, shown in Figure 9.21, is informally analyzed by cropping
highlight patches and plotting the pixels of these patches in a chromaticity dia-
gram, shown in Figure 9.22. The analysis is only informal as the camera was not

i i

i i
 i i

i i

9.9. Application: Estimating the Illuminant in an Image 523

Figure 9.21. The bull figurine is illuminated from behind by a white light and from the
front by an incandescent source, leading to two spatially separate highlights that can be
analyzed individually. (Bronze figurine by Loet Vanderveen.)

characterized, and therefore the conversion of pixels to chromaticity coordinates

can only be approximate. These figures are only included to illustrate the method.
The highlight created by the white-light source forms a straight line that in-
tersects the Planckian locus near a color temperature of 5500 K. The light source
is, in fact, a D65 daylight simulator. The patch taken around the incandescent
highlight yields a cluster in the chromaticity diagram on the right of Figure 9.22
that intersects the Planckian locus at a color temperature of under 2000 K.
In summary, the analysis of highlights in an image may be used to estimate
the color of the illuminant, assuming that the image contains highlights that can
be identified in some manner. There exists some evidence that the human visual
system employs a similar strategy. However, several other computational tech-
niques may be used instead of, or in conjunction with, highlight analysis. In any
case, all techniques require all parts of the image, including the highlights, to be
well exposed. Since the intensity of specular highlights can be arbitrarily high,
high dynamic range imaging techniques (as discussed in Part III of this book) can
be useful for illuminant estimation algorithms.

i i

i i
 i i

i i

524 9. Illuminants

y 0.9 y 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3
Planckian locus Planckian locus
0.2 0.2

0.1 0.1

0.0 0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
x x

Figure 9.22. The two highlight patches in the insets are taken from Figure 9.21. The
chromaticity coordinates of the white highlight form a line that intersects the Planckian
locus (left), whereas the incandescent highlight does not (right).

9.10 Further Reading

The IESNA lighting handbook is an invaluable resource for the lighting engi-
neer [941]. Wyszecki and Stiles have also collected a large amount of data [1262].
Many of the CIE publications are relevant to work involving illuminants.

i i

i i
 i i

i i

Chapter 10
Chromatic Adaptation

In everyday life we encounter a wide range of viewing environments, and yet our
visual experience remains relatively constant. This includes variations in illumi-
nation, which on any given day may range from starlight, to candlelight, incan-
descent and fluorescent lighting, and sunlight. Not only do these very different
lighting conditions represent a huge change in overall amounts of light, but they
also span a wide variety of colors. Objects viewed under these varying conditions
generally maintain an overall constant appearance, e.g., a white piece of paper
will generally appear white (or at least will be recognized as white).
The human visual system accommodates these changes in the environment
through a process known as adaptation. As discussed briefly in Chapter 5, there
are three common forms of adaptation: light adaptation, dark adaptation, and
color or chromatic adaptation. Light adaptation refers to the changes that occur
when we move from a very dark environment to a very bright environment, for in-
stance going from a matinee in a darkened movie theater into the sunlight. When
we first move outside, the experience is dazzling and may actually be painful to
some. We quickly adapt to the bright environment, and the visual experience re-
turns to normal. Dark adaptation is exactly the opposite, moving from a bright
environment to a dark one. At first it is difficult to see anything in the dark room,
but eventually we adapt and it is possible to see objects again.
In general the time required for dark adaptation is much longer than that re-
quired for light adaptation (see Section 3.4.2). Part of the light- and dark-adapting
process can be explained physiologically by the dilation of the pupil. If we con-
sider that the average pupil diameter can range from 2.5 mm in bright light, to
8 mm in darkness, this can account for about a 10:1 reduction in the amount of
light striking the retina (considering a circular area of approximately 5 mm2 to
525

i i

i i
 i i

i i

526 10. Chromatic Adaptation

50 mm2 ) [1262] (see also Section 4.2.4). This reduction is not enough to account
for the wide variety of dynamic ranges we typically encounter. The remainder
of the adaptation occurs at the photoreceptor and retinal level, as well as at the
cognitive level.
While changes in overall illumination levels are important for digital imaging
systems, they are generally not considered the most important form of adapta-
tion. Typically, we are more concerned with changes in the color of the lighting.
As we learned in Section 2.11.2, the overall color of everyday illumination can
vary widely, from orange sunlight at dawn or twilight, to the blue Northern sky.
Despite these varieties of illumination, the colors of objects do not vary nearly
as much. The human visual system is constantly evaluating the illumination en-
vironment and adapting its behavior accordingly. We can think of this process
as the biological equivalent of the automatic white balancing available in today’s
digital cameras and camcorders (much as we can think of the automatic exposure
adjustment akin to our light and dark adaptation).
This chapter details the process of chromatic adaptation, from examples and
experimental data to computational models that predict human perception. A
model for predicting corresponding colors, the stimuli necessary to generate a
chromatic response under one lighting condition that is identical to that under
a different lighting condition, is presented. These corresponding colors form
an important component of today’s color management systems. This chapter
also outlines how we can use the computational models of chromatic adapta-
tion in a digital-imaging system to form a basis for an automatic white-balancing
algorithm.

10.1 Changes in Illumination

The wide range of viewing conditions that we may encounter in everyday life
is staggering, and yet the diversity of conditions typically goes unnoticed. This
is, in part, because our visual system is very adept at normalizing the viewing
conditions and presenting an ultimate visual experience that is fairly consistent.
Due to this normalization behavior, we often see the visual system described as
being color constant, suggesting that object colors appear the same regardless of
their illuminating conditions.
While the human visual system does provide some form of constancy to our
everyday experiences, we are by no means color constant. Consider looking out-
side on a bright sunny day compared to looking outside on a cloudy overcast day.
On the sunny day, all of the colors appear considerably more colorful, while they

i i

i i
 i i

i i

10.1. Changes in Illumination 527

appear less chromatic on the cloudy day. Although the appearances have changed,
we do not assume that all of the objects themselves have actually changed phys-
ical properties. Rather we recognize that the lighting has influenced the overall
color appearance.
This change in color appearance has been carefully studied and quantified and
is described in detail in Chapter 11. Artists also often have a keen eye to these
overall changes in appearance caused by illumination. Claude Monet, in particu-
lar, was very fond of painting the same scene under a wide variety of illumination.
His Haystack series and Rouen Cathedral paintings of the late 19th century are ex-
cellent examples. Although the actual contents of his paintings did not vary, he
was able to capture an incredible variety of colors by painting at a variety of times
of day and seasons. Interestingly enough, the actual appearance of these scenes
may actually have looked more similar to him while he was painting through his
own chromatic adaptation.
Although we are not fully color constant, the general concept of constancy
does apply to chromatic content. A powerful example of chromatic adaptation
is that most object colors do not significantly change appearance when viewed
under different lighting conditions. This is perhaps best illustrated by thinking
of reading a magazine or newspaper in different areas of your office or home.
Under these conditions, you may have incandescent (tungsten) illumination, flu-
orescent lighting, and natural daylight through a window. Perhaps, you may also
have candlelight or a fire providing some illumination. Throughout these different
conditions, the appearance of the magazine paper will still be white. This is espe-
cially impressive considering the color of the illumination has gone from very red
for the fire, toward yellow for the tungsten, and blue for the daylight. Our visual
system is continuously adapting to these changes, allowing the paper to maintain
its white appearance.
Another example of the vast changes in illumination can be seen in Fig-
ure 10.1, where we have the same scene photographed under very different con-
ditions. The photographs in this figure were RAW digital camera images (i.e.,
minimally processed sensor data) with no automatic white balancing, so they pro-
vided an excellent approximation of the spectral radiances present in the scene.
We can see that at sunset the color of the illumination was predominantly orange,
while in the morning the color predominantly very blue.
Another interesting example is illustrated in Figure 10.2, where the wall of
a building is illuminated by direct sunlight, with a bright blue sky behind it. In
the photograph the wall looks very yellow, a perception that is enhanced by the
simultaneous contrast of the blue sky behind it. In reality, the wall is in fact white,
and the coloration is due entirely to the illumination. To an observer looking at

i i

i i
 i i

i i

528 10. Chromatic Adaptation

Figure 10.1. An example of the same scene viewed under two very different illuminating
conditions.

Figure 10.2. A “white” building illuminated by direct sunlight.

the building it would be obvious that the building was indeed white, though the
appearance of the wall would retain a slight yellow tint. In other words, if the
person were asked to choose a paint to coat another wall such that it matched
the building, they would easily select a white paint, but if they were asked to
artistically reproduce the appearance of the building, they would select a yellow
paint. Similarly to Monet and the Rouen Cathedral, they would recognize that
although the building itself was white, the overall visual experience was actually
yellow.
We have seen how the same scene can contain a huge variation in the color of
lighting depending on the time of day it is viewed. Weather itself can also play

i i

i i
 i i

i i

10.1. Changes in Illumination 529

Figure 10.3. A building photographed under two very different weather conditions.
(Photo courtesy of Ron Brinkmann.)

a large role in changing our everyday viewing conditions. Figure 10.3 shows the
same building photographed on a relatively clear day, as well as on a foggy day.
The building itself appears much more blue when illuminated on the clear day
than it does on the foggy day. Despite these large changes in overall illumination,
we do not think that the building itself is actually changing color. Chromatic
adaptation accounts for much of the visual system’s renormalization, allowing us
to maintain a relatively large degree of color constancy.
Essentially, chromatic adaptation allows for a large number of lighting condi-
tions to appear “white.” Perhaps a better explanation is to say that a white paper
is still recognized as being white when illuminated under lights of a wide range
of chromaticities. For example, in Figure 10.4 we see a variety of real-world
lights plotted in a CIE u v chromaticity diagram. Also plotted in this diagram
are the CIE Daylight coordinates. The colors of the points plotted in Figure 10.4
are the relative chromaticities when viewed in the sRGB color space with a white
point of CIE D65. All of these light sources appear white to a person viewing
them if they are viewed in isolation, due to chromatic adaptation. This is also
why it is of critical importance to not use chromaticity coordinates as a method
for describing color appearance! The color appearance of all of the light sources
plotted in Figure 10.4 would most likely all be described the same, despite their
markedly different chromaticities. Further, this is why it is important to not show

i i

i i
 i i

i i

530 10. Chromatic Adaptation

0.5

0.4

0.3

0.2

0.1

0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
u

Figure 10.4. A series of real-world light sources and the CIE Daylight illuminants plotted
in the CIE u v chromaticity diagram. A white piece of paper illuminated by any of these
light sources maintains a white color appearance.

chromaticity diagrams in color, since chromatic adaptation itself will change the
appearance of any given point in a chromaticity space.

10.2 Measuring Chromatic Adaptation

Chromatic adaptation can be considered a form of normalization that occurs in
the visual system, accounting for changes in the chromatic content of the viewing
conditions. Ideally, we would like to create a computational model that can pre-
dict these changes for use in color-imaging systems. For example, we may want
to know what colors to send to our printer in order to generate a match between
an original viewed under one lighting condition and the reproduction that will be
viewed under a different condition. Recall from Chapter 7 that colorimetry pro-
vides us with a method for determining if two colors match, with strict conditions
that the viewing conditions of the two colors are identical. Models of chromatic
adaptation can be used to extend colorimetry to predict visual matches when the
viewing conditions are under different colored illumination.
In order to develop these models, it is important to measure the actual behavior
of the chromatic adaptation of our visual system. Historically, several different
types of psychophysical experiments have been used to test chromatic adaptation.
Two types have been used with particular success: color matching and magnitude
estimation. These experiments often generate what are known as corresponding

i i

i i
 i i

i i

10.2. Measuring Chromatic Adaptation 531

colors, or stimuli that under one illumination “match” other stimuli viewed under
different illumination.
Color-matching experiments are similar in nature to those used to generate
the CIE standard observers. Essentially, stimuli are viewed under one condition
and then a match is identified (or generated) under a different viewing condition.
This can be done either through successive viewing or simultaneously using a
haploscopic approach, as explained below.
In successive-viewing experiments, an observer looks at a stimulus under one
illuminating condition and then determines a match under a different condition.
Time must be taken to adapt to each of the viewing conditions. For example, an
observer may view a color patch illuminated under daylight simulators for a pe-
riod of time and then, in turn, determine the appearance match with a second stim-
ulus illuminated with incandescent lights. Prior to making any match (or creating
a match using the method of adjustment), the observer has to adapt to the second
viewing condition. Fairchild and Reniff determined that this time course of chro-
matic adaptation should be about 60 seconds to be 90 percent complete [296]. It
should be noted that this time course was determined for viewing conditions that
only changed in chromatic content and not in absolute luminance levels, and more
time may be necessary for changes in adapting luminance. In viewing situations
at different luminances and colors, we would generally say that we are no longer
determining corresponding colors based upon chromatic adaptation but rather full
color-appearance matches. Details on models of color appearance will be found
in Chapter 11.
The inherent need to delay the determination of a color match to allow for
and to study chromatic adaptation leads some to coin this experimental technique
memory matching. This need to rely on human memory to determine a color
match adds inherent uncertainty to these measurements, when compared to di-
rect color matching such as those found using a bipartite field. Wright suggested
using a form of achromatic color matching to alleviate some of these uncertain-
ties [1255]. An achromatic color matching experiment involves determining or
generating a stimulus that appears “gray” under multiple viewing conditions. De-
termining when a stimulus contains no chromatic content should not involve color
memory and should provide a direct measurement of the state of chromatic adap-
tation. Fairchild used this technique to measure chromatic adaptation for soft-
copy (e.g., CRT) displays [302].
Another technique that attempts to overcome the uncertainties of memory
matching is called haploscopic color matching. In these experiments the observer
adapts to a different viewing condition in each eye and makes a color match
while simultaneously viewing both conditions. These matches are made under
the assumption that our state of chromatic adaptation is independent for each eye.

i i

i i
 i i

i i

532 10. Chromatic Adaptation

Fairchild provides an excellent overview of some of the historical experiments

and devices used for these types of experiments [298].
Color matching is not the only technique that can be used to generate corre-
sponding color data to study chromatic adaptation. Magnitude estimation has also
been used with success to generate such data, most noticeable the LUTCHI data
measured by Luo et al. [711, 712]. Magnitude estimation has observers generat-
ing descriptions of the overall color appearance of stimuli under multiple view-
ing conditions. Essentially, observers are asked to assign numerical values along
color-appearance attributes such as lightness, chroma, and hue. In these types
of experiments, it is important to adequately train the observers so that they un-
derstand and can assign ratings to each of the attributes. These color-appearance
attributes themselves are described later in Chapter 11.
All of these chromatic-adaptation experiments are designed to generate sets
of colors (often represented as tristimulus values or chromaticities) that appear to
match under different viewing conditions. If there were no chromatic adaptation,
we may expect that the tristimulus values would always be the same regardless
of the illuminating conditions. Due to the renormalization caused by chromatic
adaptation, what we generally see is that the color matches tend to “drift” toward
the chromaticity of the adapting conditions themselves. This is another way of
saying that our adapting white point, or the tristimulus values we consider to be
achromatic, itself shifts as a function of the viewing conditions. Computational
models that predict the degree of this shift are discussed later in this chapter.

10.3 Mechanisms of Chromatic Adaptation

In an ideal situation the models used to predict the behavior of chromatic adap-
tation would be based upon actual mechanisms of the human visual system. Al-
ternatively, empirical models can be fit to the measured corresponding color data.
These empirical models could be described as a type of black-box approach to
chromatic adaptation. For a digital-imaging system, the black-box approach may
be adequate, though it does not provide any insight into what is actually happen-
ing within the visual system with regard to chromatic adaptation. The black-box
modeling is also very difficult to modify or fix if there is poor performance. In
practice, the computational models we use should probably be considered a hy-
brid approach, as they can be considered based on our understanding of the visual
system but also designed to fit the existing experimental data.
While the experimental techniques described in the previous section can pro-
vide insight into the ultimate behavior of the human visual system, they do not

i i

i i
 i i

i i

10.3. Mechanisms of Chromatic Adaptation 533

provide much information with regard to the underlying causes of the behavior.
In fact, the full mechanisms of chromatic adaptation in the visual system are not
yet understood. Typically the mechanisms are often broken into two categories:
physiological and cognitive. Physiology itself may be too broad a description, as
that could include the cognitive processes themselves. Perhaps a better descrip-
tion, as used by Fairchild, would be sensory and cognitive mechanisms [302].
Sensory mechanisms refer to those that occur automatically as a function of
the physics of the input illumination. An example of this type of mechanism
would be the automatic dilation of the pupils as a function of the overall amount
of illumination. Other examples include the transition from rods to cones as a
function of input luminance as well as gain-control mechanisms on the photore-
ceptors themselves. Gain control will be discussed in more detail below. While
the sensory mechanisms may be automatic, they are not instantaneous. In fact, the
60-second adaptation time necessary for the memory-matching experiments de-
scribed above is necessary to allow time for the sensory mechanisms [296]. Cog-
nitive mechanisms can be considered higher-level mechanisms where the brain
actually alters perception based upon prior experiences and understanding of the
viewing conditions.
While pupil dilation and the transition from the rods to the cones plays a large
role in our overall light and dark adaptation, these sensory mechanisms do not
play a large role in chromatic adaptation. Although some imaging systems are
designed for use at low light levels, for most common situations, the rods play a
very small role in chromatic adaptation and color appearance. We will focus in-
stead on the role of the cones, specifically, on gain control of the cone photorecep-
tors. One of the simplest descriptions of and explanation for chromatic adaptation
can be thought of as an automatic and independent re-scaling of the cone signals
based upon the physical energy of the viewing condition. This theory was first
postulated by von Kries in 1902 and still forms the basis of most modern models
of chromatic adaptation. In terms of equations, this von Kries-type adaptation is
shown below:

La = α L, (10.1a)
Ma = β M, (10.1b)
Sa = γ S, (10.1c)

where L, M, and S are the cone signals, and α , β , and γ are the independent gain
components that are determined by the viewing condition. While we know today
that the hypothesis of von Kries is not entirely true, it is still remarkable how
effective such a simple model can be for predicting chromatic adaptation.

i i

i i
 i i

i i

534 10. Chromatic Adaptation

Figure 10.5. An example of sensory chromatic adaptation caused by photoreceptor gain

control.

Evidence of the receptor gain control can be seen in the after-images that we
encounter in everyday life. An example of this is shown in Figure 10.5. Focusing
on the + symbol in the top frame of Figure 10.5 will cause half of the cones
on your retina to adapt to the red stimulus while the other half will adapt to the
green stimulus (interestingly enough, due to the optics of the eye, the left half of
each retina will adapt to the green while the right half will adapt to the red). After

i i

i i
 i i

i i

10.3. Mechanisms of Chromatic Adaptation 535

focusing on the top frame for about 60 seconds, one should move one’s gaze to the
bottom frame. The image shown on the bottom should now look approximately
uniform, despite the fact that the left side is tinted red and the right side is tinted
green. This is because the photoreceptors in the retina have adapted to the red and
green fields. Staring at a blank piece of paper should result in an after-image that
is opposite to the adapting field (green/cyan on the left and reddish on the right).
See also Figure 5.20 for another example.
Another way to think about the von Kries receptor gain control is to consider
that the cones themselves become more or less sensitive based upon the physical
amount of energy that is present in a scene. For instance, if a scene is predomi-
nantly illuminated by “blue” light, then we would expect the S cones to become
less sensitive, while the L cones would become (relatively) more sensitive. Con-
versely, if there is more red energy in the illumination, such as with an incandes-
cent light bulb, then we would expect the short cones to become relatively more
sensitive than the long cones.
This example is illustrated in Figure 10.6. The spectral power distribution
(SPD) of incandescent illumination (e.g., CIE Illuminant/Source A; see Section
9.1.1) is illustrated by the green curve in Figure 10.6, while the individual cone

Solid lines: relative cone responses

Dashed lines: relative adapted cone responses
1.0
Sensitivity

CIE A
0.9
Color representing
0.8 CIE A rendered into
0.7 the sRGB color space
L
M
0.6
S
0.5
0.4
0.3
0.2
0.1
0.0
400 450 500 550 600 650 700
Wavelength (nm)

Figure 10.6. An example of von Kries-style independent photoreceptor gain control.

The relative cone responses (solid line) and the relative adapted cone responses to CIE A
(dashed) are shown. The background color represents CIE A rendered into the sRGB color
space.

i i

i i
 i i

i i

536 10. Chromatic Adaptation

functions are shown with solid and dashed lines. The relative cone responses
prior to chromatic adaptation are shown as solid lines, while the adapted cones are
represented by the dashed lines. The patch of color in Figure 10.6 represents the
spectral power distribution of CIE A approximately rendered into the sRGB color
space with a white point of CIE D65; that is, the tristimulus values of the printed
background when viewed under daylight illumination should be very similar to a
blank page viewed under incandescent lighting.
Clearly, we can see that the patch of color in Figure 10.6 is a deep yellow-
orange. The blank paper illuminated by incandescent light will not appear nearly
as orange due to the effects of chromatic adaptation. Normalizing the photore-
ceptors independently as a function of the SPD of the light-source results in the
dashed lines shown in Figure 10.6. If we were to then go ahead and integrate
those new adapted cone responses with the background SPD, we would see that
the relative integrated responses are about equal for all the cones, and thus the
adapted perceptual response is approximately white. Alternative models based
upon the von Kries-style of receptor gain control will be discussed in the next
section.
Receptor gain control is not the only form of sensory adaptation for which
we need to account. The visual system, as a generality, is a highly nonlinear
imaging device. As a first-order assumption, many often say that the visual system
behaves in a compressive logarithmic manner. While this is not entirely true,
it does describe the general compressive nature of the visual system (compare
Figure 17.9). Models of chromatic adaptation can include nonlinear compression
as well as subtractive mechanisms. The nonlinear compression can take the form
of simple power functions as well as more sophisticated hyperbolic or sigmoidal
equations. However, for speed and simplicity, most current adaptation models
rely on simple linear gain controls. More sophisticated color appearance models,
such as CIECAM02, do contain these nonlinear components.
Although independent photoreceptor gain controls cannot account entirely
for chromatic adaptation, models based upon this theory can still be very use-
ful. From a purely physiological standpoint, we can consider photoreceptor gain
control as a natural function of photopigment depletion, or bleaching, and regen-
eration. At a fundamental level, color begins when a photon of light hits a pho-
topigment molecule in the cones. The probability of the photopigment molecule
absorbing the light is determined by both the wavelength of the photon and the
properties of the molecules, which are slightly different for the photopigments in
each of the cone types.
When a photopigment molecule absorbs a photon, it undergoes a chemical
transformation known as isomerization, which eventually leads to the transmis-

i i

i i
 i i

i i

10.3. Mechanisms of Chromatic Adaptation 537

Figure 10.7. Example of a cognitive form of chromatic adaptation. In the middle frame,
the yellow candy has a green filter placed over it causing its appearance to shift towards
green. When the same green filter is placed over the entire image (bottom frame) the yellow
candy appears yellow again (adapted from [509]).

i i

i i
 i i

i i

538 10. Chromatic Adaptation

sion of an electrical current. This process essentially depletes the number of avail-
able photopigment molecules for further light absorption. If all of the molecules
in a particular receptor are depleted, then that receptor is said to be bleached.
Until that point, however, we can say that with fewer and fewer photopigment
molecules available, that particular cone will become less and less sensitive to
light. Interested readers are encouraged to find more information on this process
in vision science texts such as Human Color Vision [566] or Color Vision: From
Genes to Perception [353]. See also Section 3.4.2.
Cognitive mechanisms of chromatic adaptation are more difficult to model, or
even comprehend. It is somewhat counter-intuitive to think that our knowledge
and understanding of a particular scene can fundamentally alter our perception
of that scene, though that is indeed the case. An example of this can be seen in
Figure 10.7. We see the same image with differently colored filters placed over
it. In the top frame, we see the original image, while in the middle frame a green
filter has been cut and placed over the yellow candy. We can see that the color
appearance of the candy has been changed drastically, and now it actually looks
more green than its neighbor.
When we instead place the same green filter over the entire image, as shown in
the bottom frame of Figure 10.7, the appearance of the same candy reverts back to
yellow. This is because our brain recognizes that the color change is a result of a
change in illumination, through the information present in the entire image. This
example is based upon a classic demonstration by Hunt [509]. Cognitive effects of
chromatic adaptation are based upon our understanding of the viewing conditions
as well as past experiences, and they can be considered a higher-level form of
adaption than the pure sensory mechanisms. These higher-level mechanisms do
not necessarily follow the same time course of adaptation either and can be almost
instantaneous. Again, this can be seen in Figure 10.7 when glancing between the
bottom and middle frame. While the color of the “yellow” candy itself does not
change between the two images, the overall appearance changes immediately.

10.4 Models of Chromatic Adaptation

Understanding the behavior and some of the underlying mechanisms of chromatic
adaptation allows us to generate models that can better predict our ultimate per-
ception to a given color stimulus. In Chapter 7, we learned that colorimetry can
be a very powerful tool for specifying color matches, or predicting if two differ-
ent stimuli will appear the same. While colorimetry itself is incredibly powerful,
there are some very strict limitations. A colorimetric match does not tell us what

i i

i i
 i i

i i

10.4. Models of Chromatic Adaptation 539

a color looks like, nor can it tell us if two stimuli match when viewed under
disparate lighting conditions. Chromatic-adaptation models get us closer to be-
ing able to describe the appearance of a stimulus and form a fundamental part
of a more complete color-appearance model, as will be explained in Chapter 11.
Chromatic-adaptation models can also be used to extend colorimetry to predict if
two stimuli will match if viewed under disparate lighting conditions where only
the color of the lighting has changed. This section will outline several computa-
tional models of chromatic adaptation, including the von Kries model, CIELAB,
as well as the underlying adaptation model in CIECAM02.

10.4.1 von Kries Chromatic-Adaptation Model

As described in the previous section, a simple hypothesis of chromatic adaptation
was postulated by von Kries in the early 1900s. He theorized that each of the
photoreceptors, or “organs of vision,” were fatigued or adapted independently of
one another [623]. While we now know that this is not entirely true, his theory
still forms the basis for many adaptation models today. It is important to note that
von Kries did not supply any mathematical formulations to accompany his theory,
though the general form was shown in Equation (10.1). For digital-imaging appli-
cations, these equations are often reworked into simple linear matrix calculations:
⎡ ⎤ ⎡ ⎤⎡ ⎤
La α 0 0 L
⎣Ma ⎦ = ⎣ 0 β 0⎦ ⎣M ⎦ . (10.2)
Sa 0 0 γ S

In practice, we traditionally replace α , β , and γ with the reciprocal of the

adapting white, which would also be expressed in terms of LMS cone signals.
This forms the basis of the von Kries chromatic-adaptation model, as each cone
signal is adapted based upon the cone signals corresponding to the prevailing
illumination (the illumination of an object considered to be white), giving
⎡ ⎤
1
⎡ ⎤ ⎢L 0 0 ⎥⎡ ⎤
La ⎢ w ⎥ L
⎣Ma ⎦ = ⎢ 1 ⎥
⎢ 0 0 ⎥ ⎣M ⎦ . (10.3)
⎢ Mw ⎥
Sa ⎣ 1 ⎦ S
0 0
Sw
We can now apply the principles of colorimetry to determine if two colors
match under differently colored illumination by first calculating their adapted
cone responses. If LMSa1 = LMSa2 , then we can say that the two colors will
match. We can determine the adapted CIE tristimulus values by cascading a lin-
ear transform from XYZ to LMS prior to calculating the chromatic adaptation

i i

i i
 i i

i i

540 10. Chromatic Adaptation

and then transforming back to XYZ after adaptation. One commonly used linear
transform from XYZ to LMS and its inverse are given by
⎡ ⎤ ⎡ ⎤
L X
⎣M ⎦ = MHPE ⎣Y ⎦ (10.4a)
S Z
⎡ ⎤⎡ ⎤
0.38971 0.68898 −0.07868 X
= ⎣−0.22981 1.18340 0.04641⎦ ⎣Y ⎦ ; (10.4b)
0.00000 0.00000 1.00000 Z
⎡ ⎤ ⎡ ⎤
X L
⎣Y ⎦ = M −1 ⎣M ⎦ (10.4c)
HPE
Z S
⎡ ⎤⎡ ⎤
1.91019 −1.11214 0.20195 L
= ⎣ 0.37095 0.62905 0.00000⎦ ⎣M ⎦ . (10.4d)
0.00000 0.00000 1.00000 S

This transform is known as the Hunt-Pointer-Estevez transform [509]. Here, the

cone sensitivities are normalized such that they produce an equal LMS value when
presented with an equal-energy illuminant, or equal XYZ tristimulus values.
By combining (10.3) and (10.4), we can calculate the adapted tristimulus val-
ues for a given viewing illumination or white point:
⎡ ⎤ ⎡ ⎤
Lw Xmax
⎣Mw ⎦ = MHPE ⎣Ymax ⎦ ; (10.5a)
Sw Zmax
⎡ ⎤
1
⎡ ⎤ ⎢ Lw 0 0 ⎥ ⎡ ⎤
Xadapted ⎢ ⎥ X
⎢
⎣Yadapted ⎦ = M −1 ⎢ 0 1 ⎥ ⎣Y ⎦ .
HPE ⎢ 0 ⎥ MHPE (10.5b)
Mw ⎥
Zadapted ⎣ 1 ⎦ Z
0 0
Sw

The adapting illumination can be either measured off of a white surface in the
scene, ideally a perfect-reflecting diffuser, or alternatively for a digital image the
illumination can often be approximated as the maximum tristimulus values of
the scene, as shown in (10.5a). Since both the forward and inverse cone trans-
forms are linear 3 × 3 matrix equations, as is the von Kries normalization, the
entire chromatic adaptation transform can be cascaded into a single 3 × 3 matrix
multiplication.

i i

i i
 i i

i i

10.4. Models of Chromatic Adaptation 541

Now we know if the chromatically-adapted tristimulus values of two stimuli

match, i.e., XY Z1adapt = XY Z2adapt , then the stimuli themselves should match.
While this in itself can be useful, it is often more desirable to calculate what stim-
ulus needs to be generated under one illumination to match a given color under a
different illumination. For example, you may have a color patch illuminated by
daylight, and you want to know what tristimulus values you need to generate to
create a “matching” color patch that will be illuminated by an incandescent light.
This corresponding-colors calculation involves a round-trip or forward and in-
verse chromatic-adaptation calculation. The essence of von Kries chromatic adap-
tation is a normalization of the cones based upon the integrated energy (adapting
white) of a given illumination condition. We calculate corresponding colors by
first normalizing for one illumination and then setting these adapted signals to
the second condition. This amounts to dividing out the color of one light and
multiplying in the color of the second:
⎡ ⎤ ⎡ ⎤⎡ 1 ⎤ ⎡ ⎤
X Lw,2 0 0 Lw,1 0 0 X
⎣Y ⎦ −1 ⎣
= MHPE 0 Mw,2
⎢
0 ⎦⎣ 0
1
Mw,1 0 ⎥
⎦ MHPE ⎣Y ⎦ .
Z adapted 0 0 Sw,2 1 Z
0 0 Sw,1
(10.6)
Again, as all of these calculations are simple linear 3 × 3 matrix transforms, they
can all be cascaded into a single 3 × 3 transform for simple application to digital
images.
We can use a von Kries-type adaptation to renormalize the Hunt-Pointer-
Estevez transform itself such that it provides equal LMS values when presented
with D65 illumination. We first calculate the “normalizing” white LMS values
using Equation (10.5a) and then use those values in a von Kries adaptation on the
cone signals:
⎡ ⎤ ⎡ ⎤
L 0.9504
⎣M ⎦ = MHPE ⎣1.0000⎦ ; (10.7a)
S D65 1.0889 XYZD65
⎡ 1 ⎤
LD65 0 0
⎢ 1
0 ⎥
MHPED65 = ⎣ 0 MD65 ⎦ MHPE (10.7b)
1
0 0 SD65
⎡ ⎤
0.4002 0.7076 −0.0808
= ⎣−0.2263 1.1653 0.0457⎦ . (10.7c)
0.0000 0.0000 0.9184

i i

i i
 i i

i i

542 10. Chromatic Adaptation

The simple von Kries adaptation transform has been shown to provide a re-
markably good fit to corresponding experimental color data, such as the data gen-
erated by Breneman in 1987 [126] and Fairchild in the early 1990s [302]. Al-
though it is remarkable that such a simple model could predict the experimental
data as well as the von Kries model does, there were some discrepancies that
suggested that a better adaptation transform was possible.

10.4.2 CIELAB as a Chromatic-Adaptation Model

The CIELAB color space was introduced in Chapter 8, and provides a close-to-
perceptually-uniform opponent color space that can also be used to derive color
differences between two stimuli. While we do not generally describe the CIELAB
model as a chromatic-adaptation model, there are elements of chromatic adapta-
tion that can be considered. The CIELAB equations are presented again here, in
a slightly different form, for clarity:
 
∗ Y
L = 116 f − 16, (10.8a)
Yn
    
∗ X Y
a = 500 f −f , (10.8b)
Xn Yn
    
Y Z
b∗ = 200 f −f , (10.8c)
Yn Zn

r1/3 r > 0.008856,
f (r) = (10.8d)
7.787 r + 16/116 r ≤ 0.008856.

Notice that for each of the CIELAB equations, the CIE XYZ tristimulus val-
ues are essentially normalized by the tristimulus values of white. This normal-
ization is in essence a von Kries-type normalization, though for CIELAB this is
performed on the actual tristimulus values rather than on the LMS cone signals.
This type of transform was actually coined a wrong von Kries transform by Ter-
stiege [1122].
Interestingly, one of the known weaknesses of the CIELAB color space, which
was designed to be perceptually uniform, is changing hue when linearly increas-
ing or decreasing chroma (these terms are explained in depth in Chapter 11). That
is so say that lines of constant perceived hue are not straight in the CIELAB color
space, most noticeably in the blue and red regions, as quantified by experiments
such as those by Hung and Berns [500]. This often results in the purple-sky phe-
nomenon when boosting chroma in the CIELAB color space. An example of this

i i

i i
 i i

i i

10.4. Models of Chromatic Adaptation 543

Figure 10.8. The image on the left was modified by boosting its chroma in CIELAB color
space by a factor of two, shown on the right. Note the undesired purple color which appears
in parts of the water. The blockiness of the purple areas are JPEG compression artifacts;
Grand Bahama, 2004.

is shown in Figure 10.8, which shows an image before and after boosting chroma
by a factor of two.
In Figure 10.9 the Munsell renotation data is plotted in CIELAB for a given
lightness value. Since the Munsell data was experimentally generated to be con-
CIELAB b*

80

40

-40

-80

-120
-150 -100 -50 0 50 100 150
CIELAB a*
Figure 10.9. The Munsell Renotation data at Value 4 plotted in the CIELAB color space
showing curvature for the lines of constant hue.

i i

i i
 i i

i i

544 10. Chromatic Adaptation

stant hue, the “spokes” corresponding to lines of constant hue should be straight
[831]. Thus, the root cause of this type of color error was linked to the chromatic-
adaptation transform using CIE XYZ instead of LMS in evaluations by Liu et al.
in 1995 and Moroney in 2003 [690, 793].

10.4.3 Nayatani Chromatic-Adaptation Model

Further advances in modeling chromatic adaptation were developed by Nayatani
et al. in the early 1980s [819, 820]. In addition to a von Kries-style photorecep-
tor gain control, they added a nonlinear exponential term as well as an additive
noise term to handle low-light level adaptation for near threshold visual data. The
generalized form of this chromatic adaptation model is given by:
 
L + Ln βL
La = αL , (10.9a)
Lw + Ln
 
M + Mn βM
Ma = αM , (10.9b)
Mw + Mn
 
S + Sn βS
Sa = αS . (10.9c)
Sw + Sn

Here, LMSw are the cone excitations for the adapting field, or white point, LMSn
are the noise coefficients, typically a very small number, βLMS are the nonlinear
exponential values, and αLMS are linear scaling factors. The von Kries-style pho-
toreceptor gain control should be immediately obvious in the Nayatani equations,
especially as the noise coefficients approach zero.
The exponential values βLMS are functions of the adapting cone signals them-
selves, essentially meaning they monotonically increase as a function of the adapt-
ing luminance [820]. By combining both a photoreceptor gain control with the
nonlinear exponential functions, the Nayatani et al. model is able to predict cor-
responding colors caused by chromatic adaptation, as well as color appearance
changes caused by changes in overall luminance level. These include the increase
in apparent colorfulness and contrast as luminance levels increase. More details
on these color appearance changes are discussed in Chapter 11.

10.4.4 Fairchild Chromatic-Adaptation Model

The chromatic-adaptation model of Nayatani et al. illustrated the power of com-
bining the simple von Kries-style linear photoreceptor gain control with addi-
tional components to create an extended model of chromatic adaptation capable

i i

i i
 i i

i i

10.4. Models of Chromatic Adaptation 545

of predicting additional color-appearance phenomena. Fairchild presented a sim-

ilar extension designed for digital-imaging applications [299, 300]. The goal of
this model is to have a linear chromatic-adaptation transform that is capable of
predicting incomplete adaptation as well as the increase in colorfulness with lu-
minance level, e.g., the Hunt effect as discussed in Chapter 11.
To this point, all of the models of chromatic adaptation we have discussed as-
sume that we adapt completely to the reference viewing conditions. This implies
that the adapting “white” will always appear white. The experiments of Brene-
man as well as further experiments by Fairchild showed that this is not always the
case [126, 302].
Fairchild performed a series of experiments using a soft-copy CRT display to
test the degree of chromatic adaptation and found that particularly for soft-copy
displays, our adaptation is significantly less than complete. A simple example of
this can be illustrated by setting a computer display to have a “white” of equal
tristimulus values as CIE Illuminant A. The display itself will continue to appear
very orange, even after significant adaptation time. Furthermore, if we place a
piece of paper such that it is illuminated by the display, we perceive the paper
to appear white, or more white, while the display itself, which is the only source
of illumination for the paper, does not! This is illustrated in Figure 10.10. The

Figure 10.10. An example of discounting the illuminant: the laptop screen and the paper
have identical chromaticities, yet the laptop appears more orange.

i i

i i
 i i

i i

546 10. Chromatic Adaptation

computer screen has the same approximate chromaticity as the paper, though the
latter appears to be much more white. For demonstrative purposes, the photograph
in Figure 10.10 was imaged with an additional tungsten light source, though the
chromaticity of the lighting was closely matched to the laptop screen.
The fact that two objects with identical viewing conditions and tristimulus
values do not match in appearance suggests that this difference cannot be at the
sensory level and must rather be at the cognitive level. In color-appearance ter-
minology, one such cognitive mechanism is often referred to as discounting the
illuminant [302, 509]. In such cases, the observer recognizes that the color is a
combination of the object color itself and the illumination. Thus, in the viewing
conditions shown in Figure 10.10, the observer recognizes that the paper itself is
white while the laptop screen is providing the orange illumination. In the same
situation, the self-emitting laptop screen is not recognized as a white object with
orange illumination, and so it retains its overall orange cast.
Fairchild aimed to extend traditional colorimetry along with the von Kries
photoreceptor gain control to create a model that was capable of predicting this
incomplete chromatic adaptation. The model begins by converting the XYZ tris-
timulus values into LMS cone responses, using the Hunt-Pointer-Estevez trans-
formation normalized to CIE D65, from Equations 10.7c and 10.10b. The cone
signals then have a von Kries-style linear adaptation applied to them, as follows:

⎡ ⎤ ⎡ ⎤
L1 X1
⎣M1 ⎦ = MHPED65 ⎣Y1 ⎦ (10.10a)
S1 Z1
⎡ ⎤⎡ ⎤
0.4002 0.7076 −0.0808 X1
= ⎣−0.2263 1.1653 0.0457⎦ ⎣Y1 ⎦ ; (10.10b)
0 0 0.9184 Z1
⎡ ⎤ ⎡ ⎤
L1 L1
⎣M1 ⎦ = A ⎣M1 ⎦ (10.10c)
S1 S1
⎡ ⎤⎡ ⎤
aL 0 0 L1
= ⎣ 0 aM 0 ⎦ ⎣M1 ⎦ . (10.10d)
0 0 aS S1

The calculation of the matrix A is what sets the Fairchild chromatic-adaptation

transform apart from a traditional von Kries model. The coefficients aLMS are

i i

i i
 i i

i i

10.4. Models of Chromatic Adaptation 547

calculated using
PL
aL = , (10.11a)
LW
1 +Ywv + lE
PL = , (10.11b)
1 +Ywv + l1E

3 LLWE
lE = LW
. (10.11c)
LE +M SW
ME + SE
W

The computations of aM and aS proceed similarly. In (10.11), Ywv is the lu-

minance of the adapting white stimulus in cd/m2 with a 1/3 exponent applied,
LMSw are the cone signals of the adapting stimulus, and LMSE are the cone sig-
nals for an equal energy spectrum. What we should notice from these calculations
is that the degree of adaptation, essentially the P coefficient, is a function of both
luminance and color of the adapting white.
Essentially, as the absolute luminance level of the adapting stimulus increases,
the difference between lE and l1E becomes negligible and P approaches one, which
makes this transform identical to a standard von Kries transform. For lower lu-
minance levels, the lE terms become significant and P becomes a value less than
one. The lmsE terms decrease as the color of the adapting stimulus varies from
the equal-energy spectrum, suggesting that the further the adapting white is from
equal energy the less we will adapt to it. When the adapting stimulus is equal
energy, these terms drop out and again we have a traditional von Kries chromatic-
adaptation transform. The P term can also be manually set to 1.0 when we know
that cognitive discounting of the illuminant occurs, or can be manually set to lower
values when we know that it does not occur, such as when viewing a soft-copy
device such as a CRT.
The Fairchild chromatic-adaptation model originally included another trans-
formation to get into the post-adaptation cone signals:
⎡ ⎤ ⎡ ⎤
La L1
⎣Ma ⎦ = C ⎣M1 ⎦ (10.12a)
Sa S1
⎡ ⎤⎡  ⎤
1.0 c c L1
= ⎣ c 1.0 c ⎦ ⎣M1 ⎦ , (10.12b)
c c 1.0 S1

where
c = 0.219 − 0.0784 log10 (YW ) . (10.13)

i i

i i
 i i

i i

548 10. Chromatic Adaptation

This transformation adds a luminance-dependent interaction between all of the

cone signals to account for the boost in colorfulness when the adapting illu-
mination is high. In practice, this C-matrix calculation introduced undesirable
luminance-dependent artifacts in the predicted results and is generally eliminated
[302].
Since the Fairchild chromatic-adaptation model was designed to be used on
digital images, it was designed to collapse to a single 3 × 3 matrix transform much
like the traditional von Kries model. Thus, for calculating corresponding colors
for two different viewing conditions, the entire calculation can be expressed using
⎡ ⎤ ⎡ ⎤
X X
⎣Y ⎦ = M −1 A−1 C−1 C1 A1 MHPE ⎣Y ⎦ . (10.14)
HPE 2 2
Z 2 Z 1
It should be noted that the A and C calculations only need to be performed once for
each viewing condition, and then the entire transform can be applied to each pixel
in a digital image. Pixel-specific chromatic-adaptation transforms that change as a
function of the image content itself are an important element of image-appearance
models and will be discussed in Chapter 11. To calculate corresponding colors
using the Fairchild model, the forward A and C matrices must be determined for
both of the viewing conditions and then inverted for the second condition to obtain
A−1 −1
2 , C2 .

10.4.5 The Bradford and CIECAM97s Chromatic-Adaptation Models

The Bradford transform, first introduced to color-appearance modeling in the
LLAB model, is a von Kries-style adaptation transform that includes a nonlin-
ear component on the short-wavelength cone signals [641, 713]. This adaptation
transform also includes a calculation for discounting the illuminant, modeling a
form of incomplete chromatic adaptation. Perhaps the most important advance
introduced by the Bradford transform is the use of sharpened cone signals in-
stead of real physiologically plausible cone signals. The sharpened cones refers
to the fact that the spectral sensitivity are more narrow than those of real cone
sensitivities. The first stage of the Bradford model is a linear transform into these
sharpened cone signals, designated RGB, from normalized CIE XYZ tristimulus
values, as given by ⎡ ⎤
X
⎡ ⎤
R ⎢ Y ⎥
⎢ ⎥
⎢ ⎥ ⎢Y ⎥
⎣G⎦ = MBFD ⎢ ⎥ (10.15a)
⎢Y ⎥
B ⎣Z⎦
Y

i i

i i
 i i

i i

10.4. Models of Chromatic Adaptation 549

⎡ ⎤
X
⎡ ⎤
0.8951 0.2664 −0.1614 ⎢ Y ⎥
⎢
⎥
⎢ ⎥⎢Y ⎥
= ⎣−0.7502 1.7135 0.0367⎦ ⎢ ⎥ (10.15b)
⎢Y ⎥
0.0389 −0.0685 1.0296 ⎣ Z ⎦
Y

Normalized tristimulus values must be used in this transform because of the non-
linearity introduced to the short-wavelength B signal.
The sharpened “cone” signals are plotted along with the Hunt-Pointer-Estevez
cone signals in Figure 10.11. The first thing to note is that these cone signals
are not meant to be physiologically realistic, as evidenced by the negative lobes
present. They can be thought to represent types of interactions (both inhibition
and excitation) between the cone signals that may occur at the sensory level,
though they were not designed specifically in that manner. There is experimen-
tal evidence that does suggest that the cone signals do interact during chromatic
adaption and are not independent as von Kries suggested [244]. The sharpened
cone signals effectively allow for a constancy of both hue and saturation across
changes in chromatic adaptation.

Solid lines: Bradford

Dashed lines: Hunt-Pointer-Estevez
Relative sensitivity

1.0
L
M
0.8
S

0.6

0.4

0.2

0.0

−0.2
400 450 500 550 600 650 700
Wavelength (nm)

Figure 10.11. The sharpened cone signals of the Bradford chromatic-adaptation transform
(solid) and the Hunt-Pointer-Estevez cone signals (dashed).

i i

i i
 i i

i i

550 10. Chromatic Adaptation

The adapted cone signals are then calculated using

   
Rwr
Ra = D + 1 − D R, (10.16a)
Rw
   
Gwr
Ga = D + 1 − D G, (10.16b)
Gw
   " "
Bwr " "
Ba = D + 1 − D "Bβ " , (10.16c)
Bw
 
Bw 0.0834
β= . (10.16d)
Bwr

Note that the transform is based upon a reference illuminant of CIE D65 which
is calculated by transforming the CIE XYZ of D65 using the Bradford matrix
MBFD . In (10.16), RGBw are the sharpened cone signals of the adapting reference
conditions (reference white) and RGBwr are the cone signals of CIE D65. Note
the absolute value in calculating the Ba signal, necessary because the sharpened
cones can become negative.
The D term corresponds to the degree of discounting the illuminant. When
D is set to one, the observer is completely adapted to the viewing environment,
while when D is set to zero we would say that the observer is completely adapted
to the reference condition, in this case CIE D65.
Since D65 is always the numerator of (10.16), we can say that if D ∈ (0, 1),
the observer is in a mixed adaptation state, between the viewing condition and the
reference condition. The CIE XYZ tristimulus values for the reference D65 can be
calculated by multiplying by the inverse of the Bradford matrix and multiplying
back in the luminance value:
⎡ ⎤ ⎡ ⎤
Xa R aY
⎣Ya ⎦ = M −1 ⎣GaY ⎦ (10.17a)
BFD
Za B aY

⎡ ⎤⎡ ⎤
0.9870 −0.1471 0.1600 RaY
= ⎣ 0.4323 0.5184 0.0493⎦ ⎣GaY ⎦ . (10.17b)
−0.0085 0.0400 0.9685 BaY

The CIECAM97s color-appearance model borrowed the general form of the

Bradford chromatic-adaptation transform (or more precisely the LLAB adaptation
transform using the Bradford matrix), though they did not base the calculation on
a CIE D65 reference which made it easier to calculate corresponding colors [191].

i i

i i
 i i

i i

10.4. Models of Chromatic Adaptation 551

They also developed an empirical formula to calculate the D value from the lumi-
nance of the adapting conditions:
   
1
Ra = D + 1 − D R, (10.18a)
Rw
   
1
Ga = D + 1 − D G, (10.18b)
Gw
   
1
Ba = D p + 1 − D |B p | , (10.18c)
Bw
 
Bw 0.0834
p= , (10.18d)
1.0
F
D=F−+
 ,. (10.18e)
1/4 LA2
1 + 2 LA + 300

Here, LA is the adapting luminance, and F is a categorical surround function

usually set to 1.0 for normal conditions and 0.9 for darkened conditions. More
details on the effects of the luminance level and surround on color appearance
will be discussed in Chapter 11.
The sharpened cone responses of the Bradford chromatic-adaptation trans-
form proved to provide a better fit to the experimental correlated color data than
models built using traditional cone signals [191]. The incomplete adaptation fac-
tors as well as the nonlinearity make the full CIECAM97s transform more diffi-
cult to use in a practical application. The International Color Consortium (ICC),
the de facto standard body for color management, has recommended the use of
a linearized version of the CIECAM97s/Bradford transform with complete chro-
matic adaptation [523]. Essentially, this is a simple von Kries transform, such as
that shown in (10.6), to calculate corresponding colors from any given viewing
condition into the ICC reference-viewing condition of CIE D50.

10.4.6 CIECAT02 Chromatic-Adaptation Transform

The ICC selection of a linearized version of the Bradford transform is one exam-
ple that indicated the desire to have a practical chromatic-adaptation transform
that was easily applied and inverted. The CIE recognized that and set out to opti-
mize a linear transform that provided results that were similar to the performance
of the nonlinear transform in CIECAM97s [195]. Several linear transforms were
generated by various researchers to address this concern, many of which were
summarized by Fairchild [301].

i i

i i
 i i

i i

552 10. Chromatic Adaptation

Calabria and Fairchild showed that all of the transforms that use sharpened
cone responses behave similarly (and better than non-sharpened cones), and they
are perceptually identical to each other for most practical applications [145]. The
CIE then selected a sharpened transform that best fits the existing experimental
data and has a high degree of backwards compatibility to the nonlinear trans-
form in CIECAM97s [195]. This transform is known as CIECAT02, and it is
essentially a von Kries-style normalization using the following sharpened cone
transformation:
⎡ ⎤ ⎡ ⎤
R X
⎣G⎦ = MCAT02 ⎣Y ⎦ (10.19a)
B Z
⎡ ⎤ ⎡ ⎤
X R
⎣Y ⎦ = M −1 ⎣G⎦ (10.19b)
CAT02
Z B
⎡ ⎤
0.7328 0.4296 −0.1624
MCAT02 = ⎣−0.7036 1.6974 0.0061⎦ (10.19c)
0.0030 −0.0136 0.9834
⎡ ⎤
1.0961 −0.2789 0.1827
−1
MCAT02 = ⎣ 0.4544 0.4739 0.0721⎦ (10.19d)
−0.0096 −0.0057 1.0153

The CAT02 sharpened cone sensitivities are plotted along with the Bradford
cone sensitivities in Figure 10.12, showing that the Bradford signals are very sim-
ilar, though slightly sharper for the long wavelength cones.
To calculate corresponding colors using the CIECAT02 transform, assuming
complete adaptation to both the viewing conditions, we can then just use the
sharpened cone signals in a traditional von Kries adaptation matrix, as follows:
⎡R ⎤
w,2
⎡ ⎤ 0 0 ⎡ ⎤
⎢ Rw,1 ⎥
X2 ⎢ ⎥ X1
⎣Y2 ⎦ = M −1 ⎢ ⎥
Gw,2
CAT02 ⎢
0 0 ⎥ MCAT02 ⎣Y1 ⎦ . (10.20)
⎢ Gw,1 ⎥
Z2 ⎣ Bw,2 ⎦ Z1
0 0
Bw,1
Here, RGBw,1 and RGBw,2 are the sharpened cone responses of the adapting con-
ditions (white points).

i i

i i
 i i

i i

10.5. Application: Transforming sRGB Colors to D50 for an ICC Workflow 553

Solid lines: CIECAT02

Dashed lines: Bradford
1.0
Relative sensitivity
L
0.8 M
S
0.6

0.4

0.2

0.0

−0.2
400 450 500 550 600 650 700
Wavelength (nm)

Figure 10.12. The sharpened cone signals of the CIECAT02 chromatic-adaptation trans-
form (solid) and the Bradford cone signals (dashed).

10.5 Application: Transforming sRGB Colors

to D50 for an ICC Workflow
Chromatic-adaptation transforms (CATs) are designed to essentially calculate tris-
timulus values that have been normalized by the conditions of the viewing envi-
ronment. This can account for some degree of our perceived color constancy. By
performing a forward and inverse transform, we can generate corresponding col-
ors. These corresponding colors can be powerful tools in a digital-imaging library.
For instance, if we have one characterized display device, we can determine what
colors (tristimulus values) we need to generate on that device in order to match
another device under a different viewing condition.
An example of this is to determine what colors need to be displayed on a CRT
with a D65 white point in order to match a print viewed under D50 simulators.
Of course, this cross-media comparison brings many more factors into play, such
as the overall illumination levels and the color of the surround. Those factors are
incorporated into more advanced color-appearance models, of which chromatic-
adaptation transforms play an integral part. Traditional ICC color management
uses chromatic-adaptation transforms to move colors from the native-device white
into a standardized CIE-D50 white.

i i

i i
 i i

i i

554 10. Chromatic Adaptation

As we learned in Chapter 7, sRGB has become a “default” standard for color

management on the Internet [524]. While sRGB was originally designed to ap-
proximate the typical viewing conditions and primaries of a CRT device, it has
been adopted for a wide variety of other uses, including the encoding of digital-
camera images for many consumer cameras. The sRGB transform assumes that
the white point of most displays is equivalent to CIE D65, and so the encoded
white value (R = G = B) will produce D65 tristimulus values.
The ICC, on the other hand, recommends using a white point of CIE D50 in
the interim connection space, which is most often CIE XYZ. The sRGB trans-
forms used to convert from sRGB (encoded as an 8-bit value) into CIE XYZ are
given by
RGBs
RGBs = , (10.21a)
255.0
 
RGB = f RGBs , (10.21b)
⎧ 2.4
⎪
⎨ n + 0.055
n > 0.04045,
f (n) = 1.055 , (10.21c)
⎪
⎩12.92n n ≤ 0.04045.
⎡ ⎤ ⎡ ⎤
X R
⎣Y ⎦ = MsRGB ⎣G⎦ , (10.21d)
Z D65 B

where
⎡ ⎤
0.4124 0.3576 0.1805
MsRGB = ⎣0.2126 0.7152 0.0722⎦ (10.21e)
0.0193 0.1192 0.9505
It is important to note that the actual sRGB transform uses a piecewise-linear and
nonlinear component rather than a standard gamma of 2.2 as is often expressed
(see Section 8.1).
The forward transform from linearized RGB values into CIE XYZ tristimu-
lus values will provide us with values normalized for D65. To use these values
in an ICC color-managed workflow, we must apply a chromatic-adaptation ma-
trix, transforming into D50. We will use the linear Bradford matrix, since that is
recommended and used by the ICC [523]:
⎡ ⎤ ⎡ ρD50 ⎤ ⎡ ⎤
X ρD65 0 0 X
⎣Y ⎦ −1 ⎢ 0 γD50 ⎥ ⎣ ⎦ .
= MBFD ⎣ γD65 0 ⎦ MBFD Y (10.22)
Z D50 βD50 Z
0 0 β D65
D65

i i

i i
 i i

i i

10.6. Application: White Balancing a Digital Camera 555

To avoid confusion between the traditional RGB used to describe sharpened cone
signals, and the RGB of a digital device, we will use ρ , γ , and β in these equa-
tions. The sharpened cone signals for D50 and D65 are given by
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
ρ 0.9963 0.9642
⎣γ ⎦ = ⎣1.0204⎦ = MBFD ⎣1.0000⎦ ; (10.23a)
β D50 0.8183 0.8249 D50
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
ρ 0.9414 0.9504
⎣γ ⎦ = ⎣1.0405⎦ = MBFD ⎣1.0000⎦ . (10.23b)
β D65 1.0896 1.0889 D65

By using the values calculated in (10.23) as well as the traditional sRGB color
matrix in (10.22), we can generate a new matrix that will transform directly from
linearized sRGB into CIE XYZ adapted for a D50 white:
⎡ ⎤
0.9963
0 0
⎢ 0.9414 ⎥
−1 ⎢
⎢ 1.0204 ⎥
MsRGBD50 = MBFD ⎢ 0 0 ⎥ ⎥ MBFD MsRGB , (10.24a)
⎣ 1.0405
0.8183 ⎦
0 0
1.0896

and the resulting matrix is

⎡ ⎤
0.4361 0.3851 0.1431
MsRGBD50 = ⎣0.2225 0.7169 0.0606⎦ . (10.24b)
0.0139 0.0971 0.7140

Equation (10.24b) can now be used to directly take sRGB encoded values
into the standard ICC working space. In fact, when you read any ICC profiles
encapsulating sRGB, the “colorants” in those files will generally take the form of
the pre-adapted transform from Equation (10.24b).

10.6 Application: White Balancing a Digital Camera

Another potential application for chromatic-adaptation transforms that can be
very useful is white balancing digital camera data. Most current digital-imaging
technologies use linear sensors, or sensors that transduce photons with a direct
linear relationship to luminance. The data that are sensed must be processed in
order to produce images that are pleasing to look at.

i i

i i
 i i

i i

556 10. Chromatic Adaptation

One necessary processing step is white balancing, which attempts to mimic

the human behavior of chromatic adaptation. White balancing assures that when
an image is captured under a very “red” illuminant, such as indoors under in-
candescent lighting, the resultant image appears more neutral. Since this image
manipulation is akin to the human sensory process, it can actually be performed
using chromatic-adaptation transforms to assure that the resultant images appear
as the photographer remembers them based upon their own state of adaptation.
We will use a simulated digital camera to outline how this approach works.
The input into the system is a full spectral image, a light source, and a digital cam-
era with spectral sensitivities. A discussion of what to do when the nature of the
illuminant is not known, is given in Section 9.9. Image formulation is calculated
by multiplying these three signals together and then integrating (or summing)
over all wavelengths. For simulation purposes, relative spectral sensitivities and
spectral power distributions can be used. The image formation equation is given
by

∑ I (λ ) Rs (λ ) S (λ )
λ
R= ; (10.25a)
∑ S (λ )
λ

∑ I (λ ) Gs (λ ) S (λ )
λ
G= ; (10.25b)
∑ S (λ )
λ

∑ I (λ ) Bs (λ ) S (λ )
λ
B= . (10.25c)
∑ S (λ )
λ

This equation is normalized by the relative energy of the light source, S (λ ). The
full spectral image is represented by I (λ ) and Rs (λ ), Gs (λ ), and Bs (λ ) are the
spectral sensitivities of the digital camera.
As example input into the system, we will use the METACOW image, which
is a full-spectral synthetic image designed for imaging systems evaluation [293].
In addition to providing spectral information at each pixel, METACOW was de-
signed to be maximally metameric to test the variation of an imaging system from
the color-matching functions. Recall from Chapter 7 that a metameric pair of
stimuli are stimuli that appear to match but that have different spectral reflectance
or transmittance values. Due to the disparate spectral properties of the stimuli,
the matches often break down when viewed under different lighting (illuminant
metamerism) or when viewed by a different person or capture device (observer
metamerism).

i i

i i
 i i

i i

10.6. Application: White Balancing a Digital Camera 557

Figure 10.13. The METACOW image integrated with CIE 1931 Standard Observer and
CIE D65, then transformed into the sRGB color space.

One half of each of the objects in the METACOW image was rendered to have
the same spectral reflectance as the corresponding patch in the GretagMacbeth
ColorChecker (shown in Figure 12.30), while the other half was designed to be
a metameric match for the CIE 1931 Standard Observer when viewed under CIE
D65 [757]. An example of METACOW integrated with CIE D65 and the 1931
Standard Observer and converted to sRGB is shown in Figure 10.13. The full
spectral METACOW images are available on the DVD of this book.
For the sensor spectral sensitivities of our digital camera, we can choose a
simple three-channel RGB camera with Gaussian spectral sensitivities centered at
cr = 450 nm, cg = 550 nm, and cb = 600 nm with a full-width half-max (FWHM)
of 100 nm. The formulation for these sensitivities is given by
 
(cr − λ )2
Rs (λ ) = exp − ; (10.26a)
2σ 2
 
(cg − λ )2
Gs (λ ) = exp − ; (10.26b)
2σ 2
 
(cb − λ )2
Bs (λ ) = exp − . (10.26c)
2σ 2

i i

i i
 i i

i i

558 10. Chromatic Adaptation

2.0

Relative sensitivity
1.8

1.6

1.4 CIE A

1.2

1.0 R
0.8 G
B
0.6

0.4

0.2

0.0
400 450 500 550 600 650 700
Wavelength (nm)

Figure 10.14. The simulated spectral sensitivities of our digital camera plotted along with
CIE Illuminant A.

The full-width half-max value indicates the width of the Gaussian at half its max-
imum height. This value relates to σ according to [1232]:
√
FW HM = 2σ 2 ln 2. (10.27)

This means that the parameter σ is approximately 42.47 nm to achieve a FWHM

of 100 nm. The resulting sensitivities are plotted in Figure 10.14 along with CIE
Illuminant A, which will be our capture illuminant.
Combining the spectral image with the CIE A light source and the digital-
camera sensitivities, and integrating using (10.25) provides us with an RGB im-
age. This image is shown, with a nonlinear gamma correction of 1/2.2 for display
purposes, in Figure 10.15. There should be two things immediately evident from
this image: the overall yellow tone caused by the CIE A illumination, and the
fact that the two halves of the cows are now very different. The two halves no
longer match because of both illuminant and observer metamerism, given that the
spectral properties of the halves are very different.
Once an image has been captured the spectral information is lost, and there
is no way to regain that information and make the two halves match again. We
can, however, fix the overall yellow cast by applying a white-balancing based on
a chromatic-adaptation transform. To do this, we must first derive a relationship
between the digital camera sensors and CIE XYZ tristimulus values. This process
is known as device characterization, and it is a crucial step for color management

i i

i i
 i i

i i

10.6. Application: White Balancing a Digital Camera 559

Figure 10.15. The scene values as captured by the digital camera simulation with no white
balancing.

Figure 10.16. The digital-camera image chromatically adapted to D65 and transformed
into the sRGB color space.

i i

i i
 i i

i i

560 10. Chromatic Adaptation

(see Section 12.12). Since we know both the spectral sensitivities of our digital
camera as well as the spectral sensitivities of the CIE 1931 Standard observer, we
can generate a transform to try to map one to the other. The simplest way to do
this is to create a linear transform through least-squares linear regression. With
X̂ Ŷ Ẑ representing the predicted best estimate of CIE XYZ tristimulus values, the
desired 3 × 3 transform is
⎡ ⎤ ⎡ ⎤
X̂ R
⎣Ŷ ⎦ = M ⎣G⎦ . (10.28)
Ẑ λ B λ

We are interested in determining the optimal matrix M that will minimize

the error from predicted XYZ values to actual XYZ values. To do this, a linear
regression can be performed using a pseudo-inverse transform. Since RGB is a
3 × N matrix, we cannot take a direct inverse, and so a pseudo-inverse is required:

y = M · x, (10.29)

where
 −1
M = y · xT · x · xT , (10.30a)
⎡ ⎤ ⎛⎡ ⎤ ⎞−1
X 0 1 R 0 1
M = ⎣Y ⎦ R G B ⎝⎣G⎦ R G B ⎠ , (10.30b)
Z B

and xT is the transpose of x, x = (RGB)T , and y = (XY Z)T . Performing this

calculation between the spectral sensitivities of the digital camera and the 1931
standard observer results in the following linear transform:
⎡ ⎤
1.0198 0.0161 0.2456
Mcamera = ⎣ 0.2788 0.9160 0.0147⎦ . (10.31)
0.0582 −0.1098 1.4368
We can plot the differences between our optimized transformed camera sig-
nals and the actual 1931 Standard Observer, which is shown in Figure 10.17. The
figure shows that the best linear fit provides a decent transform to get our cam-
era into CIE XYZ, though there are some noticeable differences, especially in the
blue channel. It is also possible to determine the optimized camera matrix Mcamera
using more terms such as squares or interactions in the linear regression (result-
ing in a larger than 3 × 3 transform matrix) or by doing a linear regression on the
integrated signals of a test target. For instance, since we can calculate the XYZ
values of the METACOW image, we could actually do a regression between the
integrated RGB sensor values and the integrated XYZ values.

i i

i i
 i i

i i

10.6. Application: White Balancing a Digital Camera 561

Solid lines: Transformed digital camera signals

Dashed lines: CIE 1931 standard observer
Relative sensitivity 1.8
1.6
R
1.4 G
B
1.2
1.0
0.8
0.6
0.4
0.2
0.0
−0.2
400 450 500 550 600 650 700
wavelength (nm)

Figure 10.17. Our transformed digital-camera signals (solid) and the CIE 1931 Standard
observer (dashed).

Once we have our Mcamera transformation to CIE XYZ, we can then apply a
standard chromatic-adaptation transform, such as CIECAT02. We can also move
into a standard color space, such as sRGB, for ease of communication from our
device-specific RGB values. In order to do that, we need to choose the “white” of
our input image. We could use the integrated signals of just CIE A and our camera
sensitivities, since we know that was the illuminant we imaged under, or we can
use a white-point estimation technique. In this case, we will choose a simple
white-point estimate, which is the maximum value of each of the RGB camera
signals. The output white will be set to D65, the native white point of sRGB.
The full transform to achieve a white-balanced sRGB signal from our unbalanced
camera is then

⎡ ⎤
ρD65
⎡ ⎤ 0 0 ⎡ ⎤
⎢ ρmax ⎥
R ⎢ βD65 ⎥ R
⎢ ⎥
⎣G⎦ −1
= MsRGB −1
MCAT02 ⎢ 0 0 ⎥ MCAT02 Mcamera ⎣G⎦ .
⎢ βmax ⎥
B sRGB ⎣ γD65 ⎦ B
0 0
γmax
(10.32)
The resulting image is shown in Figure 10.16.

i i

i i
 i i

i i

562 10. Chromatic Adaptation

What we should see now is that the chromatic-adaptation transform did a very
good job of removing the overall yellow tint in the image, and the left halves of
the cows look much like the traditional Color Checker. The metameric halves
still look very strange and clearly do not match. This is because the sRGB color
space has a much smaller gamut than the metameric samples in the METACOW
image, and so the output image shows clipping in the color channels (below zero
and greater than 255).
The gamut of an output device essentially describes the range of colors that
the device is capable of producing. The process of moving colors from the gamut
of one device to that of another is known as gamut mapping. In this case, we
used a simple clipping function as our gamut-mapping algorithm. A more clever
gamut-mapping algorithm may be required to prevent such artifacts or to make
the image appear more realistic. Details of more sophisticated gamut-mapping
algorithms can be found in [798], and an overview is given in Section 15.4.
This example is meant to illustrate the power that a chromatic-adaptation
transform can have for white balancing digital cameras. Through device-
characterization and the CIECAT02 transform we are able to transform our sim-
ulated digital-camera signals into the sRGB color space with successful results.
For most cameras and spectral reflectances, these techniques should prove to be
very useful.

10.7 Application: Color-Accurate Rendering

Image synthesis was discussed previously in Section 2.10 where its basic princi-
ples were outlined. The assumption was made that each wavelength of interest
was rendered separately, and that therefore the renderer was fully spectral. In
Section 8.12 the choice of sharpened RGB color space was motivated for use in
a renderer. Such a color space allows images to be created that are perceptually
close to their spectral counterparts, without the cost associated with a spectral
renderer.
When creating images in this manner, the results will often have a strong color
cast that may be unexpected. There are at least two sources for such color casts
to occur, both relatively straightforward to correct, once their sources are known.
The first inaccuracy stems from a possible difference in measuring and render-
ing conditions for the materials in the scene. Especially when measured BRDFs
are used, it should be noted that such BRDFs are measured under a specific illu-
minant. If this illuminant is different from the ones used in the geometric model,
the appearance of the material will be inaccurate [1208]. As an example, one may

i i

i i
 i i

i i

10.7. Application: Color-Accurate Rendering 563

read the Yxy chromaticity coordinates for colored patches of a Gretag-Macbeth

color checker and convert these values to the chosen rendering color space (for
instance the sharpened RGB color space discussed in Section 8.12). However,
the Yxy coordinates listed on the Gretag-Macbeth ColorChecker chart are valid
for illumination under CIE standard illuminant C. Unless the light sources in the
scene to be rendered are also of type C, the chosen patches will appear different
from the ones on a real Macbeth color checker.
Under the assumption that the scene to be rendered has a single dominant light
source, i.e., there may be many light sources, but the ones providing the main
illumination are of the same type, it is possible to transform the specification of
all materials in the scene from the illuminant under which they were measured
to the dominant illuminant under which they will be rendered. Such a transform
involves a standard von Kries-style transformation as presented in this chapter,
which can be carried out prior to rendering as a pre-filtering process.
The preferred approach is to replace the dominant light sources in the scene
with neutral light sources of equal strength and change spectrally distinct light
sources accordingly. If in the sharpened RGB color space, the white point of
the dominant illuminant is given by (Rw , Gw , Bw ), the remaining illuminants are
transformed with
Rs
Rs = , (10.33a)
Rw
Gs
Gs = , (10.33b)
Gw
Bs
Bs = . (10.33c)
Bw
The surface materials are also transformed to account for the change in illuminant.
The transform from CIE XYZ to sharpened RGB is then given as
⎡ ⎤
1
⎡  ⎤ ⎢ 0 0 ⎥ ⎡ ⎤
Rm ⎢ Rw ⎥ Xm
⎢
⎣Gm ⎦ = ⎢ 0 1 ⎥
0 ⎥ Msharp ⎣Ym ⎦ . (10.34)
⎢ Gw ⎥
Bm ⎣ 1 ⎦ Zm
0 0
Bw
Rendering then proceeds in the sharpened RGB color space as normal.
The second inaccuracy stems from a possible difference in dominant illumi-
nant in the rendered scene and the viewing environment. This once more produces
an image with an observable, and potentially distracting, color cast. The final im-
age may therefore have to undergo a further white-balancing operation. Hence,

i i

i i
 i i

i i

564 10. Chromatic Adaptation

after rendering, the transform from sharpened RGB to display values is [1208]
⎡ ⎤ ⎡  ⎤⎡ ⎤
Rd Rw 0 0 R
⎣Gd ⎦ = Md M −1 ⎣ 0 Gw 0 ⎦ ⎣G ⎦ . (10.35)
sharp
Bd 0 0 Bw B

where Md is the transform that takes XYZ tristimulus values to the display’s color
space, and (Rw , Gw , Bw ) represents the white balance associated with the viewing
environment.
After rendering and white balancing for the display environment, we assume
that further corrections are carried out to prepare the image for display. These
will involve tone reproduction to bring the dynamic range of the image within
the range reproducible by the display device, gamma correction to account for the
nonlinear response of the display device to input voltages, and possibly correction
for the loss of contrast due to ambient illumination in the viewing environment.
Each of these issues is further discussed in Chapter 17.
While many computer-graphics practitioners do not follow this procedure, we
hope to have shown that maintaining color accuracy throughout the rendering
pipeline is not particularly difficult, although it requires some knowledge of the
illuminants chosen at each stage. The combined techniques described in this sec-
tion, as well as those in Section 8.12, enable a high level of color fidelity without
inducing extra computational cost.
Finally, the white-balancing operation carried out to account for the illumi-
nation in the viewing environment can be seen as a rudimentary form of color-
appearance modeling. Under certain circumstances, it may be possible to replace
this step with the application of a full color-appearance model, such as iCAM,
which is the topic of Chapter 11.

10.8 Further Reading

Further information on chromatic adaptation models can be found in Fairchild,
Color Appearance Models [302].

i i

i i
 i i

i i

Chapter 11
Color and Image
Appearance Models

Colorimetry, as described in Chapter 7, is a standardized method of measuring

colors and determining if color stimuli match. Using colorimetry, we can easily
communicate information and specify colors in a device-independent manner. We
know that if two stimuli have identical CIE XYZ tristimulus values and are viewed
in identical conditions, then the two stimuli will match. We do not know, however,
what the appearances of those stimuli are. Furthermore, we cannot guarantee if
those stimuli will still match if any element of the viewing environment changes.
Chromatic adaptation, as discussed in Chapter 10, is one method that our
visual system uses to alter its behavior based upon the color of the prevailing
illumination. Chromatic-adaptation models extend basic colorimetry to predict
color matches across changes in illumination color, essentially giving us device-
and illuminant-color-independent representation. We still are not able to describe
what the stimuli look like.
Color appearance extends upon chromatic adaptation and colorimetry to pro-
vide a description of how a stimulus actually appears to a human observer. In
essence, we can say that color appearance allows us to take the numbers arriving
from color measurement and translate these into words describing how a color
looks.
In order to predict the actual perception of on object, it is necessary to know
information about the viewing environment itself. As we saw with chromatic
adaptation, our visual system is constantly adapting to the environment and, thus,
the appearance of any given stimulus will be strongly influenced by these changes.
565

i i

i i
 i i

i i

566 11. Color and Image Appearance Models

Color appearance models take into account measurements, not only of the stimu-
lus itself, but also of the the viewing environment, and they generate, in essence,
viewing condition independent color descriptions. Using color appearance mod-
els, we can determine whether or not two stimuli match even when viewed under
markedly different conditions. Consider the problem of viewing an image on a
LCD display and trying to create a print that matches the image, but that will be
viewed in a very different context. Color appearance models can be used to pre-
dict the necessary changes that will be needed to generate this cross-media color
reproduction.
This chapter will provide an overview of the terminology used in color science
and color appearance, as well as discuss some well-known color appearance phe-
nomena. The CIECAM02 color appearance model and its uses will be discussed,
as well as two spatial or image appearance models: S-CIELAB and iCAM.

11.1 Vocabulary
As with all areas of study, it is important to have a consistent terminology or
vocabulary to enable accurate and precise communication. When studying color,
this is all the more important, as color is something we all grew up with and most
likely acquired different ways of describing the same thing. Color and appearance
vocabulary is often interchanged haphazardly, for example, easily swapping terms
such as luminance, lightness, and brightness. This may be more prevalent when
discussing a subject such as color, as it is easy to assume that since we have been
exposed to color our entire life, we are all experts.
Even in our education, treatment of color is highly varied. To a child, color
might be made up of their three primary paints or crayons: red, blue, and yel-
low. Many graphic artists or printers work under the premise of four primaries:
cyan, magenta, yellow, and black. Computer-graphics researchers may assume
that color is made up of red, green, and blue. Physicists might assume that color is
simply the electromagnetic spectrum, while chemists may assume it is the absorp-
tion or transmission of this spectrum. All of these assumptions can be considered
to be correct, and to each of the disciplines, all of the others may be considered
quite wrong.
In the field of color appearance, the standard vocabulary comes from the Inter-
national Lighting Vocabulary, published by the CIE, and the Standard Terminol-
ogy of Appearance, published by the American Society for Testing and Materials
(ASTM) [49, 187]. The definition of terms presented below comes directly from
these works.

i i

i i
 i i

i i

11.1. Vocabulary 567

11.1.1 Descriptions of Color

The concept of color itself is well known to almost everyone, though when asked,
few can give a precise definition of what exactly color is. Much of the confusion
when studying color appearance stems from the very nature and definition of color
itself. It is almost impossible to describe or define color without using an example
or the word color itself, as evident from the CIE definition.

Color. The attribute of visual perception consisting of any combination of chro-

matic and achromatic content. This attribute can be described by chromatic
color names such as yellow, orange, brown, red, pink, green, blue, pur-
ple, etc., or by achromatic color names such as white, gray, black, etc., and
qualified by bright, dim, light, dark, etc., or by combinations of such names.

That the definition of color requires the use of the word it is trying to define
is an uncomfortable circular reference that can lead to confusion. The original
authors of this vocabulary were clearly concerned with this potential confusion,
and they added a note that summarizes well the study of color appearance:

Perceived color depends on the spectral distribution of the color stimulus, on

the size, shape, structure, and surround of the stimulus area, on the state of
adaptation of the observer’s visual system, and on the observer’s experience
of prevailing and similar situations of observations.
Much of what is described in this note is addressed in further detail in this
chapter.
The most important aspect in the definition of color is the first sentence: color
is an attribute of visual perception. We typically say that color is a function of
the lighting, the environment, and the observer, and without the visual system’s
involvement, color cannot exist. All the color terminology discussed in this sec-
tion involves attributes of visual perception. That is to say, without the observer,
there can be no discussion of color. The study of color appearance and color
appearance models is an attempt to generate physically realizable measurements
that correlate with these perceptual attributes.
The definition of color is extended to the notion of related and unrelated col-
ors. The concept of these types of colors is simple, but also critical to gaining a
further understanding of color appearance.

Related color. Color perceived to belong to an area of object seen in relation to

other colors.
Unrelated color. Color perceived to belong to an area of object seen in isolation
from other colors.

i i

i i
 i i

i i

568 11. Color and Image Appearance Models

Related colors are viewed in the same environment or area as other stimuli,
while unrelated colors are viewed in isolation. We generally think that color stim-
uli are not viewed in complete isolation (unless you are an astronomer), and as
such most color appearance models have been designed to primarily predict re-
lated colors. The pedigree of these models, however, are the many color-vision
experiments performed that used stimuli viewed in isolation. These experiments
have been powerful tools for gaining an understanding of how the human visual
system behaves. It is also important to understand the differences between these
types of stimuli when utilizing models designed to predict one specific type of
color appearance.
There are several important color perceptions that only exist when viewed as
related or unrelated stimuli. Perhaps the most fascinating of these cases is the so-
called “drab” colors such as brown, olive, khaki, and gray. These colors can only
exist as related colors. It is impossible to find an isolated brown or gray stimulus,
as evidenced by the lack of a brown or gray light bulb. Even if the surface of
these lights looked brown or gray, they would appear either orange or white when
viewed in isolation.
This is not to say that we cannot make a stimulus look brown or gray using
only light, as it is quite easy to do so using a simple display device such as a
CRT or LCD. We can only generate those perceptions when there are other colors
being displayed as well, or if we can see other objects in the viewing environment.
When we talk about relative attributes of color appearance, such as lightness or
chroma, then by definition, these attributes only exist for related colors. The
absolute color appearance terminology can be used for either related or unrelated
colors.
Of all the color appearance attributes, hue is perhaps the easiest to understand.
When most people are asked to describe the color of an object, they will generally
describe the hue of the object. That said, it is almost impossible to define hue
without using examples of what a hue is.

Hue. Attribute of a visual sensation according to which an area appears to be

similar to one of the perceived colors: red, yellow, green, and blue, or to a
combination of two of them.

Achromatic color. Perceived color devoid of hue.

Chromatic color. Perceived color possessing a hue.

Hue is most often described as a continuous hue-circle or color-wheel, as

shown in Figure 11.1. The standard definition of hue given by the CIE contains
a notion of unique or physiological hues. Specifically these hues are red, yellow,

i i

i i
 i i

i i

11.1. Vocabulary 569

Figure 11.1. The concept of a hue-circle or wheel. The physiological hues of red, yellow,
green, and blue form a closed ring with other colors falling between (a combination of)
adjacent pairs [49].

green, and blue. These unique hues follow the opponent color theory of vision
postulated by Hering in the late 19th century [462]. It has been noted that several
of these particular hues are never perceived together, and other colors can be
described as a combination of two of them. For example, there is no stimulus that
is perceived to be reddish-green or a yellowish-blue, but purple can be described
as a reddish-blue and cyan as a bluish-green. This opponent theory suggests the
fundamental notion that human color vision is encoded into roughly red-green and
blue-yellow channels. Thus, opponent color representation is a crucial element of
all color appearance models.
The definitions for the achromatic and chromatic colors is also important, as
the achromatic colors represent the third channel, the white-black channel theo-
rized by Hering. Although hue, or more precisely hue-angle, is often described as
a location on a circle, there is no natural meaning for a color with a hue of zero,
and so that value can be considered an arbitrary assignment (as is a hue angle
of 360 degrees). Achromatic colors describe colors that do not contain any hue

i i

i i
 i i

i i

570 11. Color and Image Appearance Models

information, but this definition does not extend to a meaningful placement on the
hue circle.
The only color appearance attributes of achromatic colors are defined to be
brightness and lightness, though these descriptions also apply for chromatic col-
ors. These terms are very often confused and interchanged with each other, but it
is important to stress that they have very different meanings and definitions.

Brightness. Attribute of a visual sensation according to which an area appears to

emit more or less light.

Lightness. The brightness of an area judged relative to the brightness of a simi-

larly illuminated area that appears to be white or highly transmitting. Note
that only related colors exhibit lightness.

Brightness refers to the absolute perception of the amount of light of a stim-

ulus, while lightness can be thought of as the relative brightness. Remember that
luminance is defined as the energy of light in a scene integrated with the human
luminous photopic-efficiency function, CIE V (λ ). Luminance can be considered
to be linear with the amount of light energy present in the scene, while brightness
and lightness are our perceptual response to this increase in energy (typically a
nonlinear response). In practice, the human visual system behaves as a relative
lightness detector.
We can understand the relative lightness behavior of our visual system by
imagining the viewing of a black-and-white photograph, both indoors and out-
doors. When viewed indoors, the black portions of the image look black, and the
white portions look white. This same relationship holds when we look at the pic-
ture outdoors on a sunny day. If we were to measure the amount of light coming
off the photograph, we may find that the amount of light coming off the black ar-
eas of the photograph is hundreds of times larger than the amount of light coming
off the white areas of the photograph when viewed indoors. Yet our relative per-
ception of the lightness of the black and white areas remains constant. In practice,
we find that the black areas of the photograph actually appear darker when viewed
in bright sunlight while the white areas appear lighter, a phenomena known as the
Stevens effect. This and other color appearance phenomena are discussed in more
detail in Section 11.2.
Like the definition of color itself, the definitions of lightness and brightness
include a note, stating that only relative colors can exhibit lightness. This is be-
cause lightness is brightness judged in comparison to a similarly illuminated white
object. This is the reason that there cannot be a gray light source, because if it
were viewed in isolation it would be the brightest stimulus in the field of view and

i i

i i
 i i

i i

11.1. Vocabulary 571

would thus appear white. As a simple mathematical relation, we can describe the
general concept of lightness by

Brightness
Lightness = . (11.1)
Brightness of White
Lightness and brightness are used to describe the relative amount of light a
colored stimulus appears to reflect or emit. Hue describes the general “color”
of a given stimulus. The terms used to describe the amount of hue or chromatic
content a stimulus has are also often confused and interchanged. These terms,
colorfulness, chroma, and saturation have very distinct meanings, and care should
be taken when choosing which attribute to describe the color of a stimulus.

Colorfulness. Attribute of a visual sensation according to which the perceived

color of an area appears to be more or less chromatic. For a color stim-
ulus of a given chromaticity and, in the case of related colors, of a given
luminance factor, this attribute usually increases as the luminance is raised,
except when the brightness is very high.

Chroma. Colorfulness of an area judged as a proportion of the brightness of a

similarly illuminated area that appears white or highly transmitting. For
given viewing conditions and at luminance levels within the range of pho-
topic vision, a color stimulus perceived as a related color, of a given chro-
maticity, and from a surface having a given luminance factor, exhibits ap-
proximately constant chroma for all levels of luminance except when the
brightness is very high. In the same circumstances, at a given level of illu-
minance, if the luminance factor increases, the chroma usually increases.

Saturation. Colorfulness of an area judged in proportion to its brightness. For

given viewing conditions and at luminance levels within the range of pho-
topic vision, a color stimulus of a given chromaticity exhibits approxi-
mately constant saturation for all luminance levels, except when brightness
is very high.

Colorfulness describes the absolute amount or intensity of the hue of a given

stimulus. Similarly, chroma is to colorfulness as lightness is to brightness, that
is, a relative amount of colorfulness for a given brightness. Whereas chroma
is defined to be colorfulness of an area relative to the brightness of a similarly
illuminated white stimulus, saturation is colorfulness relative to the brightness of
the stimulus itself. So while only a related color can exhibit chroma, both related
and unrelated colors can exhibit saturation. The differences between chroma and
saturation are subtle, but distinct.

i i

i i
 i i

i i

572 11. Color and Image Appearance Models

Figure 11.2. The appearance attributes of lightness, chroma, and hue can be represented
as coordinates in a cylindrical color space. Brightness, colorfulness, and hue can also be
represented as such a space.

We can generalize colorfulness or chroma along with hue and lightness or

brightness as attributes that form a cylindrical color space, as seen in Figure 11.2.
That is, for any given lightness or brightness, increasing chroma or colorfulness
is represented by a perpendicular line away from the lightness axis. Colors at a
constant chroma or colorfulness are then represented by cylinders of varying radii
in this space.
Saturation can be best described as a shadow series of colors. This is perhaps
best explained with an example. If we have rounded colored objects that are illu-
minated at a glancing angle, these objects will have high lightness and chroma on
the side that is facing the light, and the brightness and colorfulness will decrease
as we move to the side away from the light, eventually becoming black where
the object sees no light. This is illustrated by a series of rendered globes in Fig-
ure 11.3. Although the brightness and colorfulness have both decreased, as have
the lightness and chroma, the saturation has remained constant. Since saturation
is the relationship between brightness and colorfulness of the object itself, we
can say that in a shadow series, because brightness and colorfulness co-vary, the

i i

i i
 i i

i i

11.1. Vocabulary 573

Figure 11.3. A rendered shadow series illustrating constant saturation. The glancing
illumination causes a decrease in brightness and colorfulness (as well as lightness and
chroma) across the surface of each globe, but saturation remains constant.

saturation is constant. Thus, we can think of the appearance attributes of light-

ness, hue, and saturation as forming coordinates in a conical color space, shown
in Figure 11.4. In this case lines of constant saturation form the surface of cones
of varying radii originating from a single point at black.
We can once again generalize the formulas for chroma and saturation, which
will be useful when we move forward and discuss color appearance models. Re-
call that chroma is considered relative colorfulness, while lightness is considered
relative brightness. Thus, by considering the brightness of a scene as white, we
can also define relative saturation as the relationship between a stimulus’ chroma
to its lightness. These relationships are shown in (11.2).
Colorfulness
Chroma = , (11.2a)
Brightness (white)
Colorfulness
Colorfulness Brightness(white)
Saturation = = , (11.2b)
Brightness Brightness
Brightness (white)
Chroma
Saturation = . (11.2c)
Lightness

i i

i i
 i i

i i

574 11. Color and Image Appearance Models

Saturation

Lightness

on
rati
Satu
tant
ns
Co

Hue Angle

Figure 11.4. The appearance attributes of lightness, saturation, and hue can be represented
as coordinates in a conical color space originating at a single point at black.

Similarly to our behavior with lightness and brightness, the human visual sys-
tem generally behaves more like a relative chroma detector. As the luminance of
the viewing conditions increases, the perceptions of lightness and chroma tend to
remain constant while the brightness of a white stimulus increases. In this same
situation the colorfulness also increases as a function of luminance. This can be
visualized by thinking of an typical outdoor scene. On a sunny day, everything
lo¡oks very colorful, while the same scene viewed on cloudy day appears much
less colorful. This appearance phenomenon is known as the Hunt effect and will
be discussed in more detail Section 11.2.2.

11.1.2 Relative or Absolute Color Reproduction?

When our goal is digital color imaging, reproduction, or synthesis, it is often eas-
iest and most desirable to represent color as coordinates in a trichromatic space.
The success of CIE-based colorimetry for many applications illustrates the po-
tential strengths of this approach. Colorimetry, as we recall, is really only valid

i i

i i
 i i

i i

11.1. Vocabulary 575

when dealing with predicting or generating color matches in identical viewing

conditions. If the viewing conditions change, such as when going from a CRT
monitor to a print medium or moving a print medium from indoors to outdoors,
colorimetry becomes insufficient.
Thus, for accurate color-imaging applications, it becomes necessary to spec-
ify the stimuli in terms of color appearance. Complete specification requires at
least five of the perceptual attributes described above: brightness, lightness, col-
orfulness, chroma, and hue. It is generally suggested that the specification of
saturation is not necessary, as that can be inferred from the other attributes. Hunt
has recently made a compelling argument that saturation should not be ignored,
and that it may actually provide a wealth of useful information as we start to fo-
cus on the color reproduction of three-dimensional objects [508]. His argument
stems from the fact that for a shaded object the chroma or colorfulness changes
while saturation remains constant, and humans recognize that the object’s color
also remains constant.
When designing color reproduction or rendering systems it might appear that
specifying more than the three relative color appearance attributes is also re-
dundant. This is not always the case, however, as described in detail by Nay-
atani et al. [821]. They discuss the distinction between choosing an absolute
brightness-colorfulness match or a relative lightness-chroma match. In practice,
for most color-reproduction applications, it is often sufficient and desirable to gen-
erate lightness-chroma appearance matches, rather than the absolute brightness-
colorfulness match.
We can return to our example of a photograph captured from an outdoor scene
in very bright sunlight. The absolute luminance levels and brightness values of the
outdoor scene are incredibly high. The photograph is then reproduced and viewed
in an indoor environment, at much lower luminance levels. Regardless whether
the reproduction is a hard-copy print, or just an image on a display device, it
is probably physically impossible to achieve the luminance levels necessary to
generate an absolute brightness-colorfulness match such that the measured en-
ergy reflected off the print or emitted from the display is the same as the original
outdoor environment.
Likewise, if we were to generate an image on our display device that has a lim-
ited luminance range, we may not want to create an absolute appearance match
when we display that image on a high-dynamic range display viewed in a very
bright environment. While it may now be physically possible to reproduce the ab-
solute attributes, it would be undesirable since the image would probably appear
dark and of very low contrast. Thus, for most general color-reproduction applica-
tions, it is often more desirable to create a relative lightness-chroma appearance

i i

i i
 i i

i i

576 11. Color and Image Appearance Models

match, where the relationship between objects in the scene is held constant. We
can think of this as analogous to dynamic range reduction for image capture and
rendering (discussed in Chapter 17), since we want to generate a relative color ap-
pearance match that closely matches in appearance to the original high dynamic
range scene.

11.1.3 Descriptions of Viewing Environments

When discussing color appearance and color appearance modeling, it is important
to have a consistent vocabulary to describe the viewing environment, or field,
just as it is important for the color attributes themselves. As we learned from
the definition of color itself, the perception of color is heavily dependent on the
environment that the stimulus is viewed in. In color appearance terminology, the
viewing environment is also generalized as the viewing conditions. Typically, in
color appearance, we divide the viewing field into four generalized components:
stimulus, proximal field, background, and surround. For a simple color stimulus,
the viewing field is specified as in Figure 11.5.

Stimulus. The stimulus is the color element of interest. As with standard col-
orimetry, the stimulus used for most color appearance research is a small
uniform color patch, which subtends 2◦ of visual angle.

As we learned in Chapter 7, the CIE has separate systems of colorimetry de-

signed to handle small or large color patches, the CIE 1931 Standard Observer
and the CIE 1964 supplemental Standard Observer. Most of the experimental data
used in color appearance modeling and research has been performed using small
uniform color patches, though this is a very active research field and is constantly

Surround
Stimulus
Proximal field
Background
10o

2o

Figure 11.5. The components of the viewing field used for color appearance specification.

i i

i i
 i i

i i

11.1. Vocabulary 577

being extended. Ideally, the stimulus is described by the full spectral represen-
tation in absolute radiance terms, though in practice this is difficult to achieve.
When the spectral power distribution is unavailable, the stimulus should be de-
scribed using a standard device-independent space, such as CIE XYZ tristimulus
values, or LMS cone responsivities.
As we progress and apply color appearance modeling to color imaging and
synthesis, the definition of the stimulus can get somewhat blurred. Is the stimulus
a single pixel, a region of pixels, or the entire image? While it may be most conve-
nient to simply assume that the entire image is the stimulus, that is surely an over-
simplification. Currently there is no universally correct definition of the stimulus
for spatially complex scenes such as images. Therefore when using images with
color appearance models, care should be taken to fully understand the manner in
which the models are applied. CIECAM02, the current CIE-recommended color
appearance model, is designed to be used for color management and imaging;
most applications assume that each pixel can be treated as a separate stimulus.
Spatial models of color appearance, also known as image appearance models, are
designed for use with complex images and will be discussed in more detail in
Section 11.4.

Proximal field. The proximal field is considered to be the immediate environment

extending from the stimulus for about 2◦ in all directions.

The proximal field can be useful for measuring local contrast phenomena such
as spreading and crispening, which are introduced in Section 11.2.1. Again the
question of defining the locally spatial proximal field becomes very difficult when
considering digital color images. Should the proximal field for any given pixel be
considered to be all of the neighboring pixels, or perhaps just the edge border of
the image itself? While it is an important factor in determining the color appear-
ance of a stimulus, in most real-world applications, the proximal field is assumed
to be the same as the background. Again, the ideal description of the proximal
field would be a measure of both spatial and spectral properties, though these are
often unavailable.

Background. The background is defined to be the environment extending from

the proximal field, for approximately 10◦ in all directions. If there is no
proximal field defined, then the background extends from the stimulus
itself.

Specification of the background is very important in color appearance, as

it strongly influences the appearance of any given stimulus and is necessary to

i i

i i
 i i

i i

578 11. Color and Image Appearance Models

model phenomena such as simultaneous contrast. Specifying the background

when dealing with simple color patches is relatively straightforward. At the min-
imum, the background should be specified with absolute CIE XYZ tristimulus
values, since the luminance of the background will also have an effect on color
appearance. Although the background is defined to span at least 10 degrees of
visual angle, typically it is specified using the same CIE standard observer as the
stimulus itself.
Again, specifying the background when calculating the appearance of color
images can be as confusing as specifying the stimulus and the proximal field. For
any given image pixel, the background actually consists of many of the neighbor-
ing pixels as well as the area outside the image. Generally, there are two different
assumptions that people use for specifying the background in color-imaging ap-
plications. The first and easiest is to assume that the entire image is the stimulus,
so that the background is the area extending 10◦ from the image edge. Another
assumption is that the background is constant and of some medium chromaticity
and luminance, e.g., a neutral gray. This can be chosen as a non-selective gray
viewed in the same lighting conditions or can be taken to be the mean color of the
image itself.
Many imaging applications strive to reproduce images of constant spatial
structure and size, and thus these concerns are mitigated or disappear entirely.
Care must be taken when calculating color appearances across changes in image
sizes, though. Braun and Fairchild describe the impact on some of these back-
ground decisions [124]. More recently, Nezamabadi et al. and Xiao et al. have
investigated the effect of image size on color appearance [833, 1263].

Surround. The surround is considered anything outside of the background. For

most practical applications, the surround is considered to be the entire room
that the observer is in.

For practical applications, color appearance models tend to simplify and ex-
press the surround in distinct categories: dark, dim, and average. For instance,
movie theaters are usually considered to have a dark surround, while most people
tend to view televisions in a dim surround. A typical office or computing environ-
ment is often said to be an average surround. A more detailed discussion on the
effect of the surround on color appearance is available in Section 11.2.4.

11.1.4 Modes of Viewing

Any changes to the viewing fields discussed above may quickly result in a change
in the color appearance of a stimulus. The discussion of color appearance phe-
nomena in the following section explains some of these changes in detail. Other

i i

i i
 i i

i i

11.1. Vocabulary 579

factors that cannot be readily explained using our model of a simplified viewing
field can also have a large effect on the perceived appearance of a stimulus. As we
have established by now, the perception of color cannot be adequately explained
by the physics of light alone. The human observer is the critical component that is
ultimately responsible for any sensation or perception. The human visual system
relies both upon sensory mechanisms and automatic changes based on biological
and physiological factors, as well as on cognitive interpretations.
The full scope of cognitive mechanisms is not yet understood, though we are
able to recognize and model some of the ultimate behavior and responses. One of
the most important cognitive behaviors affecting color appearance is the mode of
appearance. Like much of color itself, the mode of color appearance is a difficult
concept to grasp at first and might be best described with an example.
Often times, we catch a glimpse of colors that at first appear vibrant and
highly chromatic and then suddenly become mundane. When walking past a
semi-darkened room recently, we were struck by what appeared to be a wall that
appeared painted with a mural of abstract colored rectangles of various shades of
brown and orange. The appearance of the wall was quite striking and highly chro-
matic. Upon entering the room, it became obvious that the colors were the result
of the street-lights, light from very orange low-pressure sodium lamps, entering
through an open window. Almost immediately, the appearance of the various
highly chromatic rectangles changed and appeared to be the same white wall illu-
minated by the orange light.
This scene was captured with a RAW (non-white balanced) camera, and is
shown at the top of Figure 11.6. The approximation of the immediate change
in appearance upon realizing it was the light from the window is shown in the
bottom frame of Figure 11.6. This immediate change in appearance was entirely
cognitive, and it can be attributed to a switch from a surface mode of viewing to
an illumination mode.
There are five modes of viewing that affect color appearance: illuminant, illu-
mination, surface, volume, and film. A brief overview of these modes is presented
here, though a more complete description can be found in the texts by the Optical
Society of America [860], or more recently by Fairchild [302].
The illuminant mode of appearance is color appearance based on the per-
ception of a self-luminous source of light. Color appearance perceptions in the
illuminant mode typically involve actual light sources, and more often then not,
these are the brightest colors in the viewing field, for example, looking at a traffic
light at night or at a light fixture indoors.
The cognitive assumption that the brightest objects in a scene are light sources
can lead to some interesting appearance phenomena when non-illuminant objects

i i

i i
 i i

i i

580 11. Color and Image Appearance Models

Figure 11.6. The wall of a room that appeared to be a mural of highly chromatic different
colored squares (top) and the change in appearance once it was obvious that it was orange
light entering from an open window (bottom).

in a scene appear much brighter than the surrounding scene. These objects can
sometimes be perceived in an illuminant mode and may be described as glowing.
Evans coined this perception brilliance, though Nayatani and Heckaman have
more recently touched upon on this topic [288, 448, 822].

i i

i i
 i i

i i

11.1. Vocabulary 581

Another example of an object appearing to glow may occur when there are
fluorescent objects in a scene. Fluorescence is found in materials that absorb en-
ergy at one wavelength and emit that energy as light at much longer wavelengths.
Fluorescent objects can appear to glow, in part because they absorb energy from
non-visible UV portions of the spectrum and emit this as light in the visible por-
tions. Thus, these objects can appear to be much brighter than the objects in the
surrounding scene, and they are often said to take on the appearance of a light
source.
The illumination mode of appearance is similar to the illuminant mode, except
that perceived color appearance is thought to occur as a result of illumination
rather than the properties of the objects themselves. Consider the example given
in Figure 11.6. Upon entering the room, it became evident that it was a single
color wall that was being illuminated through a series of windows. The orange
color was instantly recognized as coming from an illumination source. There are
many clues available to a typical observer of a scene when determining if color is
a result of illumination. These clues include the color of the shadows, the color
of the entire scene, complex color interactions at edges and corners, as well as the
color of the observer themselves.
The perceived color of the wall, as described above, is an example of the
surface mode of appearance. At first it appeared that the surface of the wall
was painted with a colored mural. In this mode of appearance, the color of a
surface is perceived as belonging to the object itself. For example, an observer
may know that the skin of an apple is red, and the apple will maintain a certain
degree of “redness” despite large changes in illumination. This is an example
of discounting-the-illuminant, as we discussed in Chapter 10. Any recognizable
object provides an example of the surface mode of appearance. For this mode of
appearance, we need both a physical surface and an illuminating light source.
The volume mode of appearance is similar in concept to the surface mode,
except that the color is perceived to be a part of a bulk or volume of a transparent
substance. An example of the volume mode of appearance can be found in the
perceived color of liquids, such as a cup of coffee. Coffee is often classified into
categories such as light, medium, and dark, based upon the color of the brewed
coffee itself. The color of the coffee is not thought to be just a characteristic
of the surface, but rather it exists throughout the entire volume. Adding milk to
the coffee increases the scattering and the beverage becomes opaque. This is an
example of a volume color changing into a surface color. Volume color requires
transparency as well as a three-dimensional shape and structure.
The final mode of appearance, called either the aperture or film mode, encom-
passes all remaining modes of appearance. In the film mode, color is perceived

i i

i i
 i i

i i

582 11. Color and Image Appearance Models

as an aperture that has no connection with any object. One example is a photo-
graphic transparency on a light-box. Any object can switch from surface mode to
aperture mode, if there is a switch in focus from the surface itself. This can be
accomplished purposely, by using a pinhole or a camera lens system. The differ-
ences in our perception between a soft-copy display device and a hard-copy print
may be associated with the different modes of viewing. The print is generally
considered to be viewed as a surface mode, while the self-luminous display may
be viewed in film mode or perhaps in an illuminant mode.

11.2 Color Appearance Phenomena

This section deals with common examples of colored stimuli that do not follow
the predictions of basic colorimetry. Certain changes in the viewing field cause
changes in appearance that cannot be predicted with XYZ tristimulus matching.
The CIE system of colorimetry was developed using a color-matching experiment,
as discussed in Chapter 7. To recap, colorimetry states that if two stimuli have
identical tristimulus values then those two stimuli will match each other for a
single given viewing condition. Colorimetry does not attempt to predict if the
colors will match if any aspect of the viewing condition changes.
This section will illustrate several examples of where the color matches will
indeed breakdown, as various elements of the viewing field are changed. Among
the changes in viewing condition are changes in illumination level, illumination
color, surround, background, and viewing mode. The examples shown here il-
lustrate the limitations of basic colorimetry and the need for advanced colorime-
try, also called color appearance modeling. The foundations of most color ap-
pearance models stem from the study of these individual phenomena, so it is
important to review and understand our behavior. The recognition and under-
standing of these color appearance phenomena are also important for a color
imaging-system designer or computer-graphics researcher, as many of these ex-
amples are very common in their occurrence. This section will describe several
common color appearance phenomena, including those caused by local spatial
structure, luminance or illuminant color changes, and changes in the viewing
surround.

11.2.1 Changes in Appearance Caused by Spatial Structure

Perhaps the most easily recognized color appearance phenomenon is that of si-
multaneous contrast. This was introduced briefly in Section 5.4.2. Figure 11.7
illustrates an example of simultaneous lightness contrast. Each of the rows of the

i i

i i
 i i

i i

11.2. Color Appearance Phenomena 583

Figure 11.7. An example of simultaneous lightness contrast. Each row of gray patches
has identical luminance.

figure are gray patches of identical luminance, with each subsequent row decreas-
ing in luminance. When placed upon a background with a white-to-black gray
gradient, we see that the patches in each row actually appear of a different light-
ness (or perception of luminance) creating the illusion of a dark to light contrast
or tone scale.
Similarly, the relative contrast of the tone scale in each of the columns also
appears different. These observed differences are a result of the background in
the viewing field. It is for this reason that we should not consider the perception
of contrast to be a global property of an image measured by the ratio of maximum
white to black, also known as Michelson contrast. Rather, we should consider
contrast to be both a global and local perception. Peli gives an excellent overview
of contrast in complex images, and more recently Calabria and Fairchild measured
contrast perception for color imaging [146, 147, 888].
Simultaneous contrast is not limited to changes in lightness perception, as it
can also cause the color of a stimulus to shift in appearance when the color of

i i

i i
 i i

i i

584 11. Color and Image Appearance Models

the background changes. This is illustrated in Figure 11.8, where each of the
small patches has an identical XYZ tristimulus value. The change in color of a
stimulus tends to follow the opponent color theory of vision. The patches on the
yellow gradient should appear more blue, while the patches on the blue gradient
should appear more yellowish-green. Interesting to note in Figure 11.8 is that
the colors of any neighboring patches do not appear very different, though the
appearance change across the entire field is readily apparent. This again suggests a
very localized perception of appearance. Shevell et al. present some very striking
examples of this spatially localized simultaneous contrast; they also developed
visual models to predict this spatial relationship [67, 1039]. In general, as the
spatial frequency of a stimulus increases, the contrast effect actually ceases and
in some cases reverses and may assimilate the background color or spread its own
color to the background.

Figure 11.8. An example of simultaneous color contrast. Each of the patches has identical
CIE XYZ tristimulus values.

i i

i i
 i i

i i

11.2. Color Appearance Phenomena 585

Figure 11.9. An example of spreading. Only the lines contain color, though that color
appears in the background as well (adapted from [630]).

At a high-enough spatial frequency, simultaneous contrast is replaced with

spreading. With spreading, the color of a stimulus mixes with the color of the
background. Recall that in the case of simultaneous contrast, the color of a
stimulus takes on the opposite color of the background. One hypothesis is that
spreading is caused by optical blurring of the light coming from the stimulus and
mixing with the light coming from the background. While this might be true for
very high-frequency stimuli, such as half-tone dots, it does not fully explain the
spreading phenomenon. Spreading often occurs when the stimuli are very dis-
tinct from the background. An example is shown in Figure 11.9. The only colors
present in this figure are the black or white backgrounds, and the colored lines,
though it should appear as if the color portions of the lines have spread into the
background.
Research is ongoing to understand the transition point between simultane-
ous contrast and spreading and the overall effects of spatial frequency on color
appearance [67, 1039]. Related appearance phenomena, though arguably more
complex, include neon spreading and the watercolor effect. Neon spreading com-
bines spreading with the perceptual attribute of transparency, while the watercolor
effect generates the appearance of a surface color based entirely on edge bound-
aries. Bressan provides an excellent review of neon spreading [127], and Pinna
does the same for the watercolor effect [898].
Simultaneous contrast can also give rise to an increase in perceived color dif-
ference between color stimuli. This effect is known as crispening and can be
observed in Figure 11.10. Crispening causes an increase in perceived color differ-
ence between two stimuli when the viewing background is similar to the color of

i i

i i
 i i

i i

586 11. Color and Image Appearance Models

Figure 11.10. An example of lightness crispening. The color difference between the
two small patches is the same for all backgrounds, but should appear greatest when the
background is close in color to the stimuli.

the stimuli. In the figure, the differences between the small gray patches are the
same for all three backgrounds, but the difference looks the largest on the gray
background. Similar effects can be seen for color patches, known as chromatic
or chroma crispening. More details on lightness and chroma crispening can be
found in Semmelroth [1025] and Moroney [792].

11.2.2 Luminance-Induced Appearance Phenomena

The above phenomena involve appearance changes as a function of spatial struc-
ture or color-background viewing field. Large color shifts can also occur when
the viewing illumination changes. We have already introduced the color changes
caused by chromatic adaptation in Chapter 10. Color appearance is also dictated
by the overall amount of energy or illumination. Luminance changes are very
common in everyday life, and we are probably all familiar with the effect on
color appearance, even if we do not realize it at first. The classic example is the
difference between looking outside on a bright sunny day or on a dark overcast

i i

i i
 i i

i i

11.2. Color Appearance Phenomena 587

day. Objects tend to appear very bright and colorful on a sunny day and somewhat
subdued on an overcast day. These occurrences can be well described by both the
Hunt effect and the Stevens effect.
The Hunt effect states that as the luminance of a given color increases, its
perceived colorfulness also increases. This effect was first identified in a seminal
study by Hunt on the effects of light and dark adaptation on the perception of
color [504]. A haploscopic matching experiment was performed, similar to that
used to generate the CIE Standard Observers. However, in haploscopic matching,
observers are presented with one viewing condition in their left eye and another
in their right eye.
In the Hunt experiment, observers used the method of adjustment to create
matches on stimuli viewed at different luminance levels for each eye. When the
adjusting stimulus was present at low luminance levels, the observers required a
significant increase in the colorimetric purity to match a stimulus viewed at a very
high luminance level. This indicated that the perception of colorfulness is not in-
dependent of luminance level. If we think about this in terms of a chromaticity
diagram, the perceived chromaticity shifts toward the spectrum locus as the lu-
minance levels increase, or shifts toward the adapted white as luminance levels
decrease.
Going back to the sunny day example, the Hunt effect partially explains why
objects appear much more vivid or colorful when viewed in a bright sunny envi-
ronment. To summarize, the Hunt effect states that as absolute luminance levels
increase the colorfulness of a stimulus also increases.
Similar to the Hunt effect, we also see that brightness is a function of chro-
matic content. The perception of brightness or lightness is often erroneously as-
sumed to be a function of luminance level alone. This is not the case and is well
illustrated by the Helmholtz-Kohlrausch effect. The Helmholtz-Kohlrausch effect
shows that brightness changes as a function of the saturation of a stimulus; i.e., as
a stimulus becomes more saturated at constant luminance, its perceived brightness
also increases.
Another way to describe this effect is say that a highly chromatic stimulus
will appear brighter than an achromatic stimulus at the same luminance level.
If brightness were truly independent of chromaticity, then this effect would not
exist. It is important to note that the Helmholtz-Kohlrausch effect is a function
of hue angle as well. It is less noticeable for yellows than purples, for instance.
Essentially this means that our perception of brightness and lightness is a function
of saturation and hue, and not just luminance. Fairchild and Pirrotta published a
general review of the Helmholtz-Kohlrausch effect as well as some models for
predicting the effect [295].

i i

i i
 i i

i i

588 11. Color and Image Appearance Models

Another interesting relationship between luminance level and chromatic col-

ors is the Bezold-Brücke hue shift. This phenomenon relates the perceived hue of
monochromatic light sources with luminance level. It is often assumed that the
hue of monochromatic light can be described completely by its wavelength. This
is not the case, as the hue of a monochromatic light will shift as the luminance
of the light changes. The amount of hue shift also changes both in direction and
magnitude as a function of the wavelength. Experimental results regarding the
Bezold-Brücke hue shift can be found in work published by Purdy [929].
One very important consideration for these hue shifts is that all of the experi-
mental data were obtained using unrelated colors. Recall that unrelated colors are
stimuli viewed in complete isolation and that they occur rarely in everyday life.
Perhaps luckily for digital-imaging applications, Hunt published a report indicat-
ing that the Bezold-Brücke hue shift disappears for related colors [506]. As such,
most modern color appearance models designed for imaging applications do not
take this phenomenon into account.
A further very important color appearance phenomenon that must be account-
ed for is the overall change in contrast as a function of luminance level. This in-
crease in contrast was examined closely in a classic study by Stevens and Stevens
[1085]. Their study indicated that as luminance level increases so too does the
brightness contrast (see Section 6.4); this effect has come to be referred to as the
Stevens effect.
In their original study, observers performed magnitude-estimation experi-
ments on the brightness of stimuli across many different luminance-adapting con-
ditions. The results showed that brightness tended to follow a power-law rela-
tionship with luminance, thus forming the basis for Stevens’ power law. This
nonlinear compressive relationship between luminance and brightness, as well
as lightness, is a very important behavior of our visual system. However, this
study also shows that the exponent of the power function changes as a function of
adapting-luminance level. Essentially, as the adapting-luminance level increases,
bright colors tend to look brighter, and darker colors tend to look darker. So, as
the adapting-luminance level increases, the rate of change between the brightness
of the dark and light colors increases. This rate of change is often considered the
contrast of the scene; we can therefore say that contrast increases with absolute
luminance levels.
It is interesting to note that most of these color appearance phenomena can
influence our perception of the contrast in images. The Stevens effect is the most
obvious of these, but both the Hunt effect and the Helmholtz-Kohlrausch effect
can also increase perceived contrast. This suggests that there are strong inter-
actions between color appearance attributes, and they should not be taken to be

i i

i i
 i i

i i

11.2. Color Appearance Phenomena 589

orthogonal or independent perceptions. Any change in one of these attributes may

actually result in undesirable (or desirable) changes in the others.

11.2.3 Hue Phenomena

Clearly luminance changes can cause large shifts in the appearance of colored
stimuli, indicating the importance of absolute luminance measurements for pre-
dicting appearance rather than just relative colorimetric data. This section exam-
ines several phenomena that result from changing the hue of the viewing con-
ditions, unrelated to the overall changes caused by chromatic adaptation. These
hue changes are much less common than luminance changes and often not very
perceptible or negligible for most color-imaging applications. They are included
here as a historical note, and because many models of color appearance have been
developed that are capable of compensating for these effects.
As shown above, the Bezold-Brücke hue shift illustrated that we cannot sim-
ply use the wavelength of monochromatic light sources as a predictor of perceived
hue. As luminance levels change, the perceived hue can also change. A similar
effect is known as the Abney effect, which states that adding a broad spectrum
“white” light to a monochromatic light does not preserve constant hue. Another
way of expressing this is to say that straight lines in a chromaticity diagram, ra-
diating from the chromaticity of the white point to the spectral locus, are not
necessarily lines of constant hue. Unlike the Bezold-Brücke hue shift, this effect
is valid for related colors as well as unrelated colors. The Abney effect indicates
another area of the visual system that behaves as a nonlinear system. More details
of both the Bezold-Brücke and the Abney effect can be in found in [929].
Another interesting, though difficult to reproduce, phenomenon involving
monochromatic illumination is the Helson-Judd effect [455]. In essence, this ef-
fect says that non-selective (gray) stimuli viewed under highly chromatic illumi-
nation take on the hue of the light source if they are lighter than the background,
and take on the complementary hue if they are darker than the background. So a
dark-gray sample viewed on a medium-gray background under red illumination
will look somewhat green, while a light-gray sample would appear pinkish. This
effect almost never occurs in common practice, and it is very difficult to repro-
duce even in a controlled experiment. Some color appearance models do take this
effect into account in their predictions, and so it is worth mentioning. More re-
cently, Fairchild has shown that this effect disappears when dealing with complex
stimuli, such as images [302].

i i

i i
 i i

i i

590 11. Color and Image Appearance Models

11.2.4 Changes in Appearance Caused by the Surround

The Stevens effect demonstrated that contrast for simple patches increases as
a function of the adapting luminance and the luminance of the stimulus itself.
Bartleson and Breneman studied the effects of the luminance level of the surround
on complex stimuli, specifically images [72]. Recall that in color appearance ter-
minology, the surround of the viewing field is considered to be the entire field
outside of the background and can generally be considered the entire viewing
room. Bartleson and Breneman found results that were very similar to Stevens
with regard to changes in overall luminance level.
More interestingly, they noticed interesting results with respect to the change
in the relative luminance of the image surround when compared to the luminance
levels of the stimulus itself. It was found that perceived contrast in images in-
creases as the luminance of the surround increases but the luminance of the stim-
ulus remains constant. Another way of saying this is that when an image is viewed
in a dark surround, the black colors look lighter while the light colors remain rel-
atively constant. As the surround luminance increases, the blacks begin to look
darker, causing overall image contrast to increase. More recently, Liu found simi-
lar results using a different approach of allowing observers to adjust the luminance
and color of the surround for optimal image quality [689].
This behavior closely matches a practice that historically has taken place in
the photographic world. For optimum tone reproduction, photographic-film trans-
parencies designed for viewing in a darkened room have a much higher contrast
than those designed for viewing as a print in a bright room. Hunt and Fairchild
provide more in-depth analysis of the history and prediction of optimal tone repro-
duction for complex images, an important consideration for overall image qual-
ity [302, 509]. In their original publication, Bartleson and Breneman published
equations that predicted their results well, though these equations were later sim-
plified to create equations for calculating optimal tone reproduction. Such equa-
tions have been adapted and are included in the predictions of many models of
color appearance, including CIECAM02 [302].
Surround compensation can play a key part in the design and implementation
of a color-reproduction system, for example, producing a computer-rendered fea-
ture film that will be viewed originally in a darkened movie theater and then later
at home on a television in a dim or average surround. If care is not taken, the
movie may appear to have too much contrast when viewed at home. This phe-
nomenon actually seems counter-intuitive to most people, as many of us actually
turn off the surrounding lights when viewing movies at home to obtain a higher
contrast. While it is true that turning off the surround lights will actually reduce
the perception of contrast, what most people are actually doing is removing flare

i i

i i
 i i

i i

11.3. Color Appearance Modeling 591

from the surface of the screen. Flare can be considered stray light that is not part
of the desired stimulus; it has the effect of reducing overall contrast by essentially
increasing the luminance of the black. We should not confuse the removal of flare
with the change in perception caused by the luminance of the surround.

11.3 Color Appearance Modeling

CIE tristimulus colorimetry was originally created for a single, relatively sim-
ple, purpose, namely predicting when two stimuli will match for the average
observer, under a single viewing condition. For this task, colorimetry has been
remarkably successful for science and engineering purposes. We have already
seen the limitations of basic colorimetry with some of the color appearance phe-
nomena described above, since altering any part of the viewing field can drasti-
cally change the appearance of a stimulus. Chromatic-adaptation transforms, as
described in Chapter 10, extend basic colorimetry so that it is possible to predict
matches across changes in the color of the illumination across viewing conditions.
Chromatic-adaptation transforms do not yet provide enough information to allow
us to describe the actual color appearance of a stimulus.
To accurately describe the color appearance of a stimulus, we can use the
color terminology described above, including the relative terms of lightness, sat-
uration, and chroma, as well as the absolute terms of brightness and colorfulness.
The attribute of hue can be used to describe either absolute or relative appearance.
Even with the addition of a chromatic-adaptation transform, CIE tristimulus col-
orimetry is not able to predict any of these appearance terms. In order to generate
measurements corresponding to these attributes, we must use a color appearance
model.
So what is a color appearance model? In its simplest form, we can think
of a color appearance model as a tool to translate numbers and measurements
into words describing a color. The CIE Technical Committee TC1-34, Testing
Colour Appearance Models, defined exactly what is necessary to be considered a
color appearance model [191]. The definition agreed upon is as follows: A color
appearance model is any model that includes predictors of at least the relative
color appearance attributes of lightness, chroma, and hue [191].
This is a very lenient definition of what constitutes a color appearance model,
though it does capture the essence. At the very least, we can say that color ap-
pearance models minimally require some form of a chromatic-adaptation trans-
form, as well as some form of a uniform color space. Color appearance mod-
els that are more complicated are capable of predicting absolute attributes, such
as brightness and colorfulness, as well as luminance-dependent effects such as

i i

i i
 i i

i i

592 11. Color and Image Appearance Models

the Hunt and Stevens effects. Spatially structured phenomena, such as crispen-
ing and simultaneous contrast, require models of spatial vision as well as color
appearance.
There are many color appearance models that have been developed, each de-
signed with specific, and perhaps different, goals in mind. Some of these models
that one may come across include CIELAB, Hunt, Nayatani, ATD, RLAB, LLAB,
ZLAB, CIECAM97s, and CIECAM02. The interested reader is encouraged to ex-
amine Fairchild’s thorough examination on the history and development of many
of these models [302]. Most of these models are beyond the scope of this text,
though we will examine CIELAB as well as CIECAM02 as simple color appear-
ance predictors. The latter is the current recommended color appearance model
from the CIE; it is based upon the strengths of many of the other models men-
tioned above.

11.3.1 CIELAB as a Color Appearance Model

As we first introduced in Chapter 8, and revisited in Chapter 10, CIELAB is a
color space designed for predicting small color differences. It was designed to be
perceptually uniform for the ease of calculating these color differences. Despite
this design intention, CIELAB has been used with varying degrees of success
as a color appearance model. This is because CIELAB does include predictors
of lightness, chroma, and hue. These attributes are generated by performing a
rectangular-to-cylindrical transformation on the CIE a∗ and b∗ coordinates. These
predictors allow CIELAB to be labeled as a color appearance model. We will
use it here as a simple model to illustrate the design of more complicated color
appearance models.
CIELAB calculations require a pair of CIE XYZ tristimulus values: those of
the stimulus itself, as well as those of the reference white point. Already we see
that CIELAB requires more information about the viewing conditions than stan-
dard tristimulus color matching, as do all color appearance models. The reference
white-point values are used in a von Kries-type chromatic-adaptation transform.
Recall from Chapter 10 that this chromatic-adaptation transform is less than opti-
mal and may be responsible for some of the hue nonlinearity associated with the
CIELAB space.
The adaptation transform is followed by a compressive cube-root nonlinearity
and an opponent color transformation. The cube-root power functions attempt to
model the compressive relationship between physical measurements of luminance
and psychological perceptions. These compressive results were discussed above
with regard to the Stevens effect. The cube-root function is replaced by a linear

i i

i i
 i i

i i

11.3. Color Appearance Modeling 593

function for very dark stimuli. The CIELAB L∗ coordinate, as a correlate to

perceived lightness, is a function purely of luminance. It ranges from 0.0 for
absolute black stimuli to 100.0 for diffuse white stimuli. It is possible to calculate
L∗ values greater than 100 for specular highlights or fluorescent materials that
have a higher luminance than the input white point, though care should be taken
to fully comprehend the meaning of these values.
The a∗ and b∗ coordinates approximate, respectively, the red-green and
yellow-blue of an opponent color space. A positive a∗ value approximates red
while a negative approximates green. Similarly, a positive b∗ correlates with yel-
low while negative values correlate to blue. It should be noted that only the phys-
iological unique hue of yellow lines up with the b∗ axis while the other unique
hues are a combination of both the chromatic components. Purely achromatic
stimuli such as whites, grays, and blacks should have values of 0.0 for both a∗
and b∗ .
The definition of a color appearance model requires a minimum of predictions
for lightness, chroma, and hue. The CIELAB L∗ provides us with a lightness
prediction, but a∗ and b∗ do not provide us with correlates of chroma and hue.
These correlates can be calculated by transforming the Cartesian coordinates of
a∗ and b∗ into cylindrical coordinates of Cab ∗ and hab , where Cab ∗ represents
chroma and hab represents hue angle (see Section 8.7.2).
With the cylindrical coordinates of chroma and hue angle, we now have
enough information to predict or describe the general color appearance of a stim-
ulus. There are several important caveats, however, that should be considered.
The wrong von Kries transform is clearly a source of color appearance errors.
CIELAB is also incapable of predicting many of the color appearance phenom-
ena described above. These include all luminance, surround, background, and
discounting-the-illuminant effects. CIELAB also assumes 100 percent complete
adaptation to the white point.
Since CIELAB was really only designed to predict small color differences
between similar objects under a single viewing condition, it is impressive that it
can be used as a color appearance model at all. Many color-management utilities
actually use CIELAB as a method of specifying device-independent colors as well
as a method for transferring between different white points. For such applications,
a user may be better off using a more advanced chromatic-adaptation transform
such as those described in Chapter 10 along with CIELAB, rather than relying on
the chromatic-adaptation transform in CIELAB.

i i

i i
 i i

i i

594 11. Color and Image Appearance Models

11.3.2 The CIECAM02 Color Appearance Model

Color appearance research has been ongoing for many decades, and the fruition
of this research resulted in the formulation of many different color appearance
models. Many of these models were designed with different goals using a va-
riety of experimental data. Until the late 1990s, it was often difficult to choose
which model to use for any given task, resulting in industry confusion. This
confusion was alleviated in 1997 with the formulation of the CIE recommended
CIECAM97s color appearance model [191]. The CIECAM97s model was de-
signed to work at least as well as all of the existing models for the color appear-
ance phenomena that it predicts. It can be viewed as essentially a hybrid of the
best parts of many different models. The general pedigree of CIECAM97s arose
from components in the Hunt, Nayatani, RLAB, and LLAB models [191].
The s in CIECAM97s represented the initial attempt to create a simplified
color appearance model, for general-purpose imaging applications. The initial
goal had been to create a comprehensive (CIECAM97c) version of the model
that included predictions for a wider array of appearance phenomena, as well
as including rod contributions so it could be used for very low-level scotopic and
mesopic applications. The industry demands for the more comprehensive model
were not there, and as such this model was never developed [302].
That the year 1997 was included in CIECAM97s was an indicator that the CIE
recognized that this was an interim recommendation designed to focus color ap-
pearance research on a single unified model and provide a benchmark for further
appearance-model development. This approach was very successful and has led
directly to the creation of the CIECAM02 color appearance model. CIECAM02
is the currently recommended color appearance model, specifically designed for
color-management systems and digital color imaging [195]. CIECAM02 was de-
signed specifically to account for some weaknesses that were discovered in the
CIECAM97s model and was intended to be simpler and more easily inverted
[195]. That the year is still included in the name of CIECAM02 indicates that
the CIE still considers this an interim model in this very active field. Though it is
simpler than CIECAM97s, it is a more accurate model based upon existing visual
data and is recommended to replace CIECAM97s in most applications.
The forward and inverse calculations of CIECAM02 are presented below.
As a general workflow, the input to CIECAM02 consists of XYZ tristimulus
values. These are transformed into a sharpened cone space for the chromatic-
adaptation transform. The adapted signals are then transformed into cone sig-
nals and processed through a nonlinear hyperbolic-type compressive transform.
The compressed signals are transformed into the appearance attributes discussed
above. CIECAM02 is capable of predicting all six appearance attributes: light-

i i

i i
 i i

i i

11.3. Color Appearance Modeling 595

Surround condition c NC F
Average 0.69 1.0 1.0
Dim 0.59 0.9 0.9
Dark 0.525 0.8 0.8

Table 11.1. Input parameters for the surround conditions in CIECAM02.

ness, brightness, chroma, colorfulness, saturation, and hue. CIECAM02 is also

designed to predict the changes in color appearance as a function of luminance,
including the Hunt and Stevens effects, as well as the changes in appearance as
a function of the surround. Additionally CIECAM02 is designed to predict the
changes in lightness and brightness perception as a function of chromatic content.
The inputs to CIECAM02 include the CIE XYZ tristimulus values of the stim-
ulus (X, Y , and Z), and like CIELAB, the CIE XYZ tristimulus values of the
adapting white (Xw , Yw , and Zw ). In addition, the absolute luminance in cd/m2
of the adapting white, LA , is input. The luminance of the adapting white is typi-
cally taken to be 20 percent of the luminance of a white object in the scene, or 20
percent of the maximum output luminance of a self-emitting display. The relative
background-luminance factor, Yb , is also input into CIECAM02. The surround
conditions are generally defined to be one of three categories: average, dim, or
dark. If the user is unsure of which category to select, they can calculate the sur-
round ratio, SR , between the luminance of a white in the surround LSW and the
luminance of the scene or device white LDW as follows:

LSW
SR = . (11.3)
LDW

If the surround ratio is zero, then a dark surround should be chosen; if it is less
than 0.2, a dim surround should be selected. An average surround is selected for
all other values. The values c, Nc , and F are then input from Table 11.1. It is
possible to linearly interpolate the values from two surround conditions, though it
is generally recommended to use the three distinct categories.

11.3.3 Chromatic Adaptation

The first stage of the CIECAM02 calculations is the chromatic-adaptation trans-
form. This process is identical to that discussed in Section 10.4.6; it is repeated
here for completeness. The XYZ tristimulus values of the stimulus and the white
point are first transformed into sharpened cone signals, represented by RGB in

i i

i i
 i i

i i

596 11. Color and Image Appearance Models

CIECAM02: ⎡ ⎤ ⎡ ⎤
R X
⎣G⎦ = MCAT02 ⎣Y ⎦ , (11.4)
B Z
with MCAT02 defined as
⎡ ⎤
0.7328 0.4296 −0.1624
MCAT02 = ⎣−0.7036 1.6975 0.0061⎦ . (11.5)
0.0030 0.0136 0.9834

The next stage is to calculate the degree of chromatic adaptation D, which is

taken to be a linear scale between the adaptation to the current viewing condition
and the reference illuminant, which is CIE Illuminant E (an illuminant with an
equal-energy spectrum) for CIECAM02. A D value of 1.0 implies we are com-
pletely adapted to the current viewing condition, while a factor of 0.0 implies we
are fully adapted to the reference condition. The degree of adaptation D can be
explicitly set to a value based upon an understanding of the viewing conditions,
or can be implicitly calculated. If it is implicitly calculated, it is taken to be a
function of only the adapting luminance LA and surround factor F.
    
1 −(LA + 42)
D = F 1− exp . (11.6)
3.6 92

This equation was designed to produce a maximum value of 1.0 for a stimulus
viewed at 1000 lux.
The adapted cone signals are then calculated using a simple linear von Kries-
style adaptation transform:
 
Yw D
Rc = + 1 − D R, (11.7a)
Rw
 
Yw D
Gc = + 1 − D G, (11.7b)
Gw
 
Yw D
Bc = + 1 − D B. (11.7c)
Bw

Note that these equations multiply back in the relative luminance of the adapt-
ing white Yw . While typically this is taken to be 100, there are some applications
where this is not the case. Fairchild presents a compelling argument that Yw should
always be replaced with 100.0 in these calculations, especially when using the
chromatic-adaptation transform outside of CIECAM02 [302].

i i

i i
 i i

i i

11.3. Color Appearance Modeling 597

There are several other factors that can be considered part of the greater
chromatic-adaptation transform. These terms, although not explicitly part of the
von Kries adaptation, account for changes in chromatic adaptation as a function
of overall luminance levels, as well as some background luminance and induction
factors that can account for some color appearance changes as a function of the
background luminance. These factors are calculated using

1
k= , (11.8a)
(5LA + 1)
 2
FL = 0.2k4 (5LA ) + 0.1 1 − k4 (5LA )1/3 , (11.8b)
Yb
n= , (11.8c)
Yw
1
Nbb = Ncb = 0.725 0.2 , (11.8d)
√ n
z = 1.48 + n. (11.8e)

The adapted RGB values are then converted into cone signals prior to the non-
linear compression. This is done by inverting the MCAT02 matrix to obtain adapted
XY Z tristimulus values and then transforming those into cone signals using the
Hunt-Pointer-Estevez transform, normalized for the reference equal-energy con-
dition. Note that the RGB terminology is used in CIECAM02 rather than LMS
to maintain consistency with older color appearance models. This transform is
given by:
⎡ ⎤ ⎡ ⎤
R Rc
⎣G ⎦ = MHPE M −1 ⎣Gc ⎦ ; (11.9a)
CAT02
B Bc
⎡ ⎤
0.38971 0.68898 −0.07868
MHPE = ⎣−0.22981 1.18340 0.04641⎦ ; (11.9b)
0.00000 0.00000 1.00000
⎡ ⎤
1.096124 −0.278869 0.182745
−1
MCAT02 = ⎣ 0.454369 0.473533 0.072098⎦ . (11.9c)
−0.009628 −0.005698 1.015326

The cone signals are then processed through a hyperbolic nonlinear compres-
sion transform. This transform was designed to be effectively a power function of
a wide range of luminance levels [302]. These compressed values RGBa are then

i i

i i
 i i

i i

598 11. Color and Image Appearance Models

used to calculate the appearance correlates:

 
R 0.42
400 FL
100
Ra =   + 0.1, (11.10a)
FL R 0.42
27.13 +
100
 
G 0.42
400 FL
100
Ga =   + 0.1, (11.10b)
FL G 0.42
27.13 +
100
 
B 0.42
400 FL
100
Ba =   + 0.1. (11.10c)
FL B 0.42
27.13 +
100

11.3.4 Calculations of Hue

The adapted cone signals can now be used to calculate the appearance correlates.
The first stage is to move into a temporary opponent color space, similar to a∗
and b∗ in the CIELAB calculations. A more accurate rectangular opponent color
space will be calculated at a later stage:

12 Ga Ba
a = Ra − + , (11.11a)
11 11
 
1  
b= Ra + Ga − 2Ba . (11.11b)
9

Hue angle is then calculated from these opponent changes in a manner identi-
cal to the calculation in CIELAB:
 
−1 b
h = tan . (11.12)
a

It is important to note that the CIECAM02 calculations expect hue angle to be

represented in degrees rather than radians. Also, in practical implementations, an
atan2() function that ranges from 0–360 degrees should be used rather than -180–
180. If this function is not available, the hue angle can be calculated by adding
180 if a is a negative number and adding 360 if a is positive and b is negative.

i i

i i
 i i

i i

11.3. Color Appearance Modeling 599

Red Yellow Green Blue Red

i 1 2 3 4 5
hi 20.14 90.00 164.25 236.53 380.14
ei 0.8 0.7 1.0 1.2 0.8
Hi 0.0 100.0 200.0 300.0 400.0

Table 11.2. Values used to calculate hue quadrature.

Next, an eccentricity factor et is computed:

1+
π  ,
et = cos h + 2 + 3.8 . (11.13)
4 180
This eccentricity factor is necessary, because the four physiological hues, red,
green, yellow, and blue, are not perceptibly equidistant from each other. This is
especially true between the blue and red hues, as there is a large range of percepti-
ble purple colors in that range. Note, that for this calculation, the cosine function
π
is expecting radians while the hue angle is represented in degrees; hence, the 180
term.
The data in Table 11.2 are then used to determine hue quadrature through a
linear interpolation between the calculated hue value and the nearest unique hue:

100 (h − hi ) /ei
H = Hi + . (11.14)
(h − hi ) /ei + (hi+1 − h) /ei+1

11.3.5 Brightness and Lightness Calculations

The appearance correlates of lightness and brightness are calculated by first de-
termining the achromatic signal, which is essentially a linear combination of the
“cone” signals, along with the background-induction factor Nbb :
0 1
A = 2Ra + Ga + (1/20)Ba − 0.305 Nbb . (11.15)

The relative attribute of lightness J is then calculated with this achromatic

signal, processed through another nonlinear function that is determined by the
surround input factors c along with z, the base nonlinear component calculated
earlier:  cz
A
J = 100 . (11.16)
Aw
The absolute appearance attribute of brightness Q is then calculated from the
lightness signal. This calculation is a function of the viewing surround c along

i i

i i
 i i

i i

600 11. Color and Image Appearance Models

with the absolute luminance level function FL . This accounts for both the Stevens
effect as well as the changes in perceived contrast as a function of the surround:
 
4 J
Q= (Aw + 4) FL0.25 . (11.17)
c 100

11.3.6 Chroma, Colorfulness, and Saturation

The chromatic attributes of chroma, colorfulness and saturation can now be cal-
culated. They are functions of the lightness and brightness to account for the Hunt
effect. First, a temporary magnitude value t is calculated:
 1/2
(50000/13) Nc Ncb et a2 + b2
t= . (11.18)
Ra + Ga + (21/20)Ba
This value has components that are very similar to the CIELAB calculation of
chroma, but it also includes factors accounting for the background- and surround-
induction values, as well as for the hue-eccentricity value.
The temporary magnitude value t is then used along with the lightness calcula-
tion and the background exponent of n to ultimately calculate the relative attribute
of chroma C: 
J
C=t 0.9
(1.64 − 0.29n )0.73 . (11.19)
100
From the calculation of chroma, along with the luminance-adaptation factor
FL , we can calculate the absolute correlate of colorfulness M. We can also use the
colorfulness and brightness values to calculate a value for saturation s:
M = C FL0.25 , (11.20a)

M
s = 100 . (11.20b)
Q

11.3.7 Opponent Color Representation

Sometimes it may be desirable to have the rectangular, or Cartesian, representa-
tion of the appearance coordinates. These can be calculated from the hue angle,
along with the chroma, colorfulness, or saturation attributes:
aC = C cos(h), (11.21a)
bC = C sin(h), (11.21b)
aM = M cos(h), (11.21c)
bM = M sin(h), (11.21d)

i i

i i
 i i

i i

11.3. Color Appearance Modeling 601

as = s cos(h), (11.21e)
bs = s sin(h). (11.21f)

11.3.8 Inverting CIECAM02

Many times all we are interested in is determining the appearance attributes of a
given stimulus. We can consider these attributes as representing not only a device-
independent color, but also a viewing-condition-independent color. If all we want
is to explain what a stimulus looks like, then the forward model of CIECAM02 is
sufficient.
However, one of the most powerful applications of a color appearance model
involves the prediction of corresponding colors through both a forward and in-
verse transform. For example, suppose we have an image displayed on a com-
puter screen that we are viewing in our office, and we want to print that image.
We can use a color appearance model to account for the changes in these dis-
parate viewing conditions and calculate the tristimulus values necessary to get a
print that matches the screen. In most typical color-management situations, we
desire an appearance match rather than a direct colorimetric match.
The surround parameters of the output viewing condition are used for the
inversion, as determined from Table 11.1. The first stage is to calculate the achro-
matic signal for the output white Aw using (11.15). There may be times when we
start out with only a subset of the appearance attributes. In these cases, we must
first calculate the relative appearance attributes before inverting. As such, if the
input to the inverse model is absolute brightness Q, we must first calculate the
relative lightness response J:
 2
cQ
J = 6.25 . (11.22)
(Aw + 4) FL0.25
If we start with colorfulness M, then we first calculate chroma:
M
C= . (11.23)
FL0.25
If we start with lightness J and saturation s, then we first must calculate brightness
in order to calculate chroma:

4 J
Q= (Aw + 4) FL0.25 , (11.24a)
c 100

s 2 Q
C= . (11.24b)
100 FL0.25

i i

i i
 i i

i i

602 11. Color and Image Appearance Models

If we have the hue angle h, then we can proceed, or we can calculate it from hue
quadrature H and Table 11.2:

(H − Hi ) (ei+1 hi − ei hi+1 ) − 100 hi ei+1

h = , (11.25a)
(H − Hi ) (ei+1 − ei ) − 100 ei+1

(h − 360) if h > 360,
h= (11.25b)
h otherwise.

From there, we calculate the temporary magnitude value t and the eccentricity
value et . It is important to note that for the remainder of these equations we use
the parameters from the output viewing conditions:

⎛ ⎞1/0.9
⎜ C ⎟
t =⎜
⎝
⎟
⎠ , (11.26a)
J
(1.64 − 0.29n )0.73
100
1

π  
et = cos h + 2 + 3.8 . (11.26b)
4 180

Next we calculate the achromatic signal A, along with the viewing-condition-

dependent components p1 , p2 , and p3 :

 
J 1/cz
A = Aw , (11.27a)
100
(50000/13) Nc Ncb et
p1 = , (11.27b)
t
 
A
p2 = + 0.305, (11.27c)
Nbb
21
p3 = . (11.27d)
20

We then convert hue angle into radians:

π
hr = h . (11.28)
180

i i

i i
 i i

i i

11.3. Color Appearance Modeling 603

If |sin (hr )| ≥ |cos (hr )|,

p1
p4 = , (11.29a)
sin (hr )
 
460
p2 (2 + p3 )
1403
b=        , (11.29b)
220 cos (hr ) 27 6300
p4 + (2 + p3 ) − + p3
1403 sin (hr ) 1403 1403
 
cos (hr )
a=b . (11.29c)
sin (hr )
Otherwise, if |sin (hr )| < |cos (hr )|,
p1
p5 = , (11.30a)
cos (hr )
 
460
p2 (2 + p3 )
1403
a=        ,
220 27 6300 sin (hr )
p5 + (2 + p3 ) − − p3
1403 1403 1403 cos (hr )
(11.30b)
 
sin (hr )
b=a . (11.30c)
cos (hr )
We now calculate the RGBa signals:
460 451 288
Ra =p2 + a+ b, (11.31a)
1403 1403 1403
460 891 261
Ga = p2 − a− b, (11.31b)
1403 1403 1403
460 220 6300
Ba = p2 − a− b. (11.31c)
1403 1403 1403
Then we undo the nonlinearities, again using the FL calculation of the output
viewing condition:
 
 100 27.13 |Ra − 0.1| 1/0.42
R = , (11.32a)
FL 400 − |Ra − 0.1|
 
 100 27.13 |Ga − 0.1| 1/0.42
G = , (11.32b)
FL 400 − |Ga − 0.1|
 
100 27.13 |Ba − 0.1| 1/0.42
B = . (11.32c)
FL 400 − |Ba − 0.1|

i i

i i
 i i

i i

604 11. Color and Image Appearance Models

Note that if any of the values of Ra − 0.1, Ga − 0.1, Ba − 0.1 are negative, then the
corresponding RGB must be made negative.
At this point, we leave the Hunt-Pointer-Estevez cone space and move into
the sharpened cone space for the inverse chromatic-adaptation transform:
⎡ ⎤ ⎡ ⎤
Rc R
⎣Gc ⎦ = MCAT02 M −1 ⎣G ⎦ ; (11.33a)
HPE
Bc B
⎡ ⎤
1.910197 −1.112124 0.201908
−1
MHPE = ⎣ 0.370950 0.629054 −0.000008⎦ . (11.33b)
0.000000 0.000000 1.000000

We calculate the adapted sharpened signals, again using a degree of adapta-

tion factor D that is explicitly determined for the output viewing conditions, or
implicitly calculated using Equation 11.6:
RC
R=  , (11.34a)
Yw D
+1−D
Rw
GC
G=  , (11.34b)
Yw D
+1−D
Gw
BC
B=  . (11.34c)
Yw D
+1−D
Bw

Finally we can transform the adapted signals back into CIE XYZ tristimulus
values. These values can then be further transformed into a device-specific space
for accurate color appearance reproduction between devices:
⎡ ⎤ ⎡ ⎤
X R
⎣Y ⎦ = M −1 ⎣G⎦ . (11.35a)
CAT02
Z B

We have included a C++ implementation of both forward and reverse versions

of CIECAM02 on the DVD included with this book.

11.3.9 Final Thoughts on CIECAM02

As we can see from this section, the CIECAM02 model is a fairly complicated
color appearance model that is capable of predicting a variety of changes in

i i

i i
 i i

i i

11.4. Image Appearance Modeling 605

appearance caused by changes in viewing conditions. The CIECAM02 model

was designed specifically for color-management applications, such as reproduc-
ing prints that accurately match a monitor. When the desired outcome for device
color management is appearance match, then CIECAM02 should be considered a
valuable tool in any color-engineer’s toolkit. Although CIECAM02 was designed
specifically for image-processing use, it should not be considered a spatial model
of color appearance. CIECAM02 treats all pixels in an image as individual color
stimuli and does not allow for complex spatial interactions, or calculations of lo-
cal contrast phenomena. Several spatial extensions of color appearance models
are mentioned in the next section.

11.4 Image Appearance Modeling

Color appearance modeling has been successful in facilitating device-independent
color representation across a variety of viewing conditions and is incorporated
into many modern color-management systems. While CIECAM02 is a very pow-
erful color appearance model, there remains significant room for improvement
and extension of its capabilities. Image appearance modeling is one such tech-
nique designed to address issues with respect to spatial properties of vision, image
perception, and image quality. Just as color appearance models can be considered
extensions of basic CIE colorimetry, image appearance models can be consid-
ered further extensions. These models evolved to combine attributes of color
appearance along with attributes of spatial vision models. Most have also evolved
specifically with attributes of color difference equations in mind, for the purpose
of better predicting perceived differences between spatially complex stimuli.
Spatial vision models that have been historically used for image-quality met-
rics generally focus on black-and-white quality. As mentioned in the previous
section, color appearance models have largely ignored spatial vision while spa-
tial vision models have largely ignored color. Examples of these types of spatial
vision models include the Daly Visible Differences Predictor (VDP) and, more
recently, Watson’s Spatial Standard Observer [229, 1219, 1221]. These models
predict image differences based only upon luminance information. Some spatial
models of vision that contain both color and spatial properties of stimuli include
the Retinex model as described in Section 5.8.4 by Land and McCann and its var-
ious derivatives and the S-CIELAB model [1292]. The CVDM metric is a further
refinement of the S-CIELAB model combined with attributes of visual masking
from VDP [550]. The Retinex model was not originally designed as a complete
model of image appearance and quality, though its spatially variable mechanisms

i i

i i
 i i

i i

606 11. Color and Image Appearance Models

of chromatic adaptation and color constancy serve some of the same purposes in
image rendering and certainly provide some of the critical groundwork for recent
research in image appearance modeling.
This section will focus on two models that have their pedigree in CIE col-
orimetry and color appearance modeling: S-CIELAB and iCAM. S-CIELAB is
a simple spatial extension to the CIELAB color space, and was designed to mea-
sure color differences of complex images. Fairchild and Johnson based iCAM on
S-CIELAB, combined with a spatially variant color appearance model [291,292].

11.4.1 S-CIELAB
The S-CIELAB model was designed as a spatial pre-processor to the standard CIE
color difference equations [1292]. The intention of this model was to account for
color differences in complex color stimuli, such as those between a continuous
tone image and its half-tone representation.
The spatial pre-processing uses separable convolution kernels to approximate
the contrast-sensitivity functions (CSF) of the human visual system. These ker-
nels behave as band-pass functions on the luminance channels and low-pass func-
tions for the chromatic opponent channels. It is important to emphasize that the
CSF kernels chosen are tied heavily into the color space in which they were de-
signed to be applied. The opponent color space chosen in S-CIELAB is based
upon color appearance data from a series of visual experiments on pattern color
performed by Poirson and Wandell [908, 909]. This space, and the corresponding
measured contrast-sensitivity functions, were both fit to the visual data and should
be used together.
For S-CIELAB, the CSF serves to remove information that is imperceptible to
the human visual system and to normalize color differences at spatial frequencies
that are perceptible. This is especially useful when comparing images that have
been half-toned for printing. For most common viewing distances, these dots are
not resolvable and tend to blur into the appearance of continuous colors. A pixel-
by-pixel color-difference calculation between a continuous image and a half-tone
image results in extremely large errors, while the perceived difference can actually
be small. The spatial pre-processing stage in S-CIELAB blurs the half-tone image
so that it more closely resembles the continuous tone image.
The original S-CIELAB implementation uses separable convolution kernels
to perform the spatial filtering, along with the traditional CIELAB color dif-
ference equations. Johnson and Fairchild introduced a slight refinement to this
model based upon spatial filtering in the frequency domain and the use of the
CIEDE2000 color-difference metrics [555]. This implementation has been folded

i i

i i
 i i

i i

11.4. Image Appearance Modeling 607

Input Image

Convert to ACC
Opponent Color Space

Spatial Filter

Convert to XYZ

Calculate CIELAB
coordinates

Figure 11.11. The general flow chart of S-CIELAB spatial pre-processing.

i i

i i
 i i

i i

608 11. Color and Image Appearance Models

into the iCAM image-difference calculations and will be discussed in

Section 11.4.2. This section will focus on the original S-CIELAB calculations.
The general flowchart of the S-CIELAB model is shown in Figure 11.11 for a
single image. In practice, this process is usually performed for an image pair, and
a pixel-by-pixel color difference is calculated on the output image pair.
The first stage in the calculation is to transform the input image into an op-
ponent color space based upon the Poirson and Wandell data, known as AC1C2 or
O1 O2 O3 [1292]. This space is defined as a linear transform from CIE XYZ tris-
timulus values, so it is necessary to first transform the input image into a colori-
metric representation using device characterization. The transformation into the
opponent color space approximating an achromatic channel, a red-green channel,
and a blue-yellow channel is given by
⎡ ⎤ ⎡ ⎤⎡ ⎤
A 0.279 0.720 −0.107 X
⎣C1 ⎦ = ⎣−0.449 0.290 0.077⎦ ⎣Y ⎦ . (11.36)
C2 0.086 0.590 −0.501 Z
The convolution kernels are weighted sums of discrete Gaussian functions.
The general form of the kernels is

f (x, y) = k ∑ ωi Ei (x, y) , (11.37a)

i
  2 
x + y2
Ei (x, y) = ki exp − . (11.37b)
σi2
Note that although these equations are given in two dimensions, due to their
separability they are computed in practice using one-dimensional convolutions.
The parameters ωi and σi for (11.37) are the weighting functions and half width
expressed in terms of spatial frequency. These parameters can be found in Ta-
ble 11.3. The k parameters are designed to normalize the kernels so that they sum

Channel Weight ωi Width σi

Achromatic 0.921 0.0283
0.105 0.133
-0.108 4.336
Red-green 0.531 0.0392
0.330 0.494
Blue-yellow 0.488 0.0536
0.371 0.386

Table 11.3. Convolution kernel parameters for the S-CIELAB model.

i i

i i
 i i

i i

11.4. Image Appearance Modeling 609

1.4

Sensitivity
Achromatic
1.2 Red−green
Yellow−blue
1.0

0.8

0.6

0.4

0.2

0.0
0 10 20 30 40 50 60
Spatial Frequency (cycles−per−degree)

Figure 11.12. The convolution kernels of S-CIELAB represented in the frequency do-
main. Notice the general approximation of the human contrast-sensitivity functions.

to 1.0. This assures that when solid patches are filtered and used as stimuli, the
S-CIELAB model reduces to a standard color difference equation.
The final kernel size is determined to be a function of the viewing conditions,
primarily the resolution of the display, the number of pixels in the image, and
the viewing distance. Equation (11.38) shows this relationship for an image size
expressed in pixels, a display resolution expressed in dots-per-inch, and a viewing
distance expressed in inches. The spatial frequency of the kernel size is expressed
as a function of samples/pixels per degree of visual angle. In practice, the spread
of the Gaussian σi is multiplied by samples per degree to generate the convolution
kernel: ⎛ ⎞
size
⎜ dpi ⎟ 180
samples per degree = tan−1 ⎜ ⎟
⎝ distance ⎠ π . (11.38)

We can get a general idea of the type of spatial filtering performed by these
convolution kernels by transforming them into the frequency domain using a 1D
Fourier transform. If we do this for a large enough kernel, greater than 120 sam-
ples per degree, we can see the full extent of the filtering. This is shown in Fig-
ure 11.12. We can see that the achromatic filter behaves in a band-pass nature,
while the two chromatic filters are low-pass in nature. The achromatic filter essen-
tially blurs less than the red-green, which in turn blurs less than the blue-yellow.
This follows the generally known behavior of the contrast-sensitivity functions of

i i

i i
 i i

i i

610 11. Color and Image Appearance Models

the human visual system. Thus the S-CIELAB pre-processing provides a rela-
tively simple way to take into account the spatial properties of the visual system,
while also relying on the strength of the CIELAB color difference equations.

11.4.2 The iCAM Image Appearance Framework

Research in image appearance modeling is essentially trying to tie in our knowl-
edge of spatial vision with our knowledge of color appearance, in order to create a
single model applicable to color appearance, image rendering, and image-quality
specifications and evaluations. One initial framework for such a model for still
images, referred to as iCAM, is outlined in this section [292]. In addition to
traditional CIE colorimetry and color appearance modeling, this framework was
built upon a wide variety of previous research in the perception of images. These
areas include uniform color spaces, the importance of local image contrast and
surround, algorithms for image-difference and image-quality measurement, and
color and spatial adaptation to natural scenes. The iCAM framework is based
upon earlier models of spatial and color vision applied to color appearance prob-
lems, as well as on models for high dynamic range (HDR) imaging [292].
As we learned earlier in this chapter, a color appearance model is necessary to
extend CIE colorimetry to fully describe the appearance of color stimuli across a
variety of viewing conditions. Likewise, an image appearance model is necessary
to describe the more spatially complex color stimuli. For simplicity, we are gener-
ically using the term image to describe any complex scene that has been discreetly
sampled or rendered. Image appearance models attempt to extend upon these tra-
ditional color appearance descriptors to also predict such perceptual attributes as
sharpness, graininess, and contrast. Since image appearance also generates the
standard attributes such as lightness, chroma, and hue, these models should inher-
ently be able to predict the overall perceived color difference between complex
image stimuli.
At the heart of all CIE methods for describing color match, difference, and
appearance lies a uniform color space. For example, S-CIELAB and traditional
CIE color difference equations rely on CIELAB as their color space. A uniform
color space also lies at the heart of an image appearance model. This space,
along with per-pixel chromatic adaptation and local contrast detection are the
essence of image appearance modeling. It should be emphasized that iCAM was
never designed to be a complete model of image appearance, but rather a generic
framework or foundation to aid in the research and development of more complete
models.
The general process for using the iCAM model for predicting the overall ap-
pearance of still images is shown in Figure 11.13. Ideally, the model takes colori-

i i

i i
 i i

i i

11.4. Image Appearance Modeling 611

Colorimetric Input Image

Low-passed image for local Low-passed luminance image

chromatic adaptation effects for local contrast effects

Independent Gain Control CIECAM02 FL Function

1.00
1.2
0.90

0.80 1

0.70
Relative Response

0.8
0.60

0.50
FL

0.6

0.40
0.4
0.30

0.20
0.2
0.10

0.00 0
390 440 490 540 590 640 690 0 200 400 600 800 1000 1200
Wavelength (nm) Adapting Luminance

CIECAM02 Chromatic
Adaptation CIECAM02 FL

3x3 + Exponent Into IPT

Image appearance map in IPT,

or Lightness, Chroma, Hue

Figure 11.13. An iconic flow chart of the iCAM framework.

i i

i i
 i i

i i

612 11. Color and Image Appearance Models

metric images as input for the stimulus and surround in absolute luminance units.
Relative colorimetry can be used, along with an absolute luminance scaling or ap-
proximation of absolute luminance levels. Images are typically specified in terms
of CIE XYZ tristimulus values, or another well-understood device-independent
RGB space such as sRGB.
The adapting stimulus used for the chromatic-adaptation transform is then cal-
culated by low-pass filtering of the CIE XYZ image stimulus itself. This stimulus
can also be tagged with absolute luminance information to calculate the degree of
chromatic adaptation, or the degree of adaptation can be explicitly defined. A sec-
ond low-pass filter is performed on the absolute luminance channel, CIE Y, of the
image and is used to control various luminance-dependent aspects of the model.
This includes the Hunt and Stevens effects, as well as simultaneous contrast ef-
fects caused by the local background. This blurred luminance image representing
the local neighborhood can be identical to the low-pass filtered image used for
chromatic adaptation, but in practice it is an image that is less blurred.
In an ideal implementation, another Gaussian filtered luminance image of sig-
nificantly greater spatial extent representing the entire surround is used to control
the prediction of global image contrast. In practice, this image is generally taken
to be a single number indicating the overall luminance of the surrounding viewing
conditions, similarly to CIECAM02. In essence, this can be considered a global
contrast exponent that follows the well-established behavior of CIECAM02 for
predicting the Bartleson and Breneman equations.
The specific low-pass filters used for the adapting images depend on viewing
distance and application. A typical example might be to specify a chromatic-
adaptation image as a Gaussian blur that spans 20 cycles per degree of visual
angle, while the local surround may be a Gaussian blur that spans 10 cycles per
degree [1267]. Controlling the extent of the spatial filtering is not yet fully un-
derstood, and it is an active area of research. For instance, recent research in
high dynamic range rendering has shown that the low-pass filtering for the local
chromatic adaptation and contrast adaptation may be better served with an edge-
preserving low-pass function, such as the bilateral filter as described by Durand
and Dorsey, and also later in Section 17.5 [270, 625, 626]. Research in HDR ren-
dering is one example of application-dependence in image appearance modeling.
A strong local chromatic adaptation might be generally appropriate for predicting
actual perceived image-differences or image-quality measurements, but inappro-
priate for image-rendering situations where generating a pleasing image is the
desired outcome.
A step-by-step process of calculating image appearance attributes using the
iCAM framework as described above follows. The first stage of processing in

i i

i i
 i i

i i

11.4. Image Appearance Modeling 613

iCAM is calculating chromatic adaptation, as is the first stage for most gen-
eral color appearance models. The chromatic-adaptation transform embedded in
CIECAM02 has been adopted in iCAM, since it has been found to have excellent
performance with all available visual data. This transform is a relatively sim-
ple, and easily invertible, linear chromatic-adaptation model amenable to image-
processing applications.
The general CIECAM02 chromatic-adaptation model is a linear von Kries
normalization of sharpened RGB image signals from a sharpened RGB adaptation
white. In traditional chromatic-adaptation and color appearance modeling that we
have discussed up to this point, the adapting white is taken to be a single set of
tristimulus values, often assumed to be the brightest signal in the image or the
XYZ tristimulus values of the scene measured off a perfect reflecting diffuser.
It is at this stage that image appearance deviates from traditional color ap-
pearance models. Instead of a single set of tristimulus values, the adapting signal
is taken to be the low-pass filtered adaptation image at each pixel location. This
means that the chromatic adaptation is actually a function of the image itself.
These adapting white signals in the blurred image can also be modulated by the
global white of the scene, if that is known.
The actual calculations of the adapted signals are identical to those in
CIECAM02, but will be presented again here for clarity. The von Kries nor-
malization is modulated with a degree-of-adaptation factor D that can vary from
0.0 for no adaptation to 1.0 for complete chromatic adaptation. The calculation of
D is also identical to the CIECAM02 formulation and can be used in iCAM as a
function of adapting luminance LA for various viewing conditions. Alternatively,
as in CIECAM02, the D factor can be established explicitly.
The chromatic-adaptation model is used to compute pixel-wise corresponding
colors for CIE Illuminant D65 that are then used in the later stages of the iCAM
model. It should be noted that, while the chromatic-adaptation transformation is
identical to that in CIECAM02, the iCAM model is already significantly different
since it uses the blurred image data itself to spatially-modulate the adaptation
white point. Essentially, any given pixel is adapted to the identical pixel location
in the blurred image.

⎡ ⎤ ⎡ ⎤
R X
⎣G⎦ = MCAT02 ⎣Y ⎦ ; (11.39a)
B Z
    
1 −LA − 42
D = F 1− exp ; (11.39b)
3.6 92

i i

i i
 i i

i i

614 11. Color and Image Appearance Models

   
RD65
Rc (x, y) = D + (1 − D) R (x, y) ; (11.39c)
RW (x, y)
   
GD65
Gc (x, y) = D + (1 − D) G (x, y) ; (11.39d)
GW (x, y)
   
BD65
Bc (x, y) = D + (1 − D) B (x, y) . (11.39e)
BW (x, y)

The use of the blurred XYZ image as a spatially modulated adapting white
point implies that the content of an image itself, as well as the color of the overall
illumination, controls our state of chromatic adaptation. In this manner, iCAM
behaves similar, with regards to color constancy, to the spatial modulations of the
Retinex approach to color vision. This behavior can result in a decrease in overall
colorfulness or chroma for large uniform areas, such as a blue sky. While this may
be the correct prediction for the overall image appearance, it may produce unde-
sirable results when using this type of model for image-rendering applications.
Another example of the localized spatial behavior inherent in an image ap-
pearance model is the modulation of local and global contrast using the absolute
luminance image and surround luminance image or value. This modulation is
accomplished using the luminance level control function FL from CIECAM02.
This function is essentially a compressive function that slowly varies with abso-
lute luminance; it has been shown to predict a variety of luminance-dependent
appearance effects in CIECAM02 and earlier models [302]. As such, it was also
adopted for the first incarnation of iCAM. However, the global manner in which
the FL factor is used in CIECAM02 and the spatially localized manner used in
iCAM is quite different; more research is necessary to establish its overall appli-
cability for image appearance.
The adapted RGB signals, which have been converted to corresponding col-
ors for CIE Illuminant D65, are transformed into LMS cone responses. These
LMS cone signals are then compressed using a simple nonlinear power function
that is modulated by the per-pixel FL signal. Then, the cone signals are trans-
formed into an opponent color space, approximating the achromatic, red-green,
and yellow-blue space analogous to higher-level encoding in the human visual
system. These opponent channels are necessary for constructing a uniform per-
ceptual color space and correlates of various appearance attributes. In choosing
this transformation for the iCAM framework, simplicity, accuracy, and applica-
bility to image processing were the main considerations.
The uniform color space chosen was the IPT space previously published by
Ebner and Fairchild and introduced in Section 8.7.4 [274]. The IPT space was

i i

i i
 i i

i i

11.4. Image Appearance Modeling 615

derived specifically for image-processing applications to have a relatively simple

formulation and to have a hue-angle component with good prediction of constant
perceived hue. Predicting lines of constant hue has traditionally been very im-
portant in gamut-mapping applications and will be increasingly important with
any gamut-expansion algorithms that are desired for new high dynamic range and
wide-gamut displays. The mathematical transformation into the IPT opponent
space is far simpler than the transformations used in CIECAM02:
⎡ ⎤ ⎡ ⎤⎡ ⎤
L 0.4002 0.7075 −0.0807 XD65
⎣M ⎦ = ⎣−0.2280 1.1500 0.0612⎦ ⎣YD65 ⎦ ; (11.40a)
S 0.0000 0.0000 0.9184 ZD65

 L (x, y)0.43FL (x,y) if L (x, y) ≥ 0,
L (x, y) = (11.40b)
− |L (x, y)|0.43FL (x,y) if L (x, y) < 0;,

 M (x, y)0.43FL (x,y) if M (x, y) ≥ 0,
M (x, y) = (11.40c)
− |M (x, y)|0.43FL (x,y) if M (x, y) < 0;

 S (x, y)0.43FL (x,y) if S (x, y) ≥ 0,
S (x, y) = (11.40d)
− |S (x, y)|0.43FL (x,y) if S (x, y) < 0;
⎡ ⎤ ⎡ ⎤⎡  ⎤
I 0.4000 0.4000 0.2000 L
⎣ P ⎦ = ⎣ 4.4550 −4.8510 0.3960⎦ ⎣M  ⎦ . (11.40e)
T 0.8056 0.3572 −1.1628 S

Although it seems counter-intuitive to have negative LMS cone responses, it

is possible to have negative LMS values because of the chromatic adaptation on
sharpened RGB values as well as the linear transformation from CIE XYZ values;
care must be taken when applying the nonlinear power function.
Once the IPT values are computed for the image data, a simple coordinate
transformation from the rectangular opponent space to cylindrical coordinates is
applied to obtain image-wise predictors of the traditional color appearance at-
tributes such as lightness J, chroma C, and hue angle h:

J = I, (11.41a)

C = P2 + T 2 , (11.41b)
 
P
h = tan−1 , (11.41c)
T

i i

i i
 i i

i i

616 11. Color and Image Appearance Models

Q = FL 1/4 J, (11.41d)
M = FL 1/4C. (11.41e)

Differences in these dimensions can be used to compute differences between

two images. In some instances, correlates of the absolute appearance attributes
of brightness Q and colorfulness M are required. These are obtained by scaling
the relative attributes of lightness and chroma with the appropriate function of FL ,
again derived from the image-wise luminance map as shown in (11.41).

11.4.3 Image-Difference Calculations in iCAM

For image-difference and image-quality predictions within the iCAM framework,
it is also necessary to apply spatial filtering to the image data to eliminate any im-
age variations at spatial frequencies too high to be perceived. This is performed,
as in S-CIELAB, as a pre-processing stage. The spatial pre-processing serves to
eliminate information that is imperceptible to the visual system and to normalize
color differences at spatial frequencies that are visible.
Like the filtering performed in S-CIELAB, this computation is highly depen-
dent on viewing distance and is based on filters derived from human contrast-
sensitivity functions. Since the human contrast-sensitivity functions vary with
respect to luminance, band-pass for the low and high frequencies, and low-pass
for the chromatic channels, it is necessary to apply these filters in an opponent
color space.
The choice of opponent color space is crucial for this step as it is highly de-
sirable that the luminance and chromatic channels be mathematically orthogonal.
Problems such as chromatic fringing can arise when this is not the case. As such,
the iCAM framework performs spatial filtering in a specially designed orthogonal
color space called Y C1 C2 . More details on the development of this opponent
space can be found in [556, 1070]. The linear transform from CIE XYZ into
Y C1C2 is given by
⎡ ⎤ ⎡ ⎤⎡ ⎤
Y 0.0556 0.9981 −0.0254 X
⎣C1 ⎦ = ⎣ 0.9510 −0.9038 0.0000⎦ ⎣Y ⎦ . (11.42)
C2 0.0386 1.0822 −1.0276 Z D65

It is important to note that this space is designed to approximate an iso-

luminant and opponent space as best as possible with a linear and orthogonal
space, and requires CIE tristimulus values specified for CIE Illuminant D65.

i i

i i
 i i

i i

11.4. Image Appearance Modeling 617

1.2

Sensitivity
Achromatic
1.0 Red−green
Yellow−blue
0.8

0.6

0.4

0.2

0.0
0 10 20 30 40 50 60
Spatial Frequency (cycles−per−degree)

Figure 11.14. The spatial filters used to calculate image differences in the iCAM frame-
work.

The spatial filters are designed to be used in the frequency domain, rather than
as convolution kernels in the spatial domain. The general form for these filters is

CSFlum ( f ) = a · f c · e−b· f , (11.43a)

−b1 · f c1 −b2 · f c2
CSFchrom ( f ) = a1 · e + a2 · e . (11.43b)

The parameters, a, b, and c, in (11.43) are set to 0.63, 0.085, and 0.616,
respectively, for the luminance CSF for application on the Y  channel. In (11.43),
f is the two-dimensional spatial frequency defined in terms of cycles per degree
(cpd) of visual angle. To apply these functions as image-processing filters, f is
described as a two-dimensional map of spatial frequencies of identical size to
the image itself. For the red-green chromatic CSF, applied to the C1 dimension,
the parameters (a1 , b1 , c1 , a2 , b2 , c2 ) are set to (91.228, 0.0003, 2.803, 74.907,
0.0038, 2.601). For the blue-yellow chromatic CSF, applied to the C2 dimension,
the parameters are set to (5.623, 0.00001, 3.4066, 41.9363, 0.083, 1.3684). The
resulting spatial filters for a single dimension are shown in Figure 11.14.
The band-pass nature of the luminance contrast-sensitivity function, as well
as the low-pass nature of the two chromatic channels can be seen in Figure 11.14.
These filters should look similar to those used in S-CIELAB (see Figure 11.12),
though slightly smoother. The smoothness is a result of defining the filters in the
frequency domain rather than as discrete convolution kernels of limited extent.
Two other important features can be seen with regard to the luminance CSF: its
behavior at 0 cycles per degree (the DC component) and that the response goes
above 1.0.

i i

i i
 i i

i i

618 11. Color and Image Appearance Models

Care must be taken with the DC component when performing spatial filter-
ing in the frequency domain. The DC component contains the mean value of the
image for that particular channel. As with S-CIELAB, we would like the image-
difference metric to collapse down into a traditional color-difference metric for
solid patches. For this to happen, it is important that the mean value of any given
channel does not change. The luminance spatial filter described by (11.43) and
shown in Figure 11.14 goes to zero at the DC component. Therefore, it is neces-
sary to first subtract the mean value of the luminance channel, apply the spatial
filter, and then add the mean value back to the image.
The other important feature of the luminance CSF is that it goes above 1.0 for
a band of frequencies ranging roughly between 3 and 15 cycles per degree. This
is where the visual system is most sensitive to color differences, and, as such,
these regions are more heavily weighted. Care must be taken when applying
a frequency-domain filter that goes above 1.0, as this can often lead to severe
ringing. However, when the filter is sufficiently broad, this is often not a problem.
When the filter itself becomes very narrow, such as when applied to a large high-
resolution image, it may be necessary to normalize the luminance CSF such that
the maximum is at 1.0. Even then, care must be taken to minimize ringing artifacts
when using a narrow image-processing filter.
The images are filtered by first transforming into the frequency domain by
use of a Fast Fourier Transform, FFT, multiplying with the CSF filter, and then
inverting the FFT to return to the spatial domain:
FiltImagelum = FFT−1 {(FFT {Image − mean(Image)}) · CSFlum }
+ mean(Image), (11.44a)
−1
FiltImagechrom = FFT {(FFT {Image}) · CSFchrom } . (11.44b)

The contrast-sensitivity functions described in this framework serve to modu-

late spatial frequencies that are not perceptible and to enhance certain frequencies
that are most perceptible. Generally CSFs are measured using simple grating
stimuli with care taken to avoid spatial frequency adaptation. Spatial frequency
adaptation is very similar to chromatic adaptation and essentially decreases sen-
sitivity to certain frequencies based upon information present in the visual field.
This decrease in sensitivity to some frequencies can serve to enhance the sensi-
tivity to other frequencies, through a form of CSF normalization. An excellent
description of the mechanisms of spatial frequency adaptation can be found in
Blakemore and Campbell, and more recently in Webster [101, 1228, 1229].
It should be noted that a multi-scale, or multi-channel, spatial vision model is
not required to predict spatial frequency adaptation. Instead, all that is required is

i i

i i
 i i

i i

11.4. Image Appearance Modeling 619

that the CSF functions be allowed to change shape as a function of the adapting
field, though this is clearly an indicator of the existence of multi-scale mecha-
nisms in the visual system.
Since spatial frequency adaptation cannot be avoided in real-world viewing
conditions, several models of spatial frequency adaptation or contrast-gain con-
trol have been described for practical applications [229, 290, 310, 311, 884, 1220].
These models generally alter the nature or behavior of the CSF based upon either
assumptions on the viewing conditions, or based upon the information contained
in the images themselves. A simplified image-dependent mechanism for spatial
frequency adaptation is given by

CSF
CSFadapt = , (11.45a)
α · FFT(Image) + 1
1
α= . (11.45b)
D · Xsize ·Ysize

This model essentially normalizes the contrast-sensitivity function based upon the
amount of information present in the image itself. It can almost be thought of as
a von Kries-style frequency adaptation. More details on the development of this
spatial frequency adaptation model can be found in [294].
When using this type of model, the frequency representation of the image
itself is typically blurred to represent spatial frequency channels. This blurring
can be done by performing a Gaussian convolution on the frequency image, which
can be thought of as multiplying the spatial image by a Gaussian envelope. The
scaling function α converts the frequency representation into absolute units of
contrast at each spatial frequency.
The D factor is similar to the degree of chromatic-adaptation factor found in
CIECAM02 and is set to 1.0 for complete spatial frequency adaptation. Spatial
frequency adaptation is important when calculating image differences between
images that may have regular periodic patterns, such as stochastic half-tone pat-
tern jpeg-compressed images with an 8-pixel blocking pattern. The regular period
of these patterns reduces the visual sensitivity to the spatial frequency of the pat-
tern itself, thus making the pattern less visible.
One potential benefit of spatial frequency adaptation is the ability to predict
visual masking without the need for multi-scale approaches. If a masking fre-
quency is present in an image, the contrast-sensitivity function for that particular
frequency region will become less sensitive.

i i

i i
 i i

i i

620 11. Color and Image Appearance Models

11.4.4 Summary of Image Appearance

Just as color appearance models were created to extend CIE colorimetry for pre-
dicting changes in appearance across disparate viewing conditions, image appear-
ance models aim to further extend this ability across spatially complex images.
The goal of image appearance modeling is to augment the traditional color ap-
pearance attributes described in Section 11.1.1 with other perceptual attributes of
images such as sharpness, graininess, and contrast. S-CIELAB and iCAM were
both designed as simple extensions to traditional color appearance models. Nei-
ther of these models are yet ready to fully account for all of the unique perceptions
the human visual system has for complex stimuli. We can consider the field of
image appearance modeling to be in its infancy, and we hope for great strides
in the future. Applications for these types of models can include high dynamic
range image rendering, gamut mapping, digital watermarking, and image-quality
measurements. These models are not limited to static imagery and can easily be
extended into the temporal domain by adding spatio-temporal filtering and models
for the time-course of chromatic adaptation.

11.5 Applications of Color and Image

Appearance Models
Color appearance modeling is designed to predict the change in appearance of a
stimulus when the viewing conditions change. This makes color appearance mod-
els useful for various applications. CIECAM02 was designed with color manage-
ment applications in mind [195]. Desktop imaging allows us to view, synthesize,
and capture images from a variety of sources. These images can then be printed
and viewed in a wide array of conditions, or even posted online or emailed around
the globe. Color appearance models allow us to essentially represent colors in a
viewing-condition independent manner and allow us to calculate the changes nec-
essary to get accurate color matches in many different environments.
This section outlines one such practical application: calculating the colors
necessary to send to a printer viewed in a bright environment to match a monitor
viewed in a dark environment. This section also touches on a practical applica-
tion of a per-pixel chromatic-adaptation transform for white balancing a digital
photograph, as well as calculating image differences using the iCAM framework
described above.

i i

i i
 i i

i i

11.5. Applications of Color and Image Appearance Models 621

11.5.1 Application: Creating an Appearance Match

between a Monitor and a Printer
Perhaps the most common use for a color appearance model is for cross-media
color reproduction, e.g., color management. If we have a stimulus presented on a
computer monitor, how do we reproduce the same appearance when we send the
stimulus to a printer? In order to do that, we must take into account the viewing
conditions that the original was viewed in, as well as the viewing conditions of
the reproduction. For this example, we will start with an original image that is
viewed on a D65 sRGB display, with a maximum luminance of 100 cd/m2 . We
will assume that the screen is being viewed in a dark surround, and that the relative
luminance of the background Yb is 10 percent. The adapting luminance LA will
be taken to be 20 percent of the maximum luminance of the display. The original
image is shown in Figure 11.15.
To fully illustrate the changes in perception predicted by CIECAM02, we will
choose an output viewing condition at a very high luminance level of 1000 cd/m2 ,
with a 20 percent relative luminance of the background. Thus, the adapting lu-
minance LA is 200 cd/m2 and the background luminance Yb is 20. Since this is a
print viewed in a typical environment, we will choose an average surround with
a D50 white point. If we calculate the appearance correlates for the monitor en-
vironment using the forward CIECAM02 calculations, and then invert the model
using the print-viewing conditions, we achieve the image shown in Figure 11.16.
It should be noted that for this calculation we allowed CIECAM02 to
calculate the degree of adaptation factor D, based upon the adapting
luminance.
There should be a few things immediately evident. The first, and most obvi-
ous, change is that the entire image appears more yellow. This is because the D50
illumination of the print-viewing environment is significantly more yellow than
the D65 of the monitor. If we were to stand in a room that is illuminated with D50
simulators, especially at the light levels used in this example, we would quickly
adapt to this more yellow environment. As such, our sensitivity to yellow would
decrease.
In order to achieve an appearance match back to our monitor, it is necessary to
increase the yellow content of the image to account for this change in sensitivity.
Note that for illustrative purposes in this example we actually transformed the im-
age back into sRGB, which does have a native white of D65. In practice we send
the adapted XYZ tristimulus values to a printer, which then reproduces those de-
sired tristimulus values through a printer characterization, specifically measured
for a D50 viewing environment.

i i

i i
 i i

i i

622 11. Color and Image Appearance Models

Figure 11.15. Original image encoded in sRGB. (Photo courtesy of Ron Brinkmann.)

The next thing to notice is that both the contrast and the overall colorfulness
of the reproduced image are markedly decreased. This is because our original
was viewed at a very low luminance level in a darkened environment. When we
view the printed reproduction in an environment with a very high luminance level
and surround, the Hunt and Stevens effects, as well as the surround effects, will
increase the perceived contrast and colorfulness. It is necessary to decrease the
actual image contrast to account for these changes in the perceived contrast.

i i

i i
 i i

i i

11.5. Applications of Color and Image Appearance Models 623

Figure 11.16. Image processed in CIECAM02 from a D65 monitor viewed in a dark sur-
round to a D50 print viewed in an average surround. (Photo courtesy of Ron Brinkmann.)

This example is meant to illustrate the predictive power of using a color ap-
pearance model for cross-media color reproduction. In general, CIE colorimetric
matches will not be adequate for reproducing accurate colors across a wide va-
riety of viewing conditions. In these situations, color appearance matches are a
much more desirable outcome. CIECAM02 was designed specifically for color
management applications, such as the example described here, and is an excellent
choice for many color reproduction or rendering systems.

i i

i i
 i i

i i

624 11. Color and Image Appearance Models

Figure 11.17. An image of a white building illuminated by the setting sun.

11.5.2 Application: Per-Pixel Chromatic Adaptation

Based on Image Content
There are many occasions when we would like to automatically white balance
an image using a color appearance or chromatic-adaptation transform. This can
be very useful when there is a global color shift caused by the illumination, as
we saw in Chapter 10. Other times the color shift is more localized, and using a
global chromatic-adaptation transform can result in undesirable effects.
An example is shown in Figure 11.17 and Figure 11.18. The original white
building is illuminated with a very orange light at sunset, while the sky behind

Figure 11.18. Processing the image through the CIECAT02 global chromatic-adaptation
transform using the tristimulus values of the building itself as the adapting white.

i i

i i
 i i

i i

11.5. Applications of Color and Image Appearance Models 625

Figure 11.19. The low-pass filtered image used in a per-pixel chromatic-adaptation trans-
form.

the image is still very blue. Since we know that the building is white, we can
take the CIE XYZ tristimulus values averaged across a number of pixels in the
building and use those as our adapting white in a standard CIECAT02 chromatic-
adaptation transform. The results of this calculation are shown in Figure 11.18.
We can see immediately that the building does indeed look white, but at the same
time the sky looks far too blue. This is because the sky is also being adapted by
the orange illumination, when in reality an observer would adapt locally when
viewing this scene.
In a situation like this, we would like to use a local chromatic-adaptation trans-
form, similar to that used in the iCAM image appearance framework. This can be

Figure 11.20. The adapted image using a per-pixel chromatic-adaptation transform with
complete adaptation to the low-passed image.

i i

i i
 i i

i i

626 11. Color and Image Appearance Models

accomplished by using (11.39), along with the blurred adapting image shown in
Figure 11.19. The results of this calculation are shown in Figure 11.20.
We can see immediately that the resulting image shows some of the desired
localized adaptation results, as the building looks mostly white while the sky looks
less blue. Unfortunately, there are also severe haloing artifacts around the edges
of the image, as well as undesirable color de-saturation in areas such as the roof.
It would be better to combine some aspects of the global chromatic-adaptation
transform with some aspects of the per-pixel chromatic adaptation. This can be
accomplished by modifying the adaptation equations slightly to include a global
adaptation factor. The modified chromatic adaptation formulae are then
   
RD65
Rc (x, y) = D + (1 − D) R (x, y) , (11.46a)
RW (x, y) + Rg
   
GD65
Gc (x, y) = D + (1 − D) G (x, y) , (11.46b)
GW (x, y) + Gg

   
BD65
Bc (x, y) = D + (1 − D) B (x, y) . (11.46c)
BW (x, y) + Bg

In these modified adaptation values Rg , Gg , and Bg represent a single global

adaptation white point. We can now calculate an image that has been adapted
with a mix of local and global illumination. This is shown in Figures 11.21
and 11.22. For the calculations of Figure 11.21, we essentially assumed that the

Figure 11.21. The adapted image using a combination of the localized per-pixel
chromatic-adaptation transform as well as a global RGBg equal energy white.

i i

i i
 i i

i i

11.5. Applications of Color and Image Appearance Models 627

Figure 11.22. The adapted image using a combination of the localized per-pixel
chromatic-adaptation transform as well as a global RGBg derived from the tristimulus val-
ues of the building itself. In essence, a combination of Figure 11.18 and Figure 11.20.

image was locally adapted to the low-passed image shown in Figure 11.19 and
globally adapted to an equal energy illumination, where RGBg equaled 1.0. Fig-
ure 11.22 was calculated with the same localized adaptation, but with the global
white point set to the tristimulus values of the building itself, the same tristimulus
values used to calculate Figure 11.18.
We can see from these results that combining the per-pixel chromatic-
adaptation transform with the global adaptation based on (11.46) allows for a
wide variety of white balancing effects.

11.5.3 Application: Calculating Image Differences

The general purpose of S-CIELAB and one of the initial goals of the iCAM frame-
work is to predict overall image differences between two spatially complex stim-
uli. A classic example of why this is necessary is described in this section. An
original image is shown in Figure 11.23. Two reproductions are shown in Fig-
ure 11.24 and Figure 11.25.
One of the reproductions has a very noticeable red arch, while the other has
noise added to it. The noise was added in such a way that the mean difference
in CIE XYZ tristimulus values was exactly the same for the two reproductions.
The image with the arch is clearly different, while the image with the noise is
imperceptible. If we were to calculate a CIELAB color difference, using the CIE
ΔE94∗ color-difference equation, we would find that the noise image has an average

color difference of about 1.05 units from the original, while the red arch image has

i i

i i
 i i

i i

628 11. Color and Image Appearance Models

Figure 11.23. An original image used to study image-difference calculations. (Photo

courtesy of Ron Brinkmann.)

Figure 11.24. The same image as Figure 11.23 altered by changing the color of the arch.
(Photo courtesy of Ron Brinkmann.)

an average color difference of about 0.25 units. This means that the noise image,
though barely perceptible, has on average a four times larger color difference.
We can process these images through the iCAM image difference framework,
using 11.43–11.44, for a viewing distance corresponding to 80 cycles per degree.

i i

i i
 i i

i i

11.6. Further Reading 629

Figure 11.25. The same image as Figure 11.23 altered by adding noise such that the
colorimetric error is the same as Figure 11.24. (Photo courtesy of Ron Brinkmann.)

When we do that, the mean color difference in IPT space for the red arch image
becomes approximately 0.75, while the mean color difference for the noise image
becomes 0.25. These units have been scaled by 100 on the I dimension, and 150
on the P and T dimensions, to match the units of CIELAB [274]. We see that by
applying the spatial filters as a pre-processing stage, the image difference calcu-
lations match up with what we would intuitively expect, while a color difference
calculation without the filters does not.

11.6 Further Reading

A full account of the history and development of color appearance models is given
by Mark Fairchild in Color Appearance Modeling [302]. Robert Hunt’s book,
The Reproduction of Colour, provides in-depth information on issues pertaining
to color reproduction.

i i

i i
 i i

i i

Part III
Digital Color
Imaging

i i

i i
 i i

i i

Chapter 12
Image Capture

Previous chapters presented the theory of light propagation and the perception
of color, as well as its colorimetric representation. The final part of this book
details how man-made devices capture and display color information in the form
of images.
When devices have sensors that individually respond to the same range of
wavelengths of the electromagnetic spectrum to which human sensors are re-
sponsive, they generate images that are similar to what humans perceive when
they view their environment. Such devices enable the preservation of scenes of
value in people’s lives, in the form of images and movies. Images and videos
are also used for art, for entertainment, and as information sources. This chapter
presents image-capture techniques for electromagnetic radiation in the visible part
of the spectrum. Although we focus predominantly on digital cameras and the ra-
diometry involved with image capture, we also briefly describe photographic film,
holography, and light field techniques. We begin with a description of the optical
imaging process which enables images to be projected onto sensors.
An image starts with a measurement of optical power incident upon the cam-
era’s sensor. The camera’s response is then correlated to the amount of incident
energy for each position across the two-dimensional surface of its sensor. This is
followed by varying levels of image processing, dependent on the type of camera
as well as its settings, before the final image is written to file. A typical diagram of
the optical path of a digital single-reflex camera (DSLR) is shown in Figure 12.1.
The main optical components of a DSLR camera are:

Lens. The lens focuses light onto an image plane, where it can be recorded.

Mirror. The mirror reflects light into the optical system for the view finder. This
ensures that the image seen through the view-finder is identical to the im-
633

i i

i i
 i i

i i

634 12. Image Capture

Sensor for
Penta exposure control
prism

Finder
eyepiece

Image plane
Finder
screen Infrared rejection
Optical low pass filter
Cover glass
Microlens array
Sub Color filter array
mirror Sensor
Package
Main
Objective lens mirror
Shutter
Autofocus

Figure 12.1. The components of a typical digital single lens reflex camera (after [1141]).

age that will be captured by the sensor. During image capture, the lens is
temporarily lifted to allow light to reach the sensor; in that position, it also
blocks stray light from entering the camera through the view-finder.

View-finder. Light reflected off the mirror is passed through a penta-prism to di-
rect it toward the finder eyepiece.

Shutter. The shutter determines the duration of the exposure.

Infrared rejection filter. Since humans are not sensitive to infrared light, but cam-
era sensors are, an infrared rejection filter is typically present to ensure that
the camera only responds to light that can be detected by the human visual
system.

Optical low-pass filter. This filter ensures that high frequencies in the scene are
removed before spatial sampling takes place, in cases where these frequen-
cies would cause aliasing.

Micro-lens array. Since for each picture element (pixel) the light-sensitive area is
smaller than the total area of the pixel, a micro-lens can help ensure that all
light incident upon the pixel surface is directed towards the light-sensitive
part. A micro-lens therefore improves the efficiency of the camera.

i i

i i
 i i

i i

12.1. Optical Image Formation 635

Color filter array. To enable a single sensor to capture a color image, the sensor
is fitted with a color filter array, so that each pixel is sensitive to its own
set of wavelengths. The pattern repeats itself over small clusters of pixels,
of for instance 2 × 2 or 4 × 4 pixels. The camera’s firmware will perform
interpolation (demosaicing) to reconstruct a color image. The nature of the
filter, as well as the signal processing, determine the primaries and the white
point of the camera.

Sensor. The sensor is the light sensitive device that is placed in the image plane.
It records the incident number of photons for each of its pixels and converts
it to charge. After an exposure is complete, the charge is converted, pixel
by pixel, to a voltage and is then quantized. The camera’s firmware then ap-
plies a variety of image processing algorithms, such as demosaicing, noise
reduction, and white balancing.

Some camera types, such as point-and-shoot cameras, have no mirrors, and in-
stead have either an electronic view-finder, or have separate optics for the view-
finder. Cameras for motion imaging largely consist of the same components, bar
the mirrors, and may rely on an electronic shutter to determine the length of light
measurement for each frame. Differences can also be found in the in-camera im-
age processing algorithms. Digital video cameras, for instance, may use noise
reduction techniques that work over a sequence of frames, rather than by neces-
sity having to rely on the information present in a single frame [1000]. Digital
cinema cameras do not include elaborate signal processing circuitry and rely on
off-camera post-processing to shape and optimize the signal.

12.1 Optical Image Formation

Images are formed by optical imaging systems that focus light onto a sensor. For
any real system, not all light available in a scene can be collected and imaged.
Since camera systems collect only a segment of any wavefront, some diffraction
will occur, and this limits the achievable quality. Thus, optical imaging systems
are considered diffraction-limited. For sensor elements that are large relative to
the wavelengths being imaged, diffraction becomes relatively unimportant, and
geometrical optics can be used to model such system behavior (see Section 2.9).
In practice, the behavior of such systems can therefore be modeled using geomet-
rical optics, without the need to resort to physical optics. The basis of geometrical
optics is formed by Fermat’s principle, Snell’s law, the eikonal equation, and the
ray equation [666]; each is described in Chapter 2.

i i

i i
 i i

i i

636 12. Image Capture

Without reference to any specific imaging system, consider a point light

source, emitting light equally in all directions. In terms of ray tracing (see Sec-
tion 2.10.1), rays emanating from this light source diverge. The point source itself
is called the focus of a bundle of rays. If, by placing an optical system near a point
light source, a ray bundle can be made to converge into an arbitrary point, then
this point is also called a focus point [447].
If a ray bundle emanating from a point P0 converges in its entirety upon a
second point P1 , then the optical system that caused this convergence is called
stigmatic (sharp) for these two points. The points P0 and P1 are conjugate points;
reversing their roles would allow a perfect image to be created at P0 for rays
emanated at P1 . If the bundle of rays converges upon a small area rather than a
point, a blur spot is created, and the image is no longer perfect.
Of course, the system may allow points nearby P0 to be stigmatically imaged
to points that are nearby P1 . In an ideal optical system, the region of points that
are stigmatically imaged is called object space, whereas the region of points into
which object space is stigmatically imaged is called image space. Note that both
spaces are three-dimensional.
If we assume that a given set of points in object space spans a curve, then
this curve will be mapped to another curve in image space. If both curves are
identical, then the image is called perfect. An optical imaging system that is stig-
matic as well as perfect is called an absolute instrument. Maxwell’s theorem for
absolute instruments as well as Carathéodory’s theorem applies to such instru-
ments [112]. The former states that the optical length of any curve in object space
equals the optical length of its image. The latter theorem says that the mapping
between object and image space of an absolute instrument is either a projective
transformation, an inversion, or a combination of both.
The constraints placed upon optical systems before they can be called absolute
instruments are rather limiting. In most real imaging systems, the image space is
a region on a plane or curved surface, named the image plane (although only a
small region of the plane is used).

12.1.1 Imaging Geometry

In the following, we make the assumption that the optical imaging system is con-
structed such that all rays only make a small angle φ with respect to a reference
axis. Such rays are called paraxial. Systems for which this assumption holds,
i.e., φ is small, enable sine and cosine terms to be simplified as sin(φ ) ≈ φ and
cos(φ ) ≈ 1. This simplification to geometric optics is called linear optics.
If, in addition, all optical elements involved in the system are arranged along
the reference axis, which is then called the optical axis, and all elements are ro-

i i

i i
 i i

i i

12.1. Optical Image Formation 637

tationally symmetric with respect to this central axis, then further simplifications
are possible. The study of such systems is called Gaussian, first-order, or paraxial
optics.
According to Maxwell’s and Carathéodory’s theorems, in Gaussian optics
imaging may be approximated as a projective transformation. An object point
P = (px py pz )T generally projects to P = (px py pz )T according to [666]

m11 px + m12 py + m13 pz + m14

px = , (12.1a)
m41 px + m42 py + m43 pz + m44
m21 px + m22 py + m23 pz + m24
py = , (12.1b)
m41 px + m42 py + m43 pz + m44
m31 px + m32 py + m33 pz + m34
pz = . (12.1c)
m41 px + m42 py + m43 pz + m44

Using homogeneous coordinates (see Section A.10) and applying symmetry ar-
guments, this projection can be represented with the following transformation:
⎡ ⎤ ⎡ ⎤⎡ ⎤
px f 0 0 0 px
⎢ py ⎥ ⎢ 0 f 0 0 ⎥ ⎢ py ⎥
⎢ ⎥=⎢ ⎥⎢ ⎥, (12.2)
⎣ pz ⎦ ⎣ 0 0 z0 f f  − z0 z0 ⎦ ⎣ pz ⎦
pw 0 0 1 −z0 1

where we have chosen two focal points z0 and z0 and two focal lengths f and f  .
The 3D position of the transformed point is found by applying the homogeneous
divide, yielding P = (px /pw py /pw pz /pw )T .
As we are still discussing a general transformation effected by an arbitrary
rotationally symmetric optical system, we will only assume that this system sits
somewhere between the object point P and the image point P and is centered
around the z-axis. This geometric configuration is shown in Figure 12.2. Note that
we use a right-handed coordinate system with the positive z-axis, i.e., the optical
axis, pointing toward the right. The positive y-axis points up. The x = 0-plane
is called the meridional plane, and rays lying in this plane are called meridional
rays. All other rays are called skew rays. Meridional rays passing through an
optical system will stay in the meridional plane.
Due to the fact that the optical system is assumed to be circularly symmet-
ric around the z-axis, we can drop the x-coordinate and focus on the y- and z-
coordinates. The former is given by

f py
py = , (12.3)
z − z0

i i

i i
 i i

i i

638 12. Image Capture

Object Image
Principal Principal
Plane Plane
y
P0 = (py, pz) f’ f f f’
py P1 = (p’y, p’z)
p’y

N F H N’ F’ H’
pz z0 z’0 p’z z

N = Object nodal point N’ = Image nodal point

F = Object focal point F’ = Image focal point
H = Object principal point H’ = Image principal point

Figure 12.2. Projective transformation from P = (py pz )T to P = (py pz )T . The trans-
formation is characterized by the object focal point at z0 , the image focal point at z0 , the
object focal length f , and the image focal length f  .

which is known as Newton’s equation. The z-coordinate is given by

f f
pz − z0 = , (12.4)
z − z0
yielding the formula for the perspective transformation of a pinhole camera. A
pinhole camera is an optical system consisting of a small hole in a surface that sep-
arates image space from object space. It is a camera model that is frequently used
in computer graphics. In addition, a real pinhole camera can be built
cheaply [956].
Associated with an optical imaging system that effects this transformation are
several important points located along the z-axis. These points are collectively
referred to as cardinal points [666]:

Object focal point. Located at F = (0 0 z0 )T , it is also known as the first focal

point or the front focal point. An object located at this point will be imaged
infinitely far away. This result is obtained by substituting z = z0 in (12.4).

Image focal point. Located at F = (0 0 z0 )T , it ia also known as the back focal
point or the second focal point. From (12.4), we have that for an image
located at P1 = F , that its conjugate object point is located at infinity.

Object principal point. Located at position H = (0 0 z0 + f )T , it is also called

the first principal point or the front principal point. The plane parallel to
the x-y axis through this point is called the object principal plane. An object

i i

i i
 i i

i i

12.1. Optical Image Formation 639

located in this plane will be imaged with a lateral magnification of 1, i.e.,

its size after mapping into the corresponding image principal plane will
remain unaltered. Note that, for now, we will assume that the principal
plane is, as the name suggests, a planar surface. In a real optical system,
this is not necessarily true, leading to optical errors which are discussed in
Section 12.3.

Image principal point. This point is located at position H = (0 0 z0 + f  )T .

Object nodal point. This point is located at position N = (0 0 z0 − f  )T . A ray

passing through N at an angle θ with the optical axis will pass through
the image nodal point N at the same angle. For such rays, the angular
magnification is 1.

Image nodal point This point is located at position N = (0 0 z0 − f )T .

In a physical system, the radii of the lens elements are limited in spatial extent.
This means that only a fraction of the amount of light emitted by a point source
can ever be imaged. Further, the smallest diameter through which light passes
is either determined by fixed lens elements, or by an adjustable diaphragm. The
limiting element in this respect is called the aperture stop. Correspondingly, the
element limiting the angular extent of the objects that can be imaged is called the
field stop. This element determines the field of view of the camera.
Further important parameters characterizing an optical imaging system are the
entrance pupil and exit pupil. The entrance pupil is the aperture stop as seen from
a point both on the optical axis and on the object. An example with a single lens
and an aperture stop is shown in Figure 12.3. This means that the size of the
entrance pupil is determined both by the size of the aperture stop as well as the
lenses that may be placed between the object and the aperture stop. The axial
object point forms the apex of a cone whose base is determined by the entrance
pupil. The entrance pupil therefore determines how much light emanated from
this point is imaged by the system.
The exit pupil is the aperture stop as seen from the image plane through any
lenses that may be located between the aperture stop and the image plane (see
Figure 12.3). The exit pupil forms the base of a second cone, with the axial point
in the image plane as the apex. The entrance and exit pupils are generally not
the same, given that the presence of lenses may magnify or minify the apparent
size of the aperture stop. The ratio of the entrance and exit pupil diameters is
called the pupil magnification or pupil factor, a parameter used in determining
the radiometry associated with the optical system.

i i

i i
 i i

i i

640 12. Image Capture

Aperture
stop
Chief
Ray
Optical
axis
Mxp Mep

Object
Plane Image
Exit Plane
pupil
Entrance
pupil

Figure 12.3. The entrance and exit pupils of an optical system with a single lens and an
aperture stop. Also shown is a chief ray, which passes through the center of the aperture
stop.

Finally, rays starting from any off-axis point on the object that passes
through the center of the aperture stop are called chief rays [447]. An exam-
ple is shown in Figure 12.3. The chief ray is aimed at the center of the entrance
pupil Mep and passes through the center of the exit pupil Mxp . The marginal ray,
on the other hand, starts from the on-axis object point and passes through the rim
of the aperture stop (or entrance pupil).

12.1.2 Imaging Radiometry

A camera consists of an optical system and a sensor. The sensor measures image
irradiance Ee as a result of scene radiance Le incident upon the optical system (see
Section 6.2). The relationship between these two quantities is very important in
imaging.
We assume that the object distance is large relative to the object focal length.
The image irradiance is then proportional to the scene radiance. Further, Ee is
proportional to the size of the entrance pupil, meaning that the larger the entrance
pupil, the more light the system collects. Finally, Ee is inversely proportional
to f 2 , the square of the object’s focal length. This comes from the fact that the
lateral magnification is proportional to the focal length. Hence, the longer the
focal length, the larger the area in the image plane over which light is distributed.

i i

i i
 i i

i i

12.1. Optical Image Formation 641

+y

+x
Ψ
Ψ
dΨ
v
s r dΨ Ψ
d +y
h
θ r
d dA
dA +x
θ h
Object s
plane
+z

Exit pupil Image

Entrance pupil
plane

Figure 12.4. A differential area dA is imaged onto the area dA on the image plane by a
system characterized by its entrance and exit pupils (after [666]).

Following Lee [666], we analyze a differential area dA, located off-axis in the
object plane, which projects to a corresponding differential area dA on the image
plane. In between these two areas is the optical system, which is characterized by
the entrance and exit pupils. The geometric configuration is shown in Figure 12.4.
The chief ray starting at dA makes an angle θ with the optical axis. The distance
between dA and the entrance pupil along the optical axis is indicated by s; this
patch is located a distance h from the optical axis.
Further, d is the radius of the entrance pupil, and the differential area dΨ on
the entrance pupil is located a distance r from the optical axis. We are interested in
integrating over the entrance pupil, which we do by summing over all differential
areas dΨ. The vector v from dA to dΨ is given by
⎡ ⎤
r cos(Ψ)
v = ⎣r sin(Ψ) − h⎦ . (12.5)
s

This vector makes an angle α with the optical axis, which can be computed from

s
cos(α ) = . (12.6)
v

i i

i i
 i i

i i

642 12. Image Capture

Assuming that the differential area dA is Lambertian (see Section 2.9.2), then the
flux incident upon differential area dA is given by [666]
s
 d  2π r dΨ dr
v s
dΦ0 = Le dA (12.7a)
r=0 Ψ=0 v 2 v
 d  2π
r s2 dΨ dr
= Le
2 dA (12.7b)
r=0 Ψ=0
r2 cos2 (Ψ) + (r sin(Ψ) − h)2 + s2
 d  
2π s2 + h2 + r2 rs2 dr
= Le dA
3/2 (12.7c)
r=0
(s2 + h2 + r2 )2 − 4h2 r2
⎛ ⎞
π ⎜ s2 + h2 − d 2 ⎟
= Le dA ⎝1 −
1/2 ⎠ . (12.7d)
2
(s2 + h2 + d 2 )2 − 4h2 d 2

As shown in Figure 12.4, all quantities used to compute the flux over the
entrance pupil have equivalent quantities defined for the exit pupil; these are
all indicated with a prime. Thus, the flux over the exit pupil can be computed
analogously:
⎛ ⎞
π  ⎜ s2 + h2 − d 2 ⎟
dΦ1 = L dA ⎝1 −
1/2 ⎠ . (12.8)
2 e
(s2 + h2 + d 2 )2 − 4h2 d 2

Under the assumption that the optical system does not suffer light losses, the
flux at the entrance and exit pupils will be the same. The irradiance at the exit
pupil is then given by
dΦ0
Ee = (12.9a)
dA ⎛ ⎞
π dA ⎜ s2 + h2 − d 2 ⎟
= Le ⎝1 −
1/2 ⎠ , (12.9b)
2 dA
(s2 + h2 + d 2 )2 − 4h2 d 2

which is equivalent to
⎛ ⎞
π  ⎜ s2 + h2 − d 2 ⎟
= L ⎝1 −
1/2 ⎠ . (12.9c)
2 e
(s2 + h2 + d 2 )2 − 4h2 d 2

i i

i i
 i i

i i

12.1. Optical Image Formation 643

If the index of refraction at the object plane is n, and the index of refraction at the
image plane is n , then it can be shown that the irradiance at the exit pupil is given
by
⎛ ⎞
  2 2 2 2
π n ⎜ s +h −d ⎟
Ee = Le ⎝1 −
1/2 ⎠ (12.10a)
2 n 2 2 2 2 2 2
(s + h + d ) − 4h d
  2
π n
= Le G, (12.10b)
2 n

where

s2 + h2 − d 2
G = 1−
1/2 . (12.10c)
(s2 + h2 + d 2 )2 − 4h2 d 2

Equation (12.10a) is called the image irradiance equation. While this equations
is general, the quantities involved are difficult to measure in real systems. How-
ever, this equation can be simplified for certain special cases. In particular on-axis
imaging, as well as off-axis imaging, whereby the object distance is much larger
than the entrance pupil diameter form simplified special cases. These are dis-
cussed next.

12.1.3 On-Axis Image Irradiance

When the object of interest is located on the optical axis, then h as well as h are
equal to 0. Equation (12.10a) then simplifies to
  2  
n d 2
Ee = π Le . (12.11)
n s2 + d 2
If we consider the cone spanned by the exit pupil as the base and the on-axis point
on the image plane as the apex, then the sine of the half-angle β of this cone is
given by
d 2
sin(β ) = √ . (12.12)
s2 + d 2
We can therefore rewrite (12.11) as follows:
π Le   2
Ee = 2
n sin (β ) . (12.13)
n
The quantity n sin (β ) is called the numerical aperture. Since Ee is proportional
to the numerical aperture, we see that the larger this numerical aperture, the lighter

i i

i i
 i i

i i

644 12. Image Capture

the image will be. Thus, it is a value that can be used to indicate the speed of the
optical system.
A related measure is the relative aperture F, which is given by

1
F= . (12.14)
2n sin (β )

The relative aperture is also known as the f-number. In the special case that the
image point is at infinity, we may assume that the distance between image plane
and exit pupil s equals f  , i.e., the image focal length. The half-angle β can then
be approximated by tan−1 (d  / f  ), so that the relative aperture becomes

1
F∞ ≈ (12.15a)
2n sin (tan−1 (d  / f  ))
1 f
≈ . (12.15b)
n 2d 

By using the pupil magnification m p = d/d  , this equation can be rewritten as

1 f
F∞ ≈ . (12.15c)
m p n 2d

If the object and image planes are both in air, then the indices of refraction
at both sites are approximately equal to 1. If the magnification factor is assumed
to be close to 1 as well, then the relative aperture for an object at infinity may be
approximated by
f
F∞ = , (12.16)
D
where D is the diameter of the entrance pupil. An alternative notation for the f -
number is f /N, where N is replaced by the ratio f /D. Thus, for a lens with a focal
length of 50 mm and an aperture of 8.9 mm, the f -number is written as f/5.6. The
image irradiance Ee can now be written as
 2
π D2 Le mp
Ee = . (12.17)
4 f

We note that the quantity π D2 /4 is the area of the entrance pupil.

12.1.4 Off-Axis Image Irradiance

For object points that are not on the optical axis, we can simplify (12.10a) if the
distance of the object to the entrance pupil is much larger than the radius of the

i i

i i
 i i

i i

12.1. Optical Image Formation 645

entrance pupil, i.e., s  d. The image irradiance then becomes

s2 d 2 dA
Ee ≈ π Le (12.18a)
(s2 + d 2 + h2 )
2 dA
s2 d 2 dA
≈ π Le . (12.18b)
(s2 + h2 )
2 dA

This equation can be rewritten by noting that the cosine of the off-axis angle θ
(see Figure 12.4) is given by
s
cos (θ ) = √ . (12.19)
s2 + h2
The image irradiance is therefore approximated with
 2
d dA
Ee ≈ π Le cos4 (θ ) . (12.20)
s dA

The ratio dA/dA is related to the lateral magnification m of the lens through

dA
m= , (12.21)
dA
so that we have  2
d
Ee ≈ π Le cos (θ )
4
m2 . (12.22)
s
The lateral magnification satisfies the following relation:
m f
= . (12.23)
m−1 s
The image irradiance can therefore be approximated with
 2
 d
Ee ≈ π Le cos (θ )
4
, (12.24)
(m − 1) f 
or, in terms of the f -number,
π Le 1
Ee ≈ cos4 (Θ) . (12.25)
4 F 2 n2 (m − 1)2 m2p

If we assume that the lateral magnification m is 2, the pupil magnification m p is

1, and the index of refraction at the image plane is n = 1, then we can write

π Le π Le d 2
Ee ≈ cos 4
(Θ) ≈ cos4 (Θ) . (12.26)
4 F2 4 f2

i i

i i
 i i

i i

646 12. Image Capture

Note that the fall-off related to Θ behaves as a cosine to the fourth power. This
fall-off is a consequence of the geometric projection. Even in systems without
vignetting, this fall-off can be quite noticeable, the edges of the image appearing
significantly darker than the center [666]. However, lens assemblies may be de-
signed to counteract these effects, so that in a typical camera system the fall-off
with eccentricity does not follow the cosine to the fourth power fall-off.

12.1.5 Vignetting

Consider a simple optical configuration as shown in Figure 12.5. Although no

aperture stop is present, the finite diameters of the lenses impose an effective
aperture. The cross-section of this aperture depends on which point on the object
plane is considered. The further off-axis this point is taken, the smaller the cross-
section. This means that less light reaches the image plane dependent on the
distance between the point and the optical axis. This fall-off is independent of
and in addition to the cos4 fall-off discussed in the preceding section. It is known
as vignetting [15].
As the amount of vignetting depends on the distance to the optical axis, we
introduce a spatial dependency on points in the object plane (x, y) and correspond-
ing points on the image plane (x , y ), as well as an attenuation factor V (x , y ) that
takes vignetting into consideration. The general image irradiance equation then

Effective
aperture
stop
Chief
Ray

Optical
axis

Object Image
Plane Cross-section of off-axis Plane
ray bundle is smaller than
on-axis ray bundle

Figure 12.5. In this optical system that consists of two lenses, the effective aperture of the
off-axis ray bundle is smaller than the effective aperture of the on-axis ray bundle.

i i

i i
 i i

i i

12.1. Optical Image Formation 647

becomes
  2
π    n
Ee (x , y ) = L (x , y ) T V (x , y ) G, (12.27)
2 e n
where G is defined in (12.10c). We have also introduced the transmittance T of
the lens assembly to account for the fact that lenses absorb some light.

12.1.6 Veiling Glare

Optical systems in practice suffer from several limitations and imperfections that
are so far not accounted for in the image irradiance equation.
The lens assembly, as well as the inside of the lens barrel and the blades of the
aperture, may scatter and reflect light to some extent. Theoretically, a point light
source should be imaged onto a single pixel of the sensor, but due to limitations
of the camera design, some light will be imaged onto all other pixels according
to some distribution. This distribution is known as the glare spread function (or
GSF).
The phenomenon itself is known as veiling glare or lens flare [1246], depen-
dent on the effect it has on the resulting image. In particular, strong light sources

Figure 12.6. Example of an image showing veiling glare; Bristol Balloon Fiesta, Ashton
Court Estate, August 2007.

i i

i i
 i i

i i

648 12. Image Capture

that may even be outside the field of view may result in a spatially uniform in-
crease in image irradiance and thereby in a corresponding reduction of contrast.
This effect is called veiling glare.
Specific undesirable patterns may be imaged as a result of reflections inside
the lens assembly. These patterns are collectively called flare sources. For in-
stance, light reflected off the blades of the diaphragm may produce an image of
the diaphragm, resulting in iris flare.
An example of an image suffering from flare is shown in Figure 12.6. Under
normal exposure settings, and with good quality hardware, flare should remain
invisible. In this particular example, the sun was directly visible, causing reflec-
tions inside the camera, which are seen here as bright patches that are not part of
the environment.
Accounting for flare sources with a spatially dependent function g(x , y ), the
image irradiance is therefore more generally modeled with
  2
π    n
Ee (x , y ) = L (x , y ) T V (x , y ) G + g(x , y ). (12.28)
2 e n

The more components a lens has, the more surfaces there are that could reflect
and refract light. Zoom lenses, which can focus light onto the image plane at a
variable range of magnifications [545, 680, 873], as shown in Figure 12.7, tend
to have more internal surfaces than other lens types, so that they are also more
liable to produce flare. Within a camera, an extra source of flare is the sensor
itself, which may reflect as much as 30% of all light incident upon it. This is due
to the different indices of refraction of the Si (nSi ≈ 3) and SiO2 (nSiO2 = 1.45)
layers in the sensor itself [812]. Substituting these values into (2.138a) yields a
reflectance of R = 0.35. Most of this reflected light will re-enter the lens, possibly
to be reflected and refracted again.
To minimize the occurrence of flare, lenses can be coated with an anti-
reflective coating, as discussed further in Section 12.6.1. Similarly, the anti-

Figure 12.7. From left to right: the same scene was captured with different magnification
settings; Long Ashton, Bristol, May 2007.

i i

i i
 i i

i i

12.2. Lenses 649

reflective coating can be sandwiched between the Si/SiO2 interface to reduce re-
flections off of the sensor. Photographers may further reduce the effect of lens
flare by employing a lens hood, so that light entering the optical system from
oblique angles is blocked.
Finally, a related phenomenon is known as glare. Glare is caused by stray
reflections entering the lens directly due to strongly reflective surfaces being im-
aged [335]. If these reflections are off dielectric materials, then they will be po-
larized and can therefore be subdued by using polarization filters.

12.2 Lenses
So far we have examined optical systems more or less as a black box and under
several assumptions, such as Gaussian optics, derived the image irradiance that is
the basis of optical image formation. The components of optical systems typically
consist of refractive elements as well as an aperture. In this section we discuss
lenses as well as some of their attributes.
If a lens consists of a single element, it is called a simple lens. If it is made
up of multiple components, it is called a compound lens. Simple lenses are char-
acterized by their index of refraction as well as the shape of their front and back
surfaces. They may also be coated to improve their optical properties. For now
we concentrate on their shape.
The front and back surfaces of a lens may be either planar, spherical, or curved
in some different way. In the latter case, such lens elements are called aspherics.
Lenses with spherical surfaces are important because they are relatively straight-
forward to produce and polish, while still allowing high accuracy and minimal
artifacts [447]. Indicating the radius of the surface on the side of the object plane
with d0 and the radius of the surface facing the image plane with d1 , then depen-
dent on the sign of the radius different lens types may be distinguished, as shown
in Figure 12.8.
Note that an infinite radius means that the surface is planar. The sign of the
radius indicates the direction of curvature in relation to the vector along the optical
axis that points from the object plane to the image plane. A positive radius is for
a spherical surface that has its center on the side of the object plane, whereas a
negative radius is for a surface with its center on the side of the image plane.
Lenses that are thicker at the center, such as the three convex lenses in Fig-
ure 12.8, are variously called convex, converging, or positive. They tend to direct
light towards the optical axis. If the opposite holds, i.e., they are thinner at the
center, they are called concave, diverging, or negative [447].

i i

i i
 i i

i i

650 12. Image Capture

Bi-convex Bi-concave Planar Planar Meniscus Meniscus

d0 > 0 d0 < 0 convex concave convex concave
d1 < 0 d1 > 0 d0 = ∞ d0 = ∞ d0 > 0 d0 > 0
d1 < 0 d1 > 0 d1 > 0 d1 > 0
d0 < d1 d0 > d1

Figure 12.8. Different single lens types.

12.2.1 A Single Spherical Surface

Consider a spherical surface with radius d = P2 − C as shown in Figure 12.9.

The axial point P0 is assumed to be on the object plane. A ray with length lo
intersects the spherical surface at point P2 , and is refracted toward the surface
normal n (under the assumption that the index of refraction of the lens n1 is larger
than the index of refraction n0 of the surrounding medium) so that it intersects the
optical axis at P1 .

P2
θi
lo li
h θr
ϕr

P0 so V si C P1

|| P0 - V || = so n1
n0
|| V - P1 || = si
|| P2 - P0 || = lo
Indices of
|| P1 - P2 || = li
refraction

Figure 12.9. Geometry associated with refraction through a spherical surface.

i i

i i
 i i

i i

12.2. Lenses 651

According to Fermat’s principle (see Section 2.9.1), the optical path length of
this ray is given by
n0 lo + n1 li . (12.29)
The path lengths of the two ray segments lo and li can be expressed in terms of
the radius of the sphere d, the object distance so , and the image distance si :

lo = d 2 + (so + d)2 − 2d (so + d) cos (φ ), (12.30a)

li = d 2 + (si − d)2 + 2d (si − d) cos (φ ). (12.30b)

Substituting this result into the optical path length (12.29) yields

n0 d 2 + (so + d)2 − 2d (so + d) cos (φ )

+ n1 d 2 + (si − d)2 + 2d (si − d) cos (φ ). (12.31)

Applying Fermat’s principle by using the angle φ as the position variable results
in [447]
n0 d (so + d) sin (φ ) n1 d (si − d) sin (φ )
− = 0, (12.32)
2lo 2li
which can be rewritten to read
 
n0 n1 1 n1 si n0 so
+ = − . (12.33)
lo li d li lo
Any ray that travels from P0 to P1 by means of a single refraction through
the spherical surface obeys this relationship. Of course, there exist angles φ (and
therefore intersection points P2 ) for which the point P1 is never reached.
If we apply the principle of Gaussian optics, i.e., cos(φ ) ≈ 1, and only con-
sider paraxial rays, then we can simplify (12.30) to approximate

lo ≈ so , (12.34a)
l i ≈ si . (12.34b)

Equation (12.33) then becomes

n0 n1 n1 − n0
+ = . (12.35)
so si d
In the special case that the image point is at infinity, i.e., si = ∞, then we find
the object focal length f from (12.35):
n0
f = so = . (12.36)
d (n1 − n0 )

i i

i i
 i i

i i

652 12. Image Capture

Similarly, the image focal length is obtained by setting so = ∞, giving

d n1
f  = si = . (12.37)
n1 − n0

12.2.2 Thin Lenses

In the preceding discussion we have ignored the fact that a lens has both a front
and a back surface. We have only analyzed the front surface and derived equations
relating the object and image path lengths to the indices of refraction and the
radius of the spherical surface.
We now consider the combined effect of both the front and back surface. As-
sume that the lens is being constructed of two spherical surfaces, with radii d0 and
d1 , as shown in Figure 12.10. The analysis of refraction at both the front and back
surfaces yields the following relation:
 
n0 n0 1 1 n1 t
+ = (n1 − n0 ) − + . (12.38)
s0 si d0 d1 (sn − t) sn
If the lens is thin, we may approximate its thickness as t ≈ 0, so that the second
term on the right-hand side vanishes:
 
n0 n0 1 1
+ = (n1 − n0 ) − . (12.39)
s0 si d0 d1
Further, if the surrounding medium is air, so that n0 ≈ 1, we arrive at the thin-lens
equation, which is also known as the lens-maker’s formula:
 
1 1 1 1
+ = (n1 − 1) − . (12.40)
so si d0 d1

P3
P2

d1
d0

P’ P0 C2 V1 t V2 C1 P1
sn so si

n1
|| P0 - V1 || = so n0
|| V2 - P1 || = si Indices of
|| P’ - V1 || = sn refraction
|| V1 - V2 || = t

Figure 12.10. Two facing spherical surfaces form a thin lens, which projects light ema-
nating from on-axis point P0 to point P1 .

i i

i i
 i i

i i

12.2. Lenses 653

In the limit that the object distance so becomes infinite, the image distance
becomes the image focal length si = fi . Conversely, a similar argument can be
made for the object focal length for points projected on an infinitely far-away
image plane: so = fo . But since we are dealing with a thin lens, we may set
fi = fo . Replacing either with a generic focal length f , we have
 
1 1 1
= (n1 − 1) − . (12.41)
f d0 d1

Note that all light rays passing through the focal point, located a distance f in
front of the lens, will travel through space in parallel after passing through the
lens. Such light is called collimated. Combining (12.40) with (12.41), we find the
well-known Gaussian lens formula:
1 1 1
= + . (12.42)
f so si

Related to this formula is the concept of lateral magnification mt , which is also

known as transverse magnification and is given by
si
mt = − . (12.43)
so
This measure denotes the ratio of the size of the image to the size of the object.
The sign indicates whether the image is inverted (upside-down) or not: a positive
lateral magnification means the image is right side up, whereas a negative lateral
magnification indicates an inverted image. The object and image distance so and
si are positive for real objects and images1 , so that all real images created by a
thin lens are inverted (m p < 0).

12.2.3 Thick Lenses

Most lenses do not have a thickness that can be safely ignored. In such cases, we
speak of thick lenses. For simplicity, we think of thick lenses as consisting of two
spherical surfaces, spaced a distance t apart. Such lenses behave essentially as
an optical system and are characterized by their six cardinal points, as described
in Section 12.1.1. If the effective focal length is measured with respect to the
principal planes, then we have [447]
 
1 1 1 (n1 − 1)t
= (n1 − 1) − + . (12.44)
f d0 d1 n1 d0 d1
1 If rays leaving an optical system are converging, then the resulting image is termed real, whereas

the image is called virtual if the rays leaving the optical system are divergent.

i i

i i
 i i

i i

654 12. Image Capture

The index of refraction of a

GRIN lens varies radially.

Figure 12.11. A cylindrical gradient index lens.

12.2.4 Gradient Index Lenses

For the lenses described so far, the design parameters available are the index of
refraction, achieved through the appropriate choice of dielectric material, and the
curvature of the front and back faces of simple lenses. The placement of multiple
simple lenses and other optical elements along the path taken by the light from
the object plane to the image plane gives further freedom in lens design.
However, it is also possible to construct a lens by fabricating materials that
have a spatially varying index of refraction. Such lenses are known as gradient
index lenses [734]. Fabrication typically proceeds by immersing a rod of homo-
geneous glass in a bath of molten salts. Over a period of time, ionic diffusion then
causes the index of refraction to change. A consequence is that the lens can be a
circular disk, with planar front and back surfaces, as shown in Figure 12.11. Such
lenses are known as GRIN rods.
The analysis of gradient index lenses is more complicated than for other
lenses, as light incident at oblique angles will follow a curved ray path. Ray trac-
ing techniques may be used to analyze the behavior of such lenses by evaluating
the ray equation (2.180) as rays traverse the lens (discussed in Section 2.9.1).

12.3 Aberrations
Much of the analysis of lens systems presented so far uses approximations that
are only accurate to a first order. A full lens design requires more complicated
analysis that is beyond the scope of this book [1061, 1062]. Lens design pro-
grams have taken much of the burden of analysis out of the hands of the lens
designer. In addition, ray tracing, discussed in the context of computer graphics
in Section 2.10.1, can now be effectively used to evaluate lens design.

i i

i i
 i i

i i

12.3. Aberrations 655

Even so, deviations from idealized conditions do occur in practice. They are
collectively known as aberrations and can be classified into two main groups:
chromatic aberrations and monochromatic aberrations.
Chromatic aberrations occur when the material used to manufacture the lens
has a wavelength-dependent index of refraction. Such materials show dispersion
of light (see Section 3.6.2), which is clearly unwanted in lens design. Monochro-
matic aberrations form a group of distortions that make the image either unclear
(spherical aberration, astigmatism, and coma), or warped (distortion, Petzval field
curvature). Both classes of aberrations are discussed next.
The assumption that ray paths make only small angles with the optical axis,
i.e., the limitation to paraxial rays, enabled sine and cosine functions to be approx-
imated as linear and constant, respectively. Thus, we used only the first terms of
their respective Taylor expansions:
∞
(−1)n
sin (x) = ∑ (2n + 1)! x2n+1 ; (12.45a)
n=0
∞
(−1)n 2n
cos (x) = ∑ x . (12.45b)
n=0 (2n)!

If we were to use the first two terms of these expansions, we would get the so-
called third-order theory, which is more accurate, but more complicated. The
aberrations discussed here, however, are those that result from a deviation from
first-order (Gaussian) optics.

12.3.1 Spherical Aberrations

For a spherical refracting surface, (12.35) gave the relation between its radius,
indices of refraction, and the image- and object-distances. It made the assumption
that the ray path has the same length as the path from object to image plane along
the optical axis. If the ray path is recalculated using third-order theory, then the
difference with (12.35) is a partial measure of the error introduced by assuming
Gaussian optics. If the second term of the Taylor expansion is retained, then this
equation becomes [447]
     
n 0 n1 n1 − n0 n0 1 1 2 n1 1 1 2
+ = +h 2
+ + − . (12.46)
so si d 2so s0 d 2si d si

With respect to (12.35), the extra term varies with h2 , where h is the distance from
the point on the lens where the ray of interest strikes to the optical axis. This term
causes light passing through the periphery of the lens to be focused nearer the

i i

i i
 i i

i i

656 12. Image Capture

Optical
axis
Paraxial
focus

{
Object Longitudinal Image
Plane spherical Plane
aberration

Circle of least

{
Image
confusion Plane

Paraxial
focus
{ Transversal
spherical
aberration
Caustic

Figure 12.12. Diagram of spherical aberration.

lens than light that passes through the lens nearer the optical axis. This process
is called spherical aberration and is illustrated in Figure 12.12. Here, the rays
intersect the optical axis over a given length, and this phenomenon is therefore
called longitudinal spherical aberration. These same rays intersect the image
plane over a region, and this is called transversal spherical aberration.
Finally, in the presence of spherical aberration, the rays form a convex hull
which is curved. This shape is called a caustic. The diameter of the projected
spot is smallest when the image plane is moved to the location associated with the
so-called circle of least confusion, which in Figure 12.12 is in front of the focus
point that would be computed if only paraxial rays were taken into account.

12.3.2 Comatic Aberrations

A second type of aberration stems from the fact that the principal planes discussed
in Section 12.1.1 are only well approximated with planar surfaces very close to the
optical axis. Further away from the optical axis, the principal planes are in effect
principal curved surfaces. Even a small distance away from the optical axis, these
surfaces curve away from the idealized planar shape. The effect is called coma
or comatic aberration and is shown in Figure 12.13. If the marginal rays focus
farther from the optical axis than the principal ray (the off-axis ray that passes

i i

i i
 i i

i i

12.3. Aberrations 657

Image
Plane

Chief
Ray Optical
axis

Negative
Object coma
Plane

Image
Plane

Chief
Ray Optical
axis

Object Positive
Plane coma

Figure 12.13. Diagram of a positive and a negative coma.

through the principal points), then the result is a positive coma. If the marginal
rays hit the image plane closer to the optical axis than the principal ray, the coma
is called negative.

Figure 12.14. An example of a positive coma created by aiming an off-axis collimated

beam of light at the center of a positive lens. The colored fringes suggest the presence of
chromatic aberrations as well.

i i

i i
 i i

i i

658 12. Image Capture

An example of a coma is shown in Figure 12.14. This example was created

by aiming a collimated beam of white light at a positive lens. The light was aimed
at the center of the lens from an angle toward the right of the optical axis. The
cross-section of the beam is circular, yet its projection on the image plane is not.
The result is a positive comatic aberration.

12.3.3 Astigmatism
A further aberration, called astigmatism, occurs for off-axis object points. Recall
that the meridional plane is defined by the object point and the optical axis. The
chief ray lies in this plane, but refracts at the lens boundaries. The sagittal plane
is orthogonal to the meridional plane and is made up of a set of planar segments,
each of which intersect parts of the chief ray.
We consider a ray bundle lying in the meridional plane, as well as a separate
ray bundle lying in the sagittal plane. In the absence of other forms of aberrations,
all the rays in the meridional plane will have the same optical path length and will

y Meridional rays Image

Plane

Chief
Ray Optical
axis (z)

Focal point for

Object meridional rays
Plane

x Saggital rays Image

Plane

Focal point for

Chief saggital rays
Optical
Ray axis (z)

Object
Plane

Figure 12.15. A flat bundle of rays located in the meridional plane has a different focal
point than a flat bundle of rays lying in the sagittal plane, leading to astigmatism.

i i

i i
 i i

i i

12.3. Aberrations 659

Meridional
plane

Circular
Meridionally
cross-
elongated
Saggital section
cross-section
plane
Sagitally
elongated
cross-section

Figure 12.16. The meridional focal point is elongated, as is the sagittal focal point, albeit
in orthogonal directions. In between these focal point lies the circle of least confusion,
which has the smallest diameter.

therefore converge at the same focal point. However, this optical path length can
be different from the optical path length of the rays lying in the sagittal plane.
In turn, this causes the focal point of sagittal rays to lie either before or after the
focal point associated with meridional rays, as shown in Figure 12.15.
At the sagittal focal point, the meridional rays will still not have converged,
leading to an elongated focal point. Similarly, the meridional focal point will be
elongated. For rays that are neither sagittal nor meridional, the focal point will be
in between the sagittal and meridional focal points. Somewhere midway between
the two focal points, the cross-section of all rays passing through the lens will be
circular (see Figure 12.16). In the presence of astigmatism, this circular cross-
section will be the sharpest achievable focal point; it is called the circle of least
confusion.
12.3.4 Petval Field Curvature
For spherical lenses, the object and image planes, which were hitherto considered
to be planar, are in reality curving inwards for positive lenses and outwards for
negative lenses. In fact, both “planes” are spherical and centered around the lens.
This is known as Petzval field curvature. The planar approximation is only valid
in the paraxial region. If a flat image plane is used, for instance because the
image sensor is flat, then the image will only be sharp near the optical axis, with
increasing blur evident in the periphery.
The occurrence of field curvature can be counteracted by using a combination
of positive and negative lenses. For instance, the inward curvature of the image
plane associated with a positive lens can be corrected by placing a negative lens
near the focal point of the positive lens. The negative lens is then known as a field
flattener.

i i

i i
 i i

i i

660 12. Image Capture

Object Pincusion Barrel

distortion distortion

Figure 12.17. The effect of pincushion and barrel distortions upon the appearance of a
regularly shaped object.

12.3.5 Distortions
Another source of monochromatic aberrations is called distortion and relates to
the lateral magnification of the lens (defined for thin lenses in Section 12.2.2).
While this measure was defined as a constant for the whole lens, it is only valid
under paraxial conditions, i.e., near the optical axis. In real spherical lenses, the
lateral magnification is a function of distance to the optical axis.
If the lateral magnification increases with distance to the optical axis, then a
positive distortion or pincushion distortion results. For lenses with a decreasing
lateral magnification, the result is a negative distortion or barrel distortion. Both
distortions affect the appearance of images. In the case of a pincushion distortion,
straight lines on the object appear to be pulled toward the edge of the image.

Figure 12.18. An example of a lens with barrel distortion.

i i

i i
 i i

i i

12.3. Aberrations 661

A barrel distortion causes straight lines to be pushed toward the center of the
image as shown in the diagram of Figure 12.17 and the photographic example of
Figure 12.18 [1235].
Single thin lenses typically show very little distortion, with distortion being
more apparent in thick lenses. Pincushion distortion is associated with positive
lenses, whereas barrel distortion occurs in negative lenses. The addition of an
aperture stop into the system creates distortion as well. The placement of the
aperture stop is important; in a system with a single lens and an aperture stop, the
least amount of distortion occurs when the aperture stop is placed at the position
of the principal plane. The chief ray is then essentially the same as the principal
ray.
Placing the aperture stop after a positive lens will create pincushion distortion,
whereas placing the aperture stop in front of the lens at some distance will create
barrel distortion. For a negative lens, the position of the aperture stop has the
opposite effect. The effects of distortion in a compound lens can be minimized by
using multiple single lenses which negate each other’s effects and by placing the
aperture half way between the two lenses.
Other types of geometric distortion are decentering distortion and thin prism
distortion; the former is caused by optical elements that are not perfectly centered
with respect to each other, and the latter is due to inadequate lens design, man-
ufacturing, and assembly and is characterized by a tilt of optical elements with
respect to each other.

12.3.6 Chromatic Aberrations

As discussed in Section 3.6.2, we may expect dielectric materials, and therefore
lenses, to have a wavelength-dependent index of refraction. The focal length
of lenses in turn is dependent on the index of refraction of the lens, as shown
in (12.41) for thin lenses and (12.44) for thick lenses. Assuming a collimated
beam of white light aimed at a positive lens along its optical axis, the focal point
for wavelengths in the blue range lies closer to the lens than for wavelengths near
the red range. The distance between these two focal points along the optical axis
is called the longitudinal or axial chromatic aberration.
As light is refracted by greater or lesser angles dependent on wavelength, rays
exiting the lens from the same off-axis position will hit the image plane at dif-
ferent positions. The wavelength-dependent lateral displacement of light creates
a region along the image plane known as the lateral chromatic aberration. The
colored fringes are shown in Figure 12.14. The distinction between longitudinal
and lateral chromatic aberration is analogous to the longitudinal and transversal
effects shown for spherical aberrations in Figure 12.12.

i i

i i
 i i

i i

662 12. Image Capture

One can correct for chromatic aberration by applying a pair of thin lenses hav-
ing different refractive indices, a configuration termed thin achromatic doublets.
The wavelength-dependent focal length f (λ ) of the pair of lenses is then given by
1 1 1 d
= + − , (12.47)
f (λ ) f1 (λ ) f2 (λ ) f1 (λ ) f2 (λ )
assuming that the lenses are a distance d apart, their respective focal lengths are
f1 (λ ) and f 2 (λ ), and their indices of refraction are n1 (λ ) and n2 (λ ). If the index
of refraction of the surrounding medium is taken to be 1, then the focal lengths of
the two lenses are given by
1
= k1 (n1 (λ ) − 1) , (12.48a)
f 1 (λ )
1
= k2 (n2 (λ ) − 1) . (12.48b)
f 2 (λ )
In these equations, we have replaced the factor that depends on the radius of the
front and back surfaces of the two lenses by the constants k1 and k2 , which do not
depend on wavelength. Substituting (12.48) into (12.47) gives
1 d
= k1 (n1 (λ ) − 1) + k2 (n2 (λ ) − 1) − .
f (λ ) 1 1
k1 (n1 (λ ) − 1) k2 (n2 (λ ) − 1)
(12.49)
For the focal lengths f (λR ) and f (λB ) to be identical, the two thin lenses
should be spaced a distance d apart. This distance can be found by solving (12.49)
for d, giving
1 k1 (n1 (λB ) − n1 (λR )) + k2 (n2 (λB ) − n2 (λR ))
d= . (12.50)
k1 k2 (n1 (λB ) − 1) (n2 (λB ) − 1) − (n1 (λR ) − 1) (n2 (λR ) − 1)
In the special case that the two lenses are touching, we have that d = 0, giving
k1 n2 (λB ) − n2 (λR )
=− . (12.51)
k2 n1 (λB ) − n1 (λR )
It is now possible to specify the focal length of the compound lens as that of
yellow light, which has a wavelength nearly half way between the wavelengths
for blue and red light. The focal lengths for the two thin lenses are given by
1
= k1 (n1 (λY ) − 1) , (12.52a)
f1 (λY )
1
= k2 (n2 (λY ) − 1) . (12.52b)
f2 (λY )

i i

i i
 i i

i i

12.3. Aberrations 663

Dividing and rearranging these two equations leads to

k1 n2 (λY ) − 1 f 2 (λY )
= . (12.53)
k2 n1 (λY ) − 1 f 1 (λY )

From Equations (12.51) and (12.53), we find that

f2 (λY ) (n2 (λB ) − n2 (λR )) / (n2 (λY ) − 1) w2

= = . (12.54)
f1 (λY ) (n1 (λB ) − n1 (λR )) / (n1 (λY ) − 1) w1

The quantities w1 and w2 are the dispersive powers associated with the indices of
refraction n1 and n2 . If we take the following wavelengths,

λF = 486.1 nm, (12.55a)

λD = 589.2 nm, (12.55b)
λC = 656.3 nm, (12.55c)

that coincide with the Fraunhofer F, D, and C spectral lines (see Table 3.8 on
page 169), then we can define the dispersive power of an optical material more
precisely. Rather than dispersive power, its reciprocal V = 1/w is often used,
which is variously known as the V-number, the Abbe number, the dispersive index,
or the constringence of the material. The Abbe number V of a material is thus
defined as
nD − 1
V= , (12.56)
nF − nC

where nD = n(λD ), nF = n(λF ), and nC = n(λC ). It is desirable in lens design to

use materials that have low dispersion, and therefore high Abbe numbers, as well
as high indices of refraction.

12.3.7 Blur Circle

An ideal optical system would image a point source onto a single point on the
image plane. However, the various types of aberrations discussed above lead to
a blurred shape on the image plane, which can be approximated as a circle. This
blur circle then has a radius b that can be determined as follows.
The configuration and distances (all assumed positive) are as indicated in Fig-
ure 12.19. Note that the object is placed a height h above the optical axis, which
coincides with the radius of the aperture d of the lens. The radius b of the blur

i i

i i
 i i

i i

664 12. Image Capture

Object Chief f
Ray Image
Plane

h=d

h1 h2
Object
Plane
b
s i

so

Figure 12.19. The geometric configuration assumed for computing the blur circle.

circle is then

b = h2 − h1 (12.57a)
 
1 1
= di − (12.57b)
s so
so − s
= di . (12.57c)
s so
The Gaussian lens formula (12.42) for this configuration is
1 1 1
= + , (12.58)
f s i
and therefore we have
sf
i= . (12.59)
s− f
Substitution into (12.57c) yields
s f d so − s pfd
b= = . (12.60)
s − f s so s− f
In this equation, the quantity p = (so − s)/s so can be seen as the percentage focus
error. We use the blur circle next in the calculation of depth of field.

12.3.8 Depth of Field

Only points in the object plane are maximally sharply focused on the image plane.
However, objects in a scene are more often than not located at various distances
before and after the object plane. While all these points are imaged less sharply,

i i

i i
 i i

i i

12.3. Aberrations 665

Figure 12.20. The limited depth of field causes a narrow range of depths to appear in
focus. Here, the background is out of focus as a result. Macro lenses, in particular, have
a very narrow depth of field, sometimes extending over a range of only a few millimeters;
Orlando, FL, May 2004.

there is a region of distances between the lens and object points that are all focused
reasonably sharply. This range of depths is called depth of field. As an example,
an image with a limited depth of field is shown in Figure 12.20.
What constitutes an acceptably sharp focus depends on the size of the image
plane, the resolution of the sensor, the size at which the image is reproduced, and
all of this in relation to the angular resolution of the human visual system. As
such, there is not a simple calculation as to range of values that are considered to
be in focus.
However, if for some system it is determined that a circle with a radius b on
the image plane leads to a reproduction that displays this circle as a single point.
We then assume that the camera is focused on an object at distance so and that the
blur circle is much smaller than the aperture, i.e., d  b. The distance from the
lens to the nearest point snear that has acceptable focus is then

so f
snear = . (12.61a)
b
f + (s0 − f )
d

i i

i i
 i i

i i

666 12. Image Capture

Similarly, the furthest distance sfar still in acceptable focus is

so f
sfar = . (12.61b)
b
f − (so − f )
d
With snear < so < sfar , the depth of field is now simply the distance between the
near and far planes that produce an acceptable focus, i.e., sfar − snear .
We see from (12.61b) that the far plane will become infinite if the denominator
goes to zero, which happens if
b
f= (so − f ) . (12.62)
d
Solving for so gives the hyperfocal distance:
 
d
so = f +1 . (12.63)
b
The corresponding near plane is given by
 
f d
snear = +1 . (12.64)
2 b
Thus, if the camera is focused upon the hyperfocal plane, all objects located be-
tween this value for snear and beyond will be in sharp focus.
Finally, we see that the size of the aperture d affects the depth of field. An
example of a pair of images photographed with different aperture sizes is shown
in Figure 12.21. In particular in macro photography the depth of field can become
very limited—as small as a few millimeters in depth.

Figure 12.21. Images photographed with a large aperture (f/3.2, left) and a small aperture
(f/16, right). The depth of field is reduced to less than a centimeter in the left image, due
to the use of a 60 mm macro lens. The figurine is a Loet Vanderveen bronze.

i i

i i
 i i

i i

12.4. The Diaphragm 667

While depth of field is a concept related to the object side of the lens, a similar
concept exists for the image side of the lens. Here, it is called depth of focus or
lens-to-film tolerance and measures the range of distances behind the lens where
the image plane would be in sharp focus. The depth of focus is typically measured
in millimeters, or fractions thereof.

12.4 The Diaphragm

The aperture stop discussed in the preceding sections determines how much light
reaches the image plane. Most camera designs feature a variable aperture stop,
called the diaphragm, giving the photographer control over how light or dark the
image will be. An example is shown in Figure 12.22.
In addition to regulating the amount of light incident upon the image plane,
the diaphragm also has an impact on the severity of different types of aberrations.
An optical system with a very small aperture stop operates largely in the paraxial
region, limiting monochromatic aberrations. Thus, although the total amount of
light collected at the image sensor depends on both the size of the diaphragm
and the exposure time, for every camera there exists an optimal aperture size that
minimizes the occurrence of artifacts, including vignetting. Whether practical
conditions allow that setting to be used, depends on how much light and how
much movement is available in the scene.
The maximum aperture diameter afforded by a lens, and therefore the min-
imum f -number, determines the flux reaching the image plane. The larger this
aperture, the shorter the exposure time required to collect the desired image irra-
diance. Thus, the maximum aperture diameter determines the lens speed.

Figure 12.22. A diaphragm with three size settings. This diaphragm does not allow
more than three settings and is suitable for budget cameras. More typical diaphragms are
constructed using a set of blades which together form a variable-sized opening.

i i

i i
 i i

i i

668 12. Image Capture

Figure 12.23. A partly opened two-blade shutter.

12.5 The Shutter

The shutter is a mechanism which can be opened and closed to allow light to
reach the sensor for a pre-determined time. It can be either a physical device, or it
can be implemented electronically. In the latter case, the firmware of the camera
simply instructs the sensor to start and stop recording at the appropriate times.
When the shutter is a mechanical device, there is choice of where to place it
in the optical path. In principle, the shutter can be placed anywhere: before the
lens, in between the lens elements, or in front of the sensor. When placed in front
of the sensor, it is called the focal plane shutter; when placed in between the lens
elements, it is called an inter-lens shutter. Both types of shutter have an effect on
the type of artifacts that may occur. As such, shutters may be designed to have a
beneficial effect on the quality of the imagery [906].
As an example, an ideal shutter opens in zero time, remains fully open for the
duration of the exposure, and then shuts again in zero time. This ideal is clearly
not achievable with a mechanical shutter. However, a shutter may be designed to
open over some amount of time and then close again. If this results in the center
of the aperture being exposed longer than its margin, the glare associated with
the aperture will be softer than with an ideal shutter. This can be visually more
pleasing.
Inter-lens shutters are usually center-opening, meaning that a set of blades
parts to let light through. An example of a two-blade shutter is shown in Fig-
ure 12.23. They are normally placed as near as possible to the diaphragm, as
this is where the aperture of the system is smallest. During opening and closing,
which may take a substantial fraction of the total exposure time, both the size and
the shape of the shutter opening may vary over time. Thus, the effective aperture
of the camera becomes a function of time.

i i

i i
 i i

i i

12.6. Filters and Coatings 669

Focal plane shutters are normally made of a pair of curtains that slide across
the sensor at some distance from each other, forming a moving slit. The closer
this type of shutter is mounted to the sensor, the less diffraction off the edge of
the shutter will have an effect on the image quality. Focal plane shutters are often
used in single-lens reflex (SLR) cameras.

12.6 Filters and Coatings

In addition to lenses to focus the light, other elements may be placed in the op-
tical path. These elements consist of coatings applied to the surface of the lens
elements, as well as filters. In particular, coatings are applied to reduce unwanted
reflections of the lenses, as well as the sensors which tend to reflect a significant
portion of the incident light. Filters may be added to the optical system to reduce
aliasing, and to split the light into color bands. The latter is called the color filter
array and requires subsequent signal processing to reconstruct a color image. We
therefore defer a discussion of this filter to Section 12.8.1.

12.6.1 Anti-Reflective Coatings

The surface reflectance of a lens is typically reduced when a thin layer of coating
is applied. Under the assumption that the index of refraction for the surrounding
medium is 1, and that the angle of incidence is perpendicular to the lens surface,
the reflectance of the lens is given by (2.139a). If we introduce a single layer of
coating with an index of refraction given by nc , the reflectance of the lens becomes
 2
n − n2c
R= , (12.65)
n + n2c

where n is the refractive index of the glass used in the fabrication of the lens. We
also assume that the thickness of the coating is a quarter of a wavelength. To
reduce the reflectance to 0, we would have to choose a coating with a refractive
√
index of nc = n. For glass with a refractive index of around 1.5, this means that
a coating with an index of refraction of around 1.22 should be sought. Unfortu-
nately, such coatings do not exist. The most suitable coatings have an index of
refraction of around 1.33 [613].
For a two-layer coating, the reflectance is given by
 2
n2c,1 n − n2c,2
R= , (12.66)
n2c,1 n + n2c,2

i i

i i
 i i

i i

670 12. Image Capture

where the coating applied to the lens has a refractive index of nc,2 and the second,
outer layer has an index of refraction of nc,1 . Once more, the coatings are assumed
to have a thickness of a quarter of a well-chosen wavelength. To guarantee no
reflections, this equation shows that we should seek a combination of coatings for
which the following relation holds:
nc,2
n= . (12.67)
nc,1
In practice, solutions of this form are feasible, and lenses employed in digital
cameras are routinely coated with multiple coats to suppress unwanted reflections.

12.6.2 Optical Anti-Aliasing

The process of demosaicing2 makes implicit assumptions on the scene being im-
aged. In particular, the spatial frequencies incident upon the sensor must be low
enough to enable the spatial averaging algorithms to produce sensible results. If
the spatial frequency reaches beyond the Nyquist limit, which is half the spatial
frequency of the sensor, the result is that these high frequencies are mapped to
frequencies below the Nyquist limit. This is called aliasing and results in visible
artifacts such as jaggies along edges and Moiré patterns in places where textures
have frequencies that are too high [106, 107].
A good camera design therefore includes counter-measures against spatial fre-
quencies that cannot be represented in the digital read-out of the sensor. One of
the ways to achieve this is to apply a low-pass filter. In cameras, this is usually
achieved with an optical filter, placed in front of the sensor. Such an anti-aliasing
filter can be constructed for instance by using polarization [390]. By placing two
birefringent elements in succession, for instance quartz, calcite (see Figure 2.26),
liquid crystal, or lithium niobate [613], the optical signal can be low-pass filtered
in horizontal and vertical directions before it is projected onto the sensor.
The displacements Δx and Δy achieved by the two birefringent filters can be
computed from the thickness Δd of the filter and the two indices of refraction ne
and no of the birefringent material as follows:

n2e − n2o
Δx = Δy = Δd . (12.68)
2ne no
The displacement Δy is achieved by the second filter, which is rotated 90◦ with
respect to the first filter.
Unfortunately, high-quality birefringent elements are expensive. An alterna-
tive is to base the anti-aliasing filter on phase delay. Remember that the speed
2 Note that the “c” is pronounced as “k”.

i i

i i
 i i

i i

12.6. Filters and Coatings 671

of light propagation depends on the type of dielectric material the light passes
through. By etching a shallow pattern into a dielectric element, some of the
light passes through a thicker layer of material, emerging with some phase de-
lay. By varying the depth of the etched pattern, the amount of phase delay can be
controlled. The different amounts of phase-delayed light will interfere, thereby
canceling high-frequency content more than low-frequency content. The type of
pattern edged into the dielectric element can be a set of discs which are randomly
placed and have randomly chosen diameters. The advantage of this approach is
cost of manufacture, whereas the disadvantage is that unless the filter is carefully
designed, this type of phase-delay filter can act as a diffraction grating, causing
significant artifacts [5].

12.6.3 Infrared Rejection

Filters and other elements may be placed in the optical path to alter attributes of
the light before it hits the sensor. One reason for altering the light is to remove the
wavelengths to which the human visual system is not sensitive but to which the
camera’s sensor is sensitive. Failing to remove these wavelengths will create in-
correct colors. For instance, some flowers as well as some types of black clothing
reflect substantially in the infrared [5]. The sensing of the infrared component is
liable to turn blue in the scene into purple in the image.
Most infrared rejection filters are absorption-type, i.e., infrared wavelengths
are absorbed and converted to heat. Some filters reflect the infrared component
using a dichroic element, or they use a combination of the two techniques. The

Figure 12.24. A CCD sensor with the infrared filter fitted (left) and removed (right).

i i

i i
 i i

i i

672 12. Image Capture

placement of the infrared filter is typically between the lens and the sensor. An
example of an infrared rejection filter is shown in Figure 12.24.

12.7 Solid-State Sensors

A sensor may be constructed by placing photosensitive elements in an array. Us-
ing the photo-electric effect (see Section 6.5.3), such elements work by generat-
ing electron-hole pairs as a result of incident photons and are generically called
charge detection nodes. There are currently two dominant technologies for build-
ing image sensors, both of which have their own advantages and disadvantages.
These are charge-coupled devices (CCDs) and complementary metal-oxide semi-
conductor (CMOS) sensors. As both classes of image sensor are being actively
developed and improved, their relative merits are shifting over time [95, 103, 330,
687, 688, 992, 1123, 1252].
The quality of images produced by cameras depends strongly on the design of
the sensor. Common characteristics by which sensors can be evaluated are [1123]:
Pixel count. The trend in both digital still and video photography is towards
higher total numbers of picture elements (pixels). This sensor characteristic
is desirable, as more pixels enable the camera to produce sharper images.
Higher pixel counts can be achieved by building larger sensors, or smaller
pixels on a sensor of given size. Of course, larger sensors require more
complicated and larger lenses, as off-axis behavior of the optical elements
becomes more important. Smaller pixels on the other hand means that the
surface area over which light can be detected becomes smaller, and there-
fore the responsivity of the sensor may be negatively affected. In turn this
would require longer exposure times.
A further problem with reducing the size of pixels is that sensor designs
begin to run into the problem of diffraction. Assuming that the lens is free
of any aberrations, and that the aperture is circular, then light is imaged
onto a small disc surrounded by concentric circles, a pattern known as the
Airy disc. This phenomenon is demonstrated for a slit in Figure 2.40 on
page 75. The intensity of the first ring is 1.75% that of the center disc and
is located at a radius r of [613]
r = 1.22 λ F, (12.69)

where λ is the wavelength, and F is the f -number of the optical system.

The radius r is a measure of the resolving power of the lens, as it indicates

i i

i i
 i i

i i

12.7. Solid-State Sensors 673

the minimum spacing between two points that can just be distinguished. It
is known as the Rayleigh limit. As the Rayleigh limit depends largely on
the f -number, this result indicates that for high resolutions, small apertures
should be avoided.

Angular response. Lenses focus light on the image plane, which is where the
sensor is placed. Ideally, all light arrives at the sensor at a right angle. In
practice, this can be difficult to realize, and light will be incident upon each
position of the sensor over a given solid angle. As a consequence, it is
desirable to design sensors such that their off-axis response is matched to
the lens system in front of the sensor.

Signal-to-noise ratio. Whenever a pixel is exposed to light, it collects charge in

response to incident photons. However, charge is also collected and lost
due to various other processes that degrade the signal. The noise perfor-
mance of a sensor is therefore an important characteristic, and is discussed
in more detail in Section 12.10. A further definition used to characterize
sensors is the noise-equivalent power, which is the radiant power that pro-
duces a signal-to-noise ratio of 1. The detectivity D is the reciprocal of the
noise-equivalent power, whereas the specific detectivity D∗ is the detectivity
normalized by the area of the pixel, i.e., D∗ = D/A, where A is the surface
area of the pixel.

Dynamic range. Noise sources place a low limit on the magnitude of the signal
that can be measured by a sensor. On the other hand, the ability to collect
charge by each pixel is also limited, causing pixels to saturate if more light
is incident than can be accommodated. The range of irradiance values that
can be usefully detected by a sensor is known as its dynamic range, defined
as the ratio between the optical power at which the sensor saturates and the
noise-equivalent power. It depends on many different factors, including the
size of the pixel. As pixel dimension reduces, the supply voltages must be
lower, and this places a limit on the dynamic range [688].

Responsivity/Sensitivity. The signal the sensor produces depends on the number

of incident photons as well as further processing that may occur. The
amount of signal generated per unit of image irradiance is called the re-
sponsivity of the sensor. It is determined by the quantum efficiency of each
pixel and its fill factor. The quantum efficiency is given by the number of
photons required to generate a single electron. The fill factor of a pixel is
the amount of surface area devoted to the photo-sensor relative to the total
surface area of the pixel. Related to responsivity is the sensor’s sensitivity

i i

i i
 i i

i i

674 12. Image Capture

which is the reciprocal of the input required to generate a response that is

just above the sensor’s noise level.

Linearity. The collection of charge in response to incident photons, and the sub-
sequent conversion to a digital signal should maintain a linear relationship
between the number of photons and the value of the resulting signal.3 For
most sensors, this relationship holds to a high degree of accuracy. How-
ever, the linearity of transistors may be compromised in sensor designs that
utilize ultra-thin electrical paths (0.35 μ m and below) [688].

Pixel uniformity. Due to variations in the fabrication process, a uniformly illumi-

nated sensor may produce a signal that is spatially variant. Such undesired
non-uniformities may sometimes be corrected for by the firmware. In par-
ticular, two forms of non-uniformity are distinguished—dark signal non-
uniformity (DSNU), which can be measured in the absence of illumination,
and photo response non-uniformity (PRNU), measured when the sensor is
illuminated. DSNU is discussed further in Section 12.10.2, while PRNU is
discussed in Section 12.11.4.

In the following, we briefly discuss the operating principles of CCD and

CMOS sensors.

12.7.1 CCD Devices

In a charge-coupled device (CCD), the photosensitive elements are photodiodes
or photogates that are arranged in an array. A photodiode can absorb photons,
attracting electrons which reduce the voltage across the diode proportionally to
the amount of incident power.
When the exposure starts, these photosites collect charge as the result of in-
cident photons. Each photosite has a maximum amount of charge that can be
collected, after which the photosite saturates. This is called the full-well capacity.
After the exposure has finished, the collected charges are moved towards the ana-
log to digital (A/D) converter. The image is thus read one pixel at a time, which
is a relatively slow process.
There are three types of CCD architecture, called full-frame, frame-transfer,
and inter-line sensors. In a full-frame sensor, the full surface area of the sensor is
devoted to collecting light. A full-frame CCD does not contain an electronic shut-
ter, so that a camera with a full-frame sensor is typically fitted with a mechanical
shutter, as discussed in Section 12.5.
3 For logarithmic sensors, the relationship between output signal and number of incident photons

should be as close as possible to a logarithmic function.

i i

i i
 i i

i i

12.7. Solid-State Sensors 675

A frame-transfer architecture is composed of a light-sensitive area for charge

collection and a separate storage area that is masked by an opaque material, such
as aluminium. After the exposure finishes, the collected charge is transferred to
the masked area before read-out commences. The transfer to the opaque storage
area effectively implements an electronic shutter, despite the slow read-out pro-
cess. A disadvantage of this approach is that the size of the light-sensitive area is
only half that of the sensor, so that compared to a full-frame design, an equivalent
sensor would be twice as large and, therefore, more expensive to produce. A fur-
ther disadvantage is that the process of transferring charge from one area of the
chip to another may cause some smear. The fill factor of the light-sensitive area
can be as high as that of a full-frame sensor.
An inter-line design redistributes the storage area of a frame-transfer sensor
into strips, so that columns of pixels are alternated with columns of storage areas.
This reduces the fill-factor of each pixel to about 0.5, but this can be improved
to 0.9 or higher by placing a micro-lens that which focuses light onto the sensor
in front of each pixel, and away from the aluminium strips. As each pixel has its
own adjacent storage area, the transfer of charge requires only one step, thereby
vastly reducing the occurrence of smear.
In addition to moving charges underneath an opaque strip, it is possible to
implement a fast electronic shutter by fitting each pixel with a shutter drain. This
is an extra line which can be biased to a high voltage, thus draining away photon-
generated electrons. Lowering the voltage on this shutter drain and increasing the
voltage on the imaging-array gate then opens the shutter [940].
The read-out procedure of a CCD sensor requires the collected charges to be
moved to a single on-chip site, where the charge is amplified and presented to the
external interface pin. The basic principle used to achieve this transfer is unique to
CCDs and is shown in Figure 12.25. Each pixel contains a potential well that can
collect and hold the charge as long as a bias voltage is applied. To move the charge
from one potential well to the next, first the bias voltage for the neighboring pixel
is raised, so that the pair of pixels form a single larger potential well. Then, the
bias voltage is removed from the first pixel, so that the charge concentrates in the
potential well associated with the second pixel.
This process of charge transfer can be used to move the charges step-by-step
to one or more amplifiers, typically consisting of a few transistors, which convert
the charge to a voltage. Analog-to-digital conversion takes place off-chip, and
several further steps of image processing may then follow in the digital domain.
These are typically implemented in a separate integrated circuit (IC).
CCD sensors typically require different voltages for different parts comprising
the chip. The clocking voltage in particular needs to be relatively high, leading

i i

i i
 i i

i i

676 12. Image Capture

Photo Vertical CCD

sensors output registers

Horizontal CCD Electron-toVoltage

output register conversion

Read-out procedure:
1 - Shift charge under light shield
2 - Shift columns into horizontal output register
3 - Shift accumulated row into convertor

Figure 12.25. The read-out procedure on an inter-line CCD sensor is that first all charges
are moved underneath the opaque strip. In parallel, all columns are shifted down by one
position, moving the bottom scan-line into the horizontal read-out register. This register is
then read step-by-step. This process is repeated until all scan-lines are read.

to both high system complexity and high power consumption [103], although the
trend in CCD design is towards lower voltages and less power consumption. Ad-
ditionally, it is possible to design CCD sensors with a single power supply volt-
age, which is used to internally derive other required voltages using an integrated
charge pump [154].
Finally, when the light incident upon a pixel is very bright, the potential well
may fill up and charge may leak away into neighboring pixels. When this happens,
the result is a streak of light called bloom. Moderns CCDs have anti-blooming
mechanisms at each charge detection node; these allow extraneous charge to leak
away in a controlled manner. Such gates typically reduce the fill factor by about
30%, reducing sensor sensitivity.
Bloom can also be avoided manually, by avoiding long exposures. To obtain
the desired image, several short exposures, each without bloom, may be stacked
in software. The noise characteristics of the resulting image are not affected by
this method.

12.7.2 CMOS Devices

Complementary metal-oxide semiconductor (CMOS) processing is used to fabri-
cate standard integrated circuits for computing and memory applications, but can

i i

i i
 i i

i i

12.7. Solid-State Sensors 677

be adapted for use as sensors. With CMOS technology, it is possible to integrate

the sensor array, the control logic, and potentially analog-to-digital conversion on
the same chip. CMOS sensors are mixed-signal circuits, containing both analog
circuitry, possibly A/D converters, and digital control logic. The main building
blocks are a pixel array, analog signal processors, row and column selectors, and
a timing and control block.
The pixel array itself can be constructed in many different ways. In contem-
porary CMOS image-sensor devices, the photosensor at each pixel is augmented
with additional circuitry, such as a buffer/amplifier, yielding an active pixel sen-
sor (APS). With such designs, high frame-rates are possible [341]. As each am-
plifier is only activated during readout, the power dissipation can be kept low.
Early active pixel designs suffered from fixed pattern noise [95], which is dis-
cussed in Section 12.10.2. Many variations of active pixel sensors are currently in
existence, based upon either photodiodes or photogates. Examples of technolo-
gies that have been demonstrated include thin film on ASIC (TFA) pixel designs,
which yield high sensitivity and dynamic range [707, 708], and complementary
active pixel sensors (CAPS) [1264]. There is an ongoing effort to develop sensors
with higher fill factors, lower power consumption, better noise characteristics, and
better sensitivity.
Analog-to-digital (A/D) conversion circuitry may be included for each pixel.
The design is then called a digital pixel sensor (DPS) and benefits high-speed
imaging applications, as some of the bottlenecks associated with A/D conversion
and readout are reduced. Digital pixel sensors also potentially eliminate some of
the problems that occur with the analog circuitry associated with APS systems,
including a reduction of column fixed-pattern noise and column readout-noise
(see Section 12.10.2). The drawback of the DPS approach is that the extra cir-
cuitry causes a reduction of the fill factor, although this problem reduces as the
technology scales down.
CMOS sensors sometimes include circuitry that allows each pixel to be read in
an arbitrary order. Electronic scan circuits can be added that determine the pixel
read order. As a result, when the exposure is finished, either the whole image can
be read, or specific regions may be selected. This can, for instance, be used to
implement region-of-interest (ROI), electronic zoom, pan, and tilt.
The sensor design may incorporate additional analog systems, for instance to
improve noise, increase the signal-to-noise ratio, or perform A/D conversion. If
an A/D converter is included, then it will be subject to several design considera-
tions [330]. The throughput should be high enough to yield a frame-rate that sup-
ports the intended application. The converter should not introduce non-linearities
or other artifacts and should dissipate only a small amount of power. If too much

i i

i i
 i i

i i

678 12. Image Capture

power is used, then the heat generated by the A/D converter may create temper-
ature variations across the chip’s surface, which in turn has a negative effect on
the sensor’s noise characteristics. Finally, the A/D converter should not introduce
noise into the system through cross-talk.
To improve the dynamic range of CMOS systems, a variety of techniques may
be incorporated. For instance, the sensor may be designed to produce a logarith-
mic response [177, 330], employ conditional resets [1273], switch between pixel
circuit configurations [850], change saturation characteristics [241], use negative
feedback resets [525], employ an overflow capacitor per pixel [22], or use mul-
tiply exposed pixels [2, 748, 809, 995, 1266, 1272]. While the dynamic range of
image sensors continues to improve, photographers can also create high dynamic
range images with current digital cameras by employing the multiple exposures
techniques discussed in Section 13.1.
A range of other image processing functionalities may be included on the
same chip, including motion detection, amplification, multi-resolution imaging,
video compression, dynamic range enhancement, discrete cosine transforms, and
intensity sorting [95, 330], making CMOS technology suitable for a wide variety
of applications.

12.8 In-Camera Signal Processing

The sensor produces a signal that will require further processing to form a pre-
sentable color image. First, a color image must be reconstructed. Further process-
ing may include white balancing and sharpening amongst other. In the following,
we discuss demosaicing, white balancing, and sensitivity settings in more detail.

12.8.1 Color Filter Arrays

Digital color cameras can be constructed by placing spatially varying colored
filters in front of the image sensor. As a result, the light will be filtered such
that each pixel will be sensitive to a particular distribution of wavelengths. It is
therefore possible to create a color image with a single sensor.
Each cluster of pixels on the sensor is illuminated with a pattern of colored
light, with many patterns being available to the sensor designer [5,880]. The most
often used pattern is the Bayer mask, which contains 2 × 2 clusters of pixels with
one red, two green, and one blue filter [76]. The two green pixels are placed
diagonally with respect to each other. The double weight for the green range of
wavelengths is chosen because the human visual system has highest visual acuity

i i

i i
 i i

i i

12.8. In-Camera Signal Processing 679

Figure 12.26. The Bayer color filter array.

at medium-wavelength light (see Section 6.1). An example of a Bayer color filter

array (CFA) is shown in Figure 12.26.
The output of the sensor can be considered to be a composite, monochromatic
image, with each pixel filtered by the CFA. To reconstruct a color image, it would
be possible to take each cluster of four pixels and recombine these into a single
colored output pixel. However, such a naı̈ve approach would exhibit poor spatial
resolution.
A better alternative is to reconstruct an image at the same resolution as the
sensor. Under the assumption that object features are projected onto large sets
of pixels, interpolation schemes may be employed to reconstruct the image. This
process is known as CFA interpolation or demosaicing [5].
There are several trade-offs that a camera designer faces when reconstructing
a color image [589, 590, 633, 1104]. First, the processing taking place by the
firmware must be as fast and uncomplicated as possible. Second, the processing
should introduce as few artifacts as possible.
For the demosaicing process, this typically means that the luminance values
are taken to be estimated by the values sensed for the green channel. To obtain
luminance values for the pixels that are not green, linear interpolation may be
used. As this creates a blurry image, a better strategy is to use non-linear adaptive
averaging [7], for instance by employing a simple edge detector and ensuring
that averaging does not take place across edges [512]. The kernel size of the
edge detector is an important trade-off between processing time and quality of
the results. The edge-detection and averaging process can be further refined by
making the assumption that red, green, and blue color values are correlated within
a local image region. This means that, for instance, blue and red are related as
follows:
B(x, y) = R(x + 1, y) + k, (12.70)
where k is a constant. This assumption is not unreasonable given that strong
correlations between red, green, and blue channels exist in natural images. Statis-
tical regularities in natural images are further discussed in Section 15.1. By using

i i

i i
 i i

i i

680 12. Image Capture

Figure 12.27. Kodak’s proposal for a new color filter array, including unfiltered panchro-
matic pixels to efficiently derive an estimated luminance signal.

this particular correlation, the edge-detection process will have more values in a
small neighborhood of pixels to consider, so that a more accurate result can be
obtained [5].
In a camera design, the luminance channel (often approximated by just us-
ing the green channel) may be augmented by two chrominance channels, which
are usually taken to be red minus green (CR ) and blue minus green (CB ). Note
that the green value was computed using the CFA interpolation scheme. This
very simple solution is intended to minimize firmware processing times. To com-
pute the missing chrominance values, linear interpolation is employed between
neighboring chrominance values (the nearest neighbors may be located along the
diagonals).
Luminance and chrominance values are then converted to RGB. This con-
version, as discussed in Section 8.1, is device-dependent as a result of the dyes
chosen in the manufacture of the CFA, as well as the firmware processing applied
to the image.
The process of demosaicing is intimately related to the choice of color filter
array. While the Bayer pattern is currently favored in most camera designs, it is
not optimal in the sense that the filters absorb much of the incident light, leading
to a reduced quantum efficiency of the camera system. In the summer of 2007,
Kodak announced an alternative color filter array which consists of four types of
filter, namely red, green, blue, and panchromatic. The latter is effectively an un-
filtered pixel. The advantage of this approach is that a luminance channel can be
derived from the CFA with a much higher efficiency. In addition, the pattern is
constructed such that when the panchromatic pixels are ignored, the result is a
Bayer pattern that allows conventional color interpolation algorithms to be em-
ployed. The Kodak CFA is shown in Figure 12.27.
The placement of a CFA in the optical path significantly reduces the amount
of light that reaches the sensor, requiring an increase in aperture size, exposure
time, or sensor sensitivity to compensate. An alternative to the use of CFAs is to

i i

i i
 i i

i i

12.8. In-Camera Signal Processing 681

use multiple sensors. A beam splitter is then placed in the optical path to split the
light beam into three directions so that light can be focused onto three sensors, one
for each of the red, green, and blue ranges of wavelengths. The result is a higher
quantum efficiency and better color separation. Due to cost factors associated
with the perfect registration of the three light beams, this approach is used only in
high-end equipment such as studio video cameras.

12.8.2 White Balancing

Further in-camera processing may be required, dependent on user preferences and
camera settings. This includes, for instance, white balancing, which can be either
automatic or user-specified [275]. In the case that the white balance factor is set
to a specific color temperature, the camera will multiply the pixel values by a 3 ×
3 matrix, often derived from a pre-calculated look-up table.
When the camera is set to “auto,” the camera will have to infer the illuminant
from the image it has just captured. While algorithms to achieve this were dis-
cussed in Section 9.9, in-camera processing must be very fast, and that condition
limits the range of algorithms that can be sensibly employed. A very simple and
fast approach is to make the gray-world assumption, which states that averaging
all the colors found in an environment yields gray. Specifically, the average image
color is typically assumed to be 18% gray. Thus, the color found by computing
this average can be used as the correction factor. In motion imaging cameras, the
average color can be computed over many frames, rather than for each frame indi-
vidually. There are, of course, many situations where the gray-world assumption
fails, and in those cases the white balancing will also fail. This can be remedied by
employing a more advanced algorithm, for example by excluding very saturated
pixels during the averaging [5].
A second approach is to assume that the pixel with the highest value represents
a white surface. The color of this pixel is then assumed to be that of the light
source. This heuristic fails if self-luminous surfaces are imaged. For instance,
colored light sources such as traffic lights would cause this approach to fail.
The third approach to automatic white balancing using in cameras is to ana-
lyze the color gamut of the captured image. Based on statistical assumptions on
the average surface reflectance and the emission spectra of typical light sources,
the color gamut of the captured image can be compared to a database of color
gamuts [503].
Practical in-camera white balancing algorithms employ a combination of
these three techniques. Many professional photographers prefer to set their cam-
eras to record “raw” (unprocessed) image data and to perform processing (includ-
ing demosaicing, color correction, white balance, and picture rendering) off-line

i i

i i
 i i

i i

682 12. Image Capture

in application software such as Adobe Lightrom, Photoshop, or Apple Aperture.

This opens up the possibility of using the advanced algorithms that were described
in Section 10.6. Note that off-line white balancing may suffer from gamut lim-
itations, a problem that can be largely overcome by using high dynamic range
imaging [952].

12.8.3 ISO Sensitivity

Digital sensors usually have a parameter called ISO sensitivity that allows the
analog gain to be adjusted prior to analog-to-digital conversion [534]. For high
ISO values, more amplification is applied so that less incident light is necessary
to generate output values that lie within the camera’s operating range. However,
noise in the signal is also amplified more, and the resulting image is liable to
appear grainy. Figure 12.28 shows the impact of changing the ISO values on the
final image.

Figure 12.28. Both images were captured with a Nikon D2H; the left image with ISO 200
(exposure 1/20 s at f/20) and the right image using the Hi-2 setting, which is equivalent
to ISO 6400 (1/1000 s at f/20). The increase in sensor sensitivity is paired with a marked
increase in noise; Long Ashton, Bristol, May 2007.

12.9 A Camera Model

Mathematical models of a camera are useful for understanding the sources of error
that may accumulate during image acquisition. In the following description of a
camera model, the analysis of noise sources, and calibration scheme, we follow
the work of Healey and Kondepudy [439]. It may be assumed that the number of
electrons collected at a photosite (x, y) is obtained by integrating over the surface
area of the pixel (u, v) and over all wavelengths λ to which the sensor is sensitive.
As each electron represents a unit of charge, the total charge Q collected at a

i i

i i
 i i

i i

12.10. Sensor Noise Characteristics 683

photosite is then
  x+u  y+v
Q(x, y) = Δt Ee (x, y, λ ) Sr (x, y) ck (λ ) dx dy d λ . (12.71)
λ x y

Here, Δt is the integration time in seconds, Ee is the spectral irradiance incident
upon the sensor (W/m2 ), Sr is the spatial response of the pixel, and ck (λ ) is the
ratio of charge collected to the light energy incident during the integration (C/J).
In an ideal system, there is no noise and there are no losses, so that the charge
collected at a photosite is converted to a voltage, which may be amplified with a
gain of g, leading to a voltage V :

V (x, y) = Q(x, y) g. (12.72)

The A/D converter then converts this voltage into a digital count n. Assuming
that there are b bits, n will be in the range

0 ≤ n ≤ 2b − 1. (12.73)

If the quantization step is given by s, this means that the voltage is rounded to
D = n s, relating n, s, and V as follows:

(n(x, y) − 0.5) s < V (x, y) ≤ (n(x, y) + 0.5) s. (12.74)

The digital image n(x, y) can then be further processed in the camera firmware.
Of course, in practice, sensors do not follow this ideal model, and, in partic-
ular, noise is introduced at every stage of processing. In the following sections,
we discuss the different sources of noise, followed by an outline of a procedure to
characterize a camera’s noise response.

12.10 Sensor Noise Characteristics

One of the greatest design challenges for any type of image sensor is the suppres-
sion of noise, which can be defined as any component of the output signal that is
unwanted and does not represent the image irradiance incident upon the sensor.
The first of two classes of noise is temporal noise, which can be induced by the
various components making up the sensor. This type of noise will be different
every time an image is captured. On the other hand, fixed pattern noise arises due
to irregularities during the fabrication of the sensor and will induce an unwanted
spatially varying signal that is the same every time an image is captured.

i i

i i
 i i

i i

684 12. Image Capture

12.10.1 Reset Noise

Before image sensing can begin, the potential wells have to be reset. The proce-
dure to accomplish this requires a certain amount of time, which is known as the
reset-time constant. In high-speed applications, clock speeds may be chosen such
that the reset-time interval, the time allocated to resetting the charge-detection
nodes, is longer than the reset-time constant. This means that not all potential
wells are fully empty when the capture of the image begins. The residual charge
forms a spatially varying signal known as reset noise [521].
This type of noise can be canceled using correlated double sampling (CDS)
circuitry. First, each pixel is reset. The remaining signal is then read, yielding a
bias frame containing a measure of the reset noise. After image capture, the signal
is read again. Subtraction of the two signals yields a reset-noise-free image.

12.10.2 Fixed Pattern Noise

There are several sources of noise in a camera that show up as inaccuracies in the
camera’s output. First, when the sensor is uniformly lit, the charge collected at
every pixel will vary due to fabrication errors. This type of variation is a form
of fixed-pattern noise known as dark signal non-uniformity (DSNU). It can be
corrected for by taking a photograph with the lens cap fixed. The resulting image
is a measure of fixed-pattern noise, and it can be subtracted from every image
taken. Some cameras have a built-in function to remove this type of noise, which
is achieved by taking a second exposure with the shutter closed and subtracting it
from the real exposure.
In our camera model, the fixed pattern noise shows up as a variation in sen-
sor response Sr , as well as a variation in the quantum efficiency ck (λ ). These
two sensor characteristics are each scaled by a constant which is fixed (but differ-
ent) for each pixel. Naming these constants k1 (x, y) and k2 (x, y), we assume that
they are properties of each pixel and can therefore be taken out of the integration
of (12.71). Taking k(x, y) = k1 (x, y) k2 (x, y), our camera model now produces a
charge Qn for charge collection site (x, y):

Qn (x, y) = Q(x, y) k(x, y), (12.75)

where the subscript n is introduced to indicate that this electron/charge count in-
cludes a noise factor. The fixed pattern noise k(x, y) is taken to have a mean of
1 and a variance of σk2 which depends on the quality of the camera design. This
model is valid in the absence of bloom; i.e., we assume that charge collection at
each photosite is independent from collection at neighboring sites.

i i

i i
 i i

i i

12.10. Sensor Noise Characteristics 685

12.10.3 Dark Current and Dark Current Shot Noise

In the absence of any light, electrons may still make their way into the potential
wells that collect charge, for instance due to thermal vibrations. The uniform
distribution of electrons over the sensor collected in the dark is known as the dark
current. Post-processing may correct for the average level of dark signal, but a
certain amount of noise will be superimposed on the dark signal. In addition,
there may be a non-uniformly distributed source of noise which accumulates in
the dark [203], known as dark current shot noise, which cannot be corrected for
by post-processing, but should be minimized through appropriate sensor design
and operation. In critical applications, dark current may be reduced by cooling
the sensor [546], which is especially important in low-light photography which
for instance occurs in astronomy.
Dark current is independent of the number of photoelectrons generated. The
charge associated with the number of dark electrons Qdc produced as a result of
dark current is therefore added to the signal, giving a total charge Qn :

Qn = Q(x, y) k(x, y) + Qdc (x, y). (12.76)

12.10.4 Photon Shot Noise

The arrival times of photons at the image sensor is random and follows Poisson
statistics. This means that if the mean number of photons increases, the variance
in the signal increases by the same amount. As such, above a given threshold,
photon shot noise will dominate the other types of noise present in the output
signal. However, as all other types of noise are independent of the input signal,
below this threshold other types of noise will be larger.
The impact of photon shot noise can be reduced by collecting a higher number
of photo-electrons, for example by widening the aperture, or increasing exposure
time—up to the limit imposed by the full-well capacity of the sensor.
Under the assumption that photons adhere to a Poisson distribution, and that
the number of photoelectrons collected at each charge collection site is a linear
function of the number of incident photons, the distribution of photoelectrons
also follows Poisson statistics. As such, the uncertainty in the number of col-
lected electrons can be modeled with a zero mean random variable Qs which has
a variance that depends on the number of photoelectrons, as well as the number
of dark electrons. The total number of electrons then becomes

Qn = Q(x, y) k(x, y) + Qdc (x, y) + Qs (x, y). (12.77)

i i

i i
 i i

i i

686 12. Image Capture

12.10.5 Transfer Noise

The process of transferring the accumulated charges to the output amplifier may
cause errors, leading to transfer noise [155,156]. However, the read-out efficiency
of modern CCD devices can be greater than 0.99999 [439]. For practical pur-
poses, transfer noise can therefore be neglected. The amplifier of a CCD device
generates additional noise with zero mean. This amplifier noise is independent
of the amount of charge collected and therefore determines the noise floor of the
camera.
The amplifier applies a gain g to the signal, but also introduces noise and
applies low-pass filtering to minimize the effects of aliasing. The amplifier noise
(also known as read noise) is indicated with Q− r , so that the output voltage of the
amplifier is given by

V (x, y) = (Q(x, y) k(x, y) + Qdc (x, y) + Qs (x, y) + Qr (x, y)) g. (12.78)

If compared with (12.72), we see that the fixed pattern noise is multiplied with
the desired signal, whereas the remaining noise sources are added to the signal.
The combined signal and noise is then amplified by a factor of g.

12.10.6 Quantization Noise

The voltage V produced by the amplifier is subsequently sampled and digitized
by the A/D converter. This leads to a further additive noise source Qq , which is
independent of V , giving

D(x, y) = (Q(x, y) k(x, y) + Qdc (x, y) + Qs (x, y) + Qr (x, y)) g + Qq (x, y). (12.79)

The quantization noise Qq is a zero-mean random variable which is uniformly

distributed over the range [−0.5q, 0.5q] and has a variance q2 /12 [857, 858].

12.10.7 Implications of Noise

Most of the noise sources discussed here are independent of the image irradiance,
with the exception of photon shot noise. This means that all noise sources become
increasingly important when photographing dark environments, requiring longer
exposure times. Figure 12.29 shows a plot on log-log axes where the number of
photon electrons increases linearly with image irradiance (the slope of the curve
is 1). This trend continues until the full-well capacity is reached. At this stage,
the photosite cannot convert more photon electrons, and the value of Q does not
rise any further. All noise sources, except photon shot noise, lie on a horizontal
line in this plot, indicating independence from the image irradiance.

i i

i i
 i i

i i

12.10. Sensor Noise Characteristics 687

log Q Saturation

Q - Signal
(Slope = 1.0)

Qs - Photon shot
noise (Slope = 0.5)
Other noise sources
(Slope = 0.0)

log Ee

Figure 12.29. Signal and noise as function of image irradiance. The axis are on a log-
scale, with, for this example, arbitrary units.

The standard deviation of photon shot noise, however, increases as the square
root of the signal level, i.e., in Figure 12.29 the amount of photon shot noise is
indicated with a straight line with a slope of 0.5. To limit the effect of photon shot
noise, it is therefore desirable to design sensors with a large full-well capacity,
since at the saturation point, the impact of photon shot noise is smallest. On the
other hand, the full-well capacity of a photosite depends directly on its surface
area.
With the exception of high-end digital cameras, where full-frame sensors are
used, the trend in sensor design is generally toward smaller sensors, as this en-
ables the use of smaller optical components which are cheaper. This is useful for
applications such as cell phones and web cams. In addition, the trend is toward
higher pixel counts per sensor. This means that the surface area of each photosite
is reduced, leading to a lower full-well capacity, and therefore saturation occurs
at lower image irradiances.
In addition, the dynamic range of a sensor is related to both the full-well
capacity and the noise floor. For the design of high dynamic range cameras, it
would be desirable to have a high full-well capacity, because, even in the case
that all other noise sources are designed out of the system as much as possible,
photon shot noise will limit the achievable dynamic range.

i i

i i
 i i

i i

688 12. Image Capture

12.11 Measuring Camera Noise

In order to measure the noise introduced by a digital camera, we can image a
variety of known test targets and compare the sensor output with known val-
ues [439]. For instance, images can be taken with the lens cap fitted. Second,
we can uniformly illuminate a Lambertian surface and take images of it. If sensor
characteristics are to be estimated, it will be beneficial to defocus such an image.
In practice, the variation in sensor output will be determined by camera noise, as
well as (small) variations in illumination of the surface, and (small) variations in
the reflectance properties of this surface.
We will assume that the test surface is facing the camera and that it is nearly
uniformly illuminated with a light source. The reflected radiance Le is then given
by
ρd
Le = L, (12.80)
π
where the light source has strength L and the Lambertian surface reflects a fraction
of ρd /π into the direction of the camera. This relation was obtained by applying
Equation (2.209) to (2.206) and carrying out the integration assuming that the
light source is a point light source positioned infinitely far away.
The lens in front of the sensor then focuses the light onto the sensor so that
the irradiance incident upon a pixel is given by plugging (12.80) into (12.10c):
  2
 ρd n
Ee = L G, (12.81)
2 n
which we will rewrite as

Ee = ks ρd L. (12.82)
To model the (small) variation of the illumination and reflectance as a function
of position on the image plane, the factors in this equation are parameterized as a
function of both wavelength λ and position (x, y) on the sensor:
Ee (x, y, λ ) = ks ρd (x, y, λ ) L(x, y, λ ). (12.83)
The average illumination on the test target is L̄(λ ) and the average reflectance
is given by ρ̄d (λ ). The illumination onto the test card surface that is ultimately
projected onto pixel (x, y) is then
L(x, y, λ ) = L̄(λ ) + Lr (x, y, λ ), (12.84)
where Lr is the deviation from the average L̄(λ ) for this pixel. The expected value
of L(x, y, λ ) is then
E (L(x, y, λ )) = L̄(λ ), (12.85)

i i

i i
 i i

i i

12.11. Measuring Camera Noise 689

and the expected value of the residual Lr (x, y, λ ) is zero:

E (Lr (x, y, λ )) = 0. (12.86)

Similarly, the reflectance of the test card at the position that is projected onto
sensor location (x, y) can be split into an average ρ̄d and a zero mean deviation
ρr (x, y, λ ):

ρd (x, y, λ ) = ρ̄d (λ ) + ρr (x, y, λ ), (12.87a)

where

E (ρd (x, y, λ )) = ρ̄d (λ ), (12.87b)

E (ρr (x, y, λ )) = 0. (12.87c)

It is reasonable to assume that there is no correlation between the illumination

on the test card surface and its reflectance. This means that the image irradiance
can be written as

Ee (x, y, λ ) = ks (L̄(λ )ρ̄d (λ ) + ε (x, y, λ )) , (12.88a)

with

ε (x, y, λ ) = ρr (x, y, λ ) L̄(λ )

+ ρ̄d (λ ) Lr (x, y, λ )
+ ρr (x, y, λ ) Lr (x, y, λ ), (12.88b)

The expected value of ε is then 0. The charge collected by the sensor is given
by (12.71) and can now be split into a constant component Qc , and a spatially
varying component Qv (x, y):

Q(x, y) = Qc + Qv (x, y), (12.89)

where
  x+u  y+v
Qc = ks Δt L̄(λ )ρ̄d (λ ) Sr (x, y) ck (λ ) dx dy d λ ; (12.90a)
λ x y
  x+u  y+v
Qv (x, y) = ks Δt ε (x, y, λ ) Sr (x, y) ck (λ ) dx dy d λ . (12.90b)
λ x y

As E(ε (x, y, λ )) = 0, we have that Qv (x, y) has a zero mean and a variance that
depends on the variance in L̄(λ ) and ρ̄d (λ ). During a calibration procedure, it

i i

i i
 i i

i i

690 12. Image Capture

is important to control the illumination of the test card to achieve illumination

that is as uniform as possible. Similarly, the test card should have as uniform a
reflectance as possible to maximally reduce the variance of Qv (x, y). Most of the
variance in the resulting signal will then be due to the sensor noise, rather than to
non-uniformities in the test set-up.

12.11.1 Noise Variance

The variance of the system modeled by (12.79) consists of several components.
We will first discuss these components, leading to an expression of the total vari-
ance of the sensor. We can then estimate its value by photographing a test card
multiple times [439].
The quantized values D(x, y) can be modeled as random variables as follows:

D(x, y) = μ (x, y) + N(x, y). (12.91)

The expected value E(D(x, y)) is given by μ (x, y):

μ (x, y) = Q(x, y) k(x, y) g + E (Qdc (x, y) g) . (12.92)

The zero-mean noise is modeled by N(x, y):

N(x, y) = Qs (x, y) g + Qr (x, y) g + Qq (x, y). (12.93)

The noise sources can be split into a component that does not depend on the
level of image irradiance and a component that does. In particular, the photon
shot noise, modeled as a Poisson process, increases with irradiance (see Sec-
tion 12.10.4):
Qs (x, y) g. (12.94)

Accounting for the gain factor g, the variance associated with this Poisson process
is given by
g2 (Q(x, y) k(x, y) + E(Qdc (x, y))) . (12.95)

The signal-independent noise sources, amplifier noise and quantization noise, are
given by
Qr (x, y) g + Qq (x, y) (12.96)

and have a combined variance of

q2
g2 σr2 + , (12.97)
12

i i

i i
 i i

i i

12.11. Measuring Camera Noise 691

where σr2 is the variance of the amplifier noise. The total variance σ 2 in the noise
introduced by the sensor is then the sum of these two variances:

q2
σ 2 = g2 (Q(x, y) k(x, y) + E(Qdc (x, y))) + g2 σr2 + . (12.98)
12
The expected value of the dark current for a given pixel can be replaced by the
sum of the average expected value over the whole sensor and the deviation from
this expected value for a given pixel. Introducing the notation QE(dc) (x, y) =
E(Qdc (x, y), we can write

QE(dc) (x, y) = Q̄E(dc) + QdE(dc) (x, y), (12.99)

where Q̄E(dc) is the average expected value of the dark current and QdE(dc) (x, y) is
the deviation from the expected value. Equation (12.98) can now be rewritten as
follows:
 
σ 2 = g2 Q(x, y) + Q̄E(dc) + g2 (k(x, y) − 1) Q(x, y)
  q2
+ g2 QdE(dc) (x, y) − Q̄E(dc) + g2 σr2 + . (12.100)
12
Under the assumptions that |k(x, y) − 1|  1 and |QdE(dc) (x, y) − Q̄E(dc) |  Q̄E(dc) ,
the variance of the sensor noise can be approximated with
  q2
σ 2 ≈ g2 Q(x, y) + Q̄E(dc) + g2 σr2 + . (12.101)
12

12.11.2 Estimating the Total Variance

By photographing a uniformly illuminated test card twice, we can expect to obtain
two quantized images, both with the same expected value μ (x, y) for each pixel,
but with a differing random noise N(x, y) [439]:

D1 (x, y) = μ (x, y) + N1 (x, y); (12.102a)

D2 (x, y) = μ (x, y) + N2 (x, y). (12.102b)

By subtracting the two images, we get a difference image with zero mean and a
variance of 2σ 2 . The expected value of either image D1 and D2 is given by

μ = Q(x, y) g + Q̄E(dc) g. (12.103)

If the spatial variance of Q(x, y) is sufficiently minimized, then we may substitute

the spatial mean Q̄:
μ = Q̄ g + Q̄E(dc) g. (12.104)

i i

i i
 i i

i i

692 12. Image Capture

Using (12.101), the variance in terms of μ is then

q2
σ 2 = gμ + g2 σr2 +
. (12.105)
12
Given two captured images D1 and D2 , an estimate of μ can be computed by
X Y X Y
1 1
μ̂ =
XY ∑ ∑ D1 (x, y) + XY ∑ ∑ D2 (x, y). (12.106)
x=0 y=0 x=0 y=0

An estimate of the variance σ 2 is then computed with

X Y
1
σ̂ 2 = ∑
XY − 1 x=0 ∑ (D1 (x, y) − D2(x, y) − μ̂d (x, y))2 , (12.107)
y=0

where μ̂d is the mean of the difference image.

The illumination used for the pair of images was so far left unspecified. In
order to compute estimates of the amplifier gain g and the variance of the signal-
independent noise sources σc2 = g2 σr2 + q2 /12, we can now vary the illuminance
on the test card and collect a pair of images for each illumination level. Us-
ing (12.105) the parameters ĝ and σ̂c2 can be found by minimizing the following
expression using a line fitting technique:
 2  2
σ̂i − ĝμ̂i + σ̂c2
∑ 
var σ̂i2
 , (12.108)
i

where the sum is over the set of image pairs which are indexed by i. The variance
in σ̂i can be estimated by
 
  2 σ̂i2 2
var σ̂i2 ≈ (12.109)
XY − 1

12.11.3 Estimating Dark Current

In the absence of light, a camera will still output a signal as a result of dark current.
By fitting the lens cap, the number of photo-generated electrons will be zero, so
that the quantized value output by the sensor is given by [439]
D(x, y) = (Qdc (x, y) + Qs (x, y) + Qr (x, y)) g + Qq (x, y). (12.110)
The mean value of this signal is QE(dc) (x, y) g, as all other noise sources have
zero mean, and its variance is σ 2 (x, y). This variance can be estimated with the
technique discussed in the preceding section. By averaging a set of n images
obtained in this manner, the variance will be reduced to σ 2 (x, y)/n. For large
enough values of n, the pixel values will then converge to an accurate estimate of
the dark current QE(dc) (x, y) g.

i i

i i
 i i

i i

12.11. Measuring Camera Noise 693

12.11.4 Estimating Fixed Pattern Noise

We now have estimates of the dark current, the amplifier gain, and the total vari-
ance in the noise. To estimate the fixed pattern noise, it is not sufficient to vary the
illumination level uniformly across the test card, as this would not enable us to
disentangle the spatial variation due to fixed pattern noise and the spatial variation
as a result of non-uniform illumination and/or reflectance [439].
However, by taking a sequence of images by varying the illumination non-
uniformly, as well as varying the orientation of the test card with respect to the
camera, we can average out the variation due to illumination and retain the fixed
pattern noise k(x, y), here called photo response non-uniformity (PRNU). Suppose
that we create n1 illumination conditions and capture n2 images for each of these
conditions. As the total noise N(x, y) has zero mean, averaging the frames for
imaging condition i will lead to the following estimate:

k(x, y) Qi (x, y) g + QE(dc) (x, y) g, (12.111)

with variance σi2 (x, y)/n2 . From Section 12.11.3 we already have an estimate of
the dark current, which we can subtract. This yields the corrected estimate

d(x, y) ≈ k(x, y) Qi (x, y) g. (12.112)

This estimate varies over the sensor surface as a result of fixed pattern noise as
well as variations in reflectance and illumination. However, for relatively small
pixel neighborhoods, the variation due to reflectance and illumination of the test
card may be assumed to be small. It is therefore possible to compute the mean
¯ y) over small windows for each pixel. Regions of 9 × 9 pixels tend to work
d(x,
well [439], leading to estimates of

¯ y) ≈ Qi (x, y) g.
d(x, (12.113)

The ratio between a single pixel estimate and the windowed estimate is then a
rough approximation of the fixed pattern noise ke (x, y):

d(x, y) k(x, y) Qi (x, y) g

¯ y) ≈ ≈ ke (x, y). (12.114)
d(x, Qi (x, y) g

To refine this approximation, the average ratio over the n1 imaging conditions is
computed:
1 n1 di (x, y)
ke (x, y) ≈ ∑ d¯i (x, y) .
n1 i=0
(12.115)

A further refinement can be obtained by removing outliers from this summation.

i i

i i
 i i

i i

694 12. Image Capture

12.12 Radiometric Camera Calibration

Images produced with digital cameras are represented with device-dependent
RGB pixel values. Although camera designers know the correspondence between
the RGB values produced by the camera and the image irradiances that gave rise
to these values, this information is not generally available to users of digital cam-
eras. As shown in the preceding sections, the optics, the CFA, the sensor as well
as the processing occurring in the firmware, all contribute to distortions in the
signal.
Camera characterization involves recovering either the camera response
curves fk or the sensor’s responsivities. In this section, we discuss methods to
retrieve these functions on the basis of calibrated input. In the context of high dy-
namic range imaging, uncalibrated input in the form of a set of multiply exposed
images may also be used to recover the camera response curve (as discussed in
Section 13.2).
For most applications, it is desirable to know the relation between pixel val-
ues and scene irradiances accurately, and for many applications this information
is vital [804]. For instance, accurate color reproduction is not possible without be-
ginning with accurate capture of irradiance values. Many color-based computer
vision algorithms lose accuracy without accurate capture.
To recover a linear relationship between the irradiance values and the pixel en-
coding produced by the camera, we need to model the non-linearities introduced
by in-camera processing. The process of recovering this relationship is known as
camera characterization or color characterization [1187].
There are two general approaches to color characterization; they are based on
spectral sensitivities and color targets [480]. Methods based on spectral sensitivi-
ties require specialized equipment such as a monochromator (a device that emits
light with a single user-specified wavelength) and a radiance meter [89]. With
such equipment, the response of the camera to each wavelength can be measured,
and from it the relationship between the spectral responsivities of the camera and
the CIE color matching functions can be estimated.
Target-based techniques make use of a set of differently colored samples that
can be measured with a spectrophotometer. A digital image is then captured of
the target. For each patch, we then have measured CIE XYZ values and captured
RGB values. The most precise calibration for any given camera requires record-
ing its output for all possible stimuli and comparing it with separately measured
values for the same stimuli [1201]. However, storage of such a quantity of data is
impractical, and, therefore, the response of the device is captured for only a lim-
ited set of stimuli—perhaps between a few dozen to a few hundred patches. The

i i

i i
 i i

i i

12.12. Radiometric Camera Calibration 695

responses to these representative stimuli can then be used to calibrate the device
for input stimuli that were not measured.
To find the transformation between RGB and XYZ values, several techniques
have been developed [480], including look-up tables which can be used to in-
terpolate or extrapolate new values [501, 502], least squares polynomial model-
ing [480, 574], neural networks [572], and fitting data to models of cameras and
image formation [65, 162, 439].

12.12.1 Look-Up Tables

If the stimuli densely sample the device gamut, then the transformation for device
output to the ideal response may be estimated by three-dimensional interpolation
of measured responses [501, 575]. The dense sampling of stimuli responses is
stored in look-up tables (LUTs). The advantage of this technique over others is
that it avoids estimation of the mapping with a function, a process that is complex
and computationally expensive.
Look-up tables require a few hundred measured responses. It may not always
be possible to densely sample the entire device gamut, and only a sparse sampling
of responses to stimuli may be available. In such cases, a function has to be
estimated from known responses, providing a mapping from the output of the
device being calibrated to the “ground truth.”
An example of a sparse set of stimuli is given in Figure 12.30, which shows
a chart of different colors that can be used as input to calibrate capturing devices.

Figure 12.30. The Gretag-Macbeth ColorChecker can be used to characterize a digital

camera.

i i

i i
 i i

i i

696 12. Image Capture

CIE 1931 Yxy Munsell notation

Number Name Y x y Hue Value / Chroma
1 Dark skin 10.1 0.400 0.350 3 YR 3.7 / 3.2
2 Light skin 35.8 0.377 0.345 2.2 YR 6.47 / 4.1
3 Blue sky 19.3 0.247 0.251 4.3 PB 4.95 / 5.5
4 Foliage 13.3 0.337 0.422 6.7 GY 4.2 / 4.1
5 Blue flower 24.3 0.265 0.249 9.7 PB 5.47 / 6.7
6 Bluish green 43.1 0.261 0.343 2.5 BG 7/6
7 Orange 30.1 0.506 0.407 5 YR 6 / 11
8 Purplish blue 12.0 0.211 0.175 7.5 PB 4 / 10.7
9 Moderate red 19.8 0.453 0.306 2.5 R 5 / 10
10 Purple 6.6 0.285 0.202 5 P 3/7
11 Yellow green 44.3 0.380 0.489 5 GY 7.1 / 9.1
12 Orange yellow 43.1 0.473 0.438 10 YR 7 / 10.5
13 Blue 6.1 0.187 0.129 7.5 PB 2.9 / 12.7
14 Green 23.4 0.305 0.478 0.25 G 5.4 / 8.65
15 Red 12.0 0.539 0.313 5 R 4 / 12
16 Yellow 59.1 0.448 0.470 5 Y 8 / 11.1
17 Magenta 19.8 0.364 0.233 2.5 RP 5 / 12
18 Cyan 19.8 0.196 0.252 5 B 5/8
19 White 90.0 0.310 0.316 N 9.5 /
20 Neutral 8 59.1 0.310 0.316 N 8/
21 Neutral 6.5 36.2 0.310 0.316 N 6.5 /
22 Neutral 5 19.8 0.310 0.316 N 5/
23 Neutral 3.5 9.0 0.310 0.316 N 2.5 /
24 Black 3.1 0.310 0.316 N 2/

Table 12.1. CIE 1931 chromaticity coordinates (xy) and luminous reflectance factor Y
for each of the 24 patches of the Gretag-Macbeth ColorChecker, which is shown in Fig-
ure 12.30. Also given is the equivalent Munsell notation. All values are for patches illumi-
nated with CIE illuminant C.

This particular example is the Gretag-Macbeth ColorChecker, which is formu-

lated to minimize metamerism under different illumination conditions [757]. If
illuminated under standard illuminant C, the Yxy chromaticity coordinates for
each of the 24 patches are given in Table 12.1.4

4 Note that the photograph of Figure 12.30 was taken under D65 illumination, which was not evenly

distributed over the chart. This particular photograph is only intended as an example of a chart that
can be used for calibration—not to allow accurate color reproduction on the basis of this figure. To
achieve good color reproduction, a color rendition chart must be placed in the scene that is to be
photographed.

i i

i i
 i i

i i

12.13. Light Field Data 697

12.12.2 Error Metric

When relatively few responses have been obtained, the measured data may be
viewed in terms of the color difference (or error) in the device and ideal responses
for each stimulus. A transformation can be computed such that the difference
between the transformed device output and the ideal response is minimized. Thus,
the following function has to be minimized:
N
argmin ∑ || f k (pn ) − Pn ||2 , (12.116)
fk n=1

where pn is the pixel value recorded by the camera for the nth stimulus, Pn is the
corresponding measured response, N is the total number of samples, and fk is
the transformation being estimated for the kth color channel. The notation ||.||2
∗ color-difference metric defined in the CIELAB
is introduced to refer to the ΔEab
color space, as discussed in Chapter 8.7.1. The objective function above could be
augmented with additional constraints, for instance to enforce smoothness of the
function fk [65].
Typical techniques for finding the mapping from known data include linear
and polynomial regression, as well as neural networks [575].
Once the camera is calibrated, the recovered function fk can be applied to
all images captured with this camera. Color calibration is necessary for users of
digital cameras to generate high quality images that are faithful representations of
actual input data and are free of any unwanted effects of various processes inside
the capturing device. The data after the mapping is termed device independent.
Data captured from two different devices can be compared after their appropriate
mappings, which may be different for the two devices.

12.13 Light Field Data

The radiance at a point in a scene may be represented by the plenoptic func-
tion [11]. This is a seven-dimensional function L(x, y, z, θ , φ , λ ,t), where x, y, and
z denote the location of a point in space, θ and φ represent the direction of the
ray, λ represents the wavelength of the ray, and t represents time. The plenoptic
function therefore models any ray of any wavelength and direction, originating
at any point in space, at any time. When we consider a monochromatic version
of this function, at a fixed time t, then this function becomes five-dimensional:
L(x, y, z, θ , φ ).
The plenoptic function may be further simplified if the environment is as-
sumed to have no occlusions. The environment may be represented by a cube

i i

i i
 i i

i i

698 12. Image Capture

Figure 12.31. Rays emanating from the surface of the cube can represent the radiance at
any point within the scene (after [11]).

with rays emanating from it in all directions, as shown in Figure 12.31. Now for
any point in the cube, the value of rays may be found by intersecting them with
the two-dimensional faces of the environment cube. This simplified plenoptic
function is only four-dimensional. Rays originating from each face of the cube

(s,t) v

(u,v)
s

z
u
ray(s,t,u,v)

Figure 12.32. Parameterization of the four-dimensional plenoptic function (after [676]).

i i

i i
 i i

i i

12.13. Light Field Data 699

can be represented by L(u, v, s,t), where s and t specify the location on the cube
face, and u and v parameterize the direction of the ray, as shown in Figure 12.32.
This four-dimensional function is variously called a lumigraph, or a light
field [381, 676]. If this function is captured, then new images of the scene from
different viewpoints may be generated from a single capture. To obtain the in-
tensity of a ray emanating from a point in the scene, the intersection of the ray is
found with the (u, v) and (s,t) planes, and the value of the function L(u, v, s,t) is
the intensity of that ray.
Standard cameras are designed so that rays from several different directions
are incident on each sensor location, dependent on the focusing achieved with the
lens assembly. The sensor integrates light over a small solid angle. By modifying
standard camera designs, it is possible to reduce the solid angle over which a
sensor records, thereby separately recording each ray of light passing through
the aperture. Devices that do this are able to capture light field data, i.e., the
four-dimensional plenoptic function. Such devices are briefly introduced in the
following sections.

12.13.1 The Plenoptic Camera

The plenoptic camera was introduced by Adelson and Wang [12]. The main dif-
ference between a standard camera and the plenoptic camera is that it separates
rays before they reach the sensing array. This separation is typically achieved
by placing a micro-lens array in front of the sensing array, as illustrated in Fig-
ure 12.33. The captured light field data is sufficient to generate images of the
scene from a variety of viewpoints and can also be used to infer the depths of
objects in the scene. The latter is called single lens stereo.

Photosensor
Microlens
Array Main Lens Object

Figure 12.33. Structure of the plenoptic camera (after [12]).

i i

i i
 i i

i i

700 12. Image Capture

A portable version of the plenoptic camera, producing better quality images,

was introduced by Ng et al. [834]. From a single capture of light field data, they
generate sharp photographs, focused at different depths.

12.13.2 Artificial Compound Eye Camera

Another imaging system that captures light field data mimics an insect’s com-
pound eye structure [843, 1118, 1119]. The design of this system is similar to
that of the plenoptic camera, except that this system does not have a main lens
(Figure 12.34). Such a design results in the formation of many small images of
the scene (each corresponding to a different micro-lens in the micro-lens array).
These images together form a compound image. As shown in the figure, a unit
constitutes a single micro-lens and its corresponding sensing array. The signal
separator prevents cross-talk between adjacent units.
Once the compound image has been captured, it can be used to generate a
higher-resolution image of the scene. The specific design of such systems allows
extremely compact systems to be made; the minimum thickness achieved is no
more than a piece of paper. However, as noted by Barlow while comparing human
and insect eyes, the spatial resolution that can be resolved by small lens arrays is
severely limited by diffraction [61].

Unit

Microlens Array
Signal Separator
Imaging Device

Figure 12.34. Structure of the compound eye camera (after [1118]).

i i

i i
 i i

i i

12.14. Holography 701

12.14 Holography
Conventional image-recording techniques, for instance photography, record light
for a given exposure time. During that time, light waves impinge on the sensor,
which thereby records intensity information. The recording medium is only sen-
sitive to the power of the light wave. In the particle model of light, the number of
photons is counted. In the wave model of light, amplitude information is retained.
An image sensor does not record where the light came from, other than the
focusing effected by the lens. The image is therefore a flat representation of
the 3D world it captures. Holographic imaging is a technique that attempts to
photographically capture 3D information by recording both amplitude and phase
information.
Recall that most light sources emit incoherent light, i.e., the phase of each
of the superposed waves is uncorrelated (see Section 2.7.1). Thus, during an
exposure, light waves arrive at the photographic plate with random phases. The
exception to this are lasers, which hold the key to holographic imaging. Shining
a laser beam at a photographic plate will cause all the light to arrive in phase.
Such a beam is called the reference beam. As the cross-section of a laser beam is
non-diverging, it has the capability to illuminate only one point of a photographic
plate. However, by inserting a lens into its optical path, the beam can be made to
diverge. This is a process known as decollimation. It then illuminates the entire
photographic plate with coherent light.
If the same laser beam is split, with some part directly aimed at the photo-
graphic plate and the other part reflected by an object, then the path length taken
by the light that is reflected will be longer than the path length of the direct beam.
This second beam is called the object beam. As its path length is longer, the light
will take slightly longer to arrive at the plate and will, therefore, be out of phase
with the reference beam. Of course, each point on the object will create a slightly
different path length.
The phase differences will cause an interference pattern in space, including
on the surface of the photographic plate. Section 2.7.1 discusses interference,
which in summary, are changes in amplitude due to light waves summing or can-
celing out. As a result, the phase difference between object and reference beams
causes local changes in amplitude to which photographic materials are sensitive.
Hence, phase information is encoded as an interference pattern. Such an encod-
ing, invented by Dennis Gabor in 1948 [342], is called a hologram; an example is
shown in Figure 12.35.
To create a hologram, a reference beam of laser light must be aimed at the
photographic plate directly, as well as at the object to be imaged. To reach the

i i

i i
 i i

i i

702 12. Image Capture

Figure 12.35. A photograph of a hologram, which was lit with a D65 daylight simulator
at an angle of 45◦ from above.

entire plate, as well as the entire object, lenses are typically used to make the
laser light divergent. A beam splitter, typically a half-silvered mirror, can be used
to split the light into object and reference beams. It is not possible to use a pair of
lasers for this task, as this would create phase differences unrelated to the object.
An example configuration is shown in Figure 12.36.

Beam splitter
Laser
Mirror
Laser beam

Lens
Object

Photographic
plate

Interference

Mirror Lens

Figure 12.36. Configuration for creating a transmission hologram. The object reflects in
many more directions than indicated, but only the directions toward the photographic plate
are indicated.

i i

i i
 i i

i i

12.14. Holography 703

Virtual Photographic
Object plate

Laser beam
Observer
Diffraction
Laser toward observer
Lens

Figure 12.37. Configuration for viewing a transmission hologram.

To view the hologram, a laser is once more used to illuminate the photographic
plate, as shown in Figure 12.37. In this case, the observer and the illuminant are
located on opposite sides of the photographic plate; thus, this is called a transmis-
sion hologram [267, 427].
Alternatively, it is possible to create a hologram by using the light that is trans-
mitted through the photographic plate to illuminate the object. The light reflected
off the object will then create an interference pattern with the light directly inci-
dent upon the photographic plate. Such a configuration, shown in Figure 12.38,
requires the illumination and the observer to be located on the same side of the
holographic plate. Holograms created in this manner are called reflection holo-
grams. Instead of using a laser to illuminate the hologram, reflection holograms
can also be viewed under white light, because white light incorporates the wave-
length of the laser used to create the hologram [267, 427].
In all cases, the recording environment needs to be carefully controlled for
vibrations, because even very small movements during the recording process will
create large problems due to the nanometer scale at which interference fringes
occur. Usually, optical workbenches are used to combat vibration. In addition,
it is possible to use pulsed lasers, allowing for very short exposure times. To
enable recording of interference patterns in the first place, the spatial resolving
power of the photographic material needs to be very high. The grain of ordinary
photographic materials is too large to record holograms.
For illustrative purposes, we will assume that the object being imaged is the
size of a single point, and that both the light source and the object are far away
from the photographic plate. Then, both the object and reference beams can be
considered plane waves. As discussed in Section 2.7.1, the interference pattern
created by a pair of coherent monochromatic time-harmonic plane waves Eo and
Er , where the subscripts denote the object or reference beam, is expressed by the
irradiance at any point in space. Related to this quantity is the optical intensity,

i i

i i
 i i

i i

704 12. Image Capture

Object

Reflected
light
Laser Photographic
Lens plate
Laser beam
Creation of a reflection hologram

Diffraction
Virtual
toward observer
Object

Photographic
plate
Laser
Lens
Laser beam
Viewing of a reflection hologram

Figure 12.38. Configuration for creating (top) and viewing (bottom) a reflection holo-
gram.

which is defined as

Io = E o E ∗o = |E o |2 , (12.117a)
Ir = E r E ∗r 2
= |E r | , (12.117b)

where E o and E r are the complex amplitudes of their corresponding electric vec-
tors. Adding two waves of the same frequency causes their amplitudes to sum,
yielding the following optical intensity [427]:

I = |E o + E r |2 (12.118a)
2
= |E o | + |E r | 2
+ E o E ∗r + E ∗o E r (12.118b)
2 2 √
= |E o | + |E r | + 2 Io Ir cos (ϕo − ϕr ) . (12.118c)

The third term in this equation is called the interference term and corresponds
to the interference term discussed in Section 2.7.1. In other words, the optical
intensity at each point depends on the magnitude of each of the two waves, plus
a quantity that depends on the phase difference which is only the result of the
difference in path length between the reference and object beams.
Dependent on the quality of the laser and the total path lengths involved in the
holographic imaging process, some of the laser’s coherence may be lost along the

i i

i i
 i i

i i

12.14. Holography 705

way. The factor V ∈ [0, 1] accounts for this loss of coherence. It is known as the
van Cittert-Zernike coherence factor and is computed as follows [343]:
Imax − Imin
V= . (12.119)
Imax + Imin
The values Imin and Imax are the minimum and maximum optical intensity in an
interference pattern when the complex amplitudes of the two plane waves are
equal. The value of V is 1 in the ideal case when coherence is maximal and 0 in
the absence of coherence.
As reflection off a surface may depolarize the light, the optical intensity at a
given point in space may be reduced by an amount linear in the angle between
the two electrical vectors. This is accounted for by the cos (θ ) factor in (12.120).
If no depolarization occurs, this angle is zero and no attenuation occurs. At right
angles, the irradiance becomes zero. Accounting for the loss of coherence, as well
as depolarization, the optical intensity becomes
√
I = |E o |2 + |E r |2 + 2V Io Ir cos (ϕo − ϕr ) cos (θ ) (12.120a)
= |E o |2 + |E r |2 +V (E o E ∗r + E ∗o E r ) cos (θ ) . (12.120b)

After processing of the plate, the transmittance t(x, y) of each point of the plate is
given by a function linear in the optical intensity [427]:

t(x, y) = t0 + β Δt I, (12.121)

where t0 is a constant transmittance term, β is a constant related to the photo-

graphic material used, and Δt is the exposure time.
The complex amplitude transmitted at each point of the hologram, when it
is illuminated by a laser of the same wavelength as the laser used to create the
hologram, is then

u(x, y) = E r t(x, y) (12.122a)

√ 
= E r t0 + β Δt E r |E o |2 + |E r |2 + 2V Io Ir cos (ϕo − ϕr ) cos (θ )
(12.122b)
= E r t0 (12.122c)


+ β Δt E r |E o |2 + |E r |2
+ β Δt V E o |E r |2 cos (θ )
+ β Δt V E ∗o E 2r cos (θ ) . (12.122d)

i i

i i
 i i

i i

706 12. Image Capture

The transmitted amplitude therefore consists of several terms. The first term is the
constant background transmittance, the second term constitutes the unmodified
illuminating beam, the third term is the reconstructed image, and the fourth term
is a twin image [343].
The twin image is undesirable if it overlaps with the reconstructed image.
However, such overlap can be avoided by aiming the reference beam at the pho-
tographic plate at an angle ϑ . The amplitude Er is then given by
 
2 π i x sin (ϑ )
Er = |E r | exp . (12.123)
λ

The implications of this are that the magnitude of the reference beam’s amplitude
remains constant, i.e., |E r |2 = constant, but the square of Er becomes
 
2 4 π i x sin (ϑ )
Er = |E r | exp
2
. (12.124)
λ

This means that the diffracted image is angled away from its twin image by an
angle of 2ϑ . As long as the object being imaged subtends a solid angle smaller
than this, the twin image will not overlap with the desired reconstruction [343].
The interference fringes recorded by the photographic plate cause diffraction
when illuminated with a (coherent) light from the same direction as the reference
beam. Many different types of hologram have been developed; they can be clas-
sified according to whether the photographic plate is a thin layer or volumetric
and whether the hologram is exposed once or multiple times. Although compu-
tationally very expensive, interference patterns can be computed with modified
rendering algorithms and followed by transfer to photographic plates.

12.15 Further Reading

There are many general books on image processing [116, 376, 543, 893, 1071].
Charles Poynton’s book Digital Video and HDTV: Algorithms and Interfaces con-
tains indispensable information for the broadcast, television, and video indus-
tries [923].
Camera sensors and the signal processing in digital cameras is described by
Nakamura’s book, Image Sensors and Signal Processing for Digital Still Cam-
eras [813]. The inner workings of film cameras are described in Goldberg’s Cam-
era Technology: The Dark Side of the Lens [374]. Ansel Adams’s books The
Camera, The Negative, and The Print are indispensable for understanding pho-
tography [8–10].

i i

i i
 i i

i i

12.15. Further Reading 707

Further reading on holography is available in the following references [149,

427, 634, 1060]. Electron holography is discussed in [1184]. Non-technical (i.e.,
non-mathematical) introductions to holography are available in [267, 580, 600].

i i

i i
 i i

i i

Chapter 13
High Dynamic Range
Image Capture

The dynamic range that can be acquired using conventional cameras is limited by
their design. Images are typically gamma encoded to eight bits per color channel.
This means that a luminance range of around two orders of magnitude can be
captured and stored. Many camera manufacturers allow Raw data to be exported
from their cameras as well. These are images that have undergone minimal in-
camera processing and are typically linearly encoded to a bit-depth of around 10
to 14. Note, here, that a linear encoding of 10 bits does not present a larger range
of luminance values than an 8-bit non-linear encoding. Even if Raw data gives
a higher dynamic range than conventional 8-bit images, at this point in time, we
cannot speak of a truly high dynamic range. In other words, limitation in current
camera designs do not allow us to capture the full range of intensities available in
typical real-world scenes.
Some of the trade-offs in camera design are outlined in the preceding chapter,
and include noise characteristics, the capacity to hold a limited amount of charge
before a pixel saturates, veiling glare, and the fact that as new sensor designs
incorporate smaller pixel sizes, quantum efficiency is reduced. All these factors
contribute to a limit on the range of intensities that can be captured in any given
exposure.
To capture the vast range of light present in many scenes, we must therefore
resort to either using dedicated, unconventional hardware, or to reconstructing a
high dynamic range (HDR) image from a set of differently exposed conventional
images. The techniques used for the acquisition of high dynamic range images
are discussed in this chapter.
709

i i

i i
 i i

i i

710 13. High Dynamic Range Image Capture

13.1 Multi-Exposure Techniques

The output generated by standard cameras has several limitations. First, sensors
in standard cameras are unable to capture the wide range of radiance values of-
ten found in our environment. For instance, the response for very light scene
elements, such as specular highlights and light sources, is saturated for radiance
values that are larger than the camera is capable of capturing. The second lim-
itation is that the output is quantized, so that all displayable radiance values are
represented by a total of 256 values in the output. Third, the camera’s firmware
may introduce non-linearities before the photographer is presented with an image.
Much of the image processing that occurs in the firmware can now be avoided by
using the camera’s Raw format.
A variety of techniques and cameras have been developed to obtain high dy-
namic range data. Currently, the cheapest and therefore dominant technique, uses
a conventional camera to take multiple, differently exposed images of the same
scene. These exposures are later recombined in software into a single HDR im-
age. The most accurate approach is to vary the exposure time, rather than change
the aperture or apply neutral density filters [950]. Many cameras can be config-

Figure 13.1. A sequence of exposures taken with different exposure times; St. Barnabas
Monastery, North Cyprus, July 2006.

i i

i i
 i i

i i

13.1. Multi-Exposure Techniques 711

Figure 13.2. The final result, obtained after merging the exposures of Figure 13.1 into a
single high dynamic range image, followed by applying dynamic range reduction to enable
the image to be printed (discussed in Chapter 17); St. Barnabas Monastery, North Cyprus,
July 2006.

ured to take several exposures using auto-bracketing; in some cases up to nine

shots with a single press of the button, each spaced up to one f-stop apart. This
enables the convenient capture of a high dynamic range scene, albeit that the full
range is represented in a sequence of exposures that then need to be recombined.
An example of such a set of exposures is shown in Figure 13.1, and the combined
result is displayed in Figure 13.2. Note that the shorter exposures are necessary
to adequately capture the sky, whereas the longer exposures capture the details of
the tree. The middle exposure of the nine appears perhaps reasonable at first sight,
but note that in this image the sky is already over-exposed. This is a problem that
is resolved by the techniques discussed in this and the following chapters.
Multiple exposure techniques allow a much larger dynamic range to be cap-
tured than can be achieved with single exposures. However, the achievable range
is not infinite due to the occurrence of veiling glare (see Section 12.1.6). In a se-
quence of exposures, some of the images are deliberately under- or over-exposed.
In essence, the camera is used outside its normal operating range, and especially
for the over-exposed images, which are captured to recover dark areas of the en-
vironment, veiling glare may occur. Thus, veiling glare places a limit on the
achievable dynamic range [758,1115], which may either be mitigated by employ-
ing deconvolution techniques [1080], or by placing a mesh between the environ-

i i

i i
 i i

i i

712 13. High Dynamic Range Image Capture

ment and the camera, and recombining multiple high dynamic range images shot
through the mesh with small lateral offsets [1115].
One of the disadvantages of multiple exposure techniques is that the set of
exposures should vary only in exposure time. As such, the scene is assumed to
be static, and the camera position is assumed to be fixed. In practice, these re-
quirements may be difficult to maintain. Ensuring the camera is steady during the
sequence of exposures is relatively straightforward. The use of a tripod mitigates
the worst artifacts associated with moving the camera, while image alignment
techniques can be used as a post-process (see Section 13.5). In the case that hand-
held exposures are taken, aside from image alignment, the effects of camera shake
should be taken into account [309].
A more complicated problem is that the scene needs to remain static. This
requirement can be met in the studio, but especially outdoor scenes can be difficult
to control. In particular, movement of clouds, vegetation, as well as humans and
animals, causes each of the exposures to depict a different scene. Recombining
multiple exposures may then lead to ghosting artifacts. These problems can be
reduced by employing a ghost removal algorithm (see Section 13.4), even though
given the current state-of-the-art, it is still preferrable to ensure the environment
being captured is, in fact, static.

13.1.1 Merging Exposures

To allow multiple exposures to be merged into a single image, each of the ex-
posures needs to be linearized, which is a straightforward task if the camera’s
response curve is known. When this curve is not known, the camera response
function may be computed using the techniques discussed in this section. The
reason that recovery of the camera response function is an important step in high
dynamic range imaging (HDRI) can be seen in Figure 13.3. The response curve
of the green channel for two different cameras is compared with a linear ramp, as
well as with the encoding of the green channel in the sRGB color space. The four
curves are all different from each other. It is therefore not appropriate to assume
that the output generated by a camera is either linear or adheres to any other color
standard, such as the encoding used in sRGB.
Once the response curve is obtained, each pixel’s image irradiance value Ee
can be computed. Considering a single pixel, its value p is due to some incident
irradiance Ee , recorded for an exposure time of Δt. The relationship between the
variables Ee and p can be expressed as

p = g(Ee Δt). (13.1)

i i

i i
 i i

i i

13.1. Multi-Exposure Techniques 713

1.0

Normalized irradiance
0.9 Linear response
sRGB curve
0.8 Nikon response curve
Minolta response curve
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Normalized pixel value

Figure 13.3. Camera response curves (green channel only) compared with sRGB and
linear response functions. The Nikon response curve is for a Nikon D2H, whereas the
Minolta curve is for a DiMAGE A1 camera.

The product Ee Δt is also referred to as the exposure. The camera response func-
tion g represents the processing of the irradiance values within the camera. As
camera manufacturers typically do not make the nature of this processing
available to their customers, the camera response curve g needs to be reverse-
engineered from the captured images.
To find Ee from the corresponding pixel values p, the function g needs to be
inverted to give
g−1 (p)
Ee = . (13.2)
Δt
The following section discusses how high dynamic range imagery may be ob-
tained using multiple linearized images of the same scene; the remaining sections
discuss techniques to compute the camera response function.

13.1.2 Multiple Linearized Exposures

When an image has been linearized, its dynamic range is limited by the range that
the camera sensor is able to capture at one time. One method to obtain a higher
dynamic range is to use multiple linearized images that were captured using dif-
ferent exposures. By altering the exposure settings of the camera, it is possible to
capture a different range of radiances of the scene. Exposure settings that allow
less light to hit the sensor will cause lighter regions to be exposed correctly in the
resulting image, while the darker regions will appear too dark. Similarly, when

i i

i i
 i i

i i

714 13. High Dynamic Range Image Capture

the exposure time is increased and more light is allowed to go through the aper-
ture, darker regions of the scene will be exposed correctly while brighter regions
will appear as burnt-out areas in the image. If enough images are captured of the
same scene with different exposure times, then each region of the scene will be
correctly exposed in at least one of the captured images.
Information from all the individual exposures in the image is used to compute
the high dynamic range image for the scene. The process of image generation is
as follows:
N
g−1 (pn (x, y))
∑ w(pn (x, y)) Δt
Ee (x, y) =
n=1
N
(13.3)
∑ w(pn (x, y))
n=1
For each pixel, the pixel values pn from each of the N exposures are linearized
by applying the inverse of the camera response g. The resulting values are then
normalized by dividing by the exposure time Δt. The irradiance value Ee of the
high dynamic range image is a weighted average of these normalized values.
Both the weight function and the camera response need to be determined be-
fore this equation can be used effectively. In the following section, we discuss
possible choices for the weight function. Different algorithms to recover the cam-
era response g are discussed afterward.

13.1.3 Weight Functions

The weights in (13.3) are used to decrease the impact of pixel values that are
incorrectly exposed. For instance, low weights may be given to pixel values that
are close to 0 or 255, as these values are possibly the result of under- or over-
exposure. Following this argument, Debevec et al. [238] use a hat function for
their weights:
⎧
⎨ pn − pmin if pn ≤ pmax + pmin ,
w(pn ) = 2 (13.4)
⎩ p − p if p > pmax + pmin .
max n n
2
A possible function for the weights may be derived from the signal to noise
ratio θ , which for an irradiance value Ee is
d pn 1
θ = Ee (13.5a)
dEe σN (pn )
g−1 (pn )
= , (13.5b)
dg−1 (pn )
σN (pn )
d pn

i i

i i
 i i

i i

13.2. Response Curve Recovery 715

where σN (pn ) is the standard deviation of the measurement noise. If the assump-
tion is made that the noise is independent of pixel value pn (a valid assumption
for all noise sources bar photon shot noise, see Section 12.10.4), then the weight
function can be defined as [784]:
g−1 (pn )
w(pn ) = . (13.6)
dg−1 (pn )
d pn

13.2 Response Curve Recovery

The following sections discuss four different techniques to recover the camera re-
sponse function g and recombine images taken with multiple exposures into high
dynamic range images. The recovery of the camera response curve is necessary
to enable the construction of high dynamic range images. In addition, it is a mea-
sure of the non-linearity in the output of a camera and can, therefore, be seen as a
partial calibration of the camera.

13.2.1 Method of Mann and Picard

Mann and Picard propose to estimate the camera’s response function using as few
as two images of the same scene, captured with different exposure times [731].
Starting with a dark pixel value in the image with the shorter exposure time, this
pixel value is related to the image irradiance as specified in (13.1). The corre-
sponding pixel in the second image is related to the image irradiance as follows:
pn+1 = g(k Ee Δt), (13.7)
where the exposure Ee Δt is k times larger than the first exposure. We then look
for the pixel in the darker image that has the same value pn+1 , and we also record
the value of the corresponding pixel in the second image. This process is repeated
until the maximum output value of the function g is reached. We now have a
set of pixels that are all related to each other: (g(Ee Δt) relates to g(k Ee Δt)), and
(g(k Ee Δt) in turn relates to g(k2 Ee Δt)), while (g(k2 Ee Δt) relates to g(k3 Ee Δt)),
etc.
These pixel pairs are then used to generate a range-range plot, as shown in
Figure 13.4. To obtain the camera response function from these points, some
assumptions about the nature of this function are made. The function is assumed
to be a semi-monotonic function, and g(0) = 0. The function is also assumed to
follow a power curve [731]:
 γ
g(Ee Δt) = α + β Ee Δt . (13.8)

i i

i i
 i i

i i

716 13. High Dynamic Range Image Capture

g (k Ee’ Δt)

g (Ee’ Δt)

Figure 13.4. An illustration of the range-range plot defined by Mann and Picard (af-
ter [731]).

Here, α is 0 (from the assumption that g(0) = 0), and β and γ are scalar values.
The pixel values in each pair can be related to each other as follows:
 γ
pn+1 = β k Ee Δt (13.9a)
= k γ pn . (13.9b)
The value of γ may be computed by applying regression to the known points in
the range-range plot. There is now sufficient information to compute the camera’s
response function. Note that this solution will give results accurate up to a scale
factor.

13.2.2 Method of Debevec and Malik

Debevec and Malik propose to compute the function g by solving a system of
linear equations [238]. The input into their algorithm is also a set of N images of
the same scene, with varying exposures. The scene is assumed to be static, so that
the irradiance at each point on the sensor remains the same across the image set.
To compute the inverse camera response function g−1 , the reciprocity assump-
tion needs to hold, i.e., halving Ee in (13.1) and simultaneously doubling Δt will
result in the same exposure value. Under extreme circumstances, such as a very
long or short exposure time, this reciprocity assumption may not hold. For typi-
cal analog sensors, reciprocity holds to within one-third f-stop for exposure times
between 1/10, 000 to 10 seconds. The reciprocity assumption holds for digital
cameras under the assumption that the sensor measures the total number of inci-
dent photons in the exposure time [238]. If we make the assumption that g is a

i i

i i
 i i

i i

13.2. Response Curve Recovery 717

monotonically increasing function, its inverse becomes well defined and may be
computed given the specified input.
The linear system of equations is derived as follows. Taking the log on both
sides of (13.1), the following equation is obtained:

ln(g−1 (pn (x, y))) = ln(Ee (x, y)) + ln(Δtn ). (13.10)

In this equation, the unknowns are ln(g−1 ) and ln(Ee (x, y)). As the function g−1
is only applied to 256 values, essentially it represents 256 unknowns. The other
set of unknowns, ln(Ee (x, y)), has as many unknowns as the number of pixels in a
single image from the input. All these unknowns may be estimated by minimizing
the following objective function:

N 0 12
O = ∑ ∑ ∑ ln(g−1 (pn (x, y))) − ln(Ee (x, y)) − ln(Δtn )
x y n=1
255
+α ∑ (ln(g−1 )) (p)2 . (13.11)
p=0

Here, α is a user-specified weight that determines the relative importance of each

part of the objective function. The first term of (13.11) is derived from (13.10),
while the second term is the sum of the squared values of the second derivative of
ln(g−1 ), and it is used to ensure smoothness in the function. It is computed by

(ln(g−1 )) (p) = ln(g−1 (p − 1)) − 2 ln(g−1 (p)) + ln(g−1 (p + 1)). (13.12)

Minimizing the objective function O is a linear least-squares problem and can

be solved by singular value decomposition [163]. To ensure that the system is
over-constrained, the number of knowns needs to be larger than the number of
unknowns, which for images of dimensions h by w, is equal to 256 + hw. This
means that not all the pixels in an image need to be used as known values in
this system of equations, as long as enough pixels are used to over-constrain the
system. Pixels should be selected for inclusion in the system of equations on the
basis of how well they represent the range of displayable values. Good candidates
are those that belong to uniform regions in the image and those that cover as wide
a range of luminance as possible.
Finally, the equations in the system may be weighted, based on their prob-
ability of being correct. Debevec and Malik use the weight function described

i i

i i
 i i

i i

718 13. High Dynamic Range Image Capture

in (13.4). After incorporating these weights, the objective function becomes

N 2 0 132
O = ∑ ∑ ∑ w(pn (x, y)) ln(g−1 (pn (x, y))) − ln(Ee (x, y)) − ln(Δtn )
x y n=1
255 0 12
+α ∑ w(p) (ln(g−1 )) (p) . (13.13)
p=0

This algorithm can be used to obtain the function g−1 for a single channel. The
above process, however, needs to be repeated for each of the three color channels.

13.2.3 Method of Mitsunaga and Nayar

Another technique to compute the camera response was proposed by Mitsunaga
and Nayar [784]. This technique models the camera response with a flexible
high-order polynomial and recovers the parameters for this model using a set of
images of the same scene captured with different exposures. The advantages of
this technique are that it is able to compute an accurate camera response despite
noise in the input images. It also does not require exact values of the exposures of
images. Given rough estimates, it computes accurate exposure information along
with the camera response function. The model of the inverse of camera response
g−1 is as follows:
K
g−1 (p) = ∑ ck pk . (13.14)
k=0
The response function is represented by a polynomial of order K applied to image
values. The minimum order K required to represent a function will depend on the
complexity of the function, so finding the polynomial representation of the func-
tion will include finding the value of K as well as the values of all coefficients ck .
Whereas in most approaches the image irradiance Ee is assumed to equal the
scene radiance Le , Mitsunaga and Nayar’s technique takes into account the size
of the camera’s aperture to relate Ee to Le . The pixel value p is a function of the
image irradiance:
Ee = g−1 (p). (13.15)
The relation between Ee and Le is given by (12.26) [488]. The image values p can
now be expressed as a function of scene radiance and shutter speed as well as the
aperture size:
   
π d 2 4
p = g Le cos (Θ)Δt (13.16a)
4 f
= g(Le ρ e), (13.16b)

i i

i i
 i i

i i

13.2. Response Curve Recovery 719

where ρ is a constant that is determined by properties of the camera, and e is

the expression that encapsulates the variation in image irradiance due to both the
shutter speed and the aperture size:

cos4 (Θ)
ρ= , (13.17)
f2
 2
πd
e= Δt. (13.18)
4
The ratio R of the linearized values is used to generate the objective function. For
a pixel at location (x,y) in the nth exposure, this ratio can be written as
 (x, y)Δt
Ee,n n Le (x, y) ρ (x, y) en
 = (13.19)
Ee,n+1 (x, y)Δtn+1 Le (x, y) ρ (x, y) en+1

g−1 (pn (x, y))

= (13.20)
g−1 (pn+1 (x, y))
= Rn,n+1 . (13.21)

Substituting the inverse camera response with the polynomial model (13.14) into
(13.21), the following relationship is obtained:
K
∑ ck pkn (x, y)
k=0
Rn,n+1 = K
. (13.22)
∑ ck pkn+1 (x, y)
k=0

Using Equation 13.22, the following minimization function is obtained:

4 5
N−1 K K
O= ∑ ∑ ∑ ∑ ck pkn (x, y) − Rn,n+1 ∑ ck pkn+1 (x, y) . (13.23)
n=1 x y k=0 k=0

The minimum of the function occurs when the derivative of the function O with
respect to ck equals zero:
δO
= 0. (13.24)
δ ck
As with previous techniques, this gives a solution up to a scale factor. However,
when a single constraint is added, the system has a unique solution.
Thus, given approximate exposure ratios Rn,n+1 , the unknown parameters of
the inverse response function may be computed. After this computation, the actual
values of the exposure ratios may be found by searching in the neighborhood of

i i

i i
 i i

i i

720 13. High Dynamic Range Image Capture

the initial ratio estimates. When more than two images are provided, searching
for the ratios can be time consuming, and, therefore, an iterative version of the
above process is used. First, user-specified approximate ratio estimates are used
to compute function parameters as discussed above. Next, these parameters are
used to compute an updated value of the ratio, using the following equation:
K
∑ ck pkn (x, y)
Rn,n+1 = ∑ ∑
k=0
K
. (13.25)
∑
x y
ck pkn+1 (x, y)
k=0

The updated ratio is again used to compute function parameters, and this cycle
continues until the change in linearized values (caused by the change in func-
tion parameter estimates of successive iterations) for all pixels in all the different
images is smaller than a small threshold:
" "
" −1 "
"g(i) (pn (x, y)) − g−1
(i−1) (p n (x, y))" < ε, (13.26)

where ε is the threshold on the change across iterations, and i is the iteration
number.
The order K of the polynomial representation of the camera response function
can be computed by repeating the above process several times, each time with a
different test value of K. For instance, the function parameters may be estimated
for all K values less than 10, and the solution with the smallest error is chosen.
For color images, the algorithm may be applied separately for each channel. Per-
Normalized irradiance

Normalized irradiance

1.0 1.0
0.9 R Nikon D2H 0.9 R Konica Minolta DiMAGE A1
G G
0.8 B 0.8 B
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.0 0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Normalized pixel value Normalized pixel value

Figure 13.5. The response curves recovered for each of the red, green, and blue channels
of a Nikon D2H and a Konica Minolta DiIMAGE A1 camera.

i i

i i
 i i

i i

13.2. Response Curve Recovery 721

channel response curves for the Nikon D2H and Konica Minolta DiMAGE A1
cameras generated with this technique are shown in Figure 13.5.

13.2.4 Method of Robertson et al.

An alternative algorithm was proposed by Robertson et al. [967]. Their HDR
assembly algorithm weights each pixel by a weight factor as well as the square of
the exposure time:
N
∑ g−1 (pn (x, y)) w(pn (x, y))tn
Ee (x, y) = n=1
N
. (13.27)
∑ w(pn (x, y))tn2
n=1

Here, the summation is over each of the N exposures; tn is the exposure time for
the nth exposure, and g is the camera’s response curve. This approach is motivated
by the fact that longer exposures tend to produce a better signal-to-noise ratio.
This weighting therefore tends to produce less noise in the resulting HDR image.
The camera response curve can be recovered by solving the following objec-
tive function O, for instance using Gauss-Seidel:
N  −1 2
O= ∑ ∑ ∑ w̄(pn (x, y)) g (pn (x, y)) − tn Ee (x, y) , (13.28)
n=1 x y
 
(pn (x, y) − 127.5)2
w̄(pn (x, y)) = exp −4 . (13.29)
127.5

The weight function w(pn (x, y)) can then be set to the derivative of the camera
response function:
w(pn (x, y)) = g (pn (x, y)). (13.30)
To facilitate differentiating the camera response curve, a cubic spline is fit to
the camera response function. The appropriate selection of the spline’s knot
locations is crucial to obtain sensible results. As under- and over-exposed pixels
do not carry information, it is desirable to set the weight function for these
pixels to zero. In Robertson’s method, this means that the derivative of the
camera response needs to be close to zero for both small and large pixel
values.
To achieve this, the spline that is fit to the camera response curve must have a
slope of zero near 0 and 255. This makes the shape of the camera response curve
recovered with Robertson’s method markedly different from the curve obtained

i i

i i
 i i

i i

722 13. High Dynamic Range Image Capture

with the techniques presented in the preceding sections. Each camera can only
have one response curve, and it is therefore reasonable to expect that different
response curve recovery algorithms produce qualitatively similar results. As a re-
sult, HDR assembly may proceed according to (13.27), provided that the weights
w(pn (x, y)) are not computed as the derivative of the camera response curve.

13.3 Noise Removal

Night scenes often have a high dynamic range, for instance, due to the presence
of street lights. Due to the overall low light levels, the exposure times tend to
be relatively long. However, it is desirable to keep the exposure times relatively
short, so that object movement is kept to a minimum. One way to achieve this
is to increase the sensitivity of the sensor. Increasing the ISO setting in digital
cameras is equivalent to amplifying the analog signal before it is passed through
the A/D converter. This procedure invariably increases the noise present in the
individual exposures.
In particular, the short exposures in the sequence will show relatively high
noise levels. Due to the division by the exposure time in the HDR assembly pro-
cess (see Equation (13.3)), noise in short exposures is amplified relative to longer
exposures.
Assuming that each of the images in the sequence captures the same scene,
and that the only difference is the exposure time, a modified form of image aver-
aging may be used. In standard image averaging, a set of identical photographs
of the same scene is “averaged” into a single, low-noise image. Accounting for
the different exposure times, HDR image assembly and noise removal may be
combined [24]:
a+s
g−1 (pn (x, y)) h(pn (x, y), a)
∑ tn
n=a
ca = a+s
, (13.31)
∑ h(pn (x, y), a)
n=a
N
∑ cn w(g(tn cn ),Ycn )
Ee (x, y) = n=1
N
. (13.32)
∑ w(g(tn cn ), Icn )
n=1

Here, the noise averaging is carried out for each pixel in each exposure over a
set of corresponding pixels in s subsequent exposures. In general, using a larger
value of s will give better noise reduction without introducing artifacts. However,

i i

i i
 i i

i i

13.3. Noise Removal 723

Dynamic range
6.8
6.6
6.4
6.2
Noise averaging
6.0 Excluding frames
5.8
5.6
5.4
5.2
5.0
2 3 4 5 6
s

Figure 13.6. The dynamic range obtained for different values of s, the parameter that
determines how many exposures are averaged in the noise-reduction step.
Weight

1.0
Threshold function τ
0.8

0.6

0.4

0.2

0.0
0 50 100 150 200 250
Pixel value

Figure 13.7. The multiplier reducing the weight for over-exposed pixels.

there is a maximum value for s beyond which no further image enhancement is

obtained, as shown in Figure 13.6. Limiting the number of exposures to s, rather
than averaging over all exposures longer than the one being filtered, will therefore
save some computation time.
Figure 13.6 also shows the resulting dynamic range obtained by excluding the
s shortest exposures in the sequence without noise averaging. Here, the dynamic
range is reduced as more exposures are removed, caused by a lack of range in the
light areas of the image.
The weight function h(pn , a) is set so that the noise is minimized. We assume
that the main source of noise in dark environments is photon shot noise, which is
noise arising from the fact that photons arrive at the sensor at random intervals.
The Poisson distribution is a good model for photon shot noise [116].

i i

i i
 i i

i i

724 13. High Dynamic Range Image Capture

In such a distribution, the mean equals the variance. Assuming that a 1-second
exposure results in the recording of N photons, then the mean μ equals N, and the
variance σ also equals N. For an exposure time t times longer, the mean will
be μ = tN, and the variance is σ = tN. During HDR assembly, pixel values
are divided by the exposure time. This means that for a t-second exposure, after
division the mean and variance are [116]
tN
μ= = N. (13.33)
t
tN N
σ= 2 = . (13.34)
t t
As a result, increasing the exposure time by a factor of t reduces the variance, and,
therefore, the noise, by a factor of t. This means that the weight factor applied to
each pixel should be linear in the exposure time. The function h(pn , a) is therefore
given by 
tn n = a,
h(pn (x, y), a) = (13.35)
τ (pn (x, y))tn n = a.
Thus, if the pixel pn (x, y) is from exposure a, the exposure time itself is taken
as the weight function. For pixels from a different exposure than a, the exposure
time is multiplied by a factor τ (pn ) which accounts for over-exposed pixels:
⎧
⎪
⎪ 0 ≤ x < 200,
⎨1
τ (x) = 1 − 3k(x)2 + 2k(x)3 200 ≤ x < 250, (13.36)
⎪
⎪
⎩0 250 ≤ x ≤ 255,
250 − x
k(x) = 1 − (13.37)
50
The multiplier τ is plotted in Figure 13.7.
Equation (13.32) uses a weight function w that depends on both the pixel
value pn (x, y), which is taken to be either a red, green, or blue value. In addition, a
representation of the pixel’s luminance Ycn , derived from the recorded pixel values
using Equation (8.9), is part of this weight function to avoid saturated pixels that
might give an unnatural color cast. The weight function itself is given by:
4   5
−1  Ycn − 127.5 12
w(pn (x, y),Ycn ) = g (pn (x, y)) g (pn (x, y)) 1 − . (13.38)
127.5

A result with and without application of the noise-removal scheme is shown

in Figure 13.8.

i i

i i
 i i

i i

13.3. Noise Removal 725

Figure 13.8. The image on the left was created without noise averaging, whereas the
frames were first noise averaged for the image on the right.

Figure 13.9. Movement of scenery between the capture of multiple exposures (shown at
top) lead to ghosting artifacts in the assembled high dynamic range image (bottom). This
is apparent in the lack of definition of the sheep’s head, but also note that the movement in
the long grass has resulted in an impression of blurriness; Carnac, France, July 2007.

i i

i i
 i i

i i

726 13. High Dynamic Range Image Capture

Figure 13.10. High dynamic range action photography (top left) is not yet feasible using
multi-exposure techniques. Windy conditions spoiled the high dynamic range photograph
on the bottom left. Note that in either case, a single exposure would lose detail in the
sky, as shown in the images on the right. The images on the left were tonemapped with
the photographic operator (see Section 17.3). Top images: Bristol Balloon Fiesta, Ashton
Court Estate, August 2007. Bottom images: Carnac, France, July 2007.

13.4 Ghost Removal

As discussed earlier in this section, multiple images, with different exposures of
the same scene, may be used to generate high dynamic range images. Techniques
that merge these images to form the high dynamic range image make a few im-
plicit assumptions about the capturing process and scene. One assumption is that

i i

i i
 i i

i i

13.4. Ghost Removal 727

the scene must remain stationary while the images are being captured, so that the
scene radiance entering the camera remains a constant for each pixel across the
images. If there is movement in the scene at the time of capture, then for at least
some pixels, scene radiance will be different across the image set, and the values
will no longer be comparable (Figure 13.9). Further artifacts due to motion in the
scene at the time of image capture are illustrated in Figure 13.10. Such artifacts
are commonly known as ghosts in the image. There are a variety of algorithms
that attempt to eliminate such errors, and these are discussed here.

13.4.1 Ghost Removal using Optic Flow

The input into the algorithm proposed by Kang et al. [573] is a video of a scene,
where exposures of frames are alternatively short and long. To maintain a high
frame-rate in the resulting HDR sequence, for each short exposure, a long expo-
sure will be inferred from the two neighboring frames. Similarly, for each long
exposure, a corresponding short exposure will be inferred from the neighboring
long exposures. The result of this procedure is that for each frame, two exposures
are available—one captured and the other computed from neighbors. The two
exposures per frame can then be combined into a single HDR frame.
The computed frames thus use information from adjacent captured frames.
For the frame at time t that has a long exposure Lk , the adjacent frames will have

Sk-1 Lk Sk+1

fkF fkB fk*

Boost Boost
fkF fkB
B0
SkF0* Lk-1 Lk+1 Sk*

Interpolate

Boost
Sk* Lk*

fk*

Sk‘ *

Figure 13.11. The process used by Kang et al. to compute frames estimating the unknown
frame Sk . Dotted lines indicate that a warp is being computed, while solid lines labeled
with a warp indicate that the warp is being applied (after [573]).

i i

i i
 i i

i i

728 13. High Dynamic Range Image Capture

short exposures; we denote these by Sk−1 and Sk+1 . These two frames are used to
generate an approximation of Sk that is the kth frame with a short exposure. The
process by which this information is estimated is illustrated in Figure 13.11 and
described below.
To estimate the information contained in the missing frame Sk , first bidirec-
tional flow fields are computed using a gradient-based approach. Bidirectional
F0
flow fields consist of the forward warp fk,F that is applied to Sk−1 to give Sk∗
B0
and the backward warp fk,B that is applied to Sk+1 to give Sk∗ . To compute these
flow fields, the exposures of Sk−1 and Sk+1 are boosted to match the long expo-
sure. Then, motion estimation between these boosted frames and the kth frame is
performed to generate the flow fields. Kang et al. use a variant of the Lucas and
Kanade technique [703] to estimate motion between frames.
The two new frames Sk∗ F0 and SB0 are interpolated to give S . The exposure
k∗ k∗
of this frame is adjusted to match the exposure of Lk , to give Lk∗ . The flow field
fk∗ between Lk∗ and Lk is computed using hierarchical global registration. After
 . To compute the final HDR image
computation, it is applied to Sk∗ to give Sk∗
F0 , SB0 , and S . Note, that if the
at time t, the following frames are used: Lk , Sk∗ k∗ k∗
frame at time t has a short exposure, the same process may be used to estimate
the long exposure.
To compute an HDR image from captured and computed frames, the follow-
ing procedure may be used:
F0 , SB0 , and S are converted to radiance values by us-
• The frames Lk , Sk∗ k∗ k∗
ing the camera response function and known exposure times. The radiance
F0 , ŜB0 , and Ŝ , respec-
values of the above frames may be denoted by L̂k , Ŝk∗ k∗ k∗
tively. The response function may be computed using any of the techniques
discussed earlier in this section.

• For all pixel values in L̂k that are higher than a certain threshold, the corre-
sponding values in the HDR image are taken directly from Ŝk∗  , since high

values in L̂k are considered to be saturated and not good enough to pro-
vide reliable registration with adjacent frames. Note that in the case where
the frame at time t has a short exposure, all values in Ŝk below a specified
threshold are assumed to provide unreliable registration, and corresponding
values in the HDR image are simply taken from L̂k∗  .

• In the remaining regions, the radiance map values are computed by

fMW (Ŝ∗(x,y)
F0 , L̂
(x,y) )Ŝ∗(x,y) + fMW (Ŝ∗(x,y) , L̂(x,y) )Ŝ∗(x,y) + fW (L̂(x,y) )
F0 B0 B0
E(x,y) =
fMW (Ŝ∗(x,y)
F0 , L̂
(x,y) ) + f MW (Ŝ∗(x,y) , L̂(x,y) ) + fW (L̂(x,y) )
B0

(13.39)

i i

i i
 i i

i i

13.4. Ghost Removal 729

1 1

fW fM

0 0
0 1 0 dmax
Pixel Value d

Figure 13.12. The flow fields fW (left) and fM (right).

where the subscript k has been dropped as it is common to all frames, (x, y)
denotes pixel location, fW denotes a weight function which is illustrated in
Figure 13.12, and fMW (a, b) = f M (|a − b|) × fW (a), where f M is given by

⎧  3  2
⎪
⎨ δ δ
2 −3 + 1 if δ < δmax ,
fM (δ ) = δmax δmax (13.40)
⎪
⎩0 otherwise.

This function is also illustrated in Figure 13.12. Its main purpose is to en-
sure that if the warped radiance values are very different from the captured
radiance values, then they should be given a relatively small weight in com-
putation of the final radiance values.

13.4.2 Variance-Based Ghost Removal

Techniques that rely on accurate motion estimation to remove ghosts from HDR
images may not work well whenever motion estimation fails. Scenes in which
motion estimation cannot as yet be performed accurately, include those where
an object is visible in some frames but not in others [1206]. Motion estimation
is also prone to errors when the scene includes phenomena like inter-reflections,
specularity, and translucency [573].
Ward introduced an alternate technique that does not rely on motion estima-
tion [953]. The underlying assumption is that for each pixel location, a single
correctly exposed pixel value is adequate to represent the scene radiance at that
location. Therefore, in regions where there is movement in the scene, a single
exposure may be used to provide values for the HDR image, as opposed to a

i i

i i
 i i

i i

730 13. High Dynamic Range Image Capture

weighted average of several exposures. If this approach is followed, ghosts will

not appear in the final HDR image.
Ghost removal may thus be divided into two sub-problems. First, ghost re-
gions need to be identified in the image. Second, for each pixel in these regions,
an appropriate exposure needs to be selected from the image set that will provide
the scene radiance value for that pixel in the HDR image.
To identify the regions where ghosts occur, the weighted variance of each
pixel in the image can be computed. If a pixel has a high variance, it is possible
that this variance is due to movement in the scene at the time of capture. Ghost
regions are, therefore, identified as the set of pixels that have variance above 0.18.
Values from a single exposure may be used in the HDR image for ghost re-
gions. However, replacing regions of high dynamic range with values from a
single exposure will reduce the dynamic range of the HDR image. Another way
to deal with ghost regions is to find the most appropriate exposure for each pixel
in the ghost region and to use its value in the HDR image. A possible problem
with this approach is that a moving object may be broken up, since the most ap-
propriate pixels that represent the object may come from different exposures.
To avoid these problems, the ghost regions may first be segmented into sep-
arable patches. The most appropriate exposure that will provide radiance values
for each of these patches in the HDR image is then determined. For each patch, a
histogram is generated using the corresponding values in the original HDR image.
The top 2% are discarded as outliers, and the maximum in the remaining values
is considered. The longest exposure that includes this value in its valid range is
selected to represent that patch.
Finally, interpolation is performed between these patches and their corre-
sponding patches in the original HDR image. Each pixel’s variance is used as
the interpolation coefficient. This ensures that pixels that have low variances, but
lie within a patch identified as a ghost region, are left unaltered.

13.4.3 Ghost Removal in Panoramic Images

A similar approach to removing ghosts from images has been proposed by Uyt-
tendaele et al. [1167]. This approach is not designed specifically to remove ghosts
from high dynamic range images. However, this approach for ghost removal is
easily adapted for HDR images. Similar to the preceding algorithm, this method
is not able to handle situations in which there is high contrast within a single ghost
region.
Given a set of input images of a scene, Uyttendaele et al.’s algorithm creates
panoramas of the scene in which ghosting artifacts have been removed. Once the
images are registered, the first task is to identify regions in each input image that

i i

i i
 i i

i i

13.4. Ghost Removal 731

belong to moving objects. Once these are found, a single instance of each moving
object is selected and used in the panorama.
Regions in each input image that belong to moving objects are found by com-
paring each pixel in an image with corresponding pixels in the remaining images.
If a pixel differs from corresponding pixels by more than a threshold amount, it is
identified as representing a moving object. Once all pixels in all the images have
been evaluated in this manner, a region-extraction algorithm is applied to each
image to identify and label contiguous regions. These are known as regions of
difference (ROD).
RODs in different images are assumed to represent the same moving object if
they overlap. Thus, to have a single instance of a moving object in the panorama,
only one of the overlapping RODs is used in the final image. To find a single
instance of each object, RODs are represented by a graph. Each ROD is a vertex
in this graph, and if two RODs overlap, their corresponding vertices in the graph
are connected with an edge. To obtain a single instance of a moving object, the
vertex cover can be found and then removed from this graph. This leaves behind
a set of disconnected vertices, representing a single instance of all moving objects
in the scene.
As there are multiple instances of each moving object, any one of these may
be used in the panorama. To ensure that the best possible instance is used, each
ROD or vertex is given a weight. The closer the ROD is to the center of the image
and the larger its size, the higher its corresponding vertex weight. Thus finding
and removing the vertex cover with the minimum weight will leave behind the
most appropriate vertex or instance of object in the panorama.

13.4.4 Ghost Removal using Kernel Density Estimation

Khan et al. [586] have proposed a way to remove ghosts in high dynamic range
images even if the ghosting appears in a high contrast region. For example, this
method makes it easier to capture a candle flame. In addition, this approach is
not dependent on explicit motion detection or detection of regions within the im-
age where ghosting is likely to occur. Ghost removal is achieved by making a
straightforward change to the high dynamic range image-generation process as
given by (13.3); a change in the way that the weights are computed.
As discussed previously, a weight is assigned to a pixel based on the probabil-
ity that the pixel represents a correctly exposed value. Here, weights are computed
that are also based on the probability that pixels belong to the static part of the
scene. If pixels representing moving objects in the scene are given small weights,
then the corresponding high dynamic range image will show less ghosting.

i i

i i
 i i

i i

732 13. High Dynamic Range Image Capture

The probability that a pixel belongs to the static part of the scene may be
computed statistically by representing the static scene with a distribution and then
finding the probability that the candidate pixel belongs to this distribution. If such
a probability is very small, it will indicate that the pixel does not belong to the
static scene and, instead, represents a moving object. Such a pixel is given a
small weight. On the other hand, a pixel with a large probability is given a large
weight. As the probability is correlated with a suitable weight for each pixel, the
weight of a pixel may be set equal to this probability.
If it is assumed that most of the pixels in any local region in the image set
represent the static part of the scene, then the distribution of the static part may be
approximated by the pixels in the immediate neighborhood of the pixel for which
the weight is being computed. Using weighted kernel density estimation [881,
980], this probability is given by

∑ w(y)KH (x − y)
y∈F
P(x|F) = , (13.41)
∑ w(y)
y∈F

where F is the distribution representing the static scene, y is an element of F

that represents a pixel, and KH is the kernel function with a bandwidth specified
by H. Each element y is a five-dimensional vector, where three dimensions denote
the color and two dimensions denote the distance of the pixel from the candidate
pixel x. Khan et al. use the identity matrix to specify the bandwidth. The ini-
tial weights w(y) are based on each pixel’s probability of representing a correctly
exposed value as discussed in Section 13.1. The neighborhood of pixel x is rep-
resented by all the pixels in its immediate vicinity, in all the differently exposed
images.
The final weight of a pixel can now be computed by multiplying its initial
weight with the probability computed above, to give

wt+1 (x) = w0 (x) P(x|F). (13.42)

While this procedure reduces ghosting artifacts in the image, they are still
apparent in the final HDR image, and further processing is typically required.
The algorithm can improve weights iteratively by using the updated weights from
(13.42) as the initial weights of pixels y belonging to the distribution in (13.41).
The updated probability can be used once more to compute the weights for each
pixel, and the process can be repeated in this manner until the weights converge.
The final weights will show little if any ghosting, as illustrated by Figure 13.13.

i i

i i
 i i

i i

13.5. Image Alignment 733

Figure 13.13. Ghost removal using (in reading order): 1, 4, 7, and 10 iterations. (Khan
et al., “Ghost Removal in High Dynamic Range Imaging,” IEEE International Conference
on Image Processing, 2005–2008, c 2006 IEEE [586].)

13.5 Image Alignment

If the image set is captured with a hand-held camera, the images are often mis-
aligned due to hand motion at the time of capture. This causes objects to appear
blurred at the edges, and in cases of severe camera movement, can mean signif-
icant ghosting in the final high dynamic range image. Thus, a multiple-exposure
technique to combine several images into a single high dynamic range image
should consider alignment of the separate exposures.
While an obvious solution is to consider standard image alignment techniques,
this is often not possible, because the exposures in the set are all different. This
causes different parts of the scene to be properly exposed in different images,
making alignment a more difficult problem.
One method of aligning images is to use optic flow, as discussed in Sec-
tion 13.4.1. Motion is estimated between scenes captured by two consecutive
images. For moving objects, motion vectors may exist only for certain regions
in the image (those representing the object). For misaligned images, the motion
vectors are likely to exist for larger portions of the image and, potentially, for the
whole image.
A second technique was proposed by Ward et al. [1214]. This method is able
to align images in the case of translation of the camera at the time of image cap-
ture. It does not rely on motion estimation or a known camera response function.
First, binary maps of the images are computed by setting all of the pixels above a
threshold value to one and the remaining pixels to zero. The threshold for an im-

i i

i i
 i i

i i

734 13. High Dynamic Range Image Capture

Figure 13.14. The image on the left was generated using misaligned images. The image
on the right was created with the same set of images after applying Ward’s image alignment
technique [1214]. Insets of a representative area are shown in the middle. (Image courtesy
of Greg Ward.)

age is the median value from a low resolution histogram over the grayscale image
pixels. This conversion of images to binary values is akin to normalizing the im-
ages and making them comparable with each other. The difference between two
consecutive images can be computed by using an exclusive-or (XOR) operator to
evaluate each pixel and then summing the responses over all pixels.
If there is a misalignment, this difference will be large. An upper limit has to
be set on the magnitude of translation that is used to limit the amount of transla-
tion. This process can be sped up significantly by generating image pyramids for
each image. Image alignment may now be performed at various stages, starting
with the smallest images. The final translation obtained in one level can be used
as the starting translation in the next level. Figure 13.14 shows a result obtained
by this technique.

13.6 Single Capture High Dynamic Range Images

The previous section discusses the generation of a high dynamic range image from
multiple captures of the same scene, where each capture has a different exposure.
Such captures can be made with standard cameras. While this is an important
way of generating HDR images with inexpensive and commonly available cam-
eras, one of the drawbacks of such a scheme is that for most cameras, exposure
settings need to be changed manually after each image capture. This can be quite

i i

i i
 i i

i i

13.6. Single Capture High Dynamic Range Images 735

tedious and time consuming. Furthermore, the scene needs to remain static while
the images are being captured. As this is often not the case, movement in the
scene between captures results in the appearance of artifacts in the reconstructed
high dynamic range image. These artifacts may be removed in many cases with
techniques discussed in Section 13.4.
To avoid such shortcomings altogether, cameras have been constructed that
capture multiple exposures of the scene at the same time (with a single click of
the button). As there is no time lapse between the capture of images, movement
in the scene at the time of capture is less likely to be a problem. The captured
exposures are then used off-line to generate the final HDR image.
Such devices may capture multiple exposures of the scene in the form of mul-
tiple images, or a single image, where the exposure of pixels varies across the
image. In the latter case, the image may be used to construct multiple images
with different, uniform exposures. Multiple images obtained with a single cap-
ture are then used to generate the HDR image as discussed in the previous section.
Another important advantage of these cameras over standard cameras is that
they are amenable to recording high dynamic range video. Techniques that make
HDR video from multiple exposures need to address the issue of image registra-
tion and motion estimation for scenes where there is any motion. Neither of these
problems are completely solved, and they often generate errors. Devices that cap-
ture multiple images with a single capture tend to suffer less from these problems
and can, therefore, be used without modification in a larger set of environments.
We now briefly describes such cameras.

13.6.1 Split Aperture Imaging

As its name implies, a split aperture imaging device splits the light entering the
device into multiple beams, with the help of a beam splitter [14, 471, 526].
Multiple sensors are placed within the device, such that each beam is incident
upon a different sensor. Each beam goes through a modulator that sets the expo-
sure for the sensor that it is incident upon. Sensors’ responses are akin to multiple
captures of the same scene with varying exposures; these can be post-processed
into HDR imagery.
Split aperture imaging devices enable the capture of high dynamic range in-
formation with a single capture, but they have some disadvantages that discourage
their use. First, such devices are often very expensive, since they require multi-
ple sensors. Second, they invariably require precise optics for alignment of the
different images.

i i

i i
 i i

i i

736 13. High Dynamic Range Image Capture

e0 e3

e1 e2

Figure 13.15. Multiple exposures may be captured by placing this filter in front of the
sensing array (after [817]).

13.6.2 Spatially Varying Pixel Exposures

Devices with spatially varying pixel exposures avoid some of the problems en-
countered with split aperture imaging. These devices capture multiple exposures
of the same scene simultaneously on a single sensor. This is achieved by sampling
across both the spatial and exposure dimensions of image irradiance at the same
time [817]. This can be achieved by placing a spatially varying filter in front of
the sensor. Such a filter is shown in Figure 13.15. Here, a 2 × 2 window in this
filter will have four different transmission coefficients, where e0 < e1 < e2 < e3 .
This pattern is repeated across the filter.
If such a filter is placed in front of the camera sensor, it will affect the ir-
radiance at each sensor location. Four different exposures, with reduced spatial
resolution, will be captured on the sensor.
The post-processing of images captured with this method begins by discarding
over- and under-exposed values in the four images and then finding the surface
that best fits the undiscarded values [817]. Such a surface representation that
treats the image as a height field is then resampled to construct four complete
images with appropriate resolution. These images are then recombined into a
single HDR image.

13.6.3 Adaptive Dynamic Range Imaging

An adaptive dynamic range (ADR) device adapts to the different radiances from
different parts of the scene. This adaptation is done separately for each pixel,
and, therefore, each pixel sensor is optimally exposed for the radiance that it cap-
tures [816]. The advantage of such devices over those discussed in the previous
section is that the spatially varying filter is not static, and they can therefore cap-
ture a larger class of scenes well.
An optical attenuator is placed in front of the sensor array, and its transmit-
tance depends on two factors: its transmittance in the previous iteration and the

i i

i i
 i i

i i

13.7. Direct High Dynamic Range Capture 737

Transmitter

Transmitter
Sensor

Sensor
Modified
Scene Scene Scene
Radiance Radiance Radiance

Controller Controller

Figure 13.16. The concept of adaptively varying pixels’ exposures. At first, the attenuator
uniformly transmits the light (left). Once the sensors have accumulated some irradiance,
their response is used by the controller to adjust the optical attenuator (right) (after [816].

corresponding sensor’s value. If the pixel value is saturated, then transmittance is

decreased. Therefore, in the next iteration, the pixel’s value will have a smaller
probability of being saturated. The transmittance is continually adjusted in this
manner until pixel values are no longer saturated. A similar process is applied for
under-exposed pixels. Transmittance is increased in such cases to allow higher
irradiance for such pixels. Figure 13.16 illustrates this concept.

13.7 Direct High Dynamic Range Capture

The techniques to capture high dynamic range images introduced so far all depend
on multiple exposures, even if they are generated and recombined in-camera. A
more durable approach would be to develop cameras that can directly capture the
full dynamic range available in real-world environments. Such cameras are both
rare and expensive and have only recently started to appear on the market. Several
such devices are being developed simultaneously by different companies:

Grass Valley. One of these devices is the Viper FilmStream camera, which is pro-
duced by Grass Valley (a division of Thomson). This film camera is capable
of capturing about three orders of magnitude, which is well matched to the
dynamic range reproducible in movie theaters, but is otherwise known as
medium dynamic range.

i i

i i
 i i

i i

738 13. High Dynamic Range Image Capture

SMaL. SMaL Camera Technologies (Cambridge, Massachusetts) have developed

a CMOS sensor that is capable of capturing four orders of magnitude. They
currently sell two different products that use this sensor, marketed as secu-
rity cameras and consumer-level photo cameras.

Point Grey Research. Point Grey Research (Vancouver, Canada) offer a video
camera that captures six HDR images in a single shot. Six sensors cap-
ture HDR information from different parts of the scene, and the combined
information can be used to form environment maps.

Panavision. Panavision builds the Genesis, a film camera that outputs 10 bits
of log-encoded full resolution film, qualifying the Genesis as a medium
dynamic range camera. It serves the same market as the Viper. The sensor
is a 12.4 megapixel-wide gamut CCD device that does not employ a Bayer
pattern.

Pixim. Pixim (Mountain View, California) offer two CMOS sensors that can each
capture a dynamic range of roughly four orders of magnitude.

SpheronVR. SpheroCam HDR, a camera offered by SpheronVR (Kaiserslautern,

Germany) is the highest performance HDR camera currently on the market.
It captures nearly eight orders of magnitude of dynamic range, at a very
high spatial resolution. It uses a line scan CCD for capture and, therefore,
needs to physically rotate to capture a scene. This means that full spherical

Figure 13.17. A full 360◦ scan taken with the SpheroCam HDR camera. The result was
tone mapped with the photographic operator (see Section 17.3). Sunrise in Tübingen, Ger-
many, taken from the rooftop of the Max Planck Institute (MPI), November 2006. (Pho-
tograph taken by Narantuja Bujantogtoch as part of the MPI for Biological Cybernetics
Database of Natural Illuminations.)

i i

i i
 i i

i i

13.8. Application: Drawing Programs 739

captures can be taken, but a disadvantage is that it takes from fifteen to

thirty minutes to complete a full 360◦ scan. An example image taken with
this camera is shown in Figure 13.17.
It is anticipated that when these devices become more affordable, they will replace
other forms of HDR generation. Their use will not require HDR assembly as a
post-processing step, thereby removing a limitation that prevents multi-exposure
techniques to reach the mass market. Further HDR sensor and camera designs are
discussed in Hoefflinger’s book [475].

13.8 Application: Drawing Programs

Many of the limitations associated with conventional photography are addressed
and removed in high dynamic range imaging (HDRI). High dynamic range (HDR)
images are often called radiance maps. Being able to capture real-world lumi-
nance has many advantages in photography and film [952], as well as for novel
applications such as image-based lighting [239, 240] and image-based material
editing (discussed in Section 13.9) [587].

Figure 13.18. An interactive brush showing a histogram of the pixels under the circle.
This user interface allows the user to accurately select absolute luminance values, even if
they are outside the display range. The area under the brush is displayed directly, after ap-
propriate exposure control is applied, whereas the remainder of the image is tone-mapped
prior to display. (Background image courtesy of Paul Debevec. This image was pub-
lished in Mark Colbert, Erik Reinhard, and Charles E Hughes, “Painting in High Dynamic
Range,” Journal of Visual Communication and Image Representation,  c Elsevier 2007.)

i i

i i
 i i

i i

740 13. High Dynamic Range Image Capture

On the other hand, new challenges emerge. For instance, how can one ma-
nipulate high dynamic range images when they cannot be displayed directly on
conventional display devices? This is a new issue that arises in drawing programs
that offer support for high dynamic range images. With a luminance range that
exceeds the display range, drawing directly onto an HDR canvas can be likened
to key-hole surgery: only part of the display range can be visualized at any given
time. Typically, one may solve such an issue by drawing on tone-mapped repre-
sentations of images (see Chapter 17 on dynamic range reduction). Unfortunately,
it is not clear that this approach is either intuitive or accurate.
In such cases, an alternative visualization may help, for instance by attaching
a histogram visualization for the pixels under a region of interest (indicated by the
size and position of a brush), allowing accurate selection of absolute luminance
values. An example is shown in Figure 13.18. A qualitative approach may exploit
visual phenomena familiar to artists, such as glare to allow the user to intuitively
map a visual representation of glare back to the luminance value with which the
brush paints [202].

13.9 Application: Image-Based Material Editing

There are several advantages to using high dynamic range images that are only
just beginning to be explored. Although veiling glare may limit the range of
values that can be captured, there is no limit on the range of values that can be
represented in a suitably chosen file format. As a result, image histograms tend
to have a long tail, which is a significantly different shape than histograms of
conventional images tend to have. This makes it easier to determine, for instance,
which pixels potentially represent highlights.
Further, the lack of quantization means that a much richer amount of infor-
mation is available in images, which may simplify image processing tasks. For
instance, gradient fields computed in images are smoother. As an example, the use
of high dynamic range images for the purpose of estimating the shape of objects
(shape from shading [1293]) does not have to overcome the limitations imposed
by quantization.
These advantages, along with the fact that some distortions in images are not
easily perceived, have led to the development of an image-editing technique that
enables substantial changes to the apparent material of objects in images [587].
The input to the algorithm is a single high dynamic range image, along with an
alpha matte which defines the set of pixels belonging to the object of interest.
The algorithm then extracts an approximate 2.5D shape of the object, i.e., a
depth map, and separately computes a representation of the environment, based on

i i

i i
 i i

i i

13.10. Further Reading 741

Figure 13.19. The pixels representing the bronze statue (left) were processed to pro-
duce the impression of transparency (right); Loet Vanderveen bronze, Native Visions Art
Gallery, Winter Park, FL. (
c 2006 ACM, Inc. Included here by permission [587].)

the pixels representing the background. The object shape as well the environment
map are then plugged into the rendering equation (2.207). The user is then free
to choose an arbitrary BRDF fr , thereby changing the apparent material of the
object.
In many cases, it is beneficial to simplify this rendering procedure by filter-
ing the environment directly in several different ways before texture mapping the
result onto the object. This has the added benefit of introducing desirable high
frequency content which helps mask inaccuracies in the recovered shape, as well
as inaccuracies in the environment. It allows objects to be converted to either
transparent or translucent objects. The result is that plausible material transfor-
mations can be achieved (shown in Figure 13.19), without requiring 3D scanning
equipment or multi-view techniques to derive more physically accurate 3D geom-
etry. However, it is noted that without high dynamic range images as input, this
approach would fail.

13.10 Further Reading

Hardware for high dynamic range image and video capture is discussed in detail in
the book edited by Hoefflinger [475]. A more general discussion on high dynamic
range imaging is available in High Dynamic Range Imaging [952].

i i

i i
 i i

i i

Chapter 14
Display Technologies

In producing a sensation of color, regardless of what the source of information

is—be it a natural image, an abstract stimulus, or a computer-generated scene—
display devices are used to convey information. Displays form an interface be-
tween man and machine, allowing the latter to convey visual information. Al-
though all display devices share this fundamental goal, they achieve this in very
different ways.
Understanding the operating principles of display devices, their limitations,
commonalities, and differences is essential in displaying high-quality images.
Additionally, in many cases, one should also know how to calibrate and char-
acterize a display device to ensure accurate color reproduction.
This chapter first focuses on the different types of display technologies, both
conventional and emerging, along with their operating principles. In particular,
we focus on liquid crystal display devices, as these are currently the prominent
form of display. Then properties common to all display devices such as pri-
maries, white point, brightness, contrast, and gamma are discussed. Finally, we
focus on display characterization which is an indispensable step for accurate color
reproduction.

14.1 Cathode-Ray Tubes (CRTs)

The cathode ray tube (CRT) is the oldest type of electronic display technology
dating back to the second half of the nineteenth century [214, 473]. It has also
been the most ubiquitous display technology until recently, but has now largely
been replaced by more ergonomic flat-panel displays (see Section 14.2).
743

i i

i i
 i i

i i

744 14. Display Technologies

Cathode Focusing Vertical

anode deflection Phosphor screen
coils

Heater
Control
grid Accelerating
anode
Horizontal
deflection
coils

Figure 14.1. The components of a CRT display device.

A CRT display is made up of four key components encapsulated in an evac-

uated glass envelope: an electron gun, a deflection system, a shadow mask (or
aperture grille), and a phosphor-coated screen. Visible light is produced when the
electrons produced by the electron gun hit the phosphor coated screen at a position
determined by the deflection system. This operation is illustrated in Figure 14.1.
The electron gun is responsible for producing and focusing the electron beam.
When the cathode is heated by means of a heater, electrons escape from its sur-
face by a phenomenon called thermionic emission. The control grid has a negative
potential with respect to the cathode, pushing electrons towards its center. A posi-
tive potential (around 20 kV) is applied to the accelerating anode which causes the
electrons to exit the control grid at high velocities. These electrons are focused to
form an electron beam using the focusing anode [1120].
In the absence of vertical and horizontal deflection, such an electron beam
would hit the center of the screen causing the center to glow. However, usually
the aim is to produce visible light over the entire screen surface. Steering the
beam toward any desired point on the screen surface may be accomplished by
deflecting the beam using magnetic or electrostatic forces [1120]. Magnetic de-
flection employs two pairs of coils, one pair in the horizontal and the other pair
along the vertical axis of the electron gun. The magnetic field produced by each
pair is perpendicular, and, by adjusting the current passing through each coil, the
electron beam can be deflected towards any direction on the phosphor screen.
The deflected electron beam travels in a vacuum until it collides with the phos-
phors. Upon collision, some of the electrons in the phosphor material jump to a
higher energy state and then return back to their original states. Typical phosphor
decay times are between 20 and 50 μ s. Depending on the type of the phosphor
used, this causes the emission of a photon with a wavelength determined by the

i i

i i
 i i

i i

14.1. Cathode-Ray Tubes (CRTs) 745

Blue Electron guns

Red Green Aperture grill

Shadow mask

Phosphor screen Phosphor screen

Figure 14.2. Shadow mask and aperture grill used in color CRTs ensure that each of the
three types of phosphor receives light from the correct beam.

change in energy levels. This phenomenon is called cathodoluminescence which

is a type of luminescence (see Section 3.5.6).
The color from a CRT is obtained by using three types of phosphors that emit
light predominantly in the long-, medium, and short-wave regions of the visible
spectrum (see Figure 7.17). Each type of phosphor is excited by separate beams
produced by three electron guns. To ensure that each gun excites the correct type
of phosphor, a masking mechanism is used [464]. The most common types of
masks, the shadow mask and the aperture grill, are shown in Figure 14.2. A
close-up photograph of a CRT display is shown in Figure 3.38.
To obtain an image on the CRT screen that appears to be steady, the electron
beam must sweep through the screen multiple times every second. Convention-
ally, a sweep cycle starts from the top-left corner of the screen and scans from
left to right and top to bottom. When all of the scanlines are swept, the beam is
deflected back to the top-left corner for the next cycle. Such a systematic sweep-
ing pattern is called a raster. For television, scanning is normally performed with
very nearly the same frame rate as the alternating current power that supplies the
display device (60 Hz in the US and 50 Hz in Europe).
Sensitivity to flicker increases with ambient light level. “Flicker-free” televi-
sions use signal processing to create interpolated fields or frames and drive the
display at double the prevailing main frequency. Because they are used in rela-
tively bright environments, computer monitors have a higher refresh rate, achiev-
ing for instance 85 Hz.

i i

i i
 i i

i i

746 14. Display Technologies

A common technique to reduce information rate (“band-width”) for a given

refresh rate is called interlacing. Here, during each vertical sweep the odd- or
the even-numbered scanlines are alternately displayed. This technique exploits
the fact the the human visual system fills in the missing information in alternating
scanlines.

14.2 Liquid Crystal Displays (LCDs)

Liquid crystals (LC) are organic fluids that exhibit both liquid and crystalline-like
properties. They behave as liquids in that LC molecules can move freely within
their material. However, they can be made to maintain a specific orientation, un-
like true liquids. In addition, they exhibit two important properties that make them
suitable for use in display devices. First, liquid crystals are transparent. Second,
their orientation, and hence their effects on light polarization, can be changed by
applying an external electrical field [810] (see Section 2.3 for polarization).
These physical properties form the basis of liquid crystal display technologies.
In this section, we focus on displays that are operated in transmissive mode, i.e.,
a back-light (uniformly) illuminates an LCD screen, which modulates per pixel
how much light is transmitted. Each pixel contains either a red, a green, or a
blue filter, so that color images are formed by appropriately addressing triplets of
pixels1 . Each of the components of LCD displays are discussed in the following.

14.2.1 Back-Lights
The light sources used in LCD displays are most typically cold-cathode fluores-
cent lamps (CCFLs), although LED-based back-lights are just entering the mar-
ket. Both types are discussed here.
A CCFL consists of a glass-tube with electrodes at both ends. The tube con-
tains an inert gas and a small amount of mercury which emits ultraviolet radiation
when stimulated by an applied field. The inside of the tube is coated with a blend
of phosphors which luminesce when excited by the ultraviolet radiation, such that
the spectrum produced by the light source is seen as white, although the spectrum
may comprise of bands that are spectrally distinct (see Figure 14.3) [170]. A dif-
fuser is placed between the light source and the LCD screen to ensure uniform
illumination.
1 Note that in our definition, a pixel is a single unit which emits light according to a single spectral

power distribution. Another definition is possible, whereby a pixel consists of three units, peaked at
long, medium, and short wavelengths.

i i

i i
 i i

i i

14.2. Liquid Crystal Displays (LCDs) 747

2.5

Le (W/m2)
R
G
2.0
B

1.5

1.0

0.5

0.0
400 500 600 700

Figure 14.3. The spectral power distribution for typical phosphors used in CCFL back-
lighting (adapted from a plot provided by Louis Silverstein, VCD Sciences, Inc., Scottsdale
AZ).

In a typical LCD display, the power consumption is dominated by the CCFL.

Dynamic dimming techniques are introduced to minimize the energy use [170,
175, 349]. Based on the imagery displayed, it may be possible to dim the back-
light and increase the transmittance of the LCD panel to compensate.
Another approach which can effectively reduce power requirements is the use
of LED technology for back-lighting. This technology offers further advantages
in the form of enhanced ruggedness and stability under a wide range of environ-
ment temperatures. LEDs are mercury and lead-free.
Further, with LEDs it is possible to create much wider color gamuts, due to
the use of saturated primaries [569], and to also increase the light output [1237].
While a color gamut of around the same size as the NTSC color gamut is achiev-
able with pc-LEDs, mixed RGB LEDs can produce a wider color gamut. Recall
from Section 9.8 that LEDs can be constructed with sharp spectral peaks having
a width of around 20–30 nm. As a result, their primaries will lie close to the
spectral locus of a chromaticity diagram. In addition, it is possible to construct
back-lights with more than three primaries (for instance five or six primaries) to
further extend the available gamut [1059].
The luminance requirements of the back-light have an impact on its design.
For standard displays emitting less than 750 cd/m2 , the back-light can be mounted
along the edge of the display, whereas for brighter displays a direct-view back-
light will be required. Edge-lit and direct-view displays are diagrammed in
Figure 14.4.

i i

i i
 i i

i i

748 14. Display Technologies

Figure 14.4. Example diagrams of edge-lit (left) and direct-view back-lights (right).

One of the problems with LED back-light design is temperature dissipation.

The printed circuit board acts as a heat sink. The thermal resistance of recent
high-flux LEDs is in the order of 13–15 ◦ C/W for the Luxeon emitter, 9 ◦ C/W for
the Osram Golden Dragon, and 2.6 ◦ C/W for the Lamina BL2000. This means
for example that for a 3 W Luxeon emitter, the LED die temperature would rise
to around 39 ◦ C above ambient. Thus, thermal resistance puts an upper bound
on luminous output, as the associated temperature rise will limit device reliabil-
ity [1059].

14.2.2 Liquid Crystal Phases

Liquid crystals may be found in several phases such as nematic, cholesteric, and
smectic [1120]. For a given material, the phase it assumes depends on temper-
ature, as shown in Figure 14.5 [706]. Below temperature Tm , the material is
anisotropic and crystalline. Above temperature Tc , the material is isotropic and
liquid. In between, three phases can be distinguished, which increase from more
ordered to less ordered structures and are collectively known as the mesophase.
These phases are named smectic C, smectic A, and nematic. The molecules in

Solid Smectic C Smectic A Nematic Liquid

Tm Tc Temperature

Figure 14.5. The different phases that liquid crystals can assume depend on temperature.
Inbetween temperatures Tm and Tc the material behaves as a liquid, albeit with ordered
structure, and has a milky appearance.

i i

i i
 i i

i i

14.2. Liquid Crystal Displays (LCDs) 749

Cholesteric

Figure 14.6. In the cholesteric phase, the liquid crystal molecules are aligned parallel to
the front and back substrates (i.e., the tilt angle is 0), and form a helical structure.

the smectic phases are oriented alike, and additionally form layers. The smectic
C phases is characterized by molecules showing small random deviations from a
preferred direction which is angled with respect to the orientation of the layers.
In the smectic A phase, this tilt angle becomes perpendicular to the surfaces of
the layers. In the nematic phase, the molecules are oriented alike, but do not form
layers. All phases are anisotropic, except for the liquid phase, which due to the
random orientation of the molecules is isotropic.
It is possible to add chiral compounds, which change the nematic phase to
the cholesteric phase. In this phase, the directors of calamitic (i.e., rod-shaped)
molecules in subsequent layers form a helical structure, as shown in Figure 14.6.
The director is the vector describing the long axis of a liquid crystal molecule.
Additional phases may be distinguished, but these are less important for the man-
ufacture of liquid crystal displays.
The director of each molecule makes an angle θ with the average direction.
The wider the spread of angles around the average, the more chaotic the ordering
of the molecules, and the more isotropic the appearance of the material. The order
parameter S of a phase is a measure of the optical properties that may be expected,
and is given by [706]

S = 0.5 3 cos2 (θ ) − 1 , (14.1)

where θ is the angular deviation of individual liquid crystal molecules from the
average director, and . indicates an average over a large number of molecules.
In a fully ordered state, the molecules all perfectly align, yielding S = 1. In a fully
unordered state, we obtain S = 0. In a typical nematic phase, the order parameter
S is between 0.4 and 0.7 [706]. The ordering of calamitic molecules in smectic
and nematic phases gives rise to birefringence, as discussed in Section 2.6. This
optical property is exploited in liquid crystal displays.

i i

i i
 i i

i i

750 14. Display Technologies

y z
Pre-tilt angle
α0

x x
Rubbed alignment layer

Figure 14.7. The orientation of liquid crystal molecules in the presence of a rubbed
surface.

14.2.3 Rubbing
The orientation of liquid crystal molecules is non-random in the mesophases. In
the construction of a liquid crystal display, it is important to force the alignment
of the molecules with respect to each other, but also with respect to a substrate.
Liquid crystals are normally sandwiched between glass plates. Before any electric
field is applied, the molecules nearest to the glass plates can be oriented in a
preferred direction by creating microscopic grooves. This is achieved by coating
the glass plates with a layer of polyimide material, which is then rubbed with
a fine cloth with short fibers [1112]. The coated and rubbed polyimide layer is
called the alignment layer.
This rubbing procedure creates microscopic grooves to which liquid crystal
molecules will align. The orientation is in the plane of the glass substrate, but
also causes the molecules to orient slightly away from the substrate, known as the
pre-tilt angle α0 . The orientation of liquid crystal molecules in the rest-state is
shown in Figure 14.7.

14.2.4 Twisted Nematic Operation in the Field-Off State

In a twisted nematic display, both glass substrates are coated and rubbed. How-
ever, the rubbing direction of the second substrate is orthogonal to the direction
of the first substrate (see Figure 14.8). The molecules near the surface of the
first substrate are therefore aligned with an angle of 90◦ to the molecules near the
surface of the second substrate. The molecules in between are twisted in a helix
structure that rotates over a twist angle of β = 90◦ between the first and second
substrate. This alignment is visualized in Figure 14.9.

i i

i i
 i i

i i

14.2. Liquid Crystal Displays (LCDs) 751

Horizontal
polarization filter Grooved
alignment layers Vertical
polarization filter
Unpolarized
Glass with
backlight
colored filter
Address
transistor

Liquid crystals Colored

polarized
light

Figure 14.8. A pixel in a twisted nematic active matrix LCD display.

The next step towards understanding the operation of a twisted nematic liquid
crystal display is to consider how polarized light propagates through a twisted
nematic liquid crystal layer. We assume that the light is incident at a right angle,
i.e., along the z-axis in Figure 14.9. The birefringence of the aligned molecules
polarizes light into two orientations. By placing linear polarizing filters before
the first and after the second alignment layer, light entering and exiting the liquid
crystal layer can be linearly polarized.
The first polarizing filter is known as the polarizer, whereas the second polar-
izing filter is known as the analyzer. The polarizer is required, as without it, the
liquid crystal cell by itself would not show the desired optical effect. The amount
of light transmitted through the analyzer depends on the angle of polarization pro-
duced by the liquid crystal layer with respect to this filter (as given by Malus’ law
in (2.86)).

Layer thickness d

y
x

Figure 14.9. The orientation of liquid crystal molecules in the presence of a pair of
orthogonally oriented rubbed surfaces.

i i

i i
 i i

i i

752 14. Display Technologies

Horizontal
polarization filter Grooved
alignment layers Vertical
polarization filter
Unpolarized
backlight

Polarized light
Liquid crystals

Horizontal
polarization filter Grooved
alignment layers Vertical
polarization filter
Unpolarized
backlight

Liquid crystals

Figure 14.10. The stack of LCs takes a twisted shape when sandwiched between two glass
plates with orthogonal surface etching (top). They rotate the polarization of light passing
through them. Under an electric field, liquid crystal molecules can be realigned, and light
transmission can be blocked as a result of the light having polarization orthogonal to the
second polarizer (bottom). This example shows the normally white mode.

When propagating through the liquid crystal layer, which is assumed to have
a thickness of d (typically in the range of 3.5 to 4.5 μ m), the birefringence of
the material changes the light’s polarization from linear to circular to linear again,
provided that the layer thickness is chosen to be
√
3λ
d= , (14.2)
2 Δn
where λ is a typical wavelength, normally chosen to be yellow at 505 nm, and
Δn is the difference of the indices of refraction parallel and perpendicular to the
director, as discussed in Section 2.6 and introduced in Equation (2.141).
If the direction of the polarizing filters is the same as their corresponding
alignment layers, then light can pass through. This is called the normally white
mode, an example of which is shown in Figure 14.10. If the polarizer and align-
ment layers are crossed, then light cannot pass through, and this is called the
normally black mode.

i i

i i
 i i

i i

14.2. Liquid Crystal Displays (LCDs) 753

14.2.5 Realigning Liquid Crystals

The helically twisted shape of the liquid crystals, as determined by the two crossed
alignment layers, can be modified by applying a voltage. To this end, electrodes
can be added to the sandwich. The configuration of the electrodes across the
display is intimately linked with the addressing scheme used to switch individual
pixels. This is discussed in Section 14.2.9.
For now, we assume that a single pixel is formed by augmenting the lay-
ered structure with two electrodes, across which a voltage can be applied. Such
electrodes must be transparent to enable light to pass through. They are typically
manufactured using a thin layer of indium tin oxide (ITO), typically 100 nm thick.
Due to the anisotropic dielectric constant associated with the liquid crystal
molecules (Δε , see Section 2.6), the electric field created by applying a voltage
across the electrodes exerts a small force, which realigns the molecules to the
electric field. This alignment occurs throughout the liquid crystal layer, with the
exception of those molecules that are in close proximity to the alignment layers.
Inter-molecular forces resist this realignment, so that rotation of the
molecules begins at a threshold voltage Vth . Full alignment is reached at a voltage
Vmax , where all molecules, except those in contact with the alignment layers, are
rotated parallel to the field (assuming that Δε > 0). Note that the electric field
changes both the tilt angle α (perpendicular to the x-y plane of Figure 14.9) and
the twist angle β (in the x-y plane).
As the realignment causes light to reach the second polarizer at different an-
gles dependent on the applied voltage, the amount of light transmitted through the

T 0.5

0.4
Normally black mode

0.3

0.2

0.1 Normally white mode

0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0
V/Vth

Figure 14.11. The transmittance T of a twisted nematic cell as a function of relative

driving voltage V (after [706]). As discussed further in Section 14.2.7, the transmittance
has a maximum of 0.5.

i i

i i
 i i

i i

754 14. Display Technologies

display depends on this voltage. The transmittance T as a function of voltage V

for both the normally white and normally black modes is shown in Figure 14.11.
When a voltage is applied, the liquid crystal molecules realign themselves
as function of inter-molecular forces and the externally applied electric field.
As these molecules are relatively large, their physical movement takes a certain
amount of time. The dynamic viscosity ηd , measured in Ns/m2 , of the liquid crys-
tal material depends on the force F required to move a molecule with area A over
a distance d with velocity v, as follows:

Fd
ηd = . (14.3)
Av

The kinematic viscosity η , measured in mm2 /s, equals the dynamic viscosity ηd
divided by the density δ of the material:

ηd
η= . (14.4)
δ

However, the density of liquid crystal materials hovers around 1 Ns2 /mm4 , so
that the value of both types of viscosity in liquid crystal displays are close to
identical. The viscosity of liquid crystal materials is typically in the 20 to 45
mm2 /s range [706]. The effect of the viscous properties of liquid crystal materials
is that the time to realign the molecules for each pixel is not instant, but takes a
little while, typically at least a few milliseconds and perhaps tens of milliseconds.
In the presence of electric and magnetic fields, the equilibrium that is reached
depends on three types of deformation to the liquid crystal material. These are
splay, twist, and bend, and their torques are characterized by the elastic constants
K11 , K22 , and K33 , respectively (see Figure 14.12). The values associated with
these torques are in the order of 10−3 N [706].

Equilibrium Splay (K11) Twist (K22) Bend (K33)

Figure 14.12. Splay, twist, and bend in liquid crystal materials (after [706]).

i i

i i
 i i

i i

14.2. Liquid Crystal Displays (LCDs) 755

y y
β0 y’
E0
s
β
x’
x
z=0 β0
β

x
P z=d
s’
t t’

Figure 14.13. Coordinate system at z = 0 (left) and z = d, where d is the thickness of the
cell.

14.2.6 Light Propagation

We consider a pixel configured as a twisted nematic cell. The coordinate system
assumed puts the surface of the cell in the x-y-plane. The light incident on the cell
is then in the z-direction. The twist angles β are also in the x-y-plane, whereas
the tilt angles are perpendicular to this plane. The liquid crystal molecules form a
helical shape with its axis in the z-direction.
Further, we assume that the light incident upon the cell is linearly polarized
with the polarization oriented with an angle β0 with respect to the x-axis. The
coordinate system assumed in the following computations is sketched in Fig-
ure 14.13. Thus, we no longer require the polarizing filter at the first substrate
to be aligned with the grooves in the first alignment layer. The electric field vec-
tor E0 therefore gives rise to the Jones vector J, given by
   
J E0
J = x = R (−β0 ) , (14.5)
Jy 0
where R(θ ) is a rotation matrix,
 
cos(θ ) sin(θ )
R (θ ) = . (14.6)
− sin(θ ) cos(θ )
The Jones vector is defined in Section 2.6. The twist angle βz at position z along
the z-axis is equal to
βz = β0 z, (14.7)
so that the final twist angle after passing through the cell with thickness d is
βd = β0 d. (14.8)

i i

i i
 i i

i i

756 14. Display Technologies

The assumption made here is that the twist angle changes linearly along the z-
axis. This is a valid assumption when no voltage is applied to the cell. When
a voltage is applied, then the total twist angle (as well as the tilt angle) remains
the same. However, the twist angle then varies non-linearly with z. The model
discussed below can be extended to deal with this situation [1270].
Under the assumption of a linearly progressing twist angle, the pitch p of the
helix is constant, and is given by
2π
p= . (14.9)
β0
Although we omit the derivation, it is now possible to compute the direction of the
linear polarization with respect to the analyzer after light has passed through the
cell, which has thickness d. Assume that the analyzer is oriented in direction P,
as shown in Figure 14.13. The Jones vector, given in the s -t  coordinate system,
which is the coordinate system associated with the analyzer, is [706, 1270]
     
Js −2 π i n̄ d E0
= R (−Ψ) exp M R (−β0 ) , (14.10)
Jt  λ 0

where the key matrix M is given by

 
cos (γ ) + i a β sin (γ ) β sin (γ )
M= . (14.11)
−β sin (γ ) cos (γ ) − i a β sin (γ )

The unknowns in these equations are

n + n⊥
n̄ = , (14.12a)
2
p Δn
a= , (14.12b)
2λ
2π d
β= , (14.12c)
p

γ = β 1 + a2 , (14.12d)
2π d
s= . (14.12e)
pε

Equation (14.10) is a general expression that is valid for several types of ne-
matic display. In particular, if the tilt angle α is chosen to be 0, and the twist
angle is β = π /2, then we have a classic twisted nematic display. For α = 0 and
β > π /2, the display is called a super-twist nematic display. Finally, α = 0 and
β = π /2 yields a so-called mixed-mode nematic display.

i i

i i
 i i

i i

14.2. Liquid Crystal Displays (LCDs) 757

14.2.7 Transmittance of a Twisted-Nematic Cell

As mentioned above, a single pixel (cell) of a twisted nematic display can be
described with (14.10) where α = 0 and β = π /2. For a cell with thickness d, the
pitch p can then be derived from (14.9):

p = 4 d. (14.13)

This leads to the two parameters a and γ with values

2 d Δn
a= , (14.14a)
λ

 2
π 2 d Δn
γ= 1+ . (14.14b)
2 λ

Further, the s and t  axes are rotated over 90◦ with respect to the x- and y-axes. By
substitution of (14.14) into (14.11) and subsequently into (14.10), we obtain [706]
⎡ ⎧ ⎛ ⎞ ⎛ ⎞⎫ ⎤
⎨  2  2 ⎬
⎢ ⎝π 2 d Δn ⎠ π i d Δn
sin ⎝
π 2 d Δn ⎠
E0 ⎥
  ⎢e ⎩cos 2 1 + λ
+
λ 2
1+
λ ⎭ ⎥
Js ⎢ ⎥
=⎢ ⎥.

⎛ ⎞
Jt ⎢   ⎥

⎢ π π 2 d Δn ⎠2 ⎥
⎣ −e sin ⎝ 1+ E0 ⎦
2 2 λ
(14.15)
In this equation, we have
 
−2 π i n̄ d
e = exp , (14.16)
λ

where Δn is the difference between the ordinary and extraordinary indices of re-
fraction of the birefringent material and is taken from (2.141). The value n̄ is the
average of these indices of refraction:

n + n⊥
n̄ = . (14.17)
2

A parallel analyzer would be oriented along the t  -axis, leading to a normally

black mode. The irradiance Ee of the light passing through the analyzer is derived
from the time averaged value of Jt  , analogously to the expression for irradiance

i i

i i
 i i

i i

758 14. Display Technologies

0.5

Ee / 〈E02 〉T
0.4

0.3

0.2

0.1

0.0
0 2 4 6 8 10 12 14 16 18 20


Figure 14.14. The reduced irradiance Ee / E02 T plotted against a = (2 d Δn)/λ for the
normally black mode.

discussed in Section 2.4 [706]:

1
Ee = |J  |2 , (14.18a)
2 t ⎛ ⎞

2 ⎝π 2 d Δn 2 ⎠  2
sin 1+ E0 T
2 λ
1
=   , (14.18b)
2 2 d Δn 2
1+
λ

where the time average of E0 is taken. As can be seen from this equation, as
well as from Figure 14.14 which shows the reduced irradiance (effectively the
transmittance), the amount of light transmitted reaches a minima when the square
root becomes a multiple of 2. The first minimum is obtained for
 
2 d Δn 2
1+ = 2, (14.19)
λ
so that we find the relation √
3λ
d= , (14.20)
2 Δn
a result which was already anticipated in (14.2). Thus, a given wavelength, to-
gether with the material-dependent parameter Δn, determines the optimal wave-
length to block all light. As every wavelength in the visible range would require

i i

i i
 i i

i i

14.2. Liquid Crystal Displays (LCDs) 759

0.5
Ee / 〈E0 〉T

0.4

0.3

0.2

0.1

0.0
0 2 4 6 8 10 12 14 16 18 20
2dΔn/λ

Figure 14.15. The reduced irradiance Ee / E02 T plotted against a = (2 d Δn)/λ for the
normally white mode.

a somewhat different thickness of the liquid crystal cell, the normally black mode
can only be fully realized for one wavelength. As a result, this mode leaks light
which is seen as a bluish-yellowish color.
For the normally white mode, the analyzer is aligned with the s -axis, so that
the irradiance becomes
1
Ee = |J  |2 , (14.21a)
2 s ⎛ ⎞
   
π 2 d Δn 2 ⎠ 2 d Δn 2
cos2 ⎝ 1+ +
2 λ λ  2
1
=  2 E0 T . (14.21b)
2 2 d Δn
1+
λ

Here, maxima occur for the same values of d for which minima occur in the nor-
mally black mode. As a result, the √ thinnest cells for which maximal transmittance
happens
 2 have a thickness of d = 3 λ /2 Δn. A plot of the reduced irradiance (i.e.,
Ee / E0 T ) in the normally white mode is shown in Figure 14.15.
In both the normally black and the normally white mode, the irradiance trans-
mitted through the cell depends on wavelength, as can be determined from (14.21).
This results in color shifts that depend on the selected gray level. However, when
the maximum voltage is applied in the normally white mode, the cell becomes

i i

i i
 i i

i i

760 14. Display Technologies

Figure 14.16. Shown here is a Toshiba 32-inch LCD television and a 12-inch Apple iBook
with LCD screen, photographed at different angles. The iBook is several years older than
the television and has a pronounced angular dependence. The television is virtually free of
angular artifacts.

black independent of wavelength. As the black state of the normally white mode
does not have a colored tinge, it is the preferred mode [706].
The equations shown in this section are valid for light incident at right an-
gles. For obliquely incident light, the Jones matrix method can be refined and
extended [168, 398, 684, 706, 1274, 1284], showing that dependent on the type
of cell, the transmittance shows a more or less pronounced fall-off with viewing
angle, as shown in Figure 14.16.
Moreover, these equations assume that the cell is in the off-state, i.e., no volt-
age is applied. Under this assumption, the twist angle varies linearly along the
z-axis, and the tilt angle can be assumed to be constant (see Figure 14.17). When
a voltage is applied, the total twist angle remains the same, due to the anchoring
effect imposed by the rubbed substrates. However, the distribution of twist and

angle
β=π/2

Twist angle

Tilt angle
α
0.0
z=0 z=d

Figure 14.17. Twist and tilt angles in the off-state of a twisted nematic cell (after [1270]).

i i

i i
 i i

i i

14.2. Liquid Crystal Displays (LCDs) 761

angle
β=π/2

Twist angle

Tilt angle

α
0.0
z=0 z=d

Figure 14.18. Twist and tilt angles in the on-state of a twisted nematic cell (after [1270]).

tilt angles along the z-axis will vary non-linearly, as shown in Figure 14.18. A
computational model of these phenomena is discussed by Yamauchi [1270].

14.2.8 Multi-Domain and In-Plane Switching Cells

Improvements in viewing angle for different LCD types can be achieved with
various approaches, including the use of lenslet arrays [547], in-plane switching
(which keeps the tilt angle constant throughout the liquid crystal medium) [562],
vertical alignment [481], retardatio-film-compensated twisted nematic LCDs,
optically compensated bending (OCB), and multi-domain twisted nematic LCDs
[375]. In this section, we briefly discuss multi-domain and in-plane switching
solutions.
In the preceding sections we have omitted the analysis for off-axis viewing of
twisted nematic cells. However, it can be shown that under such circumstances
a significant fall-off in transmittance can be observed. Further, contrast reversals
may occur. One of the reasons is that the tilt angle is non-zero, meaning that
certain viewing angles coincide with the direction of the director, so that trans-
mittance is governed by a different index of refraction than other viewing angles.
Twisted nematic displays are normally oriented such that viewing the display from
below yields the most artifacts, including contrast inversion.
To minimize such viewing angle-dependent fall-off, it is therefore desirable
to construct displays such that the viewing angle and the director are either never
aligned, or such that the angle between view vector and director has a uniform
distribution over the surface area of the pixel.

i i

i i
 i i

i i

762 14. Display Technologies

Figure 14.19. An improvement in viewing angle is realized with multi-domain cells.

The latter approach is taken in multi-domain techniques, where each pixel is

subdivided into preferably four domains. A pair of sub-domains has identical tilt
angles, whereas the other pair has opposite tilt angles. Further, the tilt angle pro-
gresses pairwise through the cell in opposite directions. This scheme is visualized
in Figure 14.19. It is also possible to construct multi-domain pixels with only two
domains, or with more than four domains. However, fewer than four domains
results in lower efficiency, whereas more than four domains does not lead to a
noticeable improvement in viewing angle [706].
Finally, in-plane switching techniques are constructed with both electrodes on
one side of the cell, rather than on opposite sides, as shown in Figure 14.20. The

Analyzer

Polarizer

Electrodes
Off state On state

Light

Figure 14.20. In-plane switching cells are constructed with both electrodes on one side of
the cell, so that the tilt angle remains constant, irrespective of the applied voltage.

i i

i i
 i i

i i

14.2. Liquid Crystal Displays (LCDs) 763

advantage of this configuration is that the tilt angle is constant and close to zero,
independent of the voltage applied to the electrodes. As a result, the transmittance
T varies little with viewing angle. It is given by

sin (2β ) sin2 (γ )

T= , (14.22)
2
where β is the twist angle, and γ is given by

2 π Δn d
γ= . (14.23)
λ
The maximum transmittance of 0.5 is reached when β = π /4 and γ = π /2,
whereas for β = 0 the transmittance is 0.
While in-plane switching solves the attenuation of transmittance with viewing
angle, there are two problems associated with this technique. First, the torque
required to rotate the liquid crystal molecules is larger than with other techniques,
and this leads to a longer response time. Further, the placement of the electrodes
reduces the aperture of the system, resulting in less light being transmitted.

14.2.9 Active Matrix Addressing

To modulate the transmittance of a single cell, we have seen in Figure 14.11 that
a voltage must be applied across the liquid crystal layer. In a twisted nematic
LCD display, a large array of pixels should be individually addressed to form an
image. Although different addressing schemes exist, we focus on active matrix
addressing.
A pair of pixels with associated addressing lines is shown in Figure 14.21.
The addressing circuitry is fabricated on the glass substrate nearest the back-light

Vvideo = Vd
Column m Column m+1

Vg Row n

VDS TFT
VLC CLC CS LC pixel Storage
capacitor

Row n+1

Figure 14.21. Two pixels of an active matrix display.

i i

i i
 i i

i i

764 14. Display Technologies

Transmittance/
CRT

Emission
LCD

Tp Tr Tf Time

Figure 14.22. A comparison of LCD and CRT addressing times (after [84]). Note that
the emission time for a CRT dot is Tp ≈ 50 μ s, whereas an LCD pixel transmits light for
T f ≈ 16 ms.

using thin film transistor (TFT) technology. Pixels are addressed one row at a time
by raising the control voltage Vg for the row (n in Figure 14.21). The transistors
for the pixels in this row are now conductive.
The image signal voltage Vd applied to each of the columns then charges the
liquid crystal cells CLC for the pixels in row n only. The auxiliary capacitor CS
is also charged, which helps keep the pixel in its state during the time when all
other rows are addressed. As a consequence of charge storage in the capacitance
of each LCD pixel, the pixel modulates light until it is next addressed (unlike a
CRT, which emits light at each phosphor dot only for a brief instant).
The time to address a full frame, i.e. N rows, is T f . As a result, the time avail-
able for addressing one row of pixels is Tr = T f /N. A TFT-addressed twisted ne-
matic display reaches the desired voltage over the liquid crystal cell, and therefore
the desired transmittance, more slowly than in the per-pixel addressing of a CRT
display. The relative timing of the two technologies is shown in Figure 14.22.
In Figure 14.11 the transmittance of the liquid crystal cell against driving
voltage is shown. The relationship between the two is more or less linear over
a voltage swing of around 2 V. To construct a display that can handle 256 gray
levels, the voltage difference between two consecutive gray levels is around 7.8
mV [706]. Thus, for a stable, high quality display, the accuracy of the driving
voltages must be high.

14.2.10 Color Filters for LCD displays

The basic operation of an LCD display is by modulation of the amount of light that
is transmitted through each of the pixels/cells that make up the display2 . With-
out the use of color filters, the display would have a spectral emission mostly
determined by the back-light.
2 We limit our discussion to transmissive displays. However, it should be noted that LCD displays

can also be constructed reflective or a combination of transmissive and reflective.

i i

i i
 i i

i i

14.2. Liquid Crystal Displays (LCDs) 765

Protective layer Color filters

Black matrix
Substrate
ITO

Figure 14.23. Cross-section of a color filter layer for LCD displays (after [706]).

To construct a color display, a layer of additive color filters is added to the

substrate. A black matrix may be added as well, to help minimize light leakage
between pixels, thereby improving the reproduction of dark tones. A cross-section
of the color filter layer is shown in Figure 14.23.

Diagonal Triangular Striped

Figure 14.24. Spatial layout schemes for color filters used in LCD display devices
(after [706]).

The spatial layout of pixels depends on the application. For moving images,
a triangular or diagonal arrangement is preferred, whereas for still images (i.e.,
computer displays), a striped configuration is better. These layouts are shown in
Figure 14.24. A photograph of an LCD display is presented in Figure 14.25 and
shows a striped pattern.

Figure 14.25. A close-up photograph of a Toshiba LCD television.

i i

i i
 i i

i i

766 14. Display Technologies

T 1.0

R
0.8 G
B
0.6

0.4

0.2

0.0
400 500 600 700

Figure 14.26. Transmission of a set of representative LCD color filters (adapted from a
plot provided by Louis Silverstein, VCD Sciences, Inc., Scottsdale, AZ).

Additive color filters can be fabricated in several different ways. Regardless,

the colorimetric properties of the combination of the back-light and the filters
are chosen to conform to the display standard (e.g., sRGB). An example of a set
of spectral transmissions are shown in Figure 14.26. Further, the loss of light
through the filter should be minimized. The transmittance should be uniform
across the surface of the filter. Further requirements for color filters are related
to the manufacturing process, and include for instance thermal stability, chemical
resistance, reliability and light stability [855].
Color filters can be manufactured by dyeing, pigment dispersion and print-
ing. Pigment dispersion is now most commonly used [706]. The resulting color
gamuts for typical instances of these techniques are shown in Figure 14.27.

y 0.8
Dyeing
Printing
Pigment dispersion
0.6

0.4

0.2

0.0
0.0 0.2 0.4 0.6 0.8
x

Figure 14.27. Color gamuts for different color filter technologies (after [855]).

i i

i i
 i i

i i

14.3. Transflective Liquid Crystal Displays 767

14.2.11 Electronically Controlled Birefringence

A different approach to generating color requires a mixed-mode twisted nematic
display. In such a display, there is control over both the tilt angle α and the twist
angle β , whereas in a conventional twisted nematic display, applying a voltage
only changes the twist angle.
If the tilt angle is changed, then light incident upon the cell hits the liquid
crystal molecules under increasing angles. The apparent index of refraction is
then no longer n , but is dependent on the tilt angle and, indirectly, on the voltage
VLC applied to the cell. At the same time, this means that Δn = n − n⊥ depends
on the applied voltage. In a mixed mode twisted nematic display the reduced
irradiance is given by [706]
  
Ee π d Δn
 2 = cos β 1 +
2
. (14.24)
E0 T λ
√
The first maximum of this function occurs for π d Δn/λ = 3. As a result, the
maximum energy transmitted by the cell occurs for the wavelength λ ,

π d Δn
λ= √ . (14.25)
3
By controlling the tilt angle, and thereby Δn, the wavelength of maximum trans-
mittance can be controlled. This phenomenon is called electronically controlled
birefringence (ECB) and enables the formation of color by shifting the peak wave-
length of light transmitted through the cell.

14.3 Transflective Liquid Crystal Displays

A variation of LCD displays can be constructed such that they can be viewed
by applying a back-light as well as in reflective mode, i.e., lit from the front by
whatever ambient light is available. This means that the back-light is only used
in dark environments, which leads to a very energy efficient display that finds
application for instance in mobile devices.
In reflective mode, light incident upon the front of the display passes through
the liquid crystal layer, is reflected by a diffuse reflector, and then passes through
the liquid crystal layer once more before exiting the display. In transmissive
mode, light from the back-light passes through the display as discussed in the
preceding section. Each pixel is subdivided into two areas, with one serving trans-
missive mode and the other reflective mode.

i i

i i
 i i

i i

768 14. Display Technologies

In dual cell gap types, the gap between the front and back substrates is halved
for the pixel fragment serving reflective mode to account for the fact that light
passes through this section twice. The areas of the pixel devoted to the transmis-
sive and reflective modes are given as a ratio, which can be varied dependent on
the application. It typically ranges from 1:4 to 4:1 [1294]. The response time in
reflective mode will be much faster than for transmissive mode for a dual cell gap
type transflective display, as the time to set a pixel depends on the square of the
thickness of the liquid crystal layer (i.e., the cell gap). Typically, the response
time is four times slower for the transmissive mode of operation.
Single cell gap transflective liquid crystal displays have liquid crystal layers
with the same thickness for both transmissive and reflective layers. This makes
manufacturing more straightforward and creates a uniform response time for both
modes of operation. However, the light efficiency is now lower.

14.4 Plasma Display Panels (PDPs)

Plasma display panels [99, 1280] are based upon gas discharge (see also Sec-
tion 3.5.3). Gas discharge is a sequence of reactions that occur when a voltage is
applied through a tube or a cell that contains certain chemically inert gases. As
a result of this potential difference, electrons accelerate, and collide with neutral
gas atoms. These collisions may ionize the gas, i.e., remove electrons from it, if
electrons impart sufficient energy to the gas atoms.
The excited atom remains excited for a short duration (around 10−8 s), after
which it immediately returns to the ground state. This causes the emission of
a photon with a wavelength depending on the energy difference between the two
orbitals, which, in turn, depends on the mixture of the inert gas used in the reaction
(see Section 3.5.3).
Early monochrome plasma displays used a mixture composed of neon to
which a small percentage of argon (about 0.1%) was added [1120]. This mix-
ture caused the display to emit an orange color. One alternative to obtain other
colors is to change the composition of the gas mixture, although other inert gas
mixtures tend to be luminously less efficient than those using neon as the main
component [17].
Modern plasma display panels use UV-emitting plasma and introduce phos-
phors into the display. In this case, a gas mixture composed of helium and xenon
(≈ 5%–10%) is used [1066]. When excited by electrons, this mixture emits ultra-
violet light with a sharp peak at 147 nm and a broader emission spectrum centered
around 172 nm. This light, in turn, excites the phosphors causing glow in the vis-

i i

i i
 i i

i i

14.4. Plasma Display Panels (PDPs) 769

Back glass plate

Barrier MgO layer

rib

Address
electrode

Phosphor
Glass
Barrier
panel
rib

The cavity is filled with Sustain

a He/Xe gas mixture. electrode (ITO)

Figure 14.28. An exploded view of a red pixel of a plasma display panel.

ible region of the spectrum. If three phosphors with peaks around the red, blue,
and green parts of the spectrum are used, their combination yields a sufficiently
large color gamut.
Figure 14.28 shows the structure of a pixel in a plasma display panel. Here,
the phosphor-coated cavities contain the gas mixture. The barrier ribs minimize
light leakage between pixels; differently shaped barriers have been developed for
improved efficiency [999, 1276]. The sustain electrodes are made from trans-
parent indium tin oxide (ITO) to enable as much light to escape the cavity as
possible [276]. The phosphors are at the back side of the display. The MgO layer
protects the sustain electrodes and the dielectric layer into which these electrodes
are embedded [1160]. Due to the path taken by the light, this configuration is
known as the reflected phosphor geometry [1159].
The screen of a plasma display is composed of a large array of pixels laid out
in a matrix fashion. Each pixel is sandwiched between row and column electrodes
and can be turned on by applying a voltage to the corresponding row and column.
The display can be driven by using alternating or direct current [1120]. Pixel pitch
is typically around 800–1000 μ m2 . Plasma displays are predominantly built in
larger frames (up to a 102-inch diagonal with 1920 × 1080 pixels. PDPs typically
have peak luminance values of 300 cd/m2 with contrast ratios of 300:1, although
displays having a maximum luminance of 1000 cd/m2 and a contrast ratio of
2000:1 have been demonstrated [276]).
Plasma displays are typically power-limited: not all cells can produce maxi-
mum luminance simultaneously. Typical video material has an average luminance
of 20%; this average luminance can be distributed arbitrarily across the panel.

i i

i i
 i i

i i

770 14. Display Technologies

However, beyond some average luminance, signal processing circuits attenuate

the RGB drive signals so as to limit average luminance3 .
If discharge is taking place in a cell, dropping the sustain voltage will extin-
guish the discharge. The cell will remain off until addressed and ignited. Once
ignited, it will remain in the discharge state until the sustain voltage is dropped
again. Each pixel in a plasma display is thereby electrically bi-stable, having in
effect one bit of memory.
The intensity modulation of pixels in plasma displays is achieved by pulse-
width modulation (PWM). The number of pulses per second is far in excess of the
temporal resolution of the human visual system. The fraction of time that the pix-
els are on therefore determines which average light intensity is reproduced [258].
The phosphors used in a plasma display panel are comparable to those used in
CRTs, so the reproduction of color in PDPs approximates that of CRTs.
The main disadvantage of plasma displays is their low energy efficiency,
which is less than 1% [276]. The reasons for this low efficiency are the low
conversion from power to ultra-violet radiation (around 10% efficiency), as well
as an inefficient conversion from ultra-violet radiation to visible light by means
of phosphorescence (around 20%–40% efficiency). As a result, the luminous ef-
ficacy is less than 2 lm/W. The display industry is currently working to improve
this number to around 5 lm/W which is what CRTs achieve [276].
One of the disadvantages of PDP technology, as with all phosphor-based dis-
plays, is that the phosphors themselves degrade with use. If, for instance, a logo
or a menu-bar is displayed for prolonged periods of time, then this results in an
after-image that can not be removed. This phenomenon is called burn-in.

14.5 Light-Emitting Diode (LED) Displays

A light-emitting diode (LED) is based upon a semiconductor containing a p-n
junction as shown in Figure 14.29 (see also Section 3.5.7). The p-type region
has an excess amount of holes and the n-type has an excess amount of electrons.
Excess electrons and holes are introduced into the semiconductor by doping it
with other atoms that readily give out or accept electrons. For instance, if the
semiconductor is made of silicon (Si), which is from the IVA group of the periodic
table, doping it with a IIIA atom makes it a p-type (acceptor) and a VA atom
makes it an n-type (donor).
Under forward bias, electrons and holes diffuse toward the junction where
they may recombine. The recombination may be radiative (light is emitted), or
3 When operated in the luminance-limiting regime, PDPs behave similarly to RGB+W DLP
R

projectors.

i i

i i
 i i

i i

14.5. Light-Emitting Diode (LED) Displays 771

Light

p-type n-type
I
V
+ -

Figure 14.29. An LED is a simple semiconductor containing a p-n junction that is

driven with a forward bias. Light is emitted from the junction through electron-hole
recombination.4

non-radiative (heat is produced) (see Figure 14.30). If radiative recombination

occurs, the wavelength of the emitted light depends on the band-gap energy be-
tween the conductance and valence bands of the semiconductor material and the
impurities used. Therefore, by using different materials, different colors of light
can be obtained.
The operation of an LED involves electroluminescence, whereby a material
emits light in response to an electrical current. In this sense, an LED is the oppo-
site of a solar cell which produces electrical current in response to light.
The most common types of semiconductors used in LEDs are gallium arsenide
(GaAs) or its related ternary compounds such as GaAsP and GaAlAs. Blue LEDs
are typically based upon gallium nitride (GaN) [918, 1120].

- - -
-
Radiative - - -
recombination Conduction band
Eg
λ heat
+ + + Non-radiative
+ recombination
p-type region
+ + +
Valence band
n-type region
hc
λ= where Eg: Band-gap energy
Eg - : electron
h: Planck’s constant
c: Speed of light + : hole
λ: Wavelength

Figure 14.30. An electron and a hole may recombine radiatively or non-radiatively. Ra-
diative recombination results in light whose wavelength depends on the band-gap energy
of the semiconductor. Non-radiative recombination produces heat.

4 In this diagram, the electrons flow form the negative post of the power supply to the positive post,

which is opposite to the direction of conventional current, which is drawn here.

i i

i i
 i i

i i

772 14. Display Technologies

LEDs are usually used in the construction of very large scale displays, such
as interactive advertisement boards. They are also used as back-lights for liquid
crystal displays. For instance, Apple’s iPod uses an LCD display with LED back-
light. Another use of LEDs to back-light a display is afforded by HDR displays,
as discussed in Section 14.16.

14.6 Organic Light-Emitting Diode Displays

An organic light-emitting diode (OLED) display is a relatively recent but highly
promising display technology, which produces visible light when an electrical
current is passed through it. Light emission from an OLED is similar to inorganic
LEDs in that photons are produced when a hole-electron pair radiatively recom-
bines. However, in an OLED radiative recombination takes places in an organic
layer rather than in an inorganic p-n junction.5
The interest in OLEDs dates back to the mid-1960s [451, 452], although the
breakthrough in OLED technology came after Tang and van Slyke [1117], rec-
ognized that the efficiency of an OLED can be greatly increased by using a bi-
layered structure where each layer is responsible for either electron or hole trans-
port and injection (Figure 14.31). A multi-layer structure allows each layer to be
optimized separately for different tasks.

Electron Electron
injection transport Emission
Cathode layer layer layer

Hole Hole Transparent Transparent

transport injection anode substrate
layer layer

Figure 14.31. An OLED is composed of a stack of layers sandwiched between a cathode

and an anode. Each layer can be optimized separately.

5 Organic materials are those primarily made of compounds of carbon, hydrogen, oxygen, and

nitrogen.

i i

i i
 i i

i i

14.6. Organic Light-Emitting Diode Displays 773

In an OLED, electrons and holes injected from the cathode and anode diffuse
toward the emissive layer under an electrical field, where they form electron-
hole pairs called excitons [1275]. An exciton may radiatively recombine resulting
in emission of a photon whose wavelength depends on the band energy of the
organic material in the emissive layer [1168]. To increase the emission efficiency,
the emissive layer may be doped with other organic materials [1040].
OLEDs can be made of small organic molecules or larger conjugated poly-
mers [143]. In the former type, organic layers are prepared by depositing each
layer on a transparent substrate by thermal evaporation. As polymers are too large
to evaporate and may disintegrate upon excessive heat, polymer-based OLEDs
(PLEDs) are usually created by spin-coating the polymer liquid onto indium tin
oxide coated substrates which are solidified by heating [328]. The fabrication
process of PLEDs is therefore simpler than that for OLEDs, although the manu-
facturing of the top electrodes for PLEDs is more complicated. In addition, the
production of PLEDs is more wasteful of materials [924].
OLED displays are usually fabricated on a glass substrate, but they can also
be produced on flexible plastic [400, 1103]. In addition, OLED displays can be
transparent when in the off-state, making them useful for augmented reality ap-
plications and heads-up displays [399]. In addition, transparency is necessary in

White
Organic Transparent organic Colored Transparent
Cathodes layers Anode substrate Cathodes layer Anode filters substrate

Blue
organic Colored Transparent Organic Organic Organic Transparent
Cathodes layer Anode filters substrate Cathode layer layer layer Anode substrate

Figure 14.32. Various methods to obtain a range of colors using OLEDs.

i i

i i
 i i

i i

774 14. Display Technologies

as shown at the bottom of Figure 14.32 [401]. In either case, the anodes will also
have to be transparent [740].
A full-color OLED display device can be realized using a matrix layout simi-
lar to that of an LCD display. Different from an LCD display, OLEDs themselves
emit light and, therefore, they do not require a back-light. An OLED display
can use passive or active matrix addressing. In a passive configuration, the in-
tensity emitted for each pixel needs to be very high and depends on the number
of scanlines in the display—more scanlines means that each flash will have to be
correspondingly brighter due to the fact that only one line is active at any given
time. If there are N scanlines, then the intensity of each pixel will be N times
higher than the average perceived intensity [1026]. This places an upper limit on
the size of the display that can be addressed in passive matrix addressing.
Full-color can be produced by subdividing each pixel into three OLEDs emit-
ting red, green, and blue colors. Alternatively, an OLED emitting white light can
be filtered by three color filters, or an OLED emitting blue or ultraviolet light may
be used to illuminate three fluorescent materials which re-emit light with the de-
sired spectral composition [328]. One final method is to stack up OLEDs on top
of each other by separating them with transparent films. This allows for a higher
spatial resolution in the same display area and is, therefore, a favorable technique.
Different color reproduction schemes are illustrated in Figure 14.32.
Aging effects in OLED displays can be classified as either life-time aging
or differential aging. Currently, the organic molecules used in the fabrication of
OLEDs tend to degenerate over time, which means that OLED technology is not
yet suitable for use in displays that need to last for many years. Useful life is
normally measured as the length of time that elapses before the display becomes
half as bright as it was when new. The life span of an OLED micro-display also
depends on the luminance value at which it is operated. The product life cycle
for displays hovers around four years, and OLED displays currently fulfill the
requirement of lasting for this amount of time under normal conditions.
Differential aging results from the fact that the remaining life time of each
pixel depends on how it has been switched on. If pixels have had different lengths
of on time, then their brightness may be affected differently. The threshold of ac-
ceptable pixel differences across the display is typically set to 2–3% for graphics
applications and 5–8% for video applications [924].
OLEDs offer significant advantages that are likely to result in their widespread
use in the near future. They consume very little power, are luminously efficient
(in the order of 20 lm/W [894, 1029]), exhibit wide viewing angles, and have fast
response times (in the order of a few microseconds). OLEDs are also very thin
and lightweight, and can be deposited on flexible surfaces. This opens up the

i i

i i
 i i

i i

14.7. Field Emission Displays 775

eventual possibility for display devices that can be rolled-up or stuck onto curved
surfaces.

14.7 Field Emission Displays

Field emission displays (FEDs) work on the same principle as cathode ray tubes
in that a cathode generates electrons in a vacuum [1114]. The electrons are ac-
celerated until they hit a phosphorescent layer which, as a result, produces visible
light. Unlike a CRT, however, a field emission display incorporates one or more
cathodes per pixel which are constructed of a metal tip or more recently of carbon
nanotubes [346, 782] or diamond-like carbon [535, 682]. Different from a CRT,
electrons are generated by field emission, for which very thin, sharply pointed
electrodes are required.
Field emission displays can be made with an extremely shallow depth (around
2 mm). The power consumption is projected to be lower than that of LCDs and
CRTs [535]. The behavior of a FED is otherwise similar to a CRT, including its
advantage of a wide viewing angle (around 170◦ ).
The cathodes of developmental FED devices were made of molybdenum,
which had several disadvantages. The most important of these is that the en-
ergy required to produce light may create enough heat to melt the cathodes. As a
result, the pointed tips would be deformed, reducing the field effect and thereby
the device’s efficiency. In addition the cathodes may react with the residual gases
within the vacuum, which would further reduce the efficiency.
As a result, the traditional FED devices have never become a commercial
success. However, recent developments in nanotechnology has enabled the de-

Insulator
Gate electrode
Spacer
Substrate

Nanotubes

Phosphor
Anode
Glass panel

Figure 14.33. The components of a FED pixel.

i i

i i
 i i

i i

776 14. Display Technologies

velopment of nanotubes, which are made of carbon atoms configured in covalent

bonds (see Section 3.3).
In a FED no beam deflectors are required to steer the electron beam in scan-
line order across the display. Each pixel produces its own electron beam. A
diagram showing the components of a single pixel is shown in Figure 14.33. The
chemically stable nature of these materials has solved the problem of overheating,
melting, and reacting with residual gases. As a result, there exists renewed inter-
est in FED devices as a potential technology that may be incorporated in display
devices.

14.8 Surface-Conduction Electron-Emitter Displays

Surface-conduction electron-emitter display (SED) technology, in development
by Canon, is based on the same principle as the CRT and the FED: photons are
emitted when phosphors are bombarded by electrons. However, different from a
typical CRT and similar to FEDs, each red, green, and blue component of each
SED pixel contains its own miniature electron gun.
Reserving a separate electron gun per pixel component obviates the need to
have a beam deflection system typical of standard CRTs. SEDs can therefore be
constructed as flat panels. However, the construction of each pixel in a SED is
different from a FED as electrons are produced by a tunneling mechanism. As a
consequence, device depths of a few centimeters can be achieved [276].
Light emission from an SED pixel is illustrated in Figure 14.34. When an
electrical voltage of 16–18 volts is applied at V f , electrons jump between elec-
tron emitters owing to the tunneling effect (i.e., passage of electrons through an
insulator—in this case vacuum) [770]. The gap has dimensions on the order of

Light
Glass substrate
Phosphors
Vs

Vacuum Electrons

Electron tunneling
Glass substrate
Vf
0 0

Figure 14.34. The structure of an SED pixel. The tunneling effect and the attraction of
electrons toward the phosphors are depicted.

i i

i i
 i i

i i

14.9. Microcavity Plasma Devices 777

several nanometers. Some of the freed electrons are accelerated toward the phos-
phors as a result of the large potential difference (≈10 kV) applied at Vs . When
these electrons strike the phosphor molecules, photons are emitted due to cathodo-
luminescence. The efficacy of current prototypes reaches that of conventional
CRT displays and is of the order of 5 lm/W.
Since phosphors produce light into a wide angle, SED image quality does not
degrade at wide viewing angles. Since the phosphor technology has been success-
fully used for accurate color reproduction in standard CRTs, the SED technology
inherits this desired property as well. Finally, owing to its slim and ergonomic de-
sign, the SED technology may become an interesting competitor in the flat panel
display market in the near future. Displays with a diagonal of 36 inches have been
announced [845]. However, there remain significant challenges to high-volume
manufacturing.

14.9 Microcavity Plasma Devices

Plasma devices are efficient for large displays, but not so for smaller displays.
This is largely due to limitations related to the minimum size of each pixel. How-
ever, a recent development borrows manufacturing techniques from VLSI and
microelectromechanical systems to produce microplasma devices. These devices
consist of regularly spaced microcavities filled with a helium/xenon mixture and
surrounded by electrodes. They therefore operate on the same principle as con-
ventional plasma display devices, albeit at a much smaller scale due to the new
fabrication techniques employed.
Microcavity plasma devices (or microplasma devices for short) are still at an
early stage of development, but they promise very high packing densities (of more
than 104 pixels per square centimeter), peak luminances of over 3000 cd/m2 , and
luminous efficacies of more than 6 lm/W [276].

14.10 Interferometric Modulator (IMOD) Displays

IMOD is an emerging reflective display technology that utilizes interference to
reproduce color. As discussed in Section 2.7.1, interference causes certain wave-
lengths to cancel out if they are out of phase (i.e., destructive interference) and
other wavelengths to add if they are in phase (i.e., constructive interference). An
IMOD display utilizes both effects to reproduce either a black or a colored pixel.
Color generation from an IMOD display resembles the iridescence effect on some
butterflies’ wings—interference at the scale of visible wavelengths.

i i

i i
 i i

i i

778 14. Display Technologies

Ambient Colored Ultraviolet

light light light

Transparent
substrate

Thin-film
mirror

On-state IMOD cell Off-state IMOD cell

Figure 14.35. On and off states of an IMOD cell.

An IMOD pixel consists of a transparent substrate on top of a mechanically

deformable thin-film mirror; these are separated by a small cavity [324, 1195].
The height of this cavity is in the order of a few hundred nanometers to allow
interference of light in the visible range. Ambient light that hits the transparent
surface is both reflected off the surface and transmitted toward the thin film mir-
ror. The portion of the light that is transmitted toward the mirror goes through sev-
eral inter-reflections between the mirror and the substrate. At each inter-reflection
some portion of light exits the cavity and interferes with the light that was initially
reflected off the surface. The height of the cavity determines which wavelengths
interfere destructively, hence cancel each other out, and which wavelengths inter-
fere constructively, according to the principle shown in Figure 2.31. When the
cavity is collapsed by applying a voltage, constructive interference occurs only
in the ultraviolet range, which is outside the visible spectrum, and therefore re-
sults in a black cell (Figure 14.35). By using cavities of different heights various
primary colors can be obtained, as illustrated in Figure 14.36.

Ambient Red Ambient Green Ambient Blue

light light light light light light

hb
hg
hr

hr > hg > hg

Figure 14.36. Three primary colors can be obtained by constructing cavities of different
heights.

i i

i i
 i i

i i

14.11. Projection Displays 779

To construct a full color IMOD pixel, three types of subpixels are used with
each type having a different gap size. To reproduce various shades of color, each
subpixel type is further divided into sub-subpixels, or subpixel elements. The de-
sired intensity for each primary color can then be obtained by spatial or temporal
dithering.
The chief advantages of IMOD displays over other flat panels include a sig-
nificantly reduced power consumption and the ability to reproduce bright colors
under direct sunlight. These factors may render IMOD technology an attractive
alternative particularly for hand-held applications. In dark environments, supple-
mental illumination is provided by a low power front light. At the time of writing,
IMOD displays have a reflectivity of around 50% and a contrast ratio of 8 : 1.
This compares favorably with typical newspaper which has approximately 60%
reflectivity and a 4 : 1 contrast ratio [527].

14.11 Projection Displays

In projection displays, the final image is not displayed on the light modulator itself
but is projected on a screen internal or external to the end-user display device.
This enables a large image to be produced by a relatively small light modulating
device, often leading to low manufacturing costs.
Projection displays may be distinguished as rear and front projection displays.
A rear projector uses a transmissive screen that must have dispersive character-
istics in order to scatter the incident light into a wide viewing angle. A rear
projection display system is typically a single unit comprised of the projector, the
screen, and an optical system in between that ensures a sharp image on the screen.
In a front projection system, also called beamer, the projector is placed in
front of a diffusely reflective screen which scatters light in a wide viewing an-
gle. The distance between the projector and the screen is usually much larger in
comparison to a rear projection system. In a front projection system, the coupling
between the screen and the projector is minimal, that is, the same projector can be
used with different screens placed at different distances within the effective focal
length of the projector.
Several possible types of lamp are available for use in projection displays,
including metal halide and ultra-high performance (UHP) varieties. Metal-halide
additives improve the color properties of gas-discharge lamps, but at the same
time reduce the luminance output, making them less ideal for projection displays.
Ultra-high performance (UHP) lamps are nowadays the norm, due to their high
light output combined with a small physical size [249]. This has enabled the size
of projection systems to become significantly smaller.

i i

i i
 i i

i i

780 14. Display Technologies

0.12

Power
120 bar
0.10 290 bar

0.08

0.06

0.04

0.02

0.00
400 450 500 550 600 650 700 750

Figure 14.37. The emission spectra of an UHP lamp at different pressures (after [249]).

An UHP lamp relies on mercury discharge only: no rare earth gases are used.
This gives high brightness and high luminous efficiency [249]. The emission
spectra of UHP lamps depends on the pressure realized within the bulb. At higher
pressures, the spectrum becomes flatter, giving a better color-balanced output with
a higher color rendering index. The emission spectra of an UHP lamp at different
pressures is plotted in Figure 14.37.

14.11.1 Pulse-Width Modulation

Some of the projection display types discussed below can switch each pixel into
either an on-state or an off-state. This includes DLP projectors and LCoS devices
(Section 14.13 and 14.14). To create intermediate levels, the switching between
on- and off-states is achieved at rates much higher than the frame rate. Thus, to
achieve gradation, the bit-planes of a pixel value are considered in turn. The least
significant bit is loaded first and illuminated for a brief period of time, namely a
fraction of 1/(2n − 1) of the frame time, assuming that n is the number of bits per
color. Then the second bit is loaded and illuminated for twice the amount of time.
This is repeated until all bit-planes have been processed. The fraction of a frame
ti that the ith bit is processed is therefore

2i
ti = , 0 ≤ i < n, (14.26)
2n − 1

whereby each ti constitutes the duration of one of the n sub-frames. This approach
is feasible due to the very short switching times that are possible, combined with

i i

i i
 i i

i i

14.12. Liquid Crystal Display (LCD) Projectors 781

the comparatively slow response of the human visual system which effectively
integrates the flashes of light over periods exceeding the frame time.
In the following sections, an overview of the most commonly used projector
technologies is given.

14.12 Liquid Crystal Display (LCD) Projectors

An LCD projector contains an ultra-bright lamp which emits white light. The
wide spectrum is separated into three relatively narrow bands with the peak of
each band in the red, green, and blue regions of the spectrum. Separation is
accomplished by passing the light through multiple dichroic mirrors (see Sec-
tion 3.7), each of which transmits a certain band of wavelengths and reflects the
remainder (Figure 14.38).

Mirror Mirror
Dichroic beam
Blue light
combiner
Light Hot
source mirror
White Mirror LCs
light Dichroic
prism Red light Color
image
IR
Yellow light Green light Projection
lens
Mirror Mirror
Dichroic
prism

Figure 14.38. The main internal elements of an LCD projector.

The red, green, and blue beams are then directed toward individual liquid
crystal panels, which modulate their intensity according to the pixel values of
the frame that is to be displayed. The modulated beams are then combined by a
prism and reflected toward a lens that focuses the image at a certain distance. The
operating principle of an LCD projector is shown in Figure 14.38.

14.13 Digital Light Processing (DLP

R
) Projectors
Digital light processingR
technology is based on light reflection from individu-
ally controllable microscopic mirrors. Fabricated on a digital micromirror device

i i

i i
 i i

i i

782 14. Display Technologies

Substrate Micro-mirror

Switch

Switch

Torsion hinge

Substrate
Switch

Figure 14.39. A simplified diagram of the components of a single micromirror.

(DMD), each mirror is hinged and can be tilted between ±10◦ to change the di-
rection of the reflected light (see Figure 14.39). When tilted to +10 degrees the
incident light is reflected toward a lens which projects the light onto a screen.
This is called the on-state. If tilted by −10 degrees, the incident light is reflected
toward a light absorber (the off-state).
The intensity of the reflected light can be controlled by alternating between
the “on” and “off” states very rapidly, up to 50, 000 alternations per second. For
each tiny mirror, the ratio of the number of on- and off-states across a frame time
determines the effective intensity of light produced for that pixel.
DLP projectors can produce color using one of three different approaches that
are characterized by the number of DMDs employed. In the first approach, using a

4-color wheel
Lout 1.4 3-color wheel
equivalent
1.2

1.0

0.8

0.6

0.4

0.2

0.0
0.0 0.2 0.4 0.6 0.8 1.0
Iin

Figure 14.40. Compared with a three-color filter wheel, the extra white improves the
luminance output (after [1096]).

i i

i i
 i i

i i

14.13. Digital Light Processing (DLP

R
) Projectors 783

DMD

Prism
Integrator rod
Projection lens
Color wheel
Arc lamp Lenses

3- and 4-color wheels

Figure 14.41. The optical components of a 1-chip DLP projector (adapted from [489]).

single DMD, a color filter wheel rotates rapidly in front of a broadband bright light
source. When the blue region of the wheel is in alignment with the light source, it
only allows blue light to reach the DMD. Synchronously, the micromirrors align
themselves according to the blue channel of the image to be displayed. These
steps are repeated rapidly for the red and green channels, resulting in a full color
output. This RGB sequential 1-DMD design is generally not able to reproduce
very high luminances, due to the absorption of energy by the red, green, and blue
color filters.
To achieve higher luminance, the color wheel is frequently fitted with an ad-
ditional white channel (see Figure 14.40), and thus this approach uses a total of
four channels (RGBW) [92]. Since one DMD is used in both RGB and RGBW de-
signs, these set-ups are also called 1-chip DLP projection systems (Figure 14.41).
Their advantages include portability and low cost.
The disadvantage of using a white channel is that the color gamut has an
unusual shape, in that the volume spanned by the red, green, and blue primaries,
is augmented with a narrow peak along the luminance axis. This means that
colors reproduced at high luminances will be desaturated. At low luminances,
saturated colors can be reproduced. The lack of accuracy of RGBW systems at
high luminance values makes this approach useful for business graphics, but less
so for color-critical applications.
A second alternative, which has a higher light efficiency and optimizes the
lifetime of the light bulb, uses a 2-chip design. Bulbs with a long lifetime may
be deficient in the red end of the spectrum. To obtain a high light efficiency, red
light is continuously directed towards one DMD, whereas the second DMD al-

i i

i i
 i i

i i

784 14. Display Technologies

2-color wheel DMD 1

DMD 2

Prism
Integrator rod
Projection lens
Color wheel
Arc lamp Lenses

Figure 14.42. The optical components of a 2-chip DLP projector (adapted from [489]).

ternates between reflecting green and blue light. This is achieved by designing
a color wheel with yellow (red + green) and magenta (red + blue) segments.
The dichroic prism spatially separates the red component from alternatively the
green and blue components. A 2-chip projection system is outlined in
Figure 14.42
In general, 1- and 2-chip DLP devices suffer from artifacts which stem from
the rotation of the color wheel and the corresponding sequential reproduction of
the three color channels. If this speed is too low, then the time for a boundary
between two neighboring colors of the filter to pass through the optical path is
relatively long. When a projection is viewed, the viewer’s eye-tracking (see Sec-
tion 4.2.1) may coincide with the movement of the chromatic boundary that is
passing across the image [252]. Dependent on the relative speeds of the rotation
of the color wheel and the eye movement itself, this may be seen as a colored
fringe. The leading and trailing edges will exhibit color fringes. These color
separation artifacts are known as the rainbow effect.
The third alternative involves the use of three chips, with each chip controlling
a different color channel (Figure 14.43). Similar to an LCD projector, a broadband
light can be separated into three components using a prism, or, alternatively, red,
green, and blue LEDs can be used to illuminate each chip. As this approach
requires three DMDs, this alternative is usually more costly. However, 3-DMD
systems do not exhibit color fringing. They are used in DLP cinema projectors.

i i

i i
 i i

i i

14.14. Liquid Crystal on Silicon (LCoS) Projectors 785

DMD 1

DMD 3
DMD 2

Prism
Integrator rod
Projection lens

Arc lamp Lenses

Figure 14.43. The optical components of a 3-chip DLP projector (adapted from [489]).

14.14 Liquid Crystal on Silicon (LCoS) Projectors

An LCoS projector is very similar to that of an LCD projector (Figure 14.44).
However, instead of using transmissive LCD panels, an LCoS projector utilizes
reflective LCoS modulators. LCoS projectors can also be configured to form color
by use of a color filter wheel [1163].

LCoS

Mirror
Blue light
Light Prism
source LCoS Prism
White
light
Red light Prism Color
image
Yellow light
Projection
Green Prism lens
Mirror
light
Dichroic
prism
LCoS

Figure 14.44. The components of an LCoS projector.

i i

i i
 i i

i i

786 14. Display Technologies

LCoS projector technology involves an LCD light modulator fabricated on a

silicon “backplane” upon which active electronic circuitry has been fabricated. It
is a micro-display device that is used under optical magnification. The modula-
tion of light is achieved using liquid crystals. The liquid crystals are sandwiched
between a glass layer and a silicon layer (as opposed to two glass layers for a
typical LCD), where the latter is coated with a highly reflective metal, typically
aluminium, which is both highly conductive electrically and highly reflective op-
tically (Figure 14.45). Thus, the metal acts simultaneously as a mirror and an
electrical cathode, and in conjunction with the anode, controls the state of the
liquid crystal layer. The device works in reflective mode. The control circuitry
that turns the liquid crystal cells on and off is embedded into the silicon layer.
This allows for a closer packing of the liquid crystal elements enabling higher
resolutions.
An LCoS chip typically does not have colored filters to achieve color repro-
duction, but, if one chip is used, it is sequentially flashed with red, green, and
blue light. This is synchronized with altering the states of the chip, so that when
illuminated with red light, the image corresponding to the red channel is loaded,
and similarly for green and blue. This is called sequential color. The advantages
of this approach are that only one third the number of pixel elements are required,
and that even at high magnification, there will be no spatial separation of the red,
green, and blue channels. It should be noted that spatial accuracy is traded for the
potential for temporal artifacts, including eye-tracking induced color fringing.
Three-chip LCoS projectors do not use field sequential color, but use dichroic
mirrors to separate a white light beam into red, green, and blue beams, which are
aimed at separate LCoS chips, before being recombined into a color signal (see
Figure 14.44).
In addition to the nematic liquid crystal technology shown in Figure 14.45,
ferroelectric techniques are under development. The advantage of the latter is
that faster switching times can be achieved (changing the state of a nematic pixel
takes in the order of milliseconds, whereas ferroelectric pixels can be switched
much faster). A second difference is that the ferroelectric technology is binary, so
that pixels are either on or off. Here, pulse-width modulation (PWM) is used to
achieve intermediate levels.
LCoS chips can be used in near-to-eye applications, such as head-mounted
displays. Here, the source of light is typically a triple of red, green, and blue
LEDs that sequentially illuminate a single LCoS chip.

i i

i i
 i i

i i

14.15. Multi-Primary Display Devices 787

Silicon
substrate

Mirror

Liquid crystals
Anode Glass panel
(Indium tin
oxide)

Figure 14.45. The components of an LCoS pixel. No polarizers are required in this
configuration, as polarization has typically already occurred in the beamsplitter.

14.15 Multi-Primary Display Devices

The range of colors reproducible by a display device is called its color gamut
(see Chapter 8), and this usually corresponds to only a fraction of colors that the
human eye can distinguish. Figure 14.46 shows the gamuts of several display
standards in a CIE 1931 chromaticity diagram. As this figure shows, the standard
gamuts cover only some part of the total range of visible colors.6
In many digital imaging applications it would be desirable to display a wider
array of colors than the Rec. 709/sRGB gamut typical of most of today’s displays.
Colors outside the display gamut can be replaced with well-chosen stand-in colors
by using gamut mapping techniques (Section 15.4), or, in professional imaging,
by skilled craftspeople who can manually execute gamut mapping. However, it is
impossible to convey the original color information. A better solution would be
to increase the gamut of the display device and avoid the gamut mapping problem
altogether.
Inspection of Figure 14.46 shows that the gamut of a display device can be
increased if it uses more highly saturated primaries, which amounts to stretching
the corners of the triangle towards the spectrum locus [497,588]. This requires the
use of almost monochromatic primaries such as lasers [478] or LEDs. Another
solution is to use color filters to narrow the emission bandwidth of an excited
phosphor. However, these filters block a significant portion of light causing a
6 Limited coverage is also observed in the dimension of luminance, which is orthogonal to the

chromaticity plane drawn in Figure 14.46. However, the aim of multi-primary display devices is to
increase the coverage of reproducible colors in the chromaticity plane, rather than extend the dynamic
range of reproducible luminance values.

i i

i i
 i i

i i

788 14. Display Technologies

y 0.9
EBU
0.8 sRGB / Rec. 709
Adobe RGB 1998
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
x

Figure 14.46. Several standard gamuts are shown in a CIE 1931 chromaticity diagram.

reduction in the maximum display brightness [588, 847]. Alternatively, a diffrac-

tion grating could be inserted into the optical path to separate white light into its
constituent components [20].

y 0.9
sRGB
0.8 6 primaries
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
x

Figure 14.47. The concept of extending gamut with multiple primaries. Each vertex of
the red polygon represents the chromaticity coordinates of one of the six primaries of an
experimental six-primary display device. This approach alleviates the gamut limitations
imposed by a three-primary display. The sRGB gamut is shown for reference.

i i

i i
 i i

i i

14.15. Multi-Primary Display Devices 789

A different approach is to use display devices that use more than three pri-
maries. Such devices are called multi-primary display devices, and they have
been actively studied in recent years [131, 981, 1269]. The advantage of using
more than three primaries is that the color gamut ceases to be restricted to a trian-
gle, but may take arbitrary polygonal shapes. This affords the flexibility necessary
to better cover the horseshoe shape of the visible color gamut (see Figure 14.47).
A multi-primary display device can be constructed by superposing the images
produced by two conventional LCD projectors7 . Each projector is fitted with three
additional interference filters placed in the optical paths of the red, green, and blue

100
Transmittance (%)

80

60

40

20

0
380 480 580 680 780
Wavelength
100
Transmittance (%)

80

60

40

20

0
380 480 580 680 780
Wavelength

Figure 14.48. Low-pass filters (shown with dashed lines) are used to select the lower
bandwidth regions of the original primaries of the first LCD projector (top). High-pass
filters (bottom) are used to select the higher bandwidth regions of the primaries of the
second projector (after [21]).

7 The technique described here is currently of an experimental nature.

i i

i i
 i i

i i

790 14. Display Technologies

Normalized intensity
1.0 P3
P2
P4
0.8
P5
P1
0.6 P6

0.4

0.2

0.0
380 480 580 680 780
Wavelength

Figure 14.49. The spectral power distributions of six primaries of an experimental multi-
primary display device (after [21]).

beams. The first projector is fitted with three low-pass interference filters, while
the second projector is fitted with three high-pass filters [21]. This set-up allows
the selection and transmittance of light of only a narrow bandwidth from the spec-
trum of each original projector primary, as shown in Figure 14.48. These selected
regions then become the primaries of the composite multi-primary display device.
The primaries for a six-primary display device are shown in Figure 14.49.
The spectrum C(λ ) of the light emitted by an N-primary display device is
given by
N
C(λ ) = ∑ α j S j (λ ) + β (λ ), (14.27)
j=1

where S j is the spectral distribution of the jth primary, α j is its weight, and β (λ )
is the spectral intensity of the background light [21] (i.e., residual light emitted
for a black image). The corresponding tristimulus values are then given by
⎡ ⎤ ⎡! ⎤ ⎡! ⎤
X N S (λ ) x̄(λ ) d λ
! j !
β (λ ) x̄(λ ) d λ
⎣Y ⎦ = ∑ α j ⎣ S j (λ ) ȳ(λ ) d λ ⎦ + ⎣ β (λ ) ȳ(λ ) d λ ⎦ , (14.28a)
! !
Z j=1 S j (λ ) z̄(λ ) d λ β (λ ) z̄(λ ) d λ
⎡ ⎤ ⎡ ⎤
N PX j Xβ
= ∑ α j ⎣ PY j ⎦ + ⎣Yβ ⎦ , (14.28b)
j=1 PZ j Zβ
where PX j , PY j , and PZ j are the tristimulus values of each primary and Xβ , Yβ , Zβ
are the tristimulus values of the background light.

i i

i i
 i i

i i

14.16. High Dynamic Range Display Devices 791

To exploit the full potential of multi-primary display devices in the absence

of a wide-gamut interchange color space, they should be driven by multi-spectral
or hyper-spectral image data. For multi-spectral images, the (visible) spectrum is
divided into a maximum of typically 16 bands. Hyper-spectral images are repre-
sented with more than 16 bands8 .
Multi- and hyper-spectral images can be produced by a spectral renderer.
Spectral images can also be created with the aid of multi-spectral cameras and
capture techniques [1269]. However, in many cases one only has access to three-
channel image data (i.e., RGB pixel values) [851]. Therefore, solutions are sought
for finding the weights α for each primary from a three-dimensional input signal.
This problem can be solved by inverting (14.28b), although several degrees
of freedom exist since an N-dimensional vector [α1 , . . . , αN ]T is computed from
a three-dimensional vector [1268]. Another method is to use an N × 3 look-up
table.

14.16 High Dynamic Range Display Devices

The ratio of the maximum and the minimum luminance values a display device
can produce is usually termed its dynamic range. Most typical display devices
(CRTs, LCDs, etc.) have a dynamic range between two and three orders of mag-
nitude.9 In contrast, most real-world environments contain a significantly higher
dynamic range, approximately 10 orders of magnitude from starlight to sunlight.
Figure 14.50 shows typical luminance values found in natural environments to-
gether with what typical monitors can reproduce. This discrepancy is further
emphasized by the capability of the eye to adapt to 10 orders of dynamic range at
different times and more than 5 orders of magnitude at a single state of adapta-
tion [312].
While the range of luminances that can be produced using phosphor-based
display devices is high, some of this range is at low levels where the human visual
system is unable to detect contrasts [923]. Thus, the effective dynamic range is
limited. As for LCDs, the liquid crystal cells still transmit light even in their off
state, increasing the black level. This limits their effective dynamic range.
High dynamic range imaging is about to become main stream, as it is adopted
by key industries, including film and games [953]. Many algorithms have been
developed to compress the luminance of HDR images, so that they may be dis-
played on typical monitors (see Chapter 17). Although these algorithms enable
8 Placing the separation between multi-spectral and hyper-spectral imagery at 16 bands is somewhat

arbitrary, and other definitions may be encountered in the literature.

9 An order of magnitude represents a factor of 10.

i i

i i
 i i

i i

792 14. Display Technologies

Approximate typical
monitor display range

Starlight Moonlight Indoor lighting Sunlight

-5 -3 -1 1 3 5
10 10 10 10 10 10
Le (in cd/m2 )

Figure 14.50. Typical luminance levels found in natural environments and the range of
luminances a typical monitor can display.

us to view HDR images on standard monitors, the full potential of an HDR image
is not realized in this manner.
This problem has led to the development of experimental high dynamic range
display devices [1021]. Such devices can be constructed with a two-layered ar-
chitecture, where the first layer is emissive and the second layer is transmissive.
Both layers can be spatially modulated, resulting in a good trade-off between spa-
tial resolution and dynamic range.
Using this principle, Seetzen et al. produced two types of HDR displays
[1021]. In the first type, a DLP device is used as a back projector to provide
the required light. The front layer consists of an LCD panel used to modulate the
light from the projector. Between the layers, a Fresnel lens is installed to couple
the two layers. The projector-based HDR display provides a proof of concept for
the system. In addition, the use of a projector means that high contrasts can be
produced at high frequencies, limited only by the resolution of the projector and
the LCD panel, and the accuracy of the alignment between them. This is a distinct
advantage in applications such as psychophysical experiments. However, its large
depth makes the device impractical in more mainstream applications.
To reduce the depth, a second model was constructed by replacing the projec-
tor with an array of ultra-bright LEDs. In the first prototypes, a hexagonal grid of
760 white LEDs is used, although the currently commercialized 37-inch display
contains 1380 white LEDs. Each LED can be individually controlled. The LCD
front panel has a resolution of 1920 by 1080 pixels (Figure 14.51).
In a darkened room the minimum and maximum display luminance were mea-
sured to be 0.015 cd/m2 and 3450 cd/m2 , respectively. In theory, the minimum
luminance can be zero when all LEDs are turned off. However, in practical situ-
ations, some stray light will always reflect off the screen, thereby increasing the
black level. The minimum value corresponds to the smallest driving signal of
LEDs larger than zero. The dynamic range of the device is therefore 230, 000 : 1,
i.e., more than five orders of magnitude.

i i

i i
 i i

i i

14.16. High Dynamic Range Display Devices 793

Figure 14.51. The BrightSide HDR display at the Max Planck Institute for Biological
Cybernetics; Tübingen, Germany, November 2006.

The HDR display is addressed by deriving two sets of signals from an input
linear-light HDR image. The first set of signals drives the LEDs while the second
set controls the LCD. In principle, there are many combinations of LED and LCD
values that lead to the same net emitted luminance. It is desirable to choose the
lowest LED level possible and adjust the LCD level accordingly. This minimizes
power consumption of the display as well as aging of the LEDs.
The fact that a two-layer system allows for different combinations of LED
and LCD levels to yield the same output can also be leveraged to calibrate the
display [1023]. By adding a lightguide, some of the emitted light can be guided
towards the side of the screen, where image sensors can be fitted. A calibration
procedure can then measure the light output of each individual LED and correct
for variations due to the manufacturing process, including LED chromaticity and
non-uniformities in the LCD screen, as well as operating variations, including the
differential aging of LEDs, and thermal effects.
The low spatial resolution of the back-light has two important implications.
First, the full dynamic range can only be achieved at low spatial frequencies. In
other words, in every local neighborhood covered by a single LED, the dynamic
range will be limited to that of the LCD panel. However, this problem is sig-
nificantly alleviated by the fact that the human eye is not sensitive to extreme
dynamic ranges over small visual angles [1021, 1186].

i i

i i
 i i

i i

794 14. Display Technologies

Second, the image data presented to the LCD panel needs to be adjusted for
the low resolution produced by the LED array. This is performed by artificially
enhancing the edges of the input sent to the LCD layer, where the amount of
enhancement depends on the point spread function of the LED elements.

14.17 Electronic Ink

A reflective display can be constructed from pigments that are electrically charged
[169,205,348]. By making white pigments positively charged and black pigments
negatively charged while suspending them in a clear liquid capsule that is sand-
wiched between a pair of electrodes, a display can be constructed, as diagrammed
in Figure 14.52. The capsule can be as small as 10 μ m2 . The white ink is made
of titanium oxide.
Transparent electrode
Black particles
Clear liquid
Titanium oxide particles
Electrode
Substrate

Figure 14.52. A diagram of an electronic ink pixel.

Fabrication of electronic ink employs standard printing processes and can

therefore be constructed onto several different materials, including plastic, fab-
ric, paper, and glass. The power consumption of electronic ink devices is very
low due to the fact that energy is only required to change the state of a pixel, but
not to maintain it. Therefore, the image being displayed remains after the power
is switched off.

14.18 Display Characterization

To produce colorimetrically accurate images or stimuli, the properties of the dis-
play device must be taken into account. Display characterization refers to the
process of measuring display properties. Display calibration involves adjusting
the display device (or an associated subsystem, such as a graphics adapter), to
bring the display subsystem into a desired configuration [122].
In the most general sense, a display is said to be characterized if for every pixel
value, the corresponding CIE tristimulus values emitted by the display are known.

i i

i i
 i i

i i

14.18. Display Characterization 795

The most direct, albeit impractical, method to achieve this would be to measure
the distribution of emitted light by a spectroradiometer for all pixel values. The
CIE XYZ tristimulus values can then be found using

XRGB = x̄(λ ) f (λ ; R, G, B) d λ , (14.29a)
λ
YRGB = ȳ(λ ) f (λ ; R, G, B) d λ , (14.29b)
λ
ZRGB = z̄(λ ) f (λ ; R, G, B) d λ , (14.29c)
λ

where f (λ ; R, G, B) is the spectral power distribution given the pixel triplet

(R, G, B). The CIE color matching functions x̄, ȳ, z̄ are used here. This method
is impractical because would require approximately 16 million measurements for
a display having 8 bits per channel. Even if attempted, the characteristics of the
display device are liable to change significantly during the time taken to measure
the device [211].
The number of measurements can be significantly reduced if the display de-
vice satisfies certain assumptions. The first of these assumptions is called chan-
nel independence. This assumption of channel independence states that driving
a color channel with different signal values does not affect the distribution and
amount of the light emitted from other channels. Under this assumption, the fol-
lowing relationship holds:

f (λ , R, G, B) = fr (λ ; R) + fg (λ ; G) + fb (λ ; B) + f0 . (14.30)

Here, f0 denotes the sum of the light reflected by the display surface due to un-
wanted external light and the light emitted when all pixels are zero. This as-
sumption reduces a three-dimensional problem to three simpler one-dimensional
problems. Equation (14.30) can be rewritten using tristimulus values instead of
spectral distributions:
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
XRGB XR XG XB X0
⎣YRGB ⎦ = ⎣YR ⎦ + ⎣YG ⎦ + ⎣YB ⎦ + ⎣Y0 ⎦ . (14.31)
ZRGB ZR ZG ZB Z0

Display characterization can be further simplified if a second assumption known

as channel chromaticity constancy also holds. This assumption states that the
basic spectral shape of the emitted light remains the same regardless of the value
of a color channel. Thus, the spectral distribution only undergoes an amplitude

i i

i i
 i i

i i

796 14. Display Technologies

scaling:

fr (λ ; R) = vr (R) f r (λ ; Rmax ), (14.32a)

fg (λ ; G) = vg (G) fg (λ ; Gmax ), (14.32b)
fb (λ ; B) = vb (B) fb (λ ; Bmax ). (14.32c)

Here, fr (λ ; Rmax ) represents the emission spectra when the red channel is full on,
and vr (R) is the tone response curve when the channel is driven with R instead
of Rmax . With this formulation, all tone response curves must lie in the interval
[0, 1]. It is possible to rewrite (14.32) using the tristimulus values:
⎡ ⎤ ⎡ ⎤
XR XRmax
⎣YR ⎦ = vr (R) ⎣YRmax ⎦ , (14.33a)
ZR ZRmax
⎡ ⎤ ⎡ ⎤
XG XGmax
⎣YG ⎦ = vg (G) ⎣ YGmax ⎦ , (14.33b)
ZG ZGmax ,
⎡ ⎤ ⎡ ⎤
XB XBmax
⎣YB ⎦ = vb (B) ⎣YBmax ⎦ , (14.33c)
ZB ZBmax

where the subscripts Rmax , Gmax , and Bmax are used to distinguish the tristimulus
values produced when each channel is driven by the maximum signal value. If
a monitor satisfies both of these assumptions, a forward model can be written to
simulate its operation. For this, we can combine (14.31) and (14.33):
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
XRGB XRmax XGmax XBmax vr (R) X0
⎣YRGB ⎦ = ⎣YRmax YGmax YBmax ⎦ ⎣vg (G)⎦ + ⎣ Y0 ⎦ , (14.34a)
ZRGB ZRmax ZGmax ZBmax vb (B) Z0 ,
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
XRGB vr (R) X0
⎣YRGB ⎦ = M ⎣vg (G)⎦ + ⎣Y0 ⎦ , (14.34b)
ZRGB vb (B) Z0

where the matrix M is the 3 × 3 tristimulus matrix. The forward model is useful
to find the tristimulus value corresponding to an (R, G, B) triplet used to drive
the display. In most situations though, an inverse model is required to find the
(R , G , B ) values that yield the desired tristimulus values [1035]. This model can

i i

i i
 i i

i i

14.18. Display Characterization 797

be obtained by inverting the forward model:

⎡ ⎤ ⎡ ⎤
RXYZ X − X0
⎣GXYZ ⎦ = M −1 ⎣ Y −Y0 ⎦ , (14.35a)
BXYZ Z − Z0 ,
⎡  ⎤ ⎡ −1 ⎤
R vr (RXYZ )
⎣G ⎦ = ⎣v−1
g (GXYZ ) .
⎦ (14.35b)
B v−1
b (BXYZ )

Note that the inverse model is valid only if the matrix M as well as the tone re-
sponse curves are invertible. For M to be invertible, the individual color channels
of the monitor must be independent. This condition is satisfied by most typical
desktop computer monitors [1034, 1035], but it is not satisfied by LCD monitors
to be used in demanding graphics arts, studio video, or digital cinema production
work. Thus, 3 × 3 linear models have been found to work well with both CRT and
LCD displays [74, 92]. However, they are not good predictors for four-channel
DLP devices [1096, 1259, 1260]. In addition, tone response curves are usually
monotonically increasing functions and are therefore invertible [1035].
As (14.35b) shows, the three key components necessary for display character-
ization are

1. the color conversion matrix M −1 ;

2. the tone response curves vr , vg , vb ;

3. The sum of the emitted light when all pixels are zero and the light reflected
off the surface of the monitor.

In the following we focus on the details of how each component can be obtained.
Before carrying out any measurements, the first step must be to set the bright-
ness and contrast controls of the monitor since these controls affect the tone re-
sponse curves. For a CRT monitor, the brightness control affects the black level
of the monitor, which is the amount of light emitted when displaying a black im-
age. It would therefore be better to refer to this control as the black level setting.
It operates on non-linear RGB values and, therefore, slides the range of values
along the electro-optic conversion curve (EOCF).
This control should first be set to minimum and then slowly increased until
the display just starts to show a hint of gray for a black input [923]. The setting is
then reduced one step to full black. Setting the brightness control to a lower value
causes information to be lost in the dark regions of the displayed material. Setting

i i

i i
 i i

i i

798 14. Display Technologies

1.0

Luminance (normalized)
0.9
0.8
Higher
0.7
contrast
0.6
0.5
0.4
0.3
0.2
Lower
0.1 contrast
0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Digital counts (normalized)

Figure 14.53. The contrast control of CRTs and the brightness control of LCDs are used
to maximize the luminance range output by the display device (after [923]).

it to a higher value may significantly reduce the contrast of displayed imagery. If

the black level is set appropriately, it may be assumed for the purpose of further
calculations that the black level is zero. This simplifies the calibration procedure.
The meaning of the brightness control in an LCD display device is different
than for a CRT, as here the control affects the intensity of the back-light [923].
After setting the correct brightness value, the contrast knob can be adjusted
to maximize the available display luminance. Increasing the contrast amounts to
scaling the displayed luminance, as shown in Figure 14.53.
Once brightness and contrast settings are made, the next step involves mea-
suring the emitted light when all pixels are set to zero. Ideally, this measurement
should be made under the lighting conditions under which the display will sub-
sequently be used. The measured values are then converted to tristimulus values,
which represent the monitor’s black level.
The next step is to measure the tristimulus values corresponding to the max-
imum drive value for each channel. For this, each channel is individually driven
with the maximum pixel value and its corresponding tristimulus values are com-
puted. These are then adjusted by subtracting the tristimulus values of the black
level. The adjusted tristimulus values constitute the matrix M in (14.34b).
The last step involves the measurement of the tone response curves. This
is typically achieved by ramping each channel from zero to its maximum signal

i i

i i
 i i

i i

14.18. Display Characterization 799

value and taking measurements at several sample points:


x̄(λ ) fr (λ ; Ri )d λ
λ
vr (Ri ) =  (14.36a)
x̄(λ ) fr (λ ; Rmax )d λ
λ
XRi
= , (14.36b)
XRmax
YGi
vg (Gi ) = , (14.36c)
YGmax
ZBi
vb (Bi ) = . (14.36d)
ZBmax

Ordinarily, a power-law corresponding to the display’s intended standard will re-

sult, e.g., a 2.2 power law for an sRGB-compliant display. In principle, we could
have used the Y tristimulus value to compute all three response curves. How-
ever, using X for the red channel, Y for the green channel, and Z for the blue
channel ensures that we collect data where the emissivity of each channel is the
highest. This makes the method more robust against measurement noise [92].
The full curve can then be constructed by either linear or non-linear interpola-
tion [92]. For an 8-bit/channel display, 17 measurements are typically sufficient
(i.e., 1 ≤ i ≤ 17) [92, 704, 920, 1243], that is driving with pixel component values
0, 15, 30, . . ., 255.
With the computation of the tone response curves, all the necessary infor-
mation for monitor characterization is established, assuming that the channel in-
dependence and channel chromaticity constancy properties hold. However, it is
essential to remember that this information is valid as long as the monitor is used
with comparable settings that were used during characterization. Also if the mon-
itor will be in an illuminated room, the light reflected off the screen must be
measured and incorporated into the offset term [X0 ,Y0 , Z0 ]T of (14.35a).

14.18.1 CRT Monitors

In the previous section, we explained the most generic way of monitor charac-
terization without discussing the specifics of different display devices. The accu-
racy of this method depends on how well the display device meets the channel
independence and chromaticity constancy requirements. In the domain of CRT
monitors, these measurements are also called gun independence and phosphor
constancy [1035]. Fortunately, most CRT display devices satisfy these assump-
tions sufficiently well [91, 210].

i i

i i
 i i

i i

800 14. Display Technologies

The tone response curves of CRT monitors are often modeled by a gain-offset
gamma (GOG) model . For the red channel, this is expressed as
 γr
R
vr (R) = kg,r + ko,r , (14.37)
Rmax
where kg,r , ko,r , and γr denote the gain, offset, and gamma for the red chan-
nel [92]. In this equation, the gamma term γ models the inherent non-linearity
in the cathode ray tube. It is caused by the space charge effect due to the accu-
mulation of electrons between the cathode and the accelerating anode (see Sec-
tion 14.1) [173, 652]. This parameter, which is typically between 2.2 and 2.4,
cannot be altered by the user.
However, the other two parameters, gain and offset, can usually be adjusted by
the user by altering the contrast and brightness knobs, respectively. Under proper
settings (14.37) simplifies to
 γr
R
vr (R) = . (14.38)
Rmax
Similar equations can be written for the green and blue channels. With this sim-
plification, we can rewrite (14.35b) for a CRT monitor as follows:
⎡ ⎤ ⎡  ⎤
R (RXYZ )1/γr
⎣G⎦ = ⎣(G )1/γg ⎦ . (14.39)
XYZ
B (BXYZ )1/γb
In computer graphics, this operation is commonly called gamma correction. In
video and HDTV, gamma correction takes place at the camera and incorporates
a power function exponent that effects not only pre-compensation for the non-
linearity of the eventual display device, but also includes compensation for ap-
pearance effects at the display.

14.18.2 LCD Monitors

As discussed in Section 14.2, LCD monitors work on a different principle than
that of CRT monitors. The assumptions for CRT monitors are not necessarily
valid for LCD displays. However, several characterization studies have shown
that the channel independence and chromaticity constancy assumptions hold to a
large degree also for LCD displays [297, 1035]. Nevertheless, displays produced
by different manufacturers may differ significantly in the degree they satisfy these
assumptions, and therefore some displays may not be characterizable with the
standard technique explained in the previous section [359]. For such monitors,
more complex methods can be employed [237, 1116, 1277].

i i

i i
 i i

i i

14.18. Display Characterization 801

Normalized emitted radiance

1.0

0. 9 Measured red channel

Measured red channel
0. 8 Measured red channel
sRGB EOCF (2.2 power)
0. 7

0. 6

0. 5

0. 4

0. 3

0. 2

0. 1

0.0
0.0 0. 1 0. 2 0. 3 0. 4 0. 5 0. 6 0. 7 0. 8 0. 9 1
Normalized pixel value

Figure 14.54. Measured electro-optic conversion curve (EOCF) of the red, green, and
blue channels of a Dell UltraSharp 2007FP 20.1-inch LCD monitor. A typical gamma of
2.2 is shown for comparison.

One of the important differences between LCDs and CRTs lies in their tone re-
sponse curves. These functions are usually called electro-optic conversion curves
(EOCFs). The native response of LCDs is shown in Figure 14.11, indicating
that the intrinsic, physical electro-optic conversion curves are essentially sig-
moidal [368, 1035].
However, to make LCD monitors compliant with the existing image, video,
and PC content, and to make them compatible with CRT display devices, they
are typically equipped with look-up tables to transform their response curves to
follow a 2.2 power law [297, 1035], as standardized by the sRGB standard (IEC
61966-2-1). In Figure 14.54, the measured curves of the individual color chan-
nels of a current LCD monitor are shown. Note that both axes in this graph are
normalized, so that the black level of this display cannot be inferred from this
figure. This specific monitor closely approximates the response curve of a typical
CRT with an EOCF of 2.2. For some monitors though, critical deviations from
the gain-offset gamma model are sometimes observed, and therefore the use of
external look-up tables is recommended for color critical applications [237, 297].

i i

i i
 i i

i i

802 14. Display Technologies

The assumption of channel chromaticity constancy is only approximate for

LCD displays. With two color channels set to zero, in the ideal case, varying
the value of the remaining color channel should produce identical xy chromaticity
coordinates. In an LCD panel, however, light leakage impairs this ideal behavior.
By subtracting the tristimulus values of the black level, the resulting values are
likely to exhibit much improved channel chromaticity constancy, as for instance
shown for an Apple Studio Display [297].
Finally, dependent on the specific technology employed, the spectral distri-
bution of light emitted from an LCD panel may change significantly with the
viewing angle. Therefore, characterization performed from one viewing angle
will only be valid for that viewing angle. Significant colorimetric variations are
observed for LCDs with changes in viewing angle [1035].

14.18.3 LCD Projectors

An LCD projector can be characterized in much the same way as an LCD monitor.
However, here the contribution of the black level is more important and cannot be
neglected. Thus, for an LCD projector display, the procedure for characterization
is as follows [1096].
First, the pixel values are linearized. It should be noted that most projectors
are built for displaying video or PC content and, therefore, follow a power-law
curve. Thus, linearization can be achieved by applying (14.39), leading to the
linear triplet (R, G, B)T .
Next, the triplets for the primaries (XR ,YR , ZR )T , (XG ,YG , ZG )T , and
(XB ,YB , ZB )T , as well as the black point (XK ,YK , ZK )T are measured. Using ho-
mogeneous coordinates (see Appendix A.10), the transformation is given by
⎡ ⎤ ⎡ ⎤⎡ ⎤
X XR − XK XG − XK XB − XK XK R
⎢Y ⎥ ⎢ YR −YK YG −YK YB −YK YK ⎥ ⎢G⎥
⎢ ⎥=⎢ ⎥⎢ ⎥, (14.40)
⎣ Z ⎦ ⎣ ZR − ZK ZG − ZK ZB − ZK ZK ⎦ ⎣ B ⎦
1 0 0 0 1 1

yielding the desired transformation.

14.18.4 Calibration in Practice

Calibration of display devices is an involved procedure that typically requires
dedicated measuring equipment. This task can be simplified by using software
and/or hardware solutions, either measuring equipment provided with the display
subsystem, or third-party equipment, such as that depicted in Figure 14.55. A
sensor is placed in front of the display, measuring its output. By correlating the

i i

i i
 i i

i i

14.19. Further Reading 803

Figure 14.55. A display calibration device.

values sent to the display with the measured values, the display is characterized.
The output of the calibration procedure is typically an ICC profile, which is used
either by the operating system, or by application software to adjust imagery for
display on the characterized display. Such profiles are commonly in use to ensure
appropriate color reproduction (see Chapter 16).

14.19 Further Reading

Display devices are discussed in many textbooks. Particularly relevant to color
is Widdel et al., Color in Electronic Displays [1243]. Poynton’s Digital Video
and HDTV: Algorithms and Interfaces contains a wealth of color-related dis-
play information [923]. Specific display technologies are presented in Liquid
Crystal TV Displays: Principles and Applications of Liquid Crystal Displays by
Kaneko [571], Organic Light-Emitting Devices: A Survey by Shinar [1040], and
Flat-Panel Displays and CRTs by Tannas [1120]. The electro-optical effects in
different types of liquid crystal displays, as well as the various addressing schemes
in use for such displays, are discussed by Ernst Lueder in Liquid Crystal Displays:
Addressing Schemes and Electro-Optical Effects [706]. The desire to manufacture
flexible flat-panel displays has led to a considerable amount of research, which is
discussed in Gregory P Crawford (ed.), Flexible Flat Panel Displays [213].

i i

i i
 i i

i i

Chapter 15
Image Properties and
Image Display

In this book, the focus is on light and color, and its implications for various ap-
plications dealing with images. In previous chapters, some of the fundamentals
pertaining to image capture and image display were discussed. This leaves us with
several attributes of images that are worth knowing about. In particular, images of
existing environments tend to have statistical regularities that can be exploited in
applications such as image processing, computer vision, and computer graphics.
These properties are discussed in the following section.
Further, we discuss issues related to the measurement of dynamic range. Such
measurements are becoming increasingly important, especially in the field of high
dynamic range imaging, where knowledge about the dynamic range of images can
be used to assess how difficult it may be to display such images on a given display
device.
The third topic in this chapter relates to cross-media display. As every display
has its own specification and properties and is located in a differently illuminated
environment, the machinery required to prepare an image for accurate display on
a specific display device, should involve some measurement of both the display
device and its environment. One of the tasks involved in this process, is matching
the gamut of an image to the gamut of the display device. Finally, we discuss
gamma correction and possible adjustments for ambient light which is reflected
off the display device.

805

i i

i i
 i i

i i

806 15. Image Properties and Image Display

15.1 Natural Image Statistics

Images of natural scenes tend to exhibit statistical regularities that are important
for the understanding of the human visual system which has evolved to interpret
such natural images. In addition, these properties may prove valuable in image-
processing applications. An example is the color transfer algorithm outlined in
Section 8.10, which relies on natural image statistics to transfer the color mood
of an image to a second image.
If all greyscale images that consist of 1000 × 1000 pixels are considered,
then the space that these images occupy would be a million-dimensional space.
Most of these images never appear in our world: natural images form a sparse
subset of all possible images. In recent years, the statistics of natural images have
been studied to understand how properties of natural images influence the human
visual system. In order to assess invariance in natural images, usually a large set
of calibrated images are collected into an ensemble and statistics are computed on
these ensembles, rather than on individual images.
Natural image statistics can be characterized by their order. In particular, first-
order, second-order, and higher-order statistics are distinguished [1001]:

First-order statistics treat each pixel independently, so that, for example, the dis-
tribution of intensities encountered in natural images can be estimated.

Second-order statistics. measure the dependencies between pairs of pixels, which

are usually expressed in terms of the power spectrum (see Section 15.1.2).

Higher-order statistics are used to extract properties of natural scenes which can
not be modeled using first- and second-order statistics. These include, for
example, lines and edges.

Each of these categories is discussed in some detail in the following sections.

15.1.1 First-Order Statistics

Several simple regularities have been established regarding the statistical distri-
bution of pixel values; the gray-world assumption is often quoted in this respect.
Averaging all pixel values in a photograph tends to yield a color that is closely
related to the blackbody temperature of the dominant illuminant [902]. The inte-
grated color tends to cluster around a blackbody temperature of 4800 K.

i i

i i
 i i

i i

15.1. Natural Image Statistics 807

For outdoor scenes, the average image color remains relatively constant from
30 minutes after dawn until within 30 minutes of dusk. The correlated color
temperature then increases, giving an overall bluer color.1
The average image color tends to be distributed around the color of the dom-
inant scene illuminant according to a Gaussian distribution, if measured in log
space [665, 666]. The eigenvectors of the covariance matrix of this distribution
follow the spectral responsivities of the capture device, which for a typical color
negative film are [666]
1
L = √ (log(R) + log(G) + log(B)) , (15.1a)
3
1
s = √ (log(R) − log(B)) , (15.1b)
2
1
t = √ (log(R) − 2 log(G) + log(B)) . (15.1c)
6
For a large image database, the standard deviations in these three channels are
found to be 0.273, 0.065, and 0.030. Note that this space is similar to the Lαβ
color space discussed in Section 8.7.5. Although that space is derived from LMS
rather than RGB, both operate in log space, and both have the same channel
weights. The coefficients applied to each of the channels are identical as well,
with the only difference being that the role of the green and blue channels has
been swapped.

15.1.2 Second-Order Statistics

The most remarkable and salient natural image statistic that has been discovered
so far is that the slope of the power spectrum tends to be close to 2. The power
spectrum of an M × M image is computed as

|F(u, v)|2
S(u, v) = , (15.2)
M2
where F is the Fourier transform of the image. By representing the two-dimen-
sional frequencies u and v in polar coordinates (u = f cos φ and v = f sin φ ) and
averaging over all directions φ and all images in the image ensemble, it is found
that on log-log scale amplitude as a function of frequency, f lies approximately
on a straight line [144, 314, 983, 986, 1001]. This means that spectral power as a
function of spatial frequency behaves according to a power law function. More-
over, fitting a line through the data points yields a slope α of approximately 2 for
1 Despite the red colors observed at sunrise and sunset, most of the environment remains illumi-

nated by a blue skylight, which increases in importance as the sun loses its strength.

i i

i i
 i i

i i

808 15. Image Properties and Image Display

natural images. Although this spectral slope varies subtly between different stud-
ies [144, 259, 315, 983, 1130], and with the exception of low frequencies [651],
it appears to be extremely robust against distortions and transformations. It is
therefore concluded that this spectral behavior is a consequence of the images
themselves, rather than of particular methods of camera calibration or exact com-
putation of the spectral slope. We denote this behavior by
A A
S( f ) ∝ = 2−η , (15.3)
fα f
where A is a constant determining overall image contrast, α is the spectral slope,
and η is its deviation from 2. However, the exact value of the spectral slope
depends somewhat on the type of scenes that make up the ensembles. Most in-
terest of the natural image statistics community is in scenes containing mostly
trees and shrubs. Some studies show that the spectral slope for scenes containing
man-made objects is slightly different [1295, 1296]. Even in natural scenes, the
statistics vary, dependent on what is predominantly in the images. The second-
order statistics for sky are, for example, very different from those of trees.
One of the ways in which this becomes apparent is when the power spectra are
not circularly averaged, but when the log average power is plotted against angle.
For natural image ensembles, all angles show more or less straight power spectra,
although most power is concentrated along horizontal and vertical directions [986,
1001] (see also Figure 15.3). The horizon and the presence of tree trunks are said
to be factors, although this behavior also occurs in man-made environments.
The power spectrum is related to the auto-correlation function through the
Wiener-Khintchine theorem, which states that the auto-correlation function and
the power spectrum form a Fourier transform pair [840]. Hence, power spectral
behavior can be equivalently understood in terms of correlations between pairs of
pixel intensities.
A related image statistic is contrast, normally defined as the standard deviation
of all pixel intensities divided by the mean intensity (σ /μ ). This measure can
either be computed directly from the image data, or it can be derived from the
power spectrum through Parceval’s theorem [1001]:

σ2
= ∑ S(u, v). (15.4)
μ 2 (u,v)

This particular contrast computation can be modified to compute contrast in dif-

ferent frequency bands. Frequency-conscious variance can then be thresholded,
yielding a measure which can detect blur [313]. This is useful as lack of contrast
can also be caused by the absence of sufficient detail in a sharp image.

i i

i i
 i i

i i

15.1. Natural Image Statistics 809

The above second-order statistics are usually collected for luminance images
only, as luminance is believed to carry the greatest amount of information. How-
ever, chrominance channels are shown to exhibit similar spectral behavior [878],
and therefore, all subsequent qualitative arguments are expected to be true for
color as well.
The fact that the power spectral behavior of natural images, when plotted in
log-log space, yields a straight line with a slope of around -2 is particularly impor-
tant, since recent unrelated studies have found that in image interpretation tasks,
the human visual system performs best when the images conform to the above
second-order statistic. In one such study, images of a car and a bull were morphed
into each other, with varying distances between the images in the sequence [879].
Different sequences were generated with modified spectral slopes. The minimum
distance at which participants could still distinguish consecutive images in each
morph sequence was then measured. This distance was found to be smallest when
the spectral slope of the images in the morph sequence was close to 2. Deviation
of the spectral slope in either direction resulted in a deteriorated performance to
distinguish between morphed images.
In a different study, the effect of spectral slope on the detection of mirror
symmetry in images was assessed [933]. Here, white noise patterns with varying
degrees of vertical symmetry were created and subsequently filtered to alter the
spectral slope. An experiment, in which participants had to detect if symmetry
was present, revealed that performance was optimal for images with a spectral
slope of 2.
These studies confirm the hypothesis that the HVS is tuned to natural images.
In fact, the process of whitening (i.e., flattening the power spectrum to produce a
slope α of 0) is consistent with psychophysical measurements [52], which indi-
cates that the HVS expects to see images with a 1/ f 2 power spectrum.

15.1.3 Power Spectra

In this section, we show how the power spectrum of an ensemble of images can
be established. We used 133 images from the van Hateren database of natural
images [433]. The images in this ensemble are available as luminance images.
Example images are shown in Figure 15.1.
For this image ensemble, the power spectrum was computed and the spectral
slope was estimated. The power spectrum computation proceeds as follows.2
A window of 512 × 512 pixels was cut out of the middle of the image upon
which further processing was applied. Then, the weighted mean intensity μ was
2 This calculation closely follows the method described in [1001].

i i

i i
 i i

i i

810 15. Image Properties and Image Display

Figure 15.1. Example images drawn from the van Hateren database [433]. (Images cour-
tesy of Hans van Hateren.)

subtracted to avoid leakage from the DC-component of the image, where μ is

defined as
∑(x,y) L(x, y)w(x, y)
μ= . (15.5)
∑(x,y) w(x, y)
Next, the images were prefiltered to avoid boundary effects. This is accom-
plished by applying a circular Kaiser-Bessel window function (with parameter
α = 2) to the image [429]:
 
2 2 
x +y
I0 πα 1.0 − (N/2) 2  N
w(x, y) = , 0 ≤ x2 + y2 ≤ . (15.6)
I0 (πα ) 2

Here, I0 is the modified zero-order Bessel function of the first kind and N is
the window size (512 pixels). In addition, this weight function was normalized
by letting
∑ w(x, y)2 = 1. (15.7)
(x,y)

This windowing function was chosen for its near-optimal trade-off between
side-lobe level, main-lobe width, and computability [429]. The resulting images

i i

i i
 i i

i i

15.1. Natural Image Statistics 811

were then compressed using the Fourier transform,

L(x, y) − μ 2π i(ux+vy)
F(u, v) = ∑ μ
e . (15.8)
(x,y)

Finally, the power spectrum was computed as per (15.2) and the resulting
data points plotted. Although frequencies up to 256 cycles per image are com-
puted, only the 127 lowest frequencies were used to estimate the spectral slope.
Higher frequencies may suffer from aliasing, noise, and low modulation trans-
fer [1001]. The estimation of the spectral slope was performed by fitting a straight
line through the logarithm of these data points as a function of the logarithm of
1/ f . This method was chosen over other slope estimation techniques, such as
the Hill estimator [472] and the scaling method [215], to maintain compatibility
with [1001]. In addition, the number of data points (127 frequencies) is insuf-
ficient for the scaling method, which requires at least 1,000 data points to yield
reasonably reliable estimates.
With this method, second-order statistics were extracted. The 1.87 spectral
slope reported for the van Hateren database was confirmed (we found 1.88 for
our subset of 133 images). The deviations from this value for the artificial image
ensemble are depicted in Figure 15.2. The angular distribution of power tends to
show peaks near horizontal and vertical angles (Figure 15.3). Finally, the distribu-
tion of spectral slopes for the 133 images in this ensemble is shown in Figure 15.4.

Power vs. spatial frequency

0
log10 Power

−1 Slope −1.88
−2
−3
−4
−5
−6
−7
−8
−9
0 1 2 3
Spatial frequency (log10 cycles/image)

Figure 15.2. Spectral slope for the image ensemble. The double lines indicate ±2 standard
deviations for each ensemble.

i i

i i
 i i

i i

812 15. Image Properties and Image Display

Average power per direction

log10 average power

−4.5

−5.0

−5.5

−6.0
0 90 180 270 360
Orientation (deg)

Figure 15.3. Log average power as function of angle.

0.15
fraction

0.10

0.05

0.00
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
1/f-exponent (α)

Figure 15.4. A histogram of spectral slopes, derived from an image ensemble of 133
images taken from the van Hateren database.

15.1.4 Higher-Order Statistics

One of the disadvantages of using amplitude information in the frequency domain
is that phase information is completely discarded, thus ignoring the position of
edges and objects. For this reason, higher-order statistics have been applied to
natural image ensembles. The simplest global nth-order statistics that may capture
phase structure are the third and fourth moments, commonly referred to as skew
and kurtosis [840]:
E{x3 }
s= , (15.9a)
E{x2 }3/2
E{x4 }
k= − 3. (15.9b)
E{x2 }2
These dimensionless measures are, by definition, zero for pure Gaussian distribu-
tions. Skew can be interpreted as an indicator of the difference between the mean

i i

i i
 i i

i i

15.1. Natural Image Statistics 813

and the median of a data set. Kurtosis is based on the size of a distribution’s tails
relative to a Gaussian distribution. A positive value, associated with long tails in
the distribution of intensities, is usually associated with natural image ensembles.
This is for example evident when plotting log-contrast histograms, which plot
the probability of a particular contrast value appearing. These plots are typically
non-Gaussian with positive kurtosis [986].
Thomson [1126] has pointed out that for kurtosis to be meaningful for nat-
ural image analysis, the data should be decorrelated, or whitened, prior to any
computation of higher-order statistics. This can be accomplished by flattening the
power spectrum, which ensures that second-order statistics do not also capture
higher-order statistics. Thus, skew and kurtosis become measures of variations
in the phase spectra. Regularities are found when these measures are applied to
the pixel histogram of an image ensemble [1125, 1126]. This appears to indicate
that the HVS may exploit higher-order phase structure after whitening the image.
Understanding the phase structure is therefore an important avenue of research in
the field of natural image statistics. However, it appears that so far it is not yet
possible to attach meaning to the exact appearance of higher-order statistics.

15.1.5 Gradient Field Statistics

Rather than study the statistical properties of pixels, we can also look at derived
quantities, such as gradients. The gradient field of an image is a vector-valued
quantity ∇L, defined as

∇L = (L(x + 1, y) − L(x, y), L(x, y + 1) − L(x, y)) . (15.10)

Natural images typically have strong discontinuities in otherwise relatively

smooth regions. This means that large image areas have only small gradients, with
larger gradients happening more sparsely. Thus, the components of the gradient
field will show a characteristic distribution, which can be observed by computing
its histogram. The gradient distribution of a set of four example images, shown
in Figure 15.5, is plotted in Figure 15.6. The images used here were captured in
high dynamic range (see Chapter 17), and therefore contain much larger gradi-
ents than encountered in conventional 8-bit images. In addition, these images are
linear by construction, having a gamma of 1.0. Conventional images are likely
to have gamma correction applied to them, and this may affect the distribution of
gradients in images.

i i

i i
 i i

i i

814 15. Image Properties and Image Display

Figure 15.5. Example images used for computing gradient distributions, which are shown
in Figure 15.6. From left to right, top to bottom: Abbaye de Beauport, Paimpol, France,
July, 2007; Combourg, France, July, 2007; Fougères, France, June, 2007); Kermario,
Carnac, France, July, 2007.

0
log2 probability

Abbaye de Beauport
-2
Combourg
-4 Fougeres
Kermario (Carnac)
-6

-8

-10

-12

-14

-16

-18
-1.5 -1.0 -0.5 0.0 0.5 0.1 1.5
Gradient (× 104)

Figure 15.6. The probability distribution of gradients for the four images depicted in
Figure 15.5.

i i

i i
 i i

i i

15.1. Natural Image Statistics 815

15.1.6 Response Properties of Cortical Cells

One of the reasons to study the statistics of natural images is to understand how
the HVS codes these images. Because natural images are not completely random,
they incorporate a significant amount of redundancy. The HVS may represent
such scenes using a sparse set of active elements. These elements will then be
largely independent.
Here, it is assumed that an image can be represented as a linear superposition
of basis functions. Efficiently encoding images now involves finding a set of
basis functions that spans image space and ensures that the coefficients across an
image ensemble are statistically as independent as possible [853]. The resulting
basis functions (or filters) can then be compared to the response properties of
cortical cells in order to explain the early stages of human visual processing. This
has resulted in a number of different representations for basis functions, including
principal components analysis (PCA) [55,420], independent components analysis
(ICA) [82,83,433], Gabor basis functions [314,315], and wavelets [315,514,853].
The PCA algorithm finds a set of basis functions that maximally decorrelates
pairs of coefficients; this is achieved by computing the eigenvalues of the co-
variance matrix (between pairs of pixel intensities). The corresponding eigenvec-
tors represent a set of orthogonal coefficients, whereby the eigenvector with the
largest associated eigenvalue accounts for the largest part of the covariance. By
using only the first few coefficients, it is possible to encode an image with a large
reduction in free parameters [314]. If stationarity is assumed across all images
in an ensemble and between all regions in the images, i.e., the image statistics
are the same everywhere, then PCA produces coefficients that are similar to the
Fourier transform [113]. Indeed, PCA is strictly a second-order method assuming
Gaussian signals.
Unfortunately, decorrelation does not guarantee independence. Also, intensi-
ties in natural images do not necessarily have a Gaussian distribution, and there-
fore PCA yields basis functions that do not capture higher-order structure very
well [853]. In particular, the basis functions tend to be orientation and frequency
sensitive, but with global extent. This is in contrast to cells in the human primary
cortex which are spatially localized (as well as being localized in frequency and
orientation).
In contrast to PCA, independent components analysis constitutes a transfor-
mation resulting in basis functions that are non-orthogonal, localized in space,
and selective for frequency and orientation. They aim to extract higher-order in-
formation from images [82,83]. ICA finds basis functions that are as independent
as possible [522]. To avoid second-order statistics from influencing the result
of ICA, the data is usually first decorrelated (also called whitened), for example

i i

i i
 i i

i i

816 15. Image Properties and Image Display

using a PCA algorithm. Filters can then be found that produce extrema of the
kurtosis [1001]. A kurtotic amplitude distribution is produced by cortical simple
cells, leading to sparse coding. Hence, ICA is believed to be a better model than
PCA for the output of simple cortical cells.
The receptive fields of simple cells in the mammalian striate cortex are lo-
calized in space, oriented, and bandpass. They are therefore similar to the basis
functions of wavelet transforms [315, 853]. For natural images, strong correla-
tions between wavelet coefficients at neighboring spatial locations, orientations,
and scales, have been shown using conditional histograms of the coefficients’
log magnitudes [1049]. These results were successfully used to synthesize tex-
tures [919, 1047] and to denoise images [1048].

15.1.7 Dimensionality of Reflectance and Daylight Spectra

The reflectance spectra of large numbers of objects have been analyzed for the
purpose of determining their dimensionality. We have so far assumed that a spec-
tral distribution is sampled at regular intervals, for instance from 380 to 780 with
a 5 or 10 nm spacing. This leads to a representation of each spectrum by 43 or so
numbers.
Values in between sample points can be approximated, for instance, by linear
interpolation. However, it is also possible to use higher-order basis functions and,
thereby, perhaps reduce the number of samples. The question then becomes which
basis functions should be used, and how many are required.
It has been found that the eigenvalues of the covariance matrix decay very
fast, leading to only a few principal components that account for most of the vari-
ance [201]. Similar results have been found for daylight spectra [559], suggesting
that spectral distributions can be represented in a low-dimensional space, for in-
stance having no more than six or seven dimensions [728, 729].

15.2 Dynamic Range

The dynamic range afforded by an image encoding depends on the ratio between
smallest and largest representable value, as well as the step size between succes-
sive values. Conventional images are stored with a byte per pixel for each of the
red, green, and blue components. Thus, for low dynamic range images, there are
only 256 different values per color channel.
High dynamic range images encode a significantly larger set of possible val-
ues. The maximum representable value may be much larger, and the step size
between successive values may be much smaller. The file size of high dynamic

i i

i i
 i i

i i

15.2. Dynamic Range 817

range images is therefore generally larger as well, although at least one standard
(the OpenEXR high dynamic range file format [565]) includes a very capable
compression scheme, and JPEG-HDR essentially occupies the same amount of
space as an equivalent low dynamic range image would [1209, 1210]. To con-
struct an image in this file format, the HDR image is tone-mapped and encoded as
a standard JPEG. The ratio between the original and the JPEG image is subsam-
pled and stored in a meta-tag. HDR-aware viewers can reconstruct the HDR from
this data, whereas other image viewers simply ignore the meta-tag and display the
JPEG directly. Encoding of a residual image, such as the ratio image, can also be
used to store an extended gamut [1073].
Recently, the problem of HDR video encoding has received some attention.
This has led to MPEG-HDR, a high dynamic range video encoder/decoder
promising high compression ratios [733], as well as a second video-encoding
scheme that is based on a model of human cones [435], which was briefly dis-
cussed in Section 4.8.
In general, the combination of smallest step size and ratio of the smallest and
largest representable values determines the dynamic range that an image encoding
scheme affords. Images have a dynamic range that is bounded by the encoding
scheme used. The dynamic range of environments is generally determined by the
smallest and largest luminance found in the scene. Assuming this range is smaller
than the range afforded by the encoding scheme, such an image can be captured
and stored. The question then remains: What is the actual dynamic range of such
a scene?
A simplistic approach to measure the dynamic range of an image may, there-
fore, compute the ratio between the largest and smallest pixel value of an im-
age. Sensitivity to outliers may be reduced by ignoring a small percentage of the
darkest and brightest pixels. Alternatively, the same ratio may be expressed as a
difference in the logarithmic domain.
However, the recording device or rendering algorithm may introduce noise
that will lower the useful dynamic range. Thus, a measurement of the dynamic
range of an image should factor in noise. A better measure of dynamic range
is, therefore, a signal-to-noise ratio, expressed in decibels (dB), as used in signal
processing. This concept is discussed further in the following section.
Assuming we have a reasonable measure of dynamic range, it then becomes
possible to assess the extent of the mismatch between the dynamic range of an
image and the dynamic range of a given display device. A good tone-reproduction
operator will be able to reduce the dynamic range of an image to match that of the
display device without producing visible artifacts.

i i

i i
 i i

i i

818 15. Image Properties and Image Display

15.2.1 Signal-to-Noise Ratio

Assuming that the signal and the noise are interdependent, a measure of the signal-
to-noise ratio (SNR) may be computed as follows:
 
μs
SNR dep
= 20 log10 , (15.11)
σn

where μs is the mean average of the signal, and σn is the standard deviation of
the noise. In the case that signal and noise are independent, the SNR may be
computed by
 
σs
SNR indep
= 20 log10 . (15.12)
σn
For high dynamic range images, which typically have long-tailed histograms,
using the mean signal to compute dynamic range may not be the most sensible
approach, since small but very bright highlights or light sources tend to have little
impact on this measure. Thus, rather than use σs or μs to compute one of the above
SNR measures, it may be a better approach to compute the peak signal-to-noise
ratio, which depends on the maximum luminance Lmax of the image:
 
(Lmax )2
SNRpeak = 20 log10 . (15.13)
σn

While μs , σs , and Lmax are each readily computed for any given image, the
estimation of the noise floor σn in an image is a notoriously difficult task, for
which many solutions exist. In particular, assumptions will have to be made about
the nature of the noise.
In some specific cases, the noise distribution may be known. For instance,
photon shot noise, which is characterized by the random arrival times of photons
at image sensors, has a Poisson distribution. For a sensor exposed for T seconds,
and a rate parameter of ρ photons per second, the probability P of detecting p
photons is given by
(ρ T ) p e−ρ T
P(p) = . (15.14)
p!
The mean μn and standard deviation σn are then

μn = ρ T, (15.15a)

σn = ρ T . (15.15b)

i i

i i
 i i

i i

15.2. Dynamic Range 819

Thus, for photon shot noise, the signal-to-noise ratio can be computed by [1282]

SNRPoisson = 10 log10 (ρ T ), (15.16)

which is again measured in decibels.

When the source of the noise in an image is unknown, further assumptions
about the noise will have to be made before a signal-to-noise ratio can be com-
puted. In particular, a model of the distribution of the noise must be assumed. In
the absence of further information, one may, for instance, assume that the noise
has zero mean and has a Gaussian distribution with a standard deviation of σn .
Under this assumption, one could apply a Gaussian filter to the image L with a
kernel width of σn , producing a second, blurred image Lblur . Alternatively, rather
than blurring the image, a median filter may be applied that minimizes excessive
blur and, thereby, potentially yields a more accurate measure of dynamic range. It
would also be possible to apply more advanced noise reduction algorithms, such
as level set methods [727].
An indication of the noise floor is then given by the pixel-wise mean squared
difference between the blurred and original images:

1 N  2
σn = ∑ Li − Liblur .
N i=0
(15.17)

The accuracy of the noise floor and, therefore, the measure of dynamic range,
is dependent on the noise reduction technique, as well as on the filter parame-
ter settings chosen. This makes these techniques limited to comparisons of the
dynamic range between images that are computed with identical techniques and
parameter settings.
On the other hand, the signal-to-noise ratio may be computed by relying on
the minimum and maximum luminance (Lmin and Lmax ):
 max 
L
SNRmax = 20 log10 . (15.18)
Lmin

In this case, there is no dependency on the choice of noise-reduction technique.

Instead, there is a dependency on Lmin and Lmax , which are then taken to be rep-
resentative of the signal and noise levels. However, by definition, these values are
outliers. As this is a logarithmic measure, its sensitivity to outliers is somewhat
ameliorated compared with the common approach of simply listing the dynamic
range as Lmax /Lmin .
Finally, whether an image’s dynamic range is measured in terms of signal-
to-noise ratios or simply by comparing the maximum and minimum luminance,

i i

i i
 i i

i i

820 15. Image Properties and Image Display

there is always a trade-off between accuracy and sensitivity to outliers or param-

eter settings. In the following section we show measurements performed on high
dynamic range images using the metrics outlined in this section.

15.2.2 Gaussian Filter

In this section we show, for both the Gaussian and median filters, how their filter-
parameter settings affect the dynamic range measurements. The example images
used are shown in Figure 15.7.
For Gaussian filters, the kernel is specified with two parameters that both de-
termine the result. First, the width of the Gaussian filter is determined by σ , such
that for an image region centered around the origin, the kernel is given by

1 x +y
2 2
−
Rσ (x, y) = e 2σ2 . (15.19)
2π σ2

Figure 15.7. Example images used for assessing the impact of filter-kernel parameters
on SNR measures; Tower of Hercules Roman Lighthouse, A Coruña, Spain, August 2005;
Rennes, France, July 2005; Insel Mainau, Germany, June 2005.

i i

i i
 i i

i i

15.2. Dynamic Range 821

As any Gaussian filter has infinite non-zero extent, a second parameter determines
into how many pixels the kernel is discretized. This number, k, is typically odd.
While the continuous Gaussian filter has unit area, after discretization, the filter
no longer sums to 1. We therefore normalize the discretized filter kernel before
starting the computations.
As an example, for k = 3, 5, 7, the filter kernels are
⎡ ⎤
⎡ ⎤ 0.003 0.013 0.022 0.013 0.003
0.075 0.124 0.075 ⎢0.013 0.060 0.098 0.060 0.013⎥
⎢ ⎥
⎣0.124 0.204 0.124⎦ ; ⎢0.022 0.098 0.162 0.098 0.022⎥ ;
⎢ ⎥
0.075 0.124 0.075 ⎣0.013 0.060 0.098 0.060 0.013⎦
0.003 0.013 0.022 0.013 0.003
⎡ ⎤
0.000 0.000 0.001 0.002 0.001 0.000 0.000
⎢0.000
⎢ 0.003 0.013 0.022 0.013 0.003 0.000⎥⎥
⎢0.001 0.013 0.059 0.097 0.059 0.013 0.001⎥
⎢ ⎥
⎢0.002 0.002⎥
⎢ 0.022 0.097 0.159 0.097 0.022 ⎥.
⎢0.001 0.013 0.059 0.097 0.059 0.013 0.001⎥
⎢ ⎥
⎣0.000 0.003 0.013 0.022 0.013 0.003 0.000⎦
0.000 0.000 0.001 0.002 0.001 0.000 0.000
(dB RMS)

32
Hercules Roman Lighthouse
30
dep

28
SNR

26

24

22 σ=1
σ=2
20 σ=3
σ=4
18 σ=5
σ=6
16
0 5 10 15 20 25 30 35
k

Figure 15.8. Different values for Gaussian filter-kernel parameters k and σ computed
using the Tower of Hercules image for the SNRdep metric.

i i

i i
 i i

i i

822 15. Image Properties and Image Display

26

(dB RMS)
Hercules Roman Lighthouse
24

indep
22

SNR
20

18
σ=1
16 σ=2
σ=3
σ=4
14 σ=5
σ=6
12
0 5 10 15 20 25 30 35
k

Figure 15.9. Different values for Gaussian filter-kernel parameters k and σ computed
using the Tower of Hercules image for the SNRindep metric.

44
(dB)

Hercules Roman Lighthouse

peak

42
SNR

40

38

36

34 σ=1
σ=2
32 σ=3
σ=4
30 σ=5
σ=6
28
0 5 10 15 20 25 30 35
k

Figure 15.10. Different values for Gaussian filter-kernel parameters k and σ computed
using the Tower of Hercules image for the SNRpeak metric.

i i

i i
 i i

i i

15.2. Dynamic Range 823

To show the impact of discretization, the k parameter is increased from 3 in

steps of 2 for a Gaussian filter with widths σ = 1 . . . 6, producing a sequence of
blurred images Lk, blur
σ . These images are then used to compute the various SNR
metrics outlined in the previous section. The results for the Tower of Hercules
image of Figure 15.7 are shown in Figures 15.8, 15.9, and 15.10.
The trends that can be observed in these figures are typical, also for the other
two example images. It is expected that if the image is blurred more, by using
wider filter kernels, the difference between the blurred and unblurred versions
becomes larger, and this increases the measured noise floor. As a result, the mea-
sured dynamic range will be lower. A second effect that is to be expected with
this approach is that when a larger filter kernel is used, a larger number of samples
is required before the results become stable. These figures thus highlight the fact
that all three measures depend on the choice of parameter settings.

15.2.3 Median Filter

In computer vision, median filters are often used to remove noise in images. The
median filter may therefore also be employed to aid in the computation of the
dynamic range of an image.
The chief advantage of a median filter over a Gaussian filter is that it does not
tend to blur over edges. It therefore tends to leave the signal intact, while only
removing noise. This is achieved by considering a neighborhood of k by k pixels,
sorting the values in ascending order, and selecting the median value. Each pixel
is thus replaced by the median value of a neighborhood of pixels.
This leaves one user parameter to specify the size of the local neighborhood.
We have measured the signal-to-noise ratios as before, but now on the basis of
the median filter. For the three example images, the SNR measures are given in
Figures 15.11 through 15.13.
These figures show that wider filter kernels reduce each of the SNR measures
before settling into a stable value. However, when a stable value is reached, the
image obtained by applying a median filter with a large kernel will have deterio-
rated significantly. An example is shown in Figure 15.14. As a result, the SNR
measures that are based on a median filter affect both the noise as well as the sig-
nal. We therefore do not recommend using large filter kernels to measure noise.
In the following, we will therefore use a value of k = 3 when a median filter is
used for SNR measures.

i i

i i
 i i

i i

824 15. Image Properties and Image Display

45

SNR (dB)
Hercules Roman Lighthouse
40 peak
SNR
dep
SNR
35 indep
SNR

30

25

20

15

10
0 5 10 15 20 25 30 35
k

Figure 15.11. SNR measures plotted as a function of the median-filter parameter k for the
Tower of Hercules image.
SNR (dB)

60
Church Tower

50 peak
SNR
dep
SNR
indep
SNR
40

30

20

10

0
0 5 10 15 20 25 30 35
k

Figure 15.12. SNR measures plotted as a function of the median-filter parameter k for the
Church Tower image.

i i

i i
 i i

i i

15.2. Dynamic Range 825

50
SNR (dB)

Trees
45

40

35

30 peak
SNR
dep
25 SNR
indep
SNR
20

15

10

0
0 5 10 15 20 25 30 35
k

Figure 15.13. SNR measures plotted as a function of the median-filter parameter k for the
Tree image.

Figure 15.14. Luminance values after filtering with a median filter with kernel size k = 3
(left) and k = 15 (right).

i i

i i
 i i

i i

826 15. Image Properties and Image Display

SNR (dB)
180
Tower of Hercules
160
Church Tower
Trees
140

120

100

80 Gaussian filter Median filter

{
{
60

40

20

SNR Gauss
SNR indep

SNR indep
peak

peak

SNR peak
SNR max
SNR dep

SNR dep
SNR

Figure 15.15. Different signal-to-noise metrics compared against each other.

15.2.4 Dynamic Range Measurements

To enable further comparison, we select a relatively small kernel size (σ = 1) for
Gaussian filters, because the amount of noise present in these images is small.
For this filter, the value of k needs to be at least 7 to obtain stable results. In the
following, the median filter is applied with a kernel size of k = 3 to minimize
affecting the signal, and thereby the signal-to-noise ratio.
We also compare two different version of SNRpeak . The first is based on Lmax ,
whereas the second is based on a filtered version of Lmax . The filter used here is
a Gaussian filter with σ = 1 and k = 7. Such a small filter kernel should remove
single-pixel outliers, and is therefore likely to be a more stable measure for the
peak of the signal than the unfiltered maximum luminance.
Given these parameter settings, the various SNR measures are compared for
all three example images in Figure 15.15. There are several observations to be
made on the basis of this comparison. First, it can be seen that the SNR measure
based on the minimum and maximum luminance overestimates the dynamic range
in the images. In addition, SNRmax shows the largest range for the tree image,
which can be attributed to a spurious pixel. It is therefore not a good measure
to use in general. This also means that reporting the dynamic range as the ratio
of Lmax /Lmin is not a sensible approach (despite the common practice of using
exactly this measure).

i i

i i
 i i

i i

15.3. Cross-Media Display 827

On the other hand, the RMS-based measures rely on average values in the
image. This results in a tendency to underestimate the dynamic range, as small
but very bright features do not contribute much to this measure. However, it is
precisely these features that have a significant impact on the dynamic
range.
The best approaches therefore appear to be afforded by the peak signal-to-
noise measures. Note that each of these measures, whether guided by a Gaussian
or a median filter, rate the night scene to have the highest dynamic range. This is
commensurate with visual inspection, which would rank this scene first, followed
by the Tower of Hercules, and then the tree scene (which is expected to have the
lowest dynamic range of the three example images).
The Gaussian-based peak-SNR measures provide very similar results, inde-
pendent of whether the image is first blurred by a small filter kernel prior to deter-
mining Lmax . The median-based peak-SNR metric shows a somewhat higher dy-
namic range, presumably because this filter is better able to leave edges
intact.
As such, of the filters and measures included in this comparison, we would
advocate the use of a median filter-based measure of dynamic range, employing a
peak-SNR measure, rather than a RMS measure.

15.3 Cross-Media Display

In many cases, the origins of an image are unknown. This means that even if the
display device and the viewing environment have been appropriately character-
ized and corrected for, the image may not appear as intended. For color-accurate
reproduction of images, we need several pieces of information about the image, as
well as about the display device and its environment. These include display pri-
maries, white point, gamma encoding, as well as a characterization of the display
environment.
In the following sections, we assume that the original image is defined in
a display-defined space, rather than a scene-referred space. As a result, in the
preparation of an image for a particular display, the (non-linear) processing ap-
plied to the image must first be undone before the processing specific to the target
display can be applied.

15.3.1 LDR to LDR Display Pipeline

A low dynamic range (LDR) image will be in a given color space, including its
associated gamma encoding. The first step toward display of such an image is to

i i

i i
 i i

i i

828 15. Image Properties and Image Display

Input/ Device dependent View environment Common

output transforms dependent transforms subspace

Original Inverse gamma Color space Forward Image

(LDR) (linearize) transform appearance model enhancement

Reproduction Gamma Color space Backward Gamut

(LDR) correction transform appearance model mapping

Figure 15.16. The processing steps necessary to prepare a low dynamic range image for
display on a specific display device.

undo all the device-dependent encoding steps and bring the image into a common
subspace where further processing may occur to account for a specific rendering
intent. In addition, gamut mapping may occur in this subspace. The image is
then ready to be transformed into the device-dependent space associated with the
display device for which the image is being prepared. An overview of the steps
involved is given in Figure 15.16 [795].
The first step in the transform from original low dynamic range image to the
device-independent subspace is to linearize the data by undoing the gamma en-
coding. With known primaries and white point, the image can then be transformed
into a device-independent color space (see Chapter 8). A color appearance model
may then be applied to account for the viewing environment for which the image
was originally intended (see Chapter 11).
The image is now in a common subspace, where further image processing may
be applied, such as enhancing for accurateness, pleasantness, etc., depending on
the reason for displaying a particular image; this step may be omitted depending
on rendering intent.
The next step in the common subspace is to account for the difference in
color gamut between the display device and the gamut defined by the image’s
primaries. In particular, if the color gamut of the display device is smaller than
the image’s, as for example in printing devices, the range of colors in the image
must be shoe-horned into a smaller gamut. It should be noted here that source
and target gamuts may overlap, so that in practice gamut compression may be
applied for some colors and gamut expansion may be applied for other colors,
even within the same image. Algorithms to accomplish this are called gamut
mapping algorithms and are discussed in Section 15.4.

i i

i i
 i i

i i

15.3. Cross-Media Display 829

After gamut mapping, the linear, enhanced, and gamut-mapped image must
be encoded for the display device on which the image is to be displayed. The first
step is to apply a color appearance model in backward direction, using appearance
parameters describing the viewing environment of the display. The image is then
transformed into the display’s device-dependent color space, followed by gamma
correction to account for the non-linear relationship between input values and
light intensities reproduced on the display..
The human visual system’s adaptability is such that if one or more of the above
steps are omitted, chances are that inaccuracies go largely unnoticed. Many ap-
plications are not color critical, and this helps ameliorate the effects of inaccurate
color reproduction. However, if two color reproductions are shown side-by-side,
for instance when a print is held next to a computer monitor, large to very large
differences may become immediately obvious. For accurate cross-media color
reproduction, all the steps summarized in Figure 15.16 need to be executed.

15.3.2 HDR to LDR Display Pipeline

In the case of high dynamic range images, the steps required to prepare for display
are different. To begin, HDR images are linear by construction, so that lineariza-
tion is not required. The sequence of events leading to color-accurate reproduction
of high dynamic range images is currently not fully understood, but is an active
area of research. However, it is believed that the sequence of steps outlined in
Figure 15.17 will lead to the acceptable reproduction of HDR images.
We will assume the HDR data to be in a known device-independent color
space. However, this is by no means guaranteed, and color space transforms
may have to be applied before executing further device-independent processing.
Image processing for rendering intent of HDR images is currently not an active
research area. The same is true for gamut mapping of HDR images. As a result,
the application of standard image processing and gamut mapping to HDR images
is currently unexplored territory. The issue of HDR color is further discussed in
Section 17.10.1.

Input Common subspace View environment Device dependent Output

dependent transform transforms

Original Image Gamut/Tone Appearance Color space Gamma Reproduction

(HDR) enhancement mapping model transform correction (LDR)

Figure 15.17. The processing steps involved to display a high dynamic range image on a
specific display device.

i i

i i
 i i

i i

830 15. Image Properties and Image Display

Input Common View environment Device dependent Output

subspace dependent transform transforms

Original Image Appearance Color space Gamma Reproduction

(HDR) enhancement model transform correction (HDR)

Figure 15.18. The processing steps involved to display a high dynamic range image on a
high dynamic range display device.

The range of luminance values of an HDR image tends to be much larger

than the display range of current common display devices. Thus, recently much
effort has been expended on developing techniques to reduce the dynamic range
of HDR images for the purpose of displaying them on low dynamic range display
devices [952]. Dynamic range reduction algorithms are known as tone mapping
or tone reproduction operators. They are the topic of Chapter 17.
A color appearance model should then be applied to account for the display
environment. Here, a further interesting complication occurs, as most color ap-
pearance models were developed to account for changes in viewing environment
over a relatively small range of luminance values. Whether color appearance
models can be extended to operate over much larger luminance ranges is currently
unknown.
The final two steps, a device-dependent color space transform and gamma
correction, are identical for both low dynamic range and high dynamic range
content.

15.3.3 HDR to HDR Display Pipeline

The emergence of high dynamic range display devices is further complicating the
issue of color-accurate rendering. In theory, high dynamic range image content
can be displayed directly without the need for tone reproduction, as shown in Fig-
ure 15.18. However, if the dynamic range of the input image exceeds the display
capability, then some dynamic range reduction may still be needed. A much larger
unknown in this sequence of steps is that color appearance modeling for high dy-
namic range display devices will require a substantially different approach. The
reason for this is that under conventional lighting conditions (assume office light-
ing) the human visual system will partially adapt to the room lighting and partially
to the display device. Adaptation to the display device is not an issue normally
considered for conventional display devices, as their light output is of the same
order as the lighting in a normal office environment. It is currently unknown how
to account for partial adaptation to the display device.

i i

i i
 i i

i i

15.3. Cross-Media Display 831

Input/ Device dependent View environment Common

output transforms dependent transforms subspace

Original Inverse gamma Color space Forward Image

(LDR) (linearize) transform appearance model enhancement

Dynamic range
expansion?

Reproduction Gamma Color space Backward Gamut

(HDR) correction transform appearance model mapping

Figure 15.19. The processing steps involved to display a low dynamic range image on a
high dynamic range display device.

15.3.4 LDR to HDR Display Pipeline

Finally, the display of low dynamic range images on a high dynamic range dis-
play requires careful consideration. As the input image now has a much lower
dynamic range, an extra step involving range expansion may be required. The
reason that such a step may be necessary is that the low dynamic range image
may be a representation of a high dynamic range environment. The image may
have originally been captured in high dynamic range and subsequently stored in
a low dynamic range file format. Alternatively, the image may be created with
conventional techniques, possibly exhibiting under- and/or over-exposed areas.
In each of these cases, though, a more accurate reproduction may be obtained by
expanding the range in a suitable way. However, dynamic range expansion for
the display of conventional images, i.e., legacy content, on high dynamic range
display devices is only beginning to be explored.
Nonetheless, a relatively consistent picture is starting to emerge. User studies
have shown that the perceived image quality is positively affected by increasing
both the contrast and the absolute luminance levels [1022, 1279]. Increasing both
simultaneously yields the largest increase in perceived quality, although partic-
ipants tend to prefer an increase in luminance level first, and only if the mean
luminances of two images are the same, then the image with the higher contrast
is preferred [26].
Further, Seetzen et al. have found a relationship between the perceived qual-
ity of images and the peak luminance and contrast ratio afforded by the dis-
play [1022]. For a given contrast ratio, the perceived image quality rises with
peak luminance level up to a maximum, after which it falls again. They fit the

i i

i i
 i i

i i

832 15. Image Properties and Image Display

following equation to their psychophysical data:

 
Q(Lpeak ) = c1 Lpeak exp −c2 Lpeak , (15.20)

where Lpeak is the display’s peak luminance level (in cd/m2 ) and Q is the associ-
ated perceived image quality. The constants c1 and c2 relate to the optimal peak
luminance level Lopt and the corresponding maximum perceived image quality
Qmax as follows:
1
Lopt = , (15.21a)
b
a
Qmax = exp(−1) . (15.21b)
b
Values of Lopt and Qmax for various contrast ratios are given in Table 15.1. Further,
it was found that the optimal contrast ratio ΔLopt as a function of peak luminance
Lopt can be broadly modeled with the following relation:
 
ΔLopt ≈ 2862 ln Lopt − 16283. (15.22)

All studies suggest that simply emulating LDR display hardware when dis-
playing legacy content on HDR displays would be a sub-optimal solution. In-
creasing luminance and contrast would be preferred. Algorithms to achieve ap-
propriate up-scaling are termed inverse tone reproduction operators. Fortunately,
provided that the LDR image is properly exposed and does not contain compres-
sion artifacts, the procedure for up-scaling can be straightforward. In fact, user
studies indicate that linear up-scaling is adequate and is preferred over non-linear
scaling [26].
In line with these findings, Rempel et al. propose a linear scaling for pixels
representing reflective surfaces and a non-linear scaling for pixels representing
light sources and other scene elements that are perceived as bright [955]. A similar
technique was proposed specifically for the treatment of light sources [776, 778].

Contrast Lopt Qmax

2500 856 0.4756
5000 1881 0.5621
7500 3256 0.6602
10000 12419 1.0000

Table 15.1. Optimal display peak luminance values and corresponding maximum per-
ceived image quality for various contrast ratios (after [1022]).

i i

i i
 i i

i i

15.4. Gamut Mapping 833

In video encodings, values over 235 (the white level) are typically assumed to
represent light sources or specular highlights, and this value is therefore used to
determine which type of up-scaling is used. Given that the conversion from low
dynamic range to high dynamic range values contains two segments, there is a risk
of introducing spatial artifacts when neighboring pixels are scaled by different
methods. These effects can be mitigated by applying spatial filtering [955]. If
the input image was not linear to begin with, it is first linearized by undoing the
gamma encoding.
Input images are not always well-behaved; for instance they may contain
under- or over-exposed areas. In certain cases, it may be possible to apply texture
synthesis techniques to infer and re-synthesize the missing image areas [1207].
In addition, images may have been encoded using lossy compression. In this
case, the up-scaling to a higher display range may amplify the artifacts that were
initially below visible threshold to values that are easily perceived. Additional
processing is therefore likely to be necessary to deal with problem images.

15.4 Gamut Mapping

One of the attributes of a color space is its gamut, the volume of colors that can
be reproduced. The boundary of a color space, the gamut boundary, is a three-
dimensional shape which is different for different color spaces. Given that each
device has its own color space, it is evident that while we can expect consider-
able overlap between the gamut of different devices, there are ranges of colors in
one device that cannot be reproduced by another device [795, 798, 1095]. As an
example, Figure 8.1 shows that even for a single luminance value, the triangular
boundaries imposed by several RGB color spaces are different.
In practice, this means that the conversion between two RGB color spaces
may yield colors that cannot be represented in the target color space. Starting, for
example, with the Adobe RGB color space which has a relatively large gamut,
conversion to the smaller gamut of the sRGB color space tends to yield problem
colors, as shown in Figure 15.20. Even larger differences in color gamut exist
between display devices and printing devices. In particular, the CMYK gamut
used by softcopy display devices is much smaller than typical RGB spaces.
However, it should be noted that even if the 3D volume spanned by the gamut
boundaries leads to a smaller volume, the print gamut may be displaced with
respect to the display gamut yielding non-overlapping regions. The print gamut
may therefore include colors that the display gamut cannot reproduce (and vice
versa). The implication is that gamut mapping algorithms will have to compress

i i

i i
 i i

i i

834 15. Image Properties and Image Display

Figure 15.20. The images on the left are assumed to be specified in the Adobe RGB color
space. Conversion to the sRGB color space creates some pixels that are out-of-gamut.
These are encoded as red pixels on the right. Top: Castle Coombe, UK, September 2006.
Bottom: Stained glass window in Merchant Venturers Building, Bristol, UK, March 2006.

some colors and expand others. The discussion in this section outlines the issues
related to gamut mapping, and follows the work by Ján Morovic [795, 797, 798].
To achieve a consistent output across media it is therefore necessary to map an
image from one color space into a different one. Simple color space conversion
is appropriate for colors that are within the overlapping region in both source and
target color spaces. Colors that are outside the target color gamut need to be
mapped to colors that are within the target device’s gamut. The gray region in
Figure 15.21 indicates the set of colors that cannot be reproduced given that the
two example gamuts only partially overlap.
Remapping out-of-gamut colors to colors that are exactly at the boundary of
the target device’s gamut is called gamut clipping. This approach has the ad-
vantage that most colors remain unaffected and will therefore be accurately re-
produced. However, if only out-of-gamut colors are remapped, then undesirable
visible discontinuities may be introduced at the boundaries.
An alternative approach is to warp the entire gamut, thereby distorting all
colors. This is called gamut mapping. The aim is then to keep the distortion
of all colors to a minimum and, hopefully, below the visible threshold. Such an
approach avoids jarring discontinuities, however, the trade-off lies in the fact that
all colors will be distorted by some amount.

i i

i i
 i i

i i

15.4. Gamut Mapping 835

Original Gamut Original Gamut

Reproduction Gamut Reproduction Gamut
0.7 0.7 Effective Gamut
v’ v’ Out-of-gamut colors
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.0 0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
u’ u’

Figure 15.21. Partially overlapping gamuts (left) cause the colors located in the gray
region to be out-of-gamut.

Various alternatives to these two extremes exist. For instance, in CIELAB

space, it is possible to keep the luminance values unaffected and only warp a∗
and b∗ values. It is also possible to clip some values, leave others unaffected, and
have a non-linear mapping for intermediate values.
Gamut mapping and gamut clipping algorithms require several stages of com-
putation. First, the boundary of the reproduction gamut needs to be established.
Second, a color space is to be selected in which the mapping is carried out (such
as CIELAB). The type of mapping spans a further range of algorithms. While
gamut clipping is relatively straightforward, there are many approaches to gamut
mapping.
The image to be reproduced is given in a particular color space. However,
the image’s pixels form a color gamut that may be a subset of this color space’s
gamut. Hence, it may be beneficial to compute the image’s gamut and map that
to the reproduction device’s gamut.

15.4.1 Aims of Gamut Mapping

When converting an image between different media, for instance by preparing an
sRGB image for printing on a CMYK printer, the most appropriate gamut map-
ping algorithm may well depend on the application for which the image is printed.
In general, the aim of gamut mapping is to ensure a good correspondence of over-
all color appearance between the original and the reproduction by compensating
for the mismatch in the size, shape and location between the original and repro-
duction gamuts [795, 797].

i i

i i
 i i

i i

836 15. Image Properties and Image Display

Currently, models to quantify the appearance of images are still in their in-
fancy, although progress in this area is being made (see Chapter 11). This makes
a direct implementation of the above definition of gamut mapping more compli-
cated. Instead, a set of heuristics for gamut mapping have been identified; they
are listed here in decreasing order of importance [721, 795]:

Ensure that gray values are mapped to gray values. This means that achromatic
values in the original gamut should remain achromatic in the reproduction
gamut.

Maximize luminance contrast. As the gamut is typically reduced, some of the

loss in contrast can be recovered by using the entire range of luminance
values in the reproduction gamut. Even if this involves a shift in mean lu-
minance, it is desirable to map the image’s black and white points to the
minimum and maximum reproducible luminance.

Reduce the number of out-of-gamut colors. Although ideally all the colors are
brought within range, the overall appearance of an image can be improved
by clipping some values against the boundaries of the reproduction gamut.
In the extreme case, all the out-of-gamut colors are clipped, although this
is unlikely to be the optimal case.

Minimize hue shifts. It is desirable to leave the hue of colors unmodified.

Increase saturation. Given that the reproduction gamut is smaller than the origi-
nal gamut, it is often preferable to maximize the use of this smaller gamut,
even if this means increasing saturation. This is similar to maximizing lu-
minance contrast, albeit along a different dimension.

Maintain the relationship between colors [1095]. If the mapping of colors

requires shifts in luminance, saturation, and/or hue, then it is important
to shift related colors in similar ways.

It should be noted that this list of requirements, and their ordering, is largely based
on practice and experience in color reproduction.

15.4.2 Gamut Boundaries

Before any gamut mapping can take place, it is necessary to compute the bound-
aries of the reproduction gamut. Dependent on the device, this can be achieved
by means of the Kubelka-Munk equations [283] (see Section 3.4.6) or the Neuge-
bauer model [725]. The most common approach, however, is to use sampling
algorithms to derive gamut boundaries. Further methods for computing either

i i

i i
 i i

i i

15.4. Gamut Mapping 837

media gamuts or image gamuts have also been discussed in the literature [176,
528, 621, 796, 989].
To achieve effective gamut clipping, as well as gamut mapping, an algorithm
to intersect a line along which the mapping is to be carried out with the calcu-
lated gamut must be applied. Such problems can be solved with techniques akin
to ray tracing [1042], although specific gamut intersection algorithms have been
proposed as well [124, 467, 468, 794].

15.4.3 Color Spaces

Gamut mapping algorithms are commonly designed to preserve some perceptual
attributes, while changing others, with the intended end result that all colors fall
within the reproduction gamut. In addition, the visible difference between the
original image and the reproduced image should be minimal.
Perceptual attributes, including hue, chroma, saturation, brightness, lightness,
and colorfulness, are usually computed on the basis of a device-independent color
space such as CIE L∗ a∗ b∗ or CIE L∗ u∗ v∗ , or on the basis of a color appearance
model such as CIECAM97s.
The quality of gamut mapping then depends, in part, on how accurate the
predictors for these perceptual attributes are. For instance, CIE L∗ a∗ b∗ has defi-
ciencies in the uniformity of its hue angle predictor in the blue region, i.e., around
hue angles of 290◦ . Under ideal circumstances, the predictors for the different
perceptual attributes should be independent. However, in this case, sets of col-
ors that differ only in luminance or chroma may also cause differences in their
hue prediction. Problems in this regard are smaller with the CIECAM97s color
appearance model [795].

15.4.4 Gamut Clipping

Gamut clipping algorithms only affect colors that are within the boundaries of
the original gamut, but outside the reproduction gamut. These colors are moved
to lie on the reproduction gamut boundary. Colors that already lie within the
reproduction gamut are unaffected.
Two variations of gamut clipping are shown in Figures 15.22 and 15.23.
Note that although we show gamut clipping in a two-dimensional chromaticity
diagram, in practice gamut clipping (and gamut mapping in general) is a three-
dimensional problem. We only show two-dimensional examples because these
are simpler to illustrate.
Figure 15.22 shows an example where out-of-gamut values are moved along
the color axes until they reach the boundary. If the gamut boundary is aligned

i i

i i
 i i

i i

838 15. Image Properties and Image Display

Original Gamut
Reproduction Gamut
Effective Gamut
Out-of-gamut colors
Image color
Reproduction color
Image and reproduction color

Figure 15.22. A simple gamut clipping scheme, whereby out-of-gamut colors are moved
toward the reproduction gamut boundary along the axes of the color space in which gamut
clipping is executed.

with the color space’s axes, then this constitutes the smallest possible shift and,
therefore, the smallest possible distortion in the final result.
Gamut boundaries, however, rarely ever align with the axes of any color space.
A more accurate gamut clipping approach therefore shifts out-of-gamut colors
along a direction orthogonal to the nearest gamut boundary, as shown in Fig-
ure 15.23.

Original Gamut
Reproduction Gamut
Effective Gamut
Out-of-gamut colors
Image color
Reproduction color
Image and reproduction color

Figure 15.23. Gamut clipping where out-of-gamut colors are moved toward the nearest
point on the reproduction gamut boundary.

15.4.5 Gamut Compression

Gamut mapping strategies, other than gamut clipping, will affect colors that are
both inside and outside the reproduction gamut. It may be necessary to remap col-

i i

i i
 i i

i i

15.4. Gamut Mapping 839

Reproduction value

Out-of-gamut values
Gamut clipping
Polynomial
Piece-wise linear
Linear
Original value

Figure 15.24. Gamut compression types (adapted from [795]).

ors inside the reproduction gamut to maintain their relationship with neighboring
colors which may be outside the reproduction gamut. This means that all colors
are remapped by some amount. Typically, colors that are outside the reproduc-
tion gamut are remapped over larger distances than colors inside the reproduction
gamut. Mappings can be classified either as linear, piece-wise linear, polynomial,
or a combination of these methods. Example remapping functions are shown in
Figure 15.24.

15.4.6 Mapping Directions

Once a particular type of mapping is chosen, the gamut boundaries of the original
and reproduction gamuts are known, and the direction of the mapping needs to
be chosen. This direction can be chosen to be towards a specific point, such as
the lightness value for which a cusp occurs, as shown in the top left plot of Fig-
ure 15.25. Alternatively the middle value on the lightness axis (in L*C*h* space
for instance) can be chosen as the point toward which the mapping is directed
(middle figure on the top row of Figure 15.25).
A different strategy remaps colors such that either lightness and hue, or satura-
tion and hue remains constant (top right and bottom left of Figure 15.25). Finally,
the direction of the mapping can be toward the nearest point on the boundary of
the reproduction gamut (Figure 15.25, bottom right).

15.4.7 Comparison of RGB Spaces

The range of colors that can be reproduced in each RGB color space depends on
the chosen primaries. Converting colors from one space to another is frequently
required, for instance to prepare an image for display on a particular device. The

i i

i i
 i i

i i

840 15. Image Properties and Image Display

Cusp Center Constant Lightness

100 100 100
Lightness

Lightness

Lightness
80 80 80

60 60 60

40 Cusp 40 Cusp 40 Cusp

20 20 20

0 0 0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Chroma Chroma Chroma

Constant Saturation Minimum ΔE

100 100
Lightness

Lightness
80 80
Original Gamut
Reproduction Gamut 60 60
Image color
Reproduction color 40 Cusp 40 Cusp
20 20

0 0
0 20 40 60 80 100 0 20 40 60 80 100
Chroma Chroma

Figure 15.25. Gamut mapping strategies (adapted from [795]).

performance of different RGB color spaces has been measured in an experiment

by Henry Kang [576,577]. Prints are made of 150 patches of color using an inkjet
printer. The prints are then measured under D65 illumination to obtain CIE XYZ
tristimulus values, which are converted to CIE LAB. This ensures that the test
set covers a range of colors that are commonly encountered, as they can all be
reproduced on standard printers.
The measured XYZ values are then converted to one of the RGB color spaces
with their associated non-linear encodings. The RGB data thus obtained is subse-
quently converted back to CIE LAB, for comparison with the measured CIE LAB
values. Pairs of CIE LAB values are then subjected to the ΔEab color-difference
metric (see Section 8.7.1) to assess the magnitude of the error for each RGB
space/encoding. The percentage of colors that ended up being out-of-range after
the RGB encoding process, as well as the average and maximum ΔEab errors are
listed in Table 15.2.
If monochromatic lasers are used as primaries, then the vertices of the triangle
spanned in a chromaticity diagram lie on the spectrum locus, thus giving a much

i i

i i
 i i

i i

15.5. Gamma Correction 841

RGB space Percentage out-of-range Average ΔEab Maximum ΔEab

SMPTE-C 30.0% 2.92 27.42
sRGB 25.3% 2.22 28.26
Adobe RGB (1998) 10.0% 1.17 19.83
NTSC 6.0% 0.87 12.19
SMPTE-240M 5.3% 0.74 9.60

Table 15.2. Percentage of out-of-gamut errors accumulated for different color spaces
(after [577]).

wider gamut. This would result in fewer out-of-gamut problems. For instance, us-
ing monochromatic primaries with xy chromaticity coordinates (0.7117, 0.2882),
(0.0328, 0.8029), (0.1632, 0.0119) for red, green and blue, respectively, yields
5.3% of out-of-range colors with average and maximum errors of 1.05 and 28.26,
respectively. Although the average error is now much smaller, the maximum error
remains large. This is due to the fact that the choice of primaries is important. In
this case, the primaries are such that the line between the green and red primaries
excludes a significant portion of yellow colors [577].
Finally, it is possible to select primaries that are outside the spectral locus. An
example is the RIMM-ROMM space [1072]. In this color space, the red primary is
at 700 nm, but the green and blue primaries are outside the spectral locus, having
xy chromaticity coordinates of (0.7347, 0.2653), (0.1596, 0.8404), and (0.0366,
0.0001). In the above experiment, no patches were out-of-gamut, and the average
and maximum ΔEab errors are 0.71 and 4.29, respectively [577]. Although it is
clear that such a color space cannot be realized with any physical device, it is a
useful intermediate color space, as it largely avoids gamut conversion problems.

15.5 Gamma Correction

Normally, final step in the preparation of an image for display is gamma correc-
tion. This step attempts to correct for the non-linear relationship between input
voltage and strength of the light emitted by the phosphors of a cathode ray tube
(CRT), known as the electro-optic conversion curve (EOCF).
This non-linearity is frequently well modeled by a power law whereby the
exponent is given by the constant γ . This is especially true for CRTs. On the other
hand, LCDs can have very different EOCFs, and may not necessarily be modeled
by a power law, unless corrective look-up tables are included in the design.

i i

i i
 i i

i i

842 15. Image Properties and Image Display

Figure 15.26. Gamma chart. Displaying this chart on a monitor enables the display
gamma to be estimated (after an original idea by Paul Haeberli).

To correct for the non-linear response of power-law displays, gamma correc-

tion is applied using:
1/γ
Ld = Ld , (15.23)

where γ is the gamma value associated with the display device [923].
The EOCF of a monitor can be measured with a photometer. However, under
the assumption of power law behavior, it is also possible to estimate the gamma
exponent by displaying the image shown in Figure 15.26. The alternating black
and white scanlines on one side should produce a middle gray as these lines are
fused by the human visual system. This middle gray can be matched to an entry in
the chart on the other side, which consists of patches of gray of varying intensities.
The best match in the chart is the gray value that the monitor maps to middle gray.
From the actual value in that patch p ∈ [0, 1] the gamma exponent of the monitor
can be derived:
log(0.5)
γ= . (15.24)
log(p)

i i

i i
 i i

i i

15.6. Ambient Light 843

15.6 Ambient Light

After gamma correction, there are two further complications that should be ad-
dressed in critical applications. The appearance of an image is affected by both
the state of adaptation of the observer—a problem addressed by color appearance
modeling—and also by the fact that display devices are made of reflective mate-
rials. The former is discussed in detail in Chapter 11, while the effect of ambient
reflections is discussed in this section.
Since a typical display device is located in a lit environment, such as a home
or an office, some of that light falls on the monitor is reflected towards the viewer.
This ambient light could be crudely modeled by adding a constant value to the
light that is emitted from the screen. If Ld is the light emitted by the display
device, then Ld is the light that reaches the observer:

Ld = Ld + LR , (15.25)

where LR is the amount of reflected ambient light. While this component can once
more be measured with a photometer with the display device switched off, there
also exists a technique to estimate the reflections off a screen, similar to reading a
gamma chart, albeit slightly more complicated [251].
This technique is based on Weber’s law (see Section 5.5.2). It requires a
method to determine a just noticeable difference (JND) ΔLd for a particular dis-
play value Ld , obtained with all the light sources in the room switched off. We
can repeat this experiment with the room lights on, yielding a second JND ΔLb .
To determine the just noticeable contrast, a test pattern with a set of patches
increasing in contrast can be displayed. By choosing 1/ f noise as the test pattern
(as shown in Figure 15.27), the pattern matches the second-order image statistics
of natural scenes. It is therefore assumed that the calibration is for tasks in which
natural images are involved.
By reading this chart, the viewer is able to select which patch holds the pattern
that is just visible. It was found that this task is simple enough to be fast, yet
accurate enough to be useful in the following computations [251].
This pattern is read twice, once with the lights on, and once with the lights
off. With the two JNDs obtained using the above method, Weber’s law then states
that
ΔLd ΔLb
 =  . (15.26)
Ld Ld + LR

i i

i i
 i i

i i

844 15. Image Properties and Image Display

Figure 15.27. Test pattern used to determine the smallest amount of contrast visible
on a particular display device. (Image from [251], 
c 2006 ACM, Inc. Reprinted with
permission.)

From this equation, we can compute how much light LR is reflected off the screen:
 
 ΔLb
LR = Ld −1 , (15.27)
ΔLd

where all the variables on the right-hand side are known.

Knowing how much light is reflected from a particular display device when
placed in a specific environment is sufficient for the development of an algorithm
to correct for ambient light. As the screen reflections add light to the signal that
is propagated from the screen to the viewer, it should be possible to subtract the
same amount, i.e., LR , from all pixels prior to display. However, this leads to two
problems.
First, dark pixels may be mapped to negative values, which would then have
to be clamped to zero. This is unacceptable as it will lead to uncontrolled loss of
detail. Second, by subtracting a quantity from each pixel, the dynamic range of
the display device is under-utilized as the range between [Lmax − LR , Lmax ] is no
longer used.
Alternative solutions to subtracting LR from each pixel include modifying the
gamma value used in gamma correction to account for ambient light [1216], using
hyperbolic functions [700], or using histogram equalization [1230]. However,
none of these techniques take into account the value of LR we have just computed.
With LR valid for a display value of Ld , we seek a monotonically non-
decreasing function f that maps Ld to Ld − LR , has a slope of 1 at Ld to main-
tain contrast at this display value, and maps 0 to 0 and m to m (where m is the

i i

i i
 i i

i i

15.6. Ambient Light 845

maximum value). The constraints on this function f : [0, m] → [0, m] can be sum-
marized as follows:

f (0) = 0, (15.28a)
f (m) = m, (15.28b)
f (L) = L − LR , LR ≤ L ≤ m (15.28c)

f (L) = 1, (15.28d)

f (x) ≥ 0, (15.28e)

By splitting this function into two domains, the first spanning [0, L] and the second
spanning [L, m], these constraints can be met by setting [1006]

px
f (x) = . (15.29)
(p − 1)x + 1

We first make the appropriate substitutes to yield the following pair of functions:

(L − LR ) px
f[0,L] (x) = , (15.30a)
x(p − 1) + L
x−L
p (m − L + LR )
f [L,m] (x) = m−L + L − LR . (15.30b)
x−L
(p − 1) +1
m−L

We can now solve for the free parameter p by using the restrictions on the deriva-
tive of f . For the [0, L] range we have

 p (L − LR ) (p − 1)px (L − LR )
f[0,L] (x) = − (15.31a)
L(p − 1)x/L + 1 L2 ((p − 1)x/L + 1)2
p (L − LR ) L
= (15.31b)
(xp − x + L)2
= 1. (15.31c)

i i

i i
 i i

i i

846 15. Image Properties and Image Display

By substituting x = L, we find p[0,L] = (L − LR )/L. Similarly, for the range [L, m]

the derivative f[L,m] is given by

p (m − L + LR )
f[m,L] (x) =   (15.32a)
(p − 1) (x − L)
(m − L) +1
m−L
(p − 1) p (x − L) (m − L + LR )
=  2 (15.32b)
(p − 1) (x − L)
(m − L)2 +1
m−L
= 1. (15.32c)

Solving for p yields

m−L
p[L,m] = . (15.33)
m − L + LR
Substitution into f (x) then yields the final form of the remapping function:
⎧
⎪
⎪ (L − L )2 x
⎪ 2 R
⎪ 0 ≤ x ≤ L,
⎨ L − LR x
f (x) = x−L (15.34)
⎪
⎪ + L − LR L ≤ x ≤ m.
⎪
⎪ L R (x − L)
⎩1−
(m − L + LR ) (m − L)

Contrast correction
250
L = 12.75
L = 25.50
L = 38.25
200 L = 51.00
f(x)

L = 76.5
150 m = 255

100

50

0
0 50 100 150 200 250
x

Figure 15.28. Example remapping functions for JNDs of 5%, 10%, 15%, and 20% of
the maximum display value m. (Image from [251], 
c 2006 ACM, Inc. Reprinted with
permission.)

i i

i i
 i i

i i

15.6. Ambient Light 847

Choosing the maximum value m = 255 and the pivot point L at one third the
maximum, example mappings are plotted in Figure 15.28. It will be assumed that
the pivot point L is the same as the luminance value for which the JNDs were
computed, i.e., L = Ld . As the correction algorithm is optimal for this particular
value, and necessarily approximate for all others, the choice of pivot point, and
therefore the display value Ld for which the JNDs are computed, should be chosen
such that the displayed material is optimally corrected. For very dark images, the
JNDs should be measured for lower values of Ld , whereas for very light images,
higher values of Ld could be chosen. In the absence of knowledge of the type of
images that will be displayed after correction, the JNDs can be computed for a
middle gray, i.e., L = Ld ≈ 0.2 m.
In summary, by applying this function to display values, we get new display
values which are corrections for ambient reflections off of the screen. The light
that reaches the retina is now given by

Ld = f (Ld ) + LR . (15.35)

For the pivot point L = Ld , we have f (Ld ) = Ld − LR . The light that reaches the
eye is therefore exactly Ld = Ld , which is the desired result. For values larger or
smaller than Ld , the correction is only partial.

i i

i i
 i i

i i

Chapter 16
Color Management

Color management is arguably the most important practical application of many

of the concepts we have discussed so far. As an abstract concept, color man-
agement describes how to manipulate color to achieve appearance matches from
one physical device to another color imaging device. In order to achieve these
matches, it is necessary to combine elements of color measurement, colorime-
try, chromatic adaptation, and color appearance. This chapter will describe the
general concept of a color management system and go into some detail about the
de-facto industry standard ICC Profile Format for color management. A brief
overview of color management using color look-up tables (LUTs) is also de-
scribed. This chapter should be considered an introduction to the general concepts
of color management and should not be considered a blueprint for creating or ma-
nipulating device color profiles. Rather, we hope this chapter brings a greater
understanding of what a color profile is and what types of data it may contain.

16.1 A Generic Color Management System

Color imaging devices all inherently have a device-native color space based upon
the characteristics of the device itself. These characteristics may be the light-
producing primaries of a display device such as a CRT or LCD, the spectral sen-
sitivities of a camera or scanner, or the inks, dyes, or toners of a color printer. Even
if these devices have the same type of native color space such as an RGB space of
a camera and a computer display, there is no reason to believe that matching de-
vice units in those RGB spaces will result in identical colors. The general concept
of color management involves moving those device units from all the individual
color spaces into a unified device-independent space. We have already been in-
849

i i

i i
 i i

i i

850 16. Color Management

Input Device
Characterization

Device
Independent CIE
XYZ

Output Device
Characterization

Figure 16.1. An example of device-independent color management using CIE

colorimetry.

troduced to these types of color spaces, based upon CIE XYZ colorimetry and
extended to include color appearance spaces such as CIELAB and CIECAM02.
Recall our introduction to device-independent color from Chapter 7 using the CIE
XYZ system of colorimetry, redrawn here in Figure 16.1.
For this type of color management system, we bring the individual device
units into a common reference color space. The common reference color space,
or connection space, must be well understood and unambiguously defined in order
to be considered device-independent. We bring the colors into this space through
a process known as device characterization, which involves careful measurement
of the individual device properties. Conceptually, we can think of converting the
device units into the reference color space as a forward characterization. From
the reference color space we can then move to any output device using an inverse

i i

i i
 i i

i i

16.2. ICC Color Management 851

device-characterization. Essentially this process tells us what exact device units

are necessary to generate the same XYZ tristimulus values. Again, the inverse
characterization requires careful measurement of the properties of the output de-
vice. In reality, we may not have a specific one-directional nature for a color man-
agement system. We can think of many devices that could be considered on both
the input and output side. For instance, a digital camera may capture an image
and move that image into the reference space. From there, we may want to move
that image to a CRT or LCD monitor for digital preview and manipulation. The
final stage may involve printing the image, such that it matches our input device
or matches the image that was displayed on the computer screen. In this situation
it is necessary to have both the forward and inverse device characterization for
moving into and out of the reference color space. There are many different tech-
niques for characterizing color imaging devices, including modeling the physical
properties of the device or through the use of color look-up tables (LUTs). An
introduction to these techniques is given in this chapter; for more information the
interested reader should see [93, 577, 1035].
Recall that CIE XYZ colorimetry only guarantees a color match under a single
viewing condition; it tells us nothing about what that color actually looks like.
Almost by definition, color management requires the generation of color matches
with very disparate devices and viewing conditions, such as a computer display
viewed in a dim surround and a printer viewed in an average or bright surround. In
these situations, it may be desirable to have your reference color space be a color
appearance space, or at least use certain aspects of color appearance models such
as chromatic adaptation transforms or lightness rescaling. We may not actually
generate colorimetric matches between the input and output devices, but are rather
more interested in generating color appearance matches.

16.2 ICC Color Management

One of the de-facto industry standards for device color management is the ICC
color management system. The International Color Consortium, R or ICC, was
founded in 1993 to create an open color management system. The founding mem-
bers of the ICC were Adobe Systems Inc., Agfa-Gevaert N.V., Apple Computer,
Inc., Eastman Kodak Company, FOGRA (Honorary), Microsoft Corporation, Sil-
icon Graphics, Inc., Sun Microsystems, Inc. and Taligent, Inc. [523]. Today the
ICC has expanded to over 60 members in industry and academia. Specifically,
the ICC is focused on developing and maintaining a color profile format, as well
as the registration of specific tags and descriptions of elements for these pro-

i i

i i
 i i

i i

852 16. Color Management

files [523]. As of this printing, the current ICC profile format stands at Version
4.2.0.0, which is defined in the ICC.1:2004-10 profile specification and also at
their website (http://www.color.org).
Much like our generic color management example given above, the profile-
based system recommended by the ICC is grounded in the use of a reference
color space which is ”unambiguously defined” [523]. The reference color space
chosen by the ICC is the CIE XYZ system of colorimetry, based upon the 1931
Standard Observer. Since we know the CIE XYZ system still contains some am-
biguity with regard to the appearance of colors, further assumptions based upon
a reference viewing condition and standard color measurement techniques were
also defined. The reference viewing condition follows the recommendations of
ISO 13655:1996, Graphic Technology—Spectral Measurement and Colorimetric
Computation for Graphic Arts Images [533]. The reference medium is defined
to be a reflection print, viewed under CIE D50 illumination at 500 lux, with a
20% surround reflectance. Additionally, the reflectance media should be mea-
sured using a standard 45/0 or 0/45 measurement geometry with a black backing
placed behind the media substrate. This measurement geometry places the light
source and color measurement sensor 45 degrees apart to minimize specular re-
flectance contributions. In addition, the ICC and ISO standards define the relative
reflectance of the reflectance substrate to be 89% relative to a perfect reflecting
diffuser, with the darkest value having a reflectance of 0.30911% [523]. These
values correspond to a linear dynamic range of approximately 290–1. Addition-
ally there are specifications for viewing flare, or stray light hitting the reference
medium, of about 1.06 candelas per square meter. These standardized viewing
conditions, along with the reference color space of the 1931 CIE XYZ tristimu-
lus values, are called the Profile Connection Space (PCS). The PCS is the basis
for all the color profile based transformations used in ICC color management, as
shown in Figure 16.2. Additionally, the ICC allows the use of CIELAB as a PCS,
calculated based upon the standard CIE XYZ tristimulus values as well as on the
reference viewing conditions.
The color transforms represented by the circular T in Figure 16.2 represent
a road map for converting between specific device units and the reference CIE
XYZ color space. These transforms form the basis of the ICC color profiles.
These transforms could be based entirely on color measurements of the particu-
lar imaging devices and conform to generating CIE XYZ colorimetric matches.
Recall once again, from Chapter 7, that colorimetric matches are only guaranteed
to hold up for a single viewing condition. When performing cross-media color
reproduction, we are almost certain to attempt to generate color matches across
vastly disparate viewing conditions, using a wide variety of devices. By providing

i i

i i
 i i

i i

16.2. ICC Color Management 853

T T
T T
Profile Connection
Space (PCS)

T T
T

T = A bidirectional device to PCS color space transform

Figure 16.2. The general ICC color management workflow, using a Profile Connection
Space (PCS), redrawn from [523].

additional information for the PCS, namely by providing a reference viewing con-
dition and medium as well, the ICC has created a system for color management
that can account for these changes in color appearance. This includes changes in
the color of the viewing environment, using chromatic adaptation transforms, as
well as changes in the surround, media-white and black points, and overall tone
and contrast. How these color transforms are chosen and applied is specified by
the rendering intent of the end-user.

i i

i i
 i i

i i

854 16. Color Management

16.2.1 Rendering Intents

One of the biggest areas of confusion that many users experience when they first
encounter ICC color management is the choice of rendering intent offered by the
system. The ICC specifies four distinct rendering intents, also called rendering
styles, in their profile standard: media-relative colorimetric, absolute colorimetric,
perceptual, and saturation. This section will describe what each of these rendering
intents is trying to accomplish and the goals and compromises of these different
intents.
The colorimetric rendering intents are seemingly the most obvious. The ICC
profile standard does not actually define the absolute rendering intent in the profile
itself, but rather uses additional information contained in the profiles to calculate
ICC absolute transforms. The two colorimetric rendering intents are designed to
faithfully preserve and generate colorimetric matches of in-gamut colors, often at
the expense of out-of-gamut colors. Care should be taken when using these in-
tents with devices that have large differences in gamut volume. The actual gamut
mapping of the out-of-gamut colors is not defined by the ICC or the profile and is
left to the discretion of the device manufacturer. Although the colorimetric ren-
dering intents suggest a strict match based upon CIE XYZ colorimetry, they do
allow for some changes in viewing or measurement condition through the use of
a chromatic adaptation transform. When the input or output data are measured or
viewed under illumination that differs from the reference CIE D50, a chromatic
adaptation transform should be used to convert those data into CIE D50. The
ICC recommends the use of the linearized Bradford transform, as described in
Chapter 10.
The media-relative colorimetric rendering intent is based upon color data that
is normalized relative to the media white point for reflecting and transmitting data.
This implies that the relative media white will have CIE XYZ values of D50 in
the PCS, or have CIELAB values of 100 for L∗ , and 0, 0 for a∗ and b∗ . This trans-
form rescales the colors, accounting for chromatic adaptation into the reference
viewing conditions, such that the white matches the reference white. It should be
noted that this rescaling is technically only appropriate for the in-gamut colors.
The goal of this transform is to assure that the highlight detail in color images is
maintained, by scaling all colors into the range of the reference medium, though
this rescaling may introduce chromatic shifts for other colors [523]. To prevent
these shifts from happening, it may be desirable to use the absolute colorimetric
intent.
The absolute colorimetric intent does not necessarily rescale the data such
that the whites match for a specific device and the reference medium. One of
the requirements of an ICC profile is that it contains the PCS values of the media

i i

i i
 i i

i i

16.2. ICC Color Management 855

white, relative to a perfect reflecting diffuser. Whereas the reference viewing

condition is said to have a relative D50 white of [0.9642, 1.0, 0.8249], the actual
media white may be considerably different. In this situation the ICC absolute CIE
XYZ tristimulus values are rescaled using
 
Xmw
Xabs = Xrel , (16.1a)
XD50
 
Ymw
Yabs = Yrel , (16.1b)
YD50
 
Zmw
Zabs = Zrel , (16.1c)
ZD50

where Xmw is the CIE tristimulus value of the media white, as specified by the
mediaWhitePointTag and Xrel is the relative CIE tristimulus value being normal-
ized. It is important to note that while this is referred to as ICC absolute colori-
metric intent, it is actually closer to relative colorimetric scaling as defined by
the CIE, since we are dealing with normalized units relative to a perfect reflect-
ing diffuser. This alone has probably been a point of confusion for many users.
The ICC absolute colorimetric intent may be most useful for simulating different
output devices or mediums on a single device such as a computer LCD display,
a process generally referred to as color proofing. A similar strategy could be uti-
lized to simulate one type of printing press or technology on another printer. For
practical applications, the ICC states that the media-relative colorimetric intent
can be used to move in and out of the PCS, and absolute colorimetric transforms
can be calculated by scaling these media-relative PCS values by the ratio of the
destination media white to the source media white [523].
The ICC colorimetric intents should be based upon actual color measurements
of the imaging devices, allowing for chromatic adaptation to the reference view-
ing conditions. It is possible that these color measurements do not represent actual
color measurements but rather measurements that may have been altered in order
to minimize color error for important colors, such as skin tones, or even to cre-
ate preferred color reproduction of the in-gamut colors. Although these types of
color measurements may not represent true CIE XYZ colorimetry, they will be
interpreted as colorimetric transformations in an ICC workflow.
The saturation and perceptual intents are much less stringently defined than
the two colorimetric rendering intents. These are purposely left to be vendor-
specific, to allow for proprietary color rendering strategies. Of these two, the
saturation intent is more specific. This intent was designed to allow for the re-
production of simple images where color saturation, vividness, or chroma is most
important, such as business graphics and charts. In these situations, it may be

i i

i i
 i i

i i

856 16. Color Management

desirable to perform non-colorimetric rendering in order to maintain the overall

color saturation. An example of this may be actually changing the hue of a graphic
in order to get the most pleasing image, e.g., using the full“red” primary of one
computer display to reproduce a graphic created while viewing a second display
whose red primary is different.
The perceptual rendering intent was designed to produce color reproductions
that best match the appearance of images when viewed on different display de-
vices. It is specifically for this rendering intent that it was necessary to accurately
define the reference viewing conditions and medium, and those conditions are
used only for perceptual re-rendering. This rendering intent allows for the rescal-
ing of both the out-of-gamut and in-gamut colors, as necessary, in order to gen-
erate appearance matches. The ICC suggests that for perceptual intent the PCS
should represent the appearance of an image as viewed by a human adapted to the
reference conditions [523]. In addition, this rendering intent allows for appear-
ance changes based upon the viewing conditions themselves, such as changes in
the surround or illumination amounts, as discussed in Chapter 11. Since the rela-
tive white and black levels, and as such the linear dynamic range, are well defined
in the reference viewing conditions this allows the perceptual intent to adjust the
tone scale based upon these levels. For example, an input profile from a device
with a larger dynamic range than the reference viewing medium could use some-
thing like a smooth sigmoid function to maintain details in the shadows and high-
lights. Sophisticated gamut-mapping algorithms or gamut-expanding algorithms
can be used to shrink or expand the gamut in and out of the PCS.
As we learned in Chapter 11, viewing conditions can greatly alter the percep-
tion of color appearance. As illumination levels increase, we know that both the
colorfulness and the contrast increase via the Hunt and Stevens effects. We know
that, additionally, the surround levels can also change the apparent contrast, as
demonstrated by Bartleson and Breneman. We know that our chromatic adapta-
tion changes as a function of the mode of viewing, as well as illumination levels
and other factors. It is possible to use a color appearance model to aid in the gen-
eration of the transforms used in the perceptual rendering intent. Many times we
do not actually want our color reproductions to generate actual color appearance
matches, but perhaps would rather they produce a pleasing color image that best
utilizes the input or output device space. A conceptual flowchart of using the per-
ceptual rendering intent for a single input and output imaging device is shown in
Figure 16.3.
Whereas the colorimetric rendering intents should build the device-to-PCS
and PCS-to-device transforms based upon measurements and device characteri-
zation, this is not absolutely necessary in perceptual intent transforms. To achieve

i i

i i
 i i

i i

16.2. ICC Color Management 857

Input Image Output Image

Reference XYZ Color

Perceptual Intent Color Space & Viewing Perceptual Intent Color
Re-Rendering Based on Conditions Re-Rendering Based on
Input Viewing Output Viewing

Figure 16.3. An ICC perceptual rendering intent workflow with a single input and output
device.

any degree of color reproduction accuracy, however, many transforms do start

with colorimetric measurements as a baseline. In the example shown in Fig-
ure 16.3, an input image is re-rendered into the profile connection space, based
upon the defined reference viewing conditions. This re-rendering should take
into account the properties of the scene, as well as properties of the device itself.
For instance if the image taken is a brightly illuminated outdoor scene using a
camera capable of capturing a wide dynamic range, then the perceptual intent re-
rendering would need to rescale, or tone-map the image down into the reference
color space. Chapter 17 gives more details on how this type of dynamic range
reduction transform may be accomplished using an automatic algorithm.
A typical re-rendering would have to at least map the tone-scale into the lower
range of the reference media, taking care to preserve highlight and shadow de-
tail. Careful consideration should be given to areas of the image that cannot be
faithfully reproduced by the reference medium. For example, since the reference
medium is defined to be a typical reflection print, there is a maximum reflectance
that is bounded by the printing substrate or paper—defined to be 89% of a per-
fect reflecting diffuse surface. In many real images, the specular highlights may
be well above even a 100% diffuse surface. We can either scale these highlights
down into the range of the PCS or allow the PCS to have values greater than 1.0
for CIE Y, or 100 for CIELAB L∗ .

i i

i i
 i i

i i

858 16. Color Management

Once transformed into the PCS, we then have another color profile transform
to an output device. This is generally another re-rendering into the viewing con-
ditions of the output display device. If we are moving to a device that also is
capable of reproducing a large dynamic range, we may wish to undo some of
the tone compression that was performed on the input, provided this can be done
in a visually lossless manner. Additionally, the perceptual re-rendering can take
into account properties of the viewing conditions, adjusting the overall bright-
ness or contrast as a function of illumination level and the surround. Again, these
re-renderings are performed into the PCS space, assuming the reference viewing
conditions, and then back out of that PCS space into the viewing conditions of the
output device.
Color measurement and color appearance models may be used to generate a
baseline perceptual rendering transform, but they may still not generate an overall
desired color reproduction. As such, it is possible to manipulate the color trans-
formations in order to maximize overall image pleasantness. Just as this could
be done with the colorimetric transforms, for example, to minimize errors in spe-
cific regions of color space, the PCS will assume that color appearance matches
are being generated. As such, the creation of perceptual transforms used in the
ICC color management system can almost be classified as part science and tech-
nology and part creative or artistic choice. Commercial software is available to
generate colorimetric profiles and transforms for most common color imaging
devices, though there are far fewer resources available for generating these per-
ceptual transforms.
To reemphasize this point, remember that the perceptual rendering intent is
used to generate desired or pleasing color appearance matches, and the specific
definition of these transforms is left to the actual manufacturers, vendors, and im-
plementers of the ICC profiles. It is quite likely that the color-managed image
reproduction between multiple devices using the perceptual intent will be differ-
ent based upon the different manufacturers’ choices of implementation. Although
this does sound slightly ominous with regard to accurate color reproduction across
different device manufacturers, by forcing all the perceptual rendering transforms
to first go into the viewing conditions of the reference PCS and then using these
same reference conditions as well on output, the ICC system does mitigate in prac-
tice manufacturer differences. It should also be noted that although the reference
viewing conditions are designed to emphasize the graphic arts and printing indus-
try, similar concepts can be utilized or extended for pure soft-copy reproduction,
such as digital camera and computer display, or a film and video workflow.
All of the color transforms for these different rendering intents are actually
encoded into the color profiles themselves. Many, but not all, profiles supply all

i i

i i
 i i

i i

16.2. ICC Color Management 859

of these transforms internally and allow the color management engine to choose
the appropriate one based upon user input. This potentially makes the generation
of color profiles a complicated, and even artistic, procedure. Many end users of
color management systems are also often confused about which rendering intent
to choose. It is our goal that this chapter will help aid them in making their
decisions. The ICC also provides some guidance, generally recommending the
perceptual rendering intent when strict adherence to colorimetric matching is not
necessary, such as when creating pleasing natural image reproductions [523].
They also recommend using the perceptual transform when the input and out-
put devices have largely disparate gamuts or properties, as the perceptual intent
should do a better job of gamut-mapping and taking into account changes in view-
ing conditions. When working in a proofing environment, e.g., when simulating
one type of device on another, the ICC recommends using the colorimetric intents.
Remember that the colorimetric intents are only attempting to generate colorimet-
ric matches for colors that are within the gamut of the devices; they do not have
defined gamut-mapping algorithms associated with them.

16.2.2 The Color Management Module

In the ICC framework of color management, the profiles themselves provide all
of the information necessary to generate the color transformations for all of the
different rendering intents. These profiles are then linked together by a color man-
agement engine, or color management module (CMM). The general architecture
of the ICC color management system is shown in Figure 16.4. The job of the
CMM is to parse the profile data structure, interpret the correct set of profile tags
necessary to generate a set of color transforms based upon the rendering intent,
and link profiles together to create an end-to-end color reproduction system. The
ICC does allow for the use of third-party color management modules, as well as
private data tags in the color profiles themselves, in order to allow for proprietary
discrimination or value.
The example architecture shown in Figure 16.4 is designed to operate at the
computer operating system level, where an application sits on top of, or can refer
to, the color management interface. Additionally the application can call certain
graphics or imaging libraries that also sit on top of the color management inter-
face. The interface also handles the file reading, and perhaps parsing, of the ICC
profiles and passes that information to the color management module. In order
to function properly, there must always be a default color management module
that handles the generation and application of the appropriate color transform,
which is then passed back through the color management interface. By allowing
third-party developers to plug into the color management interface rather than to

i i

i i
 i i

i i

860 16. Color Management

Applications

Graphics Imaging
Library Library
ICC Color Profiles

Color Management Interface

Default
CMM

3rd Party 3rd Party

CMM CMM

Figure 16.4. A general overview architecture of the ICC color management framework,
redrawn from [523].

applications and libraries directly, the system allows dynamic swapping of the ac-
tual CMM used. When used at the operating system level, this type of architecture
allows all computer applications to receive identical color management and not
worry about applying the color transforms themselves.

16.2.3 ICC Profiles

As we have already discussed, the most important pieces of the ICC color man-
agement system are the color profiles themselves. These are essentially binary
files that contain the information necessary for the CMM to generate the color
transforms. These are required to contain a 128-byte header, a profile tag table,
and the profile tagged element data. These tags and the data they point to repre-
sent the essence of the profile, and a registry of them is maintained by the ICC.
The ICC specification does allow for the use of private data tags for use with
third-party color management modules. It is important to note that Version 4 of
the ICC specification allows for the use of 8-bit and 16-bit integers for the data
elements in the profile, as well as for the profile connection space. Care must be
taken when performing color transformations with limited bit-depth, as it is very
easy to introduce quantization errors. There are far too many defined tags for the
scope of this text, so this section will outline several important tags as an intro-
duction to the ICC profile format. The interested reader is encouraged to examine
the full ICC specification for more details [523].

i i

i i
 i i

i i

16.2. ICC Color Management 861

The profile header contains useful information about the profile itself, as well
as the type of transforms defined. Among other elements, it contains information
about the canonical input space, e.g., RGB or CMY space and the desired profile
connection space or canonical output space, e.g., XYZ or CIELAB. It can also
include information about the preferred CMM, the manufacturer and model of
the device being profiled, the profile creator, etc. Additionally the profile header
contains the class of the profile. There are seven distinct classes of profiles, and
these are defined to make it easier for the CMM to implement certain types of
color transforms. The seven classes are:

• input device profile;

• display device profile;

• output device profile;

• DeviceLink profile;

• ColorSpace conversion profile;

• abstract profile;

• named color profile.

Of these classes, the input, output, and display profiles are by far the most
common among standard color imaging systems. The device link profile is more
exotic, and is used to link two devices together without using the PCS. This may
be beneficial when performing color reproduction in a closed system, such as ty-
ing a scanner or camera directly to a printer. Additionally these classes of profiles
may be used for color proofing of multiple printers, while avoiding any cross-
contamination, for example, between the CMYK primaries of the two printers.
The color space conversion profile is designed for transforming between a given
color space that does not describe a specific device and the PCS. For example,
it would be possible to use this type of profile to convert from images encoded
in an LMS cone space into the PCS XYZ space. The abstract profile and named
profile are even more rare. The abstract profile can contain color transformations
from the PCS to the PCS and could be used as a form of color re-rendering. The
named color profile is used to transform from device units into a specific categor-
ical color space, corresponding to color names. This could be used for a specific
device that has a set of named colorants or primaries.
Following the profile header is the tag table, which serves as an index into
the actual transform information contained in the profile. The tag table includes

i i

i i
 i i

i i

862 16. Color Management

the number of tags, the actual signatures of each tag, the size, and location in the
actual binary file. The number of tags required by the ICC specification varies
by the type and class of the profile. There are a few common requirements for
all the classes, except for the device link: profileDescriptionTag, copyrightTag,
mediaWhitePointTag, and chromaticAdaptationTag.
The first two required tags provide information about the profile itself. As
we already learned in Section 16.2.1, the mediaWhitePointTag is used to convert
from the media-relative colorimetric intent into the ICC absolute colorimetric in-
tent, through a normalization of the XYZ or CIELAB values. Remember, the ICC
absolute colorimetric intent should be thought of as a type of relative colorime-
try, as defined by the CIE. The chromaticAdaptationTag encodes a 3 × 3 color
matrix that can be used to convert a CIE XYZ tristimulus value measured at the
device’s viewing conditions into the D50 illumination of the reference viewing
condition. Although any chromatic adaptation matrix could be used, the ICC rec-
ommends using the linearized Bradford matrix, as discussed in Section 10.4.5.
The chromaticAdaptationTag is encoded in a nine-element array:
 
a0 a1 a2 a3 a4 a5 a6 a7 a8 (16.2)

and can be applied as

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
X a0 a1 a2 X
⎣Y ⎦ = ⎣a3 a4 a5 ⎦ · ⎣Y ⎦ . (16.3)
Z PCS a6 a7 a8 Z Source

The equations for calculating a chromatic adaptation matrix using the linearized
Bradford transform for the ICC D50 reference condition are
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
R X 0.8951 0.2664 −0.1614 X
⎣G⎦ = MBFD ⎣Y ⎦ = ⎣−0.7502 1.7135 0.0367⎦ · ⎣Y ⎦ , (16.4a)
B Z 0.0389 −0.0685 1.0296 Z
⎡ ⎤
RD50
⎢ RSource 0.0 0.0 ⎥
⎢ ⎥
−1 ⎢ GD50 ⎥
Madapt = MBFD · ⎢ 0.0 0.0 ⎥ · MBFD (16.4b)
⎢ GSource ⎥
⎣ BD50 ⎦
0.0 0.0
BSource

Types of Transformations. In ICC profiles, the color transformations them-

selves can be broken down into two distinct categories: matrix and tone-

i i

i i
 i i

i i

16.2. ICC Color Management 863

"B" Curves

Red TRC
Device Space PCS

Channel 1
Green TRC Matrix X
Channel 2 Y
3x3 Z
Channel 3

Blue TRC

Figure 16.5. Workflow for converting colors from a input or display device into the PCS
using a MatTRC transform, redrawn from [523].

reproduction curves (MatTRC) and color look-up tables (cLUT or LUT). The
matrix and tone-reproduction curves are used primarily to represent the colori-
metric transforms for some input and display classes of profiles. The look-up
table type transforms can be used for the colorimetric intents for the three main
classes of profiles, as well as for the perceptual and saturation intents. It should
be noted that these transform types are not defined for the named color profiles,
which have their own special characteristics.
The MatTRC transforms should be familiar to readers, as they are similar
transforms to those discussed for converting different types of additive RGB color
spaces in Chapter 8. An example workflow for these transformations is shown in
Figure 16.5. The specific tags in the ICC profile that relate to these transforms
are redTRCTag, redMatrixColumnTag, greenTRCTag, greenMatrixColumnTag,
blueTRCTag, and blueMatrixColumnTag. The three MatrixColumnTags repre-
sent the measured XYZ values of the three red, green, and blue sensors of a input
device, such as a camera or scanner, or the RGB primaries of a display device.
As we learned from Chapters 7 and 8, additive devices such as LCD and CRT
displays can be characterized by measuring the individual CIE XYZ values of
the three primaries. Using Grassmann’s laws, we can calculate a 3 ×3 color ma-
trix transformation from an RGB device into CIE XYZ by adding the individual
CIE XYZ values of the primaries. The red, green, and blue MatrixColumnTag

i i

i i
 i i

i i

864 16. Color Management

Input value LUT array index Output value

0.0000 0 0.0000
0.1000 1 0.0063
0.2000 2 0.0290
0.3000 3 0.0707
0.4000 4 0.1332
0.5000 5 0.2176
0.6000 6 0.3250
0.7000 7 0.4563
0.8000 8 0.6121
0.9000 9 0.7931
1.0000 10 1.0000

Table 16.1. Example one-dimensional look-up table (1D-LUT) approximating a gamma

of 2.2.

elements form the first, second, and third columns of this matrix, respectively:
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
X Xred Xgreen Xblue R
⎣Y ⎦ = ⎣Yred Ygreen Yblue ⎦ · ⎣G⎦ . (16.5)
Z Zred Zgreen Zblue B device

The red, green, and blue TRCTags represent transforms for undoing any nonlinear
encoding of the device color space, and as seen in Figure 16.5, they are applied
before the linear transform to CIE XYZ. The curve types allowed by the TRCTags
include a simple power or gamma function of the form Y = X γ . These values are
stored as 16-bit fixed numbers, with eight fractional bits. In practice, this means
that the numbers encoded would be divided by 28 or 256 to get the appropriate
gamma value.
Alternatively, the TRCTags can encode sampled one-dimensional functions of
up to 16-bit unsigned integer values. These sampled functions can be thought
of as one-dimensional look-up tables, or 1D-LUTs. Perhaps the easiest way to
think about 1D-LUTs is to represent the table as an array of numbers. The in-
put color, in device space, is then the index into this array, and the output is the
value contained at that index. A simple example is given in Table 16.1, using an
11-element LUT to approximate a gamma function of 2.2. The ICC specification
states that function values are between entries; linear interpolation should be used
to calculate the intermediate entries. We can plot the 11-element 1D-LUT, lin-
early interpolated to 256 entries against an actual gamma function, as shown in
Figure 16.6.
This figure suggests that there is a reasonable relationship between the small
1D-LUT approximating the 2.2 gamma function and the actual function. This re-

i i

i i
 i i

i i

16.2. ICC Color Management 865

1
gam m a
0.9 1D LUT

0.8

0.7

0.6
Output

0.5

0.4

0.3

0.2

0.1

0
0 0.2 0.4 0.6 0.8 1
Input

Figure 16.6. The differences between an 11-element 1D-LUT interpolated to 256 ele-
ments and an actual gamma curve. Although the functions appear similar, the differences
between the two can lead to visual errors.

sult is actually slightly misleading, as we can find out by calculating the color dif-
ferences with an 8-bit image encoded with a 1/2.2 gamma function. An example
image is shown in Figure 16.7. We can convert this image into XYZ by apply-
ing the two different TRC functions, followed by a matrix transformation. From
there we can calculate CIE δ E ∗ a b color differences to get an idea of how large
the error may be. Assuming Figure 16.7 is encoded using the sRGB primaries,
with a gamma correction of 1/2.2, we find that the maximum color difference be-
tween using a real gamma function and an 11-element 1D-LUT is about 3.7. The
complete range of errors is shown in Figure 16.8. We can see that there are some
quantization artifacts that would be clearly visible. This suggests that when using
1D-LUTs, care should be taken to use as large a LUT as possible to avoid errors
introduced by the interpolation. This idea will be revisited later in the chapter
with regards to 3D-LUTs.
Again, these MatTRC transformations are only really valid for certain types
of devices and are only defined for input and display classes of profiles. They
are also only used for colorimetric transformations into and out of the CIE XYZ
version of the PCS. For the other rendering intents, or for other color imaging
devices, the transforms used can be generically referred to as the AToB and BToA
transform, for those going into and out of the PCS, respectively. The specific tags

i i

i i
 i i

i i

866 16. Color Management

Figure 16.7. An example image encoded with sRGB primaries and a 1/2.2 gamma
compression.

3.5

2.5

1.5

0.5

Figure 16.8. The CIE δ E ∗ ab color differences between using a small interpolated 1D-
LUT and an actual gamma function to move into the ICC PCS.

i i

i i
 i i

i i

16.2. ICC Color Management 867

Profile class MatTRC AToB0Tag AToB1Tag AToB2Tag

Input colorimetric device to PCS perceptual device to PCS colorimetric device to PCS saturation
Display colorimetric device to PCS perceptual device to PCS colorimetric device to PCS saturation
Output undefined device to PCS perceptual device to PCS colorimetric device to PCS saturation
ColorSpace undefined device to PCS perceptual device to PCS colorimetric device to PCS saturation
Profile class BToA0Tag BToA1Tag BToA2Tag
Input PCS to device perceptual PCS to device colorimetric PCS to device saturation
Display PCS to device perceptual PCS to device colorimetric PCS to device saturation
Output PCS to device perceptual PCS to device colorimetric PCS to device saturation
ColorSpace PCS to device perceptual PCS to device colorimetric PCS to device saturation

Table 16.2. The different color transformation tags and rendering intents available in ICC
profiles.

used by the ICC are summarized in Table 16.2, for the input, display, output, and
color space classes of transform. The three other classes have specific rules and
tags and are beyond the scope of this text.
Conceptually, we can refer to all of the AToB and BToA transforms as those
based upon color look-up tables (cLUTs). This is perhaps an oversimplification,
as we can see in Figure 16.9, as these transforms can contain 1D-LUTs, matrix
transforms, multi-dimensional cLUTs, or combinations of these types of trans-
forms. Specifically, the AToB transforms contain up to five processing elements
that are applied in a specific order: 1D-LUTs ”B” Curves, 3×3 matrix with an
offset (represented by a 3×4 matrix), 1D-LUTs ”M” curves, n-dimensional color
LUTs, and 1D-LUTs ”A” curves. The four combinations of these, shown in Fig-
ure 16.9, can be made by setting specific individual transforms to identity.
We have already discussed how the one-dimensional look-up tables can be
applied in the MatTRC ICC transforms. Conceptually, the ”A”, ”M”, and ”B”
curves work the same way. It is important to note that the ”A” curves, of which
there are as many as there are color channels in our imaging device, can only
be used in conjunction with the multi-dimensional cLUT. The matrix transform
is very similar to that used in the MatTRC, though there is an additional offset
element. This additive offset could be useful to account for differences in the
black of the device space and that of the PCS, or it could be used to account for
flare in the viewing conditions. The multi-dimensional look-up table (cLUT) is
perhaps the most important element of the AToB transforms. As such, a quick
overview of how multi-dimensional look-up tables work is in order. We should
also note that when our device-space has more than 3 primaries, it is the cLUT
that reduces the dimensionality of those channels down to the three dimensions
of the CIE XYZ or more often CIELAB PCS.
We can think of a cLUT as a dictionary that translates specific device pri-
maries from the device space into another color space. Consider a simple color

i i

i i
 i i

i i

868 16. Color Management

"B" Curves

Device Space PCS

Channel 1
X L*
Channel 2 Y or a*
Z b*
Channel 3

Device Space PCS

Channel 1
Matrix X L*
Channel 2 Y or a*
3x4 Z b*
Channel 3

"A" Curves
"B" Curves

Device Space
Multi-Dimensional
Lookup Table PCS

Channel 1
X L*
Channel 2 Y or a*
CLUT
... Z b*
Channel n

"A" Curves
"M" Curves "B" Curves

Device Space
Multi-Dimensional
Lookup Table PCS

Channel 1
Matrix X L*
Channel 2 Y or a*
CLUT
... 3x4 Z b*
Channel n

Figure 16.9. Example AToB LUT transforms from device coordinates to the ICC PCS,
redrawn from [523].

i i

i i
 i i

i i

16.2. ICC Color Management 869

printer, that has cyan, magenta, and yellow primaries (CMY). As we know, the
goal of the ICC profile is to provide a transform from those specific device pri-
maries into the PCS. We could accomplish this using mathematical and empirical
models, such as the Kubelka-Munk model discussed in Chapter 8. We could also
accomplish this using a brute-force color measurement method along with a color
look-up table. In an extreme case, if we state that our printer is capable of gen-
erating 256 levels (8 bits) of each of the cyan, magenta, and yellow colors, then
we need to print each of those colors separately, and then all of the combinations
of colors. We could then take a color measuring device such as a spectropho-
tometer and measure each one of the printed combinations. With three channels
that would equal 256 × 256 × 256, or over 16 million samples to measure! As-
suming we have the time and energy necessary to accomplish this task, however,
we would then have a direct map that could tell us for any given combination of
CMY what the associated XYZ or CIELAB values are.
Obviously, this is not the most practical approach as the amount of time, let
alone printer media, would not be conducive to such an exercise. If we consider
devices with more than three channels, we can see that the number of measure-
ments necessary to completely fill our LUT would grow exponentially. For prac-
tical implementations we can use a lower number of measurements in each of the
channels and perform some type of interpolation for values in between. As we
saw with the simple 1D-LUTs, if we use too few samples the color error can be
very large. Therefore, it is desirable to achieve a balance between the number of
samples, type of interpolation, and acceptable color error.
Figure 16.10 shows an example of building a cLUT using five sample points
for each of the CMY levels. This creates what is known as a 5×5×5 LUT, with
a total of 125 samples. We could print each of these samples and measure the
corresponding CIELAB or CIE XYZ values to populate the table. Alternatively
we can use a printer model to estimate what the sample values would be. Essen-
tially at each of the colored bulbs, or nodes, in Figure 16.10, we can store the

Figure 16.10. Mapping a device CMY space using a 5 × 5 × 5 color look-up table.

i i

i i
 i i

i i

870 16. Color Management

p110 p111
p11
p1
p10
p100 p101
p010 p011
p01 p
y p0

x p000 p00 p001

z

Figure 16.11. An example of trilinear interpolation.

corresponding color measurement information. To determine what the approxi-

mate color measurement data would have been in between the nodes, we perform
a three-dimensional interpolation.
Common forms of 3D interpolation include trilinear, prism, pyramid, and
tetrahedral. Essentially, we can think of the different types of interpolation as
having trade-offs between accuracy, speed, robustness to noise in the imaging
system, and complexity. They also have different properties with regard to the
number of individual nodes or vertices they use and their ability to maintain local
curvature in the device color space. In general, the fewer vertices used, the better
they are able to handle local changes in the color space, though this may come at
the expense of color accuracy when noise is present.
Trilinear interpolation is the three-dimensional equivalent to linear interpola-
tion; it uses all eight nodes of the three-dimensional color space. An example of
trilinear interpolation is shown in Figure 16.11. If we are interested in finding the
values of point p in Figure 16.11, we must first calculate the linear distance along
each of the three dimensions. Essentially we can think of trilinear interpolation as
performing three separate interpolations along each of the color channels. First,
we calculate the points p00 − p11 via linear interpolation along the z-axis, then
we calculate the points p0 − p1 through interpolation along the x-axis, and finally
calculate point p through interpolation along the y-axis.
Prism interpolation, as shown in Figure 16.12, first subdivides the color space
cube into two pieces by cutting across a face diagonal. If possible, this cutting
should contain the neutral axis so that interpolation along that line is maintained,
and color errors are not introduced into neutral tones. A test is then performed to
determine in which prism the desired point lies. From there, the point is projected
onto the two triangle faces and interpolated using triangle interpolation. Simple
linear interpolation is then done between two projected points. As such, prism

i i

i i
 i i

i i

16.2. ICC Color Management 871

Prism 1 Prism 2

Figure 16.12. An example of sub-dividing the color space cube for prism interpolation.

Pyramid 1 Pyramid 2 Pyramid 3

Figure 16.13. An example of subdividing the color space cube for pyramid interpolation.

interpolation uses only six nodes or vertices of our sampled color space and is
more capable of following local curvature in the color space.
Similarly, pyramid interpolation splits the color cube into three pyramids, each
containing five vertices, as shown in Figure 16.13. These pyramids all come
together at a single point, and that point should be chosen to correspond to neutral
colors. Again tests are performed to determine in which of the pyramids the point
lies, and from there the colors are interpolated.
Tetrahedral interpolation subdivides the space into six tetrahedra, each us-
ing only four vertices. All of the tetrahedra share a common edge, which ideally
should be pointed along the neutral axis. An example of this sub-division is shown
in Figure 16.14. Due to its computational simplicity, as well as ability to main-
tain the neutral axis, this is a common form of interpolation used and will be
discussed in further detail later in this chapter. It should be noted that tetrahedral
interpolation for color space transformation has been covered by certain patents
since the 1970s and 1980s, as detailed by Kang [575]. More details on specific
interpolations can also be found in Kang and Sharma et al. [575, 577, 1035].
Color look-up tables can be powerful tools for transforming colors from de-
vice specific spaces into the ICC reference color space. These cLUTs can be
constructed in such a way as to provide color re-rendering for perceptual ren-
dering intents, or they can also be used to guide certain regions of color space
to other regions. For instance, it would be possible to use cLUTs to shift most
“green” colors to a more saturated green, to create pleasing grass and tree repro-
duction. Likewise, we could shift the “blue” colors away from purple to create
pleasing sky colors. These types of manipulations may be valid for the perceptual
intent. Color look-up tables are not limited to just three dimensions, as we have
illustrated in these examples. The same concepts and interpolation techniques can

i i

i i
 i i

i i

872 16. Color Management

Tetrahedron 1 Tetrahedron 2 Tetrahedron 3

Tetrahedron 4 Tetrahedron 5 Tetrahedron 6

Figure 16.14. An example of subdividing the color space cube for tetrahedral interpola-
tion.

be easily extended to handle n-dimensional color imaging devices. The cLUT is

then used to reduce this dimensionality to the three-dimensional nature of the ICC
PCS.
What we have described so far has been limited to creating a regular sampling
grid in the device color space. Sometimes this is not ideal, due to the nonlinear
relationship between the device colorants and the PCS. Often times we would
like to have the interpolation and sampling performed in a perceptually uniform
color space to minimize visual artifacts. This is one of the reasons CIELAB is
offered as the ICC profile connection space and why many output ICC profiles
choose to use CIELAB rather than CIE XYZ as the PCS. We could use a non-
uniform sampling grid in the cLUT, by essentially changing the distance between
each of the nodes as a function of the device color space. For example, we could
place more nodes toward the dark colors, knowing that the human visual system
is more sensitive to errors in dark regions. This non-uniform sampling can make
it difficult to use some of the geometric interpolation methods as described above,
since the tests to determine which bounding geometry would be different for each
sample cube. Alternatively, we can use 1D-LUTs on the device channels to pro-
vide an expansion or compression of the channels. This allows us to use regularly
sampled grids and standard interpolation techniques, which is why the ICC al-
lows the specification and application of 1D-LUTS, shown as the ”A” curves in
Figure 16.9, prior to application of the cLUTs.
It should be quite clear how the AToB and BToA transforms offered by the
ICC specification can provide a powerful set of tools to generate color reproduc-
tion matches from device spaces into the reference color space. It should also
be quite clear that the shear number of different types of transforms that must
be generated for specific devices can be overwhelming, and therefore beyond the
scope of most casual users. As such, third party profile-creation software tools

i i

i i
 i i

i i

16.2. ICC Color Management 873

have been developed, often in association with color measuring devices. These
tools handle all the appropriate look-up table generation, color measurement, and
binary encoding of the profiles.

16.2.4 Black Point Compensation

There are many more tags than those we have described available in ICC profiles,
most of which are outside the scope of this text. One last tag that should be men-
tioned, however, is the mediaBlackPointTag. Much like the mediaWhitePointTag,
which is used to transform the data from the relative colorimetric intent to the
absolute colorimetric intent, the black point contains the measured CIE XYZ tris-
timulus values of the darkest area of the actual medium. It is recommended that
the chromaticities of the black point match those of the white, although this is
not strictly enforced. If the black has been measured under different illumina-
tion than the CIE D50 of the reference viewing conditions, then these tristimulus
values must be converted into D50 using a chromatic adaptation transform.
Ideal imaging devices would all have identical black values with a luminance
of zero. Unfortunately, real imaging devices have a substantially higher, and often
different, black value. This can come from light leakage, internal and external
flare, or the inability to absorb all the incident light. For reflection media, there
is a limit to how dark a colorant can actually be, and different colorants may
have very different limitations. As such, it is quite common to have black point
differences between two real imaging devices, especially across different types
of media. This is illustrated in Figure 16.15. Essentially this can be thought of
as a gamut-mapping issue, as many dark colors from the input are mapped or
clipped into a single black in the output. This mapping would result in a loss of
information in the shadow regions, which may be undesirable.
In many situations, we could simply request using a perceptual rendering in-
tent that should take these differences into account. Many times, however, we are
interested in maintaining relative colorimetric matches for most of the data, but
would still like to maintain details in the shadows. This correction is known as
black point compensation. Although the ICC does not specify how this transform
is performed, they do allow it to happen for relative colorimetric and saturation
rendering intents. It is not allowed for absolute colorimetric and should be already
built into the perceptual rendering intent. What black point compensation should
do is compress the shadow data in the input space, or PCS, using a simple tone-
mapping algorithm. This mapping should attempt to maintain as much linearity
for the remaining colors as possible.

i i

i i
 i i

i i

874 16. Color Management

L* L*

b* b*

Figure 16.15. Two iconic projections of device-gamuts with different blacks in the
CIELAB space.

16.3 Practical Applications

Color management, and specifically the ICC color management framework, is
in essence a practical application of many of the concepts we have discussed in
this text. It combines elements of color measurement and colorimetry, chromatic
adaptation, and color appearance to generate reproduction matches across a vari-
ety of color imaging devices. In this section, a brief overview of color manage-
ment from a digital camera to a computer display device is given. Additionally
we will discuss an example of trilinear and tetrahedral interpolation.

16.3.1 Color Management Between a Digital Camera

and a Computer Display
Today, ICC profiles can be embedded in almost all major image file formats. Most
commercial digital cameras have the option of, or default to, saving images in the
sRGB format and embedding this profile in the image file. An example of such
an image is shown in Figure 16.16. To display this image such that it is appro-
priately reproduced, we can use the standard ICC transforms. Although we are
using sRGB as an image encoding and input into our color management system,
most sRGB profiles are actually of the display class and contain the MatTRCTag
indicating that the transform into the CIE XYZ of the PCS is handled by three
tone-reproduction curves. Parsing the sRGB profile, which ships with most major

i i

i i
 i i

i i

16.3. Practical Applications 875

Figure 16.16. An image captured with a digital camera and tagged with an sRGB profile.

computer operating systems, we can examine the transforms used to move into
the PCS.
First we consider the nonlinear red, green, and blue TRCTags. Recalling from
Chapter 8, we know that the transfer function defined for sRGB is broken into
two distinct parts: a linear segment for dark colors and a gamma-offset-gain func-
tion for the remaining parts. This function is encoded as 16-bit values in a 1024
element 1D-LUT in the ICC profile as shown in Figure 16.17. We first apply this
LUT to our input sRGB data to get linear RGB units. Next we can examine the
red, green, and blue MatrixColumnTags and combine these into our 3×3 matrix
to transform into CIE XYZ:
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
X 0.4361 0.3851 0.1431 R
⎣Y ⎦ = ⎣0.2225 0.7169 0.0606⎦ · ⎣G⎦ . (16.6)
Z D50 0.0139 0.0971 0.7141 B sRGB linear

Note that this transform appears to be significantly different than the standard
sRGB transform. This is because it has been chromatically adapted from the
standard white of CIE D65 to the reference D50 transform using a linear Bradford
transform as shown in Equation (16.4). We apply this matrix to our linearized

i i

i i
 i i

i i

876 16. Color Management

0.9

0.8

0.7

0.6
Output
0.5

0.4

0.3

0.2

0.1

0
0 200 400 600 800 1000 1200
Input Index

Figure 16.17. The red, green, and blue TRCTags from an sRGB profile.

Figure 16.18. An image transformed into the ICC XYZ reference color space.

i i

i i
 i i

i i

16.3. Practical Applications 877

RGB data, and our image is now in the reference CIE XYZ PCS as shown in
Figure 16.18.
Once we are in the reference color space, we need to move back out onto a
computer display. Essentially, we want to figure out what digital counts should
be displayed on our monitor to achieve a colorimetric match to our input image.
In order to do this, we need to characterize our monitor. We can do this using
commercial software and a spectrophotometer. We first measure the red, green,
and blue primaries at their full intensity. This will help generate the RGB to
XYZ color matrix as shown in Equation (16.6). We also measure the combination
of the three primaries at maximum value to get a measurement of the white of
our computer display. Typically, the native white will be around CIE D65. It is a
good idea to perform a check of your monitor, adding up the individual tristimulus
values of each of the three channels (equivalent to passing R = G = B = 1 into the
3×3 color matrix) to see if they match the measured tristimulus values of your
white. This will determine if your display is behaving in an additive way. From
there, we need to measure the individual tone-reproduction curves of each of the
primaries. This is done by presenting individual ramps on each of the channels.
We can fit a simple gamma function to these measurements, or we can use the
measurements themselves as a 1D-LUT.
For this particular display, the profile is included on the DVD; it was found
that a simple gamma of approximately 1.8 was adequate to describe all three color
channels. The measured color conversion matrix is
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
X 0.4735 0.3447 0.1460 R
⎣Y ⎦ = ⎣0.2517 0.6591 0.0891⎦ · ⎣G⎦ . (16.7)
Z D50 0.0058 0.0677 0.7515 B display

Note that this matrix also has a native white (R=G=B=1) equivalent to CIE
D50, despite the fact that the monitor was measured to have a white of CIE XYZ
(1.0104, 1.0300, 1.0651). This is because, once again, we applied a linear Brad-
ford chromatic adaptation transform to get into the PCS color space. This chro-
matic adaptation transform is given by
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
X 1.042633 0.030930 −0.052887 X
⎣Y ⎦ = ⎣ 0.022217 1.001877 −0.021103⎦ · ⎣Y ⎦ . (16.8)
Z PCS −0.001160 −0.003433 0.761444 Z display

Recall that the ICC specification requests that these transforms are defined
for moving into the PCS and that they must be invertible. As such, we would
actually use the inverse of Equations (16.8) and (16.7), applied in that order. We
1
would then have to apply the inverse of the display gamma, essentially RGB 1.8 .

i i

i i
 i i

i i

878 16. Color Management

Figure 16.19. An sRGB image transformed into a particular monitor display color space.

After that, the image is ready to be displayed. An example of the image that
would be displayed is shown in Figure 16.19. Remember, with the ICC workflow
the underlying encoded image bits do not change, the only thing that changes is
how we interpret those bits. We can see that in order to display an sRGB image
correctly on this particular display, we must make the image more blue and with
a higher contrast.

16.3.2 More Details on Trilinear and Tetrahedral Interpolation

The two most common forms of interpolation for color look-up tables are proba-
bly trilinear and tetrahedral. An example of calculating values based upon these
interpolation techniques is given here. Recall our three-dimensional cube
bounded by eight nodes, or vertices, in the device color space. The cube is shown
again here in Figure 16.20. Hypothetically, we have the measured CIELAB or
CIE XYZ values at each of the corner nodes, represented by points p000 − p111 ,
and we are interested in knowing the PCS values of point p, with the coordinates
in a generic device space of x, y, z.
In the trilinear interpolation, we simply take the three one-dimensional inter-
polations along the device coordinate axes. Combined, this not so simple equa-

i i

i i
 i i

i i

16.3. Practical Applications 879

p110
p11 p111

x p1
p010
p01
p p10
p100
y p011 p101

p0

p000 p00 p001

z

Figure 16.20. A bounding cube in our device color space used in a color look-up table.

tion, based upon [1035], is

PCS (x, y, z) =p000 + (p001 − p000 ) z + (p010 + (p011 − p010 ) z

− (p000 + (p001 − p000 ) z)) y + {p100 + (p101 − p100 ) z
+ (p110 + (p111 − p110 ) z − (p100 + (p101 − p100 ) z)) y
− [p000 + (p001 − p000 ) z + (p010 + (p011 − p010 ) z
− (p000 + (p001 − p000 ) z)) y]} x. (16.9)

To calculate the values in the PCS, you replace the p∗∗∗ values in the equation
with the measured CIELAB or CIE XYZ values at those nodes.
For tetrahedral interpolation, as shown in Figure 16.14, we first must figure
out in which of the six tetrahedra our point lies. This is done using a series of
inequality tests on the (x, y, z)-coordinates of our color space. The interpolation
is then calculated based upon the four specific vertices on that tetrahedron. Note,
that to maintain the neutral axis as best we can, all the tetrahedra share the p000
vertex. Table 16.3 shows the inequality test and the corresponding weights used
for the interpolation. Once we have those weights, we can interpolate our values

Test Wz Wy Wx
x<z<y p110 − p010 p010 − p000 p111 − p110
z<y<x p111 − p011 p011 − p001 p001 − p000
z≤x≤y p111 − p011 p010 − p000 p011 − p010
y≤z<x p101 − p001 p111 − p101 p001 − p000
y<x≤z p100 − p000 p110 − p100 p111 − p110
x≤y≤z p100 − p000 p111 − p101 p101 − p100

Table 16.3. The test cases and vertex weights for tetrahedral interpolation.

i i

i i
 i i

i i

880 16. Color Management

in the PCS using the following relatively simple equation,

PCS (xyz) = p000 + wz z + wy y + wx x. (16.10)

This simplicity is what makes tetrahedral interpolation a popular choice for color
look-up tables.

i i

i i
 i i

i i

Chapter 17
Dynamic Range Reduction

Real-world environments typically contain a range of illumination much larger

than can be represented by conventional 8-bit images. For instance, sunlight at
noon may be as much as 100 million times brighter than starlight [312,952,1075].
The human visual system is able to detect 4 or 5 log units of illumination si-
multaneously and can adapt to a range of around ten orders of magnitude over
time [312].
On the other hand, conventional 8-bit images with values between 0 and 255,
have a useful dynamic range of around two orders of magnitude. Such images
are represented typically by one byte per pixel for each of the red, green and blue
channels. The limited dynamic range afforded by 8-bit images is well-matched to
the display capabilities of cathode-ray tubes (CRTs), and therefore to all displays
that are designed to be backwards compatible with CRTs. Their range, while be-
ing larger than two orders of magnitude, lies partially in the dark end of the range,
where the human visual system has trouble discerning very small differences un-
der normal viewing circumstances. Hence, CRTs have a useful dynamic range of
2 log units of magnitude.
Of course, the inability to store and display the full range of illumination
means that, in practice, compromises have to be sought. For instance, professional
photographers tend to select scenes with a dynamic range that can be captured
with currently available hardware. Examples of professional photographs taken
in “good light” are shown in Figure 17.1. Shooting into the light creates a much
higher dynamic range that is sometimes exploited for artistic effect, as shown in
Figure 17.2.
Although the limitations of photography can be exploited to create a desired
atmosphere, for many applications it would be preferable to capture the full dy-
namic range of the scene, independent of the lighting conditions. If necessary, the
image can then be processed for artistic effect later.
881

i i

i i
 i i

i i

882 17. Dynamic Range Reduction

Figure 17.1. Photographs taken in “good light” contain no under- or over-exposed areas.
This can be achieved by keeping the sunlight behind the camera. (Images courtesy of Kirt
Witte, Savannah School of Art and Design, http://www.TheOtherSavannah.com).

In addition, capturing the full dynamic range of a scene implies that, in many
instances, the resulting high dynamic range image cannot be directly displayed,
since its range is likely to exceed the two orders of magnitude range afforded by
conventional display devices. Examples of images that are typically difficult to
capture with conventional photography are shown in Figure 17.3.
For such images, the photographer normally chooses whether the indoor or
the outdoor part of the scene is properly exposed. Sometimes, neither can be
achieved within a single image, as shown in Figure 17.4 (left). These decisions
may be avoided by using high dynamic range imaging (Figure 17.4 (right)).
Landscapes tend to have a bright sky and a much darker foreground; such
environments also benefit from high dynamic range imaging. For this particular
type of photography, it is also possible to resort to graduated neutral density filters,
whereby the top half of the filter attenuates more than the bottom half. Positioning

Figure 17.2. Photographs shot into the light. The stark contrast and under-exposed fore-
grounds are created for dramatic effect. (Images courtesy of Kirt Witte, Savannah School
of Art and Design (http://www.TheOtherSavannah.com).)

i i

i i
 i i

i i

Dynamic Range 883

Figure 17.3. Some environments contain a larger range of luminance values than can
be captured with conventional techniques. From left to right, top to bottom: Harbor of
Konstanz, June 2005; Liebfrauenkirche, Trier, Germany, May 2006; the Abbey of Mont
St. Michel, France, June 2005, hut near Schloß Dagstuhl, Germany, May 2006.

Figure 17.4. With conventional photography, some parts of the scene may be under- or
over-exposed (left). Capturing this scene with nine exposures, and assembling these into
one high dynamic range image, affords the result shown on the right; refectory in Mont St.
Michel, France, June 2005.

i i

i i
 i i

i i

884 17. Dynamic Range Reduction

Figure 17.5. Linear scaling of high dynamic range images to fit a given display device
may cause significant detail to be lost (left and middle). The left image is linearly scaled.
In the middle image, high values are clamped. For comparison, the right image is tone-
mapped, allowing details in both bright and dark regions to be visible; Mont St. Michel,
France, June 2005.

such a filter in front of the lens allows the photographer to hold detail in both sky
and foreground. Nonetheless, a more convenient and widely applicable solution
is to capture high dynamic range data.
There are two strategies to display high dynamic range images. First, we
may develop display devices that can directly accommodate high dynamic range
imagery [1020, 1021]. Second, high dynamic range images may be prepared for
display on low dynamic range display devices by applying a tone-reproduction
operator. The purpose of tone reproduction is therefore to reduce the dynamic
range of an image such that it may be displayed on a display device.
We can use a simple compressive function to normalize an image (see Fig-
ure 17.5 (left)). This constitutes a linear scaling, which is sufficient only if the
dynamic range of the image is slightly higher than the dynamic range of the dis-
play. For images with a higher dynamic range, small intensity differences will be
quantized to the same display value and visible details are lost. In Figure 17.5
(middle), all large pixel values are clamped. This makes the normalization less
dependent on noisy outliers, but here we lose information in the bright areas of the
image. In comparison, the right image in Figure 17.5 is tone-mapped, showing
detail in both the light and dark regions.
In general linear scaling will not be appropriate for tone reproduction. The
key issue in tone reproduction is to compress an image, while at the same time
preserving one or more attributes of the image. Different tone-reproduction algo-
rithms focus on different attributes such as contrast, visible detail, brightness, or
appearance.
Ideally, displaying a tone-mapped image on a low dynamic range display de-
vice would create the same visual response in the observer as the original scene.
Given the limitations of display devices, this is, in general, not achievable, al-
though we may approximate this goal as closely as possible.

i i

i i
 i i

i i

17.1. Spatial Operators 885

An orthogonal issue is that of choosing parameters. Dependent on the chosen

tone-reproduction operator, different parameters may have to be chosen by the
user. Some operators do not require any parameter adjustment, whereas most op-
erators have a few parameter values that help control the final appearance. At the
other end of the scale, more advanced user interfaces can be designed to guide the
compression algorithm. Such interfaces may consist of scribble tools to roughly
indicate which areas are to be affected [686]. In this chapter, however, we focus
predominantly on automatic and semi-automatic operators.

17.1 Spatial Operators

In the following sections we discuss tone-reproduction operators that apply com-
pression directly on pixels. Often global and local operators are distinguished.
Tone-reproduction operators in the former class change each pixel’s luminance
values according to a compressive function that is the same for each pixel [264,
310, 780, 1007, 1152, 1211, 1213]. The term global stems from the fact that many
such functions need to be anchored to some values determined by analyzing the
full image. In practice, most operators use the geometric average L̄v to steer the
compression:  
1
N∑
L̄v = exp log (δ + Lv (x, y)) . (17.1)
x,y

The small constant δ is introduced to prevent the average from becoming zero in
the presence of black pixels. Note that absolute black tends to occur mostly in
computer-generated images. In the real world, it is difficult to find environments
with a complete absence of photons. The geometric average is normally mapped
to a predefined display value. The effect of mapping the geometric average to
different display values is shown in Figure 17.6.
Alternatively, sometimes the minimum or maximum luminance found in the
image is used. This approach tends to suffer from the same problems as dynamic
range measures that are based on the minimum and maximum luminance: they
are sensitive to outliers.
The main challenge faced in the design of a global operator lies in the choice
of compressive function. Many functions are possible, for instance based on the
image’s histogram (Section 17.7) [1211] or on data gathered from psychophysics
(Section 17.2).
On the other hand, local operators compress each pixel according to a specific
compression function that is modulated by information derived from a selection
of neighboring pixels, rather than from the full image [47, 174, 178, 270, 291, 552,

i i

i i
 i i

i i

886 17. Dynamic Range Reduction

Figure 17.6. Spatial tone mapping applied after mapping the geometric average to differ-
ent display values (left: 0.09, right: 0.18.); Schloß Dagstuhl, Germany, May 2006.

856, 883, 884, 931, 932, 950]. The rationale is that a bright pixel in a bright neigh-
borhood may be perceived differently than a bright pixel in a dim neighborhood.
Design challenges for local operators involve choosing the compressive function,
the size of the local neighborhood for each pixel, and the manner in which lo-
cal pixel values are used. In general, local operators are able to achieve better
compression than global operators (Figure 17.7), albeit at a higher computational
cost.

Figure 17.7. A local tone-reproduction operator (left) and a global tone-reproduction

operator (right) [950]. The local operator shows more detail, as for instance seen in the
insets; Mont St. Michel, France, June 2005.

i i

i i
 i i

i i

17.1. Spatial Operators 887

Both global and local operators are often inspired by the human visual system.
Most operators employ one of two distinct compressive functions; this is orthog-
onal to the distinction between local and global operators. Display values Ld (x, y)
are most commonly derived from image luminance Lv (x, y) by the following two
functional forms:
Lv (x, y)
Ld (x, y) = , (17.2a)
f (x, y)
Lvn (x, y)
Ld (x, y) = . (17.2b)
Lvn (x, y) + gn (x, y)
In these equations, f (x, y) and g(x, y) may either be constant or a function that
varies per pixel. In the former case, we have a global operator, whereas a spatially
varying function results in a local operator. The exponent n is a constant that is
either fixed, or set differently per image.
Equation (17.2a) divides each pixel’s luminance by a value derived from ei-
ther the full image or a local neighborhood. As an example, the substitution
f (x, y) = Lmax /255 in (17.2a) yields a linear scaling such that values may be di-

Figure 17.8. Halos are artifacts commonly associated with local tone-reproduction opera-
tors. Chiu’s operator is used here without smoothing iterations to demonstrate the effect of
division (left). Several mechanisms exist to minimize halos, for instance by using the scale-
selection mechanism used in photographic tone mapping (right); St. Barnabas Monastery,
North Cyprus, July 2006.

i i

i i
 i i

i i

888 17. Dynamic Range Reduction

rectly quantized into a byte, and they can therefore be displayed. A different
approach is to substitute f (x, y) = Lblur (x, y), i.e., divide each pixel by a weighted
local average, perhaps obtained by applying a Gaussian filter to the image [174].
While this local operator yields a displayable image, it highlights a classical prob-
lem whereby areas near bright spots are reproduced too dark. This is often seen
as halos, as demonstrated in Figure 17.8.
The cause of halos stems from the fact that Gaussian filters blur across sharp
contrast edges in the same way that they blur small details. If there is a large con-
trast gradient in the neighborhood of the pixel under consideration, this causes
the Gaussian blurred pixel to be significantly different from the pixel itself. By
using a very large filter kernel in a division-based approach, such large contrasts
are averaged out, and the occurrence of halos can be minimized. However, very
large filter kernels tend to compute a local average that is not substantially differ-
ent from the global average. In the limit that the size of the filter kernel tends to
infinity, the local average becomes identical to the global average and, therefore,
limits the compressive power of the operator to be no better than a global operator.
Thus, the size of the filter kernel in division-based operators presents a trade-off
between the ability to reduce the dynamic range and the visibility of artifacts.

17.2 Sigmoidal Compression

Equation 17.2b has an S-shaped curve on a log-linear plot and is called a sig-
moid for that reason. This functional form fits data obtained from measuring the
electrical response of photoreceptors to flashes of light in various species [811].
It has also provided a good fit to other electro-physiological and psychophysical
measurements of human visual function [484, 485, 594].
Sigmoids have several desirable properties. For very small luminance values,
the mapping is approximately linear, so that contrast is preserved in dark areas
of the image. The function has an asymptote at 1, which means that the output
mapping is always bounded between 0 and 1. A further advantage of this function
is that for intermediate values, the function affords an approximately logarithmic
compression. This can be seen for instance in Figure 17.9, where the middle
section of the curve is approximately linear on a log-linear plot. To illustrate,
both Ld = Lw /(Lw + 1) and Ld = 0.25 log(Lw ) + 0.5 are plotted in this figure,1
showing that these functions are very similar over a range centered around 1.
In Equation 17.2b, the function g(x, y) may be computed as a global con-
stant, or as a spatially varying function. Following common practice in electro-
1 The constants 0.25 and 0.5 were determined by equating the values as well as the derivatives of

both functions for x = 1.

i i

i i
 i i

i i

17.2. Sigmoidal Compression 889

Ld 1.0
Ld = Lw / (Lw + 1)
Ld = 0.25 log (Lw) + 0.5

0.5

0.0
-2 -1 0 1 2
log (L w )

Figure 17.9. Over the middle-range values, sigmoidal compression is approximately log-
arithmic.

1.0 Ld = Lw / (L w + 10)
Ld = Lw / (L w + 1)
Ld = Lw / (L w + 0.1)

0.5

0.0
-2 -1 0 1 2

Figure 17.10. The choice of semi-saturation constant determines how input values are
mapped to display values.

physiology, we call g(x, y) the “semi-saturation” constant. Its value determines

which values in the input image are optimally visible after tone mapping. In par-
ticular, if we assume that the exponent n equals 1, then luminance values equal to
the semi-saturation constant will be mapped to 0.5, as plotted in Figure 17.9. The
effect of choosing different semi-saturation constants is shown in Figure 17.10.
The function g(x, y) may be computed in several different ways. In its sim-
plest form, g(x, y) is set to L̄v /k, so that the geometric average is mapped to user
parameter k [950]. In this case, a good initial value for k is 0.18, which con-
forms to the photographic equivalent of middle gray. For particularly bright or
dark scenes, this value may be raised or lowered. Alternatively, its value may be
estimated from the image itself [954]. The exponent n in Equation 17.2b may be
set to 1; other choices for n are discussed below.
In this approach, the semi-saturation constant is a function of the geometric
average, and the operator is therefore global. A variation of this global opera-
tor computes the semi-saturation constant by linearly interpolating between the

i i

i i
 i i

i i

890 17. Dynamic Range Reduction

geometric average and each pixel’s luminance [948]:

g(x, y) = a Lv (x, y) + (1 − a) L̄v . (17.3)

The interpolation is governed by user parameter a ∈ [0, 1] which has the effect
of varying the amount of contrast in the displayable image (Figure 17.11). More
contrast means less visible detail in the light and dark areas and vice versa. This
interpolation may be viewed as a half-way house between a fully global and a
fully local operator by interpolating between the two extremes without resorting
to expensive blurring operations.
Although operators typically compress luminance values, this particular op-
erator may be extended to include a simple form of chromatic adaptation. It thus
presents an opportunity to adjust the level of saturation normally associated with
tone mapping, as discussed at the beginning of this chapter.
Rather than compress the luminance channel only, sigmoidal compression is
applied to each of the three color channels:
Ir (x, y)
Ir,d (x, y) = , (17.4a)
Ir (x, y) + gn (x, y)
Ig (x, y)
Ig,d (x, y) = , (17.4b)
Ig (x, y) + gn (x, y)
Ib (x, y)
Ib,d (x, y) = . (17.4c)
Ib (x, y) + gn (x, y)
The computation of g(x, y) is also modified to bilinearly interpolate between the
geometric average luminance and pixel luminance, and between each independent
color channel and the pixel’s luminance value. We therefore compute the geomet-
ric average luminance value L̄v , as well as the geometric average of the red, green,
and blue channels (I¯r , I¯g and I¯b ). From these values, we compute g(x, y) for each
pixel and for each color channel independently. The equation for the red channel
(gr (x, y)) is then

Gr (x, y) = c Ir (x, y) + (1 − c) Lv (x, y), (17.5a)

Ḡr (x, y) = c I¯r + (1 − c) L̄v , (17.5b)
gr (x, y) = a Gr (x, y) + (1 − a) Ḡr (x, y). (17.5c)

The green and blue channels are computed similarly. The interpolation parameter
a steers the amount of contrast as before, and the new interpolation parameter
c ∈ [0, 1] allows a simple form of color correction, as shown in Figure 17.11.
So far we have not discussed the value of the exponent n in Equation (17.2b).
Studies in electro-physiology report values between n = 0.2 and n = 0.9 [485].

i i

i i
 i i

i i

17.2. Sigmoidal Compression 891

Figure 17.11. Linear interpolation of a varies contrast in the tone-mapped image: this
is set to 0.0 (left) and to 1.0 (right). Linear interpolation according to c yields varying
amounts of color correction: c was varied between 0.0 (top) and 1.0 (bottom); Rennes,
France, June 2005.

While the exponent may be user-specified, for a wide variety of images we may
estimate a reasonable value from the geometric average luminance L̄v and the min-
imum and maximum luminance in the image (Lmin and Lmax ) with the following

i i

i i
 i i

i i

892 17. Dynamic Range Reduction

empirical equation:
 1.4
Lmax − L̄v
n = 0.3 + 0.7 . (17.6)
Lmax − Lmin

The different variants of sigmoidal compression shown above are all global
in nature. This has the advantage that they are fast to compute, and they are very
suitable for medium to high dynamic range images. Their simplicity makes these
operators suitable for implementation on graphics hardware as well. For very high
dynamic range images, however, it may be necessary to resort to a local operator
since this may give somewhat better compression.
A straightforward method to extend sigmoidal compression replaces the
global semi-saturation constant by a spatially varying function that also can be
computed in several different ways. Thus, g(x, y) then becomes a function of a
spatially localized average. Perhaps the simplest way to accomplish this is to
again use a Gaussian blurred image. Each pixel in a blurred image represents
a locally averaged value that may be viewed as a suitable choice for the semi-
saturation constant.2
As with division-based operators discussed in the previous section, we have
to consider haloing artifacts. If sigmoids are used with a spatially varying semi-
saturation constant, the Gaussian filter kernel is typically chosen to be very small
in order to minimize artifacts. In practice, filter kernels of only a few pixels wide
are sufficient to suppress significant artifacts while at the same time producing
more local contrast in the tone-mapped images. Such small filter kernels can
be conveniently computed in the spatial domain without losing too much per-
formance. There are, however, several different approaches to compute a local
average; these are discussed in the following section.

17.3 Local Neighborhoods

For local schemes, the compression of each pixel is affected by a local neighbor-
hood of pixels. Whether used within a division-based approach, or in sigmoidal
compression, the size of the neighborhood is important. For division-based ap-
proaches, larger kernels are typically used to minimize artifacts. For sigmoidal
techniques, small filter kernels produce the least objectionable artifacts.
In either case, halo artifacts occur when the local average is computed over a
region that contains sharp contrasts with respect to the pixel under consideration.
2 Although g(x, y) is now no longer a constant, we continue to refer to it as the semi-saturation

constant.

i i

i i
 i i

i i

17.3. Local Neighborhoods 893

It is therefore important that the local average is computed over pixel values that
are not significantly different from the pixel that is being filtered.
This suggests a strategy whereby an image is filtered such that no blurring
over such edges occurs. A simple, but computationally expensive, way is to com-
pute a stack of Gaussian blurred images with different kernel sizes, i.e., an image
pyramid. For each pixel, we may choose the largest Gaussian that does not over-
lap with a significant gradient. The scale at which this happens can be computed
as follows.
In a relatively uniform neighborhood, the value of a Gaussian blurred pixel
should be the same regardless of the filter-kernel size. Thus, in this case, the dif-
ference between a pixel filtered with two different Gaussians should be around
zero. This difference will only change significantly if the wider filter kernel over-
laps with a neighborhood containing a sharp contrast step, whereas the smaller
filter kernel does not. A difference of Gaussians (DoG) signal LiDoG (x, y) at scale
i can be computed as follows:
LiDoG (x, y) = Ri σ (x, y) − R2 i σ (x, y). (17.7)
It is now possible to find the largest neighborhood around a pixel that does not
contain sharp edges by examining differences of Gaussians at different kernel
sizes i [950]:
" DoG "
" Li (x, y) "
" "
" Ri σ (x, y) + α " > t i = 1 . . . 8. (17.8)

Here, the difference-of-Gaussians filter is divided by one of the Gaussians to nor-

malize the result and, thus, enable comparison against the constant threshold t that
determines if the neighborhood at scale i is considered to have significant detail.
The constant α is added to avoid division by zero.
For the image shown in Figure 17.12 (left), the scale selected for each pixel
is shown in Figure 17.12 (right). Such a scale selection mechanism is employed
by the photographic tone-reproduction operator [950] as well as in Ashikhmin’s
operator [47].
Once the appropriate neighborhood for each pixel is known, the Gaussian
blurred average Lblur for this neighborhood may be used to steer the semi-
saturation constant, such as for instance employed by the photographic tone-
reproduction operator:
Lw
Ld = . (17.9)
1 + Lblur
It is instructive to compare the result of this operator with its global equivalent,
defined as
Lw
Ld = . (17.10)
1 + Lw

i i

i i
 i i

i i

894 17. Dynamic Range Reduction

Figure 17.12. Scale selection mechanism. The left image shows the tone-mapped result.
The image on the right encodes the selected scale for each pixel as a gray value—the darker
the pixel, the smaller the scale. A total of eight different scales were used to compute this
image; Clifton Suspension Bridge, Bristol, UK, April 2006.

Images tone mapped with both forms are shown in Figure 17.13. In addition,
this figure shows the color difference as computed with the CIE94 color difference
metric, presented in Section 8.8.2. This image shows that the main differences oc-
cur near (but not precisely at) high-frequency high-contrast edges, predominantly

Figure 17.13. This image was tone mapped with both global and local versions of the
photographic tone-reproduction operator (top left, and top right, respectively). Below, for
∗ color difference is shown; Stonehenge, UK, October 2006.
each pixel, the CIE ΔEab

i i

i i
 i i

i i

17.4. Sub-Band Systems 895

seen in the clouds. These are the regions where more detail is produced by the
local operator.
Image pyramids may also be used to great effect to tone map an image directly.
Careful design of the filter stack may yield results that are essentially free of halos,
as discussed in the following section.
An alternative approach includes the use of edge-preserving smoothing oper-
ators that are designed specifically for removing small details while keeping sharp
contrasts intact. Such filters have the advantage that sharp discontinuities in the
filtered result coincide with the same sharp discontinuities in the input image, and
they may therefore help to prevent halos [253]. Several such filters, for example,
the bilateral filter, trilateral filter, Susan filter, the LCIS algorithm, and the mean
shift algorithm are suitable [178, 204, 270, 883, 1153], although some of them
are expensive to compute. Edge-preserving smoothing operators are discussed in
Section 17.5.

17.4 Sub-Band Systems

In the preceding section, an image pyramid was used to determine the size of a
local area around each pixel within which no sharp edges cross. The weighted
average of the pixel values in this area is considered to be representative of a local
adaptation value, and can be used as such.
Image pyramids can also be used directly for the purpose of tone reproduction,
provided the filter bank is designed carefully [681]. Here, a signal is decomposed
into a set of signals that can then be summed to reconstruct the original signal.
Such algorithms are known as sub-band systems, or alternatively as wavelet tech-
niques, multi-scale techniques or image pyramids.
An image consisting of luminance signals Lv (x, y) is split into a stack of band-
pass signals LiDoG (x, y) using (17.7) whereby the n scales, i, increase by factors
of two. The original signal Lv can then be reconstructed by simply summing the
sub-bands:
n
Lv (x, y) = ∑ LiDoG (x, y). (17.11)
i=1

Assuming that Lv is a high dynamic range image, a tone-mapped image can be

created by first applying a non-linearity to the band-pass signals. In its simplest
form, the non-linearity applied to each LiDoG (x, y) is a sigmoid [253], such as given
in (17.2b). However, as argued by Li et al., summing the filtered sub-bands then
leads to distortions in the reconstructed signal [681]. To limit such distortions,
either the filter bank may be modified or the non-linearity may be redesigned.

i i

i i
 i i

i i

896 17. Dynamic Range Reduction

Although sigmoids are smooth functions, their application to a (sub-band) sig-

nal with arbitrarily sharp discontinuities will yield signals with potentially high
frequencies. The high-frequency content of sub-band signals will cause distor-
tions in the reconstruction of the tone-mapped signal Lv [681]. The effective
gain G(x, y) applied to each sub-band as a result of applying a sigmoid can be
expressed as a pixel-wise multiplier:

LDoG
i (x, y) = LiDoG (x, y) G(x, y). (17.12)

To avoid distortions in the reconstructed image, the effective gain should have
frequencies no higher than the frequencies present in the sub-band signal. This
can be achieved by blurring the effective gain map G(x, y) before applying it to
the sub-band signal. This approach leads to a significant reduction in artifacts,
and it is an important tool in the prevention of halos.
The filter bank itself may also be adjusted to limit distortions in the recon-
structed signal. In particular, to remove undesired frequencies in each of the sub-
bands caused by applying a non-linear function, a second bank of filters may be
applied before summing the sub-bands to yield the reconstructed signal. If the
first filter bank that splits the signal into sub-bands is called the analysis filter
bank, then the second bank is called the synthesis filter bank. The non-linearity
described above can then be applied in between the two filter banks. Each of
the synthesis filters should be tuned to the same frequencies as the corresponding
analysis filters.
An efficient implementation of this approach, which produces excellent
artifact-free results, is described by Li et al. [681], and an example image is shown
in Figure 17.14. The source code for this operator is available at http://www.mit.
edu/∼yzli/hdr companding.htm. The image benefits from clamping the bottom
2% and the top 1% of the pixels; this is discussed further in Section 17.10.2.
Although this method produces excellent artifact-free images, it has the ten-
dency to over-saturate the image. This effect was ameliorated in Figure 17.14 by
desaturating the image using the technique described in Section 17.10.1 (with a
value of 0.7, which is the default value used for the sub-band approach). How-
ever, even after desaturating the image, its color fidelity remained a little too sat-
urated. Further research would be required to determine the exact cause of this
effect, which is shared with gradient-domain compression (Section 17.6). Finally,
it should be noted that the scene depicted in Figure 17.14 is a particularly chal-
lenging image for tone reproduction. The effects described here would be less
pronounced for many high dynamic range photographs.

i i

i i
 i i

i i

17.5. Edge-Preserving Smoothing Operators 897

Figure 17.14. Tone reproduction using a sub-band architecture, computed here using a
Haar filter [681]; Gymnasium of Salamis, North Cyprus, July 2006.

17.5 Edge-Preserving Smoothing Operators

An edge-preserving smoothing operator attempts to remove details from the im-
age without removing high-contrast edges. An example is the bilateral filter [871,
1131, 1231], which is a spatial Gaussian filter multiplied with a second Gaussian
operating in the intensity domain. With Lv (x, y) representing the luminance at
pixel (x, y), and using (15.19), the bilateral filter LB (x, y) is defined as
∑u ∑v Rσ1 (x − u, y − v) Rσ2 (Lv (x, y), Lv (u, v)) Lv (u, v)
LB (x, y) = . (17.13)
∑u ∑v Rσ1 (x − u, y − v) Rσ2 (Lv (x, y), Lv (u, v))
Here, σ1 is the kernel size used for the Gaussian operating in the spatial domain,
and σ2 is the kernel size of the intensity domain Gaussian filter. The bilateral filter
can be used to separate an image into “base” and “detail” layers [270]. Applying
the bilateral filter to an image results in a blurred image in which sharp edges re-
main present (Figure 17.15 (left)). Such an image is normally called a base layer.
This layer has a dynamic range similar to the input image. The high dynamic
range image may be divided pixel-wise by the base layer, obtaining a detail layer
LD (x, y), which contains all the high frequency detail, but typically does not have
a very high dynamic range (Figure 17.15 (right)):

LD (x, y) = Lv (x, y)/LB (x, y). (17.14)

By compressing the base layer before recombining into a compressed image,

a displayable low dynamic range image may be created (Figure 17.16). Compres-

i i

i i
 i i

i i

898 17. Dynamic Range Reduction

Figure 17.15. Bilateral filtering removes small details, but preserves sharp gradients (left).
The associated detail layer is shown on the right; Bristol, UK, August 2005.

Figure 17.16. An image tone mapped using bilateral filtering. The base and detail lay-
ers shown in Figure 17.15 are recombined after compressing the base layer; Bristol, UK,
August 2005.

sion of the base layer may be achieved by linear scaling. Tone reproduction on
the basis of bilateral filtering is executed in the logarithmic domain.
Aside from bilateral filtering, other operators may be applicable, such as the
mean-shift filter [204], the trilateral filter [178], anisotropic diffusion [890], an
adaptation of a Retinex-like filter [775], or even the median filter (Section 15.2.3).
Edge-preserving smoothing operators may also be used to compute a local
adaptation level for each pixel, to be applied in a spatially varying or local tone-
reproduction operator. A local operator based on sigmoidal compression can, for
instance, be created by substituting Lblur (x, y) = LB (x, y) in (17.9).

i i

i i
 i i

i i

17.6. Gradient-Domain Operators 899

Alternatively, the semi-saturation constant g(x, y) in (17.4) may be seen as a

local adaptation constant, and can therefore be locally approximated [658, 777],
for instance, with LB (x, y) or with any of the other filters mentioned above. As
shown in Figure 17.10, the choice of semi-saturation constant shifts the curve
horizontally such that its middle portion lies over a desirable range of values. In
a local operator, this shift is determined by the values of a local neighborhood
of pixels and is thus different for each pixel. This leads to a potentially better
compression mechanism than a constant value could afford.

17.6 Gradient-Domain Operators

Local adaptation provides a measure of how different a pixel is from its immedi-
ate neighborhood. If a pixel is very different from its neighborhood, it typically
needs to be more attenuated. Such a difference may also be expressed in terms of
contrast, which can be represented with image gradients (in log space):

∇L = (L(x + 1, y) − L(x, y), L(x, y + 1) − L(y)) . (17.15)

Figure 17.17. The image on the left is tone mapped using gradient domain compression.
The magnitude of the gradients ∇L is mapped to a greyscale in the right image (white is
a gradient of 0; black is the maximum gradient in the image); Mont St. Michel, June 2005.

i i

i i
 i i

i i

900 17. Dynamic Range Reduction

Here, ∇L is a vector-valued gradient field. By attenuating large gradients more

than small gradients, a tone-reproduction operator may be constructed [306]. Af-
terwards, an image can be reconstructed by integrating the gradient field to form
a tone-mapped image. Such integration must be approximated by numerical tech-
niques, by solving a Poisson equation using the Full Multigrid Method [926]. The
resulting image then needs to be linearly scaled to fit the range of the target display
device. An example of gradient domain compression is shown in Figure 17.17.

17.7 Histogram Adjustment

A simple but effective approach to tone reproduction is to derive a mapping from
input luminance to display luminance using the histogram of the input image.
Histogram equalization simply adjusts the luminance so that the probability that
each display value occurs in the output image is equal. Such a mapping is cre-
ated by computing the image’s histogram, and then integrating this histogram to
produce a cumulative histogram. This function can be used directly to map input
luminance to display values.
However, dependent on the shape of the histogram, it is possible that some
contrasts in the image are exaggerated, rather than attenuated. This produces
unnatural results, which may be overcome by restricting the cumulative histogram

L d 300

250

200

150

100

50

0
0 1 2 3 4 5 6
Log 10 L

Figure 17.18. Tone-mapping function created by reshaping the cumulative histogram of

the image shown in Figure 17.19.

i i

i i
 i i

i i

17.8. Lightness Perception 901

Figure 17.19. Image tone mapped using the histogram-adjustment technique; St. Barn-
abas monastery, North Cyprus, July 2006.

to never attain a slope that is too large. The magnitude of the slope that can
be maximally attained is determined by a threshold-versus-intensity curve (TVI)
curve (See Section 5.5.2). The method is then called histogram adjustment, rather
than histogram equalization [1211]. An example of a display mapping generated
by this method is shown in Figure 17.18. This mapping is derived from the image
shown in Figure 17.19.
This method can be extended to include the observer’s state of adaptation
and a time course of adaptation [529]. Such an operator, optimized to preserve
visibility, is then able to simulate what a visually-impaired person would see. The
TVI curve that is used in this algorithm is extended to account for adaptation,
which in turn is governed by a temporal factor modeling the amount of bleaching
occurring in the retina.

17.8 Lightness Perception

The theory of lightness perception, discussed in Chapter 5, provides a model for
the perception of surface reflectance (recall that lightness is defined as perceived
surface reflectance). To cast this into a computational model, the image needs to
be automatically decomposed into frameworks, i.e., regions of common illumina-

i i

i i
 i i

i i

902 17. Dynamic Range Reduction

tion. The influence of each framework on the total lightness needs to be estimated,
and the anchors within each framework must be computed [618–620].
It is desirable to assign a probability of belonging to a particular framework
to each pixel. This leaves the possibility of a pixel having non-zero participation
in multiple frameworks, which is somewhat different from standard segmentation
algorithms that assign a pixel to at most one segment. To compute frameworks
and probabilities for each pixel, a standard K-means clustering algorithm may be
applied. This algorithm results in a set of n centroids dotted around the image. For
a centroid with a given luminance value LC , the probability of a pixel belonging
to the framework defined by this centroid is given by
 
− log10 (LC /Lv (x, y))2
PC (x, y) = exp , (17.16)
2σ2
where σ is the maximum difference between adjacent centroids. If no pixels
belong to a centroid with a probability higher than 0.95, this centroid is merged
with a neighbor.
The probability map thus obtained is the first step towards defined frame-
works. The concept of proximity, however, is not yet accounted for. This feature
may be implemented by filtering the probability map with a bilateral filter (17.13),
with σ1 set to half of the smaller of the image width and height and σ2 set to 0.4.
Once frameworks are computed, their impact on the image’s total lightness
can be assessed. The strength of a framework depends on both its articulation
and its relative size (see Section 5.8.5). Articulation can be estimated from the
minimum and maximum luminance within each framework:
  2 
− log10 LCmax /LCmin
AC = 1 − exp , (17.17)
2σ 2

where σ = 1/3. The size factor of the relative framework XC is computed from
the number of pixels in the framework NC as follows:
 
−NC2
XC = 1 − exp , (17.18)
2σ2
where σ = 0.1. The probability maps PC are multiplied by their respective artic-
ulation and size factors to obtain the final decomposition into frameworks:

PC (x, y) = PC (x, y) AC XC . (17.19)

Afterwards, these probabilities are normalized so that, for each pixel, they sum
to 1.

i i

i i
 i i

i i

17.8. Lightness Perception 903

For each framework, the highest luminance rule may now be applied to find
an anchor. However, direct application of this rule may result in the selection of
a luminance value of a patch that is perceived as self-luminous. As the anchor
should be the highest luminance value that is not perceived as self-luminous, se-
lection of the highest luminance value should be preceded by filtering the area of
the local framework with a large Gaussian filter.
Assuming that WC is the anchor selected for local framework C, the net light-
ness of the whole image can be computed as follows:

Llightness (x, y) = 0.7 log10 (Lv (x, y)/W0 ) + 0.3 ∑ PC (x, y) log10 (Lv (x, y)/WC ) .
C
(17.20)
Here, W0 is the anchor of the global framework. This then constitutes a com-
putational model of lightness perception that can be extended for the purpose of
tone reproduction. To achieve dynamic range reduction, the above equation can
be augmented with individual scale factors fC that scale the output of each of the
frameworks independently:

log10 (Lv (x, y)/WC )

Ld (x, y) = 0.7 log10 (Lv (x, y)/W0 ) + 0.3 ∑ PC (x, y) . (17.21)
C fC

The factors fC are chosen such that each framework fits within the display capa-
bilities of the target display device.
One of the strengths of using a computational model of lightness perception
for the purpose of tone reproduction, is that traditionally difficult phenomena such
as the Gelb effect, introduced in Section 5.8.5, can be handled correctly. To

1 2 3 4 5

Top: Lightness Model

Middle: Photographic Tone Reproduction
Bottom: Bilateral Filtering

Figure 17.20. Gelb patches as reproduced by three different tone-reproduction operators.

(Figure courtesy of Grzegorz Krawczyk, Karol Myszkowski and Hans-Peter Seidel, Max
Planck Institute for Informatics, Saarbrücken, Germany.)

i i

i i
 i i

i i

904 17. Dynamic Range Reduction

Lightness
Lightness Model Photographic Tone Reproduction
0 0

-1 -1

-2 -2
-1.2 -1.0 -0.8 -0.6 -0.4 -0.2 0.0 0.2 -1.2 -1.0 -0.8 -0.6 -0.4 -0.2 0.0 0.2
log 10 L log 10 L
Lightness

0 Bilateral Filtering
1
2
3
-1 4
5

-2
-1.2 -1.0 -0.8 -0.6 -0.4 -0.2 0.0 0.2
log 10 L

Figure 17.21. Comparison of performance for three different tone-reproduction algo-

rithms with respect to reproducing the Gelb effect. The numbers refer to the test images
shown in Figure 17.20. (Figure courtesy of Grzegorz Krawczyk, Karol Myszkowski and
Hans-Peter Seidel, Max Planck Institute for Informatics, Saarbrücken, Germany.)

demonstrate, Figure 17.20 shows the output of the lightness model for a set of
stimuli and compares them against the photographic tone-reproduction operator
as well as the bilateral filter. By visual inspection, the lightness model preserves
the Gelb effect, whereas the other two tone-reproduction operators break down
to different degrees. This is confirmed by plotting the lightness values against
luminance input, as shown in Figure 17.21.
This model of lightness perception may have applications beyond tone re-
production. In particular, this model appears to be suitable for detecting pixels
that belong to self-luminous surfaces versus pixels that depict bright reflective
surfaces. It would be interesting to see this model applied within the context of
image appearance modeling, where pixels depicting self-luminous surfaces ought
to be treated differently from pixels depicting bright surfaces.
Further, this model may be used for the purpose of white balancing images
of scenes that have different areas illuminated by different light sources. For in-
stance, shops typically illuminate their displays throughout the day, even when
daylight enters through the shop window. Images taken under these circum-

i i

i i
 i i

i i

17.9. Counter Shading 905

stances may have areas illuminated by tungsten light and other areas illuminated
by daylight. The approach discussed here would partition these areas into differ-
ent frameworks and treat them separately.

17.9 Counter Shading

The design of local tone-reproduction operators often involves a trade-off between
the amount of achievable range reduction and the minimization of artifacts. As
discussed in this chapter, halos are often a by-product of local operators, unless
they are specifically designed out of the system.
This desire to remove haloing artifacts is inspired by early results as shown
in Figure 17.8. This image shows particularly large and obtrusive halos (exagger-
ated in this figure for demonstration purposes), and it is clear that these types of
artifacts are best avoided. Recent tone-reproduction operators do an excellent job
of avoiding halos.
However, at the same time, artists have used halos and related hatchings and
shadings on purpose to suggest increased contrast in drawings and paintings. In
those cases, the effect is called counter shading. Further, the existence of the
Cornsweet-Craik-O’Brien illusion (Section 5.4.4) suggests that small contrasts
can be enhanced by applying a halo-like gradient on either side of an edge.
As a result, it appears that it should be possible to apply counter shading in
a controlled fashion within a tone-reproduction operator to increase the appar-
ent contrast, without the results being perceived as artifacts. A first attempt at
deliberately applying counter shading for the purpose of enhancing contrast has

Figure 17.22. An example of algorithmic counter shading applied to chromatic channels

(left), compared with the same image without counter shading (right); Strasbourg. (Im-
ages courtesy of Kaleigh Smith, Grzegorz Krawczyk, Karol Myszkowski, and Hans-Peter
Seidel; Max Planck Institute for Informatics, Saarbrücken, Germany.)

i i

i i
 i i

i i

906 17. Dynamic Range Reduction

recently appeared [1058]. Here, the image is segmented using K-means segmen-
tation [98,724]. At the edge of each segment, contrast is enhanced with a function
that models contrasts as found in the Cornsweet illusion:
  
exp −d 2 /σ 2 (x, y) ∈ segment A,
m(x, y) = 1 + a  2 2 (17.22)
1 − exp −d /σ (x, y) ∈ segment B.
Here, d is the shortest distance to the edge between segment A and segment B,
and σ is a user parameter to specify the width of the enhanced area. It is assumed
that the average luminance in segment A is greater than in segment B. Further,
a is a measure of the amplitude of the scaling and, thus, specifies how strong
the effect is. The value of m is applied as a multiplier to enhance the u* and v*
channels in the CIE L∗ u∗ v∗ color space. Applying counter shading only to the
chromatic channels yields a subtler effect than using the luminance channel. An
example result is shown in Figure 17.22. Note that in this image, halo artifacts are
not obvious, although the overall impression is that of a more saturated image.

17.10 Post-Processing
After tone reproduction, it is possible to apply several post-processing steps to ei-
ther improve the appearance of the image, adjust its saturation, or correct for the
gamma of the display device. Here, we discuss two frequently used techniques
that have a relatively large impact on the overall appearance of the tone-mapped
results—a technique to desaturate the results and a technique to clamp a percent-
age of the lightest and darkest pixels.

17.10.1 Color in Tone Reproduction

Tone-reproduction operators normally compress luminance values, rather than
work directly on the red, green, and blue components of a color image. After these
luminance values have been compressed into display values Ld (x, y), a color im-
age may be reconstructed by keeping the ratios between color channels the same
as before compression (using s = 1) [1007]:
 
Ir (x, y) s
Ir,d (x, y) = Ld (x, y), (17.23a)
Lv (x, y)
 
Ig (x, y) s
Ig,d (x, y) = Ld (x, y), (17.23b)
Lv (x, y)
 
Ib (x, y) s
Ib,d (x, y) = Ld (x, y). (17.23c)
Lv (x, y)

i i

i i
 i i

i i

17.10. Post-Processing 907

Figure 17.23. Per-channel gamma correction may desaturate the image. The images were
desaturated with values of s = 0.1 through 0.9 with increments of 0.1; Mainau, Germany,
June 2005.

Alternatively, the saturation constant s may be chosen smaller than 1. Such per-
channel gamma correction may desaturate the results to an appropriate level, as
shown in Figure 17.23 [306].
The results of tone reproduction may appear unnatural, because human color
perception is non-linear with respect to overall luminance level. This means that
if we view an image of a bright outdoor scene on a monitor in a dim environ-
ment, we are adapted to the dim environment rather than the outdoor lighting. By
keeping color ratios constant, we do not take this effect into account. The above

i i

i i
 i i

i i

908 17. Dynamic Range Reduction

approach should therefore be seen as a limited control to account for a complex

phenomenon.
A more comprehensive solution is to incorporate ideas from the field of color
appearance modeling into tone-reproduction operators [292,884,948]. The iCAM
image appearance model is the first color appearance model operating on im-
ages [291, 292, 791]. It can also be used as a tone-reproduction operator. It there-
fore constitutes an important trend towards the incorporation of color appearance
modeling in dynamic range reduction and vice versa.
A rudimentary form of color appearance modeling within a tone-reproduction
operator is afforded by the sigmoidal compression scheme outlined in
Section 17.2 [948]. Nonetheless, we believe that further integration of tone re-
production and color appearance modeling is desirable for the purpose of properly
accounting for the differences in adaptation between scene and viewing
environments.

17.10.2 Clamping
A common post process for tone reproduction is clamping. It is for instance part of
the iCAM model, as well as the sub-band encoding scheme. Clamping is normally
applied to both very dark as well as very light pixels. Rather than specify a hard
threshold beyond which pixels are clamped, a better way is to specify a percentile
of pixels that will be clamped. This gives better control over the final appearance
of the image.
By selecting a percentile of pixels to be clamped, inevitably detail will be lost
in the dark and light areas of the image. However, the remainder of the luminance
values is spread over a larger range, and this creates better detail visibility for
large parts of the image.
The percentage of pixels clamped varies usually between 1% and 5%, depend-
ing on the image. The effect of clamping is shown in Figure 17.24. All images are
treated with the photographic tone-reproduction technique. The images on the left
show the results without clamping, and therefore all pixels are within the display
range. In the image on the top right, the darkest 7% of the pixels are clamped,
as well as the lightest 2% of the pixels. This has resulted in an image that has
reduced visible detail in the steps of the amphitheater, as well as in the wall and
the bushes in the background. However, the overall appearance of the image has
improved, and the clamped image conveys the atmosphere of the environment
better than the directly tone-mapped image.
The bottom-right image of Figure 17.24 also conveys the atmosphere of the
scene better. This photograph was taken during a bright day. In this case, the

i i

i i
 i i

i i

17.10. Post-Processing 909

Figure 17.24. Examples of clamping. All images were tone mapped using photographic
tone reproduction. The left images were not clamped, whereas 7% of the darkest pixels and
2% of the lightest pixels were clamped in the top-right image. Top: Salamis amphitheater,
North Cyprus, July 2006. Bottom: Lala Mustafa Pasha Mosque, Famagusta, North Cyprus,
July 2006.

i i

i i
 i i

i i

910 17. Dynamic Range Reduction

clamping has allowed more detail to become visible in the back wall around the
window.
To preserve the appearance of the environment, the disadvantage of losing
detail in the lightest and darkest areas is frequently outweighed by a better overall
impression of brightness. The photographs shown in Figure 17.24 were taken
during a very bright day, and this is not conveyed well in the unclamped images.
Finally, the effect of clamping has a relatively large effect on the results. For
typical applications, it is an attractive proposition to add this technique to any
tone-reproduction operator. However, as only a few tone-reproduction operators
incorporate this feature as standard, it also clouds the ability to assess the quality
of tone-reproduction operators. The difference between operators appears to be
of similar magnitude as the effect of clamping.

17.11 Validation and Comparison

A recent topic of interest in tone reproduction is validation. It is generally recog-
nized that visually comparing results obtained with different operators is a hap-
hazard way of assessing their quality. A problem with this approach is that dif-
ferent observers apply different criteria to decide which operator they appreciate
best. The sample size, i.e., the number of images that would have to be used to
draw conclusions, is usually too large to allow meaningful discussion in research
papers that are almost always limited in length. With small numbers of images
involved in a visual comparison, there is a distinct danger that the choice of image
has a more profound impact on the outcome than the choice of tone-reproduction
operator.
Other problems with the visual assessment of tone reproduction operators is
that it is sometimes unclear which qualities of a tone reproduction are desirable
to achieve a specific task. Hence, the influence of the task for which a tone-
reproduction operator is chosen is non-negligible. Consider, for instance, two
disparate tasks such as assessment of medical imagery for the purpose of diag-
nosis versus the use of tone-reproduction operators in entertainment (film, video,
games). For the latter, a visually pleasing tone-mapped result may well be all
that is required. Simple visual inspection may be sufficient to decide between the
many tone-reproduction operators currently available.
However, the same could not be said in the case of medical imaging. Here,
the tone-reproduction operator does not have to produce visually appealing re-
sults, but should enable the task—diagnosis of patients—to be carried out with
maximum confidence. Simple visual inspection is likely to be inadequate for this
task.

i i

i i
 i i

i i

17.11. Validation and Comparison 911

Nonetheless, visual comparison does allow expert observers to infer certain

behaviors of tone-reproduction operators. In addition, it is possible to include
more tone-reproduction operators in such a comparison than could be realistically
included in a validation study. As such, there is still value in presenting a side-
by-side comparison of a set of tone-reproduction operators. We therefore present
and discuss a selection of tone-mapped images in Section 17.11.2.
A less subjective approach is to define specific image features that are prefer-
entially preserved in tone-mapped images. It is then possible to measure a high
dynamic range image and its tone-mapped counterpart and apply the metric to the
ratio. If this ratio is close to 1, then the particular image feature is said to be pre-
served, whereas smaller values indicate deviation from the ideal. This approach,
discussed in Section 17.11.3, is less prone to interpretation than visual inspection,
yet is not as involved as a full-blown user study.
As mentioned, for many application areas, the performance of a tone-
reproduction operator is intimately linked to the task for which it is designed.
The notion that a tone-reproduction operator is likely to be good for some specific
tasks, but not all tasks, is currently dominant. It should therefore be possible to
identify real-world tasks for which tone-reproduction operators may be used and
to derive laboratory experiments from these tasks.
The task performance thus recorded under laboratory conditions is then an
indication of how well the real-world task may be performed. Unfortunately,
such validation studies are time consuming, and this is why only recently a few
such studies have appeared. These are discussed in Section 17.11.4.

17.11.1 Operators
For each of the studies discussed in the following, we refer to the original pa-
pers for detailed descriptions of the experimental set-up and analysis. We aim to
provide an overview as well as the main conclusions of each of the studies. The
algorithms included are summarized as follows:

Linear scaling. A global operator, whereby the image is linearly scaled to fit into
the range of the display.

Histogram adjustment. A global operator that derives luminance-mapping func-

tions from the cumulative histogram of the high dynamic range input
[1211].

Time-dependent operator. A global operator that takes the time course of adapta-
tion into account [885].

i i

i i
 i i

i i

912 17. Dynamic Range Reduction

Adaptive logarithmic operator. A global tone-reproduction operator that essen-

tially takes the logarithm for each pixel, but adjusts the base of the loga-
rithm on a per-pixel basis for enhanced performance [264].

Photographic tone reproduction. A tone-reproduction operator based on photo-

graphic practice. It exists in both local and global variants and is discussed
in Section 17.3 [950].

Ashikhmin operator. A local tone-reproduction operator inspired by human vi-

sual perception [47]. It uses a scale-selection mechanism similar to the
local photographic operator.

Bilateral filtering. A local tone-reproduction operator, discussed in Section 17.5

[270].

Trilateral filtering. A refinement of bilateral filtering, it involves the application

of the bilateral filter twice, but in different domains [178].

LCIS. A general image-processing technique to optimize contrast in images.

Similar to anisotropic diffusion, it can be used for the purpose of tone re-
production [1153].

iCAM. A local tone-reproduction operator and color appearance model,

discussed in Chapter 11 [291, 292].

Sub-band encoding. A local tone-reproduction operator, discussed in Section

17.4 [681].

Gradient-domain compression. Tone reproduction based on compressing gradi-

ent fields, discussed in Section 17.6. [306].

Photoreceptor-based compression. Sigmoidal compression, inspired by photore-

ceptors, as discussed in Section 17.2 [948].

Tumblin-Rushmeier operator. A global tone-reproduction operator, based on pre-

serving brightness [1152].

Local eye adaptation. Sigmoidal compression using local adaptation computed

with a bilateral filter, as discussed in Section 17.5. It includes a model of
time-dependent adaptation as well [658].

i i

i i
 i i

i i

17.11. Validation and Comparison 913

17.11.2 Comparison
In this section, we present side-by-side comparisons of several recent tone-
reproduction operators. As discussed above, the successful visual comparison
depends on several parameters:

• Correctness of the implementation of each operator. Where possible, we

have used implementations provided by the authors. However, in most
cases the implementations are taken from the supplemental material of High
Dynamic Range Imaging: Acquisition, Display and Image-Based Light-
ing [952]. These implementations are not known to be faulty, although
correctness cannot be guaranteed.

• Choice of user parameters. Most tone-reproduction operators provide user

parameters that need to be set appropriately to obtain meaningful results.
Only a few operators are either fully automatic, or provide an automatic
mode of operation. Where user parameters need to be specified, they were
chosen so as to produce a flattering result. However, it is impossible to
guarantee in every case that a different combination of parameter settings
would not lead to a better image.

• Choice of test images. The image used in this comparison is only one of
many that we could have chosen. It is well known that some images suit
certain operators better than others. An exhaustive comparison, or even
a visual comparison with a reasonable number of test images, is beyond
the scope of this book. We have therefore tried to select a reasonable test
image, but note that for specific applications, a better test image could have
been chosen.

• Experience of the observer. Even the experts are sometimes unclear as to

what it is they are looking for in a good tone-mapped image. There appears
to be a correlation between images that are most liked and images that look
natural, but this correlation may be weaker than expected. In addition, if a
tone-mapped image could be compared against the real environment from
which it was derived, the choice of which is the most natural-looking result
may change [44].

With these caveats in mind, images of several tone-reproduction operators are

shown in Figure 17.25. The first five images are created with local operators,
followed by linear scaling. The bottom row consists of global operators. For
these particular images, and with the parameter settings chosen by the author,
the local operators all preserve details in the dark areas very well. Details are also

i i

i i
 i i

i i

914 17. Dynamic Range Reduction

Figure 17.25. Visual comparison between several different tone-reproduction opera-

tors. First row: sub-band encoding, bilateral filtering, gradient domain compression;
Second row: Ashikhmin, photographic, linear scaling; Third row: histogram adjustment,
photoreceptor-based, Tumblin-Rushmeier.

i i

i i
 i i

i i

17.11. Validation and Comparison 915

well preserved in the bright areas, except by photographic tone reproduction. This
operator was designed to let bright areas burn out, to maintain a good impression
of brightness. Hence, this operator affords a different trade-off between detail
visibility and overall perception of brightness.
We applied clamping to the linearly scaled operator with 10% of the bottom
pixels and 5% of the top pixels clamped. This has resulted in significant burn-out
in the windows and loss of detail in the dark areas. However, this operator has
created a well exposed mid-range, presenting the walls and the floor well. Without
clamping, this image would appear to be largely black, with the windows still
burnt-out. The linear operator is effectively a baseline result, the quality of which
all other operators will have to exceed to justify their computational expense.
Both the histogram-adjustment operator and the photoreceptor-inspired opera-
tor present sufficient details in the dark areas; however, the histogram-adjustment
operator let the bright area burn out. The bright windows are reasonably well
represented by the photoreceptor-based operator. The overall impression that this
window is very bright in the real environment is well conveyed by both operators.
The Tumblin-Rushmeier operator trades detail visibility for global contrast.
This operator would benefit from additional clamping to improve detail visibility
of the mid-range of luminance values. However, we applied clamping only if
this was part of the original operator, which for Figure 17.25 was only the sub-
band encoding scheme and the linear scaling. Similarly, gamma correction was
applied to all images, except the sub-band encoding technique, as this operator
was designed to handle gamma issues directly.

17.11.3 Evaluation Metrics

In addition to visual comparison, it is possible to derive metrics of desirable im-
age features. Two such metrics are described here: global contrast and detail
visibility [1058].
The impression of global contrast is best evaluated in the brightness domain,
which itself can be approximated by taking the logarithm (see also Section 6.4).
In Section 17.4, an effective gain map G(x, y) was introduced for each pixel. Tone
mapping can thus be seen as multiplying the HDR image by G(x,y). Alternatively,
it is possible to approximate the effective gain map by a constant and view tone
reproduction as a linear function in the log domain:

log10 (Ld (x, y)) ≈ f (log10 (Lv (x, y))) = G1 log10 (Lv (x, y)) + G2 , (17.24)

with G1 and G2 coefficients estimated through linear regression for all tone-
mapped pixel values. Here, f is the estimated tone-reproduction function (lin-
ear in the log domain) that is clamped to the minimum and maximum display

i i

i i
 i i

i i

916 17. Dynamic Range Reduction

Algorithm
photoreceptor
photographic (global)
photographic (local)
lightness perception
bilateral filtering
gradient domain compression
adaptive logarithmic mapping
gamma 2.2
0 0.2 0.4 0.6 0.8 1
dC g

Figure 17.26. Change in global contrast dCg for different tone-reproduction operators
(after Smith, Krawczyk, Myszkowski, and Seidel [1058]).

luminance. The global contrast Δ log10 (Ld ) in the brightness domain can then be
derived using the minimum and maximum brightness in the tone-mapped image:

Δ log10 (Ld ) = f (max(log10 (Lv ))) − f (min(log10 (Lv ))). (17.25)

The change in global contrast dCg as a result of applying the tone-reproduction

operator (approximated by applying the function f ) is then given by
f (max(log10 (Lv ))) − f (min(log10 (Lv )))
dCg = . (17.26)
max(log10 (Lv )) − min(log10 (Lv ))
Note that dCg is equal to G1 as a result of the linear regression applied above, i.e.,
it represents the slope of the line obtained through linear regression.
Thus, the change in global contrast can be interpreted as the ratio of the bright-
ness range in the tone-mapped image and the input HDR image. If this ratio is
close to 1, then global contrast is preserved. For values less than 1, global contrast
is reduced. It is therefore desirable for tone-reproduction operators to show as lit-
tle change in global contrast as possible. In practice, some operators preserve this
feature better than others, as shown in Figure 17.26.
Figure 17.26 shows that many of the operators do a reasonable job of pre-
serving global contrast, with the photographic operator being closest to the ideal
value of 1. On the other end of the scale is the gradient-domain operator that can
only give a very limited impression of global contrast. It makes a clear trade-off
in favor of local detail visibility, as we will show next.
By compressing the global range of values, more headroom is left for ex-
pressing local details. Both global contrast and detail visibility are desirable, and
a trade-off between the two appears to be unavoidable. A metric to compute detail
visibility may be devised as follows.

i i

i i
 i i

i i

17.11. Validation and Comparison 917

In Section 17.5 we argued that a bilateral filter could be used to compute a

local adaptation luminance for each pixel. The output of the bilateral filter LB (x, y)
can be used to compute a measure of local contrast Cl :

max (L(x, y), LB (x, y))

Cl (L, x, y) = log10 . (17.27)
min (L(x, y), LB (x, y))

A model of the response of the human visual system to a given contrast can be
approximated with the following transducer function T (Cl ) [733]:

T (Cl (L, x, y)) = 54.09288Cl (L, x, y)0.41850 . (17.28)

The constants in this equation are chosen such that T (0) = 0 and, for an appro-
priately chosen threshold contrast Ct , we have T (Ct ) = 1. The threshold contrast
was chosen to be 1%, so that Ct = log10 (1.01). For high dynamic range images,
this assumption may be inaccurate. To correct for the unique conditions posed by
high dynamic range imaging, a correction factor may be applied to T (Cl ), leading
to the following model of response to local contrast:

Ct
Cresp (x, y) = T (Cl (L, x, y))  . (17.29)
LB (x, y) + tvi (LB (x, y))
log10
LB (x, y)

This response function depends on a threshold-versus-intensity function, abbrevi-

ated as TVI, that models the smallest change in intensity that a human can observe
as a function of the background intensity (see Section 5.5.2).
Algorithm

photoreceptor
photographic (global)
photographic (local)
lightness perception
bilateral filtering
gradient domain compression
adaptive logarithmic mapping
gamma 2.2
-0.05 0 0.05 0.1 0.15 0.2 0.25
Relative dark area with loss of detail

Figure 17.27. Relative area for which the change in visible detail dCresp went from > 1
to < 1 due to tone reproduction. This graph is for dark areas of the image (after Smith,
Krawczyk, Myszkowski, and Seidel [1058]).

i i

i i
 i i

i i

918 17. Dynamic Range Reduction

Algorithm
photoreceptor
photographic (global)
photographic (local)
lightness perception
bilateral filtering
gradient domain compression
adaptive logarithmic mapping
gamma 2.2

-0.05 0 0.05 0.1 0.15 0.2 0.25

Relative light area with loss of detail

Figure 17.28. Relative area for which the change in visible detail dCresp went from > 1
to < 1 due to tone reproduction. This graph is for light areas of the image (after Smith,
Krawczyk, Myszkowski, and Seidel [1058]).

The change in hypothetical response that occurs when a high dynamic range
image is tone mapped is thus a relevant metric to evaluate how well local details
are preserved. As the unit of measurement of Cresp is in just noticeable differences
(JNDs), values less than 1 are imperceptible and can therefore be set to 0. Also,
if tone reproduction causes the contrast to be reduced from > 1 JND to < 1 JND,
the change is set to 1. In all other cases, the change in detail visibility is simply
the difference in contrast response:
⎧
⎪
⎨1 if Cresp (Lv , x, y) > 1 > Cresp (Ld , x, y),
" "
dCresp (x, y) = 0 if "Cresp (Lv , x, y) −Cresp (Ld , x, y)" < 1,
⎪
⎩
Cresp (Lv , x, y) −Cresp (Ld , x, y) otherwise.
(17.30)
This metric is typically applied directly to pairs of HDR and tone-mapped images.
However, it would be instructive to separate the images into light and dark areas
and apply this metric separately. The threshold is arbitrarily set to the log average
luminance of the HDR image.
The analysis, shown for dark areas in Figure 17.27 and for light areas in Fig-
ure 17.28 measures the relative area of the image where local contrast visible in
the HDR input is mapped to less than one JND by the tone-reproduction operator.
This is the area where visible contrast is lost and, therefore, needs to be as small
as possible.

17.11.4 Validation
Tone reproduction validation studies aim to rank operators in terms of suitability
for a particular task. The design of such experiments thus involve a task that

i i

i i
 i i

i i

17.11. Validation and Comparison 919

observers are asked to perform. The accuracy with which the task is performed is
then said to be a measure for how suitable a tone reproduction is for that particular
task. It is therefore important to select a sensible task, so that practicality of tone-
reproduction operators can be inferred from task performance.
This particular requirement causes several difficulties. For instance, the game
industry is currently moving to high dynamic range imaging. This means that
rendered content needs to be tone mapped first before it can be displayed. The
real-world task is then “entertainment.” Maybe the aim of the game designer is
to achieve a high level of involvement of the player; or perhaps the goal is to
enable players to achieve high scores. In any case, it will be difficult to devise a
laboratory experiment that measures entertainment, involvement, or engagement.
For other tasks—medical visualization was mentioned earlier—the validation
study may be easier to define, but such application-specific validation studies are
currently in their infancy. A notable exception is Park and Montag’s study, which
evaluates tone reproduction operators for their ability to represent non-pictorial
data [872]. As a result, validation studies tend to address issues related to prefer-
ence and/or similarity.
Preference ratings, which measure whether one image is liked better than
another, have various advantages [263, 625, 627]. It is, for instance, not neces-
sary to compare tone-mapped images to real scenes, as only pair-wise compar-
isons between tone-mapped images are required. On the downside, not having
seen the original environment can be viewed as a serious limitation [44]. The
most preferred image may still deviate significantly from the real environment
from which it was derived. Preference scalings will not be able to signal such
deviations.
On the other hand, similarity ratings are designed to determine which tone-
mapped result looks most natural, i.e., most like the real environment that it de-
picts [627, 659, 1278]. Such studies benefit from having the real environment
present; it has been shown that including the real environment in the study may
change observers’ opinions [44].
The first validation study that assessed tone-reproduction operators appeared
in 2002 [263] and measured image preference and, in a second experiment, nat-
uralness. They found that photographic tone reproduction consistently scored
highest, followed by the histogram-adjustment method; LCIS scored low.
In a separate experiment, a set of seven tone-reproduction operators were
compared against two real scenes [1278]. A total of 14 observers rated the tone-
mapped images for overall brightness, overall contrast, visible detail in dark re-
gions, visible detail in light regions, and naturalness. Results of this experiment
are reproduced in Figure 17.29.

i i

i i
 i
i

i
i

920

Naturalness Detail Brightness

-2.0
-1.0
0.0
1.0
2.0
-2.0
-1.0
0.0
1.0
2.0
-2.0
-1.0
0.0
1.0
2.0

Ti Li Ti Li Ti Li
m m ne m ne
e- nea e- e-
de r de ar de ar
H pe
d H pe
d H pe
d
ist ist ist
og Ash ent og Ash ent og Ash ent
ra ra ra
m ikhm m ikhm m ikhm
A in A in A in
dj dj dj
us us us
tm tm tm
Ph en Ph en Ph en
ot Bi t ot Bi t ot Bi t
A ogr la A ogr la A ogr la
da ap te da ap te da ap te
pt h ra pt h ra pt h ra
iv ic ( l iv ic ( l iv ic ( l
e L lo e L lo e L lo
og c a og c a og c a
ar l) ar l) ar l)
ith ith ith

Naturalness (F = 8.11, p < 0.01)

m m m
Detail (dark) (F = 3.18, p < 0.01)
ic ic ic
Overall Brightness (F = 46.14, p < 0.01)

Detail Contrast
-3.0
-2.0
-1.0
0.0
1.0
2.0
3.0

-2.0
-1.0
0.0
1.0

Ti Li
Ti
m
Li
ne m
e- nea
e-
de ar de r
H pe H pe
de
ist d ist
og A
og Ash ent sh nt
ra
ra
m i k m ikhm
A
hm A in
dj in dj
us us
tm tm
Ph en
Ph
ot
en
t ot Bi t
Bi
l A ogr la
A ogr da ap t er
da ap ater
a pt h a
pt h iv ic ( l
iv ic ( l e L lo
e L lo
og cal
og cal
ar ) ar )
ith
ith
m m
ic
Detail (light) (F = 30.36, p < 0.01)
Overall Contrast (F = 8.74, p < 0.01)

ic

ures 17.26, 17.27, and 17.28 (after Yoshida, Blanz, Myszkowski, and Seidel [1278]).
ent image attributes. The green colors correspond to the operators also included in Fig-
Figure 17.29. Ratings for seven tone-reproduction operators with respect to five differ-
17. Dynamic Range Reduction

i
i

i
i
 i i

i i

17.11. Validation and Comparison 921

This experiment reveals that the global operators (linear, histogram adjust-
ment, time-dependent and adaptive logarithmic) are perceived to be overall
brighter than the local operators (bilateral filtering, Ashikhmin, and photographic).
In addition, the global operators have higher overall contrast than the local oper-
ators. Details in dark regions are best preserved by the Ashikhmin and adaptive
logarithmic operators according to this experiment. Details in light regions are
best preserved by the local operators. The histogram-adjustment, photographic
and adaptive-logarithmic operators are perceived to be the most natural reproduc-
tions of the real scenes.
Correlations between each of brightness, contrast, and detail visibility on the
one hand, and naturalness on the other were computed. It was found that none of
these attributes alone can sufficiently explain the naturalness results. It is there-
fore hypothesized that a combination of these attributes, and possibly others, are
responsible for naturalness.
A second, large-scale validation study also asked the question of which op-
erators produces images with the most natural appearance [659]. Images were
ranked in pairs with respect to a reference image, which was displayed on a high
dynamic range display device. Thus, for each trial, the participant viewed a ref-
erence image as well as two tone-mapped images that were created with different
operators. Preference for one of the two tone-mapped images was recorded in the
context of a specific image attribute that varied between experiments. In the first
experiment, overall similarity was tested using 23 scenes. In the next two exper-
iments, visibility and reproduction of detail was tested in light and dark regions
of the image using 10 scenes each. The experiments were carried out for six dif-
ferent tone-reproduction operators, and the analysis produced a ranking for each
image. In Figure 17.30, we summarize the results by showing, for each operator,
how many times it was ranked first, second, and so on.
The overall similarity ranking in this figure indicates that the iCAM model
produces images that were perceived to be closest to the high dynamic range
images, followed by photographic tone reproduction. In this experiment, the
bilateral-filter and the adaptive-logarithmic mapping did not perform as well.
Multiple comparison scores were computed on the data, which revealed that all
operators perform differently in a statistically significant manner, with the excep-
tion of histogram adjustment and the local eye-adaptation model which together
form one group.
The overall significance test was repeated for gray-scale images. Here it was
found that the photographic operator and the iCAM model perform equivalently.
It is speculated that this change in ranking may result from iCAM’s roots in color
appearance modeling, as it produces particularly natural looking color images.

i i

i i
 i i

i i

922 17. Dynamic Range Reduction

Overall Similarity
25
Votes

20

Bilateral
15 Photographic (local)
Adaptive Logarithmic
Histogram Adjustment
10 iCAM
Local Eye Adaptation
5

0
1 2 3 4 5 6
Rank
Similarity in Dark Regions Similarity in Light Regions
Votes

Votes
10 10
9 9
8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
0 0
1 2 3 4 5 6 1 2 3 4 5 6
Rank Rank

Figure 17.30. Number of images ranked for each tone-reproduction operator in experi-
ments performed by Ledda et al. [659].

In bright regions, the iCAM model still outperforms the other operators, but
the photographic operator does not reproduce details in dark regions as well, be-
ing an average performer in this experiment and performing better in dark re-
gions. Each of Ledda’s experiments provides a ranking, which is summarized in
Figure 17.31, showing that in these experiments, iCAM and the photographic op-
erator tend to produce images that are perceived to be closer to the corresponding
high dynamic range images than can be achieved with other operators.
A further user study asking observers to rate images in terms of naturalness
was recently presented [44]. This study assessed in terms of preference (Exper-
iment I). Operators were also compared for naturalness in the absence of a real
scene (Experiment II) and in the presence of a real scene (Experiment III). The

i i

i i
 i i

i i

17.11. Validation and Comparison 923

Overal Similarity Overal Similarity

(Color) (Grayscale)

1 iCAM 1 Photographic (local)

2 Photographic (local) 2 iCAM
3 Histogram Adjustment Local Eye Adaptation
Local Eye Adaptation 3 Histogram Adjustment
4 Adaptive Logarithmic 4 Adaptive Logarithmic
5 Bilateral 5 Bilateral

Detail (Light) Detail (Dark)

1 iCAM 1 Photographic (local)
2 Local Eye Adaptation Local Eye Adaptation
3 Photographic (local) 2 iCAM
Histogram Adjustment 3 Adaptive Logarithmic
4 Bilateral Histogram Adjustment
5 Adaptive Logarithmic 4 Bilateral

Figure 17.31. Ranking of operators in Ledda’s experiments [659]. The boxes around pairs
of operators indicate that these operators were not different within statistical significance.

Experiment I: Image Preference Experiment II: Naturalness

Perceived Quality

Perceived Quality

1.2 1.2

0.8 0.8

0.4 0.4

0.0 0.0
1 2 3 4 5 1 2 3 4 5
Scene Scene
Experiment III: Closest to real scene
1.2
Perceived Quality

Gradient Domain Compression

0.8 Trilateral
Photographic (local)
0.4 Adaptive Logarithmic
Histogram Adjustment

0.0
1 2 3 4
Scene

Figure 17.32. Ratings of naturalness and preference obtained with Ashikhmin and Goral’s
experiment [44].

i i

i i
 i i

i i

924 17. Dynamic Range Reduction

results for each of these experiments are reproduced in Figure 17.32 for five dif-
ferent tone-reproduction operators.
As can be seen, in this particular experiment, the absence of a real scene
caused operators to perform inconsistently across scenes, suggesting that the
choice of scene has a non-negligible influence on the outcome. However, the
results are more consistent if the images could be compared against the real envi-
ronment.
When compared directly against real environments, in Experiment III the
gradient-domain compression technique performs well, followed by adaptive-
logarithmic mapping. The trilateral filter does not produce images that are close
to the real environments in this particular experiment.
For the purpose of assessing different tone reproduction operators, as well as
for the development of an evaluation protocol for testing such operators, Kuang et
al developed a series of experiments centered around visual preference and ren-
dering accuracy in the context of a digital photography task [627].In all three of
their experiments the bilateral filter performed well, with the photographic oper-
ator scoring similarly in two of the three experiments.
In the presence of a real scene, it was found that visual preference correlates
highly with perceived accuracy, suggesting that testing for visual preference is a
reasonable approach, and enables inferences to be made about the accuracy of
tone reproduction algorithms. Further the results obtained with rating scales cor-
relate well with the much more time-consuming paired comparison tests. This
leads to the conclusion that one may use the more convenient rating scales proce-
dure to assess tone reproduction operators.
Finally, a recent study proposed not to use real images, but rather a care-
fully constructed stimulus that enables assessment of tone-reproduction opera-
tors in terms of contrast preservation [25]. It is argued that the strength of the
Cornsweet-Craik-O’Brien illusion, discussed in Section 5.4.4, depends on its lu-
minance profile. If such an illusion is constructed in a high dynamic range image,
then tone-reproduction operators are likely to distort this luminance profile and,
therefore, weaken the strength of this illusion.
It is now possible to infer the illusion’s strength in a two-alternative forced-
choice experiment by comparing the tone-mapped image against a step function
with contrast that is variable between trials. The smaller measured step sizes
are caused by weaker illusions. In turn, the strength of the illusion after tone
reproduction is taken as a measure of how well contrast is preserved.
Although many detailed conclusions can be drawn from this experiment, the
results are generally in agreement with previous tone-reproduction operators. We
infer that the preservation of contrast therefore correlates with people’s subjective

i i

i i
 i i

i i

17.11. Validation and Comparison 925

preference. The reproduction of contrast should therefore be an important design

criterion in the development of tone-reproduction operators, although it cannot be
the only criterion.

17.11.5 User Strategies

Using a high dynamic range display device, it is possible to artificially limit the
dynamic range of the display. This affords the opportunity to assess how users
adjust generic image features, such as overall brightness and contrast, for different
display conditions. Such an experiment has revealed the general strategy that
users apply to adjust an image for display [1279].
In general it was found that parameter adjustments proceed in a similar way
irrespective of the target dynamic range of the display device. This suggests that
tone-reproduction operators do not need to be tailored to specific display devices.
In particular, when high dynamic range display devices become available, it may
not be necessary to adopt new algorithms to prepare images for such displays
when accounting for user preference. Of course, issues relating to visual adapta-
tion to the display device itself, as well as color appearance issues, do still need
to be taken into account.
The strategy adopted by most users is to first adjust the image such that an
image-dependent percentage of pixels is burnt out. Typically, image areas that are
perceived as self-luminous are mapped to the maximum display value. Hence, for
images with a large directly visible light source, users chose a high percentage
of pixels to burn out. For images with small light sources, such as decorated
Christmas trees, a lower percentage of pixels is typically chosen.
Then, contrast is adjusted according to which task is set. If the goal is to
produce a good-looking image, the contrast tends to be exaggerated by up to four
times the contrast of the original high dynamic range image, even if this means
that a large percentage of the darkest pixels needs to be clipped. If the task is
to produce a realistic-looking image, as close as possible to a reference image,
then users avoid clipping light and dark regions, and contrast is not exaggerated.
However, irrespective of task, users tend to over-saturate pixels that belong to
self-luminous objects.
This result indicates that users tend to have a strong preference for both bright
and high-contrast images. This favors algorithms that achieve good detail repro-
duction in both dark and light areas of the image. Tone-reproduction operators
tailored to follow user preferences should be able to achieve this. However, tone-
reproduction operators intended for high fidelity reproduction should not clamp
such large numbers of pixels, especially not in the dark areas.

i i

i i
 i i

i i

926 17. Dynamic Range Reduction

A note of caution is warranted here, though. Dependent on circumstances,

different users may have completely different preferences, so that averaging re-
sponses across participants in subjective experiments may not always be appro-
priate. For instance, some users prefer losing details in dark regions rather than
light ones. Given that participants are relatively consistent in setting the reference
white, their contrast and brightness preferences may be different. As a result, this
would alter their attitude toward losing details in dark regions.

17.11.6 Discussion
Now that more validation studies are becoming available, some careful overall
conclusions may be drawn. It is clear that operators that perform well in one
validation study are not guaranteed to perform well in all validation studies. This
hints at the possibility that these operators are not general purpose, but are suitable
only for certain specific tasks.
On the other hand, it could also be argued that the choice of task, types of
images, as well as further particulars of the experimental designs, have an impact
on the outcome. Ashikhmin’s study has found that having a real scene included
in the experimental paradigm, for instance, significantly changes the result.
Further, there are many different operators, and each validation study neces-
sarily includes only a relatively small subset of these. There are a few operators
that have been included in most of the user studies to date. Of these, the pho-
tographic operator appears to perform consistently well, except in Ashikhmin’s
study. The adaptive-logarithmic mapping performed well in Yoshida’s study as
well as in Ashikhmin’s study, but performed poorly in Ledda’s study.
Finally, Ashikhmin’s study, as well as Ledda’s study, did not find a clear dis-
tinction between the included local and global operators. On the other hand,
Yoshida’s methodology did allow local operators to be distinguished from global
operators. This suggests that perhaps for some combinations of tasks and images,
a computationally expensive local operator would not yield better results than a
cheaper global operator. However, this does not mean that there does not exist a
class of particularly challenging images for which a local operator would bring
significant benefits. However, such scenes would also be challenging to include
in validation studies.

17.12 Further Reading

High dynamic range imaging is a relatively new field, with so far only two books
available on the topic. Reinhard et al.’s book presents the high dynamic range

i i

i i
 i i

i i

17.12. Further Reading 927

imaging pipeline, with an emphasis on capture, storage, and display algorithms. It

also includes a chapter on image-based lighting [952]. Hoefflinger’s book focuses
on high dynamic range imaging hardware, including cameras, sensors, lenses, and
display hardware [475], but also includes chapters on tone reproduction and HDR
data compression.
The DVD accompanying this book contains a large collection of high dynamic
range images. Further links to image resources are provided on the DVD as well.
Of particular importance is the HDR Photographic Survey [303], which contains
a large set of calibrated images, each annotated with appearance and GPS data
(Available at http://www.cis.rit.edu/fairchild/HDR.html).

i i

i i
 i i

i i

Part IV
Appendices

i i

i i
 i i

i i

Appendix A
Vectors and Matrices

While we assume basic vector algebra, Maxwell’s equations in differential form,

as presented in Chapter 2, are defined on vector fields. In this appendix, we
give a brief account of the basic gradient, divergence, and curl operators, as well
as vector integration and differentiation. An introduction of vector calculus in
the context of Maxwell’s equations can be found in Div, Grad, Curl, and All
That [1004].
All operations assume a Cartesian coordinate system. In the following, we
indicate unit vectors along each of the x-, y- and z-axes with ex , ey and ez .

A.1 Cross and Dot Product

The vector dot product between two vectors A and B is defined as follows:

A · B = Ax Bx + Ay By + Az Bz . (A.1)

The dot product is therefore a scalar quantity. The geometric interpretation is that
the dot product represents the cosine of the angle between two vectors times the
length of both vectors (see Figure A.1):

A · B = A B cos(α ). (A.2)

This implies that the dot product will be zero for perpendicular vectors. Vectors
pointing in the same direction have a dot product with a value equal to the square
of their length:
A · A = A 2 . (A.3)

931

i i

i i
 i i

i i

932 A. Vectors and Matrices

B
α

A. B = ||A|| ||B|| cos (α)

Figure A.1. Dot product between vectors A and B.

Another implication is that to find the angle between two vectors, we may nor-
malize both vectors and take the dot product:
A B
cos(α ) = · . (A.4)
A B

The cross product is defined as

⎛ ⎞
Az By − Ay Bz
A × B = ⎝ Ax Bz − Az Bx ⎠ (A.5)
Ay Bx − Ax By

and results in a new vector which is perpendicular to both A and B, which may be
expressed as

A · (A × B) = 0, (A.6a)
B · (A × B) = 0. (A.6b)

As shown in Figure A.2, the right-hand rule applies so that A × B = −B × A.

Finally, the length of the resulting vector may be interpreted as the sine of the

A
A‚B
B
α

||A ‚ B|| = ||A|| ||B|| sin (α)

B‚A

Figure A.2. Cross product between vectors A and B.

i i

i i
 i i

i i

A.2. Vector Differentiation 933

angle between both vectors times the length of the vectors:

A × B = A B sin(α ). (A.7)

While this identity could be used to compute the angle between two normalized
vectors, it is cheaper to use the dot product as indicated above in (A.4).

A.2 Vector Differentiation

A vector A may be defined as a function of an independent variable t. By making a
change in the value of t, the magnitude and direction of A will generally change.
By increasing or decreasing the value of t, the vector A will draw a contour in
three-dimensional space.1
The differential change dt of A with respect to variable t is then defined as

dA A(t + Δt) − A(t)

= lim . (A.8)
dt Δt→0 Δt
dA
The resulting vector dt is tangential to the contour drawn by A(t).

A.3 Gradient of a Scalar Function

A scalar function may be defined in three-dimensional space. This function will
return a scalar value for each point in space. Let T be such a scalar function.
We are interested in the value of this function at two nearby points (x, y, z) and
(x + Δx, y + Δy, z + Δz). Assuming that T is continuous and that the increments
Δx, Δy, and Δz are small, the values T1 at position P1 = (x, y, z) and T2 at position
P2 = (x + Δx, y + Δy, z + Δz) are related by a Taylor series expansion:
" " "
∂ T "" ∂ T "" ∂ T ""
T2 = T1 + Δ x + Δ y + Δ z + higher-order terms. (A.9)
∂ x "P1 ∂ y "P1 ∂ z "P1

As the two points are close to one another, the higher-order terms can be ne-
glected, so that the difference ΔT between T1 and T2 is given by
" " "
∂ T "" ∂ T "" ∂ T ""
ΔT = T2 − T1 = Δ x + Δ y + Δ z. (A.10)
∂ x "P1 ∂ y "P1 ∂ z "P1
1 We assume throughout that we are dealing with three dimensions.

i i

i i
 i i

i i

934 A. Vectors and Matrices

In the limit that the distance between the two points goes to zero, this becomes
∂T ∂T ∂T
lim ΔT = dT = dx + dy + dz. (A.11)
Δx→0 ∂x ∂y ∂z
Δy→0
Δz→0

We may write this in vector notation as a dot product:

 
∂T ∂T ∂T
dT = ex + ey + ez · (dx ex + dy ey + dz ez ) . (A.12)
∂x ∂y ∂z
With the position vector given as r, the differential change in position can be
written as
dr = dx ex + dy ey + dz ez . (A.13)
At the same time, the first term of the dot product in (A.12) represents the gradient
of T . The gradient operator (or vector differential operator) ∇ may then be defined
as
∂ ∂ ∂
∇ = grad = ex + ey + ez . (A.14)
∂x ∂y ∂z
With this definition, (A.12) becomes simply dT = ∇ T · dr. The interpretation
given to the gradient operator is that it indicates rate of change. The gradient
operator takes a scalar function and returns a vector.

A.4 Divergence
Considering a small volume Δv of space bounded by a surface Δs, the influence
of a vector field F may cause a flux, i.e., a measurable quantity flows into the
volume (in-flux) or out of the volume (out-flux). If the volume goes to zero,
the differential out-flux is represented by the divergence of a vector field. The
divergence ∇· of a vector F is then defined as
∂ Fx ∂ Fy ∂ Fz
∇·F = div F = + + (A.15)
∂x ∂y ∂z
provided that F and its first partial derivatives are continuous within Δv. The
divergence operator applied to a vector, returns a vector.

A.5 Gauss’ Theorem

Consider a vector field acting upon a volume v bounded by a closed surface s. At
each point on the surface, the vector field may be decomposed into a component

i i

i i
 i i

i i

A.6. Curl 935

dV2
dV1

dV4 V
dV3

Figure A.3. A volume V is split into four subvolumes dV1 through dV4 . The flux emanating
from dV1 flows partially into neighboring subvolumes and partially out of volume V . Only
subvolumes at the boundary contribute to flux emanating from the volume.

normal to the surface and a component tangential to the surface. The surface
integral of the component normal to the surface, taken over the surface s, may be
related to the volume integral of the divergence of the vector field, taken over the
volume v:  
F · n ds = ∇·F dv. (A.16)
s v

This relation is known as the divergence theorem or Gauss’ theorem. Its use lies
in the fact that the triple integral over the volume v may be replaced by a double
integral over the surface s that bounds the volume.
If a volume V is broken up into a set of small subvolumes, then the flux ema-
nating from a given subvolume will flow into its neighbors, as illustrated in Fig-
ure A.3. Thus for subvolumes located in the interior of V , the net flux is zero.
Only subvolumes located on the boundary of V will contribute to the flux associ-
ated with the volume. This contribution is related to the normal component of the
vector field at the surface. Hence, Gauss’ theorem, as stated above.

A.6 Curl
Suppose water flows through a river. The drag of the banks of the river cause the
water to flow slower near the banks than in the middle of the river. If an object is
released in the river, the water will flow faster along one side of the object than
the opposite side, causing the object to rotate; the water flow on either side of
the object is non-uniform. The term curl is a measure of non-uniformity, i.e., the
ability to cause rotation.

i i

i i
 i i

i i

936 A. Vectors and Matrices

Modeling the flow in the river with a vector field, there will be variation in the
strength of the field dependent on location. At each point in a vector field F, the
curl measures its non-uniformity. Curl is defined as
     
∂ Fz ∂ Fy ∂ Fx ∂ Fz ∂ Fy ∂ Fx
∇× F = curl F = − ex + − ey + − ez . (A.17)
∂y ∂z ∂z ∂x ∂x ∂y
The curl operator takes a vector and produces a new vector.

A.7 Stokes’ Theorem

The divergence theorem outlined above relates a volume integral to a surface in-
tegral over a closed surface. A similar theorem, called Stokes’ Theorem, links the
line integral over a closed contour c to the surface integral over an open surface
s enclosed by a contour c. Examples of contours and open surfaces are shown in
Figure A.4.
capping surfaces
contours c

Figure A.4. Examples of open (or capped) surfaces with associated contours.

When an open surface is subdivided into small differential surface areas, then
the integral of a vector quantity normal to each surface area may be approximated
with a line integral over the contour. The contribution of each contour cancels
out for adjacent surface areas, as shown in Figure A.5. Thus, only the contour
segments that coincide with the contour of the entire surface do not cancel out.
This leads to Stokes’ Theorem:
 
∇× F · n ds = F · n dc. (A.18)
s c

Adjacent contour segments cancel

c1 c2
ds 1 ds 2
s

Figure A.5. Neighboring differential surface areas with associated contours. Integrals over
contours c1 and c2 cancel where they are adjacent and are oriented in opposite directions.

i i

i i
 i i

i i

A.8. Laplacian 937

A.8 Laplacian
The Laplacian operator on a scalar function T is given by

∂ 2T ∂ 2T ∂ 2T
∇2 T = + 2 + 2. (A.19)
∂ x2 ∂y ∂z
For a vector-valued function F, this operator is defined by applying the above
operator on each of the three x, y, and z components independently, yielding the
Laplacian of vector F.

∂ 2 Fx ∂ 2 Fx ∂ 2 Fx
∇2 Fx = + + , (A.20a)
∂ x2 ∂ y2 ∂ z2
∂ 2 Fy ∂ 2 Fy ∂ 2 Fy
∇2 Fy = + + , (A.20b)
∂ x2 ∂ y2 ∂ z2
∂ 2 Fz ∂ 2 Fz ∂ 2 Fz
∇ 2 Fz = + + . (A.20c)
∂ x2 ∂ y2 ∂ z2
This Laplacian operator thus takes a vector and produces a new vector.

A.9 Vector Identities

For convenience, we present several identities that hold for a given scalar t and
vector F:

∇·(tF) = t∇·F + F · ∇t, (A.21)

∇×(tF) = t ∇× F + (∇t) × F, (A.22)
∇× ∇× F = ∇(∇·F) − ∇ F. 2
(A.23)

For clarity, we repeat these three equations in alternate notation:

div tF = t div F + F · grad t, (A.24)

curl tF = t curl F + (grad t) × F, (A.25)
curl curl F = grad div F − ∇ F. 2
(A.26)

A.10 Homogeneous Coordinates

Affine transformations of vectors can be seen as a linear transform followed by a
translation:
x → Ax + b. (A.27)

i i

i i
 i i

i i

938 A. Vectors and Matrices

Such transformations can be conveniently represented by a single matrix if the

vectors and matrices are augmented by an extra dimension. Thus, a three-dimensional
vector x = (x y z)T can be represented with x = (x y z 0)T . An affine transforma-
tion can then be represented by

x → Mx , (A.28)

where M is a 4 × 4 matrix. The elements of the four-vector x are then called

homogeneous coordinates.
It should be noted that in homogeneous coordinates, vectors, points, and sur-
face normals are each treated differently. For instance, under translation, a point
will be mapped to a different point in space, whereas a vector (indicating a direc-
tion) will continue to point in the same direction. Hence, a vector is represented
in homogeneous coordinates with

v = (x y z 0)T , (A.29)

whereas a point in space is represented in homogeneous coordinates with

p = (x y z 1)T . (A.30)

If an object were to undergo an affine transformation and the surface normals

were represented as vectors, after the transformation the surface normals would
no longer point away from the surface as they should. To apply an affine transfor-
mation to a surface normal using homogeneous coordinates, the surface normal is
represented as a vector, i.e., n = (x, y, z, 0)T , but instead of using the transforma-
tion matrix M, the transpose of its inverse should be used [1154]:

n → M−1 T n. (A.31)

To convert a vector from its homogeneous representation back to a conven-

tional (three-dimensional) vector, each component is divided by its fourth com-
ponent. A vector v = (x y z w )T is thus converted to v = (x y z)T using
⎡ ⎤ ⎡  ⎤
x x /w
⎣y⎦ = ⎣y /w ⎦ . (A.32)
z z /w

This procedure is known as the homogeneous divide.

i i

i i
 i i

i i

Appendix B
Trigonometry

In this appendix, we show some of the trigonometric relations that are used else-
where in this book. While we assume basic familiarity with trigonometry, we
review some of the lesser-known relations.

B.1 Sum and Difference Formulae

Starting with two unit vectors Pu and Pv , the dot product between these two vec-
tors will equal the cosine of the angle between them. If these vectors make an
angle of u and v with the positive x-axis, then we have
Pu · Pv = cos(u − v). (B.1)
However, by the definition of the dot product, we also have
Pu · Pv = cos(u) cos(v) + sin(u) sin(v). (B.2)

y Pu . Pv = cos (u-v)
= cos(u) cos(v) + sin(u) sin(v)
1 Pu
u-v
sin(u) Pv
u
v 1
sin(v) x
cos(u)
cos(v)

Figure B.1. The cosine of the difference between two angles may be found by computing
the dot product between two unit vectors.

939

i i

i i
 i i

i i

940 B. Trigonometry

By equating these two expressions, we find the first difference formula (as shown
in Figure B.1):
cos(u − v) = cos(u) cos(v) + sin(u) sin(v). (B.3)
A similar expression may be derived for the sine of the difference of two
angles. Here, we use the definition of the cross product and define two three-
dimensional vectors as
⎛ ⎞ ⎛ ⎞
cos(u) cos(v)
Pu = ⎝ sin(u) ⎠ , Pv = ⎝ sin(v) ⎠ (B.4)
0 0.
The length of Pu × Pv is then equal to sin(u − v) (see (A.7)). At the same time,
this vector is given by
⎛ ⎞
0
Pu × Pv = ⎝ 0 ⎠. (B.5)
sin(u) cos(v) − cos(u) sin(v)
The length of this vector is equal to its z-component. We thus have the following
identity:
sin(u − v) = sin(u) cos(v) − cos(u) sin(v). (B.6)
The above two difference formulas may be extended for the cosine and sine of the
sum of two angles. Here, we use the identities cos(−v) = cos(v) and sin(−v) =
−sin(v):
cos(u + v) = cos(u) cos(v) − sin(u) sin(v), (B.7a)
sin(u + v) = sin(u) cos(v) + cos(u) sin(v). (B.7b)

B.2 Product Identities

From (B.3) and (B.7a), we can derive identities for the product of two cosines by
adding these two equations:
cos(u + v) + cos(u − v)
cos(u) cos(v) = . (B.8a)
2
Similarly we have
sin(u + v) + sin(u − v)
sin(u) cos(v) = , (B.8b)
2
cos(u − v) + cos(u − v)
sin(u) sin(v) = . (B.8c)
2

i i

i i
 i i

i i

B.3. Double-Angle Formulae 941

B.3 Double-Angle Formulae

The last two equations may be used to derive double-angle formulae for cos(2u)
and sin(2u) by noting that 2u = u + u:

cos(2u) = cos(u) cos(u) − sin(u) sin(u) (B.9a)

= cos (u) − sin (u)
2 2
(B.9b)
= 2 cos (u) − 1
2
(B.9c)
= 1 − 2 sin (u).
2
(B.9d)
sin(2u) = sin(u) cos(u) + cos(u) sin(u) (B.9e)
= 2 sin(u)cos(u). (B.9f)

B.4 Half-Angle Formulae

Two derive formulae for sin(u/2) and cos(u/2), we substitute u/2 in (B.9), and
solve for cos(u/2) and sin(u/2), respectively:

1 + cos(u)
cos(u/2) = , (B.10a)
2

1 − cos(u)
sin(u/2) = . (B.10b)
2

B.5 Sum Identities

The derivation of the sum of two sines is more involved. We begin by using the
identity:
sin2 (u) + cos2 (u) = 1 (B.11)
leading to:
 
sin(u) + sin(v) = 1 − cos2 (u) + 1 − cos2 (v). (B.12a)

Rewriting yields
   
sin(u) + sin(v) = 1 − cos(u) 1 + cos(u) + 1 − cos(v) 1 + cos(v).
(B.12b)

i i

i i
 i i

i i

942 B. Trigonometry

We then introduce the constant 2:

    
1 − cos(u) 1 + cos(u) 1 − cos(v) 1 + cos(v)
sin(u) + sin(v) = 2 +
2 2 2 2
(B.12c)

and then apply the half-angle rules:

sin(u) + sin(v) = 2 sin(u/2) cos(u/2) + 2 sin(v/2) cos(v/2). (B.12d)

This expression can be expanded using sin2 (u/2) + cos2 (u/2) = 1 once more:
 
sin(u) + sin(v) = 2 sin(u/2) cos(u/2) sin2 (u/2) + cos2 (u/2) (B.12e)
 
+ 2 sin(v/2) cos(v/2) sin2 (v/2) + cos2 (v/2) (B.12f)

By rearranging terms we find

sin(u) + sin(v) = 2 (sin(u/2) cos(v/2) + cos(u/2) sin(v/2)) (B.12g)

× (cos(u/2) cos(v/2) + sin(u/2) sin(u/2)) . (B.12h)

We can then apply sum and difference formulae to find the desired expression:
   
u+v u−v
sin(u) + sin(v) = 2 sin cos . (B.13a)
2 2

Similarly, the following relations may be derived:

   
u+v u−v
sin(u) − sin(v) = 2 cos sin , (B.13b)
2 2
   
u+v u−v
cos(u) + cos(v) = 2 cos cos , (B.13c)
2 2
   
u+v u−v
cos(u) − cos(v) = −2 sin sin . (B.13d)
2 2

B.6 Solid Angle

Solid angles can be seen as the 3D equivalents of 2D angles. The solid angle
spanned by a surface around a point is equal to the area of the surface projected
on a unit sphere surrounding the point as shown in Figure B.2. Solid angles are
measured in steradians (sr), and their largest value can be 4π which is the surface
area of the full sphere.

i i

i i
 i i

i i

B.6. Solid Angle 943

Surface area: A

A
Steradian: ω =
r2

Figure B.2. The definition of solid angle ω is a conical segment of a sphere. One steradian
is the angle subtended at the center of this sphere by an area of surface equal to the square
of the radius.

Ω
Surface area
Solid angle
Projected solid angle

(p)
Ω

Figure B.3. The solid angle spanned by the surface A is equal to Ω. The projected solid
angle is shown by Ω(p) .

dθ

sin (θ) dφ

θ
dΩ

dΩ = sin (θ) dφ dθ
θ: elevational angle
φ: azimuthal angle

Figure B.4. Two differential solid angles are shown on a unit sphere. Note that the size
of the differential solid angle changes based on its vertical position on the sphere. The
relationship is given by dΩ = sin(θ )d φ d θ .

i i

i i
 i i

i i

944 B. Trigonometry

On the other hand, the projected solid angle is the area of the solid angle
projected onto the base of a unit sphere (Figure B.3). The largest projected solid
angle can be π which is equal to the full area of the base.
The differential solid angle is the smallest area element on a sphere. Its value
depends on the elevation angle at which it is computed (see Figure B.4). The sum
of all differential solid angles is equal to the surface area of the sphere:
 2π  π
sin(θ ) d φ d θ = 4π . (B.14)
0 0

i i

i i
 i i

i i

Appendix C
Complex Numbers

In this appendix, we review some basic formulae and theorems relating to com-
plex numbers.

C.1 Definition
A complex number z is defined as

z = x + iy, (C.1)
√
where x and y are real numbers, and i is the imaginary number (i = −1). Com-
plex numbers may be represented by points on a complex plane, as shown in
Figure C.1. Here, the point z may either be represented in Cartesian coordinates
(x, y) or in polar coordinates (r, ϕ ). To convert a complex number from Cartesian

Imaginary z =x+iy
= r cos (ϕ) + i r sin (ϕ)
x = r cos (ϕ) = r exp (i ϕ)
z

r
y = r sin (ϕ)

Real

Figure C.1. A complex number P may be represented as a point in the complex plane.

945

i i

i i
 i i

i i

946 C. Complex Numbers

coordinates to polar coordinates:


r= x 2 + y2 , (C.2a)
−1
ϕ = tan (y/x) . (C.2b)

For a complex number z, the real part is indicated with x = Re{z} and the
imaginary part is given by y = Im{z}. From (C.1), it can be seen that the operators
Re{z} and Im{z} are defined as follows:

z + z∗
Re{z} = , (C.3)
2
z−z ∗
Im{z} = , (C.4)
2i

where z∗ indicates the complex conjugate of z, i.e., z = x − iy = Re{z} − iIm{z}.

The magnitude |z| of a complex number is given by

√ 
|z| = z z∗ = Re{z}2 + Im{z}2 (C.5)

An frequently occurring quantity is the square of the magnitude of a complex

number, which can be computed by omitting the square root in the above equation:

|z|2 = z z∗ = Re{z}2 + Im{z}2 (C.6)

C.2 Euler’s Formula

If a complex number is represented in polar coordinates, we may apply Euler’s
formula and rewrite the equation as a complex exponential function:

z = x + iy (C.7a)
= r cos(ϕ ) + ir sin(ϕ ) (C.7b)
iϕ
= re . (C.7c)

A proof of this result may be obtained by expanding each of the exp, cos and
sin functions into a Taylor series and substituting back into Equations (C.7b)
and (C.7c).

i i

i i
 i i

i i

C.3. Theorems 947

Imaginary
(1,0) 1
i
(0,i) i
(-1,0) -1
-1 (0,-i) -i
1 Real

-i Unit circle

Figure C.2. Values of eiϕ may be deduced for specific values of ϕ by observing the unit
circle in the complex plane.

C.3 Theorems
By drawing a unit circle around the origin in the complex plane, we obtain several
useful expressions for complex exponential functions (see Figure C.2):

e0 = 1, (C.8a)
eiπ /2 = i, (C.8b)
iπ
e = −1, (C.8c)
3iπ /2
e = −i. (C.8d)

Further identities are as follows:

e−iϕ = cos(ϕ ) − i sin(ϕ ) = z∗ (C.9a)

1  iϕ 
cos(ϕ ) = e + e−iϕ , (C.9b)
2
1  iϕ 
sin(ϕ ) = e − e−iϕ , (C.9c)
2i
r1 eiϕ1 r2 eiϕ2 = r1 r2 ei(ϕ1 +ϕ2 ) , (C.9d)
r1eiϕ1
= r1 r2 ei(ϕ1 −ϕ2 ) . (C.9e)
r2 eiϕ2

Finally, an important property of the exponential function e is that for all

values of ϕ , the function eiϕ is a measure of the rate of change of eiϕ , i.e., its
derivative:
d iϕ
e = iϕ eiϕ . (C.10)
dϕ

i i

i i
 i i

i i

948 C. Complex Numbers

C.4 Time-Harmonic Quantities

One of the solutions to the wave equations discussed in Chapter 2 is in the form of
sinusoidal waveforms. While it is possible to compute solutions in terms of sines
and cosines, it is often more convenient to use complex notation. A sinusoidal
waveform f (t) may be defined as

f (t) = a cos(ω t + ϕ ). (C.11)

Here, a is the amplitude and ω t + ϕ is the phase. In most cases, we view t as

the time variable and ϕ is an arbitrary phase constant. We may complement this
function by its complex counterpart to find the complex function f (t):

f (t) = a (cos(ω t + ϕ ) + i sin(ω t + ϕ )) . (C.12)

By applying Euler’s formula we get

f (t) = a ei (ω t+ϕ ) (C.13)

i ϕ i ωt
= ae e (C.14)
i ωt
= Ae . (C.15)

Thus, the complex function f (t) may be written as the product of the complex
phasor A = a ei ϕ and the time factor ei ω t . This is an important result that al-
lows the time dependency to be eliminated from partial differential equations. It
thereby reduces partial differential equations to ordinary differential equations.
These results are used in finding solutions to the wave equations for plane waves
in time-harmonic fields in Section 2.2.3.
To reconstitute a solution, it suffices to take the real part:

f (t) = Re{ f (t)}. (C.16)

i i

i i
 i i

i i

Appendix D
Units and Constants
In this appendix, we enumerate the constants, symbols and units used in this book.
Also given are conversions between commonly encountered units, insofar as they
deviate from SI units. The basic quantities are given in Table D.1. Some relevant
units outside the SI system are given in Table D.2. The 22 derived SI units are
enumerated in Table D.3. Contants (in SI units) are given in Table D.4. For further
information on SI units, see the NIST website (http://physics.nist.gov/cuu/Units/
index.html.)

Unit Symbol Physical quantity

Meter m Length
Kilogram kg Mass
Second s Time
Ampere A Electric current
Kelvin K Thermodynamic temperature
Mole mol Amount of substance
Candela cd luminous intensity

Table D.1. Fundamental SI units.

Name Symbol SI equivalent

minute (time) min 60 s
hour h 3600 s
day d 86,400 s
degree (angle) ◦ (π /180) rad
minute (angle) ’ (π /10, 800) rad
second (angle) ” (π /648, 000) rad
liter L 10−3 m3
metric ton t 103 kg
neper Np l
bel B (1/2) ln(10N p)
electronvolt eV 1.602 18 ×10−19 J
unified atomic mass unit u 1.660 54 ×10−27 kg
astonomical unit ua 1.495 98 ×1011 m

Table D.2. Units outside the SI system.

949

i i

i i
 i i

i i

950 D. Units and Constants

Unit Symbol Definition Physical quantity

radian rad m m−1 =1 plane angle
steradian sr m2 m−2 = 1 solid angle
hertz Hz s−1 frequency
newton N kg m s−2 force
pascal Pa N m−2 pressure
joule J Nm energy
watt W J s−1 power
coulomb C sA charge
volt V W A−1 potential
farad F C V−1 capacitance
ohm Ω V A−1 electric resistance
siemens S A V−1 electric conductance
weber Wb Vs magnetic flux
tesla T Wb m−1 magnetic flux density
henry H Wb A−1 inductance
degree Celsius ◦C K Celsius temperature
lumen lm cd sr luminous flux
lux lx lm m−2 illuminance
becquerel Bq s−1 activity
gray Gy J kg−1 absorbed dose
sievert Sv J kg−1 dose equivalent
katal kat mol s−1 catalytic activity

Table D.3. Derived SI units used in eletromagnetic theory.

Constant Symbol Value Unit

1
Permittivity of vacuum ε0 × 10−9 F m−1
36 π
Permeability of vacuum μ0 4 π · 10−7 H m−1
Stefan-Boltzmann constant σ 5.670 × 10−8 W m−2 K−4
Speed of light in vacuum c 2.998 × 108 m s−1
Planck’s constant h 6.626 × 10−34 Js
Boltzmann’s constant kB 1.380 × 10−23 J K−1

Table D.4. Constants.

i i

i i
 i i

i i

Appendix E
The CIE Luminous
Efficiency Functions

λ (nm) + 0 nm + 1 nm + 2nm + 3 nm + 4 nm + 5 nm + 6 nm + 7nm + 8 nm + 9 nm

360 3.9e-06 4.4e-06 4.9e-06 5.5e-06 6.2e-06 7.0e-06 7.8e-06 8.8e-06 9.8e-06 1.1e-05
370 1.2e-05 1.4e-05 1.6e-05 1.7e-05 2.0e-05 2.2e-05 2.5e-05 2.8e-05 3.2e-05 3.5e-05
380 3.9e-05 4.3e-05 4.7e-05 5.2e-05 5.7e-05 6.4e-05 7.2e-05 8.2e-05 9.4e-05 1.1e-04
390 1.2e-04 1.3e-04 1.5e-04 1.7e-04 1.9e-04 2.2e-04 2.5e-04 2.8e-04 3.2e-04 3.6e-04
400 4.0e-04 4.3e-04 4.7e-04 5.2e-04 5.7e-04 6.4e-04 7.2e-04 8.3e-04 9.4e-04 1.1e-03
410 1.2e-03 1.4e-03 1.5e-03 1.7e-03 1.9e-03 2.2e-03 2.5e-03 2.8e-03 3.1e-03 3.5e-03
420 4.0e-03 4.5e-03 5.2e-03 5.8e-03 6.5e-03 7.3e-03 8.1e-03 8.9e-03 9.8e-03 1.1e-02
430 1.2e-02 1.3e-02 1.4e-02 1.5e-02 1.6e-02 1.7e-02 1.8e-02 1.9e-02 2.0e-02 2.2e-02
440 2.3e-02 2.4e-02 2.6e-02 2.7e-02 2.8e-02 3.0e-02 3.1e-02 3.3e-02 3.5e-02 3.6e-02
450 3.8e-02 4.0e-02 4.2e-02 4.4e-02 4.6e-02 4.8e-02 5.0e-02 5.3e-02 5.5e-02 5.7e-02
460 6.0e-02 6.3e-02 6.5e-02 6.8e-02 7.1e-02 7.4e-02 7.7e-02 8.0e-02 8.4e-02 8.7e-02
470 9.1e-02 9.5e-02 9.9e-02 1.0e-01 1.1e-01 1.1e-01 1.2e-01 1.2e-01 1.3e-01 1.3e-01
480 1.4e-01 1.4e-01 1.5e-01 1.6e-01 1.6e-01 1.7e-01 1.8e-01 1.8e-01 1.9e-01 2.0e-01
490 2.1e-01 2.2e-01 2.3e-01 2.4e-01 2.5e-01 2.6e-01 2.7e-01 2.8e-01 3.0e-01 3.1e-01
500 3.2e-01 3.4e-01 3.5e-01 3.7e-01 3.9e-01 4.1e-01 4.3e-01 4.4e-01 4.6e-01 4.8e-01
510 5.0e-01 5.2e-01 5.4e-01 5.7e-01 5.9e-01 6.1e-01 6.3e-01 6.5e-01 6.7e-01 6.9e-01
520 7.1e-01 7.3e-01 7.5e-01 7.6e-01 7.8e-01 7.9e-01 8.1e-01 8.2e-01 8.4e-01 8.5e-01
530 8.6e-01 8.7e-01 8.8e-01 9.0e-01 9.1e-01 9.1e-01 9.2e-01 9.3e-01 9.4e-01 9.5e-01
540 9.5e-01 9.6e-01 9.7e-01 9.7e-01 9.8e-01 9.8e-01 9.8e-01 9.9e-01 9.9e-01 9.9e-01
550 9.9e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00
560 9.9e-01 9.9e-01 9.9e-01 9.9e-01 9.8e-01 9.8e-01 9.7e-01 9.7e-01 9.6e-01 9.6e-01
570 9.5e-01 9.5e-01 9.4e-01 9.3e-01 9.2e-01 9.2e-01 9.1e-01 9.0e-01 8.9e-01 8.8e-01
580 8.7e-01 8.6e-01 8.5e-01 8.4e-01 8.3e-01 8.2e-01 8.0e-01 7.9e-01 7.8e-01 7.7e-01
590 7.6e-01 7.4e-01 7.3e-01 7.2e-01 7.1e-01 6.9e-01 6.8e-01 6.7e-01 6.6e-01 6.4e-01
600 6.3e-01 6.2e-01 6.1e-01 5.9e-01 5.8e-01 5.7e-01 5.5e-01 5.4e-01 5.3e-01 5.2e-01
610 5.0e-01 4.9e-01 4.8e-01 4.7e-01 4.5e-01 4.4e-01 4.3e-01 4.2e-01 4.1e-01 3.9e-01

Table E.1. The CIE 1924 Photopic luminous efficiency function, V (λ ).

951

i i

i i
 i i

i i

952 E. The CIE Luminous Efficiency Functions

λ (nm) + 0 nm + 1 nm + 2nm + 3 nm + 4 nm + 5 nm + 6 nm + 7nm + 8 nm + 9 nm

620 3.8e-01 3.7e-01 3.6e-01 3.4e-01 3.3e-01 3.2e-01 3.1e-01 3.0e-01 2.9e-01 2.8e-01
630 2.7e-01 2.5e-01 2.4e-01 2.4e-01 2.3e-01 2.2e-01 2.1e-01 2.0e-01 1.9e-01 1.8e-01
640 1.7e-01 1.7e-01 1.6e-01 1.5e-01 1.5e-01 1.4e-01 1.3e-01 1.3e-01 1.2e-01 1.1e-01
650 1.1e-01 1.0e-01 9.6e-02 9.1e-02 8.6e-02 8.2e-02 7.7e-02 7.3e-02 6.9e-02 6.5e-02
660 6.1e-02 5.7e-02 5.4e-02 5.1e-02 4.8e-02 4.5e-02 4.2e-02 3.9e-02 3.7e-02 3.4e-02
670 3.2e-02 3.0e-02 2.8e-02 2.6e-02 2.5e-02 2.3e-02 2.2e-02 2.1e-02 1.9e-02 1.8e-02
680 1.7e-02 1.6e-02 1.5e-02 1.4e-02 1.3e-02 1.2e-02 1.1e-02 1.0e-02 9.5e-03 8.8e-03
690 8.2e-03 7.6e-03 7.1e-03 6.6e-03 6.1e-03 5.7e-03 5.3e-03 5.0e-03 4.7e-03 4.4e-03
700 4.1e-03 3.8e-03 3.6e-03 3.4e-03 3.1e-03 2.9e-03 2.7e-03 2.6e-03 2.4e-03 2.2e-03
710 2.1e-03 2.0e-03 1.8e-03 1.7e-03 1.6e-03 1.5e-03 1.4e-03 1.3e-03 1.2e-03 1.1e-03
720 1.0e-03 9.8e-04 9.1e-04 8.5e-04 7.9e-04 7.4e-04 6.9e-04 6.4e-04 6.0e-04 5.6e-04
730 5.2e-04 4.8e-04 4.5e-04 4.2e-04 3.9e-04 3.6e-04 3.4e-04 3.1e-04 2.9e-04 2.7e-04
740 2.5e-04 2.3e-04 2.1e-04 2.0e-04 1.9e-04 1.7e-04 1.6e-04 1.5e-04 1.4e-04 1.3e-04
750 1.2e-04 1.1e-04 1.0e-04 9.7e-05 9.1e-05 8.5e-05 7.9e-05 7.4e-05 6.9e-05 6.4e-05
760 6.0e-05 5.6e-05 5.2e-05 4.9e-05 4.5e-05 4.2e-05 4.0e-05 3.7e-05 3.4e-05 3.2e-05
770 3.0e-05 2.8e-05 2.6e-05 2.4e-05 2.3e-05 2.1e-05 2.0e-05 1.8e-05 1.7e-05 1.6e-05
780 1.5e-05 1.4e-05 1.3e-05 1.2e-05 1.1e-05 1.1e-05 9.9e-06 9.2e-06 8.6e-06 8.0e-06
790 7.5e-06 7.0e-06 6.5e-06 6.0e-06 5.6e-06 5.3e-06 4.9e-06 4.6e-06 4.3e-06 4.0e-06
800 3.7e-06 3.5e-06 3.2e-06 3.0e-06 2.8e-06 2.6e-06 2.4e-06 2.3e-06 2.1e-06 2.0e-06
810 1.8e-06 1.7e-06 1.6e-06 1.5e-06 1.4e-06 1.3e-06 1.2e-06 1.1e-06 1.0e-06 9.8e-07
820 9.1e-07 8.5e-07 7.9e-07 7.4e-07 6.9e-07 6.4e-07 6.0e-07 5.6e-07 5.2e-07 4.8e-07
830 4.5e-07

Table E.1. (continued) The CIE 1924 Photopic luminous efficiency function, V (λ ).

λ (nm) + 0 nm + 1 nm + 2nm + 3 nm + 4 nm + 5 nm + 6 nm + 7nm + 8 nm + 9 nm

380 5.9e-04 6.7e-04 7.5e-04 8.5e-04 9.7e-04 1.1e-03 1.3e-03 1.5e-03 1.7e-03 1.9e-03
390 2.2e-03 2.5e-03 2.9e-03 3.4e-03 3.9e-03 4.5e-03 5.2e-03 6.0e-03 7.0e-03 8.1e-03
400 9.3e-03 1.1e-02 1.2e-02 1.4e-02 1.6e-02 1.9e-02 2.1e-02 2.4e-02 2.7e-02 3.1e-02
410 3.5e-02 3.9e-02 4.4e-02 4.9e-02 5.4e-02 6.0e-02 6.7e-02 7.4e-02 8.1e-02 8.8e-02
420 9.7e-02 1.1e-01 1.1e-01 1.2e-01 1.3e-01 1.4e-01 1.5e-01 1.7e-01 1.8e-01 1.9e-01
430 2.0e-01 2.1e-01 2.2e-01 2.4e-01 2.5e-01 2.6e-01 2.8e-01 2.9e-01 3.0e-01 3.1e-01
440 3.3e-01 3.4e-01 3.5e-01 3.7e-01 3.8e-01 3.9e-01 4.1e-01 4.2e-01 4.3e-01 4.4e-01
450 4.6e-01 4.7e-01 4.8e-01 4.9e-01 5.0e-01 5.1e-01 5.2e-01 5.4e-01 5.5e-01 5.6e-01
460 5.7e-01 5.8e-01 5.9e-01 6.0e-01 6.1e-01 6.2e-01 6.3e-01 6.4e-01 6.5e-01 6.6e-01
470 6.8e-01 6.9e-01 7.0e-01 7.1e-01 7.2e-01 7.3e-01 7.4e-01 7.6e-01 7.7e-01 7.8e-01
480 7.9e-01 8.1e-01 8.2e-01 8.3e-01 8.4e-01 8.5e-01 8.6e-01 8.7e-01 8.8e-01 8.9e-01
490 9.0e-01 9.1e-01 9.2e-01 9.3e-01 9.4e-01 9.5e-01 9.6e-01 9.6e-01 9.7e-01 9.8e-01
500 9.8e-01 9.9e-01 9.9e-01 9.9e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00
510 1.0e+00 9.9e-01 9.9e-01 9.9e-01 9.8e-01 9.7e-01 9.7e-01 9.6e-01 9.5e-01 9.4e-01
520 9.4e-01 9.3e-01 9.2e-01 9.0e-01 8.9e-01 8.8e-01 8.7e-01 8.5e-01 8.4e-01 8.3e-01
530 8.1e-01 8.0e-01 7.8e-01 7.7e-01 7.5e-01 7.3e-01 7.2e-01 7.0e-01 6.8e-01 6.7e-01
540 6.5e-01 6.3e-01 6.2e-01 6.0e-01 5.8e-01 5.6e-01 5.5e-01 5.3e-01 5.1e-01 5.0e-01
550 4.8e-01 4.7e-01 4.5e-01 4.3e-01 4.2e-01 4.0e-01 3.9e-01 3.7e-01 3.6e-01 3.4e-01
560 3.3e-01 3.2e-01 3.0e-01 2.9e-01 2.8e-01 2.6e-01 2.5e-01 2.4e-01 2.3e-01 2.2e-01
570 2.1e-01 2.0e-01 1.9e-01 1.8e-01 1.7e-01 1.6e-01 1.5e-01 1.4e-01 1.4e-01 1.3e-01
580 1.2e-01 1.1e-01 1.1e-01 1.0e-01 9.6e-02 9.0e-02 8.5e-02 7.9e-02 7.4e-02 7.0e-02
590 6.6e-02 6.1e-02 5.7e-02 5.4e-02 5.0e-02 4.7e-02 4.4e-02 4.1e-02 3.8e-02 3.6e-02

Table E.2. The CIE 1951 Scotopic luminous efficiency function, V  (λ ).

i i

i i
 i i

i i

953

λ (nm) + 0 nm + 1 nm + 2nm + 3 nm + 4 nm + 5 nm + 6 nm + 7nm + 8 nm + 9 nm

600 3.3e-02 3.1e-02 2.9e-02 2.7e-02 2.5e-02 2.3e-02 2.1e-02 2.0e-02 1.9e-02 1.7e-02
610 1.6e-02 1.5e-02 1.4e-02 1.3e-02 1.2e-02 1.1e-02 1.0e-02 9.3e-03 8.6e-03 8.0e-03
620 7.4e-03 6.8e-03 6.3e-03 5.8e-03 5.4e-03 5.0e-03 4.6e-03 4.2e-03 3.9e-03 3.6e-03
630 3.3e-03 3.1e-03 2.8e-03 2.6e-03 2.4e-03 2.2e-03 2.1e-03 1.9e-03 1.8e-03 1.6e-03
640 1.5e-03 1.4e-03 1.3e-03 1.2e-03 1.1e-03 1.0e-03 9.3e-04 8.6e-04 7.9e-04 7.3e-04
650 6.8e-04 6.3e-04 5.8e-04 5.4e-04 5.0e-04 4.6e-04 4.2e-04 3.9e-04 3.6e-04 3.4e-04
660 3.1e-04 2.9e-04 2.7e-04 2.5e-04 2.3e-04 2.1e-04 2.0e-04 1.8e-04 1.7e-04 1.6e-04
670 1.5e-04 1.4e-04 1.3e-04 1.2e-04 1.1e-04 1.0e-04 9.5e-05 8.9e-05 8.3e-05 7.7e-05
680 7.2e-05 6.7e-05 6.2e-05 5.8e-05 5.4e-05 5.0e-05 4.7e-05 4.4e-05 4.1e-05 3.8e-05
690 3.5e-05 3.3e-05 3.1e-05 2.9e-05 2.7e-05 2.5e-05 2.3e-05 2.2e-05 2.0e-05 1.9e-05
700 1.8e-05 1.7e-05 1.6e-05 1.5e-05 1.4e-05 1.3e-05 1.2e-05 1.1e-05 1.0e-05 9.8e-06
710 9.1e-06 8.6e-06 8.0e-06 7.5e-06 7.0e-06 6.6e-06 6.2e-06 5.8e-06 5.4e-06 5.1e-06
720 4.8e-06 4.5e-06 4.2e-06 4.0e-06 3.7e-06 3.5e-06 3.3e-06 3.1e-06 2.9e-06 2.7e-06
730 2.5e-06 2.4e-06 2.3e-06 2.1e-06 2.0e-06 1.9e-06 1.8e-06 1.7e-06 1.6e-06 1.5e-06
740 1.4e-06 1.3e-06 1.2e-06 1.2e-06 1.1e-06 1.0e-06 9.6e-07 9.1e-07 8.5e-07 8.1e-07
750 7.6e-07 7.2e-07 6.7e-07 6.4e-07 6.0e-07 5.7e-07 5.3e-07 5.1e-07 4.8e-07 4.5e-07
760 4.3e-07 4.0e-07 3.8e-07 3.6e-07 3.4e-07 3.2e-07 3.0e-07 2.9e-07 2.7e-07 2.6e-07
770 2.4e-07 2.3e-07 2.2e-07 2.0e-07 1.9e-07 1.8e-07 1.7e-07 1.6e-07 1.6e-07 1.5e-07
780 1.4e-07

Table E.2. (continued) The CIE 1951 Scotopic luminous efficiency function, V  (λ ).

λ (nm) + 0 nm + 1 nm + 2nm + 3 nm + 4 nm + 5 nm + 6 nm + 7nm + 8 nm + 9 nm

380 2.0e-04 2.3e-04 2.6e-04 3.0e-04 3.4e-04 4.0e-04 4.6e-04 5.2e-04 6.0e-04 7.0e-04
390 8.0e-04 9.2e-04 1.0e-03 1.2e-03 1.4e-03 1.5e-03 1.8e-03 1.9e-03 2.2e-03 2.5e-03
400 2.8e-03 3.1e-03 3.5e-03 3.8e-03 4.2e-03 4.7e-03 5.1e-03 5.6e-03 6.2e-03 6.8e-03
410 7.4e-03 8.1e-03 9.0e-03 9.8e-03 1.1e-02 1.2e-02 1.3e-02 1.4e-02 1.5e-02 1.6e-02
420 1.8e-02 1.9e-02 2.0e-02 2.1e-02 2.2e-02 2.3e-02 2.4e-02 2.5e-02 2.5e-02 2.6e-02
430 2.7e-02 2.8e-02 2.9e-02 3.0e-02 3.2e-02 3.3e-02 3.4e-02 3.5e-02 3.6e-02 3.7e-02
440 3.8e-02 3.9e-02 4.0e-02 4.1e-02 4.2e-02 4.2e-02 4.3e-02 4.4e-02 4.5e-02 4.6e-02
450 4.7e-02 4.8e-02 4.9e-02 5.0e-02 5.1e-02 5.2e-02 5.3e-02 5.5e-02 5.6e-02 5.8e-02
460 6.0e-02 6.3e-02 6.5e-02 6.8e-02 7.1e-02 7.4e-02 7.7e-02 8.0e-02 8.4e-02 8.7e-02
470 9.1e-02 9.5e-02 9.9e-02 1.0e-01 1.1e-01 1.1e-01 1.2e-01 1.2e-01 1.3e-01 1.3e-01
480 1.4e-01 1.4e-01 1.5e-01 1.6e-01 1.6e-01 1.7e-01 1.8e-01 1.8e-01 1.9e-01 2.0e-01
490 2.1e-01 2.2e-01 2.3e-01 2.4e-01 2.5e-01 2.6e-01 2.7e-01 2.8e-01 3.0e-01 3.1e-01
500 3.2e-01 3.4e-01 3.5e-01 3.7e-01 3.9e-01 4.1e-01 4.3e-01 4.4e-01 4.6e-01 4.8e-01
510 5.0e-01 5.2e-01 5.4e-01 5.7e-01 5.9e-01 6.1e-01 6.3e-01 6.5e-01 6.7e-01 6.9e-01
520 7.1e-01 7.3e-01 7.5e-01 7.6e-01 7.8e-01 7.9e-01 8.1e-01 8.2e-01 8.4e-01 8.5e-01
530 8.6e-01 8.7e-01 8.8e-01 9.0e-01 9.1e-01 9.1e-01 9.2e-01 9.3e-01 9.4e-01 9.5e-01
540 9.5e-01 9.6e-01 9.7e-01 9.7e-01 9.8e-01 9.8e-01 9.8e-01 9.9e-01 9.9e-01 9.9e-01
550 9.9e-01 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00 1.0e+00
560 9.9e-01 9.9e-01 9.9e-01 9.9e-01 9.8e-01 9.8e-01 9.7e-01 9.7e-01 9.6e-01 9.6e-01
570 9.5e-01 9.5e-01 9.4e-01 9.3e-01 9.2e-01 9.2e-01 9.1e-01 9.0e-01 8.9e-01 8.8e-01
580 8.7e-01 8.6e-01 8.5e-01 8.4e-01 8.3e-01 8.2e-01 8.0e-01 7.9e-01 7.8e-01 7.7e-01
590 7.6e-01 7.4e-01 7.3e-01 7.2e-01 7.1e-01 6.9e-01 6.8e-01 6.7e-01 6.6e-01 6.4e-01

Table E.3. The CIE 1988 Photopic luminous efficiency function, VM (λ )

i i

i i
 i i

i i

954 E. The CIE Luminous Efficiency Functions

λ (nm) + 0 nm + 1 nm + 2nm + 3 nm + 4 nm + 5 nm + 6 nm + 7nm + 8 nm + 9 nm

600 6.3e-01 6.2e-01 6.1e-01 5.9e-01 5.8e-01 5.7e-01 5.5e-01 5.4e-01 5.3e-01 5.2e-01
610 5.0e-01 4.9e-01 4.8e-01 4.7e-01 4.5e-01 4.4e-01 4.3e-01 4.2e-01 4.1e-01 3.9e-01
620 3.8e-01 3.7e-01 3.6e-01 3.4e-01 3.3e-01 3.2e-01 3.1e-01 3.0e-01 2.9e-01 2.8e-01
630 2.7e-01 2.5e-01 2.4e-01 2.4e-01 2.3e-01 2.2e-01 2.1e-01 2.0e-01 1.9e-01 1.8e-01
640 1.7e-01 1.7e-01 1.6e-01 1.5e-01 1.5e-01 1.4e-01 1.3e-01 1.3e-01 1.2e-01 1.1e-01
650 1.1e-01 1.0e-01 9.6e-02 9.1e-02 8.6e-02 8.2e-02 7.7e-02 7.3e-02 6.9e-02 6.5e-02
660 6.1e-02 5.7e-02 5.4e-02 5.1e-02 4.8e-02 4.5e-02 4.2e-02 3.9e-02 3.7e-02 3.4e-02
670 3.2e-02 3.0e-02 2.8e-02 2.6e-02 2.5e-02 2.3e-02 2.2e-02 2.1e-02 1.9e-02 1.8e-02
680 1.7e-02 1.6e-02 1.5e-02 1.4e-02 1.3e-02 1.2e-02 1.1e-02 1.0e-02 9.5e-03 8.8e-03
690 8.2e-03 7.6e-03 7.1e-03 6.6e-03 6.1e-03 5.7e-03 5.3e-03 5.0e-03 4.7e-03 4.4e-03
700 4.1e-03 3.8e-03 3.6e-03 3.4e-03 3.1e-03 2.9e-03 2.7e-03 2.6e-03 2.4e-03 2.2e-03
710 2.1e-03 2.0e-03 1.8e-03 1.7e-03 1.6e-03 1.5e-03 1.4e-03 1.3e-03 1.2e-03 1.1e-03
720 1.0e-03 9.8e-04 9.1e-04 8.5e-04 7.9e-04 7.4e-04 6.9e-04 6.4e-04 6.0e-04 5.6e-04
730 5.2e-04 4.8e-04 4.5e-04 4.2e-04 3.9e-04 3.6e-04 3.4e-04 3.1e-04 2.9e-04 2.7e-04
740 2.5e-04 2.3e-04 2.1e-04 2.0e-04 1.9e-04 1.7e-04 1.6e-04 1.5e-04 1.4e-04 1.3e-04
750 1.2e-04 1.1e-04 1.0e-04 9.7e-05 9.1e-05 8.5e-05 7.9e-05 7.4e-05 6.9e-05 6.4e-05
760 6.0e-05 5.6e-05 5.2e-05 4.9e-05 4.5e-05 4.2e-05 4.0e-05 3.7e-05 3.4e-05 3.2e-05
770 3.0e-05 2.8e-05 2.6e-05 2.4e-05 2.3e-05 2.1e-05 2.0e-05 1.8e-05 1.7e-05 1.6e-05
780 1.5e-05

Table E.3. (continued) The CIE 1988 Photopic luminous efficiency function, VM (λ ).

i i

i i
 i i

i i

Appendix F
CIE Illuminants

λ (nm) S0 (λ ) S1 (λ ) S2 (λ ) D50 D55 D65 D75

300 0.04 0.02 0.00 0.02 0.02 0.03 0.04
305 3.02 2.26 1.00 1.04 1.05 1.66 2.59
310 6.00 4.50 2.00 2.05 2.07 3.29 5.13
315 17.80 13.45 3.00 4.92 6.65 11.77 17.47
320 29.60 22.40 4.00 7.78 11.23 20.24 29.82
325 42.45 32.20 6.25 11.27 15.94 28.64 42.38
330 55.30 42.00 8.50 14.76 20.66 37.05 54.94
335 56.30 41.30 8.15 16.36 22.27 38.50 56.11
340 57.30 40.60 7.80 17.96 23.89 39.95 57.27
345 59.55 41.10 7.25 19.49 25.86 42.43 60.01
350 61.80 41.60 6.70 21.02 27.83 44.91 62.75
355 61.65 39.80 6.00 22.49 29.23 45.77 62.87
360 61.50 38.00 5.30 23.95 30.63 46.64 62.99
365 65.15 40.20 5.70 25.46 32.47 49.36 66.66
370 68.80 42.40 6.10 26.97 34.32 52.09 70.33
375 66.10 40.45 4.55 25.74 33.45 51.03 68.52
380 63.40 38.50 3.00 24.50 32.59 49.98 66.71
385 64.60 36.75 2.10 27.19 35.34 52.31 68.34
390 65.80 35.00 1.20 29.88 38.09 54.65 69.97
395 80.30 39.20 0.05 39.60 49.53 68.70 85.96
400 94.80 43.40 -1.10 49.32 60.96 82.75 101.94
405 99.80 44.85 -0.80 52.93 64.76 87.12 106.92
410 104.80 46.30 -0.50 56.53 68.56 91.49 111.91
415 105.35 45.10 -0.60 58.29 70.07 92.46 112.36
420 105.90 43.90 -0.70 60.05 71.59 93.43 112.81
425 101.35 40.50 -0.95 58.94 69.75 90.06 107.96
430 96.80 37.10 -1.20 57.83 67.92 86.68 103.10
435 105.35 36.90 -1.90 66.33 76.77 95.77 112.15
440 113.90 36.70 -2.60 74.84 85.61 104.86 121.21

Table F.1. Components S0 (λ ), S1 (λ ), and S2 (λ ) used to compute CIE illuminants D.

955

i i

i i
 i i

i i

956 F. CIE Illuminants

λ (nm) S0 (λ ) S1 (λ ) S2 (λ ) D50 D55 D65 D75

445 119.75 36.30 -2.75 81.05 91.81 110.94 127.11
450 125.60 35.90 -2.90 87.26 98.00 117.01 133.02
455 125.55 34.25 -2.85 88.94 99.24 117.41 132.69
460 125.50 32.60 -2.80 90.62 100.47 117.81 132.36
465 123.40 30.25 -2.70 91.00 100.19 116.34 129.85
470 121.30 27.90 -2.60 91.38 99.92 114.86 127.33
475 121.30 26.10 -2.60 93.25 101.33 115.39 127.07
480 121.30 24.30 -2.60 95.12 102.74 115.92 126.81
485 117.40 22.20 -2.20 93.54 100.41 112.37 122.30
490 113.50 20.10 -1.80 91.97 98.08 108.81 117.79
495 113.30 18.15 -1.65 93.85 99.38 109.08 117.19
500 113.10 16.20 -1.50 95.73 100.68 109.35 116.59
505 111.95 14.70 -1.40 96.17 100.69 108.58 115.15
510 110.80 13.20 -1.30 96.62 100.70 107.80 113.71
515 108.65 10.90 -1.25 96.88 100.34 106.30 111.18
520 106.50 8.60 -1.20 97.13 99.99 104.79 108.66
525 107.65 7.35 -1.10 99.62 102.10 106.24 109.55
530 108.80 6.10 -1.00 102.10 104.21 107.69 110.45
535 107.05 5.15 -0.75 101.43 103.16 106.05 108.37
540 105.30 4.20 -0.50 100.76 102.10 104.41 106.29
545 104.85 3.05 -0.40 101.54 102.54 104.22 105.60
550 104.40 1.90 -0.30 102.32 102.97 104.05 104.90
555 102.20 0.95 -0.15 101.16 101.48 102.02 102.45
560 100.00 0.00 0.00 100.00 100.00 100.00 100.00
565 98.00 -0.80 0.10 98.87 98.61 98.17 97.81
570 96.00 -1.60 0.20 97.73 97.22 96.33 95.62
575 95.55 -2.55 0.35 98.33 97.48 96.06 94.91
580 95.10 -3.50 0.50 98.92 97.75 95.79 94.21
585 92.10 -3.50 1.30 96.21 94.59 92.24 90.60
590 89.10 -3.50 2.10 93.50 91.43 88.69 87.00
595 89.80 -4.65 2.65 95.59 92.92 89.35 87.11
600 90.50 -5.80 3.20 97.68 94.42 90.01 87.23
605 90.40 -6.50 3.65 98.48 94.78 89.80 86.68
610 90.30 -7.20 4.10 99.27 95.14 89.60 86.14
615 89.35 -7.90 4.40 99.15 94.68 88.65 84.86
620 88.40 -8.60 4.70 99.04 94.22 87.70 83.58
625 86.20 -9.05 4.90 97.38 92.33 85.49 81.16
630 84.00 -9.50 5.10 95.72 90.45 83.29 78.75
635 84.55 -10.20 5.90 97.28 91.39 83.49 78.59
640 85.10 -10.90 6.70 98.85 92.33 83.70 78.43
645 83.50 -10.80 7.00 97.26 90.59 81.86 76.61
650 81.90 -10.70 7.30 95.66 88.85 80.03 74.80
655 82.25 -11.35 7.95 96.92 89.58 80.12 74.56
660 82.60 -12.00 8.60 98.18 90.31 80.21 74.32
665 83.75 -13.00 9.20 100.59 92.13 81.25 74.87
670 84.90 -14.00 9.80 103.00 93.95 82.28 75.42

Table F.1. (continued) Components S0 (λ ), S1 (λ ), and S2 (λ ) used to compute CIE illumi-

nants D.

i i

i i
 i i

i i

957

λ (nm) S0 (λ ) S1 (λ ) S2 (λ ) D50 D55 D65 D75

675 83.10 -13.80 10.00 101.06 91.95 80.28 73.50
680 81.30 -13.60 10.20 99.13 89.95 78.28 71.57
685 76.60 -12.80 9.25 93.25 84.81 74.00 67.71
690 71.90 -12.00 8.30 87.37 79.67 69.72 63.85
695 73.10 -12.65 8.95 89.48 81.26 70.67 64.46
700 74.30 -13.30 9.60 91.60 82.84 71.61 65.07
705 75.35 -13.10 9.05 92.24 83.84 72.98 66.57
710 76.40 -12.90 8.50 92.88 84.84 74.35 68.07
715 69.85 -11.75 7.75 84.86 77.54 67.98 62.25
720 63.30 -10.60 7.00 76.85 70.23 61.60 56.44
725 67.50 -11.10 7.30 81.68 74.77 65.74 60.34
730 71.70 -11.60 7.60 86.50 79.30 69.89 64.24
735 74.35 -11.90 7.80 89.54 82.14 72.49 66.69
740 77.00 -12.20 8.00 92.57 84.99 75.09 69.15
745 71.10 -11.20 7.35 85.40 78.43 69.34 63.89
750 65.20 -10.20 6.70 78.22 71.88 63.59 58.63
755 56.45 -9.00 5.95 67.96 62.34 55.01 50.62
760 47.70 -7.80 5.20 57.69 52.79 46.42 42.62
765 58.15 -9.50 6.30 70.30 64.36 56.61 51.98
770 68.60 -11.20 7.40 82.92 75.92 66.81 61.35
775 66.80 -10.80 7.10 80.59 73.87 65.09 59.84
780 65.00 -10.40 6.80 78.27 71.82 63.38 58.32
785 65.50 -10.50 6.90 78.91 72.37 63.84 58.73
790 66.00 -10.60 7.00 79.55 72.93 64.30 59.14
795 63.50 -10.15 6.70 76.47 70.14 61.88 56.93
800 61.00 -9.70 6.40 73.40 67.35 59.45 54.73
805 57.15 -9.00 5.95 68.66 63.04 55.71 51.32
810 53.30 -8.30 5.50 63.92 58.72 51.96 47.92
815 56.10 -8.80 5.80 67.34 61.86 54.70 50.41
820 58.90 -9.30 6.10 70.77 64.99 57.44 52.91
825 60.40 -9.55 6.30 72.60 66.65 58.88 54.23
830 61.90 -9.80 6.50 74.44 68.30 60.31 55.54

Table F.1. (continued) Components S0 (λ ), S1 (λ ), and S2 (λ ) used to compute CIE illumi-

nants D.

i i

i i
 i i

i i

Appendix G
Chromaticity Coordinates
of Paints

In this appendix, we list a set of chromaticity coordinates for paints. Table G.1
is after Barnes, who measured the reflectances of both organic and inorganic pig-
ments used for painting [68]. This table is useful for specifying colors in render-
ing equations. It should be noted that these values were obtained under standard
illuminant C.

Pigment Y x y Pigment Y x y
Red lead 32.76 0.5321 0.3695 Burnt umber 5.08 0.4000 0.3540
English vermillion 22.26 0.5197 0.3309 Malachite 41.77 0.2924 0.3493
Cadmium red 20.78 0.5375 0.3402 Chrome green (medium) 16.04 0.3133 0.4410
Madder lake 33.55 0.3985 0.2756 Verdigris 16.67 0.1696 0.2843
Alizarin crimson 6.61 0.5361 0.3038 Emerald green 39.12 0.2446 0.4215
Carmine lake 5.04 0.4929 0.3107 Terre verte 29.04 0.3092 0.3510
Dragon’s blood 4.94 0.4460 0.3228 Cobalt green 19.99 0.2339 0.3346
Realgar 32.35 0.4868 0.3734 Viridian 9.85 0.2167 0.3635
Venetian red 13.12 0.4672 0.3462 Cobalt blue 16.81 0.1798 0.1641
Indian red 10.34 0.3797 0.3194 Genuine ultramarine 18.64 0.2126 0.2016
Cadmium orange (medium) 42.18 0.5245 0.4260 Cerulean blue 18.19 0.1931 0.2096
Camboge 29.87 0.4955 0.4335 Azurite 9.26 0.2062 0.2008
Yellow ochre 41.15 0.4303 0.4045 Manganese violet 26.99 0.3073 0.2612
Zinc yellow 82.57 0.4486 0.4746 Cobalt violet 9.34 0.2817 0.1821
Indian yellow 40.01 0.4902 0.4519 Smalt 8.25 0.1898 0.1306
Saffron 33.45 0.4470 0.4345 Blue verditer 16.46 0.1899 0.1880
Yellow lake 27.24 0.4410 0.4217 French ultramarine blue 7.84 0.1747 0.1151
Strontium lemon yellow 84.36 0.4157 0.4732 Indigo 3.62 0.3180 0.2728
Hansa yellow 5 G 75.09 0.4447 0.5020 Prussian blue 1.30 0.2883 0.2453
Chrome yellow (medium) 63.07 0.4843 0.4470 Titanium white A 95.75 0.3124 0.3199
Cadmium yellow (light) 76.66 0.4500 0.4819 Titanium white B 95.66 0.3122 0.3196
Orpiment 64.45 0.4477 0.4402 Zinc white 94.88 0.3122 0.3200
Cobalt yellow 50.48 0.4472 0.4652 White lead 87.32 0.3176 0.3242
Raw sienna 20.03 0.4507 0.4043 Lamp black 4.08 0.3081 0.3157
Burnt sienna 7.55 0.4347 0.3414 Ivory black 2.22 0.3055 0.3176
Raw umber 7.44 0.3802 0.3693 .

Table G.1. Yxy coordinates for a selection of organic and inorganic paints.

959

i i

i i
 i i

i i

Bibliography

[1] E. Aas and J. Bogen. “Colors of Glacier Water.” Water Resources Research 24:4 (1988), 561–
565.
[2] P. M. Acosta-Serafini, I. Masaki, and C. G. Sodini. “A 1/3” VGA Linear Wide Dynamic Range
CMOS Image Sensor Implementing a Predictive Multiple Sampling Algorithm with Overlap-
ping Integration Intervals.” IEEE Journal of Solid-State Circuits 39:9 (2004), 1487–1496.
[3] A. Adams and G. Haegerstrom-Portnoy. “Color Deficiency.” In Diagnosis and Management in
Vision Care. Boston: Butterworth, 1987.
[4] D. L. Adams and S. Zeki. “Functional Organization of Macaque V3 for Stereoscopic Depth.”
Journal of Neurophysiology 86:5 (2001), 2195–2203.
[5] J. Adams, K. Parulski, and K. Spaulding. “Color Processing in Digital Cameras.” IEEE Micro
18:6 (1998), 20–30.
[6] M. M. Adams, P. R. Hof, R. Gattass, M. J. Webster, and L. G. Ungerleider. “Visual Cortical
Projections and Chemoarchitecture of Macaque Monkey Pulvinar.” Journal of Comparative
Neurology 419:3 (2000), 377–393.
[7] J. E. Adams Jr. “Interactions between Color Plane Interpolation and Other Image Processing
Functions in Electronic Photography.” In Proceedings of the SPIE 2416, pp. 144–155. Belling-
ham, WA: SPIE, 1995.
[8] A. Adams. The camera. The Ansel Adams Photography series, Little, Brown and Company,
1980.
[9] A. Adams. The negative. The Ansel Adams Photography series, Little, Brown and Company,
1981.
[10] A. Adams. The print. The Ansel Adams Photography series, Little, Brown and Company, 1983.
[11] E. H. Adelson and J. R. Bergen. “The Plenoptic Function and the Elements of Early Vision.” In
Computational Models of Visual Processing, pp. 3–20. Cambridge, MA: MIT Press, 1991.
[12] E. H. Adelson and J. Y. A. Wang. “The Plenoptic Camera.” IEEE Transactions on Pattern
Analysis and Machine Intelligence 14:2 (1992), 99–106.
[13] E. A. Adelson. “Lightness Perception and Lightness Illusions.” In The New Cognitive Neuro-
sciences, edited by M. Gazzaniga, 2nd edition, pp. 339–351. Cambridge, MA: MIT Press, 2000.
[14] M. Aggarwal and N. Ahuja. “Split Aperture Imaging for High Dynamic Range.” International
Journal of Computer Vision 58:1 (2004), 7–17.
[15] M. Aggarwal, H. Hua, and N. Ahuja. “On Cosine-Fourth and Vignetting Effects in Real Lenses.”
In Proceedings of the Eighth IEEE International Conference on Computer Vision (ICCV), 1, 1,
pp. 472–479, 2001.

961

i i

i i
 i i

i i

962 Bibliography

[16] M. Aguilar and W. S. Stiles. “Saturation of the Rod Mechanism of the Retina at High Levels of
Stimulation.” Optica Acta 1:1 (1954), 59–65.
[17] W. E. Ahearn and O. Sahni. “The Dependence of the Spectral and Electrical Properties of AC
Plasma Panels on the Choice and Purity of the Gas Mixture.” SID International Symposium
Digest 7 (1978), 44–45.
[18] L. von Ahn, M. Blum, N. J. Hopper, and J. Langford. “Captcha: Telling Humans and Computers
Apart Automatically.” In Advances in Cryptology - Proceedings of Eurocrypt 2003, Lecture
Notes in Computer Science 2656, pp. 294–311. Berlin: Springer, 2003.
[19] L. von Ahn, M. Blum, N. J. Hopper, and J. Langford. “Telling Humans and Computers Apart
Automatically.” Communications of the ACM 47:2 (2004), 57–60.
[20] T. Ajito, T. Obi, M. Yamaguchi, and N. Ohyama. “Multiprimary Color Display for Liquid
Crystal Display Projectors using Diffraction Grating.” Optical Engineering 38:11 (1999), 1883–
1888.
[21] T. Ajito, T. Obi, M. Yamaguchi, and N. Ohyama. “Expanded Color Gamut Reproduced by Six-
Primary Projection Display.” In Proceedings of the SPIE 3954, edited by M. H. Wu, pp. 130–137.
Bellingham, WA: SPIE, 2000.
[22] N. Akahane, S. Sugawa, S. Adachi, K. Mori, T. Ishiuchi, and K. Mizobuchi. “A Sensitivity
and Linearity Improvement of a 100-dB Dynamic Range CMOS Image Sensor using a Lateral
Overflow Integration Capacitor.” IEEE Journal of Solid-State Circuits 41:4 (2006), 851–858.
[23] T. Akenine-Möller and E. Haines. Real-Time Rendering, 2nd edition. Natick, MA: A K Peters,
2002.
[24] A. O. Akyüz and E. Reinhard. “Noise Reduction in High Dynamic Range Imaging.” Journal of
Visual Communication and Image Representation 18:5 (2007), 366–376.
[25] A. O. Akyüz and E. Reinhard. “Perceptual Evaluation of Tone Reproduction Operators using
the Cornsweet-Craik-O’Brien Illusion.” ACM Transactions on Applied Perception 4:4 (2007),
20–1 – 20–29.
[26] A. O. Akyüz, R. Fleming, B. Riecke, E. Reinhard, and H. Bülthoff. “Do HDR Displays Support
LDR Content? A Psychophysical Investigation.” ACM Transactions on Graphics 26:3 (2007),
38–1 – 38–7.
[27] E. Allen. “Analytical Color Matching.” Journal of Paint Technology 39:509 (1967), 368–376.
[28] E. Allen. “Basic Equations used in Computer Color matching II. Tristimulus Match, Two-
Constant Theory.” Journal of the Optical Society of America 64:7 (1974), 991–993.
[29] D. H. Alman. “CIE Technical Committee 1-29, Industrial Color Difference Evaluation Progress
Report.” Colour Research and Application 18 (1993), 137–139.
[30] M. Alpern and J. Moeller. “The Red and Green Cone Visual Pigments of Deuteranomalous
Trichromacy.” Journal of Physiology 266 (1977), 647–675.
[31] M. Alpern and S. Torii. “The Luminosity Curve of the Deuteranomalous Fovea.” Journal of
General Physiology 52:5 (1968), 738–749.
[32] M. Alpern and S. Torii. “The Luminosity Curve of the Protanomalous Fovea.” Journal of
General Physiology 52:5 (1968), 717–737.
[33] M. Alpern and T. Wake. “Cone Pigment in Human Deutan Color Vision Defects.” Journal of
Physiology 266 (1977), 595–612.
[34] M. Alpern. “The Stiles-Crawford Effect of the Second Kind (SCII): A Review.” Perception 15:6
(1986), 785–799.
[35] B. Anderson and J. Winawer. “Image Segmentation and Lightness Perception.” Nature 434
(2005), 79–83.

i i

i i
 i i

i i

Bibliography 963

[36] R. Anderson. “Matrix Description of Radiometric Quantities.” Applied Optics 30:7 (1991),
858–867.
[37] B. L. Anderson. “Perceptual Organization and White’s Illusion.” Perception 32 (2003), 269–284.
[38] A. Angelucci, J. B. Levitt, E. J. S. Walton, J.-M. Hupé, J. Bullier, and J. S. Lund. “Circuits for
Local and Global Signal Integration in Primary Visual Cortex.” Journal of Neuroscience 22:19
(2002), 8633–8646.
[39] A. Angström. “The Albedo of Various Surfaces of Ground.” Geografiska Annalen 7 (1925),
323–342.
[40] R. A. Applegate and V. Lakshminarayanan. “Parametric Representation of Stiles-Crawford
Functions: Normal Variation of Peak Location and Directionality.” Journal of the Optical Soci-
ety of America A 10:7 (1993), 1611–1623.
[41] G. B. Arden. “The Importance of Measuring Contrast Sensitivity in Cases of Visual Distur-
bances.” British Journal of Ophthalmology 62 (1978), 198–209.
[42] L. Arend and A. Reeves. “Simultaneous Color Constancy.” Journal of the Optical Society of
America A 3:10 (1986), 1743–1751.
[43] P. Artal, E. Berrio, A. Guirao, and P. Piers. “Contribution of the Cornea and Internal Surfaces to
the Change of Ocular Aberrations with Age.” Journal of the Optical Society of America A 19:1
(2002), 137–143.
[44] M. Ashikhmin and J. Goral. “A Reality Check for Tone Mapping Operators.” ACM Transactions
on Applied Perception 3:4 (2006), 399–411.
[45] M. Ashikhmin and P. Shirley. “An Anisotropic Phong BRDF Model.” journal of graphics tools
5:2 (2002), 25–32.
[46] M. Ashikhmin, S. Premoz̆e, and P. Shirley. “A Microfacet-Based BRDF Generator.” In Proceed-
ings of ACM SIGGRAPH 2000, Computer Graphics Proceedings, Annual Conference Series,
edited by K. Akeley, pp. 65–74. Reading, MA: Addison-Wesley, 2000.
[47] M. Ashikhmin. “A Tone Mapping Algorithm for High Contrast Images.” In Proceedings of
the 13th Eurographics Workshop on Rendering, pp. 145–155. Aire-la-Ville, Switzerland: Euro-
graphics Association, 2002.
[48] ASTM. “Standard Tables for Reference Solar Spectral Irradiances: Direct Normal and Hemi-
spherical on 37◦ Tilted Surface.” Technical Report G173-03, American Society for Testing and
Materials, West Conshohocken, PA, 2003.
[49] ASTM. “Standard Terminology of Appearance.” Technical Report E 284, American Society for
Testing and Materials, West Conshohocken, PA, 2006.
[50] D. A. Atchison and G. Smith. “Chromatic Dispersions of the Ocular Media of Human Eyes.”
Journal of the Optical Society of America A 22:1 (2005), 29–37.
[51] D. A. Atchison, G. Smith, and N. Efron. “The Effect of Pupil Size on Visual Acuity in Uncor-
rected and Corrected Myopia.” American Journal of Optometry and Physiological Optics 56:5
(1979), 315–323.
[52] J. J. Atick and N. A. Redlich. “What Does the Retina Know about Natural Scenes?” Neural
Computation 4 (1992), 196–210.
[53] H. Aubert. Physiologie der Netzhaut. Breslau, Germany: E Morgenstern, 1865.
[54] V. A. Babenko, L. G. Astafyeva, and V. N. Kuzmin. Electromagnetic Scattering in Disperse
Media: Inhomogeneous and Anisotropic Particles. Berlin: Springer-Verlag, 2003.
[55] R. J. Baddeley and P. J. B. Hancock. “A Statistical Analysis of Natural Images Matches Psy-
chophysically Derived Orientation Tuning Curves.” Proceedings of the Royal Society of London
B 246 (1991), 219–223.

i i

i i
 i i

i i

964 Bibliography

[56] R. Bajcsy, S. W. Lee, and A. Leonardis. “Color Image Segmentation with Detection of High-
lights and Local Illumination Induced by Interreflection.” In Proceedings of the International
Conference on Pattern Recognition, pp. 785–790. Washington, DC: IEEE, 1990.
[57] R. Bajcsy, S. W. Lee, and A. Leonardis. “Detection of Diffuse and Specular Interface Reflections
and Inter-reflections by Color Image Segmentation.” International Journal of Computer Vision
17:3 (1996), 241–272.
[58] R. M. Balboa, C. W. Tyler, and N. M. Grzywacz. “Occlusions Contribute to Scaling in Natural
Images.” Vision Research 41:7 (2001), 955–964.
[59] G. V. G. Baranoski, P. Shirley, J. G. Rokne, T. Trondsen, and R. Bastos. “Simulating the Aurora
Borealis.” In Eighth Pacific Conference on Computer Graphics and Applications, pp. 2–14. Los
Alamitos, CA: IEEE Computer Society Press, 2000.
[60] H. B. Barlow, R. M. Hill, and W. R. Levick. “Retinal Ganglion Cells Responding Selectively
to Direction and Speed of Image Motion in the Rabbit.” Journal of Physiology (London) 173
(1964), 377–407.
[61] H. B. Barlow. “The Size of Ommatidia in Apposition Eyes.” Journal of Experimental Biology
29:4 (1952), 667–674.
[62] H. B. Barlow. “Summation and Inhibition in the Frog’s Retina.” Journal of Physiology (London)
119 (1953), 69–88.
[63] H. B. Barlow. “Dark and Light Adaptation: Psychophysics.” In Handbook of Sensory Physiology
VII/4, edited by D. Jameson and L. M. Hurvich, pp. 1–28. Berlin: Springer-Verlag, 1972.
[64] K. Barnard and B. Funt. “Investigations into Multi-Scale Retinex (MSR).” In Color Imaging:
Vision and Technology, edited by L. W. MacDonald and M. R. Luo, pp. 9–17. New York: John
Wiley and Sons, 1999.
[65] K. Barnard and B. Funt. “Camera Characterization for Color Research.” Color Research and
Application 27 (2002), 152–163.
[66] K. Barnard, L. Martin, A. Coath, and B. Funt. “A Comparison of Computational Color Con-
stancy Algorithms — Part II: Experiments with Image Cata.” IEEE Transactions on Image
Processing 11:9 (2002), 985–996.
[67] C. Barnes, J. Wei, and S. K. Shevell. “Chromatic Induction with Remote Chromatic Contrast
Varied in Magnitude, Spatial Frequency, and Chromaticity.” Vision Research 39 (1999), 3561–
3574.
[68] N. F. Barnes. “Color Characteristics of Artists’ Paints.” Journal of the Optical Society of America
29:5 (1939), 208–214.
[69] P. Barone, A. Batardiere, K. Knoblauch, and H. Kennedy. “Laminar Distribution of Neurons
in Extrastriate Areas Projecting to Visual Areas V1 and V4 Correlates with the Hierarchical
Rank and Indicates the Operation of a Distance Rule.” Journal of Neuroscience 20:9 (2000),
3263–3281.
[70] B. Barsky, L. Chu, and S. Klein. “Cylindrical Coordinate Representations for Modeling Surfaces
of the Cornea and Contact Lenses.” In Proceedings of the International Conference on Shape
Modeling and Applications (SMI-99), edited by B. Werner, pp. 98–115. Los Alamitos, CA: IEEE
Computer Society, 1999.
[71] A. Bartels and S. Zeki. “The Architecture of the Colour Centre in the Human Visual Brain: New
Results and a Review.” European Journal of Neuroscience 12:1 (2000), 172–193.
[72] C. Bartleson and E. Breneman. “Brightness Perception in Complex Fields.” Journal of the
Optical Society of America 57 (1967), 953–957.
[73] M. Bass, E. W. van Stryland, D. R. Williams, and W. L. Wolfe, editors. Handbook of Optics:
Fundamentals, Techniques and Design, Second edition. New York: McGraw-Hill, 1995.

i i

i i
 i i

i i

Bibliography 965

[74] B. Bastani, W. Cressman, and B. Funt. “Calibrated Color Mapping Between LCD and CRT
Displays: A Case Study.” Colour Research and Application 30:6 (2005), 438–447.
[75] B. Baxter, H. Ravindra, and R. A. Normann. “Changes in Lesion Detectability Caused by Light
Adaptation in Retinal Photoreceptors.” Investigative Radiology 17 (1982), 394–401.
[76] B. E. Bayer. “Color Imaging Array.”, 1976. U.S. Patent 3,971,065.
[77] D. A. Baylor, B. J. Nunn, and J. L. Schnapf. “The Photocurrent, Noise and Spectral Sensitivity
of Rods of the Monkey Macaca fascicularis.” Journal of Physiology 357:1 (1984), 576–607.
[78] D. A. Baylor, B. J. Nunn, and J. L. Schnapf. “Spectral Sensitivity of Cones of the Monkey
Macaca fascicularis.” Journal of Physiology 390:1 (1987), 145–160.
[79] H. Becker, H. Vestweber, A. Gerhard, P. Stoessel, and R. Fortte. “Novel Host Materials for Ef-
ficient and Stable Phosphorescent OLED Devices.” In Proceedings of the International Display
Manufacturing Conference, pp. 329–330. San Jose, CA: Society for Information Display, 2005.
[80] R. E. Bedford and G. Wyszecki. “Axial Chromatic Aberration of the Human Eye.” Journal of
the Optical Society of America 47 (1957), 564–565.
[81] P. R. Bélanger. “Linear-Programming Approach to Color-Recipe Formulations.” Journal of the
Optical Society of America 64:11 (1974), 1541–1544.
[82] A. J. Bell and T. J. Sejnowski. “Edges Are the ’Independent Components’ of Natural Scenes.” In
Advances in Neural Information Processing Systems, pp. 831–837. Cambridge, MA: MIT Press,
1996.
[83] A. J. Bell and T. J. Sejnowski. “The Independent Components of Natural Scenes Are Edge
Filters.” Vision Research 37 (1997), 3327–3338.
[84] G. Berbecel. Digital Image Display: Algorithms and Implementation. Chichester: John Wiley
and Sons, 2003.
[85] T. T. J. M. Berendschot, J. van de Kraats, and D. van Norren. “Wavelength Dependence of
the Stiles-Crawford Effect Explained by Perception of Backscattered Light from the Choroid.”
Journal of the Optical Society of America A 18:7 (2001), 1445–1451.
[86] L. D. Bergman, B. E. Rogowitz, and L. A. Treinish. “A Rule-Based Tool for Assisting Col-
ormap Selection.” In Proceedings of IEEE Visualization, pp. 118–125. Los Alamitos, CA: IEEE
Computer Society, 1995.
[87] B. Berlin and P. Kay. Basic Color Terms: Their Universality and Evolution. Berkeley, CA:
University of California Press, 1969.
[88] B. J. Berne and R. Pecora. Dynamic Light Scattering (with Applications to Chemistry, Biology
and Physics). Mineola, NY: Dover Publications, 2000.
[89] R. S. Berns and M. J. Shiyu. “Colorimetric Characterization of a Desk-Top Drum Scanner using
a Spectral Model.” Journal of Electronic Imaging 4 (1995), 360–372.
[90] R. S. Berns, D. H. Alman, L. Reniff, G. D. Snyder, and M. R. Balonon-Rosen. “Visual Determi-
nation of Suprathreshold Color-Difference Tolerances using Probit Analysis.” Colour Research
and Application 16 (1991), 297–316.
[91] R. S. Berns, M. E. Gorzynski, and R. J. Motta. “CRT Colorimetry – Part II: Metrology.” Color
Research and Application 18:5 (1993), 315–325.
[92] R. S. Berns. “Methods for Characterizing CRT Displays.” Displays 16:4 (1996), 173–182.
[93] R. S. Berns. Billmeyer and Saltzman’s Principles of Color Technology, Third edition. New York:
John Wiley and Sons, 2000.
[94] I. Biederman and P. Kalocsai. “Neural and Psychological Analysis of Object and Face Recog-
nition.” In Face Recognition: From Theory to Applications, pp. 3–25. Berlin: Springer-Verlag,
1998.

i i

i i
 i i

i i

966 Bibliography

[95] M. Bigas, E. Cabruja, J. Forest, and J. Salvi. “Review of CMOS Image Sensors.” Microelec-
tronics Journal 37:5 (2006), 433–451.
[96] J. Birnstock, J. Blässing, A. Hunze, M. Scheffel, M. Stößel, K. Heuser, G. Wittmann, J. Wörle,
and A. Winnacker. “Screen-Printed Passive Matrix Displays Based on Light Emitting Polymers.”
Applied Physics Letters 78:24 (2001), 3905–3907.
[97] J. Birnstock, J. Blässing, A. Hunze, M. Scheffel, M. Stößel, K. Heuser, J. Wörle, G. Wittmann,
and A. Winnacker. “Screen-Printed Passive Matrix Displays and Multicolor Devices.” In Pro-
ceedings of the SPIE 4464, pp. 68–75. Bellingham, WA: SPIE, 2002.
[98] C. M. Bishop. Neural Networks for Pattern Recognition. Oxford: Oxford University Press,
1995.
[99] D. L. Bitzer and H. G. Slottow. “The Plasma Display Panel — A Digitally Addressable Display
with Inherent Memory.” In Proceedings of the AFIPS Conference 29, pp. 541–547. American
Federation of Information Processing Societies, 1966.
[100] H. R. Blackwell. “Luminance difference thresholds.” In Handbook of Sensory Physiology,
VII/4, edited by D. Jameson and L. M. Hurvich, pp. 78–101. Berlin: Springer-Verlag, 1972.
[101] C. Blakemore and F. Campbell. “On the Existence of Neurons in the Human Visual System
Electively Sensitive to Orientation and Size of Retinal Images.” Journal of Physiology 203
(1969), 237–260.
[102] B. Blakeslee and M. E. McCourt. “A Multi-Scale Spatial Filtering Account of the White Effect,
Simultaneous Brightness Contrast and Grating Induction.” Vision Research 39 (1999), 4361–
4377.
[103] N. Blanc. “CCD versus CMOS — Has CCD Imaging Come to an End?” In Photogrammetric
Week 01, edited by D. Fritsch and R. Spiller, pp. 131–137. Heidelberg: Wichmann Verlag, 2001.
[104] G. G. Blasdel and J. S. Lund. “Termination of Afferent Axons in Macaque Striate Cortex.”
Journal of Neuroscience 3:7 (1983), 1389–1413.
[105] J. F. Blinn. “Models of Light Reflection for Computer Synthesized Pictures.” Proc. SIGGRAPH
’77, Computer Graphics 11 (1977), 192–198.
[106] J. F. Blinn. “Return of the Jaggy.” IEEE Computer Graphics and Applications 9:2 (1989),
82–89.
[107] J. F. Blinn. “What We Need Around Here Is More Aliasing.” IEEE Computer Graphics and
Applications 9:1 (1989), 75–79.
[108] C. F. Bohren. “Colors of the Sea.” Weatherwise 35:5 (1982), 256–260.
[109] G. F. Bohren. “The Green Flash.” Weatherwise 35:6 (1982), 271–275.
[110] C. F. Bohren. “More About Colors of the Sea.” Weatherwise 36:6 (1983), 311–316.
[111] C. F. Bohren. “Scattering by Particles.” In Handbook of Optics: Fundamentals, Techniques
and Design, Volume 1, edited by M. Bass, E. W. van Stryland, D. R. Williams, and W. L. Wolfe,
Second edition. New York: McGraw-Hill, 1995.
[112] M. Born and E. Wolf. Principles of Optics, Seventh edition. Cambridge, UK: Cambridge
University Press, 1999.
[113] T. Bossomaier and A. W. Snyder. “Why Spatial Frequency Processing in the Visual Cortex?”
Vision Research 26 (1986), 1307–1309.
[114] P. Bouguer. Traité d’optique sur la gradation de la lumiere: Ouvrage posthume. Paris: M
l’Abbé de la Caille, 1760.
[115] D. Boussaoud, L. G. Ungerleider, and R. Desimone. “Pathways for Motion Analysis: Cortical
Connections of the Medial Superior Temporal and Fundus of the Superior Temporal Visual Areas
in the Macaque.” Journal of Comparative Neurology 296:3 (1990), 462–495.

i i

i i
 i i

i i

Bibliography 967

[116] A. Bovik, editor. Handbook of Image and Video Processing. San Diego, CA: Academic Press,
2000.
[117] R. M. Boynton and D. N. Whitten. “Visual Adaptation in Monkey Cones: Recordings of Late
Receptor Potentials.” Science 170 (1970), 1423–1426.
[118] R. M. Boynton. Human Color Vision. New York: Holt, Rinehart and Winston, 1979.
[119] D. H. Brainard and W. T. Freeman. “Bayesean Color Constancy.” Journal of the Optical Society
of America A 14:7 (1997), 1393–1411.
[120] D. H. Brainard, W. A. Brunt, and J. M. Speigle. “Color Constancy in the Nearly Natural Image.
I. Assymetric Matches.” Journal of the Optical Society of America A 14:9 (1997), 2091–2110.
[121] D. H. Brainard, A. Roorda, Y. Yamauchi, J. B. Calderone, A. Metha, M. Neitz, J. Neitz, D. R.
Williams, and G. H. Jacobs. “Functional Consequences of the Relative Numbers of L and M
Cones.” Journal of the Optical Society of America A 17:3 (2000), 607–614.
[122] D. H. Brainard, D. G. Pelli, and T. Robson. “Display Characterization.” In Encylopedia of
Imaging Science and Technology, edited by J. Hornak, pp. 172–188. New York: John Wiley and
Sons, 2002.
[123] D. H. Brainard. “Color Constancy in the Nearly Natural Image. II. Achromatic Loci.” Journal
of the Optical Society of America A 15 (1998), 307–325.
[124] G. J. Braun and M. D. Fairchild. “Techniques for Gamut Surface Definition and Visualization.”
In Proceedings of the 5th IS&T/SID Color Imaging Conference, pp. 147–152. Springfield, VA:
IS&T, 1997.
[125] G. Brelstaff and F. Chessa. “Practical Application of Visual Illusions: Errare Humanum Est.” In
Proceedings of the 2nd Symposium on Applied Perception in Graphics and Visualization (APGV),
pp. 161–161. New York: ACM, 2005.
[126] E. Breneman. “Corresponding Chromaticities for Different States of Adaptation to Complex
Visual Fields.” Journal of the Optical Society of America, A 4 (1987), 1115–1129.
[127] P. Bressan, E. Mingolla, L. Spillmann, and T. Watanabe. “Neon Color Spreading: A Review.”
Perception 26:11 (1997), 1353–1366.
[128] H. Brettel and F. Viénot. “Web Design for the Colour-Blind User.” In Colour Imaging: Vision
and Technology, edited by L. W. MacDonald and M. R. Luo, pp. 55–71. Chichester, UK: John
Wiley and Sons, Ltd., 1999.
[129] H. Brettel, F. Viénot, and J. D. Mollon. “Computerized Simulation of Color Appearance for
Dichromats.” Journal of the Optical Society of America A 14:10 (1997), 2647–2655.
[130] D. Brewster. “A Note Explaining the Cause of an Optical Phenomenon Observed by the Rev.
W. Selwyn.” In Report of the Fourteenth Meeting of the British Association for the Advancement
of Science, edited by J. Murray. London, 1844.
[131] M. H. Brill and J. Larimer. “Avoiding On-Screen Metamerism in N-Primary Displays.” Journal
of the Society for Information Display 13:6 (2005), 509–516.
[132] M. H. Brill. “Image Segmentation by Object Color: A Unifying Framework and Connection
to Color Constancy.” Journal of the Optical Society of America A 7:10 (1986), 2041–2047.
[133] G. S. Brindley. “The Discrimination of After-Images.” Journal of Physiology 147:1 (1959),
194–203.
[134] R. W. Brislin. “The Ponzo Illusion: Additional Cues, Age, Orientation, and Culture.” Journal
of Cross-Cultural Psychology 5:2 (1974), 139–161.
[135] K. H. Britten, M. N. Shadlen, W. T. Newsome, and J. A. Movshon. “The Analysis of Visual
Motion: A Comparison of Neuronal and Psychophysical Performance.” Journal of Neuroscience
12:12 (1992), 4745–4765.

i i

i i
 i i

i i

968 Bibliography

[136] A. Brockes. “Vergleich der Metamerie-Indizes bei Lichtartwechsel von Tageslicht zur
Glühlampe and zu verschiedenen Leuchtstofflampen.” Die Farbe 18:223.
[137] A. Brockes. “Vergleich von berechneten Metamerie-Indizes mit Abmusterungsergebnissen.”
Die Farbe 19:135 (1970), 1–10.
[138] K. Brodmann. Vergleichende Lokalisationslehre der Großhirnrhinde. Leipzig, Germany: Barth
Verlag, 1909. Translated by L J Garey, Localisation in the Cerebral Cortex (Smith-Gordon,
London, 1994).
[139] J. Broerse, T. Vladusich, and R. P. O’Shea. “Colour at Edges and Colour Spreading in McCol-
lough Effects.” Vision Research 39 (1999), 1305–1320.
[140] N. Bruno, P. Bernardis, and J. Schirillo. “Lightness, Equivalent Backgrounds and Anchoring.”
Perception and Psychophysics 59:5 (1997), 643–654.
[141] G. Buchsbaum. “A Spatial Processor Model for Object Color Perception.” Journal of the
Franklin Institute 310 (1980), 1–26.
[142] Bureau International des Poids et Mesures. “Le Système International d’Unités (The Interna-
tional System of Units).”, 2006.
[143] J. H. Burroughes, D. D. C. Bradley, A. R. Brown, R. N. Marks, K. Mackay, R. H. Friend, P. L.
Burns, and A. B. Holmes. “Light-Emitting Diodes based on Conjugated Polymers.” Nature
347:6293 (1990), 539–541.
[144] G. J. Burton and I. R. Moorhead. “Color and Spatial Structure in Natural Scenes.” Applied
Optics 26:1 (1987), 157–170.
[145] A. Calabria and M. D. Fairchild. “Herding CATs: A Comparison of Linear Chromatic Adap-
tation Transforms for CIECAM97s.” In IS&T/SID 9th Color Imaging Conference, pp. 174–178.
Springfield, VA: Society for Imaging Science and Technology, 2001.
[146] A. J. Calabria and M. D. Fairchild. “Perceived Image Contrast and Observer Preference I: The
Effects of Lightness, Chroma, and Sharpness Manipulations on Contrast Perception.” Journal
of Imaging Science and Technology 47 (2003), 479–493.
[147] A. Calabria and M. D. Fairchild. “Perceived Image Contrast and Observer Preference II: Empir-
ical Modeling of Perceived Image Contrast and Observer Preference Data.” Journal of Imaging
Science and Technology 47 (2003), 494–508.
[148] E. M. Callaway and A. K. Wiser. “Contributions of Individual Layer 2-5 Spiny Neurons to
Local Circuits in Macaque Primary Visual Cortex.” Visual Neuroscience 13:5 (1996), 907–922.
[149] E. Camalini, editor. Optical and Acoustical Holography. New York: Plenum Press, 1972.
[150] F. W. Campbell and A. H. Gregory. “Effect of Pupil Size on Acuity.” Nature (London) 187:4743
(1960), 1121–1123.
[151] F. W. Campbell and R. W. Gubisch. “Optical Quality of the Human Eye.” Journal of Physiology
186 (1966), 558–578.
[152] F. W. Campbell and J. G. Robson. “Application of Fourier Analysis to the Visibility of Grat-
ings.” Journal of Physiology (London) 197 (1968), 551–566.
[153] V. C. Cardei and B. Funt. “Color Correcting Uncalibrated Digital Camera.” Journal of Imaging
Science and Technology 44:4 (2000), 288–294.
[154] B. S. Carlson. “Comparison of Modern CCD and CMOS Image Sensor Technologies and
Systems for Low Resolution Imaging.” In Proceedings of IEEE Sensors, 1, 1, pp. 171–176,
2002.
[155] J. E. Carnes and W. F. Kosonocky. “Noise Sources in Charge-Coupled Devices.” RCA Review
33 (1972), 327–343.

i i

i i
 i i

i i

Bibliography 969

[156] J. E. Carnes and W. F. Kosonocky. “Noise Sources in Charge-Coupled Devices.” In SPIE

Milestone Series 177, pp. 101–109. Bellingham, WA: SPIE, 2003.
[157] V. A. Casagrande. “A Third Parallel Pathway to the Primate Retina V1.” Trends in Neuroscience
17 (1994), 305–310.
[158] J. Cataliotti and A. Gilchrist. “Local and Global Processes in Surface Lightness Perception.”
Perception and Psychophysics 57:2 (1995), 125–135.
[159] P. Cavanagh and S. Anstis. “The Contribution of Color and Motion in Normal and Color-
Deficient Observers.” Vision Research 31:12 (1991), 2109–2148.
[160] P. Cavanagh and Y. G. Leclerc. “Shape from Shadows.” Journal of Experimental Psychology
15:1 (1989), 3–27.
[161] A. Chalmers, T. Davis, and E. Reinhard, editors. Practical Parallel Rendering. Natick, MA:
A K Peters, 2002.
[162] Y.-C. Chang and J. F. Reid. “RGB Calibration for Color Image Analysis in Machine Vision.”
IEEE Transactions on Image Processing 5:10 (1996), 1414–1422.
[163] G. Charles and G. van Loan. Matrix Computations, Third edition. Baltimore, MD: The John
Hopkins University Press, 1996.
[164] W. N. Charman. “The Flawed Retinal Image. Part I.” Optician 190:5020 (1985), 31–34.
[165] W. N. Charman. “The Flawed Retinal Image. Part II.” Optician 190:5022 (1985), 16–20.
[166] W. N. Charman. “Optics of the Eye.” In Handbook of Optics: Fundamentals, Techniques and
Design, Volume 1, edited by M. Bass, E. W. van Stryland, D. R. Williams, and W. L. Wolfe,
Second edition. New York: McGraw-Hill, 1995.
[167] B. Chen and W. Makous. “Light Capture by Human Cones.” Journal of Physiology 414 (1989),
89–109.
[168] C.-J. Chen, A. Lien, and M. I. Nathan. “4 × 4 and 2 × 2 Matrix Formulations for the Optics
in Stratified and Biaxial Media.” Journal of the Optical Society of America A 14:11 (1997),
3125–3134.
[169] Y. Chen, J. Au, P. Kazlas, A. Ritenour, H. Gates, and J. Goodman. “Ultra-Thin, High-
Resolution, Flexible Electronic Ink Displays Addressed by a-Si Active-Matrix TFT Backplanes
on Stainless Steel Foil.” In Electron Devices Meeting, IEDM ’02, pp. 389–392, 2002.
[170] W.-C. Cheng and M. Pedram. “Power Minimization in a Backlit TFT=LCD Display by Con-
current Brightness and Contrast Scaling.” IEEE Transactions on Consumer Electronics 50:1
(2004), 25–32.
[171] M. E. Chevreul. De la loi du contraste simultane des couleurs. Paris: Pitois Levreault, 1839.
[172] C.-C. Chiao, T. W. Cronin, and D. Osorio. “Color Signals in Natural Scenes: Characteristics
of Reflectance Spectra and the Effects of Natural Illuminants.” Journal of the Optical Society of
America A 17:2 (2000), 218–224.
[173] C. D. Child. “Discharge from Hot CaO.” Physical Review I 32:5 (1911), 492–511.
[174] K. Chiu, M. Herf, P. Shirley, S. Swamy, C. Wang, and K. Zimmerman. “Spatially Nonuni-
form Scaling Functions for High Contrast Images.” In Proceedings of Graphics Interface ’93,
pp. 245–253. Toronto, Canada: Canadian Information Processing Society, 1993.
[175] I. Choi, H. Shim, and N. Chang. “Low Power Color TFT-LCD Display for Hand-Held Em-
bedded Systems.” In Proceedings of the Symposium on Low Power Electronics and Design,
pp. 112–117, 2002.
[176] T. J. Cholewo and S. Love. “Gamut Boundary Determination using Alpha-Shapes.” In Pro-
ceedings of the 7th IS&T/SID Color Imaging Conference, pp. 200–204. Springfield, VA: IS&T,
1999.

i i

i i
 i i

i i

970 Bibliography

[177] B. Choubey, S. Aoyoma, S. Otim, D. Joseph, and S. Collins. “An Electronic Calibration
Scheme for Logarithmic CMOS Pixels.” IEEE Sensors Journal 6:4 (2006), 950–056.
[178] P. Choudhury and J. Tumblin. “The Trilateral Filter for High Contrast Images and Meshes.”
In EGRW ’03: Proceedings of the 14th Eurographics Workshop on Rendering, pp. 186–196.
Aire-la-Ville, Switzerland: Eurographics Association, 2003.
[179] R. E. Christoffersen. Basic Principles and Techniques of Molecular Quantum Mechanics. New
York: Springer-Verlag, 1989.
[180] C. Chubb, G. Sperling, and J. A. Solomon. “Texture Interactions Determine Perceived Con-
trast.” Proceedings of the National Academy of Sciences of the United States of America 86:23
(1989), 9631–9635.
[181] J. W. Chung, H. R. Guo, C. T. Wu, K. C. Wang, W. J. Hsieh, T. M. Wu, and C. T. Chung. “Long-
Operating Lifetime of Green Phosphorescence Top-Emitting Organic Light Emitting Devices.”
In Proceedings of the International Display Manufacturing Conference, pp. 278–280. San Jose,
CA: Society for Information Display, 2005.
[182] CIE. “CIE Proceedings 1951.” Technical Report Vol. 1, Sec 4; Vol 3, p. 37, Commision
Internationale De L’Eclairage, Vienna, 1951.
[183] CIE. “Light as a True Visual Quantity: Principles of Measurement.” Technical Report Publ.
No. 41 (TC-1.4), Commision Internationale De L’Eclairage, Vienna, 1978.
[184] CIE. “A Method For Assessing the Quality of Daylight Simulators for Colorimetry.” Technical
Report CIE 51-1981, Commision Internationale De L’Eclairage, Vienna, 1981.
[185] CIE. “The Basis of Physical Photometry.” Technical Report Publ. No. 18.2 (TC-1.2), Commi-
sion Internationale De L’Eclairage, Vienna, 1983.
[186] CIE. “Colorimetry, Second Edition.” Technical Report CIE 15.2, Commision Internationale
De L’Eclairage, Vienna, 1986.
[187] CIE. “International Lighting Vocabulary.” Technical Report 17.4, Commision Internationale
De L’Eclairage, Vienna, 1987.
[188] CIE. “Special Metamerism Index: Change in Observer.” Technical Report CIE 80-1989,
Commision Internationale De L’Eclairage, Vienna, 1989.
[189] CIE. “CIE 1988 2◦ Spectral Luminous Efficiency Function for Photopic Vision.” Technical
Report Publ. No. 86, Commision Internationale De L’Eclairage, Vienna, 1990.
[190] CIE. “Method of Measuring and Specifying Colour Rendering Properties of Light Sources.”
Technical Report CIE 13.3-1995, Commision Internationale De L’Eclairage, Vienna, 1995.
[191] CIE. “The CIE 1997 Interim Colour Appearance Model (Simple Version), CIECAM97s.”
Technical Report CIE 131-1998, Commision Internationale De L’Eclairage, Vienna, 1998.
[192] CIE. “Virtual Metamers for Assessing the Quality of Simulators of CIE Illuminant D50.” Tech-
nical Report Supplement 1-1999 to CIE 51-1981, Commision Internationale De L’Eclairage,
Vienna, 1999.
[193] CIE. “CIE Standard Illuminants for Colorimetry, Standard CIE S005/E-1998.” Technical
Report Standard CIE S005/E-1998, Commision Internationale De L’Eclairage, Vienna, 2004.
Published also as ISO 10526/CIE S 005/E-1999.
[194] CIE. “Colorimetry, Third Edition.” Technical Report CIE Publ. No. 15:2004, Commision
Internationale De L’Eclairage, Vienna, 2004.
[195] CIE. “A Colour Appearance Model for Colour Management Systems: CIECAM02.” Technical
Report CIE 159:2004, Commision Internationale De L’Eclairage, Vienna, 2004.
[196] F. R. Clapper and J. A. C. Yule. “The Effect of Multiple Internal Reflections on the Densities
of Halftone Prints on Paper.” Journal of the Optical Society of America 43 (1953), 600–603.

i i

i i
 i i

i i

Bibliography 971

[197] F. R. Clapper and J. A. C. Yule. “Reproduction of Color with Halftone Images.” In Proceedings
of the Technical Association of Graphic Arts, pp. 1–12. Chicago, IL: Technical Assn. of the
Graphic Arts, 1955.
[198] F. J. J. Clark, R. McDonald, and B. Rigg. “Modifications to the JPC79 Colour-Difference
Formula.” Journal of the Society of Dyers and Colourists 100 (1984), 128–132, (Errata: 281–
282).
[199] S. Coe, W.-K. Woo, M. Bawendi, and V. Bulovic. “Electroluminescence from Single Mono-
layers of Nanocrystals in Molecular Organic Devices.” Nature 420:6917 (2002), 800–803.
[200] M. F. Cohen and J. R. Wallace. Radiosity and Realistic Image Synthesis. Cambridge, MA:
Academic Press, Inc., 1993.
[201] J. Cohen. “Dependency of the Spectral Reflectance Curves of the Munsell Color Chips.” Psy-
chonomic Science 1 (1964), 369–370.
[202] M. Colbert, E. Reinhard, and C. E. Hughes. “Painting in High Dynamic Range.” Journal of
Visual Communication and Image Representation 18:5 (2007), 387–396.
[203] M. Collet. “Solid-State Image Sensors.” Sensors and Actuators 10 (1986), 287–302.
[204] D. Comaniciu and P. Meer. “Mean Shift: A Robust Approach toward Feature Space Analysis.”
IEEE Transactions on Pattern Analysis and Machine Intelligence 24:5 (2002), 603–619.
[205] B. Comiskey, J. D. Albert, H. Yoshizawa, and J. Jacobson. “An Electrophoretic Ink for All-
Printed Reflective Electronic Displays.” Nature 394:6690 (1998), 253–255.
[206] M. Corbetta, F. M. Miezin, S. Dobmeyer, G. L. Shulman, and S. E. Petersen. “Selective and Di-
vided Attention during Visual Discriminations of Shape, Color and Speed: Functional Anatomy
by Positron Emission Tomography.” Journal of Neuroscience 11:8 (1991), 2383–2402.
[207] T. N. Cornsweet and D. Teller. “Relation of Increment Thresholds to Brightness and Lumi-
nance.” Journal of the Optical Society of America 55 (1965), 1303–1308.
[208] T. N. Cornsweet. Visual Perception. New York: Academic Press, 1970.
[209] N. P. Cottaris and R. L. de Valois. “Temporal Dynamics of Chromatic Tuning in Macaque
Primary Visual Cortex.” Nature 395:6705 (1998), 896–900.
[210] W. B. Cowan and N. Rowell. “On the Gun Independence and Phosphor Constancy of Color
Video Monitors.” Color Research and Application 11 (1986), Supplement 34–38.
[211] W. B. Cowan. “An Inexpensive Scheme for Calibration of a Colour Monitor in Terms of CIE
Standard Coordinates.” Proc. SIGGRAPH ’83 Computer Graphics 17:3 (1983), 315–321.
[212] B. H. Crawford. “The Scotopic Visibility Function.” Proceedings of the Physical Society of
London 62:5 (1949), 321–334.
[213] G. P. Crawford, editor. Flexible Flat Panel Displays. Chichester, UK: John Wiley and Sons,
2005.
[214] W. Crookes. “The Bakerian lecture: On the Illumination of Lines of Electrical Pressure, and
the Trajectory of Molecules.” Philosophical Transactions of the Royal Society of London 170
(1879), 135–164.
[215] M. E. Crovella and M. S. Taqqu. “Estimating the Heavy Tail Index from Scaling Properties.”
Methodology and Computing in Applied Probability 1:1 (1999), 55–79.
[216] R. Cruz-Coke. Colour Blindness — An Evolutionary Approach. Springfield, IL: Charles C
Springfield, 1970.
[217] R. Cucchiara, C. Grana, M. Piccardi, A. Prati, and S. Sirotti. “Improving Shadow Suppression
in Moving Object Detection with HSV Color Information.” In IEEE Intelligent Transportation
Systems, pp. 334–339. IEEE Press, 2001.

i i

i i
 i i

i i

972 Bibliography

[218] J. A. Curcio and C. C. Petty. “The Near-Infrared Absorption Spectrum of Liquid Water.”
Journal of the Optical Society of America 41:5 (1951), 302–304.
[219] C. A. Curcio, K. R. Sloan, R. E. Kalina, and A. E. Hendrickson. “Human Photoreceptor
Topography.” Journal of Comparative Neurology 292:4 (1990), 497–523.
[220] C. A. Curcio, K. A. Allen, K. R. Sloan, C. L. Lerea, J. B. Hurley, I. B. Klock, and A. H. Milam.
“Distribution and Morphology of Human Cone Photoreceptors Stained with Anti-Blue Opsin.”
Journal of Comparative Neurology 312:4 (1991), 610–624.
[221] D. M. Dacey and B. B. Lee. “The ’Blue-On’ Opponent Pathways in Primate Retina Originates
from a Distinct Bistratified Ganglion Cell.” Nature 367:6465 (1994), 731–735.
[222] D. M. Dacey and B. B. Lee. “Cone Inputs to the Receptive Field of Midget Ganglion Cells
in the Periphery of the Macaque Retina.” Investigative Ophthalmology and Visual Science,
Supplement 38 (1997), S708.
[223] D. M. Dacey and B. B. Lee. “Functional Architecture of Cone Signal Pathways in the Primate
Retina.” In Color Vision: From Genes to Perception, edited by K. R. Gegenfurtner and L. Sharpe,
pp. 181–202. Cambridge, UK: Cambridge University Press, 1999.
[224] D. M. Dacey and M. R. Petersen. “Dendritic Field Size and Morphology of Midget and Parasol
Ganglion Cells of the Human Retina.” Proceedings of the National Academy of Sciences of the
United States of America 89:20 (1992), 9666–9670.
[225] D. M. Dacey. “Morphology of a Small-Field Bistratified Ganglion Cell Type in the Macaque
and Human Retina.” Visual Neuroscience 10 (1993), 1081–1098.
[226] D. M. Dacey. “The Mosaic of Midget Ganglion Cells in the Human Retina.” Journal of
Neuroscience 13:12 (1993), 5334–5355.
[227] D. M. Dacey. “Parallel Pathways for Spectral Coding in Primate Retina.” Annual Review of
Neuroscience 23 (2000), 743–775.
[228] S. C. Dakin and P. J. Bex. “Natural Image Statistics Mediate Brightness Filling In.” Proceed-
ings of the Royal Society of London, B 270 (2003), 2341–2348.
[229] S. Daly. “The Visible Differences Predictor: An Algorithm for the Assessment of Image Fi-
delity.” In Digital Images and Human Vision, edited by A. Watson, pp. 179–206. Cambridge,
MA: MIT Press, 1993.
[230] A. Damasio, T. Yamada, H. Damasio, J. Corbett, and J. McKnee. “Central Achromatopsia:
Behavioral, Anatomic, and Physiologic Aspects.” Neurology 30:10 (1980), 1064–1071.
[231] K. J. Dana, B. van Ginneken, S. K. Nayar, and J. J. Koenderink. “Reflectance and Texture of
Real World Surfaces.” ACM Transactions on Computer Graphics 18:1 (1999), 1–34.
[232] K. J. Dana. “BRDF/BTF Measurement Device.” In Proceedings of the Eighth IEEE Interna-
tional Conference on Computer Vision (ICCV), pp. 460–466. Washington, DC: IEEE, 2001.
[233] J. L. Dannemiller. “Spectral Reflectance of Natural Objects: How Many Basis Functions Are
Necessary?” Journal of the Optical Society of America A 9:4 (1992), 507–515.
[234] J. Daugman. “High Confidence Visual Recognition of Persons by a Test of Statistical Inde-
pendence.” IEEE Transactions on Pattern Analysis and Machine Intelligence 15:11 (1993),
1148–1161.
[235] R. Davis and K. S. Gibson. Filters for the Reproduction of Sunlight and Daylight and the
Determination of Color Temperature. Bureau of Standards, Miscellaneous. Publication 114,
Washington, DC, 1931.
[236] H. Davson. Physiology of the Eye, Fifth edition. Pergamon Press, 1990.
[237] E. A. Day, L. Taplin, and R. S. Berns. “Colorimetric Characterization of a Computer-Controlled
Liquid Crystal Display.” Color Research and Application 29:5 (2004), 365–373.

i i

i i
 i i

i i

Bibliography 973

[238] P. Debevec and J. Malik. “Recovering High Dynamic Range Radiance Maps from Pho-
tographs.” In Proceedings SIGGRAPH ’97, Computer Graphics Proceedings, Annual Confer-
ence Series, pp. 369–378. Reading, MA: Addison-Wesley, 1997.
[239] P. E. Debevec. “Rendering Synthetic Objects into Real Scenes: Bridging Traditional and
Image-Based Graphics with Illumination and High Dynamic Range Photography.” In Proceed-
ings of SIGGRAPH ’98, Computer Graphics Proceedings, Annual Conference Series, pp. 45–50.
Reading, MA: Addison-Wesley, 1998.
[240] P. E. Debevec. “A Tutorial on Image-Based Lighting.” IEEE Computer Graphics and Applica-
tions 22:2 (2002), 26–34.
[241] S. Decker, R. McGrath, K. Brehmer, and C. Sodini. “A 256 × 256 CMOS Imaging Array
with Wide Dynamic Range Pixels and Column-Parallel Digital Output.” In IEEE International
Solid-State Circuits Conference (ISSCC), pp. 176–177, 1998.
[242] C. DeCusatis, editor. Handbook of Applied Photometry. Philadelphia: American Institute of
Physics, 1997.
[243] M. F. Deering. “A Photon Accurate Model of the Human Eye.” ACM Transactions on Graphics
24:3 (2005), 649–658.
[244] P. Delahunt and D. Brainard. “Control of Chromatic Adaptation Signals from Separate Cone
Classes Interact.” Vision Research 40 (2000), 2885–2903.
[245] F. Delamare and B. Guineau. Colors: The Story of Dyes and Pigments. New York: Harry N
Abrams, 2000.
[246] E. Demichel. Le Procédé 26 (1924), 17–21.
[247] E. Demichel. Le Procédé 26 (1924), 26–27.
[248] E. J. Denton. “The Contributions of the Orientated Photosensitive and Other Molecules to
the Absorption of the Whole Retina.” Proceedings of the Royal Society of London, B 150:938
(1959), 78–94.
[249] G. Derra, H. Moensch, E. Fischer, H. Giese, U. Hechtfischer, G. Heusler, A. Koerber, U. Nie-
mann, F.-C. Noertemann, P. Pekarski, J. Pollmann-Retsch, A. Ritz, and U. Weichmann. “UHP
Lamp Systems for Projection Applications.” Journal of Physics D: Applied Physics 38 (2005),
2995–3010.
[250] A. M. Derrington, J. Krauskopf, and P. Lennie. “Chromatic Mechanisms in Lateral Geniculate
Nucleus of Macaque.” Journal of Physiology (London) 357 (1984), 241–265.
[251] K. Devlin, A. Chalmers, and E. Reinhard. “Visual Self-Calibration and Correction for Ambient
Illumination.” ACM Transactions on Applied Perception 3:4 (2006), 429–452.
[252] D. S. Dewald, S. M. Penn, and M. Davis. “Sequential Color Recapture and Dynamic Filtering:
A Method of Scrolling Color.” SID Symposium Digest of Technical Papers 32:1 (2001), 1076–
1079.
[253] J. M. DiCarlo and B. A. Wandell. “Rendering High Dynamic Range Images.” In Proceedings
of the SPIE 3965 (Electronic Imaging 2000 Conference), pp. 392–401. Bellingham, WA: SPIE,
2000.
[254] J. Dillon, R. H. Wang, and S. J. Atherton. “Photochemical and Photophysical Studies on Human
Lens Constituents.” Photochemistry and Photobiology 52 (1990), 849–854.
[255] J. Dillon. “Photochemical Mechanisms in the Lens.” In The Ocular Lens, edited by H. Maisel,
pp. 349–366. New York: Marcel Dekker, 1985.
[256] R. W. Ditchburn. Eye Movements and Visual Perception. Oxford, UK: Clarendon Press, 1973.
[257] K. R. Dobkins. “Moving Colors in the Lime Light.” Neuron 25 (2000), 15–18.

i i

i i
 i i

i i

974 Bibliography

[258] D. Doherty and G. Hewlett. “Pulse Width Modulation Control in DLP Projectors.” Texas
Instruments Technical Journal, pp. 115–121.
[259] D. W. Dong and J. J. Atick. “Statistics of Natural Time-Varying Images.” Network: Computa-
tion in Neural Systems 6:3 (1995), 345–358.
[260] R. W. Doty. “Nongeniculate Afferents to Striate Cortex in Macaques.” Journal of Comparative
Neurology 218:2 (1983), 159–173.
[261] R. F. Dougherty and A. R. Wade. Available online (http://www.vischeck.com/daltonize).
[262] J. E. Dowling. The Retina: An Approachable Part of the Brain. Cambridge, MA: Belknap
Press, 1987.
[263] F. Drago, W. L. Martens, K. Myszkowski, and H.-P. Seidel. “Perceptual Evaluation of Tone
Mapping Operators with Regard to Similarity and Preference.” Technical Report MPI-I-2002-
4-002, Max Plank Institut für Informatik, Saarbrücken, Germany, 2002.
[264] F. Drago, K. Myszkowski, T. Annen, and N. Chiba. “Adaptive Logarithmic Mapping for Dis-
playing High Contrast Scenes.” Computer Graphics Forum 22:3 (2003), 419–426.
[265] M. Drew and B. Funt. “Natural Metamers.” Computer Vision, Graphics and Image Processing
56:2 (1992), 139–151.
[266] M. S. Drew, J. Wei, and Z. N. Li. “Illumination-Invariant Image Retrieval and Video Segmen-
tation.” Pattern Recognition 32:8 (1999), 1369–1388.
[267] D. D. Dudley. Holography: A Survey. Washington, DC: Technology Utilization Office, Na-
tional Aeronautics and Space Administration, 1973.
[268] R. O. Duncan and G. M. Boynton. “Cortical Magnification within Human Primary Visual
Cortex Correlates with Acuity Thresholds.” Neuron 38:4 (2003), 659–671.
[269] A. Dür. “An Improved normalization for the Ward Reflectance Model.” journal of graphics
tools 11:1 (2006), 51–59.
[270] F. Durand and J. Dorsey. “Fast Bilateral Filtering for the Display of High-Dynamic-Range
Images.” ACM Transactions on Graphics 21:3 (2002), 257–266.
[271] P. Dutré, P. Bekaert, and K. Bala. Advanced Globel Illumination. Natick, MA: A K Peters,
2003.
[272] D. M. Eagleman. “Visual Illusions and Neurobiology.” Nature Reviews Neuroscience 2:12
(2001), 920–926.
[273] F. Ebner and M. D. Fairchild. “Constant Hue Surfaces in Color Space.” In Proceedings of the
SPIE 3300, pp. 107–117. Bellingham, WA: SPIE, 1998.
[274] F. Ebner and M. D. Fairchild. “Development and Testing of a Color Space (IPT) with Improved
Hue Uniformity.” In IS&T/SID Sixth Color Imaging Conference: Color Science, Systems and
Applications, pp. 8–13. Springfield, VA: Society for Imaging Science & Technology, 1998.
[275] M. Ebner. Color Constancy. Chichester, UK: John Wiley and Sons, Ltd., 2007.
[276] J. G. Eden. “Information Display Early in the 21st Century: Overview of Selected Emissive
Display Technologies.” Proceedings of the IEEE 94:3 (2006), 567–574.
[277] W. Ehrenstein. “Über Abwandlungen der L Hermannschen Heiligkeitserscheinung.” Zeitschrift
für Psychologie 150 (1941), 83–91.
[278] W. Ehrenstein. “Modifications of the Brightness Phenomenon of L Hermann.” In The Percep-
tion of Illusory Contours, edited by S. Petry and G. E. Meyer. New York: Springer-Verlag, 1987.
Translated by A Hogg from [277].
[279] A. Einstein. “Ist die Trägheit eines Körpers von seinem Energieinhalt abhängig?” Annalen der
Physik 18:13 (1905), 639–641.

i i

i i
 i i

i i

Bibliography 975

[280] J. H. Elder. “Are Edges Incomplete?” International Journal of Computer Vision 34:2/3 (1999),
97–122.
[281] H. D. Ellis. “Introduction to Aspects of Face Processing: Ten questions in Need of Answers.”
In Aspects of Face Processing, edited by H. Ellis, M. Jeeves, F. Newcombe, and A. Young,
pp. 3–13. Dordrecht, The Netherlands: Nijhoff, 1986.
[282] S. A. Engel, X. Zhang, and B. A. Wandell. “Color Tuning in Human Visual Cortex Measured
using Functional Magnetic Resonance Imaging.” Nature 388:6637 (1997), 68–71.
[283] P. G. Engeldrum. “Computing Color Gamuts of Ink-Jet Printing Systems.” Proceedings of the
Society for Information Display 27 (1986), 25–30.
[284] C. Enroth-Cugell and J. G. Robson. “The Contrast Sensitivity of Retinal Ganglion Cells of the
Cat.” Journal of Physiology (London) 187 (1966), 517–552.
[285] C. J. Erkerlers, J. van der Steen, R. M. Steinman, and H. Collewijn. “Ocular Vergence under
Natural Conditions I: Continuous Changes of Target Distance along the Median Plane.” Pro-
ceedings of the Royal Society of London B326 (1989), 417–440.
[286] C. J. Erkerlers, R. M. Steinman, and H. Collewijn. “Ocular Vergence under Natural Conditions
II: Gaze Shifts between Real Targets Differing in Distance and Direction.” Proceedings of the
Royal Society of London B326 (1989), 441–465.
[287] D. C. van Essen and J. L. Gallant. “Neural Mechanisms of Form and Motion Processing in the
Primate Visual System.” Neuron 13:1 (1994), 1–10.
[288] R. Evans. The Perception of Color. New York: John Wiley & Sons, 1974.
[289] G. L. Fain, H. R. Matthews, M. C. Cornwall, and Y. Koutalos. “Adaptation in Vertebrate
Photoreceptors.” Physiological Review 81 (2001), 117–151.
[290] M. D. Fairchild and G. M. Johnson. “On Contrast Sensitivity in an Image Difference Model.”
In IS&T PICS Conference, pp. 18–23. Springfield, VA: Society for Imaging Science and Tech-
nology, 2001.
[291] M. D. Fairchild and G. Johnson. “Meet iCAM: A Next-Generation Color Appearance Model.”
In Proceedings of the IS&T/SID 10th Color Imaging Conference, pp. 33–38. Springfield, VA:
Society for Imaging Science and Technology, 2002.
[292] M. D. Fairchild and G. M. Johnson. “The iCAM Framework for Image Appearance, Image
Differences, and Image Quality.” Journal of Electronic Imaging 13 (2004), 126–138.
[293] M. D. Fairchild and G. M. Johnson. “METACOW: A Public-Domain, High-Resolution, Fully-
Digital, Extended-Dynamic-Range, Spectral Test Target for Imaging System Analysis and Sim-
ulation.” In IS&T/SID 12th Color Imaging Conference, pp. 239–245. Springfield, VA: Society
for Imaging Science and Technology, 2004.
[294] M. D. Fairchild and G. M. Johnson. “On the Salience of Novel Stimuli: Adaptation and Image
Noise.” In IS&T/SID 13th Color Imaging Conference, pp. 333–338. Springfield, VA: Society
for Imaging Science and Technology, 2005.
[295] M. D. Fairchild and E. Pirrotta. “Predicting the Lightness of Chromatic Object Colors using
CIELAB.” Color Research and Application 16 (1991), 385–393.
[296] M. D. Fairchild and L. Reniff. “Time-Course of Chromatic Adaptation for Color-Appearance
Judgements.” Journal of the Optical Society of America 12 (1995), 824–833.
[297] M. D. Fairchild and D. R. Wyble. “Colorimetric Characterization of the Apple Studio Display
(Flat Panel LCD).” Technical report, Munsell Color Science Laboratory, Rochester, NY, 1998.
[298] M. D. Fairchild, E. Pirrotta, and T. G. Kim. “Successive-Ganzfeld Haploscopic Viewing Tech-
nique for Color-Appearance Research.” Color Research and Application 19 (1994), 214–221.

i i

i i
 i i

i i

976 Bibliography

[299] M. D. Fairchild. “Formulation and Testing of an Incomplete-Chromatic-Adaptation Model.”

Color Research and Application 16 (1991), 243–250.
[300] M. D. Fairchild. “A Model of Incomplete Chromatic Adaptation.” In Proceedings 22nd Session
of the CIE, pp. 33–34. Vienna: Commission Internationale de l’Eclairage, 1991.
[301] M. D. Fairchild. “A Revision of CIECAM97s for Practical Applications.” Color Research and
Application 26 (2001), 418–427.
[302] M. D. Fairchild. Color Appearance Models, Second edition. Chichester, UK: John Wiley and
Sons, Ltd., 2005.
[303] M. D. Fairchild. “The HDR Photographic Survey.” In Proceedings of the 15th Color Imaging
Conference: Color Science and Engineering Systems, Technologies, and Applications, pp. 233–
238. Albuquerque, NM, 2007.
[304] H. Fairman, M. Brill, and H. Hemmendinger. “How the CIE 1931 Color-Matching Functions
Were Derived from Wright-Guild Data.” Color Research and Application 22 (1997), 11–23.
[305] R. J. Farrell and J. M. Booth. Design Handbook for Imagery Interpretation Equipment, Section
3.2. Seattle: Boeing Aerospace Co., 1984.
[306] R. Fattal, D. Lischinski, and M. Werman. “Gradient Domain High Dynamic Range Compres-
sion.” ACM Transactions on Graphics 21:3 (2002), 249–256.
[307] R. Fedkiw, J. Stam, and H. W. Jensen. “Visual Simulation of Smoke.” In Proceedings of SIG-
GRAPH 2001, Computer Graphics Proceedings, Annual Conference Series, edited by E. Fiume,
pp. 15–22. Reading, MA: Addison-Wesley, 2001.
[308] D. J. Felleman and D. C. van Essen. “Receptive Field Properties of Neurons in Area V3 of
Macaque Monkey Extrastriate Cortex.” Journal of Neurophysiology 57:4 (1987), 889–920.
[309] R. Fergus, B. Singh, A. Hertzmann, S. T. Roweis, and W. T. Freeman. “Removing Camera
Shake from a Single Photograph.” ACM Transactions on Graphics 25:3 (2006), 787–794.
[310] J. A. Ferwerda, S. Pattanaik, P. Shirley, and D. P. Greenberg. “A Model of Visual Adapta-
tion for Realistic Image Synthesis.” In Proceedings of SIGGRAPH 96, Computer Graphics
Proceedings, Annual Conference Series, edited by H. Rushmeier, pp. 249–258. Reading, MA:
Addison-Wesley, 1996.
[311] J. A. Ferwerda, S. N. Pattanaik, P. Shirley, and D. P. Greenberg. “A Model of Visual Masking
for Computer Graphics.” In Proceedings of SIGGRAPH ’97, Computer Graphics Proceedings,
Annual Conference Series, pp. 143–152. Reading, MA: Addison-Wesley, 1997.
[312] J. A. Ferwerda. “Elements of Early Vision for Computer Graphics.” IEEE Computer Graphics
and Applications 21:5 (2001), 22–33.
[313] D. J. Field and N. Brady. “Visual Sensitivity, Blur and the Sources of Variability in the Ampli-
tude Spectra of Natural Scenes.” Vision Research 37:23 (1997), 3367–3383.
[314] D. J. Field. “Relations between the Statistics of Natural Images and the Response Properties of
Cortical Cells.” Journal of the Optical Society of America A 4:12 (1987), 2379–2394.
[315] D. J. Field. “Scale-Invariance and Self-Similar ‘Wavelet’ Transforms: An Analysis of Natural
Scenes and Mammalian Visual Systems.” In Wavelets, Fractals and Fourier Transforms, edited
by M. Farge, J. C. R. Hunt, and J. C. Vassilicos, pp. 151–193. Oxford, UK: Clarendon Press,
1993.
[316] E. F. Fincham. “Defects of the Colour-Sense Mechanism as Indicated by the Accommodation
Reflect.” Journal of Physiology 121 (1953), 570–580.
[317] G. D. Finlayson and S. D. Hordley. “Color Constancy at a Pixel.” Journal of the Optical Society
of America A 18:2 (2001), 253–264.

i i

i i
 i i

i i

Bibliography 977

[318] G. D. Finlayson and S. Süsstrunk. “Color Ratios and Chromatic Adaptation.” In Proceedings
of IS&T CGIV, First European Conference on Color Graphics, Imaging and Vision, pp. 7–10.
Springfield, VA: Society for Imaging Science and Technology, 2002.
[319] G. D. Finlayson, S. D. Hordley, and P. M. Hubel. “Color by Correlation: A Simple, Unify-
ing Framework for Color Constancy.” IEEE Transactions on Patterns Analysis and Machine
Intelligence 23:11 (2001), 1209–1221.
[320] G. D. Finlayson, S. D. Hordley, and M. S. Drew. “Removing Shadows from Images using
Retinex.” In Proceedings IS&T Color Imaging Conference, pp. 73–79. Springfield, VA: Society
for Imaging Science and Technology, 2002.
[321] A. Fiorentini and A. M. Ercoles. “Vision of Oscillating Visual Fields.” Optica Acta 4 (1957),
370.
[322] D. Fitzpatrick, K. Itoh, and I. T. Diamond. “The Laminar Organization of the Lateral Geniculate
Body and the Striate Cortex in the Squirrel Monkey (Saimiri sciureus).” Journal of Neuroscience
3:4 (1983), 673–702.
[323] D. Fitzpatrick, W. M. Usrey, B. R. Schofield, and G. Einstein. “The Sublaminar Organization
of Corticogeniculate Neurons in Layer 6 of Macaque Striate Cortex.” Visual Neuroscience 11
(1994), 307–315.
[324] P. D. Floyd, D. Heald, B. Arbuckle, A. Lewis, M. Kothari, B. J. Gally, B. Cummings, B. R.
Natarajan, L. Palmateer, J. Bos, D. Chang, J. Chiang, D. Chu, L.-M. Wang, E. Pao, F. Su,
V. Huang, W.-J. Lin, W.-C. Tang, J.-J. Yeh, C.-C. Chan, F.-A. Shu, and Y.-D. Ju. “IMOD
Display Manufacturing.” SID Symposium Digest of Technical Papers 37:1 (2006), 1980–1983.
[325] J. M. Foley and M. E. McCourt. “Visual Grating Induction.” Journal of the Optical Society of
America A 2 (1985), 1220–1230.
[326] J. Foley, A. van Dam, S. Feiner, and J. Hughes. Computer Graphics Principles and Practice,
Second edition. Reading, MA: Addison-Wesley, 1990.
[327] A. Ford and A. Roberts. “Colour Space Conversions.”, 1998. Available online (http://www.
poynton.com/PFDs/coloureq.pdf).
[328] S. Forrest, P. Burrows, and M. Thompson. “The Dawn of Organic Electronics.” IEEE Spectrum
37:8 (2000), 29–34.
[329] J. Forrester, A. Dick, P. McMenamin, and W. Lee. The Eye: Basic Sciences in Practice.
London: W. B. Saunders Company Ltd., 2001.
[330] E. R. Fossum. “CMOS Image Sensors: Electronics Camera-On-A-Chip.” IEEE Transactions
on Electron Devices 44:10 (1997), 1689–1698.
[331] D. H. Foster and S. M. C. Nascimento. “Relational Color Constancy from Invariant Cone-
Excitation Ratios.” Proceedings of the Royal Society: Biological Sciences 257:1349 (1994),
115–121.
[332] J. Frankle and J. McCann. “Method and Apparatus for Lightness Imaging.”, 1983. U.S. patent
#4,384,336.
[333] W. Fries and H. Distel. “Large Layer V1 Neurons of Monkey Striate Cortex (Meynert Cells)
Project to the Superior Colliculus.” Proceedings of the Royal Society of London B, Biological
Sciences 219 (1983), 53–59.
[334] W. Fries. “Pontine Projection from Striate and Prestriate Visual Cortex in the Macaque Mon-
key: An Anterograde Study.” Visual Neuroscience 4 (1990), 205–216.
[335] K. Fritsche. Faults in Photography. London: Focal Press, 1968.
[336] G. A. Fry and M. Alpern. “The Effect of a Peripheral Glare Source upon the Apparent Bright-
ness of an Object.” Journal of the Optical Society of America 43:3 (1953), 189–195.

i i

i i
 i i

i i

978 Bibliography

[337] G. D. Funka-Lea. “The Visual Recognition of Shadows by an Active Observer.” Ph.D. thesis,
University of Pennsylvania, 1994.
[338] B. V. Funt and G. Finlayson. “Color Constant Color Indexing.” IEEE Transactions on Pattern
Analysis and Machine Intelligence 17 (1995), 522–529.
[339] B. V. Funt, M. S. Drew, and J. Ho. “Color Constancy from Mutual Reflection.” International
Journal of Computer Vision 6 (1991), 5–24.
[340] B. Funt, F. Ciurea, and J. McCann. “Retinex in Matlab.” In Proceedings of the IS&T/SID Eighth
Color Imaging Conference: Color Science, Systems and Applications, pp. 112–121. Springfield,
VA: Society for Imaging Science and Technology, 2000.
[341] M. Furuta, Y. Nishikawa, T. Inoue, and S. Kawahito. “A High-Speed, High-Sensitivity Digital
CMOS Image Sensor with a Global Shutter and 12-Bit Column-Parallel Cyclic A/D Converters.”
IEEE Journal of Solid-State Circuits 42:4 (2007), 766–774.
[342] D. Gabor. “A New Microscopic Principle.” Nature 161 (1948), 777–778.
[343] D. Gabor. “Wavefront Reconstruction.” In Optical and Acoustical Holography, edited by
E. Camatini, pp. 15–21. New York: Plenum Press, 1972.
[344] E. R. Gaillard, L. Zheng, J. C. Merriam, and J. Dillon. “Age-Related Changes in the Absorption
Characteristics of the Primate Lens.” Investigative Ophthalmology and Visual Science 41:6
(2000), 1454–1459.
[345] L. Gall. “Computer Color Matching.” In Colour 73: Second Congress International Colour
Association, pp. 153–178. London: Adam Hilger, 1973.
[346] L. Ganglof, E. Minoux, K. B. K. Teo, P. Vincent, V. T. Semet, V. T. Binh, M. H. Yang, I. Y. Y.
Bu, R. G. Lacerda, G. Pirio, J. P. Schnell, D. Pribat, D. G. Hasko, G. A. J. Amaratunga, W. Milne,
and P. Legagneux. “Self-Aligned, Gated Arrays of Individual Nanotube and Nanowire Emitters.”
Nano Letters 4 (2004), 1575–1579.
[347] M. J. Gastinger, J. J. O’Brien, J. N. B. Larsen, and D. W. Marshak. “Histamine Immunoreactive
Axons in the Macaque Retina.” Investigative Ophthalmology and Visual Science 40:2 (1999),
487–495.
[348] H. Gates, R. Zehner, H. Doshi, and J. Au. “A5 Sized Electronic Paper Display for Document
Viewing.” SID Symposium Digest of Technical Papers 36:1 (2005), 1214–1217.
[349] F. Gatti, A. Acquaviva, L. Benini, and B. Ricco. “Low Power Control Techniques for TFT-
LCD Displays.” In Proceedings of the International Conference on Compilers, Architecture,
and Synthesis for Embedded Systems, pp. 218–224, 2002.
[350] I. Gauthier, M. J. Tarr, A. W. Anderson, P. Skudlarski, and J. C. Gore. “Activation of the
Middle Fusiform ’Face Area’ Increases with Expertise in Recognizing Novel Objects.” Nature
Neuroscience 2:6 (1999), 568–573.
[351] W. J. Geeraets and E. R. Berry. “Ocular Spectral Characteristics as Related to Hazards from
Lasers and Other Sources.” American Journal of Ophthalmology 66:1 (1968), 15–20.
[352] K. Gegenfurtner and D. C. Kiper. “Color Vision.” Annual Review of Neuroscience 26 (2003),
181–206.
[353] K. R. Gegenfurtner and L. Sharpe, editors. Color Vision: From Genes to Perception. Cam-
bridge, UK: Cambridge University Press, 1999.
[354] J. Geier, L. Sera, and L. Bernath. “Stopping the Hermann Grid Illusion by Simple Sine Distor-
tion.” Perception, ECVP 2004 supplement 33 (2004), 53.
[355] W. S. Geisler and M. S. Banks. “Visual Performance.” In Handbook of Optics: Fundamentals,
Techniques and Design, Volume 1, edited by M. Bass, E. W. van Stryland, D. R. Williams, and
W. L. Wolfe, Second edition. New York: McGraw-Hill, 1995.

i i

i i
 i i

i i

Bibliography 979

[356] N. George, R. J. Dolan, G. R. Fink, G. C. Baylis, C. Russell, and J. Driver. “Contrast Polarity
and Face Recognition in the Human Fusiform Gyrus.” Nature Neuroscience 2:6 (1999), 574–
580.
[357] T. Gevers and H. Stokman. “Classifying Color Edges in Video into Shadow-Geometry, High-
light, or Material Transitions.” IEEE Transactions on Multimedia 5:2 (2003), 237–243.
[358] K. S. Gibson. “The Relative Visibility Function.” CIE Compte Rendu des Séances, Sixième
Session, Genève, 1924.
[359] J. E. Gibson and M. D. Fairchild. “Colorimetric Characterization of Three Computer Displays
(LCD and CRT).” Technical report, Munsell Color Science Laboratory, Rochester, NY, 2000.
[360] K. S. Gibson and E. P. T. Tyndall. “Visibility of Radiant Energy.” Scientific Papers of the
Bureau of Standards 19 (1923), 131–191.
[361] I. M. Gibson. “Visual Mechanisms in a Cone Monochromat.” Journal of Physiology 161
(1962), 10–11.
[362] A. Gilchrist and J. Cataliotti. “Anchoring of Surface Lightness with Multiple Illumination
Levels.” Investigative Ophthalmology and Visual Science 35:4 (1994), 2165–2165.
[363] A. Gilchrist, C. Kossyfidis, F. Bonato, T. Agostini, J. Cataliotti, X. Li, B. Spehar, V. Annan, and
E. Economou. “An Anchoring Theory of Lightness Perception.” Psychological Review 106:4
(1999), 795–834.
[364] A. L. Gilchrist. “Lightness Contrast and Failures of Constancy: A Common Explanation.”
Perception and Psychophysics 43:5 (1988), 415–424.
[365] A. L. Gilchrist. “The Importance of Errors in Perception.” In Colour Perception: Mind and the
Physical World, edited by R. M. D. Heyer, pp. 437–452. Oxford, UK: Oxford University Press,
2003.
[366] A. L. Gilchrist. “Lightness Perception: Seeing One Color through Another.” Current Biology
15:9 (2005), R330–R332.
[367] A. Glasser and M. C. W. Campbell. “Presbyopia and the Optical Changes in the Human Crys-
talline Lens with Age.” Vision Research 38 (1998), 209–229.
[368] J. Glasser. “Principles of Display Measurement and Calibration.” In Display Systems: Design
and Applications, edited by L. W. MacDonald and A. C. Lowe. Chichester: John Wiley and
Sons, 1997.
[369] A. S. Glassner. “How to Derive a Spectrum from an RGB Triplet.” IEEE Computer Graphics
and Applications 9:4 (1989), 95–99.
[370] A. S. Glassner. Principles of Digital Image Synthesis. San Fransisco, CA: Morgan Kaufmann,
1995.
[371] A. Glassner. “Computer-Generated Solar Halos and Sun Dogs.” IEEE Computer Graphics and
Applications 16:2 (1996), 77–81.
[372] A. Glassner. “Solar Halos and Sun Dogs.” IEEE Computer Graphics and Applications 16:1
(1996), 83–87.
[373] J. W. von Goethe. Zur Farbenlehre. Tübingen, Germany: Freies Geistesleben, 1810.
[374] N. Goldberg. Camera Technology: The Dark Side of the Lens. Boston: Academic Press, Inc.,
1992.
[375] S. Gong, J. Kanicki, G. Xu, and J. Z. Z. Zhong. “A Novel Structure to Improve the Viewing An-
gle Characteristics of Twisted-Nematic Liquid Crystal Displays.” Japanese Journal of Applied
Physics 48 (1999), 4110–4116.
[376] R. C. Gonzales and R. E. Woods. Digital Image Processing, Second edition. Upper Saddle
River, NJ: Prentice-Hall, 2002.

i i

i i
 i i

i i

980 Bibliography

[377] J.-C. Gonzato and S. Marchand. “Photo-Realistic Simulation and Rendering of Halos.” In Win-
ter School of Computer Graphics (WSCG ’01) Proceedings. Plzen, Czech Republic: University
of West Bohemia, 2001.
[378] A. Gooch, S. C. Olsen, J. Tumblin, and B. Gooch. “Color2Gray: Salience-Preserving Color
Removal.” ACM Transactions on Graphics 24:3 (2005), 634–639.
[379] R. H. Good, Jr and T. J. Nelson. Classical Theory of Electric and Magnetic Fields. New York:
Academic Press, 1971.
[380] C. M. Goral, K. E. Torrance, D. P. Greenberg, and B. Battaile. “Modeling the Interaction
of Light Between Diffuse Surfaces.” Computer Graphics (SIGGRAPH ’84 Proceedings) 18:3
(1984), 213–222.
[381] S. J. Gortler, R. Grzeszczuk, R. Szeliski, and M. F. Cohen. “The Lumigraph.” In Proceed-
ings SIGGRAPH ’96, Computer Graphics Proceedings, Annual Conference Series, pp. 43 –54.
Reading, MA: Addison-Wesley, 1996.
[382] P. Gouras. “Identification of Cone Mechanisms in Monkey Ganglion Cells.” Journal of Physi-
ology (London) 199 (1968), 533–547.
[383] C. F. O. Graeff, G. B. Silva, F. Nüesch, and L. Zuppiroli. “Transport and Recombination in Or-
ganic Light-Emitting Diodes Studied by Electrically Detected Magnetic Resonance.” European
Physical Journal E 18 (2005), 21–28.
[384] J. Graham. “Some Topographical Connections of the Striate Cortex with Subcortical Structures
in Macaca Fascicularis.” Experimental Brain Research 47 (1982), 1–14.
[385] T. Granlund, L. A. A. Petterson, M. R. Anderson, and O. Inganäs. “Interference Phenomenon
Determines the Color in an Organic Light Emitting Diode.” Journal of Applied Physics 81:12
(1997), 8097–8104.
[386] H. Grassmann. “Zur Theorie der Farbernmischung.” Ann. Phys. Chem. 89 (1853), 69–84.
[387] R. Greenler. Rainbows, Halos, and Glories. Cambridge, UK: Cambridge University Press,
1980.
[388] R. L. Gregory and P. F. Heard. “Border Locking and the Café Wall Illusion.” Perception 8:4
(1979), 365–380.
[389] R. L. Gregory and P. F. Heard. “Visual Dissociations of Movement, Position, and Stereo Depth:
Some Phenomenal Phenomena.” Quarterly Journal of Experimental Psychology 35A (1983),
217–237.
[390] J. Greivenkamp. “Color Dependent Optical Filter for the Suppression of Aliasing Artifacts.”
Applied Optics 29:5 (1990), 676–684.
[391] L. D. Griffin. “Optimality of the Basic Color Categories for Classification.” Journal of the
Royal Society, Interface 3:6 (2006), 71–85.
[392] K. Grill-Spector and R. Malach. “The Human Visual Cortex.” Annual Review of Neuroscience
27 (2004), 647–677.
[393] K. Grill-Spector, T. Kushnir, S. Edelman, Y. Itzchak, and R. Malach. “Cue-Invariant Activation
in Object-Related Areas of the Human Occipital Lobe.” Neuron 21:1 (1998), 191–202.
[394] K. Grill-Spector, T. Kushnir, T. Hendler, S. Edelman, Y. Itzchak, and R. Malach. “A Sequence
of Object-Processing Stages Revealed by fMRI in the Human Occipital Lobe.” Human Brain
Mapping 6:4 (1998), 316–328.
[395] S. Grossberg and E. Mingolla. “Neural Dynamics of Form Perception: Boundary Adaptation,
Illusory Figures, and Neon Color Spreading.” Psychological Review 92 (1985), 173–211.
[396] S. Grossberg and J. Todorović. “Neural Dynamics of 1-D and 2-D Brightness Perception: A
Unified Model of Classical and Recent Phenomena.” Perception and Psychophysics 43 (1988),
241–277.

i i

i i
 i i

i i

Bibliography 981

[397] F. Grum and R. J. Becherer. Optical Radiation Measurements. New York: Academic Press,
Inc., 1979.
[398] C. Gu and P. Yeh. “Extended Jones Matrix Method. II.” Journal of the Optical Society of
America A 10:5 (1993), 966–973.
[399] G. Gu, V. Boluvić, P. E. Burrows, S. R. Forrest, and M. E. Thompson. “Transparent Organic
Light Emitting Devices.” Applied Physics Letters 68:19 (1996), 2606–2608.
[400] G. Gu, P. E. Burrows, S. Venkatesh, S. R. Forrest, and M. E. Thompson. “Vacuum-Deposited,
Nonpolymeric Flexible Organic Light-Emitting Devices.” Optics Letters 22:3 (1997), 172–174.
[401] G. Gu, G. Parthasarathy, P. E. Burrows, P. Tian, I. G. Hill, A. Kahn, and S. R. Forrest. “Trans-
parent Stacked Organic Light Emitting Devices. I. Design Principles and Transparent Compound
Electrodes.” Journal of Applied Physics 86:8 (1999), 4067–4075.
[402] R. W. Gubbisch. “Optical Performance of the Eye.” Journal of Physiology 186 (1966), 558–
578.
[403] C. Gueymard, D. Myers, and K. Emery. “Proposed Reference Irradiance Spectra for Solar
Energy Systems Testing.” Solar Energy 73:6 (2002), 443–467.
[404] C. Gueymard. “The Sun’s Total and Spectral Irradiance for Solar Energy Applications and
Solar Radiation Models.” Solar Energy 76:4 (2004), 423–453.
[405] J. Guild. “The Colorimetric Properties of the Spectrum.” Phil. Trans. Roy. Soc. A 230 (1931),
149–187.
[406] R. W. Guillery and M. Colonnier. “Synaptic Patterns in the Dorsal Lateral Geniculate Nu-
cleus of Cat and Monkey: A Brief Review.” Zeitschrift für Zellforschung und mikroskopische
Anatomie 103 (1970), 90–108.
[407] B. Gulyas, C. A. Heywood, D. A. Popplewell, P. E. Roland, and A. Cowey. “Visual Form
Discrimination from Color or Motion Cues: Functional Anatomy by Positron Emission Tomog-
raphy.” Proceedings of the National Academy of Sciences of the United States of America 91:21
(1994), 9965–9969.
[408] M. Gur, I. Kagan, and D. M. Snodderly. “Orientation and Direction Selectivity of Neurons in
V1 of Alert Monkeys: Functional Relationships and Laminar Distributions.” Cerebral Cortex
15:8 (2005), 1207–1221.
[409] C. Gutierrez and C. G. Cusick. “Area V1 in Macaque Monkeys Projects to Multiple Histo-
chemically Defined Subdivisions of the Inferior Pulvinar Complex.” Brain Research 765 (1997),
349–356.
[410] D. Gutierrez, F. J. Seron, A. Munoz, and O. Anson. “Chasing the Green Flash: A Global
Illumination Solution for Inhomogeneous Media.” In Proceedings of the Spring Conference on
Computer Graphics, pp. 95–103. New York: ACM Press, 2004.
[411] D. Gutierrez, A. Munoz, O. Anson, and F. J. Seron. “Non-Linear Volume Photon Mapping.”
In Rendering Techniques 2005: Eurographics Symposium on Rendering, edited by K. Bala and
P. Dutré, pp. 291–300. Aire-la-Ville, Switzerland: Eurographics Association, 2005.
[412] C. S. Haase and G. W. Meyer. “Modeling Pigmented Materials for Realistic Image Synthesis.”
ACM Transactions on Graphics 11:4 (1992), 305–335.
[413] J. Haber, M. Magnor, and H.-P. Seidel. “Physically-Based Simulation of Twilight Phenomena.”
ACM Transactions on Graphics 24:4 (2005), 1353–1373.
[414] R. A. Hall and D. P. Greenberg. “A Testbed for Realistic Image Synthesis.” IEEE Computer
Graphics and Applications 3:6 (1983), 10–20.
[415] R. A. Hall. Illumination and Color in Computer Generated Imagery. Monographs in Visual
Communication, New York: Springer-Verlag, 1989.

i i

i i
 i i

i i

982 Bibliography

[416] R. A. Hall. “Comparing Spectral Color Computation Methods.” IEEE Computer Graphics and
Applications 19:4 (1999), 36–45.
[417] P. E. Hallet. “The Variations in Visual Threshold Measurement.” Journal of Physiology 202
(1969), 403–419.
[418] M. Halstead, B. Barsky, S. Klein, and R. Mandell. “Reconstructing Curved Surfaces from
Specular Reflection Patterns using Spline Surface Fitting of Normals.” In Proceedings of SIG-
GRAPH 96, Computer Graphics Proceedings, Annual Conference Series, edited by H. Rush-
meier, pp. 335–342. Reading, MA: Addison-Wesley, 1996.
[419] A. Hanazawa, H. Komatsu, and I. Murakami. “Neural Selectivity for Hue and Saturation of
Colour in the Primary Visual Cortex of the Monkey.” European Journal of Neuroscience 12:5
(2000), 1753–1763.
[420] P. J. B. Hancock, R. J. Baddeley, and L. S. Smith. “The Principle Components of Natural
Images.” Network 3 (1992), 61–70.
[421] J. A. Hanley and B. J. McNeil. “The Meaning and Use of the Area under the Receiver Operating
Characteristic (ROC) Curve.” Radiology 143:1 (1982), 29–36.
[422] J. H. Hannay. “The Clausius-Mossotti Equation: An Alternative Derivation.” European Journal
of Physics 4 (1983), 141–143.
[423] A. Hård and L. Sivik. “NCS — Natural Color System: A Swedish Standard for Color Nota-
tion.” Colour Research and Application 6 (1981), 129–138.
[424] A. Hård, L. Sivik, and G. Tonnquist. “NCS, Natural Color System — From Concepts to
Research and Applications. Part I.” Colour Research and Application 21 (1996), 180–205.
[425] A. Hård, L. Sivik, and G. Tonnquist. “NCS, Natural Color System — From Concepts to
Research and Applications. Part II.” Colour Research and Application 21 (1996), 206–220.
[426] A. C. Hardy and F. L. Wurzburg, Jr. “Color Correction in Color Printing.” Journal of the
Optical Society of America 38:4 (1948), 300–307.
[427] P. Hariharan. Basics of Holography. Cambridge, UK: Cambridge University Press, 2002.
[428] H. F. Harmuth, R. N. Boules, and M. G. M. Hussain. Electromagnetic Dignals: Reflection,
Focussing, Distortion, and their Practical Applications. New York: Kluwer Academic / Plenum
Publishers, 1999.
[429] F. J. Harris. “On the Use of Windows for Harmonic Analysis with the Discrete Fourier Trans-
form.” Proceedings of the IEEE 66:1 (1978), 51–84.
[430] H. K. Hartline. “The Response of Single Optic Nerve Fibers of the Vertebrate Eye to Illumina-
tion of the Retina.” American Journal of Physiology 121 (1938), 400–415.
[431] E. I. Haskal, M. Buechel, J. F. Dijksman, P. C. Duineveld, E. A. Meulenkamp, C. A. H. A.
Mutsaers, A. Sempel, P. Snijder, S. I. E. Vulto, P. van de Weijer, and S. H. P. M. de Winter.
“Ink jet Printing of Passive Matrix Polymer Light-Emitting Displays.” SID Symposium Digest
of Technical Papers 33:1 (2002), 776–779.
[432] U. Hasson, M. Harel, I. Levy, and R. Malach. “Large-Scale Mirror-Symmetry Organization of
Human Occipito-Temporal Object Areas.” Neuron 37:6 (2003), 1027–1041.
[433] J. A. van Hateren and A. van der Schaaf. “Independent Component Filters of Natural Images
Compared with Simple Cells in Primary Visual Cortex.” Proceedings of the Royal Society of
London B 265 (1998), 359–366.
[434] J. H. van Hateren. “A Cellular and Molecular Model of Response Kinetics and Adaptation in
Primate Cones and Horizontal Cells.” Journal of Vision 5 (2005), 331–347.
[435] J. H. van Hateren. “Encoding of High Dynamic Range Video with a Model of Human Cones.”
ACM Transactions on Graphics 25:4 (2006), 1380–1399.

i i

i i
 i i

i i

Bibliography 983

[436] W. Hauser. Introduction to the Principles of Electromagnetism. Reading: Addison-Wesley,

1971.
[437] V. Havran. “Heuristic Ray Shooting Algorithms.” Ph.D. thesis, Department of Computer Sci-
ence and Engineering, Faculty of Electrical Engineering, Czech Technical University in Prague,
2000.
[438] M. J. Hawken, A. J. Parker, and J. S. Lund. “Laminar Organization and Contrast Sensitiv-
ity of Direction-Selective Cells in the Striate Cortex of the Old World Monkey.” Journal of
Neuroscience 8:10 (1988), 3541–3548.
[439] G. E. Healey and R. Kondepudy. “Radiometric CCD Camera Calibration and Noise Estima-
tion.” IEEE Transactions on Pattern Analysis and Machine Intelligence 16:3 (1994), 267–276.
[440] G. Healey and D. Slater. “Global Color Constancy: Recognition of Objects by Use of
Illumination-Invariant Properties of Color Distributions.” Journal of the Optical Society of Amer-
ica A 11:11 (1994), 3003–3010.
[441] G. Healey. “Estimating Spectral Reflectance using Highlights.” Image Vision Computing 9:5
(1991), 333–337.
[442] G. Healey. “Segmenting Images using Normalized Color.” IEEE Transactions on Systems,
Man, and Cybernetics 22 (1992), 64–73.
[443] S. Hecht and J. Mandelbaum. “The Relation between Vitamin A and Dark Adaptation.” Journal
of the American Medical Association 112 (1939), 1910–1916.
[444] S. Hecht, C. Haig, and G. Wald. “The Dark Adaptation of Retinal Fields of Different Size and
Location.” Journal of General Physiology 19:2 (1935), 321–327.
[445] S. Hecht, C. Haig, and A. M. Chase. “The Influence of Light Adaptation on Subsequent Dark
Adaptation of the Eye.” Journal of General Physiology 20:6 (1937), 831–850.
[446] S. Hecht, S. Schlaer, and M. H. Pirenne. “Energy, Quanta and Vision.” Journal of General
Physiology 25 (1942), 819–840.
[447] E. Hecht. Optics, Fourth edition. San Francisco: Addison-Wesley, 2002.
[448] R. Heckaman. “Talking about Color...Brilliance.” Color Research and Application 30 (2005),
250–151.
[449] E. R. Heider and D. C. Olivier. “The Structure of the Color Space in Naming and Memory for
Two Languages.” Cognitive Psychology 3 (1972), 337–354.
[450] E. Heider. “Universals in Color Naming and Memory.” Journal of Experimental Psychology
93:1 (1972), 10–20.
[451] W. Helfrich and W. G. Schneider. “Recombination Radiation in Anthracene Crystals.” Physical
Review Letters 14:7 (1965), 229–231.
[452] W. Helfrich and W. G. Schneider. “Transients of vVolume-cControlled Current and of Recom-
bination Radiation in Anthracene.” The Journal of Chemical Physics 44:8 (1966), 2902–2909.
[453] H. L. F. von Helmholtz. Handbuch der physiologischen Optik. Leipzig, Germany: Leopold
Voss, 1867.
[454] H. Helson and D. B. Judd. “An Experimental and Theoretical Study of Changes in Surface
Colors under Changing Illuminations.” Psychological Bulletin 33 (1936), 740–741.
[455] H. Helson. “Fundamental Problmes in Color vision I. The Principle Governing Changes in
Hue, Saturation, and Lightness of Non-Selective Samples in Chromatic Illuminantion.” Journal
of Experimental Psychology 23 (1938), 439–477.
[456] H. Helson. Adaptation-Level Theory. New York: Harper and Row, 1964.
[457] S. H. C. Hendry and D. J. Calkins. “Neuronal Chemistry and Functional Organization in the
Primate Visual System.” Trends in Neuroscience 21:8 (1998), 344–349.

i i

i i
 i i

i i

984 Bibliography

[458] S. H. C. Hendry and R. C. Reid. “The Koniocellular Pathway in Primate Vision.” Annual
Review of Neuroscience 23 (2000), 127–153.
[459] S. H. Hendry and T. Yoshioka. “A Neurochemically Distinct Third Channel in the Macaque
Dorsal Lateral Geniculate Nucleus.” Science 264 (1994), 575–577.
[460] L. Henyey and J. Greenstein. “Diffuse Radiation in the Galaxy.” Astrophysics Journal 93
(1941), 70–83.
[461] E. Hering. “Zur Lehre vom Lichtsinne. IV. Über die sogenannte Intensität der Lichtempfindung
und über die Empfindung des Schwarzen.” Sitzungsberichte / Akademie der Wissenschaften in
Wien, Mathematisch-Naturwissenschaftliche Klasse Abteilung III, Anatomie und Physiology des
Menschen und der Tiere sowie theoretische Medizin 69 (1874), 85–104.
[462] E. Hering. Outlines of a Theory of the Light Sense (Translation from German: Zur Lehre vom
Lichtsinne, 1878). Cambridge, MA: Harvard University Press, 1920.
[463] L. Hermann. “Eine Erscheinung simultanen Contrastes.” Pflügers Archiv für die gesamte
Physiologie des Menschen und Tiere 3 (1870), 13–15.
[464] E. W. Herold. “History and Development of the Color Picture Tube.” Proceedings of the
Society for Information Display 15:4 (1974), 141–149.
[465] R. D. Hersch, F. Collaud, and P. Emmel. “Reproducing Color Images with Embedded Metallic
Patterns.” ACM Transactions on Graphics 22:3 (2003), 427–434.
[466] M. Herzberger. Modern Geometrical Optics. New York: Interscience, 1958.
[467] P. G. Herzog. “Analytical Color Gamut Representations.” Journal for Imaging Science and
Technology 40 (1996), 516–521.
[468] P. G. Herzog. “Further Developments of the Analytical Color Gamut Representations.” In
Proceedings of the SPIE 3300, pp. 118–128. Bellingham, WA: SPIE, 1998.
[469] M. Hess and M. Wiegner. “COP: A Data Library of Optical Properties of Hexagonal Ice
Crystals.” Applied Optics 33 (1994), 7740–7749.
[470] S. Hesselgren. Hesselgrens färgatla med kortfattad färglära. Stockholm, Sweden: T Palmer,
AB, 1952.
[471] D. Hideaki, H. Y. K. Yukio, and S. Masataka. “Image Data Processing Apparatus for Process-
ing Combined Image Signals in order to Extend Dynamic Range.” In Image Sensor. 8223491,
Japanese Patent, 1996.
[472] B. M. Hill. “A Simple General Approach to Inference about the Tail of a Distribution.” The
Annals of Statistics 3:5 (1975), 1163–1174.
[473] J. W. Hittorf. “Über die Electricitätsleitung der Gase. Erste Mitteilungen.” Annalen der Physik
und Chemie 136 (1869), 1–31, 197–234.
[474] J. Ho, B. V. Funt, and M. S. Drew. “Separating a Color Signal into Illumination and Surface
Reflectance Components: Theory and Applications.” IEEE Transactions on Patterns Analysis
and Machine Intelligence 12:10 (1990), 966–977.
[475] B. Hoefflinger, editor. High-Dynamic-Range (HDR) Vision: Microelectronics, Image Process-
ing, Computer Graphics. Springer Series in Advanced Microelectronics, Berlin: Springer, 2007.
[476] H. Hofer, B. Singer, and D. R. Williams. “Different Sensations from Cones with the Same
Photopigment.” Journal of Vision 5 (2005), 444–454.
[477] A. Hohmann and C. von der Malsburg. “McCollough Effect and Eye Optics.” Perception 7
(1978), 551–555.
[478] G. Hollemann, B. Braun, P. Heist, J. Symanowski, U. Krause, J. Kränert, and C. Deter. “High-
Power Laser Projection Displays.” In Proceedings of the SPIE 4294, edited by M. H. Wu,
pp. 36–46. Bellingham, WA: SPIE, 2001.

i i

i i
 i i

i i

Bibliography 985

[479] N. Holonyak Jr. and S. F. Bevaqua. “Coherent (Visible) Light Emission from GaAs1−∞ P∞
Junctions.” Applied Physics Letters 1 (1962), 82–83.
[480] G. Hong, M. R. Luo, and P. A. Rhodes. “A Study of Digital Camera Colorimetric Characteri-
zation Based on Polynomial Modeling.” Color Research and Application 26:1 (2001), 76–84.
[481] Q. Hong, T. X. Wu, X. Zhu, R. Lu, and S.-T. Wu. “Extraordinarily High-Contrast and Wide-
View Liquid-Crystal Displays.” Applied Physics Letters 86 (2005), 121107–1–121107–3.
[482] F. M. Honrubia and J. H. Elliott. “Efferent Innervation of the Retina I: Morphologic Study of
the Human Retina.” Archives of Ophthalmology 80 (1968), 98–103.
[483] D. C. Hood and D. G. Birch. “A Quantitative Measure of the Electrical Activity of Human Rod
Photoreceptors using Electroretinography.” Visual Neuroscience 5 (1990), 379–387.
[484] D. C. Hood and M. A. Finkelstein. “Comparison of Changes in Sensitivity and Sensation:
Implications for the Response-Intensity Function of the Human Photopic System.” Journal of
Experimental Psychology: Human Perceptual Performance 5:3 (1979), 391–405.
[485] D. C. Hood, M. A. Finkelstein, and E. Buckingham. “Psychophysical Tests of Models of the
Response Function.” Vision Research 19:4 (1979), 401–406.
[486] S. J. Hook. “ASTER Spectral Library.”, 1999. Available online (http://speclib.jpl.nasa.gov).
[487] S. D. Hordley and G. D. Finlayson. “Reevaluation of Color Constancy Algorithm Perfor-
mance.” Journal of the Optical Society of America A 23:5 (2006), 1008–1020.
[488] B. K. P. Horn. Robot Vision. Cambridge, MA: MIT Press, 1986.
[489] L. J. Hornbeck. “From Cathode Rays to Digital Micromirrors: A History of Electronic Projec-
tion Display Technology.” Texas Instruments Technical Journal 15:3 (1998), 7–46.
[490] J. C. Horton and L. C. Sincich. “How Specific is V1 Input to V2 Thin Stripes?” Society for
Neuroscience Abstracts 34 (2004), 18.1.
[491] G. D. Horwitz, E. J. Chichilnisky, and T. D. Albright. “Spatial Opponency and Color Tuning
Dynamics in Macaque V1.” Society for Neuroscience Abstracts 34 (2004), 370.9.
[492] C. Hou, M. W. Pettet, V. Sampath, T. R. Candy, and A. M. Norcia. “Development of the Spatial
Organization and Dynamics of Lateral Interactions in the Human Visual System.” Journal of
Neuroscience 23:25 (2003), 8630–8640.
[493] D. H. Hubel and T. N. Wiesel. “Receptive Fields, Binocular Interaction and Functional Archi-
tecture in the Cat’s Visual Cortex.” Journal of Physiology 160 (1962), 106–154.
[494] D. H. Hubel and T. N. Wiesel. “Ferrier Lecture: Functional Architecture of Macaque Monkey
Visual Cortex.” Proceedings of the Royal Society of London B 198:1130 (1977), 1–59.
[495] P. M. Hubel, J. Holm, and G. Finlayson. “Illuminant Estimation and Color Correction.” In
Color Imaging, pp. 73–95. New York: Wiley, 1999.
[496] D. H. Hubel. Eye, Brain, and Vision, Reprint edition, see also
http://neuro.med.harvard.edu/site/dh/ edition. New York: W. H. Freeman and Company,
1995.
[497] A. J. Hughes. “Controlled Illumination for Birefringent Colour LCDs.” Displays 8:3 (1987),
139–141.
[498] A. C. Huk, D. Ress, and D. J. Heeger. “Neuronal Basis of the Motion Aftereffect Reconsid-
ered.” Neuron 32:1 (2001), 161–172.
[499] H. C. van de Hulst. Light Scattering by Small Particles. Mineola, NY: Dover Publications,
1981.
[500] P.-C. Hung and R. S. Berns. “Determination of Constant Hue Loci for a CRT Gamut and
Their Predictions using Color Appearance Spaces.” Color Research and Application 20 (1995),
285–295.

i i

i i
 i i

i i

986 Bibliography

[501] P.-C. Hung. “Colorimetric Calibration for Scanners and Media.” In Proceedings of the SPIE
1498, edited by W. Chang and J. R. Milch, pp. 164–174. Bellingham, WA: SPIE, 1991.
[502] P.-C. Hung. “Colorimetric Calibration in Electronic Imaging Devices using a Look-Up Table
Model and Interpolations.” Journal of Electronic Imaging 2 (1993), 53–61.
[503] P.-C. Hung. “Color Theory and its Application to Digital Still Cameras.” In Image Sensors and
Signal Processing for Digital Still Cameras, edited by J. Nakamura, pp. 205–221. Boca Raton,
FL: Taylor and Francis, 2006.
[504] R. W. G. Hunt. “Light and Dark Adaptation and the Perception of Color.” Journal of the
Optical Society of America 42 (1952), 190–199.
[505] R. W. G. Hunt. “The Strange Journey from Retina to Brain.” Journal of the Royal Television
Society 11 (1967), 220–229.
[506] R. W. G. Hunt. “Hue Shifts in Unrelated and Related Colors.” Color Research and Application
14 (1989), 235–239.
[507] R. W. G. Hunt. “Why is Black and White So Important in Colour?” In Colour Imaging: Vision
and Technology, edited by L. W. MacDonald and M. R. Luo, pp. 3–15. Chichester, UK: John
Wiley and Sons, Ltd., 1999.
[508] R. W. G. Hunt. “Saturation: Superfluous or Superior?” In IS&T/SID 9th Color Imaging
Conference, pp. 1–5. Springfield, VA: Society for Imaging Science and Technology, 2001.
[509] R. W. G. Hunt. The Reproduction of Colour, Sixth edition. Chichester, UK: John Wiley and
Sons Ltd., 2004.
[510] J.-M. Hupé, A. C. James, B. R. Payne, S. G. Lomber, P. Girard, and J. Bullier. “Cortical
Feedback Improves Discrimination between Figure and Ground.” Nature 394:6695 (1998), 784–
787.
[511] J.-M. Hupé, A. C. James, P. Girard, and J. Bullier. “Response Modulations by Static Texture
Surround in Area V1 of the Macaque Monkey Do Not Depend on Feedback Connections from
V2.” Journal of Neurophysiology 85:1 (2001), 146–163.
[512] B. S. Hur and M. G. Kang. “High Definition Color Interpolation Scheme for Progressive Scan
CCD Image Sensor.” IEEE Transactions on Consumer Electronics 47:1 (2001), 179–186.
[513] A. Hurlbert. “Formal Connections between Lightness Algorithms.” Journal of the Optical
Society of America A 3 (1986), 1684–1693.
[514] J. Hurri, A. Hyvärinen, and E. Oja. “Wavelets and Natural Image Statistics.” In Proceedings of
the 10th Scandinavian Conference on Image Analysis, pp. 13–18. International Association for
Pattern Recognition, 1997.
[515] L. M. Hurvich and D. Jameson. “Some Quantitative Aspects of Opponent-Colors Theory. IV.
A Psychological Color Specification System.” Journal of the Optical Society of America 46:6
(1956), 416–421.
[516] L. M. Hurvich and D. Jameson. “The Opponent Process Theory of Color Vision.” Psychologi-
cal Review 64 (1957), 384–404.
[517] L. M. Hurvich. Color Vision. Sunderland, MA: Sinauer Associates, 1981.
[518] J. B. Hutchings. “Colour and Appearance in Nature, Part 1.” Colour Research and Application
11 (1986), 107–111.
[519] J. B. Hutchings. “Color in Anthopology and Folklore.” In Color for Science, Art and Technol-
ogy, edited by K. Nassau. Amsterdam: Elsevier Science B.V., 1998.
[520] J. B. Hutchings. “Color in Plants, Animals and Man.” In Color for Science, Art and Technology,
edited by K. Nassau. Amsterdam: Elsevier Science B.V., 1998.

i i

i i
 i i

i i

Bibliography 987

[521] J. Hynecek. “CDS Noise Reduction of Partially Reset Charge-Detection Nodes.” IEEE Trans-
actions on Circuits and Systems — I: Fundamental Theory and Applications 49:3 (2002), 276–
280.
[522] A. Hyvärinen. “Survey on Independent Components Analysis.” Neural Computing Surveys 2
(1999), 94–128.
[523] ICC. “Image Technology Colour Management - Architecture, Profile Format, and Data Struc-
ture.” Technical Report ICC.1:2004-10, International Color Consortium, 2004. Available online
(http://www.color.org).
[524] IEC. “Part 2-1: Colour Management - Default RGB Colour Space - sRGB.” Technical Report
61966, International Electrotechnical Commission, Geneva, Switzerland, 1999.
[525] M. Ikebe and K. Saito. “A Wide-Dynamic-Range Compression Image Sensor with Negative-
Feedback Resetting.” IEEE Sensors Journal 7:5 (2007), 897–904.
[526] E. Ikeda. “Image Data Processing Apparatus for Processing Combined Image Signals in order
to Extend Dynamic Range.” In 5801773, United States Patent, 2003.
[527] Q. M. T. Inc. “Interferometric Modulator (IMOD) Technology Overview.”, 2007. White paper.
[528] M. Inui. “Fast Algorithm for Computing Color Gamuts.” Color Research and Application 18
(1993), 341–348.
[529] P. Irawan, J. A. Ferwerda, and S. R. Marshner. “Perceptually Based Tone Mapping of High
Dynamic Range Image Streams.” In Eurographics Symposium on Rendering, pp. 231–242.
Aire-la-Ville, Switzerland: Eurographics Association, 2005.
[530] S. Ishida, Y. Yamashita, T. Matsuishi, M. Ohshima, T. Ohshima, K. Kato, and H. Maeda.
“Photosensitive Sseizures Provoked while Viewing ”Pocket Monster”, a Made-for-Television
Animation Program in Japan.” Epilepsia 39 (1998), 1340–1344.
[531] S. Ishihara. Tests for Colour-Blindness. Tokyo: Hongo Harukicho, 1917.
[532] M. F. Iskander. Electromagnetic Fields and Waves. Upper Saddle River, NJ: Prentice Hall,
1992.
[533] ISO. “Graphic technology - Spectral measurement and colorimetric computation for graphic
arts images.” Technical Report ISO 13655:1996, International Organization for Standardization,
1996.
[534] ISO. “Photography – Digital Still Cameras – Determination of Exposure Index, ISO Speed
Ratings, Standard Output Sensitivity, and Recommended Exposure Index.” 12232:2006.
[535] S. Itoh and M. Tanaka. “Current Status of Field-Emission Displays.” Proceedings of the IEEE
90:4 (2002), 514–520.
[536] ITU-R. “Parameter Values for HDTV Standards for Production and International Programme
Exchange,.” Technical report, BT.709-2, International Telecommunication Union, Geneva,
Switzerland, 1995.
[537] A. V. Ivashchenko. Dichroic Dyes for Liquid Crystal Displays. Boca Raton, Florida: CRC
Press, 1994.
[538] H. Ives. “The Transformation of Color-Mixture Equations from One System to Another.”
Journal of Franklin Inst. 16 (1915), 673–701.
[539] H. E. Ives. “The Resolution of Mixed Colors by Differential Visual Activity.” Philosophical
Magazine 35 (1918), 413–421.
[540] J. Iwao. “Errors in Color Calculations Due to Fluorescence when using the Neugebauer Equa-
tions.” In Proceedings of the Technical Association of the Graphical Arts, pp. 254–266. Sewick-
ley, PA: Technical Association of the Graphic Arts, 1973.

i i

i i
 i i

i i

988 Bibliography

[541] D. Jackèl and B. Walter. “Modeling and Rendering of the Atmosphere using Mie Scattering.”
Computer Graphics Forum 16:4 (1997), 201–210.
[542] D. Jackèl and B. Walter. “Simulation and Visualization of Halos.” In Proceedings of ANI-
GRAPH ’97, 1997.
[543] A. K. Jain. Fundamentals of Digital Image Processing. Upper Saddle River, NJ: Prentice-Hall,
1989.
[544] D. Jameson and L. M. Hurvich. “Color Adaptation: Sensitivity, Contrast and After-Images.”
In Handbook of Sensory Physiology, VII/4, edited by D. Jameson and L. M. Hurvich, VII/4,
pp. 568–581. Berlin: Springer-Verlag, 1972.
[545] T. H. Jamieson. “Thin-Lens Theory of Zoom Systems.” Optica Acta 17:8 (1970), 565–584.
[546] J. Janesick, T. Elliott, S. Collins, M. Blouke, and J. Freeman. “Scientific Charge-Coupled
Devices.” Optical Engineering 26:8 (1987), 692–715.
[547] J.-S. Jang and B. Javidi. “Improvement of Viewing Angle in Integral Imaging by Use of Moving
Lenslet Arrays with Low Fill Factor.” Applied Optics 42:11 (2003), 1996–2002.
[548] H. W. Jensen, J. Legakis, and J. Dorsey. “Rendering of Wet Materials.” In Proceedings of
the 10th Eurographics Symposium on Rendering, edited by D. Lischinski and G. W. Larson,
pp. 273–281. Vienna: Springer-Verlag, 1999.
[549] H. W. Jensen. Realistic Image Synthesis using Photon Mapping. Natick, MA: A K Peters,
2001.
[550] E. Jin, X. Feng, and J. Newell. “The Development of a Color Visual Difference Model
(CVDM).” In IS&T PICS Conference, pp. 154–158. Springfield, VA: Society for Imaging Sci-
ence and Technology, 1998.
[551] L. Jing and K. Urahama. “Image Recoloring by Eigenvector Mapping.” In Proceedings of the
International Workshop on Advanced Image Technology, 2006.
[552] D. J. Jobson, Z. Rahman, and G. A. Woodell. “Retinex Image Processing: Improved Fidelity
to Direct Visual Observation.” In Proceedings of the IS&T Fourth Color Imaging Conference:
Color Science, Systems, and Applications, pp. 124–125. Springfield, VA: Society for Imaging
Science and Technology, 1995.
[553] D. J. Jobson, Z. Rahman, and G. A. Woodell. “A Multi-Scale Retinex for Bridging the Gap
between Color Images and Human Observation of Scenes.” IEEE Transactions on Image Pro-
cessing 6:7 (1997), 965–976.
[554] T. Johansson. Färg. Stockholm, Sweden: Lindfors Bokförlag, AB, 1937.
[555] G. M. Johnson and M. D. Fairchild. “A Top Down Description of S-CIELAB and CIEDE2000.”
Color Research and Application 28 (2003), 425–435.
[556] G. M. Johnson and F. Mark D. “The Effect of Opponent Noise on Image Quality.” In SPIE
Proceedings 5668 (Electronic Imaging Conference), pp. 82–89. Bellingham, WA: SPIE, 2005.
[557] P. D. Jones and D. H. Holding. “Extremely Long-Term Persistence of the McCollough Effect.”
Journal of Experimental Psychology 1:4 (1975), 323–327.
[558] D. B. Judd and G. Wyszecki. Color in business, science and industry, Third edition. New York:
Wiley Interscience, 1975.
[559] D. B. Judd, D. L. MacAdam, and G. Wyszecki. “Spectral Distribution of Typical Light as
a Function of Correlated Color Temperature.” Journal of the Optical Society of America 54:8
(1964), 1031–1040.
[560] D. B. Judd. “Chromaticity Sensibility to Stimulus Differences.” Journal of the Optical Society
of America 22 (1932), 72–108.

i i

i i
 i i

i i

Bibliography 989

[561] D. B. Judd. “Sensibility to Color-Temperature Change as a Function of Temperature.” Journal

of the Optical Society of America 23 (1933), 7.
[562] T. B. Jung, J. C. Kim, and S. H. Lee. “Wide-Viewing-Angle Transflective Display Associated
with a Fring-Field Driven Homogeneously Aligned Nematic Liquid Crystal Display.” Japanese
Journal of Applied Physics 42:5A (2003), L464–L467.
[563] J. H. Kaas and D. C. Lyon. “Visual Cortex Organization in Primates: Theories of V3 and
Adjoining Visual Areas.” Progress in Brain Research 134 (2001), 285–295.
[564] J. H. Kaas, M. F. Huerta, J. T. Weber, and J. K. Harting. “Patterns of Retinal Terminations and
Laminar Organization of the Lateral Geniculate Nucleus of Primates.” Journal of Comparative
Neurology 182:3 (1978), 517–554.
[565] F. Kainz, R. Bogart, and D. Hess. “The OpenEXR Image File Format.” In SIGGRAPH ’03
Technical Sketches, 2003. See also: http://www.openexr.com.
[566] P. K. Kaiser and R. M. Boynton, editors. Human Color Vision, Second edition. Washington,
DC: Optical Society of America, 1996.
[567] P. K. Kaiser. “Photopic and Mesopic Photometry: Yesterday, Today, and Tomorrow.” In Golden
Jubilee of Colour in the CIE : Proceedings of a Symposium held by The Colour Group (Great
Britain) at Imperial College, London. Bradford, UK: The Society of Dyers and Colourists, 1981.
[568] J. T. Kajiya. “The Rendering Equation.” Computer Graphics (SIGGRAPH ’86 Proceedings)
20:4 (1986), 143–150.
[569] K. Kakinuma. “Technology of Wide Color Gamut Backlight with Light-Emitting Diode for
Liquid Crystal Display Television.” Japanese Journal of Applied Physics 45:5B (2006), 4330–
4334.
[570] M. Kalloniatis and C. Luu. “Psychophysics of Vision.”, 2006.
http://webvision.med.utah.edu/KallSpatial.html.
[571] E. Kaneko. Liquid crystal TV displays: Principles and applications of liquid crystal displays.
Tokyo: KTK Scientific Publishers, 1987. In co-publication with D Reidel Publishing Company,
Dordrecht, Holland.
[572] H. R. Kang and P. G. Anderson. “Neural Network Applications to the Colour Scanner and
Printer Calibrations.” Journal of Electronic Imaging 1 (1992), 125–134.
[573] S. B. Kang, M. Uttendaele, S. Winder, and R. Szeliski. “High Dynamic Range Video.” ACM
Transactions on Graphics 22:3 (2003), 319–325.
[574] H. R. Kang. “Colour Scanner Calibration.” Journal of Imaging Science and Technology 36
(1992), 162–170.
[575] H. R. Kang. Color Technology for Electronic Imaging Devices. Belingham, Washington: SPIE
Optical Engineering Press, 1997.
[576] H. R. Kang. “Computational Accuracy of RGB Encoding Standards.” In IS&Ts NIP16: In-
ternational Conference on Digital Printing Technologies, pp. 661–664. Springfield, VA: Society
for Imaging Science and Technology, 2000.
[577] H. R. Kang. Computational Color Technology. Bellingham, WA: SPIE Press, 2006.
[578] R. Kanthack. Tables of Refractive Indices, II. London: Hilger, 1921.
[579] K. Karhunen. “Über lineare Methoden in der Warscheinlichkeitsrechnung.” Annales
Academiae Scientiarium Fennicae 37 (1947), 3–79.
[580] J. E. Kasper and S. A. Feller. The Complete Book of Holograms: How They Work and How to
Make Them. New York: John Wiley and Sons, Inc., 1987.

i i

i i
 i i

i i

990 Bibliography

[581] N. Katoh and K. Nakabayashi. “Applying Mixed Adaptation to Various Chromatic Adapta-
tion Transformation (CAT) Models.” In IS&T PICS Conference, pp. 299–305. Springfield, VA:
Society for Imaging Science and Technology, 2001.
[582] J. Kautz and M. McCool. “Interactive Rendering with Arbitrary BRDFs using Separable Ap-
proximations.” In Proceedings of the 10th Eurographics Symposium on Rendering, edited by
D. Lischinski and G. W. Larson, pp. 247–260. Vienna: Springer-Verlag, 1999.
[583] R. Kawakami, J. Takamatsu, and K. Ikeuchi. “Color Constancy from Blackbody Illumination.”
Journal of the Optical Society of America A. (to appear), 2007.
[584] P. Kay and C. K. McDaniel. “The Linguistic Significance of the Meanings of Basic Color
Terms.” Language 54:3 (1978), 610–646.
[585] E. A. Khan and E. Reinhard. “Evaluation of Color Spaces for Edge Classification in Outdoor
Scenes.” In IEEE International Conference on Image Processing, pp. 952–955. Washington,
DC: IEEE Press, 2005.
[586] E. A. Khan, A. O. Akyüz, and E. Reinhard. “Ghost Removal in High Dynamic Range Images.”
In IEEE International Conference on Image Processing, pp. 2005–2008. Washington, DC: IEEE
Press, 2006.
[587] E. A. Khan, E. Reinhard, R. Fleming, and H. Bülthoff. “Image-Based Material Editing.” ACM
Transactions on Graphics 25:3 (2006), 654–663.
[588] J.-H. Kim and J. P. Allebach. “Color Filters for CRT-Based Rear Projection Television.” IEEE
Transactions on Consumer Electronics 42:4 (1996), 1050–1054.
[589] H. Kim, J.-Y. Kim, S. H. Hwang, I.-C. Park, and C.-M. Kyung. “Digital Signal Processor with
Efficient RGB Interpolation and Histogram Accumulation.” IEEE Transactions on Consumer
Electronics 44:4 (1998), 1389–1395.
[590] R. Kimmel. “Demosaicing: Image Reconstruction from Color CCD Samples.” IEEE Transac-
tions on Image Processing 8:9 (1999), 1221–1228.
[591] G. Kindlmann, E. Reinhard, and S. Creem. “Face-Based Luminance Matching for Perceptual
Colormap Generation.” In Proceedings of IEEE Visualization, pp. 309–406. Washington, DC:
IEEE, 2002.
[592] D. C. Kiper, J. B. Levitt, and K. R. Gegenfurtner. “Chromatic Signals in Extrastriate Areas
V2 and V3.” In Color Vision: From Genes to Perception, edited by K. R. Gegenfurtner and
L. Sharpe, pp. 249–268. Cambridge, UK: Cambridge University Press, 1999.
[593] A. Kitaoka, J. Gyoba, H. Kawabata, and K. Sakurai. “Two Competing Mechanisms Underlying
Neon Color Spreading, Visual Phantoms and Grating Induction.” Vision Research 41:18 (2001),
2347–2354.
[594] J. Kleinschmidt and J. E. Dowling. “Intracellular Recordings from Gecko Photoreceptors dur-
ing Light and Dark Adaptation.” Journal of General Physiology 66:5 (1975), 617–648.
[595] G. J. Klinker, S. A. Shafer, and T. Kanade. “Image Segmentation and Reflectance Analy-
sis through Color.” In Proceedings of the SPIE 937 (Application of Artificial Intelligence VI),
pp. 229–244. Bellingham, WA: SPIE, 1988.
[596] G. J. Klinker, S. A. Shafer, and T. Kanade. “A Physical Approach to Color Image Understand-
ing.” International Journal of Computer Vision 4:1 (1990), 7–38.
[597] K. Klug, N. Tiv, Y. Tsukamoto, P. Sterling, and S. Schein. “Blue Cones Contact OFF-Midget
Bipolar Cells.” Society for Neuroscience Abstracts 18 (1992), 838.
[598] K. Klug, Y. Tsukamoto, P. Sterling, and S. Schein. “Blue Cone OFF-Midget Ganglion Cells in
Macaque.” Investigative Ophthalmology and Visual Science, Supplement 34 (1993), 986–986.
[599] R. Knight, S. L. Buck, G. A. Fowler, and A. Nguyen. “Rods Affect S-Cone Discrimination on
the Farnsworth-Munsell 100-Hue Test.” Vision Research 38:21 (1998), 3477–3481.

i i

i i
 i i

i i

Bibliography 991

[600] W. E. Kock. Lasers and Holography, Second edition. New York: Dover Publications, Inc,
1981.
[601] J. J. Koenderink, A. van Doorn, and M. Stavridi. “Bidirectional Reflection Distribution Func-
tion Expressed in Terms of Surface Scattering Modes.” In European Conference on Computer
Vision, pp. 28–39. London: Springer-Verlag, 1996.
[602] K. Koffka. Principles of Gestallt Psychology. New York: Harcourt, Brace, and World, 1935.
[603] A. J. F. Kok. “Ray Tracing and Radiosity Algorithms for Photorealistic Image Synthesis.”
Ph.D. thesis, Delft University of Technology, 1994.
[604] H. Kolb and L. Dekorver. “Midget Ganglion Cells of the Parafovea of the Human Retina: A
Study by Electron Microscopy and Serial Section Reconstructions.” Journal of Comparative
Neurology 303:4 (1991), 617–636.
[605] H. Kolb and E. V. Famiglietti. “Rod and Cone Pathways in the Inner Plexiform Layer of the
Cat Retina.” Science 186 (1974), 47–49.
[606] H. Kolb, K. A. Linberg, and S. K. Fisher. “Neurons of the Human Retina: A Golgi Study.”
Journal of Comparative Neurology 318:2 (1992), 147–187.
[607] H. Kolb. “The Inner Plexiform Layer in the Retina of the Cat: Electron Microscopic Observa-
tions.” Journal of Neurocytology 8 (1979), 295–329.
[608] H. Kolb. “Anatomical Pathways for Color Vision in the Human Retina.” Visual Neuroscience
7 (1991), 61–74.
[609] H. Kolb. “How the Retina Works.” American Scientist 91 (2003), 28–35.
[610] A. König. “Über den Menschlichen Sehpurpur und seine Bedeutung für das Sehen.” Sitzungs-
berichte der Akademie der Wissenschaften, Berlin, pp. 577–598.
[611] A. König. Gesammelte Abhandlungen zur Physiologischen Optik. Leipzig: Barth, 1903.
[612] N. Kouyama and D. W. Marshak. “Bipolar Cells Specific for Blue Cones in the Macaque
Retina.” Journal of Neuroscience 12:4 (1992), 1233–1252.
[613] T. Koyama. “Optics in Digital Still Cameras.” In Image Sensors and Signal Processing for
Digital Still Cameras, edited by J. Nakamura, pp. 21–51. Boca Raton, FL: Taylor and Francis,
2006.
[614] J. Krauskopf, Q. Zaidi, and M. B. Mandler. “Mechanisms of Simultaneous Color Induction.”
Journal of the Optical Society of America A 3:10 (1986), 1752–1757.
[615] J. Krauskopf. “Light Distribution in Human Retinal Images.” Journal of the Optical Society of
America 52:9 (1962), 1046–1050.
[616] J. Krauskopf. “Effect of Retinal Stabilization on the Appearance of Heterochromatic Targets.”
Journal of the Optical Society of America 53 (1963), 741–744.
[617] Y. A. Kravtsov and Y. I. Orlov. Geometrical Optics of Inhomogeneous Media. Springer Series
on Wave Phenomena, Berlin: Springer-Verlag, 1990.
[618] G. Krawczyk, R. Mantiuk, K. Myszkowski, and H.-P. Seidel. “Lightness Perception Inspired
Tone Mapping.” In APGV ’04: Proceedings of the 1st Symposium on Applied Perception in
Graphics and Visualization, pp. 172–172. New York: ACM Press, 2004.
[619] G. Krawczyk, K. Myszkowski, and H.-P. Seidel. “Lightness Perception in Tone Reproduction
for High Dynamic Range Images.” Computer Graphics Forum 24:3 (2005), 635–645.
[620] G. Krawczyk, K. Myszkowski, and H.-P. Seidel. “Computational Model of Lightness Per-
ception in High Dynamic Range Imaging.” In Proceedings of SPIE 6057 (Human Vision and
Electronic Imaging). Bellingham, WA: SPIE, 2006.

i i

i i
 i i

i i

992 Bibliography

[621] W. Kress and M. Stevens. “Derivation of 3-Dimensional Gamut Descriptors for Graphic Arts
Output Devices.” In Proceedings of the Technical Association of Graphical Arts, pp. 199–214.
Sewickley, PA: Technical Assocation of the Graphics Arts, 1994.
[622] J. von Kries. “Die Gesichtsempfindungen.” In Handbuch der Physiologie des Menschen, edited
by W. Nagel, pp. 109–282. Braunschweig, Germany: Vieweg, 1905.
[623] J. von Kries. “Chromatic Adaptation.” In Sources of Color Vision, edited by D. L. MacAdam.
Cambridge, MA: MIT Press, 1970. Originally published in Festschrift der Albrecht-Ludwig-
Universitat, 1902.
[624] J. Kruger and P. Gouras. “Spectral Selectivity of Cells and Its Dependence on Slit Length in
Monkey Visual Cortex.” Journal of Neurophysiology 43 (1979), 1055–1069.
[625] J. Kuang, H. Yamaguchi, G. M. Johnson, and M. D. Fairchild. “Testing HDR Image Render-
ing Algorithms.” In Proceedings of IS&T/SID 12th Color Imaging Conference, pp. 315–320.
Springfield, VA: Society for Imaging Science and Technology, 2004.
[626] J. Kuang, G. M. Johnson, and M. D. Fairchild. “Image Preference Scaling for HDR Image
Rendering.” In IS&T/SID 13th Color Imaging Conference, pp. 8–13. Springfield, VA: Society
for Imaging Science and Technology, 2005.
[627] J. Kuang, H. Yamaguchi, C. Liu, G. M. Johnson, and M. D. Fairchild. “Evaluating HDR
Rendering Algorithms.” ACM Transactions on Applied Perception 4:2 (2007), 9–1 – 9–27.
[628] P. Kubelka and F. Munk. “Ein Beitrag zur Optik der Farbanstriche.” Zeitschrift für technische
Physik 12 (1931), 593–601.
[629] R. G. Kuehni. Computer Colorant Formulation. Lexington, MA: Lexington Books, 1975.
[630] R. G. Kuehni. Color: An Introduction to Practice and Principles. New York: John Wiley &
Sons, 1997.
[631] R. G. Kuehni. Color Space and Its Divisions: Color Order from Antiquity to the Present. New
York: Wiley-Interscience, 2003.
[632] S. W. Kuffler. “Discharge Patterns and Functional Organization of Mammalian Retina.” Jour-
nal of Neurophysiology 16 (1953), 37–68.
[633] T. Kuno, H. Sugiura, and N. Matoba. “New Interpolation Method using Discriminated Color
Correction for Digital Still Cameras.” IEEE Transactions on Consumer Electronics 45:1 (1999),
259–267.
[634] C. J. Kuo and M. H. Tsai, editors. Three-Dimensional Holographic Imaging. New York: John
Wiley and Sons, Inc., 2002.
[635] E. A. Lachica and V. A. Casagrande. “Direct W-Like Geniculate Projections to the
Cytochrome-Oxydase (CO) Blobs in Primate Visual Cortex: Axon Morphology.” Journal of
Comparative Neurology 319:1 (1992), 141–158.
[636] E. A. Lachica, P. D. Beck, and V. A. Casagrande. “Parallel Pathways in Macaque Monkey Stri-
ate Cortex: Anatomically Defined Columns in Layer III.” Proceedings of the National Academy
of Sciences of the United States of America 89:8 (1992), 3566–3560.
[637] E. P. F. Lafortune, S.-C. Foo, K. E. Torrance, and D. P. Greenberg. “Non-Linear Approximation
of Reflectance Functions.” In Proceedings of SIGGRAPH’97, Computer Graphics Proceedings,
Annual Conference Series, pp. 117–126. Reading, MA: Addison-Wesley, 1997.
[638] R. Lakowski. “Colorimetric and Photometric Data for the 10th Edition of the Ishihara Plates.”
The British Journal of Physiological Optics 22:4 (1965), 195–207.
[639] P. Lalonde and A. Fournier. “Filtered Local Shading in the Wavelet Domain.” In Proceedings
of the 8th Eurographics Symposium on Rendering, pp. 163–174. Vienna: Springer-Verlag, 1997.

i i

i i
 i i

i i

Bibliography 993

[640] P. Lalonde and A. Fournier. “A Wavelet Representation of Reflectance Functions.” IEEE

Transactions on Visualization and Computer Graphics 3:4 (1997), 329–336.
[641] K. Lam. “Metamerism and Colour Constancy.” Ph.D. thesis, University of Bradford, 1985.
[642] T. D. Lamb and J. Bourriau, editors. Colour: Art and Science. Cambridge, UK: Cambridge
University Press, 1995.
[643] T. D. Lamb. “Photoreceptor Spectral Sensitivities: Common Shape in the Long Wavelength
Region.” Vision Research 35:22 (1995), 3083–3091.
[644] T. D. Lamb. “Photopigments and the Biophysics of Transduction in Cone Photoreceptors.” In
Color Vision: From Genes to Perception, edited by K. R. Gegenfurtner and L. Sharpe, pp. 89–
101. Cambridge, UK: Cambridge University Press, 1999.
[645] J. H. Lambert. “Photometria sive de mensura et gradibus luminis, colorum et umbrae.”, 1760.
[646] D. Lamming. “Contrast Sensitivity.” In Vision and Visual Dysfunction, edited by J. R. Cronley-
Dillon. London: Macmillan Press, 1991.
[647] E. Land and J. McCann. “Lightness and Retinex Theory.” Journal of the Optical Society of
America 61:1 (1971), 1–11.
[648] E. Land. “The Retinex Theory of Color Vision.” Scientific American 237:6 (1977), 108–128.
[649] E. Land. “Recent Advances in Retinex Theory.” Vision Research 26:1 (1986), 7–21.
[650] L. D. Landau and E. M. Lifshitz. Fluid Mechanics, Course of Theoretical Physics, 6, Second
edition. Oxford: Pergamon, 1987.
[651] M. S. Langer. “Large-Scale Failures of f −α Scaling in Natural Image Spectra.” Journal of the
Optical Society of America A 17:1 (2000), 28–33.
[652] I. Langmuir. “The Effect of Space Charge and Residual Gases on Thermionic Currents in High
Vacuum.” Physical Review 2:6 (1913), 450–486.
[653] R. D. Larrabee. “Spectral Emissivity of Tungsten.” Journal of the Optical Society of America
49:6 (1959), 619–625.
[654] P. Laven. “Mieplot.” Available online (http://www.philiplaven.com).
[655] J. Lawrence, S. Rusinkiewicz, and R. Ramamoorthi. “Efficient BRDF Importance Sampling
using a Factored Representation.” ACM Transactions on Graphics 23:3 (2004), 496–505.
[656] Y. Le Grand. Light, Colour and Vision, Second edition. Somerset, N.J.: Halsted Press, 1968.
Translated by R. W. G. Hunt, J. W. T. Walsh and F. R. W. Hunt.
[657] S. J. Leat, R. V. North, and H. Bryson. “Do Long Wavelength Pass Filters Improve Low Vision
Performance?” Ophthalmic and Physiological Optics 10:3 (1991), 219–224.
[658] P. Ledda, L.-P. Santos, and A. Chalmers. “A Local Model of Eye Adaptation for High Dynamic
Range Images.” In Proceedings of ACM Afrigraph, pp. 151–160. New York: ACM, 2004.
[659] P. Ledda, A. Chalmers, T. Troscianko, and H. Seetzen. “Evaluation of Tone Mapping Operators
using a High Dynamic Range Display.” ACM Transactions on Graphics 24:3 (2005), 640–648.
[660] B. B. Lee, P. R. Martin, and A. Valberg. “The Physiological Basis of Heterochromatic Flicker
Photometry Demonstrated in the Ganglion Cells of the Macaque Retina.” Journal of Physiology
404 (1988), 323–347.
[661] B. B. Lee, J. Kremers, and T. Yeh. “Receptive Fields of Primate Retinal Ganglion Cells Studied
with a Novel Technique.” Visual Neuroscience 15:1 (1998), 161–175.
[662] C. H. Lee, B. J. Moon, H. Y. Lee, E. Y. Chung, and Y. H. Ha. “Estimation of Spectral Distri-
bution of Scene Illumination from a Single Image.” Journal of Imaging Science and Technology
44:4 (2000), 308–320.

i i

i i
 i i

i i

994 Bibliography

[663] T. Lee, J. Zaumseil, Z. Bao, J. W. P. Hsu, and J. A. Rogers. “Organic Light-Emitting Diodes
Formed by Soft Contact Lamination.” Proceedings of the National Academy of Sciences of the
United States of America 101:2 (2004), 429–433.
[664] H.-C. Lee. “Method for Computing the Scene-Illuminant Chromaticity from Specular High-
light.” Journal of the Optical Society of America A 3:10 (1986), 1694–1699.
[665] H.-C. Lee. “Internet Color Imaging.” In Proceedings of the SPIE 3080, pp. 122–135. Belling-
ham, WA: SPIE, 2000.
[666] H.-C. Lee. Introduction to Color Imaging Science. Cambridge, UK: Cambridge University
Press, 2005.
[667] A. Lefohn, R. Caruso, E. Reinhard, B. Budge, and P. Shirley. “An Ocularist’s Approach to
Human Iris Synthesis.” IEEE Computer Graphics and Applications 23:6 (2003), 70–75.
[668] J. Lekner and M. C. Dorf. “Why Some Things Are Darker when Wet.” Applied Optics 27:7
(1988), 1278–1280.
[669] P. Lennie, J. Krauskopf, and G. Sclar. “Chromatic Mechanisms in Striate Cortex of Macaque.”
Journal of Neuroscience 10:2 (1990), 649–669.
[670] P. Lennie, J. Pokorny, and V. C. Smith. “Luminance.” Journal of the Optical Society of America
A 10:6 (1993), 1283–1293.
[671] H. Lensch, J. Kautz, M. Goessele, W. Heidrich, and H. Seidel. “Image-Based Reconstruction
of Spatially Varying Materials.” In Proceedings of the 12th Eurographics Workshop on Ren-
dering, edited by S. J. Gortler and K. Myszkowski, pp. 104–115. Aire-la-Ville, Switzerland:
Eurographics Association, 2001.
[672] S. Lerman, J. M. Megaw, and M. N. Moran. “Further Studies on the Effects of UV Radiation
on the Human Lens.” Ophthalmic Research 17:6 (1985), 354–361.
[673] A. G. Leventhal, R. W. Rodieck, and B. Dreher. “Retinal Ganglion Cell Classes in the Old
World Monkey: Morphology and Central Projections.” Science 213 (1981), 1139–1142.
[674] A. G. Leventhal, K. G. Thompson, D. Liu, Y. Zhou, and S. J. Ault. “Concomitant Sensitivity
to Orientation, Direction, and Color of Cells in Layers 2, 3, and 4 of Monkey Striate Cortex.”
Journal of Neuroscience 15:3 (1995), 1808–1818.
[675] H. W. Leverenz. An Introduction to Luminescence of Solids. Mineola, NY: Dover Publications,
1968.
[676] M. Levoy and P. Hanrahan. “Light Field Rendering.” In Proceedings SIGGRAPH ’96, Com-
puter Graphics Proceedings, Annual Conference Series, pp. 31–42. Reading, MA: Addison-
Wesley, 1996.
[677] I. Levy, U. Hasson, G. Avidan, T. Hendler, and R. Malach. “Center-Periphery Organization of
Human Object Areas.” Nature Neuroscience 4:5 (2001), 533–539.
[678] P. A. Lewis. “Colorants: Organic and Inorganic Pigments.” In Color for Science, Art and
Technology, edited by K. Nassau. Amsterdam: Elsevier Science B.V., 1998.
[679] X. Li and A. Gilchrist. “Relative Area and Relative Luminance Combine to Anchor Surface
Lightness Values.” Perception and Psychophysics 61 (1999), 771–785.
[680] M. Li and J.-M. Lavest. “Some Aspects of Zoom Lens Camera Calibration.” IEEE Transactions
on Pattern Analysis and Machine Intelligence 18:11 (1996), 1105–1110.
[681] Y. Li, L. Sharan, and E. H. Adelson. “Compressing and Companding High Dynamic Range
Images with Subband Architectures.” ACM Transactions on Graphics 24:3 (2005), 836–844.
[682] Z. Li-Tolt, R. L. Fink, and Z. Yaniv. “Electron Emission from Patterned Diamond Flat Cath-
odes.” J. Vac. Sci. Tech. B 16:3 (1998), 1197–1198.

i i

i i
 i i

i i

Bibliography 995

[683] A. Liberles. Introduction to Molecular-Orbital Theory. New York: Holt, Rinehart and Winston,
Inc., 1966.
[684] A. Lien. “A Detailed Derivation of Extended Jones Matrix Representation for Twisted Nematic
Liquid Crystal Displays.” Liquid Crystals 22:2 (1997), 171–175.
[685] B. J. Lindbloom, 2005. Available online (http://www.brucelindbloom.com).
[686] D. Lischinski, Z. Farbman, M. Uyttendaele, and R. Szeliski. “Interactive Local Adjustment of
Tonal Values.” ACM Transactions on Graphics 25:3 (2006), 646–653.
[687] D. Litwiller. “CCD vs. CMOS: Facts and Fiction.” Photonics Spectra 35:1 (2001), 151–154.
[688] D. Litwiller. “CCD vs. CMOS: Maturing Technologies, Maturing Markets.” Photonics Spectra
39:8 (2005), 54–58.
[689] C. Liu and M. D. Fairchild. “Measuring the Relationship between Pereived Image Contrast and
Surround Illuminantion.” In IS&T/SID 12th Color Imaging Conference, pp. 282–288. Spring-
field, VA: Society for Imaging Science and Technology, 2004.
[690] Y. Liu, J. Shigley, E. Fritsch, and S. Hemphill. “Abnormal Hue-Angle Change of the Gemstone
Tanzanite between CIE Illuminants D65 and A in CIELAB Color Space.” Color Research and
Application 20 (1995), 245–250.
[691] C. Liu, P. R. Jonas, and C. P. R. Saunders. “Pyramidal Ice Crystal Scattering Phase Functions
and Concentric Halos.” Annales Geophysicae 14 (1996), 1192–1197.
[692] M. S. Livingstone and D. H. Hubel. “Anatomy and Physiology of a Color System in the Primate
Visual Cortex.” Journal of Neuroscience 4:1 (1984), 309–356.
[693] M. S. Livingstone and D. H. Hubel. “Connections between Layer 4B of Area 17 and the Thick
Cytochrome Oxidase Stripes of Area 18 in the Squirrel Monkey.” Journal of Neuroscience 7:11
(1987), 3371–3377.
[694] M. S. Livingstone and D. H. Hubel. “Segregation of Form, Color, Movement, and Depth:
Anatomy, Physiology, and Perception.” Science 240:4853 (1988), 740–749.
[695] M. S. Livingstone. Vision and Art: The Biology of Seeing. New York: Harry N Abrams, 2002.
[696] M. M. Loève. Probability Theory. Princeton: Van Nostrand, 1955.
[697] B. London and J. Upton. Photography, Sixth edition. London: Longman, 1998.
[698] A. V. Loughren. “Recommendations of the National Television System Committee for a Color
Television Signal.” Journal of the SMPTE 60 (1953), 321–326, 596.
[699] O. Lowenstein and I. E. Loewenfeld. “The Pupil.” In The Eye, Volume 3, edited by H. Davson,
pp. 255–337. New York: Academic Press, 1969.
[700] J. Lu and D. M. Healey. “Contrast Enhancement via Multiscale Gradient Transformation.”
In Proceedings of the 16th IEEE International Conference on Image Processing, Volume II,
pp. 482–486. Washington, DC: IEEE, 1992.
[701] R. Lu, J. J. Koenderink, and A. M. L. Kappers. “Optical Properties (Bidirectional Reflectance
Distribution Functions) of Velvet.” Applied Optics 37:25 (1998), 5974–5984.
[702] R. Lu, J. J. Koenderink, and A. M. L. Kappers. “Specularities on Surfaces with Tangential
Hairs or Grooves.” Computer Vision and Image Understanding 78 (1999), 320–335.
[703] B. D. Lucas and T. Kanade. “An Iterative Image Registration Technique with an Application to
Stereo Vision.” In Proceedings of the 1981 DARPA Image Understanding Workshop, pp. 121–
130, 1981.
[704] M. P. Lucassen and J. Walraven. “Evaluation of a Simple Method for Color Monitor Recali-
bration.” Color Research and Application 15:6 (1990), 321–326.

i i

i i
 i i

i i

996 Bibliography

[705] C. J. Lueck, S. Zeki, K. J. Friston, M.-P. Deiber, P. Cope, V. J. Cunningham, A. A. Lam-

mertsma, C. Kennard, and R. S. J. Frackowiak. “The Colour Centre in the Cerebral Cortex of
Man.” Nature 340:6232 (1989), 386–389.
[706] E. Lueder. Liquid Crystal Displays: Addressing Schemes and Electro-Optical Effects. Chich-
ester, UK: John Wiley and Sons, Ltd., 2001.
[707] T. Lulé, B. Schneider, and M. Böhm. “Design and Fabrication of a High-Dynamic-Range
Image Sensor in TFA Technology.” IEEE Journal of Solid-State Circuits 34:5 (1999), 704–711.
[708] T. Lulé, M. Wagner, M. Verhoeven, H. Keller, and M. Böhm. “100000-Pixel, 120-dB Imager
in TFA Technology.” IEEE Journal of Solid-State Circuits 35:5 (2000), 732–739.
[709] J. S. Lund, R. D. Lund, A. E. Hendrickson, A. H. Bunt, and A. F. Fuchs. “The Origin of
Efferent Pathways from the Primary Visual Cortex, Area 17, of the Macaque monkey as Shown
by Retrogade Transport of Horseradish Peroxidase.” Journal of Comparative Neurology 164:3
(1975), 287–303.
[710] J. S. Lund. “Organization of Neurons in the Visual Cortex, Area 17, of the Monkey (Macaca
mulatta).” Journal of Comparative Neurology 147:4 (1973), 455–496.
[711] M. Luo, A. Clark, P. Rhodes, A. Schappo, S. Scrivner, and C. Tait. “Quantifying Colour
Appearance. Part I. LUTCHI Colour Appearance Data.” Color Research and Application 16
(1991), 166–180.
[712] M. Luo, A. Clark, P. Rhodes, A. Schappo, S. Scrivner, and C. Tait. “Quantifying Colour
Appearance. Part II. Testing Colour Models Performance using LUTCHI Color Appearance
Data.” Color Research and Application 16 (1991), 181–197.
[713] M. Luo, M.-C. Lo, and W.-G. Kuo. “The LLAB(l:c) Colour Model.” Color Research and
Application 21 (1996), 412–429.
[714] M. R. Luo, G. Cui, and B. Rigg. “The Development of CIE2000 Colour-Difference Formula:
CIEDE2000.” Color Research and Application 26:5 (2001), 340–350.
[715] R. Luther. “Aus dem Gebiet der Farbreizmetrik.” Zeitschrift für Technische Physik 8 (1927),
540–558.
[716] D. K. Lynch and W. Livingstone. Color and Light in Nature, Second edition. Cambridge, UK:
Cambridge University Press, 2001.
[717] D. L. MacAdam. Sources of Color Science. Cambridge, MA: MIT Press, 1970.
[718] D. L. MacAdam. “Uniform Color Scales.” Journal of the Optical Society of America 64 (1974),
1619–1702.
[719] D. L. MacAdam. “Colorimetric Data for Samples of OSA Uniform Color Scales.” Journal of
the Optical Society of America 68 (1978), 121–130.
[720] D. L. MacAdam. “Introduction.” In Selected Papers on Colorimetry — Fundamentals, edited
by D. L. MacAdam. Bellingham, Washington: SPIE Optical Engineering Press, 1993.
[721] L. W. MacDonald. “Gamut Mapping in Perceptual Color Space.” In Proceedings of the First
IS&T/SID Color Imaging Conference, pp. 193–196. Spriongfield, VA: Society for Imaging Sci-
ence and Technology, 1993.
[722] D. I. A. MacLeod and R. M. Boynton. “Chromaticity Diagram Showing Cone Excitation by
Stimuli of Equal Luminance.” Journal of the Optical Society of America 69:8 (1979), 1183–
1186.
[723] M. A. MacNeil and R. H. Masland. “Extreme Diversity among Amacrine Cells: Implications
for Function.” Neuron 20 (1998), 971–982.
[724] J. B. MacQueen. “Some Methods for Classification and Analysis of Multivariate Observa-
tions.” In Proceedings of the 5th Berkeley Symposium on Mathematical Statistics and Probabil-
ity, pp. 281–297. Berkeley: University of California Press, 1967.

i i

i i
 i i

i i

Bibliography 997

[725] M. Mahy. “Calculation of Color Gamuts Based on the Neugebauer Model.” Color Research
and Application 22 (1997), 365–374.
[726] R. Malach, R. B. Tootell, and D. Malonek. “Relationship between Orientation Domains, Cy-
tochrome Oxydase Strips, and Intrinsic Horizontal Connections in Squirrel Monkey Area V2.”
Cerebral Cortex 4:2 (1994), 151–165.
[727] R. Malladi and J. A. Sethian. “Image Processing via Level Set Curvature Flow.” Proceedings
of the National Academy of Science 92:15 (1995), 7046–7050.
[728] L. T. Maloney and B. A. Wandell. “Color Constancy: A Method for Recovering Surface
Spectral Reflectance.” Journal of the Optical Society of America A 3:1 (1986), 29–33.
[729] L. T. Maloney. “Evaluation of Linear Models of Surface Spectral Reflectance with Small
Number of Parameters.” Journal of the Optical Society of America A 3 (1986), 1673–1683.
[730] Étienne. L. Malus. “Sur la mesure du pouvoir réfringent des corps opaques.” Journal de l’École
Polytechnique 15:8 (1809), 219–228.
[731] S. Mann and R. W. Picard. “On Being ‘Undigital’ with Digital Cameras: Extending Dy-
namic Range by Combining Differently Exposed Pictures.” In IS&T’s 48th Annual Conference,
pp. 422–428. Springfield, VA: Society for Imaging Science and Technology, 1995.
[732] R. J. W. Mansfield. “Primate Photopigments and Cone Mechanisms.” In The Visual System,
edited by A. Fein and J. S. Levine, pp. 89–106. New York: Alan R Liss, 1985.
[733] R. Mantiuk, A. Efremov, K. Myszkowski, and H.-P. Seidel. “Backward Compatible High
Dynamic Range MPEG Video Compression.” ACM Transactions on Graphics 25:3 (2006),
713–723.
[734] E. W. Marchand. “Gradient Index Lenses.” In Progress in Optics, XI, edited by E. Wolf, XI,
pp. 305–337. Amsterdam: North Holland, 1973.
[735] R. F. Marcinik and K. M. Revak. “Effects of New Cool White Fluorescent Lamps on Viewing
and Measuring Color.” Textile Chemist and Colorist 30:1 (1998), 14–16.
[736] D. Marcuse. Light Transmission Optics, Second edition. New York: Van Nostrand Reinhold
Company, 1982.
[737] D. Marini and A. Rizzi. “A Computational Approach to Color Adaptation Effects.” Image and
Vision Computing 18:13 (2000), 1005–1014.
[738] D. Marini, A. Rizzi, and C. Carati. “Color Constancy Effects Measurement of the Retinex The-
ory.” In Proceedings of the IS&T/SPIE Conference on Electronic Imaging, pp. 23–29. Spring-
field, VA: Society for Imaging Science and Technology, 1999.
[739] K. Z. Markov. “Elementary Micromechanisms of Heterogeneous Media.” In Heterogeneous
Media: Modelling and Simulation, edited by K. Z. Markov and L. Preziosi, pp. 1–162. Boston:
Birkhauser, 1999.
[740] T. J. Marks, J. G. C. Veinot, J. Cui, H. Yan, A. Wang, N. L. Edlerman, J. Ni, Q. Huang, P. Lee,
and N. R. Armstrong. “Progress in High Work Function TCO OLED Anode Alternatives and
OLED Nanopixelation.” Synthetic Metals 127 (2002), 29–35.
[741] G. H. Markstein. Nonsteady Flame Propagation. Oxford, UK: Pergamon, 1964.
[742] D. Marr. Vision, A Computational Investigation into the Human Representation and Processing
of Visual Information. San Fransisco: W. H. Freeman and Company, 1982.
[743] J. S. Marsh and D. A. Plant. “Instrumental Colour Match Prediction using Organic Pigments.”
Journal of the Oil and Colour Chemists Association 47 (1964), 554–575.
[744] S. R. Marshner, S. H. Westin, E. P. F. Lafortune, K. E. Torrance, and D. P. Greenberg. “Image-
Based BRDF Measurement Including Human Skin.” In Proceedings of the 10th Eurographics
Workshop on Rendering, pp. 139–152. Aire-la-Ville, Switzerland: Eurographics Association,
1999.

i i

i i
 i i

i i

998 Bibliography

[745] S. R. Marshner, S. H. Westin, E. P. F. Lafortune, and K. Torrance. “Image-Based Measurement

of the Bidirectional Reflectance Distribution Function.” Applied Optics 39:16 (2000), 2592–
2600.
[746] P. R. Martin, B. B. Lee, A. J. R. White, S. G. Solomon, and L. Rüttiger. “Chromatic Sensitivity
of Ganglion Cells in the Peripheral Primate Retina.” Nature 410:6831 (2001), 933–936.
[747] K. A. C. Martin. “Microcircuits in Visual Cortex.” Current Opinion in Neurobiology 12:4
(2002), 418–425.
[748] M. Mase, S. Kawahito, M. Sasaki, and Y. Wakamori. “A 19.5b Dynamic Range CMOS Image
Sensor with 12b Column-Parallel Cyclic A/D Converters.” In IEEE International Solid-State
Circuits Conference (ISSCC), pp. 350–351, 2005.
[749] R. H. Masland. “Processing and Encoding of Visual Information in the Retina.” Current
Opinion in Neurobiology 6 (1996), 467–474.
[750] M. Massimi. Pauli’s Exclusion Principle: The Origin and Validation of a Scientific Principle.
Cambridge, UK: Cambridge University Press, 2005.
[751] W. Matusik, H. Pfister, M. Brand, and L. McMillan. “A Data-Driven Reflectance Model.” ACM
Transactions on Graphics 22:3 (2003), 759–769.
[752] W. Matusik, H. Pfister, M. Brand, and L. McMillan. “Efficient Isotropic BRDF Measurement.”
In Proceedings of the 14th Eurographics Symposium on Rendering, pp. 241–316. Aire-la-Ville,
Switzerland: Eurographics Association, 2003.
[753] J. H. R. Maunsell and D. C. van Essen. “The Connections of the Middle Temporal Visual
Area (MT) and Their Relationship to a Cortical Hierarchy in the Macaque Monkey.” Journal of
Neuroscience 3:12 (1983), 2563–2586.
[754] J. H. R. Maunsell and W. T. Newsome. “Visual Processing in Monkey Extrastriate Cortex.”
Annual Review of Neuroscience 10 (1987), 363–401.
[755] D. M. Maurice. “The Structure and Transparency of the Cornea.” Journal of Physiology 136:2
(1957), 263–286.
[756] J. C. Maxwell. “Experiments on Color, as Perceived by the Eye, with Remarks on Color-
Blindness.” Transactions of the Royal Society of Edinburgh 21 (1855), 275–298.
[757] C. S. McCamy, H. Marcus, and J. G. Davidson. “A Color Rendition Chart.” Journal of Applied
Photographic Engineering 2:3 (1976), 95–99.
[758] J. J. McCann and A. Rizzi. “Veiling Glare: The Dynamic Range Limit of HDR Images.” In
Human Vision and Electronic Imaging XII, 6492, 6492. SPIE, 2007.
[759] J. McCann. “Lessons Learned from Mondrians Applied to Real Images and Color Gamuts.” In
Proceedings of the IS&T Seventh Color Imaging Conference, pp. 1–8. Springfield, VA: Society
for Imaging Science and Technology, 1999.
[760] E. J. McCartney. Optics of the Atmosphere: Scattering by Molecules and Particles. New York:
John Wiley and Sons, Inc., 1976.
[761] W. R. McCluney. Introduction to Radiometry and Photometry. Boston: Artech House Publish-
ers, 1994.
[762] C. McCollough. “Color Adaptation of Edge Detectors in the Human Visual System.” Science
149:3688 (1965), 1115–1116.
[763] C. McCollough. “Do McCollough Effects Provide Evidence for Global Pattern Processing?”
Perception and Psychophysics 62:2 (2000), 350–362.
[764] M. McCool, J. Ang, and A. Ahmad. “Homomorphic Factorization of BRDFs for High-
Performance Rendering.” In Proceedings of SIGGRAPH 2001, Computer Graphics Proceed-
ings, Annual Conference Series, pp. 171–178. Resding, MA: Addison-Wesley, 2001.

i i

i i
 i i

i i

Bibliography 999

[765] R. McDonald. “Acceptibility and Perceptibility Decisions using the CMC Colour Difference
Formula.” Textile Chemist and Colorist 20:6 (1988), 31–31.
[766] D. McIntyre. Colour Blindness: Causes and Effects. Austin, TX: Dalton Publishing, 2002.
[767] K. McLaren. The Colour Science of Dyes and Pigments, Second edition. Bristol, UK: Adam
Hilger Ltd., 1986.
[768] D. McQuarrie. Quantum Chemistry. Mill Valley, CA: University Science Books, 1983.
[769] H. McTavish. “A Demonstration of Photosynthetic State Transitions in Nature.” Photosynthesis
Research 17 (1988), 247–247.
[770] C. A. Mead. “Operation of Tunnel-Emission Devices.” Journal of Applied Physics 32:4 (1961),
646–652.
[771] J. A. Medeiros. Cone Shape and Color Vision: Unification of Structure and Perception.
Blountsville, AL: Fifth Estate Publishers, 2006.
[772] J. Mellerio. “Yellowing of the Human Lens: Nuclear and Cortical Contributions.” Vision
Research 27:9 (1987), 1581–1587.
[773] C. E. Metz. “Basic Principles of ROC Analysis.” Seminars in Nuclear Medicine 8 (1978),
283–298.
[774] L. Meylan and S. Süsstrunk. “Color Image Enhancement using a Retinex-Based Adaptive Fil-
ter.” In Proceedings of the 2nd IS&T European Conference on Color in Graphics, Image, and
Vision (CGIV 2004), pp. 359–363. Springfield, VA: Society for Imaging Science and Technol-
ogy, 2004.
[775] L. Meylan and S. Süsstrunk. “High Dynamic Range Image Rendering with a Retinex-Based
Adaptive Filter.” IEEE Transactions on Image Processing 15:9 (2006), 2820–2930.
[776] L. Meylan, S. Daly, and S. Süsstrunk. “The Reproduction of Specular Highlights on High
Dynamic Range Displays.” In Proceedings of the 14th IS&T/SID Color Imaging Conference.
Springfield, VA: Society for Imaging Science and Technology, 2006.
[777] L. Meylan, D. Alleysson, and S. Süsstrunk. “A Model of Retinal Local Adaptation for the Tone
Mapping of CFA Images.” Journal of the Optical Society of America A.
[778] L. Meylan, S. Daly, and S. Süsstrunk. “Tone Mapping for High Dynamic Range Displays.” In
Proceedings of SPIE 6492 (Human Vision and Electronic Imaging XII). Belingham, WA: SPIE,
2007.
[779] G. Mie. “Beiträge zur Optik trüber Medien, speziell kolloidaler Metallösungen.” Annalen der
Physik 25:4 (1908), 377–377.
[780] N. J. Miller, P. Y. Ngai, and D. D. Miller. “The Application of Computer Graphics in Lighting
Design.” Journal of the IES 14 (1984), 6–26.
[781] J. Millman. “Electronic Energy Bands in Metallic Lithium.” Physical Review 47:4 (1935),
286–290.
[782] W. I. Milne and K. B. K. Teo. “Growth and Characterization of Carbon Nanotubes for Field
Emission Display Applications.” Journal of the Society for Information Display 12 (2004),
289–292.
[783] M. G. J. Minnaert. Light and Color in the Outdoors. New York: Springer-Verlag, 1993.
[784] T. Mitsunaga and S. K. Nayar. “Radiometric Self Calibration.” In International Conference on
Computer Vision and Pattern Recognition, pp. 374–380. Washington, DC: IEEE Press, 1999.
[785] J. D. Mollon and J. K. Bowmaker. “The Spatial Arrangement of Cones in the Primate Retina.”
Nature 360:6405 (1992), 677–679.
[786] J. D. Mollon and G. Estévez. “Tyndall’s Paradox of Hue Discrimination.” Journal of the
Optical Society of America A 5 (1988), 151–159.

i i

i i
 i i

i i

1000 Bibliography

[787] J. D. Mollon. “‘Tho’ She Kneel’d in that Place where They Grew...’, The Uses and Origins of
Primate Colour Vision.” Journal of Experimental Biology 146 (1989), 21–38.
[788] F. M. De Monasterio and P. Gouras. “Functional Properties of Ganglion Cells of the Rhesus
Monkey Retina.” Journal of Physiology (London) 251 (1975), 167–195.
[789] P. Moon and D. E. Spencer. “On the Stiles-Crawford effect.” Journal of the Optical Society of
America 34:6 (1944), 319–329.
[790] G. Mori and J. Malik. “Recognizing Objects in Adversarial Clutter: Breaking a Visual
Captcha.” In Proceedings of CVPR, pp. 134–141. Los Alamitos, CA: IEEE Computer Soci-
ety, 2003.
[791] N. Moroney and I. Tastl. “A Comparison of Retinex and iCAM for Scene Rendering.” Journal
of Electronic Imaging 13:1 (2004), 139–145.
[792] N. Moroney. “Chroma Scaling and Crispening.” In IS&T/SID 9th Color Imaging Conference,
pp. 97–101. Springfield, VA: Society for Imaging Science and Technology, 2001.
[793] N. Moroney. “A Hypothesis Regarding the Poor Blue Constancy of CIELAB.” Color Research
and Application 28 (2003), 371–378.
[794] J. Morovic and M. R. Luo. “Gamut Mapping Algorithms Based on Psychophysical Experi-
ment.” In Proceedings of the 5th IS&T/SID Color Imaging Conference, pp. 44–49. Springfield,
VA: Society for Imaging Science and Technology, 1997.
[795] J. Morovic and M. R. Luo. “The Fundamentals of Gamut Mapping: A Survey.” Journal of
Imaging Science and Technology 45:3 (2001), 283–290.
[796] J. Morovic and P. L. Sun. “The Influence of Image Gamuts on Cross-Media Colour Image
Reproduction.” In Proceedings of the 8th IS&T/SID Color Imaging Conference, pp. 324–329.
Springfield, VA: Society for Imaging Science and Technology, 2000.
[797] J. Morovic. “To Develop a Universal Gamut Mapping Algorithm.” Ph.D. thesis, University of
Derby, Derby, UK, 1998.
[798] J. Morovic. “Gamut Mapping.” In Digital Color Imaging Handbook, edited by G. Sharma,
Chapter 10, pp. 639–685. Boca Raton, FL: CRC Press, 2003.
[799] B. Moulden and F. A. A. Kingdom. “White’s Effect: A Dual Mechanism.” Vision Research 29
(1989), 1245–1236.
[800] K. Moutoussis and S. Zeki. “Responses of Spectrally Selective Cells in Macaque Area V2 to
Wavelengths and Colors.” Journal of Neurophysiology 87:4 (2002), 2104–2112.
[801] K. T. Mullen and F. A. A. Kingdom. “Losses in Periferal Colour Sensitivity Predicted from
“Hit and Miss” Post-Receptoral Cone Connections.” Vision Research 36:13 (1996), 1995–2000.
[802] C. D. Müller, A. Falcou, N. Reckefuss, M. Rojahn, V. Wiederhirn, P. Rudati, H. Frohne,
O. Nuyken, H. Becker, and K. Meerholtz. “Multi-Colour Organic Light Emitting Displays by
Solution Processing.” Nature 421 (2003), 829–833.
[803] H. Müller. “Über die entoptische Wahrnehmung der Netzhautgefässe, insbesondere als Be-
weismittel für die Lichtperzeption durch die nach hinten gelegenen Netzhautelemente.” Ver-
handlungen der Physikalisch Medizinische Gesellschaft in Würzburg 5 (1854), 411–411.
[804] J. C. Mullikin, L. J. van Vliet, H. Netten, F. R. Boddeke, G. van der Feltz, and I. T. Young.
“Methods for CCD Camera Characterization.” In Image Acquisition and Scientific Imaging
Systems, SPIE vol. 2173, edited by H. C. Titus and A. Waks, SPIE vol. 2173, pp. 73–84, 1994.
[805] A. H. Munsell. A Color Notation, First edition. Baltimore, MD: Munsell Color Company,
1905.
[806] A. H. Munsell. Atlas of the Munsell Color System. Malden, MA: Wadsworth-Howland and
Company, 1915.

i i

i i
 i i

i i

Bibliography 1001

[807] A. H. Munsell. Munsell Book for Color. Baltimore, MD: Munsell Color Company, 1929.
[808] A. H. Munsell. A Color Notation: An Illustrated System Defining All Colors and Their Rela-
tions by Measured Scales of Hue, Value, and Chroma, Ninth edition. Baltimore, MD: Munsell
Color Company, 1941.
[809] Y. Muramatsu, S. Kurosawa, M. Furumiya, H. Ohkubo, and Y. Nakashiba. “A Signal-
Processing CMOS Image Sensor using a Simple Analog Operation.” In IEEE International
Solid-State Circuits Conference (ISSCC), pp. 98–99, 2001.
[810] S. Musa. “Active-Matrix Liquid-Crystal Displays.” Scientific American November.
[811] K. I. Naka and W. A. H. Rushton. “S-Potentials from Luminosity Units in the Retina of Fish
(Cyprinidae).” Journal of Physiology (London) 185:3 (1966), 587–599.
[812] J. Nakamura. “Basics of Image Sensors.” In Image Sensors and Signal Processing for Digital
Still Cameras, edited by J. Nakamura, pp. 53–93. Boca Raton, FL: Taylor and Francis, 2006.
[813] J. Nakamura, editor. Image Sensors and Signal Processing for Digital Still Cameras. Boca
Raton, FL: Taylor and Francis, 2006.
[814] K. Nassau. “Fundamentals of Color Science.” In Color for Science, Art and Technology, edited
by K. Nassau. Amsterdam: Elsevier Science B.V., 1998.
[815] K. Nassau. The Physics and Chemistry of Color: The Fifteen Causes of Color. New York:
John Wiley and Sons, 2001.
[816] S. K. Nayar and V. Branzoi. “Adaptive Dynamic Range Imaging: Optical Control of Pixel
Exposures over Space and Time.” In International Conference on Computer Vision, Volume 2,
pp. 1168–1175. Washington, DC: IEEE Press, 2003.
[817] S. K. Nayar and T. Mitsunaga. “High Dynamic Range Imaging: Spatially Varying Pixel Ex-
posures.” In International Conference on Computer Vision and Pattern Recognition, Volume 1,
pp. 472–479. Washington, DC: IEEE Press, 2000.
[818] S. K. Nayar, G. Krishnan, M. D. Grossberg, and R. Raskar. “Fast Separation of Direct and
Global Components of a Scene using High Frequency Illumination.” ACM Transactions on
Graphics 25:3 (2006), 935–944.
[819] Y. Nayatani, K. Takahama, and H. Sobagaki. “Formulation of a Nonlinear Model of Chromatic
Adaptation.” Color Research and Application 6 (1981), 161–171.
[820] Y. Nayatani, K. Takahama, H. Sobagaki, and J. Hirono. “On Exponents of a Nonlinear Model
of Chromatic Adaptation.” Color Research and Application 7 (1982), 34–45.
[821] Y. Nayatani, T. Mori, K. Hashimoto, and H. Sobagaki. “Comparison of Color Appearance
Models.” Color Research and Application 15 (1990), 272–284.
[822] Y. Nayatani. “Proposal of an Opponent Colors System Based on Color Appearance and Color
Vision Studies.” Color Research and Application 29 (2004), 135–150.
[823] M. Neitz, J. Neitz, and G. H. Jacobs. “Spectral Tuning of Pigments Underlying Red-Green
Color Vision.” Science 252 (1991), 971–974.
[824] J. Neitz, J. Carroll, Y. Yamauchi, M. Neitz, and D. R. Williams. “Color Perception Is Mediated
by a Plastic Neural Mechanism that Is Adjustable in Adults.” Neuron 35:4 (2002), 783–792.
[825] R. Nelson and H. Kolb. “A17: A Broad-Field Amacrine Cell of the Rod System in the Retina
of the Cat.” Joural of Neurophysiology 54 (1985), 592–614.
[826] R. Nelson. “Visual Responses of Ganglion Cells.”, 2006. Webvision Article. Available online
(http://webvision.med.utah.edu).
[827] F. L. van Ness and M. A. Bouman. “Spatial Modulation Transfer in the Human Eye.” Journal
of the Optical Society of America 57 (1967), 401–406.

i i

i i
 i i

i i

1002 Bibliography

[828] H. E. J. Neugebauer. “Die theoretischen Grundlagen des Mehrfarbenbuchdrucks.” Zeitschrift

für wissenschaftliche Photographie, Photophysik und Photochemie 36:4 (1937), 73–89.
[829] H. J. Neugebauer. “Clausius-Mosotti Equation for Anisotropic Crystals.” Physical Review 88:5
(1952), 1210–1210.
[830] H. Neumann, L. Pessoa, and T. Hansen. “Visual Filling-In for Computing Perceptual Surface
Properties.” Biological Cybernetics 85 (2001), 355–369.
[831] S. Newhall. “Preliminary Report of the O.S.A. Subcommittee on the Spacing of the Munsell
Colors.” Journal of the Optical Society of America 30 (1940), 617–645.
[832] R. G. Newton. Scattering Theory of Waves and Particles, Second edition. Mineola, NY: Dover
Publications, 2002.
[833] M. Nezamabadi and R. S. Berns. “The Effect of Image Size on the Color Appearance of Image
Reproduction using Colorimetrically Calibrated LCD and DLP Displays.” Journal of the Society
for Information Display 14:9 (2006), 773–783.
[834] R. Ng, M. Levoy, M. Bredif, G. Duval, M. Horowitz, and P. Hanrahan. “Light Field Photogra-
phy with a Hand-Held Plenoptic Camera.”, 2005.
[835] D. Q. Nguyen, R. Fedkiw, and H. W. Jensen. “Physically Based Modeling and Animation of
Fire.” ACM Transactions on Graphics 21:3 (2002), 721–728.
[836] D. Nickerson. “Munsell Renotations for Samples of OSA Uniform Color Scales.” Journal of
the Optical Society of America 68 (1978), 1343–1347.
[837] D. Nickerson. “OSA Uniform Color Scales Samples — A Unique Set.” Colour Research and
Application 6 (1981), 7–33.
[838] F. E. Nicodemus, J. C. Richmond, J. J. Hsia, I. W. Ginsberg, and T. Limperis. “Geometrical
Considerations and Nomenclature for Reflectance.” Technical Report Monograph 160, National
Bureau of Standards (US), Gaithersburg, MD, 1977.
[839] F. E. Nicodemus, editor. Self-Study Manual on Optical Radiation Measurements. National
Bureau of Standards, 1976. Technical Note 910-1.
[840] C. L. Nikias and A. P. Petropulu. Higher-Order Spectra Analysis. Signal Processing Series,
Englewood Cliffs, NJ: Prentice Hall, 1993.
[841] M. Nonako, T. Kashmina, and Y. Knodo. “A 5-Meter Integrating Sphere.” Applied Optics 6:4
(1967), 757–771.
[842] V. O’Brien. “Contour Perception, Illusion and Reality.” Journal of the Optical Society of
America 49 (1958), 112–119.
[843] S. Ogata, J. Ishida, and T. Sasano. “Optical Sensor Array in an Artificial Compound Eye.”
Optical Engineering 33:11 (1994), 3649–3655.
[844] M. P. Ogren and A. E. Hendrickson. “Pathways between Striate Cortex and Subcortical Re-
gions in Macaca mulatta and Saimiri sciureus: Evidence for a Reciprocal Pulvinar Connection.”
Journal of Comparative Neurology 53 (1976), 780–800.
[845] T. Oguchi, E. Yamaguchi, K. Sazaki, K. Suzuki, S. Uzawa, and K. Hatanaka. “A 36 Inch
Surface-Conduction Electron-Emitter Display (SED).” In Technical Digest of the Society for In-
formation Display (SID), pp. 1929–1931. San Jose, CA: Society for Information Display, 2005.
[846] K. Ohki, S. Chung, Y. H. Ch’ng, P. Kara, and R. C. Reid. “Functional Imaging with Cellular
Resolution Reveals Precise Micro-Architecture in Visual Cortex.” Nature 433:7026 (2005),
597–603.
[847] K. Ohno and T. Kusunoki. “The Effect of Ultrafine Pigment Color Filters on Cathode Ray Tube
Brightness, Contrast, and Color Purity.” Journal of the Electrochemical Society 143:3 (1996),
1063–1067.

i i

i i
 i i

i i

Bibliography 1003

[848] Y. Ohno. “Spectral Design Considerations for White LED Color Rendering.” Optical Engi-
neering 44:11 (2005), 111302–1 – 111302–9.
[849] N. Ohta and A. R. Robertson. Colorimetry: Fundamentals and Applications. Chichester, UK:
John Wiley and Sons, 2005.
[850] R. Oi and K. Aizawa. “Wide Dynamic Range Imaging by Sensitivity Adjustable CMOS Image
Sensor.” In Proceedings of the IEEE International Conference on Image Processing (ICIP),
pp. 583–586, 2003.
[851] H. W. Ok, S. D. Lee, W. H. Choe, D. S. Park, and C. Y. Kim. “Color Processing for
Multi-Primary Display Devices.” In IEEE International Conference on Image Processing 2005,
pp. 980–983. Washington, DC: IEEE, 2005.
[852] K. Okinaka, A. Saitoh, T. Kawai, A. Senoo, and K. Ueno. “High Performable Green-Emitting
OLEDs.” SID Symposium Digest of Technical Papers 35:1 (2004), 686–689.
[853] B. A. Olshausen and D. J. Field. “Emergence of Simple-Cell Receptive Field Properties by
Learning a Sparse Code for Natural Images.” Nature 381 (1996), 607–609.
[854] B. A. Olshausen and D. J. Field. “How Close Are We to Understanding V1?” Neural Compu-
tation 17:8 (2005), 1665–1699.
[855] W. O’Mara. Liquid Crystal Flat Panel Displays: Manufacturing Science and Technology. New
York: Van Nostrand Reinhold, 1993.
[856] A. V. Oppenheim, R. Schafer, and T. Stockham. “Nonlinear Filtering of Multiplied and Con-
volved Signals.” Proceedings of the IEEE 56:8 (1968), 1264–1291.
[857] A. V. Opppenheim and R. Schafer. Digital Signal Processing. Englewood Cliffs, NJ: Prentice
Hall, 1975.
[858] A. V. Opppenheim, R. Schafer, and J. R. Buck. Discrete-Time Signal Processing, Second
edition. Englewood Cliffs, NJ: Prentice-Hall, 1999.
[859] M. Oren and S. K. Nayar. “Generalization of Lambert’s Reflectance Model.” In Proceed-
ings of SIGGRAPH ’94, Computer Graphics Proceedings, Annual Conference Series, edited by
A. Glassner, pp. 239–246. New York: ACM Press, 1994.
[860] OSA. The Science of Color. Washington: Optical Society of America, 1963.
[861] D. Osorio, D. L. Ruderman, and T. W. Cronin. “Estimation of Errors in Luminance Signals
Encoded by Primate Retina Resulting from Sampling of Natural Images with Red and Green
Cones.” Journal of the Optical Society of America A 15:1 (1998), 16–22.
[862] G. Østerberg. “Topography of the Layer of Rods and Cones in the Human Retina.” Acta
Ophthalmologica, Supplement 6 (1935), 1–103.
[863] V. Ostromoukhov, C. Donohue, and P. Jodoin. “Fast Hierarchical Importance Sampling with
Blue Noise Properties.” ACM Transactions on Graphics 23:3 (2004), 488–495.
[864] V. Ostromoukhov. “Chromaticity Gamut Enhancement by Heptatone Multi-Color Printing.” In
Proceedings IS&T/SPIE 1909, 1993 International Symposium on Electronic Imaging: Science
& Technology, pp. 139–151. Bellingham, WA: SPIE, 1993.
[865] C. A. Padgham and J. E. Saunders. The Perception of Light and Color. New York: Academic
Press, 1975.
[866] R. H. Page, K. I. Schaffers, P. A. Waide, J. B. Tassano, S. A. Payne, W. F. Krupke, and W. K.
Bischel. “Upconversion-Pumped Luminescence Efficiency of Rare-Earth-Doped Hosts Sen-
sitized with Trivalent Ytterbium.” Journal of the Optical Society of America B 15:3 (1998),
996–1008.
[867] G. Palmer. Theory of Colors and Vision. London: Leacroft, 1777. Reprinted in David L
MacAdam (ed.), Selected Papers on Colorimetry — Fundamentals, SPIE Optical Engineering
Press, 1993.

i i

i i
 i i

i i

1004 Bibliography

[868] S. Palmer. Vision Science: Photons to Phenomenology. Cambridge MA: MIT Press, 1999.
[869] M. A. Paradiso and K. Nakayama. “Brightness Perception and Filling-In.” Vision Research 31
(1991), 1221–1236.
[870] D. A. Pardo, G. E. Jabbour, and N. Peyghambarian. “Application of Screen Printing in the
Fabrication of Organic Light Emitting Diodes.” Advanced Materials 12:17 (2000), 1249–1252.
[871] S. Paris and F. Durand. “A Fast Approximation of the Bilateral Filter using a Signal Process-
ing Approach.” In Computer Vision- ECCV 2006, Lecture Notes in Computer Science. Berlin:
Springer-Verlag, 2006.
[872] S. H. Park and E. D. Montag. “Evaluating Tone Mapping Algorithms for Rendering Non-
Pictorial (Scientific) High-Dynamic-Range Images.” Journal of Visual Communication and Im-
age Representation 18:5 (2007), 415–428.
[873] S. C. Park and R. R. Shannon. “Zoom Lens Design using Lens Modules.” Optical Engineering
35:6 (1996), 1668–1676.
[874] R. H. Park and E. I. Stearns. “Spectrophotometric Formulation.” Journal of the Optical Society
of America 34:2 (1944), 112–113.
[875] S. Parker, W. Martin, P.-P. Sloan, P. Shirley, B. Smits, and C. Hansen. “Interactive Ray Tracing.”
In Symposium on Interactive 3D Computer Graphics, pp. 119–126. New York: ACM Press,
1999.
[876] S. Parker, M. Parker, Y. Livnat, P.-P. Sloan, C. Hansen, and P. Shirley. “Interactive Ray Tracing
for Volume Visualization.” IEEE Transactions on Visualization and Computer Graphics 5:3
(1999), 238–250.
[877] A. C. Parr. “A National Measurement System for Radiometry, Photometry, and Pyrometry
based upon Absolute Detectors, (Version 1.0) [Online].” Originally published as: NIST Techni-
cal Note 1421, September 1996. Available online (http://physics.nist.gov/Pubs/TN1421/contents.
html).
[878] C. A. Párraga, G. Brelstaff, T. Troscianko, and I. R. Moorhead. “Color and Luminance Infor-
mation in Natural Scenes.” Journal of the Optical Society of America A 15:3 (1998), 563–569.
[879] C. A. Párraga, T. Troscianko, and D. J. Tolhurst. “The Human Visual System Is Optimised
for Processing the Spatial Information in Natural Visual Images.” Current Biology 10:1 (2000),
35–38.
[880] K. A. Parulski. “Color Filter Arrays and Processing Alternatives for One-Chip Cameras.” IEEE
Transactions on Electron Devices 32:8 (1985), 1381–1389.
[881] E. Parzen. “On Estimation of a Probability Density and Mode.” Annals of Mathematical
Statistics 33 (1962), 1065–1076.
[882] D. Pascale. “A Review of RGB Color Spaces.” Technical report, The BabelColor Company,
2003. Available online (http://www.babelcolor.com).
[883] S. N. Pattanaik and H. Yee. “Adaptive Gain Control for High Dynamic Range Image Display.”
In Proceedings of Spring Conference in Computer Graphics (SCCG2002), pp. 24–27. New York:
ACM Press, 2002.
[884] S. N. Pattanaik, J. A. Ferwerda, M. D. Fairchild, and D. P. Greenberg. “A Multiscale Model
of Adaptation and Spatial Vision for Realistic Image Display.” In Proceedings SIGGRAPH
’98, Computer Graphics Proceedings, Annual Conference Series, pp. 287–298. Reading, MA:
Addison-Wesley, 1998.
[885] S. N. Pattanaik, J. Tumblin, H. Yee, and D. P. Greenberg. “Time-Dependent Visual Adaptation
for Fast Realistic Display.” In Proceedings SIGGRAPH 2000, Computer Graphics Proceedings,
Annual Conference Series, pp. 47–54. Reading, MA: Addison-Wesley, 2000.

i i

i i
 i i

i i

Bibliography 1005

[886] W. Pauli. “Über den Zusammenhang des Abschlusses von Elektronengruppen im Atom mit der
Komplexstructure der Spektren.” Zeitschrift für Physik 31 (1925), 765–783.
[887] A. L. Pearlman, J. Birch, and J. C. Meadows. “Cerebral Color Blindness: An Acquired Defect
in Hue Discrimination.” Annals of Neurology 5:3 (1979), 253–261.
[888] E. Peli. “Contrast in Complex Images.” Journal of the Optical Society of America 7:10 (1990),
2032–2040.
[889] D. J. Perkel, J. Bullier, and H. Kennedy. “Topography of the Afferent Connectivity of Area 17
in the Macaque Monkey: A Double Labelling Study.” Journal of Comparative Neurology 253:3
(1986), 374–402.
[890] P. Perona and J. Malik. “Scale-Space and Edge Detection using Anisotropic Diffusion.” IEEE
Transactions on Pattern Analysis and Machine Intelligence 12:7 (1990), 629–639.
[891] V. H. Perry, R. Oehler, and A. Cowey. “Retinal Ganglion Cells that Project to the Dorsal Lateral
Geniculate Nucleus in the Macaque Monkey.” Neuroscience 12:4 (1984), 1101–1123.
[892] B. B. Petersen and D. M. Dacey. “Morphology of Retinal Ganglion Ccells with Intra-Retinal
Axon Collaterals.” Visual Neuroscience 15:2 (1998), 377–387.
[893] M. Petrou and P. Bosdogianni. Image Processing: The Fundamentals. Chichester, UK: John
Wiley and Sons, Ltd., 1999.
[894] M. Pfeiffer, S. R. Forrest, X. Zhou, and K. Leo. “A Low Drive Voltage, Transparent, Metal-Free
n-i-p Electrophosphorescent Light Emitting Diode.” Organic Electronics 4 (2003), 21–26.
[895] M. Pharr and G. Hunphreys. Physically Based Rendering: From Theory to Implementation.
San Francisco, CA: Morgan Kaufmann Publishers, 2004.
[896] M. L. F. Phillips, M. P. Hehlen, K. Nguyen, J. M. Sheldon, and N. J. Cockroft. “Upconversion
Phosphors: Recent Advances and New Applications.” In Physics and Chemistry of Lumines-
cent Materials. Proceedings of the Eighth International Symposium (Electrochemical Society
Proceedings Vol 99-40), pp. 123–129. Pennington, NJ: The Electrochemical Society, 2000.
[897] B. T. Phong. “Illumination for Computer Generated Pictures.” Communications of the ACM
18:6 (1975), 311–317.
[898] B. Pinna, G. Brelstaff, and L. Spillmann. “Surface Color from Boundaries: A New ‘Watercolor’
Illusion.” Vision Research 41 (2001), 2669–2676.
[899] M. H. Pirenne. “Dark-Adaptation and Night Vision.” In The Eye, Volume 2, edited by
H. Davson, pp. 93–122. New York: Academic Press, 1962.
[900] M. H. Pirenne. “Liminal Brightness Increments.” In The Eye, Volume 2, edited by H. Davson,
pp. 159–174. New York: Academic Press, 1962.
[901] M. H. Pirenne. Vision and the Eye, Second edition. London: Associated Book Publishers,
1967.
[902] F. H. G. Pitt and E. W. H. Selwyn. “Colour of Outdoor Photographic Objects.” The Photo-
graphic Journal 78 (1938), 115–121.
[903] F. H. G. Pitt. “Characteristics of Dichromatic Vision with an Appendix on Anomalous Trichro-
matic Vision.” In Great Britain Medical Report Series, 200. London: His Majesty’s Stationery
Office, 1935.
[904] D. G. Pitts. “The Ocular Effects of Ultraviolet Radiation.” American Journal of Optometry and
Physiological Optics 55 (1978), 19–35.
[905] G. N. Plass, T. J. Humphreys, and G. W. Kattawar. “Color of the Ocean.” Applied Optics 17:9
(1978), 1432–1446.
[906] W. T. Plummer. “Photographic Shutters: Better Pictures with a Reconsideration of Shutter
Efficiency.” Applied Optics 16:7 (1917), 1914–1917.

i i

i i
 i i

i i

1006 Bibliography

[907] I. Pobbaravski. “Methods of Computing Inks Amount to Produce a Scale of Neutrals for
Photomechanical Reproduction.” In Proceedings of the Technical Association of the Graphical
Arts, pp. 10–33. Sewickley, PA: Technical Association of the Graphic Arts, 1966.
[908] A. B. Poirson and B. A. Wandell. “The Appearance of Colored Patterns: Pattern-Color Sepa-
rability.” Journal of the Optical Society of America A 10:12 (1993), 2458–2471.
[909] A. Poirson and B. Wandell. “Pattern-Color Separable Pathways Predicts Sensitivity to Simple
Colored Patterns.” Vision Research 35:2 (1996), 239–254.
[910] J. Pokorny, V. C. Smith, G. Verriest, and A. J. L. G. Pinkers. Congenital and Acquired Color
Vision Defects. New York: Grune and Stratton, 1979.
[911] J. Pokorny, V. C. Smith, and M. Wesner. “Variability in Cone Populations and Implications.”
In From Pigments to Perception, edited by A. Valberg and B. B. Lee, pp. 23–34. New York:
Plenum, 1991.
[912] F. Pollak. “Masking for Halftone.” Journal of Photographic Science 3 (1955), 180–188.
[913] F. Pollak. “The Relationship between the Densities and Dot Size of Halftone Multicolor Im-
ages.” Journal of Photographic Science 3 (1955), 112–116.
[914] F. Pollak. “New Thoughts on Halftone Color Masking.” Penrose Annual 50 (1956), 106–110.
[915] S. L. Polyak. The Vertebrate Visual System. Chicago, IL: Chicago University Press, 1957.
[916] M. Ponzo. “Rapports entre quelque illusions visuelles de contraste angulaire et l’appréciation
de grandeur des astres à l’horizon.” Archives Italiennes de Biologie 58 (1913), 327–329.
[917] R. M. Pope and E. S. Fry. “Absorption Spectrum (380–700 nm) of Pure Water. II. Integrating
Cavity Measurements.” Applied Optics 36:33 (1997), 8710–8723.
[918] D. Pope. “The Elusive Blue Laser.” The Industrial Physicist 3:3 (1997), 16.
[919] J. Portilla and E. P. Simoncelli. “A Parametric Texture Model based on Joint Statistics of
Complex Wavelet Coefficients.” International Journal of Computer Vision 40:1 (2000), 49–71.
Available online ({www.cns.nyu.edu/∼eero/publications.html}).
[920] D. L. Post and C. S. Calhoun. “An eEvaluation of Methods for Producing Desired Colors on
CRT Monitors.” Color Research and Application 14:4 (1989), 172–186.
[921] P. Poulin and A. Fournier. “A Model for Anisotropic Reflection.” Computer Graphics (Pro-
ceedings of SIGGRAPH ’90) 24:4 (1990), 273–282.
[922] C. Poynton. “Color FAQ.”, 2002. Available online (http://www.poynton.com/ColorFAQ.html).
[923] C. Poynton. Digital Video and HDTV: Algorithms and Interfaces. San Francisco: Morgan
Kaufmann Publishers, 2003.
[924] O. Prache. “Active Matrix Molecular OLED Microdisplays.” Displays 22 (2001), 49–56.
[925] A. J. Preetham, P. Shirley, and B. Smits. “A Practical Analytic Model for Daylight.” In
Proceedings of SIGGRAPH ’99, Computer Graphics Proceedings, Annual Conference Series,
pp. 91–100. Reading, MA: Addison-Wesley Longman, 1999.
[926] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery. Numerical Recipes in C:
The Art of Scientific Computing, Second edition. Cambridge, UK: Cambridge University Press,
1992.
[927] I. G. Priest. “A Proposed Scale for Use in Specifying the Chromaticity of Incandescent Illu-
minants and Various Phases of Daylight.” Journal of the Optical Society of America 23 (1933),
141–141.
[928] E. N. Pugh Jr. and T. D. Lamb. “Phototransduction in Vertebrate Rods and Cones: Molecular
Mechanisms of Amplification, Recovery and Light Adaptation.” In Handbook of Biological
Physics, Volume 3, edited by D. G. S. et al., pp. 183–254. Amsterdam: Elsevier, 2000.

i i

i i
 i i

i i

Bibliography 1007

[929] M. D. Purdy. “Spectral Hue as a Fucntion of Intensity.” American Journal of Psychology 43

(1931), 541–559.
[930] J. E. Purkinje. Neue Beiträge zur Kenntnis des Sehens in subjectiver Hinsicht. Berlin: G
Reimer, 1825.
[931] Z. Rahman, D. J. Jobson, and G. A. Woodell. “A Multiscale Retinex for Color Rendition and
Dynamic Range Compression.” In Proceedings of the SPIE 2847 (Applications of Digital Image
Processing XIX). Bellingham, WA: SPIE, 1996.
[932] Z. Rahman, G. A. Woodell, and D. J. Jobson. “A Comparison of the Multiscale Retinex with
Other Image Enhancement Techniques.” In IS&T’s 50th Annual Conference: A Celebration of
All Imaging, pp. 426–431. Springfield, VA: Society for Imaging Science and Technology, 1997.
[933] S. J. M. Rainville and F. A. A. Kingdom. “Spatial-Scale Contribution to the Detection of
Mirror Symmetry in Fractal Noise.” Journal of the Optical Society of America A 16:9 (1999),
2112–2123.
[934] V. S. Ramachandran. “Perception of Shape from Shading.” Nature 331:6152 (1988), 163–166.
[935] S. Rámon y Cajal. “La rétine des vertébrés.” Cellule 9 (1893), 119–255.
[936] A. Rapaport, J. Milliez, M. Bass, A. Cassanho, and H. Jenssen. “Review of the Properties of
Up-Conversion Phosphors for New Emissive Displays.” SID Symposium Digest of Technical
Papers 35:1 (2004), 1237–1239.
[937] A. Rapaport, J. Milliez, M. Bass, A. Cassanho, and H. Jenssen. “Review of the Properties of
Up-Conversion Phosphors for New Emissive Displays.” IEEE/OSA Journal of Display Technol-
ogy 2:1 (2006), 68–78.
[938] K. Rasche, R. Geist, and J. Westall. “Detail Preserving Reproduction of Color Images for
Monochromats and Dichromats.” IEEE Computer Graphics and Applications 25:3 (2005), 22–
30.
[939] K. Rasche, R. Geist, and J. Westall. “Re-Coloring Images for Gamuts of Lower Dimension.”
Eurographic Forum 24:3 (2005), 423–432.
[940] D. D. Rathman, D. M. O’Mara, M. K. Rose, R. M. Osgood III, and R. K. Reich. “Dynamic
Response of an Electronically Shuttered CCD Imager.” IEEE Transactions on Electron Devices
51:6.
[941] M. S. Rea, editor. The IESNA Lighting Handbook: Reference and Application, Ninth edition.
New York: The Illuminating Engineering Society of North America, 2000.
[942] C. Redies and L. Spillmann. “The Neon Color Effect in the Ehrenstein Illusion.” Perception
10:6 (1981), 667–681.
[943] C. Redies, L. Spillmann, and K. Kunz. “Colored Neon Flanks and Line Gap Enhancement.”
Vision Research 24:10 (1984), 1301–1309.
[944] B. C. Regan, C. Julliot, B. Simmen, F. Viénot, P. Charles-Dominique, and J. D. Mollon. “Fruits,
Foliage and the Evolution of Primate Colour Vision.” Philosophical Transactions of the Royal
Society of London, B 356:1407 (2001), 229–283.
[945] T. Regier, P. Kay, and N. Khetarpal. “Color Naming Reflects Optimal Partitions of Color
Space.” Proceedings of the National Academy of Sciences 104:4 (2007), 1436–1441.
[946] R. C. Reid and R. M. Shapley. “Spatial Structure of Cone Inputs to Receptive Fields in Primate
Lateral Geniculate Nucleus.” Nature 356:6371 (1992), 716–718.
[947] R. C. Reid and R. M. Shapley. “Space and Time Maps of Cone Photoreceptor Signals in
Macaque Lateral Geniculate Nucleus.” Journal of Neuroscience 22:14 (2002), 6158–6175.
[948] E. Reinhard and K. Devlin. “Dynamic Range Reduction Inspired by Photoreceptor Physiol-
ogy.” IEEE Transactions on Visualization and Computer Graphics 11:1 (2005), 13–24.

i i

i i
 i i

i i

1008 Bibliography

[949] E. Reinhard, M. Ashikhmin, B. Gooch, and P. Shirley. “Color Transfer between Images.” IEEE
Computer Graphics and Applications 21 (2001), 34–41.
[950] E. Reinhard, M. Stark, P. Shirley, and J. Ferwerda. “Photographic Tone Reproduction for
Digital Images.” ACM Transactions on Graphics 21:3 (2002), 267–276.
[951] E. Reinhard, A. O. Akyüz, M. Colbert, C. E. Hughes, and M. O’Connor. “Real-Time Color
Blending of Rendered and Captured Video.” In Proceedings of the Interservice/Industry Train-
ing, Simulation and Education Conference. Arlington, VA: National Defense Industrial Associ-
ation, 2004.
[952] E. Reinhard, G. Ward, S. Pattanaik, and P. Debevec. High Dynamic Range Imaging: Acquisi-
tion, Display and Image-Based Lighting. San Francisco: Morgan Kaufmann Publishers, 2005.
[953] E. Reinhard, P. Debevec, G. Ward, K. Myszkowski, H. Seetzen, D. Hess, G. McTaggart, and
H. Zargarpour. “High Dynamic Range Imaging: Theory and Practice.”, 2006. (SIGGRAPH
2006 Course).
[954] E. Reinhard. “Parameter Estimation for Photographic Tone Reproduction.” journal of graphics
tools 7:1 (2003), 45–51.
[955] A. G. Rempel, M. Trentacoste, H. Seetzen, H. D. Young, W. Heidrich, L. Whitehead, and
G. Ward. “Ldr2Hdr: On-the-fly Reverse Tone Mapping of Legacy Video and Photographs.”
ACM Transactions on Graphics 26:3 (2007), 39–1 – 39–6.
[956] E. Renner. Pinhole Photography: Rediscovering a Historic Technique, Third edition. Focal
Press, 2004.
[957] M. Rezak and L. A. Benevento. “A Comparison of the Organization of the Projections of
the Dorsal Lateral Geniculate Nucleus, the Inferior Pulvinar and Adjacent Lateral Pulvinar to
Primary Visual Cortex (Area 17) in the Macaque Monkey.” Brain Research 167 (1979), 19–40.
[958] P. Rheingans. “Color, Change, and Control for Quantitative Data Display.” In Proceedings of
IEEE Visualization, pp. 252–258. Los Alamitos, CA: IEEE Press, 1999.
[959] P. Rheingans. “Task-Based Color Scale Design.” In SPIE Proceedings (Applied Image and
Pattern Recognition), pp. 33–43. Bellingham, WA: SPIE, 1999.
[960] D. C. Rich. “Light Sources and Illuminants: How to Standardize Retail Lighting.” Textile
Chemist and Colorist 30:1 (1998), 8–12.
[961] D. W. Rickman and N. C. Brecha. “Morphologies of Somatostatin-Immunoreactive Neurons
in the Rabbit Retina.” In Neurobiology of the Inner Retina, H31, edited by R. Weiler and N. N.
Osborne, H31, pp. 461–468. Berlin: Springer-Verlag, 1989.
[962] L. A. Riggs. “Visual Acuity.” In Vision and Visual Perception, edited by C. H. Graham. New
York: John Wiley and Sons, Inc., 1965.
[963] K. Riley, D. S. Ebert, M. Kraus, J. Tessendorf, and C. Hansen. “Efficient Rendering of Atmo-
spheric Phenomena.” In Eurographics Symposium on Rendering, edited by H. W. Jensen and
A. Keller, pp. 375–386. Aire-la-Ville, Switzerland: Eurographics Association, 2004.
[964] D. L. Ringach, R. M. Shapley, and M. J. Hawken. “Orientation Selectivity in Macaque V1:
Diversity and Laminar Dependence.” Journal of Neuroscience 22:13 (2002), 5639–5651.
[965] A. Rizzi, C. Gatta, and D. Marini. “From Retinex to Automatic Color Equalization: Issues in
Developing a New Algorithm for Unsupervised Color Equalization.” SPIE Journal of Electronic
Imaging 13:1 (2004), 75–84.
[966] P. K. Robertson and J. F. O’Callaghan. “The Generation of Color Sequences for Univariate and
Bivariate Mapping.” IEEE Computer Graphics and Applications 6:2 (1986), 24–32.
[967] M. Robertson, S. Borman, and R. Stevenson. “Estimation-Theoretic Approach to Dynamic
Range Enhancement using Multiple Exposures.” Journal of Electronic Imaging 12:2 (2003),
219–228.

i i

i i
 i i

i i

Bibliography 1009

[968] A. H. Robins. Perspectives on Human Pigmentation. Cambridge, UK: Cambridge University

Press, 1991.
[969] D. A. Robinson. “The Mechanics of Human Saccadic Eye Movements.” Journal of Physiology
174 (1964), 245–264.
[970] K. S. Rockland, K. S. Saleem, and K. Tanaka. “Divergent Feedback Connections from Areas
V4 and TEO in the Macaque.” Visual Neuroscience 11 (1994), 579–600.
[971] K. S. Rockland. “Elements of Cortical Architecture Revisited.” In Cerebral Cortex, edited by
K. S. Rockland, J. H. Kaas, and A. Peters. New York: Plenum, 1997.
[972] R. W. Rodieck. “The Primate Retina.” Comparative Primate Biology 4 (1988), 203–278.
[973] H. R. Rodman, K. M. Sorenson, A. J. Shim, and D. P. Hexter. “Calbindin Immunoreactivity
in the Geniculo-Extrastriate System of the Macaque: Implications for Heterogeneity in the Ko-
niocellular Pathway and Recovery from Cortical Damage.” Journal of Comparative Neurology
431:2 (2001), 168–181.
[974] B. E. Rogowitz and A. D. Kalvin. “The ’Which Blair Project:’ A Quick Visual Method for
Evaluating Perceptual Color Maps.” In Proceedings of IEEE Visualization, pp. 183–190. Los
Alamitos, CA: IEEE Computer Society, 2001.
[975] J. Roizen and R. Lipkin. “SECAM: An End to Crosstalk.” Electronics 38 (1965), 104–105.
[976] A. K. Romney and R. G. D’Andrade. “Modeling Lateral Geniculate Nucleus Cell Response
Spectra and Munsell Reflectance Spectra with Cone Sensitivity Curves.” Proceedings of the
National Academy of Sciences of the United States of America 102:45 (2005), 16512–16517.
[977] A. Roorda and D. R. Williams. “The Arrangement of the Three Cone Classes in the Living
Human Eye.” Nature 397:6719 (1999), 520–522.
[978] E. Rosche. “Natural Categories.” Cognitive Psychology 4:3 (1973), 328–350.
[979] A. Rose. “The Relative Sensitivities of Television Pick-Up Tubes, Photographics Film, and the
Human Eye.” In Proceedings of the Institute of Radio Engineers, 293–300, 293–300. Washing-
ton, DC: IEEE Press, 1942.
[980] M. Rosenblatt. “Remarks on Some Nonparametric Estimates of Density Functions.” Annals of
Mathematical Statistics 27:3 (1956), 832–837.
[981] S. Roth and W. Caldwell. “Four Primary Color Projection Display.” SID Symposium Digest of
Technical Papers 36:1 (2005), 1818–1821.
[982] J. M. Rubin and W. A. Richards. “Color Vision and Image Intensities: When Are Changes
Material?” Biological Cybernetics 45:3 (1982), 215–226.
[983] D. L. Ruderman and W. Bialek. “Statistics of Natural Images: Scaling in the Woods.” Physical
Review Letters 73:6 (1994), 814–817.
[984] D. L. Ruderman, T. W. Cronin, and C.-C. Chiao. “Statistics of Cone Responses to Natural
Images: Implications for Visual Coding.” Journal of the Optical Society of America A 15:8
(1998), 2036–2045.
[985] D. L. Ruderman. “Origins of Scaling in Natural Images.” Vision Research 37 (1997), 3385–
3398.
[986] D. L. Ruderman. “The Statistics of Natural Images.” Network: Computation in Neural Systems
5:4 (1997), 517–548.
[987] W. A. H. Rushton. “Visual Pigments in Man.” In Handbook of Sensory Physiology, VII/1,
edited by H. J. A. Dartnal. New York: Springer-Verlag, 1972.
[988] S. M. Sagar. “Somatostatin-Like Immunoreactive Material in the Rabbit Retina: Immunohis-
tochemical staining using monoclonal antibodies.” Journal of Comparative Neurology 266:2
(1987), 291–299.

i i

i i
 i i

i i

1010 Bibliography

[989] R. Saito and H. Kotera. “Extraction of Image Gamut Surface and Calculations of Its Volume.”
In Proceedings of the 8th IS&T/SID Color Imaging Conference, pp. 330–334. Springfield, VA:
Society for Imaging Science and Technology, 2000.
[990] E. Salvador, A. Cavallaro, and T. Ebrahimi. “Shadow Identification and Classification using In-
variant Color Models.” In International Conference on Acoustics, Speech, and Signal Processing
(ICASSP), Volume 3, pp. 1545–1548. Washington, DC: IEEE Press, 2001.
[991] C. D. Salzman, C. M. Murasugi, K. H. Britten, and W. T. Newsome. “Microstimulation in
Visual Area MT: Effects on Direction Discrimination Performance.” Journal of Neuroscience
12:6 (1992), 2331–2355.
[992] F. Samadzadegan, M. Hahn, M. Sarpulaki, and N. Mostofi. “Geometric and Radiometric Evalu-
ation of the Potential of a High Resolution CMOS-Camera.” In Proceedings of the International
Society for Photogrammetry and Remote Sensing (ISPRS), XXXV, B3, XXXV, B3, pp. 488–493.
Istanbul, Turkey, 2004.
[993] J. H. Sandell and P. H. Schiller. “Effect of Cooling Area 18 on Striate Cortex Cells in the
Squirrel Monkey.” Journal of Neurophysiology 48:1 (1982), 38–48.
[994] G. Sapiro. “Color and Illuminant Voting.” IEEE Transactions on Pattern Analysis and Machine
Intelligence 21 (1999), 1210–1215.
[995] M. Sasaki, M. Mase, S. Kawahito, and Y. Tadokoro. “A Wide-Dynamic-Range CMOS Im-
age Sensor based on Multiple Short Exposure-Time Readout with Multiple-Resolution Column
Parallel ADC.” IEEE Sensors Journal 7:1 (2007), 151–158.
[996] K. Sassen, N. C. Knight, Y. Takano, and A. J. Heymsfield. “Effects of Ice-Crystal Structure
on Halo Formation: Cirrus Cloud Experimental and Ray-Tracing Modeling Studies.” Applied
Optics 33:21 (1994), 4590–4601.
[997] K. Sassen. “Halos in Cirrus Clouds: Why Are Classic Displays So Rare?” Applied Optics
44:27 (2005), 5684–5687.
[998] Y. Sato, M. D. Wheeler, and K. Ikeuchi. “Object Shape and Reflectance Modeling from Ob-
servation.” In Proceedings of SIGGRAPH’97, Computer Graphics Proceedings, Annual Con-
ference Series, pp. 379–387. Reading, MA: Addison-Wesley, 1997.
[999] Y. Sato, K. Amemiya, and M. Uchidoi. “Recent Progress in Device Performance and Picture
Quality of Color Plasma Displays.” Journal of the Society for Information Display 10 (2002),
17–23.
[1000] K. Sato. “Image-Processing Algorithms.” In Image Sensors and Signal Processing for Digital
Still Cameras, edited by J. Nakamura, pp. 223–253. Boca Raton, FL: Taylor and Francis, 2006.
[1001] A. van der Schaaf. “Natural Image Statistics and Visual Processing.” Ph.D. thesis, Rijksuni-
versiteit Groningen, The Netherlands, 1998. Available online ({http://www.ub.rug.nl/eldoc/dis/
science/a.v.d.schaaf/}).
[1002] J. Schanda. “Current CIE Work to Achieve Physiologically-Correct Color Metrics.” In Color
Vision: Perspecitves from Different Disciplines, edited by W. Backhaus, R. Kliegl, and J. S.
Werner, pp. 307–318. Berlin: Walter de Gruyter, 1998.
[1003] R. Schettini, G. Ciocca, and I. Gagliardi. “Content-Based Color Image Retrieval with
Relevance Feedback.” In Proceedings of the International Conference on Image Processing,
pp. 27AS2.8–27AS2.8. Los Alamitos, CA: IEEE Computer Society, 1999.
[1004] H. M. Schey. Div, Grad, Curl, and All That. New York: W. W. Norton and Company, 1973.
[1005] C. Schlick. “A Customizable Reflectance Model for Everyday Rendering.” In Fourth Euro-
graphics Workshop on Rendering, pp. 73–84. Aire-la-Ville, Switzerland: Eurographics Associ-
ation, 1993.

i i

i i
 i i

i i

Bibliography 1011

[1006] C. Schlick. “Fast Alternatives to Perlin’s Bias and Gain Functions.” In Graphics Gems IV,
pp. 401–403. San Diego: Academic Press, 1994.
[1007] C. Schlick. “Quantization Techniques for the Visualization of High Dynamic Range Pictures.”
In Photorealistic Rendering Techniques, edited by P. Shirley, G. Sakas, and S. Müller, pp. 7–20.
Berlin: Springer-Verlag, 1994.
[1008] M. Schmolesky. “The Primary Visual Cortex.”, 2006. Available online (http://webvision.med.
utah.edu/VisualCortex.html).
[1009] J. L. Schnapf, T. W. Kraft, and D. A. Baylor. “Spectral Sensitivity of Human Cone Photore-
ceptors.” Nature 325:6103 (1987), 439–441.
[1010] J. L. Schnapf, T. W. Kraft, B. J. Nunn, and D. A. Baylor. “Spectral Sensitivity of Primate
Photoreceptors.” Visual Neuroscience 1 (1988), 255–261.
[1011] M. Schrauf, B. Lingelbach, E. Lingelbach, and E. R. Wist. “The Hermann Grid and the
Scintillation Effect.” Perception 24, supplement (1995), 88–89.
[1012] M. Schrauf, B. Lingelbach, and E. R. Wist. “The Scintillating Grid Illusion.” Vision Research
37:8 (1997), 1033–1038.
[1013] P. Schröder and W. Sweldens. “Spherical Wavelets: Efficiently Representing Functions on
the Sphere.” In Proceedings of SIGGRAPH ’95, Computer Graphics Proceedings, Annual Con-
ference Series, pp. 161–172. Reading, MA: Addison-Wesley, 1995.
[1014] E. F. Schubert and J. K. Kim. “Solid-State Light Sources Getting Smart.” Science 308 (2005),
1274–1278.
[1015] M. Schultze. “Zur Anatomie und Physiologie der Retina.” Archiv für mikroskopische
Anatomie und Entwicklungsmechanik 2 (1866), 165–286.
[1016] A. J. Schwab. Field Theory Concepts: Electromagnetic Fields, Maxwell’s Equations, Grad,
Curl, Div, etc. Berlin: Springer-Verlag, 1988.
[1017] E. Schwartz, R. B. Tootell, M. S. Silverman, E. Switkes, and R. L. de Valois. “On the Math-
ematical Structure of the Visuotopic Mapping of Macaque Striate Cortex.” Science 227 (1985),
1065–1066.
[1018] E. A. Schwartz. “Voltage Noise Observed in Rods of the Turtle Retina.” Journal of Physiology
272 (1977), 217–246.
[1019] S. H. Schwartz. Visual Perception: A Clinical Orientation. New York: McGraw-Hill, 2004.
[1020] H. Seetzen, L. A. Whitehead, and G. Ward. “A High Dynamic Range Display using Low
and High Resolution Modulators.” SID Symposium Digest of Technical Papers 34:1 (2003),
1450–1453.
[1021] H. Seetzen, W. Heidrich, W. Stuerzlinger, G. Ward, L. Whitehead, M. Trentacoste, A. Ghosh,
and A. Vorozcovs. “High Dynamic Range Display Systems.” ACM Transactions on Graphics
23:3 (2004), 760–768.
[1022] H. Seetzen, H. Li, G. Ward, L. Whitehead, and W. Heidrich. “Guidelines for Contrast, Bright-
ness, and Amplitude Resolution of Displays.” In Society for Information Display (SID) Digest,
pp. 1229–1233, 2006.
[1023] H. Seetzen, S. Makki, H. Ip, T. Wan, V. Kwong, G. Ward, W. Heidrich, and L. Whitehead.
“Self-Calibrating Wide Color Gamut High Dynamic Range Display.” In Human Vision and
Electronic Imaging XII, 6492, edited by B. E. Rogowitz, T. N. Pappas, and S. J. Daly, 6492,
p. 64920Z. SPIE, 2007.
[1024] R. G. Seippel. Optoelectronics for Technicians and Engineering. Englewood Cliffs, NJ:
Prentice Hall, 1989.

i i

i i
 i i

i i

1012 Bibliography

[1025] C. C. Semmelroth. “Prediction of Lightness and Brightness on Different Backgrounds.” Jour-

nal of the Optical Society of America 60 (1970), 1685–1689.
[1026] A. Sempel and M. Büchel. “Design Aspects of Low Power Polymer/OLED Passive-Matrix
Displays.” Organic Electronics 3 (2002), 89–92.
[1027] J. Sethian. “A Fast Marching Level Set Method for Monotonically Advancing Fronts.” Pro-
ceedings of the National Academy of Sciences of the United States of America 93 (1996), 1591–
1595.
[1028] S. A. Shafer. “Using Color to Separate Reflection Components.” Color Research and Appli-
cation 10:4 (1985), 210–218.
[1029] S. E. Shaheen, G. E. Jabbour, B. Kippelen, N. Peyghambarian, J. D. Anderson, S. R. Marder,
N. R. Armstrong, E. Bellmann, and R. H. Grubbs. “Organic Light-Emitting Diode with 20 lm/W
Efficiency using a Triphenyldiamine Side-Group Polymer as the Hole Transport Layer.” Applied
Physics Letters 74:21 (1999), 3212–3214.
[1030] R. Shapley and M. Hawken. “Neural Mechanisms for Color Perception in the Primary Visual
Cortex.” Current Opinion in Neurobiology 12:4 (2002), 426–432.
[1031] R. Shapley and J. Victor. “Hyperacuity in Cat Retinal Ganglion Cells.” Science 231 (1986),
999–1002.
[1032] R. Shapley. “Visual Sensitivity and Parallel Retinocortical Channels.” Annual Review of
Psychology 41 (1990), 635–658.
[1033] R. Shapley. “A New View of the Primary Visual Cortex.” Neural Networks 17:5–6 (2004),
615–623.
[1034] G. Sharma and H. J. Trussell. “Digital Color Imaging.” IEEE Transactions on Image Pro-
cessing 6:7 (1997), 901–932.
[1035] G. Sharma. “LCDs Versus CRTs: Color Calibration and Gamut Considerations.” Proceedings
of the IEEE: Special Issue on Flat Panel Display Technologies 90:4 (2002), 605–622.
[1036] L. T. Sharpe, H. J. Andrew Stockman, and J. Nathans. “Opsin Genes, Cone Photopigments,
Color Vision and Color Blindness.” In Color Vision: From Genes to Perception, edited by K. R.
Gegenfurtner and L. T. Sharpe, pp. 3–51. Cambridge, UK: Cambridge University Press, 1999.
[1037] J. H. Shaw. “Solar Radiation.” The Ohio Journal of Science 53:5 (1953), 258–271.
[1038] S. L. Sherwood, editor. The Nature of Psychology: A Selection of Papers, Essays and Other
Writings by Kenneth J W Craik. Cambridge: Cambridge University Press, 1966.
[1039] S. K. Shevell and J. Wei. “Chromatic Induction: Border Contrast or Adaptation to Surround-
ing Light.” Vision Research 38 (1998), 1561–1566.
[1040] J. Shinar, editor. Organic Light-Emitting Devices: A Survey. New York: Springer-Verlag,
2003.
[1041] S. Shipp and S. Zeki. “Segregation of Pathways Leading from Area V2 to Areas V4 and V5
of Macaque Monkey Visual Cortex.” Nature 315:6017 (1985), 322–324.
[1042] P. Shirley and R. K. Morley. Realistic Ray Tracing, Second edition. Wellesley, Massachusetts:
A K Peters, 2003.
[1043] P. Shirley, M. Ashikhmin, S. R. Marschner, E. Reinhard, K. Sung, W. B. Thompson, and
P. Willemsen. Fundamentals of Computer Graphics, Second edition. Wellesley, Massachusetts:
A K Peters, 2005.
[1044] A. Shmuel, M. Korman, A. Sterkin, M. Harel, S. Ullman, R. Malach, and A. Grinvald.
“Retinotopic Axis Specificity and Selective Clustering of Feedback Projections from V2 to V1
in the Owl Monkey.” Journal of Neuroscience 25:8 (2005), 2117–2131.

i i

i i
 i i

i i

Bibliography 1013

[1045] A. E. Siegman. Lasers. Sausalito, CA: University Science Books, 1986.

[1046] F. X. Sillion and C. Puech. Radiosity and Global Illumination. San Francisco, CA: Morgan
Kaufmann Publishers, Inc., 1994.
[1047] E. P. Simoncelli and J. Portilla. “Texture Characterization via Joint Statistics of Wavelet
Coefficient Magnitudes.” In Fifth IEEE Int’l Conf on Image Processing, Volume I, pp. 62–66.
Washington, DC: IEEE Press, 1998.
[1048] E. P. Simoncelli. “Bayesian Denoising of Visual Images in the Wavelet Domain.” In Bayesian
Inference in Wavelet Based Models, Lecture Notes in Statistics 141, edited by P. Müller and
B. Vidakovic, Chapter 18, pp. 291–308. New York: Springer-Verlag, 1999.
[1049] E. P. Simoncelli. “Modelling the Joint Statistics of Images in the Wavelet Domain.” In
Proceedings of the SPIE 3818 (44th SPIE Annual Meeting), pp. 188–195. Bellingham, WA:
SPIE, 1999.
[1050] L. C. Sincich and J. C. Horton. “Divided by Cytochrome Oxydase: A Map of the Projections
from V1 to V2 in Macaques.” Science 295:5560 (2002), 1734–1737.
[1051] L. C. Sincich and J. C. Horton. “The Circuitry of V1 and V2: Integration of Color, Form and
Motion.” Annual Review of Neuroscience 28 (2005), 303–326.
[1052] L. C. Sincich, K. F. Park, M. J. Wohlgemuth, and J. C. Horton. “Bypassing V1: A Direct
Geniculate Input to Area MT.” Nature Neuroscience 7:10 (2004), 1123–1128.
[1053] L. Sivik. “Systems for Descriptive Colour Notations — Implications of Definitions and
Methodology.” Farbe 40 (1994), 37–49.
[1054] G. Smith and D. A. Atchison. The Eye and Visual Optical Instruments. New York: Cambridge
University Press, 1997.
[1055] R. C. Smith and K. A. Baker. “Optical Properties of the Clearest Natural Waters (200-800
nm).” Applied Optics 20:2 (1981), 177–184.
[1056] R. C. Smith and J. E. Tyler. “Optical Properties of Clear Natural Water.” Journal of the
Optical Society of America 57:5 (1967), 589–601.
[1057] V. C. Smith, J. Pokorny, B. B. Lee, and D. M. Dacey. “Primate Horizontal Cell Dynamics: An
Analysis of Sensitivity Regulation in the Outer Retina.” Journal of Neurophysiology 85 (2001),
545–558.
[1058] K. Smith, G. Krawczyk, K. Myszkowski, and H.-P. Seidel. “Beyond Tone Mapping: En-
hanced Depiction of Tone Mapped HDR Images.” Computer Graphics Forum 25:3 (2006),
427–438.
[1059] R. Smith-Gillespie. “Design Considerations for LED Backlights in Large Format Color
LCDs.” In LEDs in Displays — SID Technical Symposium, 2006.
[1060] H. M. Smith. Principles of Holography, Second edition. New York, NY: John Wiley and
Sons,, Inc., 1975.
[1061] W. J. Smith. Modern Optical Engineering, Third edition. New York: McGraw-Hill, 2000.
[1062] W. J. Smith. Modern Lens Design, Second edition. New York: McGraw Hill, 2005.
[1063] B. Smits. “An RGB to Spectrum Conversion for Reflectances.” journal of graphics tools 4:4
(1999), 11–22.
[1064] E. Snitzer. “Cylindrical Dielectric Waveguide Modes.” Journal of the Optical Society of
America 51:5 (1961), 491–498.
[1065] D. M. Snodderly, J. D. Auran, and F. C. Delori. “The Macular Pigment II: Spatial Distribution
in Primate Retinas.” Investigative Ophthalmology and Visual Science 25:6 (1984), 674–685.
[1066] A. Sobel. “Television’s Bright New Technology.” Scientific American.

i i

i i
 i i

i i

1014 Bibliography

[1067] Society of Dyers and Colourists. Colour Index. Bradford, UK: Society of Dyers and
Colourists, 1971. Available online (http://www.colour-index.org).
[1068] F. M. Sogandares and E. S. Fry. “Absorption Spectrum (340–640 nm) of Pure Water. II.
Photothermal Measurements.” Applied Optics 36:33 (1997), 8699–8709.
[1069] J. A. Solomon, G. Sperling, and C. Chubb. “The Lateral Inhibition of Perceived Contrast Is
Indifferent to On-Center/Off-Center Segregation, but Specific to Orientation.” Vision Research
33:18 (1993), 2671–2683.
[1070] X. Song, G. M. Johnson, and M. D. Fairchild. “Minimizing the Perception of Chromatic Noise
in Digital Images.” In IS&T/SID 12th Color Imaging Conference, pp. 340–346. Springfield, VA:
Society for Imaging Science and Technology, 2004.
[1071] M. Sonka, V. Hlavac, and R. Boyle. Image Processing, Analysis, and Machine Vision, Second
edition. Pacific Grove, CA: PWS Publishing, 1999.
[1072] K. E. Spaulding, E. J. Giorgianni, and G. Woolfe. “Reference Input/Output Medium Metric
RGB Color Encoding (RIMM-ROMM RGB).” In PICS2000: Image Processing, Image Quality,
Image Capture, Systems Conference, pp. 155–163. Springfield, VA: Society for Imaging Science
and Technology, 2000.
[1073] K. E. Spauling, G. J. Woolfe, and R. L. Joshi. “Extending the Color Gamut and Dynamic
Range of an sRGB Image using a Residual Image.” Color Research and Application 25:4 (2003),
251–266.
[1074] N. Speranskaya. “Determination of Spectrum Color Co-ordinates for Twenty-Seven Normal
Observers.” Optics Spectrosc. 7 (1959), 424–428.
[1075] L. Spillmann and J. S. Werner, editors. Visual Perception: The Neurological Foundations.
San Diego: Academic Press, 1990.
[1076] L. Spillmann. “The Hermann Grid Illusion: A Tool for Studying Human Receptive Field
Organization.” Perception 23 (1994), 691–708.
[1077] B. Stabell and U. Stabell. “Chromatic Rod-Cone Interaction during Dark Adaptation.” Journal
of the Optical Society of America A 15:11 (1998), 2809–2815.
[1078] B. Stabell and U. Stabell. “Effects of Rod Activity on Color Perception with Light Adapta-
tion.” Journal of the Optical Society of America A 19:7 (2002), 1248–1258.
[1079] J. Stam. “Stable Fluids.” In Proceedings SIGGRAPH ’99, Computer Graphics Proceed-
ings,Annual Conference Series, pp. 121–128. Reading, MA: Addison-Wesley Longman, 1999.
[1080] J. Starck, E. Pantin, and F. Murtagh. “Deconvolution in Astronomy: A Review.” Publications
of the Astronomical Society of the Pacific 114 (2002), 1051–1069.
[1081] O. N. Stavroudis. “Orthomic Systems of Rays in Inhomogeneous Isotropic Media.” Applied
Optics 10:2 (1971), 260–263.
[1082] O. N. Stavroudis. The Optics of Rays, Wavefronts and Caustics. New York: Academic Press,
1972.
[1083] D. A. Steigerwald, J. C. Bhat, D. Collins, R. M. Fletcher, M. O. Holcomb, M. J. Ludowise,
P. S. Martin, and S. L. Rudaz. “Illumination with Solid State Lighting Technology.” IEEE
Journal on Selected Topics in Quantum Electronics 8:2 (2002), 310–320.
[1084] J. Steinhoff and D. Underhill. “Modifications of the Euler Equations for “Vorticity Confine-
ment”: Application to the Computation of Interacting Vortex Rings.” Physics of Fluids 6:8
(1994), 2738–2744.
[1085] J. C. Stevens and S. S. Stevens. “Brightness fFunctions: Effects of Adaptation.” Journal of
the Optical Society of America 53 (1963), 375–385.
[1086] S. S. Stevens. “To Honor Fechner and Repeal His Law.” Science 133:3446 (1961), 80–86.

i i

i i
 i i

i i

Bibliography 1015

[1087] W. Stiles and J. Burch. “N.P.L Colour-Matching Investigation: Final Report.” Optica Acta 6
(1959), 1–26.
[1088] W. S. Stiles and B. H. Crawford. “The Luminous Efficiency of Rays Entering the Eye Pupil
at Different Points.” Proceedings of the Royal Society of London, B 112:778 (1933), 428–450.
[1089] W. S. Stiles. “The Luminous Sensitivity of Monochromatic Rays Entering the Eye Pupil at
Different Points and a New Colour Effect.” Proceedings of the Royal Society of London, B
123:830 (1937), 64–118.
[1090] A. Stimson. Photometry and Radiometry for Engineers. New York: John Wiley and Sons,
Inc., 1974.
[1091] A. Stockman and L. T. Sharpe. “The Spectral Sensitivities of the Middle- and Long-
Wavelength-Sensitive Cones Derived from Measurements on Observers of Known Genotype.”
Vision Research 40:13 (2000), 1711–1737.
[1092] A. Stockman, D. I. MacLeod, and N. E. Johnson. “Spectral Sensitivities of the Human Cones.”
Journal of the Optical Society of America A 10:12 (1993), 2491–2521.
[1093] E. J. Stollnitz, V. Ostromoukhov, and D. H. Salesin. “Reproducing Color Images using Cus-
tom Inks.” In Proceedings SIGGRAPH ’98, Computer Graphics Proceedings, Annual Confer-
ence Series, pp. 267–274. Reading, MA: Addison-Wesley, 1998.
[1094] R. I. Stolyarevskaya. “The 25th Meeting of CIE and Measurements on LED Optical Parame-
ters.” Measurement Techniques 47:3 (2004), 316–320.
[1095] M. C. Stone, W. B. Cowan, and J. C. Beatty. “Color Gamut Mapping and the Printing of
Digital Color Images.” ACM Transactions on Graphics 7:4 (1988), 249–292.
[1096] M. C. Stone. “Color Balancing Experimental Projection Displays.” In Proceedings of the 9th
Color Imaging Conference: Color Science and Engineering: Systems, Technologies, Applica-
tions, pp. 342–347. Springfield, VA: Society for Imaging Science and Technology, 2001.
[1097] M. C. Stone. A Field Guide to Digital Color. Natick, MA: A K Peters, 2003.
[1098] M. Störring, E. Granum, and H. J. Andersen. “Estimation of the Illuminant Colour using
Highlights from Human Skin.” In International Conference on Color in Graphics and Image
Processing, pp. 45–50. Paris: Lavoisier, 2000.
[1099] A. Streitwieser Jr, C. H. Heathcock, and E. M. Kosower. Introduction to Organic Chemistry,
Fourth edition. New York: Macmillan, 1992.
[1100] C. F. Stromeyer. “Form-Color After Effects in Human Vision.” In Handbook of Sensory
Physiology: Perception, VIII, edited by R. Held, H. W. Leibowitz, and H. L. Teuber, pp. 97–142.
New York: Springer-Verlag, 1978.
[1101] S. Subramanian and I. Biederman. “Does Contrast Reversal Affect Object Identification?”
Investigative Ophthalmology and Visual Science 38:998.
[1102] K. Suffern. Ray Tracing from the Ground Up. Wellesley, MA: A K Peters, 2007.
[1103] A. Sugimoto, H. Ochi, S. Fujimura, A. Yoshida, T. Miyadera, and M. Tsuchida. “Flexible
OLED Displays using Plastic Substrates.” IEEE Journal of Selected Topics in Quantum Elec-
tronics 10:1 (2004), 107–114.
[1104] H. Sugiura, K. Sakawa, and J. Fujimo. “False Color Signal Reduction Method for Single-Chip
Color Video Cameras.” IEEE Transactions on Consumer Electronics 40:2 (1994), 100–106.
[1105] W. E. Sumpner. “The Diffusion of Light.” Proceedings of the Physical Society of London 12
(1892), 10–29.
[1106] Y. Sun, D. Fracchia, T. W. Calvert, and M. S. Drew. “Deriving Spectra from Colors and
Rendering Light Interference.” IEEE Computer Graphics and Applications 19:4 (1999), 61–67.

i i

i i
 i i

i i

1016 Bibliography

[1107] Y. Sun. “Rendering Biological Iridescences with RGB-Based Renderers.” ACM Transactions
on Graphics 25:1 (2006), 100–129.
[1108] W. Suzuki, K. S. Saleem, and K. Tanaka. “Divergent Backward Projections from the An-
terior Part of the Inferotemporal Cortex (Area TE) in the Macaque.” Journal of Comparative
Neurology 422:2 (2000), 206–228.
[1109] H. Suzuki. Electronic Absorption Spectra and Geometry of Organic Molecules: An Applica-
tion of Molecular Orbital Theory. New York: Academic Press, 1967.
[1110] M. J. Swain and D. H. Ballard. “Color Indexing.” International Journal of Computer Vision
7:1 (1991), 11–32.
[1111] Swedish Standards Institute. Colour Notation System, Second edition. Stockholm: Swedish
Standards Institute, 1990.
[1112] K. Takatoh, M. Hasegawa, M. Koden, N. Itoh, R. Hasegawa, and M. Sakamoto. Alignment
Technologies and Applications of Liquid Crystal Devices. London: Taylor and Francis, 2005.
[1113] H. Takeshi, H. Onoe, H. Shizuno, E. Yoshikawa, N. Sadato, H. Tsukada, and Y. Watanabe.
“Mapping of Cortical Areas Involved in Color Vision in Non-Human Primates.” Neuroscience
Letters 230:1 (1997), 17–20.
[1114] A. A. Talin, K. A. Dean, and J. E. Jaskie. “Field Emission Displays: A Critical Review.”
Solid-State Electronics 45 (2001), 963–976.
[1115] E.-V. Talvala, A. Adams, M. Horowitz, and M. Levoy. “Veiling Glare in High Dynamic Range
Imaging.” ACM Transactions on Graphics 26:3 (2007), 37–1 – 37–9.
[1116] N. Tamura, N. Tsumura, and Y. Miyake. “Masking Model for Accurate Colorimetric Charac-
terization of LCD.” In Proceedings of the IS&T/SID Tenth Color Imaging Conference, pp. 312–
316. Springfield, VA: Society of Imaging Science and Technology, 2002.
[1117] C. W. Tang and S. A. VanSlyke. “Organic Electroluminescent Diode.” Applied Physics Letters
51:12 (1987), 913–915.
[1118] J. Tanida and K. Yamada. “TOMBO: Thin Observation by Bound Optics.” Applied Optics
40:11 (2001), 1806–1813.
[1119] J. Tanida, R. Shogenji, Y. Kitamura, M. Miyamoto, and S. Miyatake. “Color Imaging with
an Integrated Compound Imaging System.” Journal of the Optical Society of America A 11:18
(2003), 2109–2117.
[1120] L. E. Tannas, editor. Flat-Panel Displays and CRTs. New York: Van Nostrand Reinhold
Company, 1985.
[1121] A. H. Taylor and E. B. Rosa. “Theory, Construction, and Use of the Photometric Integrating
Sphere.” Scientific Papers of the Bureau of Standards 18 (1922), 280–325.
[1122] H. Terstiege. “Chromatic Adaptation: A State-of-the-Art Report.” Journal of Color and
Appearance 1 (1972), 19–23.
[1123] A. J. P. Theuwissen. “CCD or CMOS Image Sensors for Consumer Digital Still Photography.”
In IEEE International Symposium on VLSI Technology, Systems, and Applications, pp. 168–171,
2001.
[1124] W. B. Thompson, P. Shirley, and J. Ferwerda. “A Spatial Post-Processing Algorithm for
Images of Night Scenes.” journal of graphics tools 7:1 (2002), 1–12.
[1125] M. G. A. Thomson. “Higher-Order Structure in Natural Scenes.” Journal of the Optical
Society of America A 16:7 (1999), 1549–1553.
[1126] M. G. A. Thomson. “Visual Coding and the Phase Structure of Natural Scenes.” Network:
Computation in Neural Systems 10:2 (1999), 123–132.

i i

i i
 i i

i i

Bibliography 1017

[1127] L. G. Thorell, R. L. de Valois, and D. G. Albrecht. “Spatial Mapping of Monkey V1 Cells

with Pure Color and Luminance Stimuli.” Vision Research 24:7 (1984), 751–769.
[1128] R. Tilley. Colour and the Optical Properties of Materials. Chichester, UK: John Wiley and
Sons, 1999.
[1129] D. Todorović. “Lightness and Junctions.” Perception 26 (1997), 379–395.
[1130] D. J. Tolhurst, Y. Tadmor, and T. Chiao. “Amplitude Spectra of Natural Images.” Ophthalmic
and Physiological Optics 12 (1992), 229–232.
[1131] C. Tomasi and R. Manduchi. “Bilateral Filtering for Gray and Color Images.” In Proceedings
of the IEEE International Conference on Computer Vision, pp. 836–846. Washington, DC: IEEE
Press, 1998.
[1132] S. Tominaga and B. A. Wandell. “The Standard Surface Reflectance Model and Illuminant
Estimation.” Journal of the Optical Society of America A 6:4 (1989), 576–584.
[1133] S. Tominaga and B. A. Wandell. “Natural Scene-Illuminant Estimation using the Sensor
Correlation.” Proceedings of the IEEE 90:1 (2002), 42–56.
[1134] S. Tominaga, S. Ebisui, and B. A. Wandell. “Scene Illuminant Classification: Brighter is
Better.” Journal of the Optical Society of America A 18:1 (2001), 55–64.
[1135] S. Tominaga. “Surface Identification using the Dichromatic Reflection Model.” IEEE Trans-
actions on Pattern Analysis and Machine Intelligence 13 (1991), 658–670.
[1136] S. Tominaga. “Dichromatic Reflection Models for a Variety of Materials.” Color Research
and Application 19:4 (1994), 277–285.
[1137] S. Tominaga. “Multichannel Vision System for Estimation Surface and Illumination Func-
tions.” Journal of the Optical Society of America A 13:11 (1996), 2163–2173.
[1138] R. B. H. Tootell, J. B. Reppas, A. M. Dale, R. B. Look, M. I. Sereno, R. Malach, T. J.
Brady, and B. R. Rosen. “Visual Motion Aftereffect in Human Cortical Area MT Revealed by
Functional Magnetic Resonance Imaging.” Nature 375:6527 (1995), 139–141.
[1139] R. B. H. Tootell, K. Nelissen, W. Vanduffel, and G. A. Orban. “Search for Color ’Center(s)’
in Macaque Visual Cortex.” Cerebral Cortex 14:4 (2004), 353–363.
[1140] K. E. Torrance and E. M. Sparrow. “Theory for Off-Specular Reflection from Roughened
Surfaces.” Journal of the Optical Society of America 57:9 (1967), 1105–1114.
[1141] K. Toyoda. “Digital Still Cameras at a Glance.” In Image Sensors and Signal Processing for
Digital Still Cameras, edited by J. Nakamura, pp. 1–19. Boca Raton, FL: Taylor and Francis,
2006.
[1142] C. van Trigt. “Color Rendering, A Reassessment.” Color Research and Application 24:3
(1999), 197–206.
[1143] T. Troscianko and I. Low. “A Technique for Presenting Isoluminant Stimuli using a Micro-
computer.” Spatial Vision 1 (1985), 197–202.
[1144] T. Troscianko, R. Baddeley, C. Párraga, U. Leonards, and J. Troscianko. “Visual Encoding of
Green Leaves in Primate Vision.” Journal of Vision 3 (2003), 137a.
[1145] T. Troscianko. “A Given Visual Field Location Has a Wide Range of Receptive Field Sizes.”
Vision Research 22 (1982), 1363–1369.
[1146] J. Tsitsiklis. “Efficient Algorithms for Globally Optimal Trajectories.” IEEE Transactions on
Automatic Control 40 (1995), 1528–1538.
[1147] D. Y. Ts’o and C. D. Gilbert. “The Organization of Chromatic and Spatial Interactions in the
Primate Striate Cortex.” Journal of Neuroscience 8:5 (1988), 1712–1727.

i i

i i
 i i

i i

1018 Bibliography

[1148] T. Tsuji, S. Kawami, S. Miyaguchi, T. Naijo, T. Yuki, S. Matsuo, and H. Miyazaki. “Red-
Phosphorescent OLEDs Employing bis-(8-Quinolinolato)-Phenolato-Aluminum (III) Com-
plexes as Emission-Layer Hosts.” SID Symposium Digest of Technical Papers 35:1 (2004),
900–903.
[1149] M. Tsukada and Y. Ohta. “An Approach to Color Constancy from Mutual Reflection.” In
Proceedings of the 3rd International Conference on Computer Vision, pp. 385–389. Washington,
DC: IEEE Press, 1990.
[1150] T. Tsutsui. “A Light-Emitting Sandwich Filling.” Nature 420:6917 (2002), 752–755.
[1151] H. F. J. M. van Tuijl. “A New Visual Illusion: Neonlike Color Spreading and Complementary
Color Induction between Subjective Contours.” Acta Psychologica 39 (1975), 441–445.
[1152] J. Tumblin and H. Rushmeier. “Tone Reproduction for Computer Generated Images.” IEEE
Computer Graphics and Applications 13:6 (1993), 42–48.
[1153] J. Tumblin and G. Turk. “LCIS: A Boundary Hierarchy for Detail-Preserving Contrast Re-
duction.” In Proceedings SIGGRAPH ’99, Computer Graphics Proceedings, Annual Conference
Series, edited by A. Rockwood, pp. 83–90. Reading, MA: Addison-Wesley Longman, 1999.
[1154] K. Turkowski. “Transformations of Surface Normal Vectors with Applications to Three Di-
mensional Computer Graphics.” Technical Report 22, Apple Computer, Inc., 1990.
[1155] S. R. Turns. An Introduction to Combustion: Concepts and Applications, Second edition.
New York: McGraw-Hill, 1996.
[1156] S. A. Twomey, C. F. Bohren, and J. L. Mergenthaler. “Reflectance and Albedo Differences
between Wet and Dry Surfaces.” Applied Optics 25:3 (1986), 431–437.
[1157] C. W. Tyler. “Analysis of Human Receptor Density.” In Basic and Clinical Applications of
Vision Science, edited by V. Lakshminarayanan, pp. 63–71. Norwood, MA: Kluwer Academic
Publishers, 1997.
[1158] E. P. T. Tyndall. “Chromaticity Sensibility to Wavelength Difference as a Function of Purity.”
Journal of the Optical Society of America 23:1 (1933), 15–22.
[1159] H. Uchiike and T. Hirakawa. “Color Plasma Displays.” Proceedings of the IEEE 90:4 (2002),
533–539.
[1160] H. Uchiike, K. Miura, N. Nakayama, and T. Shinoda. “Secondary Electron Emission Charac-
teristics of Dielectric Materials in AC-Operated Plasma Display Panels.” IEEE Transactions on
Electron Devices ED-23 (1976), 1211–1217.
[1161] Z. Ulanowski. “Ice Analog Halos.” Applied Optics 44:27 (2005), 5754–5758.
[1162] R. Ulbricht. “Die Bestimmung der mittleren raumlichen Lichtintensitat durch nur eine Mes-
sung.” Elektrotechnische Zeitschrift 21 (1900), 595–597.
[1163] I. Underwood. “A Review of Microdisplay Technologies.” In The Annual Conference of
the UK and Ireland Chapter of the Society for Information Display. San Jose, CA: Society for
Information Display, 2000.
[1164] L. G. Ungerleider and R. Desimone. “Cortical Connections of Visual Area MT in the
Macaque.” Journal of Comparative Neurology 248:2 (1986), 190–222.
[1165] L. G. Ungerleider and R. Desimone. “Projections to the Superior Temporal Sulcus from the
Central and Peripheral Field Representations of V1 and V2.” Journal of Comparative Neurology
248:2 (1986), 147–163.
[1166] United States Committee on Extension to the Standard Atmosphere. U.S. Standard Atmo-
sphere, 1976. Washington, D.C.: National Oceanic and Atmospheric Administration, National
Aeronautics and Space Administration, United States Air Force, 1976.

i i

i i
 i i

i i

Bibliography 1019

[1167] M. Uyttendaele, A. Eden, and R. Szeliski. “Eliminating Ghosting and Exposure Artifacts
in Image Mosaics.” In IEEE Computer Society Conference on Computer Vision and Pattern
Recognition, Volume 2, p. 509. Washington, DC: IEEE Press, 2001.
[1168] K. M. Vaeth. “OLED-display technology.” Information Display 19:6 (2003), 12–17.
[1169] J. M. Valeton and D. van Norren. “Light-Adaptation of Primate Cones: An Analysis Based
on Extracellular Data.” Vision Research 23:12 (1982), 1539–1547.
[1170] R. L. de Valois and K. K. de Valois. “Neural Coding of Color.” In Handbook of Perception V,
edited by E. C. Carterette and M. P. Friedman, pp. 117–166. New York: Academic Press, 1975.
[1171] R. L. de Valois and K. K. de Valois. “A Multi-Stage Color Model.” Vision Research 33 (1993),
1053–1065.
[1172] R. L. de Valois, I. Abramov, and G. H. Jacobs. “Analysis of Response Patterns of LGN Cells.”
Journal of the Optical Society of America 56:7 (1966), 966–977.
[1173] R. L. de Valois, D. M. Snodderly Jr, E. W. Yund, and N. K. Hepler. “Responses of Macaque
Lateral Geniculate Cells to Luminance and Color Figures.” Sensory Processes 1:3 (1977), 244–
259.
[1174] R. L. de Valois, K. K. de Valois, E. Switkes, and L. E. Mahon. “Hue Scaling of Isoluminant
and Cone-Specific Lights.” Vision Research 37:7 (1997), 885–897.
[1175] R. L. de Valois, N. P. Cottaris, S. D. Elfar, L. E. Mahon, and J. A. Wilson. “Some Transfor-
mations of Color Information from Lateral Geniculate Nucleus to Striate Cortex.” Proceedings
of the National Academy of Sciences of the United States of America 97:9 (2000), 4997–5002.
[1176] R. L. de Valois, K. K. de Valois, and L. E. Mahon. “Contribution of S Opponent Cells to
Color Appearance.” Proceedings of the National Academy of Sciences of the United States of
America 97:1 (2000), 512–517.
[1177] D. I. Vaney. “The Mosaic of Amacrine Cells in the Mammalian Retina.” Progress in Retinal
and Eye Research 9 (1990), 49–100.
[1178] R. VanRullen and T. Dong. “Attention and Scintillation.” Vision Research 43:21 (2003),
2191–2196.
[1179] D. Varin. “Fenomeni di contrasto e diffusione cromatica nell organizzazione spaziale del
campo percettivo.” Rivista di Psicologia 65 (1971), 101–128.
[1180] R. G. Vautin and B. M. Dow. “Color Cell Groups in Foveal Striate Cortex of the Behaving
Macaque.” Journal of Neurophysiology 54:2 (1985), 273–292.
[1181] T. R. Vidyasagar, J. J. Kulikowski, D. M. Lipnicki, and B. Dreher. “Convergence of Parvocel-
lular and Magnocellular Information Channels in the Primary Visual Cortex of the Macaque.”
European Journal of Neuroscience 16:5 (2002), 945–956.
[1182] J. A. S. Viggiano. “Modeling the Color of Multi-Colored Halftones.” In Proceedings of the
Technical Association of Graphical Arts, pp. 44–62. Sewickley, PA: Technical Association of
the Graphic Arts, 1990.
[1183] B. Vohnsen. “Photoreceptor Waveguides and Effective Retinal Image Quality.” Journal of the
Optical Society of America A 24:3 (2007), 597–607.
[1184] E. Völkl, L. F. Allard, and D. C. Joy, editors. Introduction to Electron Holography. New
York: Kluwer Academic / Plenum Publishers, 1999.
[1185] J. J. Vos, J. Walraven, and A. van Meeteren. “Light Profiles of the Foveal Image of a Point
Source.” Vision Research 16:2 (1976), 215–219.
[1186] J. Vos. “Disability Glare—A State of the Art Report.” CIE Journal 3:2 (1984), 39–53.
[1187] M. J. Vrhel and H. J. Trussell. “Color Device Calibration: A Mathematical Formulation.”
IEEE Transactions on Image Processing 8:12 (1999), 1796–1806.

i i

i i
 i i

i i

1020 Bibliography

[1188] H. L. de Vries. “The Quantum Character of Light and Its Bearing upon the Threshold of
Vision, the Differential Sensitivity and Acuity of the Eye.” Physica 10:7 (1943), 553–564.
[1189] T. Wachtler, T. J. Sejnowski, and T. D. Albright. “Representation of Color Stimuli in Awake
Macaque Primary Visual Cortex.” Neuron 37:4 (2003), 681–691.
[1190] H. G. Wagner, E. F. MacNichol, and M. L. Wolbarsht. “Three Response Properties of Single
Ganglion Cells in the Goldfish Retina.” Journal of General Physiology 43 (1960), 45–62.
[1191] G. Wald and D. R. Griffin. “The Change in Refractive Power of the Human Eye in Dim and
Bright Light.” Journal of the Optical Society of America 37 (1946), 321–336.
[1192] I. Wald and P. Slusallek. “State of the Art in Interactive Ray Tracing.” In Eurographics STAR -
State of the Art Reports, pp. 21–42. Aire-la-Ville, Switzerland: Eurographics Association, 2001.
[1193] I. Wald, P. Slusallek, and C. Benthin. “Interactive Distributed Ray Tracing of Highly Complex
Models.” In Proceedings of the 12th Eurographics Workshop on Rendering, edited by S. J.
Gortler and K. Myszkowski, pp. 274–285. Aire-la-Ville, Switzerland: Eurographics Association,
2001.
[1194] G. Wald. “Human Vision and the Spectrum.” Science 101, pp. 653–658.
[1195] M. M. Waldrop. “Brilliant Displays.” Scientific American 297:3 (2007), 94–97.
[1196] H. Wallach. “Brightness Constancy and the Nature of Achromatic Colors.” Journal of Exper-
imental Psychology 38 (1948), 310–324.
[1197] G. Walls. “The Filling-In Process.” American Journal of Optometry 31 (1954), 329–340.
[1198] E. Walowit, C. J. McCarthy, and R. S. Berns. “Spectrophotometric Color Matching Based on
Two-Constant Kubelka-Munk Theory.” Color Research and Application 13:6 (1988), 358–362.
[1199] J. W. T. Walsh. Photometry, Third edition. New York: Dover Publications, Inc., 1965.
[1200] B. Walter. “Notes on the Ward BRDF.” Technical Report PCG-05-06, Cornell Program of
Computer Graphics, Ithaca, NY, 2005.
[1201] B. A. Wandell and J. E. Farrell. “Water into Wine: Converting Scanner RGB to Tristimulus
XYZ [1909-11].” In Proceedings of SPIE 1909, pp. 92–101. Bellingham, WA: SPIE, 1993.
[1202] B. A. Wandell, H. A. Baseler, A. B. Poirson, G. M. Boynton, and S. A. Engel. “Computational
Neuroimaging: Color Tuning in Two Human Cortical Areas Measured using fMRI.” In Color
Vision: From Genes to Perception, edited by K. R. Gegenfurtner and L. Sharpe, pp. 269–282.
Cambridge, UK: Cambridge University Press, 1999.
[1203] B. A. Wandell. “Color Rendering of Color Camera Data.” Color Research and Application
11 (1986), S30–S33.
[1204] B. A. Wandell. “The Synthesis and Analysis of Color Images.” IEEE Transactions on Patterns
Analysis and Machine Intelligence 9:1 (1987), 2–13.
[1205] B. A. Wandell. Foundations of Vision. Sunderland, Massachusetts: Sinauer Associates, inc.,
1995.
[1206] J. Y. A. Wang and E. H. Adelson. “Representing Moving Images with Layers.” IEEE Trans-
actions on Image Processing 3:5 (1994), 625–638.
[1207] L. Wang, L.-Y. Wei, K. Zhou, B. Guo, and H.-Y. Shum. “High Dynamic Range Image Hallu-
cination.” In Rendering Techniques ’07 (Proceedings of the Eurographics Symposium on Ren-
dering), 2007.
[1208] G. Ward and E. Eydelberg-Vileshin. “Picture Perfect RGB Rendering using Spectral Prefilter-
ing and Sharp Color Primaries.” In Thirteenth Eurographics Workshop on Rendering, edited by
P. Debevec and S. Gibson, pp. 123–130. Aire-la-Ville, Switzerland: Eurographics Association,
2002.

i i

i i
 i i

i i

Bibliography 1021

[1209] G. Ward and M. Simmons. “Subband Encoding of High Dynamic Range Imagery.” In APGV
’04: Proceedings of the 1st Symposium on Applied Perception in Graphics and Visualization,
pp. 83–90. New York: ACM Press, 2004.
[1210] G. Ward and M. Simmons. “JPEG-HDR: A Backwards Compatible, High Dynamic Range
Extension to JPEG.” In Proceedings of the Thirteenth Color Imaging Conference, pp. 283–290.
Springfield, VA: Society for Imaging Science and Technology, 2005.
[1211] G. Ward, H. Rushmeier, and C. Piatko. “A Visibility Matching Tone Reproduction Operator
for High Dynamic Range Scenes.” IEEE Transactions on Visualization and Computer Graphics
3:4 (1997), 291–306.
[1212] G. Ward. “Measuring and Modeling Anisotropic Reflection.” Proc. SIGGRAPH ’92, Com-
puter Graphics 26:2 (1992), 265–272.
[1213] G. Ward. “A Contrast-Based Scalefactor for Luminance Display.” In Graphics Gems IV,
edited by P. Heckbert, pp. 415–421. Boston: Academic Press, 1994.
[1214] G. Ward. “Fast, Robust Image Registration for Compositing High Dynamic Range Pho-
tographs from Hand-Held Exposures.” journal of graphics tools 8:2 (2003), 17–30.
[1215] C. Ware. “Color Sequences for Univariate Maps: Theory, Experiments, and Principles.” IEEE
Computer Graphics and Applications 8:5 (1988), 41–49.
[1216] C. Ware. Information Visualization: Perception for Design. Morgan Kaufmann, 2000.
[1217] H. Wässle and B. B. Boycott. “Functional Architecture of the Mammalian Retina.” Physio-
logical Reviews 71:2 (1991), 447–480.
[1218] H. Wässle, U. Grünert, M.-H. Chun, and B. B. Boycott. “The Rod Pathway of the Macaque
Monkey Retina: Identification of AII-Amacrine Cells with Antibodies against Calretinin.” Jour-
nal of Comparative Neurology 361:3 (1995), 537–551.
[1219] A. Watson and A. Ahumada. “A Standard Model for Foveal Detection of Spatial Contrast.”
Journal of Vision 5:9 (2005), 717–740.
[1220] A. Watson and J. Solomon. “A Model of Visual Contrast Gain Control and Pattern Masking.”
Journal of the Optical Society of America 14 (1997), 2378–2390.
[1221] A. Watson. “The Spatial Standard Observer: A Human Vision Model for Display Inspection.”
SID Symposium Digest of Tehcnical Papers 37 (2006), 1312–1315.
[1222] R. A. Weale. “Human Lenticular Fluorescence and Transmissivity, and Their Effects on Vi-
sion.” Experimental Eye Research 41:4 (1985), 457–473.
[1223] R. A. Weale. “Age and Transmittance of the Human Crystalline Lens.” Journal of Physiology
395 (1988), 577–587.
[1224] M. S. Weaver, V. Adamovich, M. Hack, R. Kwong, and J. J. Brown. “High Efficiency and
Long Lifetime Phosphorescent OLEDs.” In Proceedings of the International Conference on
Electroluminescence of Molecular Materials and Related Phenomena, pp. 10–35. Washington,
DC: IEEE, 2003.
[1225] E. H. Weber. “De pulsu, resorptione, auditu et tactu.” In Annotationes Anatomicae et Physio-
logicae. Lipsiae: C F Köhler, 1834.
[1226] E. H. Weber. “On the Sense of Touch and Common Sensibility.” In E H Weber On the Tactile
Senses, edited by H. E. Ross and D. J. Murray, Second edition. Hove, UK: Erlbaum, 1996.
[1227] M. A. Webster and J. D. Mollon. “Color Constancy Influenced by Contrast Adaptation.”
Nature 373:6516 (1995), 694–698.
[1228] M. Webster. “Light Adaptation, Contrast Adaptation, and Human Colour Vision.” In Colour
Perception: Mind and the Physical World, edited by R. Mausfeld and D. Heyer, pp. 67–110.
Oxford, UK: Oxford University Press, 2003.

i i

i i
 i i

i i

1022 Bibliography

[1229] M. Webster. “Pattern-Selective Adaptation in Color and Form Perception.” In The Visual
Neurosciences, Volume 2, edited by L. Chalupa and J. Werner, pp. 936–947. Cambridge, MA:
MIT Press, 2004.
[1230] A. R. Weeks. Fundamentals of Image Processing. New York: Wiley-IEEE Press, 1996.
[1231] B. Weiss. “Fast Median and Bilateral Filtering.” ACM Transactions on Graphics 25:3 (2006),
519–526.
[1232] E. W. Weisstein. “Full Width at Half Maximum.” In MathWorld — A Wolfram Web Resource.
Wolfram Research, 1999. Available online (http://mathworld.wolfram.com/).
[1233] E. W. Weisstein. “Mie Scattering.” In MathWorld — A Wolfram Web Resource. Wolfram
Research, 1999. Available online (http://mathworld.wolfram.com/).
[1234] S.-W. Wen, M.-T. Lee, and C. H. Chen. “Recent Development of Blue Fluorescent OLED
Materials and Devices.” IEEE/OSA Journal of Display Technology 1:1 (2005), 90–99.
[1235] J. Weng, P. Cohen, and M. Herniou. “Camera Calibration with Distortion Models and Ac-
curacy Evaluation.” IEEE Transactions on Pattern Analysis and Machine Intelligence 14:10
(1994), 965–980.
[1236] S. Wesolkowski, S. Tominaga, and R. D. Dony. “Shading and Highlight Invariant Color Image
Segmentation using the MPC Algorithm.” In Proceedings of the SPIE (Color Imaging: Device-
Independent Color, Color Hard Copy and Graphic Arts VI), pp. 229–240. Bellingham, WA:
SPIE, 2001.
[1237] R. S. West, H. Konijn, W. Sillevis-Smitt, S. Kuppens, N. Pfeffer, Y. Martynov, Y. Takaaki,
S. Eberle, G. Harbers, T. W. Tan, and C. E. Chan. “High Brightness Direct LED Backlight for
LCD-TV.” SID Symposium Digest of Technical Papers 34:1 (2003), 1262–1265.
[1238] G. Westheimer. “Visual Acuity.” Annual Review of Psychology 16 (1965), 359–380.
[1239] G. Westheimer. “Visual Acuity.” In Adler’s Physiology of the Eye, edited by R. A. Moses and
W. M. Hart. St. Louis: The C V Mosby Company, 1987.
[1240] S. Westin, J. Arvo, and K. Torrance. “Predicting Reflectance Functions from Complex Sur-
faces.” In Proceedings SIGGRAPH ’99, Computer Graphics Proceedings, Annual Conference
Series, pp. 255–264. Reading, MA: Addison-Wesley Longman, 1999.
[1241] M. White. “A New Effect on Perceived Lightness.” Perception 8:4 (1979), 413–416.
[1242] T. Whitted. “An Improved Illumination Model for Shaded Display.” Communications of the
ACM 23:6 (1980), 343–349.
[1243] H. Widdel, D. L. Grossman, D. L. Post, and J. Walraven, editors. Color in Electronic Displays.
Dordrecht, Netherlands: Kluwer Academic / Plenum Publishers, 1992.
[1244] T. N. Wiesel and D. H. Hubel. “Spatial and Chromatic Interactions in the Lateral Geniculate
Body of the Rhesus Monkey.” Journal of Neurophysiology 29 (1966), 1115–1156.
[1245] D. R. Williams and A. Roorda. “The Trichromatic Cone Mosaic in the Human Eye.” In Color
Vision: From Genes to Perception, edited by K. R. Gegenfurtner and L. T. Sharpe, pp. 113–122.
Cambridge, UK: Cambridge University Press, 1999.
[1246] T. L. Williams. The Optical Transfer Function of Imaging Systems. Bristol, UK: Institute of
Physics Publishing, 1999.
[1247] S. J. Williamson and H. Z. Cummins. Light and Color in Nature and Art. New York: John
Wiley and Sons, 1983.
[1248] E. N. Wilmer and W. D. Wright. “Colour Sensitivity of the Fovea Centralis.” Nature 156
(1945), 119–121.
[1249] K. Witt. “Modified CIELAB Formula Tested using a Textile Pass/Fail Data Set.” Colour
Research and Application 19 (1994), 273–285.

i i

i i
 i i

i i

Bibliography 1023

[1250] L. B. Wolff. “Polarization-Based Material Classification from Specular Reflection.” IEEE

Transactions on Patterns Analysis and Machine Intelligence 12:11 (1990), 1059–1071.
[1251] M. Wong-Riley. “Columnar Cortico-Cortical Interconnections within the Visual System of
the Squirrel and Macaque Monkeys.” Brain Research 162:2 (1979), 201–217.
[1252] H.-S. P. Wong. “CMOS Imge Sensors — Recent Advances and Device Scaling Considera-
tions.” In International Electron Devices Meeting Technical Digest, pp. 201–204, 1997.
[1253] W. Wright. “A Red-Determination of the Trichromatic Coefficients of the Spectral Colours.”
Trans. Opt. Soc. 30 (1928–1929), 141–161.
[1254] W. D. Wright. “The Characteristics of Tritanopia.” Journal of the Optical Society of America
42 (1952), 509–520.
[1255] W. Wright. “Why and How Chromatic Adaptation Has Been Studied.” Color Research and
Application 6 (1981), 147–152.
[1256] W. Wu and J. P. Allebach. “Imaging Colorimetry using a Digital Camera.” Journal of Imaging
Science and Technology 44:4 (2000), 267–279.
[1257] S. Wurmfeld. “Color in Abstract Painting.” In Color for Science, Art and Technology, edited
by K. Nassau. Amsterdam: Elsevier Science B.V., 1998.
[1258] D. R. Wyble and R. Berns. “A Critical Review of Spectral Models Applied to Binary Color
Printing.” Color Research and Application 25 (2000), 4–19.
[1259] D. R. Wyble and M. R. Rosen. “Color Management of DLP Projectors.” In Proceedings of the
12th IS&T/SID Color Imaging Conference, pp. 228–232. Springfield, VA: Society for Imaging
Science and Technology, 2004.
[1260] D. R. Wyble and H. Zhang. “Colorimetric Characterization Model for DLP Projectors.” In
Proceedings of the 11th IS&T/SID Color Imaging Conference, pp. 346–350. Springfield, VA:
Society for Imaging Science and Technology, 2003.
[1261] G. Wyszecki and W. S. Stiles. Color Science: Concepts and Methods, Quantitative Data and
Formulae. New York: John Wiley & Sons, 1982.
[1262] G. Wyszecki and W. S. Stiles. Color science: Concepts and methods, quantitative data and
formulae, Second edition. John Wiley and Sons, Ltd, 2000.
[1263] K. Xiao, C. Li, M. Luo, and C. Taylor. “Color Appearance for Dissimilar Sizes.” In IS&T
CGIV, 2nd European Conference on Color in Graphics, Imaging, and Vision, pp. 12–16. Spring-
field, VA: Society for Imaging Science and Technology, 2004.
[1264] C. Xu, W. Zhang, W.-H. Ki, and M. Chan. “A 1.0-V VDD CMOS Active-Pixel Sensor
with Complementary Pixel Architecture and Pulsewidth Modulation Fabricated with a 0.25 μ m
CMOS Process.” IEEE Journal of Solid-State Circuits 37:12 (2002), 1853–1859.
[1265] N. H. Yabuta and E. M. Callaway. “Functional Streams and Local Connections of Layer 4C
Neurons in Primary Visual Cortex of the Macaque Monkey.” Journal of Neuroscience 18:22
(1998), 9489–9499.
[1266] O. Yadid-Pecht and E. R. Fossum. “Wide Intrascene Dynamic Range CMOS APS using Dual
Sampling.” IEEE Transactions on Electron Devices 44:10 (1997), 1721–1723.
[1267] H. Yamaguchi and M. D. Fairchild. “A Study of Simultaneous Lightness Perception for Stim-
uli with Multiple Illumination Levels.” In IS&T/SID 12th Color Imaging Conference, pp. 22–28.
Springfield, VA: Society for Imaging Science and Technology, 2004.
[1268] M. Yamaguchi, T. Ajito, and N. Ohyama. “Multiprimary Color Display using Holographic
Optical Element.” In Proceedings of the SPIE 3293, edited by S. A. Benton, pp. 70–77. Belling-
ham, WA: SPIE, 1998.

i i

i i
 i i

i i

1024 Bibliography

[1269] M. Yamaguchi, H. Haneishi, H. Fukuda, J. Kishimoto, H. Kanazawa, M. Tsuchida, R. Iwama,

and N. Ohyama. “High-Fidelity Video and Still-Image Communication based on Spectral Infor-
mation: Natural Vision System and Its Applications.” In Proceedings of SPIE 6062, edited by
M. R. Rosen, F. H. Imai, and S. Tominaga, pp. 60620G–1–60620G–12. Bellingham, WA: SPIE,
2006.
[1270] M. Yamauchi. “Jones-Matrix Models for Twisted-Nematic Liquid-Crystal Devices.” Applied
Optics 44:21 (2005), 4484–4493.
[1271] J. N. Yang and L. T. Maloney. “Illuminant Cues in Surface Color Perception: Tests of Three
Candidate Cues.” Vision Research 41 (2001), 2581–2600.
[1272] D. X. D. Yang, A. E. Gamal, B. Fowler, and H. Tian. “A 640 × 512 CMOS Image Sensor
with Ultrawide Dynamic Range Floating-Point Pixel-Level ADC.” IEEE Journal of Solid-State
Circuits 34:12 (1999), 1821–1834.
[1273] T. Yasuda, T. Hamamoto, and K. Aizawa. “Adaptive-Integration-Time Sensor with Real-Time
Reconstruction Function.” IEEE Transactions on Electron Devices 50:1 (2003), 111–120.
[1274] P. Yeh. “Extended Jones Matrix Method.” Journal of the Optical Society of America 72:4
(1982), 507–513.
[1275] H. Yersin. “Triplet emitters for OLED applications. Mechanisms of exciton trapping and
control of emission properties.” In Transition Metal and Rare Earth Compounds, Topics in
Current Chemistry 241, edited by H. Yersin. Berlin: Springer-Verlag, 2004.
[1276] M.-S. Yoo, J.-N. Kim, D.-J. Woo, S.-H. Yim, and Y.-H. Cho. “Development of Performance
and Cost-Oriented HD PDP TV with High Picture Quality using a High Efficiency Hexagonal
Array Structure.” Journal of the Society for Information Display 12 (2004), 251–256.
[1277] Y. Yoshida and Y. Yamamoto. “Color Calibration of LCDs.” In Proceedings of the IS&T/SID
Tenth Color Imaging Conference, pp. 305–311. Springfield, VA: Society of Imaging Science and
Technology, 2002.
[1278] A. Yoshida, V. Blanz, K. Myszkowski, and H.-P. Seidel. “Perceptual Evaluation of Tone
Mapping Operators with Real-World Scenes.” In Proceedings of SPIE 5666 (Human Vision and
Electronic Imaging X), pp. 192–203. Bellingham, WA: SPIE, 2005.
[1279] A. Yoshida, R. Mantiuk, K. Myszkowski, and H.-P. Seidel. “Analysis of Reproducing Real-
World Appearance on Displays of Varying Dynamic Range.” Computer Graphics Forum 25:3
(2006), 415–426.
[1280] K. Yoshikawa, Y. Kanazawa, M. Wakitani, T. Shinoda, and A. Ohtsuka. “A Full Color AC
Plasma Display with 256 Gray Scale.” In Proceedings of Asia Display, pp. 605–608. San Jose,
CA: Society for Information Display, 1992.
[1281] T. Yoshioka, J. B. Levitt, and J. S. Lund. “Independence and Merger of Thalamocortical Chan-
nels within Macaque Monkey Primary Visual Ccortex: Anatomy of Interlaminar Projections.”
Visual Neuroscience 11:3 (1994), 467–489.
[1282] I. T. Young, J. J. Gerbrands, and L. J. van Vliet. “Fundamentals of Image Processing (Online
Image Processing Course at www.ph.tn.tudelft.nl/Courses/FIP/).”, 2006.
[1283] T. Young. “On the Theory of Light and Colors.” Philosophical Transactions of the Royal
Society of London 92 (1802), 20–21. Reprinted in David L MacAdam (ed.), Selected Papers on
Colorimetry — Fundamentals, SPIE Optical Engineering Press, 1993.
[1284] F. H. Yu and H. S. Kwok. “Comparison of Extended Jones Matrices for Twisted Nematic
Liquid-Crystal Displays at Oblique Angles of Incidence.” Journal of the Optical Society of
America A 16:11 (1999), 2772–2780.
[1285] Y. Yu, P. Debevec, J. Malik, and T. Hawkins. “Inverse Global Illumination: Recovering Re-
flectance Models of Real Scenes from Photographs.” In Proceedings SIGGRAPH ’99, Computer
Graphics Proceedings, Annual Conference Series, pp. 215–224. Reading, MA: Addison-Wesley
Longman, 1999.

i i

i i
 i i

i i

Bibliography 1025

[1286] M. Yukie and E. Iwai. “Direct Projection from the Dorsal Lateral Geniculate Nucleus to
the Prestriate Cortex in Macaque Monkeys.” Journal of Comparative Neurology 201:1 (1981),
81–97.
[1287] J. A. C. Yule. Principles of Color Reproduction. New York: Wiley, 1967.
[1288] Q. Zaidi, B. Spehar, and M. Shy. “Induced Effects of Background and Foregrounds in Three-
Dimensional Configurations: The Role of T-Junctions.” Perception 26 (1997), 395–408.
[1289] S. A. Zeki. “A Century of Cerebral Achromatopsia.” Brain 113:6 (1990), 1721–1777.
[1290] S. A. Zeki. Vision of the Brain. Oxford, UK: Blackwell Publishers, 1993.
[1291] S. A. Zeki. Inner Vision: An Exploration of Art and the Brain. Oxford, UK: Oxford University
Press, 2000.
[1292] X. M. Zhang and B. Wandell. “A Spatial Extension to CIELAB for Digital Color Image
Reproductions.” Proceedings of the SID Symposium Digest 27 (1996), 731–734.
[1293] R. Zhang, P. Tsai, J. Cryer, and M. Shah. “Shape from Shading: A Survey.” IEEE Transactions
on Pattern Analysis and Machine Intelligence 21:8 (1999), 690–706.
[1294] X. Zhu and S.-T. Wu. “Overview on Transflective Liquid Crystal Displays.” In The 16th
Annual Meeting of the IEEE Lasers and Electro-Optical Society, Volume 2, pp. 953–954. Wash-
ington, DC: IEEE Press, 2003.
[1295] C. Ziegaus and E. W. Lang. “Statistics of Natural and Urban Images.” In Proceedings of the
7th International Conference on Artificial Neural Networks, Lecture Notes in Computer Science
1327, pp. 219–224. Berlin: Springer-Verlag, 1997.
[1296] C. Ziegaus and E. W. Lang. “Statistical Invariances in Artificial, Natural and Urban Images.”
Zeitschrift für Naturforschung A 53a:12 (1998), 1009–1021.
[1297] S. Zigman. “Vision Enhancement using a Short Wavelength Light Absorbing Filter.” Optom-
etry and Vision Science 67:2 (1990), 100–104.
[1298] M. D’Zmura and G. Iverson. “Color Constancy. I. Basic Theory of Two-Stage Linear Recov-
ery of Spectral Description for Lights and Surfaces.” Journal of the Optical Society of America
A 10:10 (1993), 2148–2165.
[1299] M. D’Zmura and G. Iverson. “Color Constancy. II. Results for Two-Stage Linear Recovery
of Spectral Description for Lights and Surfaces.” Journal of the Optical Society of America A
10:10 (1993), 2166–2180.
[1300] M. D’Zmura and P. Lennie. “Mechanisms of Color Constancy.” Journal of the Optical Society
of America A 3:10 (1986), 1662–1673.
[1301] M. D’Zmura, G. Iverson, and B. Singer. “Probabilistic Color Constancy.” In Geometric Rep-
resentations of Perceptual Phenomena, pp. 187–202. Mahwah, NJ: Lawrence Erlbaum, 1995.
[1302] J. A. Zuclich, R. D. Glickman, and A. R. Menendez. “In Situ Measurements of Lens Fluo-
rescence and Its Interference With Visual Function.” Investigative Ophthalmology and Visual
Science 33:2 (1992), 410–415.

i i

i i