Aesthetic Complexity Practice and Percep

Aesthetic Complexity:
Practice and Perception

in Art & Design
Guy Birkin
A thesis submitted in partial fulfilment of the requirements of

Nottingham Trent University for the degree of Doctor of Philosophy
October 2010
Copyright Notice
This work is the intellectual property of the author. You may copy up to 5% of this
work for private study, or personal, non-commercial research. Any re-use of the
information contained within this document should be fully referenced, quoting the author,
title, university, degree level and pagination. Queries or requests for any other use, or if a
more substantial copy is required, should be directed in the owner of the Intellectual
Property Rights.
ii
Acknowledgements
This thesis is dedicated to Mum for her unfailing support and to Dad for instilling an
appreciation of nature and craft.
To the following people I am particularly grateful:

The Vice-Chancellor of Nottingham Trent University, Professor Neil T. Gorman,
who provided the bursary that funded this research project.
My supervisory team, Judith Mottram, Terry Shave and Andrew Dunn, whose
knowledge, experience and patience I could not have done without.
All the test participants, who must remain anonymous in this publication.
The people who generously lent their artwork for testing: Oliver Avison,
Isaac Avison, Jane Bartholomew, Amanda Briggs-Goode, Mike Ford,
Stuart Grenville-Price, Bommsoon Lee, Simon Lewis, Geoff Litherland, Judith Mottram,
Amanda Smith, Derek Sprawson, Alex Stevenson, Kath Townsend, Jonathan Willett.
The psychologists who provided valuable advice and papers: Don C. Donderi,
Alexandra Forsythe, Paul J. Locher, Rickard Nicki and Rolf Reber.
The artists and musicians who kindly answered my questions about aesthetics and
complexity: Taylor Deupree, Mark Fell and Robert Henke.
The support staff: Julie Bradshaw, Donna Swann and Melanie Bentham-Hill.
My fellow research students, especially: Jonathan Willett, Honar Issa and
Matthew Malpass, for their advice, assistance and friendship.
My friends in Sneinton and beyond who have maintained my sanity (and tested it).
Special thanks to Kirstie for her support in difficult times.
iii
Abstract
My research investigates the aesthetics of visual complexity in the practice and
perception of visual art and design. The aim is to understand visual complexity in terms of
the relationship between the objective properties of images and subjective properties of
perception. I take a computational and empirical approach to this subject, incorporating
methods from information theory, computer graphics, complexity theory and experimental
psychology. For testing, I create cellular automata programs to generate stimulus images,
and borrow other types of visual material from students and professional artists, designers
and craftspeople. Visual complexity is measured in two ways: Firstly, an objective measure
of complexity is based on the compression of digital image files, which provides an
information-based scale of order to randomness. Secondly, psychophysical techniques are
employed to measure the subjective complexity of the images and other aesthetic
judgements. Research in complex systems theory and experimental aesthetics suggests that
we can expect an inverted U correlation between the two measures of complexity.
This project makes an original contribution to knowledge with empirical evidence for
the hypothetical correlation of information-based and perceived complexity. With cellular
automata images from simple to complex the results show an inverted U correlation; the
measures diverge as images approach randomness. The file compression measure fares less
well with art and design images in these tests, however, perhaps because of the wide variety
of visual material. Preference is more variable than judgements of complexity, and arttrained participants rated images higher than untrained participants. The implication is that
although the file compression measure does not entirely correspond with human
perception, the correlation we have found tells us that we can understand visual complexity
as a mixture of order and chaos. A balance of complexity allows for visual exploration and
pattern-finding which contributes to aesthetic value. The findings also provide a basis for
creative experimentation in art and design practice.
iv
Contents
Aesthetic Complexity: Practice and Perception in Art & Design
Copyright Notice............................................................................................................................... ii
Acknowledgements.......................................................................................................................... iii
Abstract.............................................................................................................................................. iv
Contents.............................................................................................................................................. v
Chapter 1
Aesthetics & Complexity
Introduction ............................................................................................................................ 1
Visual Complexity........................................................................................................ 2
Background: Art Practice and Research ................................................................... 7
Research Focus ..........................................................................................................10
Aesthetics............................................................................................................................... 13
Perception ...................................................................................................................13
Beauty .......................................................................................................................... 14
Art and Design ........................................................................................................... 15
Taste............................................................................................................................. 16
The Aesthetic..............................................................................................................18
Working Model of the Aesthetic ............................................................................. 21
Complexity ............................................................................................................................ 24
Identifying Complexity .............................................................................................26
Computational Complexity............................................................................ 28
Cellular Automata ........................................................................................... 29
Mapping Complexity.................................................................................................31
Uniformity........................................................................................................34
Order and Information .................................................................................. 34
Fractals..............................................................................................................36
Chaos and Randomness................................................................................. 38
Complexity .......................................................................................................40
Measuring Complexity ..............................................................................................42
Measuring Visual Complexity........................................................................ 44
Project Outline...................................................................................................................... 46
Planned Contribution to Knowledge...................................................................... 47

Chapter 2
Mapping Theory and Practice
Introduction .......................................................................................................................... 49
Aesthetic Theory................................................................................................................... 50
Perceptual Aesthetics ................................................................................................51
Environmental Aesthetics and Affordances ............................................... 52
Aesthetic Theories of Art ......................................................................................... 55
The Open Concept of Art ............................................................................. 56
The Decline of Perception in Aesthetic Theory......................................... 57
The Institutional Theory of Art.................................................................... 59
Visual Models of Art ......................................................................................60
Relational Aesthetics.......................................................................................64
Summary .....................................................................................................................66
Empirical Aesthetics ............................................................................................................ 68
Establishment of the Field .......................................................................................68
The Measurement of Perceptions................................................................. 69
The Measurement of Perceived Complexity............................................... 75
The Influence of Information Theory.................................................................... 81
Concepts from Information Theory ............................................................ 82
Visual Order and Randomness ..................................................................... 84
The Influence of the Computer.................................................................... 87
The Computational Approach to Visual Perception ................................. 91
Aesthetic Measures .................................................................................................... 95
Eye-Tracking....................................................................................................95
Aesthetic Preference.......................................................................................96
Edge Detection ............................................................................................... 98
Fractal Measures..............................................................................................99
Data Compression ........................................................................................102
Summary ...................................................................................................................106
Aesthetic Practice ...............................................................................................................107
Simplicity and Minimalism .....................................................................................108
The Grid....................................................................................................................116
vi
Computer Art ...........................................................................................................119

Fractal Art .................................................................................................................122
Visualizing Complex Information.........................................................................126
Audio-Visual Art......................................................................................................132
My Art Practice ........................................................................................................138
Summary of Context..........................................................................................................147
Identifying the Components of Visual Complexity............................................151
Visual Complexity and Image File Compression................................................153
Locating Visual Complexity ...................................................................................155
Chapter 3
Approaching Visual Complexity in Art & Design
Research Focus ...................................................................................................................158
Methodology .......................................................................................................................159
Ontology and Epistemology ..................................................................................159
Problems with Empirical Aesthetics.....................................................................160
Philosophical Problems................................................................................160
Practical Problems ........................................................................................163
Methodological Approach......................................................................................165
Methods ...............................................................................................................................167
Quantitative Methods..............................................................................................168
Measuring Objective Complexity ...............................................................168
Measuring Subjective Complexity ..............................................................169
Qualitative Methods ................................................................................................170
Semi-Structured Interviews .........................................................................170
Transcription Coding and Analysis ............................................................170
Plan ............................................................................................................................172
Chapter 4
Investigating the Spectrum of Visual Complexity
Trial 1: Objective Measure ................................................................................................174
Trial 2: Subjective Measure ...............................................................................................178
Test 1: Monochrome CA and Random Images.............................................................180
Method ......................................................................................................................180
Participants.....................................................................................................181
vii
Materials .........................................................................................................181
Procedure .......................................................................................................184
Results .......................................................................................................................184
Data Transformation....................................................................................187
Analysis......................................................................................................................190
File Types .......................................................................................................191
Preference and Complexity Ratings ...........................................................193
Average Responses .......................................................................................193
Influence of Art Training.............................................................................194
Inferential Statistics.......................................................................................196
Statistical significance ...................................................................................199
Summary ...................................................................................................................200
Discussion: Visual Resolution.....................................................................201
Test 2: CA and Random Images in Colour ....................................................................202
Method ......................................................................................................................203
Materials .........................................................................................................203
Participants.....................................................................................................205
Procedure .......................................................................................................205
Results .......................................................................................................................205
Analysis......................................................................................................................206
Effect of Art Experience .............................................................................208
Discussion.................................................................................................................209
Summary ...................................................................................................................211
Chapter 5
Visual Complexity and Aesthetic Judgement
Introduction ........................................................................................................................212
Test 3: Images from Art and Design ....................................................................213
Hypotheses ...............................................................................................................213
Test 3a: Psychology Laboratory .......................................................................................214
Method ......................................................................................................................214
Materials .........................................................................................................214
Stimuli .............................................................................................................216
Participants.....................................................................................................217
viii
Setting .............................................................................................................217
Variables and Units of Measurement.........................................................218
Procedure ..................................................................................................................219
Results .......................................................................................................................219
Discussion.................................................................................................................223
Compression Ratio .......................................................................................223
The Problem with Borders ..........................................................................224
Interviews with Participants..............................................................................................227
Interview Procedure ................................................................................................228
Transcription ............................................................................................................229
Coding .......................................................................................................................229
Structural Coding ..........................................................................................230
Pattern / Focused Coding ...........................................................................230
Results .......................................................................................................................233
Questions about the Task ............................................................................234
Comfort with the Procedure .......................................................................234
Complexity Criteria.......................................................................................235
Preference Criteria ........................................................................................239
Quality Criteria ..............................................................................................242
Art Training ...................................................................................................243
Comments about the Task...........................................................................244
Analysis: Categorizing the Codes ..........................................................................245
Summary ...................................................................................................................248
Test 3b: Art Gallery............................................................................................................249
Method ......................................................................................................................249
Participants.....................................................................................................249
Setting .............................................................................................................250
Procedure ..................................................................................................................251
Results .......................................................................................................................251
Ranked Average Scores................................................................................251
Correlations....................................................................................................253
Test 3 Findings ...................................................................................................................255
Comparing Results between Tests 3a and 3b......................................................255
ix
Combined Results from Tests 3a and 3b .............................................................258

Visualizing Data ............................................................................................258
Artists vs. Non-Artists..................................................................................263
Hypothesis Evaluation ............................................................................................264
Summary ...................................................................................................................265
Chapter 6
Conclusions
Summary of Findings.........................................................................................................267
Answers to Research Questions ............................................................................267
Focal Theory..................................................................................................268
Contextual Theory ........................................................................................269
Data Theory...................................................................................................270
Contribution to Knowledge ...................................................................................272
Implications.........................................................................................................................274
Measures of Complexity .........................................................................................274
Visual Information ..................................................................................................277
The Spectrum of Complexity and the Window of Visibility.............................279
The Spectrum of Complexity and Colour Spectra..............................................282
Maps, Measures and Models of Complexity........................................................283
Impact on My Art Practice................................................................................................285
Future Directions ...............................................................................................................289
References ......................................................................................................................................292
Appendices.....................................................................................................................................305
Appendix A: Test 1 Instructions......................................................................................305
Appendix B: Copyright Law .............................................................................................305
Appendix C: Interview Transcript ...................................................................................306
Appendix D: List of Codes ...............................................................................................310
Appendix E: Coding Results.............................................................................................312
Chapter 1
Introduction
This project investigates the aesthetics of visual complexity. If you look around now,
you will probably see mostly plain and simple areas (such as a clear sky or plain walls),
regular patterns in smaller areas (clothing and furnishings), and perhaps some irregular
patterns (carpet, grass or clouds). Besides anything living (people, pets and plants), the
most complex visual things around are probably objects of aesthetic value. Why artworks
tend to be complex, and how visual complexity relates to aesthetic value, are the questions
I aim to answer in this research.
The first chapter outlines the scope and direction of the research. It begins with a
general introduction to its focus the aesthetics of visual complexity. This includes an
account of the personal context for the research, including a brief description of previous
academic work and the art practice from which the current research developed. My art
practice uses computational models from complexity theory as creative tools, and it
supports the current project as a source of visual material. Before the contextual literature
review in Chapter 2, it is necessary to clarify the meanings of aesthetics and complexity. The
context of the project can be described as the intersection of these two fields. They are
dealt with in two sections in Chapter 1 which summarise the development of the fields,
delineate their main concerns and offer definitions of terms which are used throughout the
thesis. The section on aesthetics also introduces a working model of the aesthetic which is
used in the thesis as an explanatory and analytical tool. This preliminary discussion of the
contextual theory serves to outline the project and develop its research questions.
Introduction
1
Some of the most significant issues are identified earlier in Chapter 1, and are
discussed further throughout the thesis. Chapter 2 presents a review of contextual theory
and practice in the field of aesthetics. It is divided into three sections, covering theoretical
contributions to the field, empirical investigations, and creative aesthetic practice. Chapter
3 identifies the methodology of the project and outlines the chosen methods. Chapter 4
comprises the initial empirical investigations, which focus on the visual complexity of
stimuli based on images of cellular automata programs. The purpose of these tests is to
examine a broad range of visual information in order to obtain an overall picture of the
spectrum of complexity. An initial contribution to knowledge is made by extending the
range of visual complexity under investigation with the chosen methods. The tests in
Chapter 5 develop the findings of the earlier tests and focus on images from contemporary
art and design practices. Chapter 6 summarizes the results and describes how the tests
contribute to knowledge by extending the psychological and computational methods into
the field of art and design. An original contribution to knowledge is made with empirical
evidence for a perceptual threshold of visual complexity: Perceived complexity describes an
inverted U-shape correlation with visual complexity as measured by image file
compression. In conjunction with measures of aesthetic perception, the information-based
measure reveals that complexity can be understood as a balance of order and chaos.
Visual Complexity
A variety of patterns at different levels is a reasonable preliminary characterization of
visual complexity. For example, a painting exhibits patterns at many levels: Below the range
of our perception, patterns of protons, neutrons and electrons form specific chemical
elements which combine into hundreds of different molecules in the pigment, binder and
support. The mechanical and optical properties of these minerals, oils and fibres determine
the paintings construction by the artist and its appearance to the viewer. At the visual level
their crystalline, fluid and woven textures combine to form a record of the artistic process
which is revealed in the layers of paint and the traces of brushstrokes (Figure 1).
Introduction
2
Figure 1 Titian, The Martyrdom of Saint Lawrence (15481557), oil on canvas (crosssection, magnified 120). This section, from part of a figure in a red cloak, reveals
layers of red lacquer and particles of carbon black, lead white and cinnabar pigments
in oil. The opaque white layer reflects light that reaches it through the translucent red
glaze, intensifying the surface colour.
On the surface of a painting we perceive forms, which may be abstract or

representational, well-defined or ambiguous. These forms have colour, texture and position
relative to each other and to the whole picture. The way that these formal relations have
been handled by the artist may be recognised as a style. From the visual level, we may
notice themes in the content of the work that which is signified implicitly as opposed to
being represented explicitly. We may also perceive the relation of this particular work to the
artists oeuvre, and see how it fits the pattern of the artists output. Similarly, we may see
how the work relates to its genre and the rest of the art world, to art history, and if not also
to current political or philosophical concerns, then almost certainly to our own experience.
These are all examples of various patterns at different levels in a painting, and each of them
has a bearing on the perception and judgement of visual art. The perception of these
complex aesthetic patterns and their inter-relations provides the basis of our understanding
and appreciation of visual art and design. At the centre of them all is the level of visual
perception, and this is the focus of the current research project.
Introduction
3
In common with this project, Rudolph Arnheims book Art and Visual Perception
(1974) employs psychological findings to contribute to understanding the aesthetic
perception of art. Using Gestalt psychology, Arnheim explains how we derive meaning
from our sensory impressions. Gestalt means complete form in German, and the theory
consists of laws that describe how we perceive whole shapes by grouping individual
sensory elements. We make sense of art by perceiving forms and patterns from pictorial
elements that coexist in a hierarchy of visual levels. Arnheim describes this kind of
aesthetic comprehension as an attempt to derive generalities from the particular in a work
of art, a task which is
laborious, but not different in principle from trying to describe the nature of other
complex things, such as the physical or mental make-up of living creatures. Art is the
product of organisms and therefore probably neither more nor less complex than
these organisms themselves. (Arnheim, 1974 p.2)
Arnheim suggests that art is as complex as the beings that create it, but how does this
complexity arise? The complexity of living beings stems partly from an accumulation of
evolutionary history, in a progression from simple bacteria to complex social animals. The
parts and mechanisms that endow beings with the property of life are among the most
complex things we know, and they also provide some of the richest visual material, as
anyone knows who has tried to paint flesh, fur, feathers or foliage. 1 Therefore, one source
of complexity in art is the pictorial representation of these visually-complex beings. Nonliving things, on the other hand, tend to be simpler in construction and in appearance.
Amongst both living and non-living things, some of the most common forms of naturallyoccurring pattern are fractals (Figure 2). These self-similar patterns can be seen in clouds,
ferns, coastlines, lightning, snowflakes, and trees (Figure 3).
Drawing and painting can offer a deeper understanding of the visual world than, say, photography, because
they involve a hands-on re-creation of perceptual objects and effects. To draw (make) a picture of a scene is to
draw (take) visual information from it.
Introduction
4
Figure 2 Artificial fractals. These images show variations of two parameters: branching
angle (left to right), and number of branches (top to bottom).
Figure 3 Natural fractals: branches and leaves of oak and beech trees.
Introduction
5
Images of fractals are less common in art than they are in nature. Unlike art, which
often has a variety of different patterns, a fractal image has the same pattern at all levels, so
it can be described more succinctly, and is less complex. Artificial fractals are a little less
complex than natural fractals because they tend to have more regularity, whereas natural
fractals do not repeat exactly, which makes them more complex, and visually more
interesting. Images of natural fractals have even been shown to reduce physiological stress
(Taylor, 2006). Nevertheless, despite this particular physiological benefit they have less
aesthetic value than works of art and design, which are generally more complex. For
example, Mandelbrot (2004a) notes that Hokusais Great Wave off Kanagawa (Figure 4) is
striking partly because it is one of the earliest visualizations of natural fractals in art, but
what makes it more visually interesting is that it also has plain areas, repetitive patterns and
irregular shapes. My hypothesis is that visual complexity sustains interest in an image by
allowing our visual system to engage in pattern-finding and interpretation. The images that
provide the richest visual information for this perceptual interest are those with a variety of
patterns at different levels. In general, then, artworks are more complex than clouds, but by
how much? That is what the project aims to find out.
Figure 4 Katsushika Hokusai, The Great Wave off Kanagawa (18291832), woodcut print.
Introduction
6
To find out how complex art is, we need a way of measuring visual complexity.
Arnheim gives us a clue as to how we might begin to make these measurements through
a process of description. For example, we could start by listing the visual elements in an
image and their formal properties. Instead of a purely qualitative approach like Arnheims,
however, this project employs concepts from information theory to formalize a model of
visual complexity and support a quantitative analysis of images. With a quantitative
approach, the size of a description provides a basic measure of complexity, since something
that is more complex takes longer to describe. In this project, the descriptions are based
on digital image files, which allows for a computational analysis of aesthetics. Conceptual
tools from complexity theory and techniques from experimental psychology are brought to
bear on this computational approach to visual complexity. In this way, we can proceed to
investigate the relation of complexity to the perception and understanding of visual art by
using quantitative measurement of objective and subjective aesthetic properties as a basis
for carrying out tests on a variety of images. How we choose to select or create images for
the tests will be critical to interpreting the data (a matter discussed further in Chapter 3),
but what is most important for this study is that the material should represent a range of
visual complexity and reflect contemporary visual art practice. The next section briefly
describes the role of my own art practice in developing and supporting the current research
focus.
Background: Art Practice and Research

I am led to this research by my interest in the natural world and our relation to it
through art and science. I have always been fascinated by the aesthetic experiences of art
and nature, and by the way that scientific understanding enriches aesthetic perception.
During my first degree, a lecture on the concept of entropy in art led me to investigate the
science of chaos. Part of the appeal lay in the images used to describe these processes,
which came from computational models, so I began to explore ways of using these models
as tools for visual creativity. This was my introduction to cellular automata programs
Introduction
7
composed of many cells that interact according to simple rules to produce surprisingly
complex behaviour. Because cellular automata are easily visualized on grids, these programs
now form a significant part of my art practice, and they are used extensively in this project.
At first I calculated the programs by hand and then later on computers, but of these early
experiments the only exhibited works were needlepoint pieces (Figure 5). Images of these
new computational tools from complexity theory could look dated on digital screens and
prints, but as hand-stitched artefacts they were more original and more visually interesting.
From the use of craft techniques to visualize computational processes, my masters
degree involved an examination of craft and its relation to fine art. For this work, however,
I rejected the use of a computer, for two reasons: Firstly, calculating the programs by hand
makes the point that complex patterns can be generated by very simple processes (a
computer just speeds things up considerably), with the artwork itself acting as visual
evidence. Secondly, making the pieces by hand, instead of mechanical or digital production,
gives a more appealing visual quality the aura of an original artefact. Small imperfections
and natural variations increase the formal complexity of the surface pattern, resulting in a
visual characteristic similar to the appealing irregularity of natural fractals. This led to a reinterpretation of Walter Benjamins (1936) concept of aura in aesthetic terms, and to the
realisation that in digital reproduction there is no change in aura, unlike in mechanical
reproduction, since digital copies are exactly identical. The study culminated in the
suggestion that the visual appeal of the hand-made is a result of the enriched complexity of
surface pattern that comes with manual techniques compared to mechanical or digital
production. In relation to this conjecture, the current project explores the more general
question of how visual complexity relates to aesthetic value.
Looking at the needlepoint pattern in Figure 5, it is not obvious how it is designed,
but it is apparent that it is designed and not just the result of chance occurrences. The
appearance of living things has a similar kind of visual property, which sometimes leads to
the misconception that they too are designed rather than the result of a complex
Introduction
8
evolutionary history. In fact, the needlepoint pattern is based on a cellular automaton

program of just a few simple rules, which starts from a single stitch and gets progressively
more complex.
Figure 5 Elementary Cellular Automata #110 (2003), needlepoint (detail, work in

progress).
In complex systems theory, Stephen Wolfram (2007) and Stuart Kauffman (2003;
2008) argue that the value of computational models such as cellular automata is that they
offer an explanation for the complexity of life on earth that complements the explanatory
power of Darwinian evolutionary theory, which by itself does not account for many of the
complex structures we see in nature: the snowflakes delicate sixfold symmetry tells us
Introduction
9
that order can arise without the benefit of natural selection (Kauffman, 2003 p.47). My
understanding is that although Darwins theory explains the mechanism by which particular
phenotypes are selected over others, it fails to explain how their complex structures are
generated in the first place.
I consider the attraction of visual complexity to be similar to the appeal we find in
natural forms, which are neither too ordered and dull, nor too chaotic and confusing. My
inference is that they hold our interest because their perceptible organisation invites a visual
exploration which engages us aesthetically. The formal structure of an image is what is
represented by the concept of visual complexity. In general, a more complex image
provides a more fertile ground for visual exploration. Therefore, images that are too
complex to understand in one glance invite us to look again, but unless they repay further
exploration by offering a meaningful relation of visual elements (as opposed to a random
arrangement), we are likely to lose interest. I suggest that visual complexity is what holds
our attention, once it is captured initially, by affording an opportunity to exercise our
powerful visual senses.
Research Focus
I am not only interested in finding out what visual complexity is, but also what it does.
So in addition to studying objective formal properties, the project also aims to investigate
the subjective perception of visual complexity, in order to create a balanced study of both
sides of the aesthetic. We need to establish what it is, however, before we can look at what
it does. Therefore, measuring visual complexity and determining its formal properties, as an
initial step, provides a route to understanding how it is produced by the artist and perceived
by the audience. From this starting-point, the research questions are formulated. To
describe the stages of research, I have adapted the model of Philips & Pugh 2 (2000): Firstly,
the focal theory sets out a plausible argument for the scope and purpose of the project;
2
Philips and Pughs model comprises four similarly-named categories focal theory, background theory, data theory
and contribution with a slightly different emphasis on the content of the research stages.
Introduction
10
contextual theory comprises a structured critique of relevant material and a mapping of the
field; this leads to the data theory, which justifies the methods, details the test procedures and
presents the results; finally the contribution interprets the results, identifies their contribution
to knowledge, and discusses the impact of the findings for the field. These stages are
discussed in more detail next, and are summarised in the form of research questions at the
end of this chapter.
The project focus is based on the premise that visual complexity is a significant
aesthetic element in the creation and perception of art and design. The hypothesis is that
visual complexity provides a ground for visual exploration (pattern-finding and
interpretation) that sustains interest and contributes to aesthetic value. Whereas too little
complexity (such as simple repetitive patterns) can be uninteresting, and too much
complexity (chaos or randomness) can be confusing, we tend to find a middle ground of
complexity in the images that hang in our homes and galleries. From this initial
observation, we arrive at the two-sided question of what makes visual complexity
aesthetically valuable, and why valued aesthetic work is often visually complex. This can be
re-formulated into a single research question for the focal theory: What is the relationship
between visual complexity and aesthetic value?
We can begin to develop a structured answer by measuring visual complexity and
aesthetic judgements of a set of visual artefacts. Once a suitable measure of complexity is
established, having tested it for feasibility, the aim is to proceed through a series of
empirical tests on images of varying complexity. A statistical analysis of the data will
determine the correlation between measures of complexity and its perception, and give us a
quantitative answer to the question of aesthetic value and visual complexity. In addition,
qualitative methods can inform these results through interviews with test participants. Due
to the practical difficulty of being able to represent and measure the visual properties of
three-dimensional artefacts, the study is limited to two-dimensional visual material.
Introduction
11
To support this line of enquiry, we need to develop a theoretical base by reviewing

literature from the field and mapping the practices that form the context of this project.
This stage constitutes the contextual theory of the research. Primarily, it involves an
examination of contemporary aesthetics to determine current knowledge of the perception
of visual complexity. This material includes not only the traditional philosophical side of
aesthetics, art discourse and art practice, but also recent scientific contributions from the
emerging field of empirical aesthetics. The early stages of this project identified that the
majority of publications on visual complexity come from scientific fields. Amongst these
diverse topics are: visual development in infants, mapping whale behaviour patterns, and
optimising the readability of radar images and naval charts. The review of empirical
aesthetics necessitates a brief summary of information theory, which underpins the
techniques from experimental psychology and digital image processing that form the basis
of this projects methods. In addition, the literature review covers aspects of complexity
theory that pertain to visual perception, and examples of visual art practice that engage with
or manifest aesthetic complexity. These examples of practice are not limited to the most
complex artworks, but include artefacts and designs that represent samples from a wide
spectrum of visual complexity, from minimalism to horror vacui. 3 The contextual review
answers the question of what is known about the aesthetics of visual complexity, and with
what methods this knowledge can be acquired.
With the contextual theory in place, we are able to make decisions about the data
theory and its research methods: in this case, how to measure the complexity of an image
and the aesthetic judgements of a viewer. Once the methods are selected and trialled, we
can start to design the tests and prepare the necessary apparatus. Only then can we begin to
acquire our data and start the analysis, from which we will be able to determine the
complexity of the visual artefacts, and how their visual complexity relates to aesthetic
judgments. The final analytical stage evaluates the tests according to the methodology,
3
In an artistic composition, horror vacui is the absence of empty spaces (literally fear of emptiness).
Introduction
12
allows refinement of the procedures, and develops a sustained line of enquiry between the
experiments. Knowledge structured from the contextual research is used to interpret the
results and is re-evaluated in light of the projects findings. A final summary assesses the
impact of the new results on the field, identifies remaining gaps in knowledge, and
highlights potential areas for extending the research in the future.
Aesthetics
Before we go any further into the details of this project, it is necessary to clarify its
context to provide a foundation for the remaining research. The focus is visual complexity
how it is used by artists and designers and perceived by an audience. Its context,
therefore, consists of the areas in which the making and perception of art are analysed and
discussed. These areas are united by the term aesthetics. The following section examines the
development of aesthetics as an academic discipline, identifying its central issues and
elaborating their relevance to this project.
Perception
The word aesthetic has its origins in the ancient Greek (aestheta), meaning
things perceptible by the senses, as opposed to (noeta) things thinkable or
immaterial (OED, 2009). Aesthetic is an adjective, whereas aesthetics is the name for the
field of study concerned with aesthetic experiences and aesthetic properties. The field
acquired this modern meaning primarily from Alexander Baumgartens use of the word in
his two-volume book Aesthetica of 1750/1758: things known are to be known [] as the
object of logic; things perceived are to be known [] as the object of the science of
perception, or aesthetic (Baumgarten, in: Bennett, 1996, p.47). At that time, the idea of
aesthetics as a science never really took hold, but the word stuck with what had already
become a philosophical discipline that centred on questions relating to perception.
Sensory perception has a variety of modalities. The traditional division of the senses
into five categories represents the individual natures of these modes of perception: vision,
Aesthetics
13
hearing, smell, taste and touch. This list of five appears to be based on the visible sense
organs: eyes, ears, nose, mouth and skin. Science now shows that each of these organs is
the receptor part of a sensory system that includes specific neural pathways and areas of the
brain. An alternative way of categorizing the senses is by the type of stimulus, which we
could reduce to three: mechanical (hearing, touch), chemical (smell, taste) and
electromagnetic (vision). Perceptually, though, we experience more than five sensory
modes, including temperature, pressure, hunger, pain, balance, and spatial orientation. The
categorisation of sensory modalities is sensitive to the description of their objective and
subjective features, and therefore it is critical to a study of perception, especially one that
tries to relate the physical properties of objects to subjective sensory perceptions. In the
contemporary scientific study of perception, this field is known as psychophysics, and it
provides a basis for the methods in the current project. We will see later on in the
contemporary literature how sensory modalities are treated in relation to aesthetic
questions. Some of the themes explored relate to general issues that apply to all sensory
modes of the aesthetic, but the primary concern of this project is with the perceptual
aesthetics of vision.
Beauty
When we turn our attention to the senses, one of the first things we notice is that
some perceptions are pleasing and some are not. Our name for this sensory pleasure is
beauty, and the objects or events from which this pleasure is derived we call beautiful. The
concept of beauty provided a foundation for aesthetics as a discursive practice that can be
traced from Plato and Aristotle to the beginning of the nineteenth century. Initially, beauty
was considered to be a property of objects that is revealed to us via the senses. In Platos
aesthetics, for example, beauty is an ideal (transcendental) form of reality that is actualised
in objects to greater or lesser degrees: God designs forms, nature or craft makes them, whilst
art represents these things. Plato denounced art because he regarded it as mere imitation of
objects that are themselves imperfect manifestations of beautiful ideal forms; thus art is at a
Aesthetics
14
third remove from reality. In this ancient Greek thought, we see the beginnings of
aesthetics as being concerned with the core subject of perception via the concepts of art
and beauty.
As the discipline of aesthetics develops, the idea of beauty is gradually transformed
from being an objective property of artefacts to a subjective property of perceptions. In the
eighteenth century, the aesthetics of philosophers in the midst of this transformation
such as Shaftesbury, Hutcheson, Burke and Hume are partially subjectivized, but still
remain rooted in objective formal properties (Dickie 1971, p.30). By the nineteenth
century, beauty no longer resides in the object but in the eye of the beholder. The
philosopher Arthur Schopenhauer, for example, thought that anything can be beautiful it
just depends on directing aesthetic attention, and William Blake said Every eye sees
differently. As the eye, such the object. (Frye 1969, p.19). The objective-subjective duality
of perception remains closely connected to issues in contemporary aesthetics, and has a
bearing on the current project in its methodological choices and interpretation of results.
This project does not focus on the concept of beauty per se, but it does investigate
subjective aesthetic judgements of preference (how much an image is liked) and quality
(how good a work of art is) in relation to the perception of visual complexity in art.
Art and Design

Aesthetics is naturally associated with the practices of art and design because the arts
are understood to involve objects and events that please the senses. So in addition to
questions about perception in general, aesthetics uses concepts such as beauty to explore
questions about the perception of visual art and design. This broad category of aesthetics is
also known as the philosophy of art. Art affords aesthetic experiences by presenting
perceptible properties. Aesthetic theories of art attempt to explain how we experience,
understand and evaluate this visual material. Consequently, aesthetics brings together
science and philosophy in the study of the arts. This makes the aesthetics of art a farreaching discipline whose area of study can be approached from many different angles, and
Aesthetics
15
makes for a place of lively debate. How well-integrated are its different areas, and the
extent to which communications between them flow, is presented in the contextual review
that follows this chapter. For now, we note that this combination of art, science and
philosophy means that the aesthetics of art has the potential to be a truly inter-disciplinary
field. The current project is designed to integrate creative, empirical and critical research
methods in an investigation of the aesthetics of visual complexity.
A central occupation of philosophical aesthetics is the definition of art, which can be
divided into two distinct functions: classificatory and evaluative. While the former relates to
the conditions that make an artefact a work of art, the latter deals with what makes a good
work of art. Evaluating aesthetic objects in these terms is fundamental to the practices of
art and design. For aesthetic philosophy, therefore, the criteria for such evaluations provide
a basis for defining art practices and products. Inasmuch as the definition of art relies on
criticism rather than art objects, the aesthetics of art constitutes a form of meta-criticism.
The discourses of art and its aesthetics thus run parallel in a similar way to that of literature
and literary criticism their readerships are quite different (and this is perhaps why so
many books on visual aesthetics have so few illustrations). There was a time when the
Western concept of art was limited to so few practices that it was a relatively easy task to
identify a work of art amongst other objects. Today the range of practice is much wider,
and the criteria employed by artists, critics, buyers and curators are less easy to identify. In
an artworld that supposedly accepts any kind of object or event as a candidate for art
status, being able to recognise and use these criteria is a basic requirement for engaging in
its practice and discourse. We see in Chapter 2 how this situation is addressed in the
contemporary philosophy of art and design.
Taste
In the 18th century, the concepts of beauty and taste occupied a central place in
aesthetics, not only in the philosophy of sensory perception but also in the philosophy of
art. At this time, the meaning of aesthetic is further developed by Immanuel Kant in the
Aesthetics
16
Critique of Judgement (originally published 1790). Kant restricted its use to denote matters of
taste, that is, the judgement of sensory perceptions (Levinson, 2003). He begins his analysis
by saying, The judgement of taste is not cognitive [] and so not logical, but is aesthetic
which means that it is one whose determining ground cannot be other than subjective. (2007,
p.35, original emphasis). It is the subjective aspect of aesthetic judgement that distinguishes
it from judgements of reason and ethics (which had been treated in two previous critiques
by Kant) and that defines the aesthetic as a judgement of taste. Kant describes four types
of subjective judgement: the agreeable, the good, the beautiful and the sublime. Judgements
of the agreeable and the good are both interested (tied to desires), whereas judgements of
the beautiful and the sublime are not, and on this basis they are defined as aesthetic
judgements of taste:
Taste is the faculty of judging an object or mode of representation by means of a
delight or aversion apart from any interest. The object of such a delight is called beautiful.
(Kant 2007, p.42, original emphasis).
Here, the concept of beauty is employed by Kant to distinguish judgements of taste

from rational or ethical judgements. At this point in 18th century aesthetics, beauty has
become a tool of philosophical inquiry rather than a constraint to its focus. In Kants
emphasis, besides beauty, we also find the first conception of aesthetic judgement as being
disinterested, which is one of the defining characteristics of his aesthetic, along with
subjectivity and universality. This formulation of aesthetics is amongst the most influential
and long-lasting, but the concept of taste loses favour in aesthetic discourse as it comes to
be associated with the privileged power relations of authoritative institutions and
individuals.
Today, taste is a term rarely used in talking about art. In a contemporary artworld in
which the aesthetic judgements of gallerists and curators apparently have a great deal of
influence (as evidenced in Art Reviews annual Power 100 list, for example), it is perhaps
worth re-evaluating the concept of taste and re-introducing the term in todays art
Aesthetics
17
discourse. I believe that the concept of taste still has currency in the production and
consumption of aesthetic material, whether it is talked about or not. It would explain the
success of those who make a living based on their judgement and selection of aesthetic
material, such as D.J.s in dance music culture who may develop a higher profile than many
of the artists whose work they present. It would also explain the almost reverential attitude
towards figures such as the late radio D.J. John Peel, whose listeners trusted his judgement
and were willing to appreciate (or endure) the various types of new and experimental music
that he was known for championing. Ostensibly, the purpose of a critic is to proclaim what is
good and what is not, but the value of a critics function improves as we get to know their
taste, whether we share it or not. We could say similar things about the job of the curator.
As the roles of artist and curator become increasingly blurred in contemporary practices,
such as that of Nicolas Bourriaud, whose role we could describe as a D.J. of the artworld
(see Postproduction, 2005), aesthetic judgement remains a significant element in the day-today transactions of art, whether or not we choose to call it taste.
The Aesthetic
As the concepts of beauty and taste lose currency in aesthetic philosophy, they are
gradually replaced by discourse about what defines the aesthetic. In language, aesthetic
serves as a predicate to a particular kind of experience, object, property, concept or
judgement. In trying to construct a definition of the aesthetic, philosophy attempts to
clarify the relationship between these various aspects and to identify which is the most
fundamental. Nick Zangwill (2008) argues for the primacy of aesthetic judgements, as
follows:
Aesthetic properties are those that are ascribed in aesthetic judgments; aesthetic
experiences are those that ground aesthetic judgments; aesthetic concepts are those
that are deployed in aesthetic judgments; and aesthetic words are those that are
typically used in the linguistic expression of aesthetic judgments.
Nevertheless, it is plausible that aesthetic concepts are more fundamental, because if

they are deployed in aesthetic judgements, then they must be prior to and independent of
Aesthetics
18
those judgements, and aesthetic language too. On the other hand, one could also argue for
the primacy of aesthetic experience, without which there would be little to discuss at all. With
the current projects focus on perception, aesthetic experience is perhaps the more
fundamental, but actually this matter is not directly relevant to the project because it has
little practical consequence in terms of its methods. What is most significant is the fact that
all these aspects of the aesthetic are closely interconnected, and we should be mindful of
their relationships in our analysis of visual complexity.
Kants concept of the aesthetic is defined by disinterestedness in practical desires
except for an appreciation of the aesthetic experiences that art affords. Kant describes his
aesthetic as subjective (personal) and universal (normative) (2007, p.42). When we make an
aesthetic judgement, it cannot be anything other than subjective because it is not based on
reason or ethics, but at the same time it is also an evaluation. In this sense, aesthetic
judgements are not just explicit statements of preference; implicitly they are universal
invitations for others to agree with us. If we say I like this wine, we make what Kant
called a judgement of agreeableness, but if we say that painting is beautiful, we make a
judgement of taste. We make such a statement as if beauty were a property of that painting,
even though we also know it as a judgement based on subjective experience. To state that
one likes something is a claim beyond dispute, but to state that something is beautiful is to
make a claim that can be argued against. Aesthetic judgements are thus normative because
they make this claim to truthfulness. Unlike objective universal judgements, which are
based on shared concepts or criteria, aesthetic judgements are grounded in personal
experience. Kants normative aesthetic forms a central element of analytical philosophy,
but in continental philosophy the subjectivity of aesthetic judgement undermines its claim
to truth, which leads to the denial of normativity and to the view that all such judgements
are only relative. We return to this issue in the methodological discussion in Chapter 3,
but for now we note that the current project rejects the continental view and accepts the
normative aspect of the aesthetic.
Aesthetics
19
In philosophical terms, aesthetics tries to define the necessary and sufficient conditions
of the aesthetic, that is, to identify all the properties that things must have to be properly
called aesthetic. An example of this approach is that of Monroe Beardsley (1958), for
whom the value of an aesthetic experience depends on having three properties: unity,
intensity and complexity. For Beardsley, these three properties are the essences of the
aesthetic. The value of this essentialist approach is that it has the potential to say whether
an artefact is aesthetic and why it is worth looking at, but its usefulness crucially depends
on its relation to actual practices in the artworld. From Kant onwards, the field had
become focused on essentialist definitions of the various aspects of the aesthetic
(experiences, properties, etc.) and of art. By the 1950s, the increasing diversification of art
objects and the development of new practices strained the relationship with essentialist
definitions of art. Similarly, perceptual definitions of the aesthetic also suffered amidst an
increasingly conceptual artworld. Beardsleys aesthetic straddles the division between the
theory discussed thus far and the theories dominant in contemporary aesthetics, which are
examined in the following chapter. The dividing line is marked by the influence of
Wittgenstein, who challenged the philosophical practice of essentialist definition.
Beardsleys (1958) theory represents the first response to the neo-Wittgensteinian challenge
to avoid defining art at all (he only defined the aesthetic). We take up this discussion in
the next chapter, beginning with the more constructive responses to this challenge.
To summarise this preliminary discussion of aesthetics, we can say that the field is
chiefly concerned with perceptions (which are natural phenomena), properties (of art and
nature), and practices (of making, perceiving and understanding art). The field is divided
along the lines of the various methods employed in its study and the purposes for which
those studies are undertaken. Here we have introduced the philosophy of aesthetics, and
have focused on the discursive tradition of this field. In the next chapter we develop this
line of enquiry into contemporary culture, and expand the context into the artistic and
scientific explorations of the aesthetic. Broadly speaking, all aesthetics is connected with
Aesthetics
20
the study of perception, since almost everything we know comes to us via the senses, but
discourse in the philosophical field has shifted in parallel with art practice from the
perceptual to the conceptual. This move seems to be out of line with a contemporary
cultural environment that is increasingly designed to be perceived, and in which there
appears to be a tendency towards the aestheticization of everyday objects. The aesthetics of
technological products, for example, is now considered much more than it used to be.
Dieter Rams designs for Braun led the way in this regard; the third of his ten design
principles is good design is aesthetic (Design Museum, 2007). Jonathan Ive presently
continues the trend with sleek, functional and colourful designs for Apple Inc., such as the
iMac, iPod and iPhone. Given this situation, perhaps aesthetics should look again at how we
sense and make sense of the surrounding visual culture, in order to understand a little more
about why we appreciate the things that we do in terms of perception. This is one
motivation for the current project to re-evaluate perceptual aesthetics in general, but to
focus specifically on the property of visual complexity in the practice of art and design.
Working Model of the Aesthetic

I now present a model of the aesthetic that provides the conceptual framework for
this project (Figure 6). The model is based on three key elements artist, artwork and
audience from which a fourth element is developed the artworld. These elements
provide an ontological base for the thesis that allows for the identification of the aesthetic.
Of these four the artwork is the primary element, but it is also important to recognise that
behind every artwork is an artist, and in front of the work there must be an audience.
Including these roles reflects the position that unless an artist engages with the artworld
public, an artefact cannot lay claim to being a work of art. That the audience contributes to
an artwork is a well-established principle, as described by Marcel Duchamp in his lecture
The Creative Act, 1957:
Aesthetics
21
All in all, the creative act is not performed by the artist alone; the spectator brings the
work in contact with the external world by deciphering and interpreting its inner
qualification and thus adds his contribution to the creative act. (Stiles & Selz, 1996,
p.819)
In addition, the model illustrates the idea that art is essentially a form of
communication, and that whatever the medium it always carries a message. The arrows in
the illustration signify both the direction of communication and the temporal sequence of
the process. Communication is an essential element of art, but the current project does not
focus on interpretation and meaning; it is less concerned with semantics than with the syntax
of visual complexity. The scope is limited to the aesthetic perception of visual complexity, a
formal property of the art object, which is why the artwork is central to this model.
Artist
Audience
Artwork
Aesthetic
Aesthetic
Production
Perception
Artworld
Figure 6 The working model of the aesthetic.
In this model, the aesthetic is identified as residing in the interactions with an

artwork firstly by those who produce it and secondly by those who perceive it. The
aesthetic is an active process of experiencing art and making judgements based on the
evaluation of those experiences. The artist uses her judgement in the production of an
artwork to make decisions about how the creation of the work should proceed. More often
than not, this involves a kind of continual adjustment until the aesthetic properties of the
artwork meet the requirements of the intended aesthetic experience that the work is to
Aesthetics
22
afford the audience. In turn, the audience makes aesthetic judgements about an artwork
and the artist based on their sensory perceptions, which include the most basic judgement
of whether to continue looking or not. This projects hypothesis is that visual complexity
sustains aesthetic interest in looking by providing a ground for visual exploration.
Together, the processes of production and perception constitute the artworld the
culture in which art is made and appreciated. In this conceptual model, the artworld is
comprised of many such processes of aesthetic production and perception. Figure 6 shows
only a single artwork for the sake of clarity, but the model symbolizes many artists
producing multiple artworks which are seen by lots of people in various times and
locations. Each of these historical or potential situations constitutes a unique context for
the perception of an artwork which affects the quality of aesthetic experience it affords.
The model does not attempt to represent this kind of detail, but it is implied. Also implied
are the many other roles besides artist and audience that populate the artworld, such as
curator, critic, dealer and gallerist, to name but four. Instead, the model focuses on the core
structure of art practice and is centred on the artwork. We return to the model in Chapter
2, where it is used to structure the methodology, and in Chapter 3 it provides an analytical
tool for aesthetic theory.
A key part of the model is the location of the aesthetic in the interaction between
object (artwork) and subject (artist or audience). This objective-subjective duality is present
in many aspects of the aesthetic described thus far attributes of sensory modes, aesthetic
properties, theories of art, and so on and the way in which we deal with this issue affects
the nature of the investigation. The current project investigates visual complexity in both
objective and subjective terms, and the relation between these two aspects of the aesthetic
is a significant element in the development of this thesis. The focus of the thesis can be
described as the intersection of the fields of aesthetics and complexity. Having introduced
some key concepts of aesthetics, we next take a look at complexity.
Aesthetics
23
Complexity
First, a word about terminology: Complexity stems from the Latin root plexus,
meaning plaited or woven, and so complex literally means plaited together or
interwoven (OED, 2009). In aesthetic terms, what is woven together are the various
patterns at different levels that are available to visual perception, which we described at the
beginning of this chapter. In this thesis, the term complex carries its usual meaning as the
opposite of simple. Complexity comes in degrees, however, and the thesis explores visual
complexity as a scale from simple to complex. So while complex generally refers to one
extreme of this scale, complexity can refer to the whole range. Similarly, complexity
theory is a term used to describe many different areas of study, some of which focus only
on the most complex phenomena whilst others investigate a wide range of complexity.
This project explores a wide range of visual complexity. To avoid ambiguity I occasionally
refer to this range with the phrase spectrum of complexity.
By looking at how complexity is understood, we may begin to learn more about its
manifestation in visual art and design. In this way, the scientific literature on complexity
can contribute to understanding its production and perception in art. The objects of study
in this field of science are complex systems. One of the first people to propose a definition
of complex systems was the economist Herbert Simon, who in 1969 described them as
being made up of a large number of parts that have many interactions (1996, p.183).
Examples of complex systems include an ants nest, the stock market, the weather system
and the brain. Each of these systems displays some kind of autonomy from the workings
of its constituent elements. For example, we do not find the blueprints for the ants nest in
any individual, and yet it emerges in their collective behaviour. Similarly, many people have
tried to find rules that characterize the behaviour of a stock market, the pattern of which
depends on many variable interactions of individual brokers, bankers and businesses. In
contrast to these examples, the objects of study in the current project are two-dimensional
artworks, and these are not complex systems. The artworks under investigation are static
Complexity
24
objects, not dynamic systems. The implication is that what we can learn from complexity
science may be limited by the scope of the project.
Visual complexity does involve complex systems, however, in its perception.
Perceiving visual complexity requires very complex structures (the visual system) and
produces complex phenomena (conscious sensory awareness), but the project focus is
limited to the formal complexity of visual images, and not the complexity of these
biological and psychological mechanisms. Therefore, whilst some of the dynamic aspects of
complexity theory may be inapplicable to this project, we may yet find use for some of its
concepts and models. Indeed, one of the central concerns of complexity science is to
unearth patterns of behaviour that are common to many different systems, such that
understanding mechanisms of complexity in one type of process may inform activity in
other fields. The proof of the applicability of such concepts to other fields is the
responsibility of each individual researcher, and it often requires an element of faith or
intuition to begin such an investigation. Nevertheless, there are suggestions that it is
possible, at least in principle, such as this comment by Herbert Simon which also supports
the hypothesis of the current project: The aesthetics of natural sciences and mathematics
is at one with the aesthetics of music and painting both inhere in the discovery of a
partially concealed pattern (1996, p.2). For the purposes of the present study of visual
complexity,
The field of complex systems provides a number of sophisticated tools, some of them
concepts that help us think about these systems, some of them analytical for studying
these systems in greater depth, and some of them computer based for describing,
modelling or simulating these systems. (Nicolis & Rouvalis-Nicolis, 2007)
The aim of this section is to present some of the useful concepts that can help us think
about visual complexity in preparation for the contextual review in the next chapter. The
review serves to identify the analytical methods appropriate for a study of visual complexity in
the field of art. The present section also introduces computational tools for modelling complexity
which are used in my art practice and in this projects tests.
Complexity
25
Identifying Complexity
When we come across a system that is neither regular nor random when it has a
structure that is not easy to comprehend then we call it complex. The most complex
things we know are living beings. Indeed, life itself appears to be a special property of these
most complex organic structures:
Life is a level of complexity that almost lies outside our vision; is so far beyond
anything we have any means of understanding that we just think of it as a different
class of object, a different class of matter (Adams, 1998)
The same could be said of consciousness of sensory capabilities and mental

faculties in terms of a Cartesian dualist philosophy: The complexity of mental
phenomena seems so far removed from the material realm that we understand mind and
matter as separate classes. To a complexity theorist, however, life and consciousness are
regarded as emergent properties of complex systems, and the physical world does not simply
divide along the lines of our limited understanding. Emergence is a crucial concept in
complexity theory, and though it is difficult to define it can be summed up in the idea that
the whole is greater than the sum of its parts. To take a basic example, the physical state of matter
as solid, liquid or gas is an emergent collective property of a group of molecules. A single
molecule of a substance cannot be said to be in any one state without reference to its
relation to other molecules, and so the concept becomes redundant when a single molecule
is considered in isolation. In terms of visual perception, we could say that the Gestalt
grouping of picture elements into recognisable shapes or patterns is an emergent
phenomenon. A perceived pattern emerges from the relationship of constituent elements,
and cannot be found in the properties of the elements alone.
Often, emergent properties cannot be predicted, even when armed with complete
knowledge of the constituent elements. For that reason, emergence can be understood as
the introduction of novelty to a system; what emerges is something new, not present until the
elements in the system interact. This creative aspect of complex systems is evidenced as
much in the evolutionary history of life on earth as it is in the development of art. In this
Complexity
26
sense, the artworld can be understood as a complex system an idea which forms the
subject of Niklas Luhmanns Art as a Social System (2000). From the complexity-based
theory of biologists Maturana and Varela (1980), Luhmann borrows the concept of
autopoiesis (literally self-making) to describe the self-sustaining organisation of the artworld.
The same concept also applies to the working model of the aesthetic outlined earlier, in
which the artworld is represented as a system composed of many interacting feedback
processes of art production and perception. Autopoiesis can be understood as another
characteristic of complex systems; living things make themselves, whilst other complex
systems sustain yet more systems and beget new properties. Of course, the idea can be
applied to other social systems too (it has found currency in management theory, for
example), and this demonstrates the transferability of knowledge from complexity science
to fields such as art and aesthetics. In transferring ideas from complexity theory we should
always take account of particular details, therefore, but we may also be forced to talk in
more general terms or in terms of patterns at different levels of abstraction.
Some complexity theorists see their discipline as not just a part of traditional science,
but as a new way of approaching scientific problems in general. For example, the title and
contents of Wolframs A New Kind of Science (2002) reflect this aim. Others who share this
view have described complexity theory in terms of Thomas Kuhns (1996) concept of a
paradigm, that is, a shift in the beliefs that undergird the practice of science:
Complexity is emerging as a post-Newtonian paradigm for approaching from a
unifying point of view a large body of phenomena occurring in systems constituted by
several subunits, at the crossroads of physical, engineering, environmental, life and
human sciences. (Nicolis and Rouvas-Nicolis, 2007)
What has changed in this new approach is the scientific view of complex systems,
from being regarded as prohibitively problematic to being legitimate objects of scientific
study. Now that computers offer a means of dealing quickly with large quantities of data,
the way is open for an analysis of complexity in the many different fields in which it arises.
Because it appears in so many different guises, there is no hard and fast definition of
Complexity
27
complexity agreed upon by all partly also because it is still an emerging field. At this stage,
we have identified some of the characteristics of complexity, namely: interactivity, selforganisation, emergence and autopoiesis. While these are useful concepts in the theory of
dynamic systems, they are less applicable to the aesthetics of static two-dimensional art.
The gap between visual art and models of dynamic systems can be bridged with a
computational approach, which is outlined in the following section.
Computational Complexity
Complex systems are the kind of thing that science had struggled to cope with until
the rise of computing, because they were just too complicated to be amenable to analysis.
Digital computers enable the modelling of complex systems, and offer a way of dealing
with problems that were previously intractable: Part of the rise of the complex systems
research agenda can be tied to the use of theoretical computation as a new way to explore
such systems (Nicolis & Rouvalis-Nicolis, 2007). In this context, computers are used less
for their capacity to perform sophisticated calculations than for their ability to carry out a
great number of operations in a very short time. Cellular automata, for example, are based
on very simple operations, but what makes them difficult to compute is the sheer number
of cells and their interactions. So for the study of complex systems, the computer is more
of a labour-saving device than a particularly clever machine. But the computer also
provides a measure of complexity in terms of the amount of time or the number of
operations that it takes to perform a task, and this is what is meant by computational
complexity.
The computational approach to complexity is exemplified by the work of Stephen
Wolfram, a scientist who gained a PhD on particle physics at the age of 20. Wolfram
contributed to the field of complexity mainly through his work on cellular automata, which
are computational models of complex systems. In 1986, he founded the first journal in the
field Complex Systems, but is a somewhat controversial figure in the scientific community
partly because his recent method of disseminating findings does not conform to the usual
Complexity
28
peer-reviewed publication system, choosing instead to self-publish funded by a successful

software business. The two ventures are related, since Wolframs Mathematica software was
initially developed to aid his research into complex systems. Mathematica is used in the
current project to generate images, analyse data and visualize the results.
Published in 2002, A New Kind of Science (NKS) presents Wolframs 20-year
investigation of complexity and computation. Its method is to exhaustively examine various
types of simple programs, such as cellular automata, and to explore their behaviour over a
range of complexity, thus providing a kind of natural history of the computational domain.
NKS shows that different kinds of simple program exhibit the same range of behavioural
patterns, and makes the claim that all computational systems above a certain threshold of
complexity are essentially equivalent. Wolfram proposes that we can understand everything
in the natural world as a computation, and that doing so allows us to apply what we learn
from the study of simple programs to situations in real life. The current project is
supported by Wolframs work, since it uses his ideas about complexity to develop images
for testing which are made with CA programs implemented in Mathematica. Because cellular
automata play a significant part in the context and the content of the current project, we
introduce them next.
Cellular Automata
A cellular automaton is a system of cells that evolve in discrete time steps according
to a set of rules (Weisstein 2009). In the simplest possible CA, each cell can have two
different states, and these can be represented by two colours: black for on and white for
off, for example. At each time step, every cell uses a set of rules to determine its next
state, based on its current state and those of its neighbours. The set of rules has to include
every potential configuration of states, but can often be described simply in words, for
example: If both neighbouring cells share the same state, then remain in the same state;
otherwise change state on the next step. The simplest CA comprises a single row of cells
it is a one-dimensional CA. With states represented as colours, we would see cells change
Complexity
29
state at each time step, but with a 1D CA, instead of watching a row of cells change colour
over time, we could stack the rows in sequence, adding them successively below the last. In
this way, we can visualise the evolution of a 1D CA over a portion of time in a single 2D
image. This is what the rule specification looks like for the CA we described in words:
Figure 7 Visualization of the 1D CA rule If neighbouring cells have different states,

then change state on the next step, rule number 150 of Wolframs elementary CA.
The rule as depicted above can be interpreted as follows: The top part (1 cell + 2
neighbours) specifies all potential current conditions of the cell, and the lower part specifies
the resulting state. This rule is known as number 150 in Wolframs numbering scheme,
derived from the rules unique pattern of cells in the lower half of the rule specification.
Black cells are interpreted as 1 and white as 0 to give the binary number 10010110, which
is 150 in decimal. Figure 8 shows how the rule evolves when starting with initial conditions
of a row of white off cells with a single black on cell in the centre.
Figure 8 Evolution of elementary CA rule 150.
It is often impossible to predict the outcome (the type of pattern that will result)
even when the rules are known and understood. This is because we are talking of an
emergent property. Of the simplest 1D CA, with only two states and two neighbours taken
into account, there are 256 different rules. Wolfram calls these the elementary cellular
Complexity
30
automata (eCA). Because they are relatively few in number, are simple to generate and
understand, and because they display a wide range of complex behaviour/patterns, they
make a good model for the science of complex systems. The fact that they are also easily
visualized means that CA are ideal for use in the current investigation. Next we look at
ways of mapping the variety of CA patterns that comprise a spectrum of visual complexity.
Mapping Complexity
Wolfram (1984) mapped out four levels of complexity based on the visual
appearance of patterns generated by cellular automata:
Class 1. Simple: homogeneous, no variation in structure.
Class 2. Periodic: simple sequences, regular or nested patterns
Class 3. Chaotic: aperiodic, irregular patterns.
Class 4. Complex: complicated localized structures, mixture of order and randomness.
These classes are represented in Figure 9. In the first image, the class 1 CA rule
specifies that every possible configuration leads to a black cell, and the resulting image is
completely black after a few steps. The class 2 CA image has a regular diagonal pattern, and
repeats in a square tessellation. The class 3 image is a non-repeating pattern with an
overall texture but with chaotic variations on a small scale. The unpredictability of this rule
(number 30 in the eCA) allows for its use in Wolframs Mathematica software as the basis of
its random number generator. The last image shows complex structures within regular
patterns, and is the only form of class 4 rule in the elementary CA.
Figure 9 Examples of elementary cellular automata in Wolframs four classes of

complexity.
Complexity
31
Wolframs classification constitutes a qualitative analysis of complexity, in which the

categories are not rigorously defined but are based on the perception of visual properties
that reflect the structure and operation of CA. Li and Packard (1990) examine the
distribution of eCA types in rule space (an abstraction of the possible forms that CA
rules can take), and organise the same rules into five categories (A, B, C, D, and E). Their
classification makes a distinction between two types of periodic (B, C) and chaotic (D, E)
rule, where Wolfram has only one of each (classes 2 and 3), and places Wolframs class 4
rules into either a periodic or chaotic group depending on which aspect of the pattern is
taken into consideration (which reflects the characterisation of class 4 eCA as having both
order and randomness). Li and Packards scheme elaborates, rather than contradicts,
Wolframs classification. For the purposes of the current investigation, Wolframs system is
favourable, since it is based on visual properties, rather than the more mathematical
scheme of Packard and Li. Wolfram also investigated the proportions of the four classes of
complexity in various types of CA. The elementary CA contain approximately 50% class 1,
and 25% each for classes 2 and 3. The results for CA with more colours and more
neighbours are illustrated in Figure 10. It shows that random patterns are the most
common in many of the CA types, and complex patterns are the rarest.
Figure 10 Proportions of complexity classes in various types of CA. Wolfram, 2002,

p.948. (k = number of colours, r = number of neighbours on either side)
Christopher Langtons Ph.D. thesis (1991) comprised a quantitative examination of

the same patterns of behaviour in cellular automata by defining and manipulating a value
he called the (lambda) parameter which measured the activity of the systems. As the value
of the parameter was increased, Langton observed a phase transition between highly ordered
and highly disordered dynamics, analogous to the phase transition between the solid and fluid
Complexity
32
states of matter (1990 p.13, original emphasis). Increasing the lambda parameter is like
turning up the temperature of a system, in which a substance changes from an ordered
crystalline solid to a chaotic boiling liquid. The result is that CAs exhibiting the most
complex behaviour both qualitatively and quantitatively are found generically in the
vicinity of this phase transition (Langton, 1990 p.13). Langton coined a phrase to describe
the location of these most complex structures the edge of chaos. His results show that
as the parameter is increased, the types of pattern encountered can be recognised from
Wolframs classification, except that the sequence in which they appear is: 1, 2, 4, 3 (see
Figure 11). The significant point is that the complex CA (class 4) are found between
regions of order (classes 1 and 2) and disorder (class 3).
Figure 11 Location of the Wolfram classes in Langtons lambda space. (Langton,

1990, p.32)
It is fair to describe the work of Wolfram, Li & Packard, and Langton as mappings of
complexity, because they look at the entire spectrum of complexity and establish what is to
be found where. This classificatory exercise is more than just taxonomy though, because
Langtons mapping is structured according to critical quantitative parameters of the
programs. Cellular automata are particularly useful in this regard because their visual display
allows us to use our sensory perception as a form of analysis (in other words, there appears
to be a correspondence between their computational structure and their appearance).
Complexity
33
Wolframs qualitative mapping is thus augmented by Langtons quantitative measurement,

and the location of different types of visual pattern is given meaning in the context of a
measurable spectrum of complexity. The following sections examine the types of visual
forms and patterns that are mapped out by classification systems of Wolfram and Langton.
Uniformity
Uniform images are the simplest possible. They are not particularly interesting in
themselves but they are important in establishing a base level for a measure of complexity.
All uniform visual arrays are essentially identical in terms of structure they may differ in
colour (type of picture element), but the defining characteristic is that all the elements of a
uniform class I image are identical; there is no differentiation of picture elements.
Therefore, a description of this type of image can be quite short. We need to specify the
colour of the elements just once, and the only other parameter that really makes a
difference to analysis is the number of elements (the dimensions of the array). Altering the
number of elements in a uniform image rarely alters it perceptually, however, since all the
elements are identical. Thus a single-pixel image looks identical to a uniform image of the
same physical size composed of many pixels. The idea that a description of a pattern can be
shortened and simplified is central to the field of information theory. The following section
discusses this concept in relation to class 2 patterns.
Order and Information
In a paper that established the field of information theory, Claude Shannon (1948)
measured information in bits (binary digits), and defined information content in terms of
order, or rather its complement, entropy. In physics, entropy is the disorder of a system; in
information theory, the entropy of a message is inversely proportional to its predictability.
Regular patterns can be predicted because they have order they provide little new
information, and so we can use a compact description of a highly ordered pattern in place
of an element-by-element description of its content. Irregular patterns are hard to predict
and describe, and so they have low order and high information content. This principle of
Complexity
34
information theory provides the basis of data compression in digital image files (Salomon
2004). Those parts of a message that we can afford to discard in a short description are
what is redundant. Regular patterns are redundant and can be compressed, whereas
irregularity is less redundant and less compressible. The size of a compressed file thus
reflects the amount of disorder or information content in the original data. This fact proves
to be highly relevant to the current investigation, as it provides a basis for measuring the
complexity of digital image files, and is taken up again in the following chapters review of
contextual theory and practices. Before we go further, it is worth noting some of the
subtleties of order as understood in information theory.
The information-theory concept of order is not quite as intuitive as it may seem.
Consider the two strings of digits {0, 1, 2, 3, 4, 5, 6, 7, 8, 9} and {1, 0, 3, 2, 5, 4, 7, 6, 9, 8}.
Which is more ordered? We are likely to say the first one, but why? The second string is
constructed from the first by swapping pairs of adjacent digits, so it is not un-structured
there is no randomness here. So why should it seem less ordered? Actually, it is a
misleading situation, because in fact the two strings are structurally identical in terms of
order. The confusion stems from the meaning we attach to the symbols used. The first string
presents the digits in ascending sequence of their numerical value. But in information
theory, we disregard the meaning of symbols and focus only on their quantity and variety.
If we measure information by the length of a description, then each of those two strings
take the same amount of description (for each we must specify ten symbols and ten
locations), and are thus equal in terms of order and information. If we had used ten
meaningless symbols in place of Arabic numerals, we would be unable to perceive any
difference in order between the two strings. Order enters the frame only when there are
multiples of things to be ordered (strictly speaking, when there are more possible positions
than there are types of element). Thus, for example {0, 1, 0, 1, 0, 1, 0, 1, 0, 1} is more
ordered than {0, 0, 0, 1, 0, 1, 1, 0, 0, 1} because we can form a more compact description
of the first string ({0, 1} 5). In the same way that information theory disregards the
Complexity
35
meaning of elements in the measurement of information, the current project disregards the
meaning and interpretation of images, and focuses only on the objective formal properties
of visual complexity and their subjective aesthetic perception.
Fractals
Besides simple repetitive patterns such as stripes and checks, Wolframs class 2 CA
images include another type of ordered image, in which pattern repetitions are nested
within themselves. These intricate self-repeating patterns are called fractals, and are
exemplified by the Mandelbrot set (Figure 12).
Figure 12 The Mandelbrot set, overview (left) and detail (right). The extraordinary
variety of patterns in its infinitely complicated boundary reflects the fact that it is a
superset of many fractals, where each point in the set represents a unique fractal
pattern called a Julia set.
Benoit Mandelbrot brought the geometry of these wiggly, grainy and wrinkled forms
to the field of mathematics, demonstrating how they can be methodically generated and
precisely measured. The value of this new geometry was that it applied to many of the
forms we see in nature, which are poorly modelled by the more familiar and regular
mathematical constructs: Clouds are not spheres, mountains are not cones, coastlines are
not circles, and bark is not smooth, nor does lightning travel in a straight line
(Mandelbrot, 1982, p.1). The computational study of fractal geometry makes possible the
Complexity
36
analysis of forms that used to be viewed as pathologically problematic but which are
common in nature. Traditional Euclidean geometry cannot cope with such rough forms:
the language of mathematics and its characters are triangles, circles and other
geometric figures, without which one wanders about in a dark labyrinth. [] But they
have turned out not to be sufficient, merely because most of the world is of
infinitely great roughness and complexity. (Mandelbrot, 2004b)
Mandelbrot cites Katsushika Hokusais famous print The Great Wave off Kanagawa
(Figure 4) as one of the earliest visualizations of fractals: Hokusai holds a central role in
my current view of fractals as a notion familiar to Man, in one form or another, since time
immemorial (2004b, p.26). Whereas artificial fractals repeat their pattern regularly, natural
fractals tend to be more irregular. The Mandelbrot set (Figure 12) is irregular in this sense,
which may explain why this pattern has been widely reproduced beyond the mathematical
domain from which it came. Because of the property of self-similarity, it can be difficult to
judge the scale of fractals (in the absence of a visual cue) because they look similar whether
nearby or far away. For this reason, the image of fractal tree branches in Figure 3, which
was taken from a distance of around 12 feet, can also be perceived as a view of the earth
seen from a great height, like a satellite image of mountains and rivers. These two
alternative visual interpretations are possible because both are characteristically fractal in
form. Because fractals have a similar structure at all levels, they are fairly simple, and so a
description of them need only be fairly short. Indeed, the Mandelbrot set is constructed
from a surprisingly short algorithm: zn+1 = zn2 + C. The intricacy of a fractal image such as
this comes from the depth of iteration of the algorithm (see Figure 2, in which progressive
stages of iteration are illustrated). Fractals may thus have great depth of detail which in
theory is infinite, but in practice is limited by physical constraints but because fractals
have only one type of pattern at all levels, they are essentially quite simple. Unlike the many
different patterns in an oil painting described in the introduction, magnifying part of a
fractal reveals no new information, but the same pattern repeated at smaller scales.
Complexity
37
Chaos and Randomness

Chaos theory began in the 1960s when the meteorologist Edward Lorenz used
computers to model the weather (Gleick, 1993). When Lorenz resumed a weather
simulation from halfway, the digits he entered for the systems variables had unknowingly
been rounded off, and so the conditions were slightly different from those when it had
been stopped. This went unnoticed until the system began to diverge from the previous
run until it no longer resembled the earlier weather model. Lorenz understood its
significance for complex systems: small differences can dramatically alter the behaviour of a
chaotic system, so modelling such a system will always be uncertain. Lorenzs contribution
has become known as the butterfly effect. It demonstrates that systems such as the
weather are sensitive to initial conditions, and this is a characteristic of chaos.
The butterfly effect is a common expression for the sensitivity of chaotic systems, an
old idea which is embodied in the rhyme that begins, For the want of a nail, the shoe was
lost. The problem with this expression is that it only stresses one aspect of chaos: while it
is true that tiny events can have large consequences, there are many such small and large
events happening all the time. Most of these perturbations do have long-term effects on
the state of the system, but thankfully not every butterfly causes a tornado! What is most
important in the idea of the butterfly effect is the instability and unpredictability of the
system, for these are the characteristics of chaos. Instability is a property of dynamic
systems, and so this is not applicable to the formal properties of images, whereas
unpredictability can be applied to images by using the information theory concept of order.
Information theory treats this visual information as a streamed message in which the
elements are specified one after another. Regular patterns in the data stream are
predictable, which makes them redundant, and so they can be compressed (a shorter
description will suffice). Conversely, irregular patterns are unpredictable, so they have less
redundancy, and are less compressible.
Complexity
38
Figure 13 The Lorenz attractor (top, front and side views).
Lorenz modelled the evolution of simplified weather systems by reducing the

number of variables. The Lorenz attractor (Figure 13) represents the changing state of a
three-variable system by plotting its co-varying values in three dimensions. The shape
described by this path in phase space is fractal, and so it comes under the category of
strange attractors. What this means is that the system does not settle down to a steady state
(which would describe a single point in phase space) as in class 1 systems, or a periodic
state (a loop) as in class 2, but that it is constantly and unpredictably changing. A chaotic
system is not random, because it follows definite rules or laws (weather obeys the laws of
physics, for example), but the results may be indistinguishable from randomness. In a book
that brought this science to the wider public, James Gleick characterised chaos as order
masquerading as randomness (1993, p. 20). Upon reading Gleicks Chaos (1993), I tried
out for myself an example described in the book: I dropped a leaf into a stream, and
watched its meandering path downstream. Additional leaves dropped in the same place
followed a similar path to start with, but gradually diverged until they no longer resembled
the first. This is in effect a repetition of Lorenzs experiment, which demonstrates the
sensitivity to initial conditions of a chaotic system; immeasurably or imperceptibly small
differences in initial conditions get magnified by the system and lead to unpredictability.
Randomness is also characterised by unpredictability from one moment to the next,
but it is expected to be more predictable than chaos overall: A truly random string of
numbers contains equal proportions of each digit. In that sense, randomness is predictable
Complexity
39
and smooth, whereas chaos is lumpy. Randomness means that there is no pattern from
one element to the next, but it conforms to statistical regularity overall, just as a fair coin
toss converges to 50 percent heads over time. Similarly, the heights of a randomly selected
population can be expected to conform to a normal distribution (a bell-shaped curve). It is
sometimes thought that humans are unable to generate truly random number sequences,
but there is evidence to the contrary (Persaud, 2005). Because randomness by definition
has no regularity, a random arrangement contains a maximum of information content,
lacks redundancy, and cannot be compressed.
Complexity
The types of computer-generated pattern that Wolfram classified can also be found
in nature. After Wolfram published his results, he received many such examples sent to
him by the public, the most common of which was a particular type of sea snail shell. The
textile cone shell bears a striking resemblance to some of the CA patterns, with the same
kind of triangle structures seen in eCA rule number 22 (Figure 14). Wolfram (2002)
suggests that the shells colouration derives from an analogue of the CA rule in the
interaction of chemical pigments as the snail shell grows along its edge.
Figure 14 Textile cone shell (Textile conus) and rule 22 elementary cellular automata.
The example of the cone shell displays an irregular pattern of the class 3 type. Like
the description of randomness earlier, this pattern has local disorder (irregular arrangement
and size of elements) and global order (similar overall texture). Of the 256 elementary
cellular automata, Wolfram identifies only one of the 256 elementary CA as complex (class
Complexity
40
4). This rule (number 110) is illustrated in Figure 15. It is plain to see that rule 110 has a
mixture of class 2 regularities and class 3 randomness, which by Wolframs definition
makes it a class 4 pattern. By the information theory measure of entropy, the rule 110
pattern is mid way between order and randomness. The image also has the greatest number
of different types of pattern of all the elementary CAs, which makes it the most complex in
terms of the difficulty or size of its description. If we think about the types of twodimensional artwork in galleries, we can probably find examples to fit each of the four
classes of complexity, but I suggest that the majority could fairly be described as containing
elements of both order and chaos. This project aims to measure the complexity of a range
of artwork to see just how complex it is in relation to the classes of pattern we have
identified. Measuring visual complexity allows us to investigate its relation to the aesthetic
perception of art. The next section looks at some of these measures.
Figure 15 Elementary CA rule 110. This image shows a thousand cells (width)
evolving over a thousand time steps (downwards), a total of 1 million cells/pixels.
Complexity
41
Measuring Complexity
To be measurable, a thing must be quantifiable, but what exactly do we quantify
when it comes to complexity? Langton (1990) not only made a qualitative mapping of
complexity in cellular automata, but also measured complexity by quantifying the activity
of each rule (the lambda parameter). An alternative proposal is to measure complexity by
quantifying difficulty:
The meaning of this quantity should be very close to certain measures of difficulty
concerning the object or the system in question: the difficulty of constructing an
object, the difficulty of describing a system, the difficulty of reaching a goal, the
difficulty of performing a task, and so on. (Li 1991 p.381)
From this selection, we can see that measures based on the difficulty of making and
describing are applicable to art objects. Prominent examples of measures based on
difficulty-of-description include: information, entropy, algorithmic complexity, and
minimum description length (Flake 1990; Wolfram 2002). There is no single measure of
complexity accepted in all areas of complexity science, but amongst the most common is a
measure of the difficulty-of-description type developed independently by Solomonoff
(1964), Kolmogorov (1965) and Chaitin (1966). It is related to Shannons (1948) measure
of information, and is defined as the length of the shortest algorithm that generates a given
pattern of data (Flake, 1990). In the field of complex systems, the measure is called
Kolmogorov complexity, or algorithmic information complexity (AIC). A simple pattern
requires only a short algorithm to reproduce it, whereas a more complex pattern requires a
larger one, hence the algorithms size is a measure of the patterns complexity.
Murray Gell-Mann, co-founder of the Santa Fe Institute for research in complex
systems, describes AIC as a crude measure of complexity because it does not entirely
agree with intuitive notions of complexity (2004, p.34). At the top of the AIC scale are
structures that cannot be compressed, which have no regularities and which are therefore
random. The problem is that randomness is not the same as complexity, which is generally
Complexity
42
understood to involve a meaningful (non-random) relation of structural elements, as GellMann points out:
This property of AIC, which leads to its being called, on occasion, algorithmic
randomness, reveals the unsuitability of the quantity as a measure of complexity,
since the works of Shakespeare have a lower AIC than random gibberish of the same
length that would typically be typed by the proverbial roomful of monkeys. (GellMann, 1995, p.16)
Gell-Mann (2004) proposes a refinement of this crude measure, which he calls

effective complexity (EC). EC is almost the same as AIC, except that it is based on the
shortest description of the regularities in a system. With both measures, a uniform structure
is rated lowest since it has the least information and maximum redundancy. A random
structure has maximum information, but no regularities (patterns), and thus rates low in
terms of EC, instead of being rated highest by AIC. A complex structure with lots of
different patterns will require a larger description, and so it would be rated high in EC, in
accordance with our intuition. Figure 16 illustrates the relationship between the two
measures. The similarity to Langtons map of complexity (Figure 11) implies the similarity
of the lambda parameter scale to an information-based measure of complexity such as AIC.
EC
AIC
Ordered
Random
Low information content
High information content
High compressibility
Low compressibility
Figure 16 Effective complexity and algorithmic information complexity. Adapted

from Flake (1990) p.135 and Gell-Mann (2004) p.59.
Complexity
43
Technically, AIC is incomputable because we can never be certain that we have

found the shortest algorithm. Nevertheless, it can be closely approximated with a
complexity measure known as minimum description length (MDL), which provides a
practicable equivalent to AIC. The MDL measure can be implemented with the
information theory method of data compression, which reduces data to its smallest amount
by removing redundancy. Thus data compression gives us a measure of complexity based
on difficulty-of-description, which in this case is quantified by the length of the shortest
description.
The difference between information-based (AIC, lambda parameter) and intuitive
(EC, Wolframs classification) measures is significant to the current project because it
encapsulates the objective-subjective duality of complexity and highlights the potential
difficulties in measuring and understanding complexity in terms of aesthetic perception.
Nonetheless, there are suggestions of a similarity between human perception and
computational measures of aesthetic complexity: Wolfram (2002, p.559) says that
by far the most common way in which we determine levels of complexity is by
using our eyes and our powers of visual perception. So in practice what we most often
mean when we say that something seems complex is that the particular processes that
are involved in human visual perception have failed to extract a short description.
Measuring Visual Complexity

The early stages of this research project located the majority of publications on visual
complexity in the sciences of perception, rather than in aesthetics or the science of
complex systems. Amongst the many different methods of measuring visual complexity,
one stood out from the others since it seemed the most straightforward and compatible
with images of artworks. Don C. Donderi (2006a) finds a correlation between subjective
estimations of visual complexity and the size of compressed digital image files. Donderis
method began with research conducted for the Canadian Ministry of Defence, which
measured the difficulty of reading naval charts and radar images in order to determine the
optimal information density of visual displays. Donderi quantified the complexity of images
Complexity
44
based on their information content by measuring the size of compressed image files. The
subjective complexity of the visual displays was measured by magnitude estimation scaling
(MES), a psychological technique in which participants report a quantitative estimate of
their perceptions. The results show a strong correlation (0.78, p < 0.001) between
subjective complexity ratings and the logarithm of compressed file sizes. We now know
that data compression is a form of the MDL measure of complexity, which is equivalent to
AIC. Donderis finding supports the notion of an information-based measure of visual
complexity and suggests that it may work for images of art as well as the simpler graphic
displays in those experiments.
Donderis methods for measuring objective complexity by file compression and
subjective complexity by MES may be suitable for the current investigation. However, we
have seen that these information-based measures rate randomness the highest, and that
what we understand as complexity is actually somewhere in between order and
randomness. The first implication is that Donderis experiments examined only the lower
range of visual complexity, a point acknowledged by Donderi & McFadden (2005 p.78). It
also implies that the correlation between subjective and objective complexity may break
down for the most information-dense (random) images, in the same way that the
correspondence between AIC and EC diverges with increasing randomness (Figure 16).
Donderi (2006b, pp.8788) makes the link with Wolframs statement of the issue
(2002, p.559) in terms of a threshold of perception: We fail to perceive the inherent
patterns (which can be revealed by mathematical analysis) in some chaotic images such as
class 3 CA images, and so they appear random because they are beyond our threshold of
perception. The peak of the graph in Figure 16 represents this perceptual threshold,
beyond which our perception of complexity diverges from information-based measures.
The suggestion is that the correlation between objective complexity (AIC) and subjective
complexity (MES) will diverge for the most complex/random images. Empirical evidence
for this hypothesis of a perceptual threshold is yet to be seen. Thus we have identified a
Complexity
45
gap in knowledge of visual complexity which has the potential to contribute to

understanding its role in the aesthetics of visual art. Being able to determine the
performance of Donderis file-size measure over a wider range of complexity mapping
the perceptual correlation with informational complexity also provides a firm footing for
the investigation into the perception of complexity in art and design artefacts because it
maps the region of the spectrum of complexity that these artefacts are assumed to occupy.
Project Outline
The focus of this project is on the aesthetic properties of visual complexity in art and
design. The hypothesis is that visual complexity sustains interest and contributes to
aesthetic value by affording grounds for visual exploration (pattern-finding). The aim is to
understand the relationship between objective properties of images with the subjective
properties of perception. The objective is to find a method of measuring the visual
complexity of contemporary art and design in order to contribute to this understanding.
The philosophy of aesthetics helps us to understand the practice and perception of visual
art. With knowledge of complexity from systems theory and order from information
theory, we have a collection of tools that we can use to approach the subject of visual
complexity. The computational approach to visual complexity necessitates a quantitative
formulation of aesthetic properties which at first may seem alien to the more qualitative
studies undertaken in the tradition of philosophical aesthetics, but importantly it allows for
the measurement of aesthetic properties and empirical investigation.
In summary, the focus of the research is as follows:
Focal theory: Why is visual complexity interesting? Why are good artworks often
visually complex? What is the relationship between visual complexity and aesthetic value?
Contextual theory: What is visual complexity? How do we perceive it? Where is visual
complexity in art and design? Where is contemporary aesthetic discourse?
Project Outline
46
Data theory: How can we measure visual complexity? How can we measure aesthetic
perception? Do the measures work with the whole range of complexity, including images
from art and design?
Contribution: What is the relationship between objective and subjective visual
complexity? Does art training affect perception of and preference for complexity? How
does the complexity of an image relate to preference and quality?
Planned Contribution to Knowledge

This research project aims to make a contribution to knowledge initially by mapping
aesthetic discourse across the domains of art, philosophy and science. The literature review
presents a structured critique of theory and practice relating to the aesthetics and science of
visual complexity. Its aim is to collate a novel review of relevant material that is
unrestricted by a disciplinary focus, insofar as it goes beyond the range of material written
by and for art practitioners to include foundational texts and current papers from
experimental psychology, computer science and complex systems theory. The purpose of
this review is to establish a map of current knowledge in order to identify the gaps to which
this project can contribute and to identify the methods by which this knowledge can be
created.
The principal contribution to knowledge is made by extending empirical research
methods to the field of art and design. Techniques from experimental psychology and
computer science are applied to material from contemporary creative practices. The aim is
to add to knowledge by testing these methods with novel material and to extend the range
of material, specifically in terms of the range of visual complexity, to assess the
performance of visual complexity measures. One of the objectives is to investigate a wide
range of complexity to test the hypothesis that there is a perceptual threshold of
complexity, and to determine whether an information-based measure will support the
hypothesis, as predicted by Donderi (2006b). Examining a wide range of visual complexity
Project Outline
47
by using cellular automata images allows for a mapping that can be compared with the
findings of Wolfram and Langton. Using the method of MES with these images then
allows for the results to be correlated with the subjective aesthetic response. This provides
a connection between findings in complexity theory and empirical psychology, and
establishes a base for further investigations on visual material from contemporary visual art
practices.
In addition to the empirical investigation of visual complexity, this project draws on
art practice to explore novel methods of creating artwork. Creative practice plays a role in
exploring the aesthetics of visual complexity from the artists point of view, but also
provides a source of visual material for use as stimuli in the empirical tests. The tests offer
a novel method of evaluating this artwork, allowing for precise quantitative measurement
which is complemented by the qualitative data gathered in interviews with test participants.
Together, the creative and empirical work make for a balanced analysis of the aesthetics of
visual complexity by investigating both objective and subjective aspects of the topic, and by
considering both the production and the perception of visual art.
Project Outline
48
Chapter 2
Introduction
The aim of this chapter is to support the development of the research through an
exposition of the historical, theoretical and practical contributions to the field of aesthetics.
In keeping with the trans-disciplinary methodology of the project, this contextual review
covers the various manifestations of aesthetics as it appears in the domains of art,
philosophy and science. The material centres on the aesthetics of visual perception and
visual art, which together define the context of the study. Within this context, the critical
review examines ways of understanding, realizing, and analysing visual complexity.
Evaluating these practices provides a basis for selecting the appropriate research methods.
The aim is to map contemporary aesthetic discourse and situate the project within the field,
in order to define its contribution to knowledge and its engagement with current practice.
The structure of this review is designed to progress from the general foundation of
aesthetics and complexity discussed in the previous chapter to the more specific questions
concerning visual complexity which are at the core of this project. In trying to provide a
solid grounding for the research, this overview of contemporary aesthetic discourse takes
us from its philosophy, through scientific investigations, to artistic engagement. The
sections are arranged in order from the theoretical pursuit of the aesthetic, to the practical
analysis of visual complexity, and finally the creative application of aesthetics. This
structure aims to present the relevant issues in an order which will allow for a coherent
development of the thesis. Accordingly, it involves a narrowing of focus towards the issues
surrounding the aesthetic perception of visual complexity in two-dimensional art and
design.
Introduction
49
Aesthetic Theory
The previous chapter introduced some of the concepts central to aesthetic discourse.
It brought us up to the point where the main concern of the field was to identify the
essential features of art and the aesthetic that allow for philosophical definitions in terms of
necessary and sufficient conditions. For the purposes of this project, the division between
aesthetics then and now is marked by the influence of Wittgenstein. The new line of
aesthetic thought developed in what has been called analytic philosophy a generally
Anglo-American discourse in the Kantian tradition, contrasted with the continental
philosophy of central Europe. As described by Nol Carroll,
The purpose of the analytic philosophy of art is to explore the concepts that make
creating and thinking about art possible. Some of these concepts include: the very
concept of art itself, as well as the concepts of representation, expression, artistic
form, and aesthetics. (1995, p.5)
This thesis follows the analytical tradition of aesthetic philosophy. In part, this is a
pragmatic decision, because the continental philosophies are of less utility for this project.
More significantly, the continental attachment to post-modern scepticism of scientific
method is incompatible with the methodology adopted in this project. A more detailed
examination of this issue is presented in the methodological discussion in Chapter 3.
From the preliminary discussion in the previous chapter, we have identified the
central concerns of aesthetics as perceptions, properties and practices. Each of these
concerns has both objective and subjective aspects which apply in varying degrees to the
natural and/or artificial. Jerrold Levinson (2003) identifies three foci of aesthetics: art,
aesthetic properties and aesthetic experiences. Yet the aesthetic properties Levinson lists,
such as beauty, ugliness, and sublimity, are in fact indistinguishable from aesthetic
experiences; we certainly attribute them to objects, but that doesnt mean that they are
intrinsic properties of the objects. As such, these attributes may be described as subjective
aesthetic properties (in the sense that they are personal and variable) as plausibly as they
Aesthetic Theory
50
may be called objective properties. Therefore, the way in which we structure the field
depends on how we negotiate the objective-subjective duality of aesthetics. Instead of
getting embroiled in the finer points of this distinction, my stance is to accept the duality of
aesthetic experiences and to regard the aesthetic as the interaction between object and
subject. This position is represented in the working model of aesthetics, which identifies
the aesthetic with these interactions.
Perceptual Aesthetics
Philosophical aesthetics can be divided into the study of art and the study of
perception. The former pertains to questions about art objects: how we recognise them as
art, why we find them pleasing, and how they should be evaluated. The other side of
aesthetics concerns the phenomena that comprise aesthetic experiences, which include
natural forms and scenes as well as human-made things. A proponent of perceptual
aesthetics of art is Monroe Beardsley, who classifies an artwork as an arrangement of
conditions intended to be capable of affording an experience with marked aesthetic
character (1982, p.299). For Beardsley, aesthetic objects are a subset of perceptual objects,
some of whose qualities, at least, are open to direct sensory awareness (1982, p. 31).
A key notion in perceptual aesthetics is experience. In Beardsleys philosophy
aesthetic experiences have value when they fulfil at least one of three primary aesthetic
criteria: unity, intensity and complexity. Artworks are good if they afford these aesthetic
experiences. Richard Lind suggests that perceptual interest should be distinguished from
theoretical interest, because while the latter is an inquisitiveness about hidden organization,
the former is a curiosity about manifest organization. (1995, p.120). It is this kind of
manifest organisation of visual art and design that is under investigation in this project. The
hypothesis that visual complexity sustains interest is supported by Linds description of
aesthetic value in terms of perceptual interest:
Aesthetic Theory
51
Satisfying perceptual interest is a cognitive skill; infants spend hours perceptually

making sense of simple objects. The perceptual skills of adults are so well honed,
however, that we hardly ever notice we are habitually satisfying this goal-oriented
interest. Only when an unusual arrangement defies instantaneous clarification do we
find ourselves consciously interested in making sense of it, as when, for instance, we
are confronted for the first time with the non-objective complexity of a Jackson
Pollock. (Lind, 1995, p.120)
Lind claims that this formulation solves several of the traditional problems of
philosophical aesthetics. One of these is the problem of the disinterested attitude which,
since Kant, has often been maintained as an essential aspect of aesthetic perception
because it helps to mark the distinction from non-aesthetic interest. Lind dispenses with
disinterest as a necessary feature of aesthetic judgement, replacing it with an interest in the
act of perception itself. The fact that Lind describes this aesthetic interest as a cognitive
skill suggests that it is something that can be learned. We might assume that the teaching
of this kind of skill is to be found in art and design schools, if anywhere, but can it actually
be taught or is it something that is only implicitly developed in an artistic education? This
question is beyond the scope of the present investigation, although it is possible to explore
its implications: If aesthetic perception is a skill, and not merely a given sub-personal
mechanism, then regardless of how it is learned we might expect to find differences in
perception between people who have undergone an education at art school and those who
have not. The tests in this project investigate the difference in aesthetic perception between
two groups those who practice art and design and have received an art education, and
those who have not.
Environmental Aesthetics and Affordances
The way in which a skill is acquired depends on the learning environment, which
includes environmental conditions as well as the social and cultural context. Since
perception in humans predates the culture of art, the environment plays a large part in
shaping our perception through the evolution of perceptual mechanisms. Thus we are
perceptually sensitive to the environment in which we have evolved, and we can
Aesthetic Theory
52
understand more about aesthetic perception by looking at how we perceive the

environment. This approach can be called environmental aesthetics.
An example of this kind of understanding is The Ecological Approach to Visual Perception
(1986) by J.J. Gibson, which describes our relation to the perceptual environment in terms
of affordances. According to Gibson the purpose of perception is action, and since the
need for action varies from time to time, the affordances to be acted upon also vary.
Gibson pictured the entire visual world as a dynamic visual array. He described a perceiver
whose moving eyes passively pick up a stream of visual information from the array. The
invariants in the stream signify stable objects and the flux signifies movement or change,
either of the observer or the environment. Visual systems tend to be tuned to recognise or
respond to the perceptible properties of affordances. This explains why we tend to like
pictures of natural scenes with wide open spaces and clumped trees with a source of
running water; such environments afford food and shelter, with a good vantage point to
spot both predators and prey. Constables The Hay Wain (1821) falls into this category of
images at least it does so for Westerners, with its North-European landscape. Other
cultures have preferences for their own familiar terrain, but we all share the same basic
understanding of the affordances of the environment and have inherited the same
perceptual mechanisms. In addition, we all share a similar ability to perceive affordances,
not just in natural environments but also in designed artefacts.
Gibsons ecological theory (1971; 1978) of passive pick-up of visual information
challenges the conventional view that perception involves representation and
interpretation. As such, Gibsons view represents a minority amongst psychological
theories of perception (Bruce, Green and Georgeson 2003, p.80). As a perceptual theory,
Gibsons method stands outside the psychological consensus because it is non-empirical,
but it is also at odds with art discourse: During the 1970s, Ernst Gombrich developed an
argument with Gibson in Leonardo, the journal of art, science and technology. Gombrichs
contribution to psychological discourse on perception is presented in Art and Illusion (1977),
Aesthetic Theory
53
which draws on the conventional psychological theories that Gibson rejects. Nevertheless,
Gombrich described Gibsons Perception of the Visual World (1950) as exciting, and he
recognised that psychology has come alive to the immense complexity of the processes of
perception, and no one claims to understand them completely (1977, p.21).
Despite the arguments against Gibsons approach to aesthetics from art and
psychology, the concept of affordances retains currency in the discourse of design. In the
context of visual display design, Donderi (2006b, p.81) interprets Gibsons affordances as
visual information that allows the perceiver to act which are in the visual information
stream whether or not the observer acts on them. Graphic design and product design can
take advantage of our ability to perceive affordances by using visual cues which allow the
user to interact with a designed artefact. The idea of designing an artefact that can be used
in this way is embodied in one of Dieter Rams design principles: good design makes a
product understandable (Design Museum, 2007). The psychologist Don Norman
introduced Gibsons concept of affordances to the field of design in The Psychology of
Everyday Things (1988), later published as The Design of Everyday Things (1998). Norman says
that the concept has caught on, but not always with true understanding, and admits that
some of the blame lies with himself for failing to stress the point that affordances are those
that are perceived, and not just those properties that are inherent in a physical artefact. For
that reason, Norman suggests that perceived affordances might have been a more accurate
phrase to use, because
in design, we care much more about what the user perceives than what is actually
true. What the designer cares about is whether the user perceives that some action is
possible (or in the case of perceived non-affordances, not possible). (Norman 2009).
The application of these principles is particularly relevant to interface design in

human-computer interaction (HCI). Besides the distinction of real and perceived
affordances, Norman is also careful to distinguish between affordances and cultural
conventions. An example of a cultural convention in HCI is the scroll bar, which nowadays
Aesthetic Theory
54
is generally moved in one direction to make the text on screen scroll in the opposite
direction. Norman suggests that good design should follow cultural conventions, but what
is most important in usable products is that the desired controls should be perceived so
that the desired actions can be discovered. This understanding of affordances is also
applicable to interactive art, in which the audience has some control over the appearance or
behaviour of the artwork. Indeed, the manipulation of perceivable controls and their
actions may be a focal point for the interactive artist, who is freer to play with affordances
and the expectations of the audience than is the designer of marketable products. On the
other hand, the emerging field of critical design (the term coined by Dunne 1999) is less
constrained by market forces and company briefs and thus, like art, it is able to exploit the
perception of affordances in the creation of challenging artefacts and situations.
Aesthetic Theories of Art

Except for the kind of interactive art mentioned above, looking at art does not
usually require action of the sort Gibson described, and so the ecological theory of
affordances is not wholly applicable to the aesthetics of art. The aesthetic theory of art
focuses on a very particular aspect of the human environment, but what it shares with the
ecological theory is a grounding in perception. We discussed in the previous chapter the
division of aesthetics into the study of perception and the study of art. Aesthetic theories of
art are those that tie together these two different aspects and enable a definition of art in
terms of perception:
the aesthetic definition of art puts one in a position to say why art is valuable. Art is
valuable because it affords aesthetic experience. Thus if we can say why having
aesthetic experiences are valuable, then we are also on our way to saying why art is
valuable. The value of art will be derived from the value of having aesthetic
experiences. (Carroll, 1995, p.167)
We also discussed previously the move away from objective properties to subjective
properties of perception, and noted that the definition of art a traditional aspect of
Aesthetic Theory
55
aesthetics has also shifted in this direction. The following section examines this shift in
more detail.
The Open Concept of Art
Although Wittgenstein had considered aesthetic matters in some of his writings, his
influence on philosophical aesthetics is indirect since it stems from interpretations of his
more general philosophy. Of his writings on aesthetics, one notable remark is that The
subject is both very big and entirely misunderstood as I see it. (Wittgenstein 1966). The
idea that influenced the course of aesthetics comes from the publication in 1953 of
Philosophical Investigations (Wittgenstein 2001) on language and knowledge, and it is
summarised as follows:
The idea that in order to get clear about the meaning of a general term one had to find
the common element in all its applications has shackled philosophical investigation.
(Wittgenstein 1972, p.19)
Wittgenstein was referring to the philosophical practice of forming closed concepts

based on common elements or essences. Game is an open concept because there are no
essential features shared by all the things we know as games (football, chess, computer
games, hide-and-seek), and yet we are able to use the word without ambiguity. What games
share is a kind of family resemblance, claimed Wittgenstein. Only three years after
Wittgensteins Investigations, Morris Weitz published the influential paper The Role of Theory in
Aesthetics (1956). Weitz considered previous aesthetic theories of art as sharing the
philosophical practice of essentialist definition. The problem with an essentialist definition
of art is that the practices and products of art change, and so an aesthetic definition
becomes redundant as a gap appears between theory and practice. To Weitz, therefore, any
attempt to define art in terms of essential physical features of artworks is ultimately
doomed to fail. As he saw it, The problem is not what is art?, but what sort of concept
is art? (1956, p.30). Weitz arrived at the conclusion that a closed concept of art is
incompatible with the open practice of art; therefore art is an open concept.
Aesthetic Theory
56
Weitz used Wittgensteins open concept because art changes (but we still call it art),
and used the family resemblance argument because there are a number of overlapping
properties (essences) that are shared by some but not all works of art. Hence, there are
shared characteristics which are too sparse and mutable to be relied upon for the task of
traditional essentialist definitions of art, and yet we can still use the word art to refer to the
practices and products that possess these essences. Strengths and weaknesses in the family
resemblance argument can be illuminated with biological analogy: Much of the time,
resemblances reveal actual connections. The taxonomy of species at first relied on visual
identification, and only later conformed to relationships based on understanding of
hereditary theory and genetics. We can see that chimps are very much like us, and indeed
we share an almost identical genetic makeup. But resemblances may derive from family
inheritance or from convergent evolution 4, and we cannot always tell on the surface which
is the case. A true family relation is one of shared inheritance of the sort that could be
validated with genetic tests (such tests on fossilized and extant animals reveal our natural
history and the evolution of species). So, as in biology, it is possible to be misled by
appearances, and the family resemblance concept is not quite strong enough to get to the
actual aesthetic inheritance. This is the argument used by Maurice Mandelbaum (1965) to
criticise Weitzs employment of Wittgensteins family resemblance concept, noting its lack
of genetic connections.
The Decline of Perception in Aesthetic Theory
In the same way that Wittgenstein sought to end philosophy and failed heroically,
Weitzs argument to end the definition of art by a closed concept also failed. Weitz had a
valid point, though, about the ever-changing nature of art and the failure of aesthetic
theories to survive much beyond their contemporary artistic practices. Although the family
An example of convergent evolution is the similar structure of octopus and human eyes. This fact baffled
biologists for some time, since the two species are so far apart on the tree of life. The reason is that
adaptation to similar environmental pressures led to the development of similar systems, and so the family
resemblance of these eyes belies the actual evolutionary path of the two species.
Aesthetic Theory
57
resemblance argument has been challenged for its weakness, the point about an aesthetic
theory having to cope with the expansive variety of art is increasingly relevant. The
strongest theories since Weitz have accepted this point, and in doing so they have moved
the aesthetics of art further from the aesthetics of visual perception. For a brief period after
Weitzs essay, there was a hiatus on the definition of art, but this did not last for long:
had it not been for the neo-Wittgentseinians the temptation to define art might
never have appeared so inviting. For nothing taunts a philosopher so well as the claim
that something is impossible (Carroll, 2000, p.5).
The aesthetic definition of art resumed with Arthur Dantos concept of the artworld
(1964) and George Dickies institutional theory (1974). Their theories came about in a
situation where the distinction between ordinary objects and artworks could no longer be
taken for granted because of artists appropriation of everyday objects (the readymade)
and the incursion of art into sites beyond the gallery. Dantos essence of art is something
that the eye cannot descry an atmosphere of artistic theory, a knowledge of the history of
art: an artworld (1964, p.580). For Danto, the artworld is a world consisting of works of
art, a self-enriching community of ontologically complex objects (1993, p.204). This
context allows us to distinguish between two objects which are perceptually identical, of
which only one is a work of art. Reciprocally, the artworld also provides a context in which
the transfiguration of the commonplace (Danto 1981) can take place; everyday objects are
transformed into works of art by their incorporation into the institutional mechanism of
the artworld. Dantos canonical example is a mundane cardboard box used for containing
and transporting Brillo soap pads, and Warhols work of art Brillo Box (1964). The artworld
is both what makes this box a work of art and what enables us to recognise it as such for
the purposes of philosophical definition.
Weitzs open concept of art is evaluative, whereas Danto and Dickie offer
classificatory or descriptive definitions of art. Classificatory definitions allow us to answer
the question is this art? In a sense, this has been a non-question in art practice for around
Aesthetic Theory
58
a hundred years, since Duchamp made the point that the more interesting and more
difficult question is what is good art? Dantos theory avoids the difficult problem of
classifying contemporary art in terms of the perceptual properties of art products, and
instead provides a definition based on practice. As a defining essence of art, the concept
of the artworld is explicitly non-visual and thus non-aesthetic in the traditional sense of
visual perception. Similarly, Dickies aesthetic theory is not based on the sensory perception
of art, but relies on the identification of the social mechanisms of art. The difference is that
Danto emphasises the importance of art history and art theory for ontologically identifying
art, whereas Dickie emphasises the sociological aspects of the art world. Both are
concerned with the context of art, and both are non-perceptual.
The Institutional Theory of Art
Dickie opposes the idea of a special aesthetic attitude, as adopted by Beardsley. He
argues against Beardsleys three aesthetic conditions of value (unity, intensity and
complexity), and against essentialist approaches to the aesthetic, saying that aesthetic
perception differs only in motivation and not in perceptual qualities. In other words, Dickie
thinks that in the appreciation of art we do not use anything other than our usual powers
of perception. Like Danto, Dickies theory is non-perceptual; it is based on the same things
that the eye cannot descry namely, the artworld. The institutional theory has undergone
a few revisions, the latest of which Dickie calls the art circle because it is based on a
circular definition. Normally in philosophy, such definitions are avoided as being a vicious
circle in logic, but Dickie defends the theory as a virtuous circle. The art circle (1997)
comprises these five statements:
1. An artist is a person who participates with understanding in the making of a work of art.
2. A work of art is an artefact of a kind to be presented to an artworld public.
3. A public is a set of persons the members of which are prepared in some degree to
understand an object which is presented to them.
Aesthetic Theory
59
4. The artworld is the totality of all artworld systems.

5. An artworld system is a framework for the presentation of a work of art by an artist to
an artworld public.
The five statements of the art circle flesh out the institutional theory of art by
identifying the operative elements and their relationships. Substituting Dantos somewhat
vague notion of the artworld for a precise definition strengthens the theory overall by
providing a more usable framework for classifying art. Despite its greater precision, the art
circle still has the flexibility to accommodate new art practices and products. Note that
although Dickies theory avoids the problematic essentialist definition of artwork
properties, it nevertheless relies on the identification of the essential ontological elements
that comprise the artworld: artist, artwork and audience. It is of course possible that these
elements might change in the future artworld, just as the properties of artworks have
changed, in which case the theory may need to be revised.
Visual Models of Art
We can visualise the circular relation of Dickies theory as follows (Figure 17):
Artwork
Artworld public
Artist
3
5
Artworld
Public
Figure 17 Visual representation of Dickies (1997) art circle.

Aesthetic Theory
60
This diagram illustrates the definitional connections of Dickies five statements (in
numbered circles) that link the elements of the artworld (in shaded boxes). By slightly rearranging the visualisation of the art circle, we can see that it is structurally similar to the
working model of the aesthetic (Figure 18). The two models share a common set of
ontological categories artist, artwork, audience or artworld public, and artworld. From
these visualizations, we can see that Dickies statement 3 leads nowhere. It fails to
adequately define public in general terms, and equally fails to separate fully from the
ontological category artworld public. For this reason, it seems superfluous to the theory
and it weakens the definitional relation of persons in the artworld with Dantos concept of
the artworld. If we were to modify the theory to fit our working model, we would replace
the compound term artworld public with the singular audience, which means that we
could dispense with the superfluous statement 3. Statement 4 would then need to be
reformulated along the lines of the audience is the set of persons who, by engaging with
artwork (i.e., appreciating, presenting, trading, criticising), populate the artworld, and
statement 5 as the artworld is constituted by the practices and products of those who make
and engage with artworks. The core theoretical structure would thus remain in place apart
from the excision of the third statement.
Artist
Artwork
Artworld
public
Public
Artworld
Figure 18 Visual representation of Dickies art circle mapped onto the working model
of aesthetics.
Aesthetic Theory
61
It seems that Dickie never thought to visualize the art circle, since the texts in which
the theory is published lack any such diagrams. In contrast, Arnold Berleant did illustrate
his concept of the aesthetic field, which is related to the institutional theory of art. The
central thesis of Berleant (1970) is that all factors of the aesthetic should be taken into
consideration in regard to understanding the appreciation of art, thus aesthetics should be
considered as a complex field of interacting contextual and aesthetic properties. This is
certainly a significant point and an attractive idea, but the diagram itself (Figure 19) is
difficult to read, and adds little in terms of explanation to Berleants thesis. Each element is
connected to every other, and the arrows connect in both directions at once, which means
that it actually says very little at all about the relationships between the elements except to
say that they all influence each other at all times. In contrast to Dickies institutional theory,
Berleants model employs four ontological categories: artist, artwork, audience and
performer. The last of these is not present in Dickie (1997), whereas Berleant considers it
an essential aspect of the aesthetics of classical music, for example, in which the composer
occupies the role of artist and the musicians are the performers. The role of performer is of
less relevance to the field of visual art on which the institutional theory is based, since it is
not a defining element common to its practice. For the purposes of the current project,
therefore, the role of the performer can be justifiably disregarded, as can Berleants model.
Aesthetic Theory
62
Figure 19 Berleants (1971) diagram of the aesthetic field.
In examining the practice of computer art, Stroud Cornock and Ernest Edmonds
(1973) visualize a series of art practices. They develop a theory of different art systems
using a series of images to describe the relationship between artwork and audience (Figure
20). The art systems represented are dynamic in varying degrees, from the traditional
passive viewing of an artwork to a fully interactive relationship. These diagrams do not
include the role of the artist, because the purpose of the study concerns a particular form
of art practice and does not aim to offer a universal theory of art. Instead, the diagrams
illustrate what goes on in the everyday experience of an artwork, where the artist is not
usually present and is not a necessary part of the work. It would be possible to incorporate
these models into the working model of aesthetics, which would place the dynamic art
systems in the larger context of the artworld and refine the working model to a more
specific kind of practice. But because the focus of the current project excludes dynamic and
interactive art, however, Cornock and Edmonds model is not deployed in this thesis.
Aesthetic Theory
63
Figure 20 Cornock & Edmonds (1973). Visualization of art systems: (a) static; (b)
dynamic-passive; (c) dynamic-interactive; (d) dynamic-interactive (varying); (e) matrix.
Key: A = artwork, S = spectator, E = environment, T = time, P = participant, M =
modifier.
Relational Aesthetics
By avoiding the problems of defining art based on essential properties, Dickie claims
that the art circle offers a workable theory that allows room for changing art practices:
The institutional theory allows the freedom for which Weitz quite correctly is so anxious
to preserve in his attack on traditional theories of art. (Dickie, 1997, p.110). In so doing,
Dickies theory distances the aesthetic from the perceptual, as does Dantos concept of the
artworld. A more recent contribution to this development comes from Nicolas Bourriauds
Relational Aesthetics, which is an aesthetic theory consisting in judging artworks on the basis
of the inter-human relations which they represent, produce or prompt (2002, p.112). Like
Danto and Dickie, Bourriaud attempts to grapple with the incursion of the everyday into
the world of art, the difference being that Bourriauds theory mainly addresses art events
rather than art objects. The theory focuses on the non-perceptual social relations that are
engendered in art practices of the 1990s (notably, of those practices represented in
Bourriauds own curatorial projects). The examples of relational art include the work of
Rirkrit Tiravanija, Felix Gonzales-Torres, Vanessa Beecroft, and Liam Gillick. Because
some of this artwork would be indistinguishable from everyday happenings were it not for
Aesthetic Theory
64
the context of the artworld, they represent the type of practice for which the institutional
theory was developed. According to Bourriaud, what makes these practices relational is
that they take as their theoretical horizon the realm of human interactions and its social
context, rather than the assertion of an independent and private symbolic space of, for
example, Modernist art (2000, p.14). This relational art, Bourriaud says, takes Duchamps
notion of the audience as contributor to the artwork a step further, by postulating
dialogue as the actual origin of the image-making process (2002, p. 26).
We can use the working model of aesthetics to understand Bourriauds concept. This
can be illustrated with a personal story of when I met the artist Xavier Roux at an
exhibition in Washington DC in 2007. Xaviers artwork involved sitting with him on a
portable wheelbarrow/chair that he had made, and engaging in discussion with the artist.
(Part of the work required those who engaged with the artist to report back if dreams about
the wheelbarrow were experienced later on.) In our discussion I explained my research,
drawing the working model of the aesthetic (Figure 6). He re-drew the diagram to illustrate
his own practice he omitted the artwork from the diagram, joining the artist and audience
with a single line. For Xavier, he explained, this direct interaction is the artwork, and is the
kind of practice described in Bourriauds relational aesthetics.
Bourriaud defines the practice of art as an activity consisting in producing
relationships with the world with the help of signs, forms, actions and objects (2002,
p.107). Whilst this definition is certainly inclusive in the way that Weitz (1956) advocated, it
is difficult to see how it is useful in separating what artists do from any other field of
human enquiry or creativity (admittedly, Bourriaud does not rely on this definition to
support his theory). A more serious criticism of Bourriaud comes from Claire Bishop
(2004), who questions what kind of interaction is actually desirable in relational aesthetics:
Is the meaning (and therefore also value) of relational art actually developed collectively,
as Bourriaud suggests, and if so, why is it that the artist still appears to be in control and
remains the one person to whom these artworks are attributed? Looking back at Dickies
Aesthetic Theory
65
art circle, an artist is someone who participates with understanding in the making of a work
of art (Dickie 1997, my emphasis). Only with understanding comes the right to
authorship. The audience-participants in Bourriauds examples of relational art cannot be
said to have full understanding. In fact the opposite seems to be the case with the work of
Tiravanija and Gillick, in which the audience has no instruction for participation and the
rules of the game are unclear if not actually obscured or withheld. Bishop also notes that
the quality of the relationships in relational aesthetics are never examined or called into
question. (2004, p.65). It is a reasonable point, and it seems the same could be said for the
similarly non-perceptual theories of Danto and Dickie: Both relational aesthetics and the
institutional theory of art lack the explanatory power of the earlier closed-concept
definitions of art such as Beardsleys. These non-perceptual aesthetic theories of art are
unable to explain why we value some artworks over others or why we should do so. The
current situation is that in the context of non-perceptual aesthetics, political, moral, and
ethical judgments have come to fill the vacuum of aesthetic judgment (Bishop, 2004,
p.77). My thesis is that an empirical investigation of aesthetics has the opportunity to fill
some of these gaps in the non-perceptual theories of art. But whether a return to a
perceptual theory of art also has the potential to account for the aesthetic judgement of
everyday artworks and relational art practices remains to be seen.
Summary
Historically, aesthetic discourse has served to theorize the practices of making and
appreciating art, whether it is concerned with perception in general or with art in particular.
Philosophy has provided the ontological and epistemological foundations for an
investigation of aesthetic questions. Contemporary aesthetic theories have to contend with
an artworld which is more diverse than ever before, simultaneously made more interconnected and more fragmented in the process of globalization. The most successful
aesthetic theories are those that survive as art history accumulates and the density of
available information increases. Amongst the most successful today include the theories
Aesthetic Theory
66
that incorporated Wittgensteins ideas on definition, in which meaning is defined by use.

This idea has shaped aesthetic theory by changing its focus from the products to the
practices of art, and in so doing the field has moved away from perception.
Contemporary aesthetics is dominated by the a-perceptual theories of Weitz, Danto,
Dickie and now Bourriaud. Most of these theories are essentially answers to the question
what is art?, but we have noted that this is an old question in art, and that the more
interesting question is what makes good artworks valued in the way that they are.
Bourriauds theory claims to provide such a criteria, but it only applies to a specific subset
of contemporary art practice, and even in that regard it is not without problems. In an
article entitled Is Psychology relevant to the Arts? (1961), Dickie comes to the firm conclusion
that it is not: I am convinced that the problem of the description of the nature of
aesthetic experience is not a task to which the techniques of empirical science are
relevant (p.302). On the contrary, I believe that empirical investigation can fill the gaps in
non-perceptual art theory by explaining why we value art and what makes it good. Even as
contemporary art practice and theory have moved away from the traditional perceptual
core of aesthetics, we still perceive and understand art through our senses. For this reason,
I suggest that we should look again at the perceptual aspects of visual art which are
neglected in contemporary aesthetic theory.
My argument is that visual complexity is a contributing factor in aesthetic value, and
that a theory of aesthetic value has the opportunity to fill the gaps in the non-perceptual
theories of art. Visual complexity is just one of these perceptual aspects, but it is a valuable
one because it encompasses other, more fundamental perceptual properties, such as the
quantity, variety and order of picture elements. If we can understand how visual complexity
is manifested objectively and perceived subjectively, then we may be able to learn how it
contributes to the aesthetic perception of visual art. The next section of this chapter
presents the findings of empirical aesthetics that explore these questions in terms of the
measurable objective and subjective properties of visual complexity.
Aesthetic Theory
67
Empirical Aesthetics
This section provides a history and evaluation of developments in the scientific study
of aesthetic perception. It is presented in approximately chronological order, beginning
with an introduction to the issue of psychophysical measurement. The review narrows
from the general to the more specific concerns of this project from the foundation of
empirical aesthetics as an academic discipline to the development of experimental methods
within the field. Because the focus is on visual perception, this review is largely restricted to
aesthetics of visual art. The discussion incorporates the application of concepts from
complex systems and information theory which were introduced in Chapter 1, and puts
these in relation to empirical knowledge of visual perception.
Establishment of the Field

One aspect that distinguishes empirical aesthetics from philosophical and artistic
aesthetics is a willingness to adopt a quantitative approach to the subject. Whether aesthetic
objects and aesthetic perceptions can be justly quantified without losing significant
properties has remained in debate since the beginnings of empirical aesthetics. Dickie
(1961), for example, rejects psychological findings as irrelevant to aesthetic questions, and
would not admit of their value to art theory. We negotiate some of these difficulties in this
chapter and the next. Primarily, these problems arise due to the subjectivity of perception.
The subjectivity of sensory perception causes two problems for empirical studies:
Firstly, perceptions are not directly accessible to inspection for experimental purposes, and
secondly they are highly variable. If we can accept a general abstraction of aesthetic
perceptions, then the latter problem can be resolved through duplication of tests and
averaging of results. In this way we can approximate the trends of an individual or a group.
But we are still left with the former difficulty, namely, how to measure these subjective
perceptions in the first place. What is required is some way of quantifying the psychological
events of perception. Perceptions generally vary along a continuum, e.g. from dark to
bright or cold to hot, and the challenge is to partition the continua in such a way that they
68
can be measured. If they can be measured, then their relationship to the intensity of
physical stimuli can be determined. The field that concentrates on the relationship of
physical properties to perceptual properties is known as psychophysics. Thus the primary
task of early psychophysical measurement was the construction of a perceptual scale, which
is where the following section begins the review of empirical aesthetics.
The Measurement of Perceptions
One of the oldest psychophysical procedures is called equisection the partition of
continua into equal-appearing divisions. If a scale is divided in two, the method is called
bisection. This method was invented by Joseph Plateau (18011883), when he gave eight of
his artistic friends black and white oil paints and asked them to paint a mid grey between
the two (Stevens, 1975, p.154). Even though the artists carried out this task rather
unscientifically in different locations and lighting conditions, the greys they produced were
very similar when viewed in the same setting (Laming & Laming, 1996). From this simple
beginning, it is then only a small step to repeat such tasks and to build a scale by continued
subdivision.
Just how far can we take this process of subdivision? This is what concerned Ernst
Heinrich Weber (17951878), who formulated the idea of the just-noticeable difference
(JND). Webers method was to present a subject with a stimulus and to adjust it slightly
until the subject reported that they could perceive a difference. For example, in one
experiment a blindfolded subject holds a weight and reports when they notice a difference
as weights are added. The results showed that the JND grows in proportion to the
stimulus, i.e. if the weight being held is doubled, so does the amount that needs to be
added to achieve a noticeable difference. JND values are given with reference to the
percentage of trials in which the subject detects a difference usually 50%, although other
values are sometimes used. Through these experiments, Weber had created a law that
related physical stimuli to psychological perceptions. He had found that the JND is a
constant proportion of the strength of the stimulus, which can be written as:
69

=k
where (phi) is the stimulus intensity, is the JND ( means difference), and k is a
constant that signifies the proportion of the JND to the stimulus. N.B. Even though it
relates to perceptions, the JND is measured in the physical units of the stimulus, i.e. how
much of the stimulus is perceived.
The JND amounts to a perceptual threshold; it represents the discriminating ability
of our senses. Although Webers formula describes a relative threshold, the same idea can
be applied to the extremes of sensory continua in order to determine an absolute
perceptual threshold (e.g. how quiet a sound we can hear), which provides a useful upper
or lower point of reference for a perceptual scale. In psychophysics, there are three
common methods used to find perceptual thresholds: the method of limits, the method of
constant stimuli, and the method of adjustment.
The method of limits was invented by Wilhelm Wundt (18321920). It involves the
steady increase or decrease of stimuli until they are reported as being perceived or not.
Different results are found depending on whether the ascending or descending method of
limits is used, so some experimenters use both methods and then average the results. The
biophysicist Georg von Bekesy (18991972) offered an alternative called the staircase
method: The stimulus is reduced until it passes below the threshold, then increased until it
is perceived again, and so on until a steady value is found. In order to reduce the errors of
habituation and expectation, the method of constant stimuli presents a random succession
of various intensities, and the subject reports whether they are perceived. In the method of
adjustment, the subject controls the stimulus level, and their task is either to match another
reference level or to adjust it to a given threshold.
Weber is often credited as the father of psychophysics for his foundational work on
the JND, but others claim the title for Gustav Theodor Fechner (18011887) because he
was the first to bring a quantitative approach to the problem and to measure perceived
70
sensations. After he became aware of Webers work, Fechner made a significant

contribution to the establishment of psychophysics as a science via the publication of
Elemente der Psychophysik in 1860 (Fechner 1966). What we currently recognise as
psychophysics forms only one of two domains prescribed by Fechner; the one that he
called outer psychophysics, which relates physical events to the mental. Inner
psychophysics would deal with the relation between mental and neural events. Although it
does not exist as yet, a field of inner psychophysics may be realised as the relatively young
disciplines of neuroscience and consciousness studies continue to develop.
Fechner formulated the first psychophysical function a quantitative relation of
stimulus intensity to perceived sensation which had a great influence on psychophysics
(Stevens, 1975, p.7). He reasoned that if we could create equal units of sensation, then by
counting the number of units between two stimuli levels as in those found by equisection
we would be able to measure sensations. To Fechner, Webers JND provided such a unit.
He named his psychophysical function in honour of Weber, but it is now known as the
Weber-Fechner law:
= k log
represents sensation, k a constant, and the stimulus intensity. N.B., the Greek letter
(psi) denotes psychological measurements, while the letter (phi) denotes physical values.
In formulating this function, Fechner had made the assumption that the JND
formed perceptually equal divisions of a sensory scale: He assumed that as the physical size of
the JND () increases with stimulus intensity (), as Weber had found, its perceptual size
() does not. This meant that his psychophysical equation described a logarithmic
relationship between stimulus and sensation. In other words, as one value (the stimulus
intensity) is multiplied, the other (perceived sensation) is only added. It turns out that
Fechners logarithmic equation only holds true for a few perceptions, because his
assumption about the JND was wrong. Despite suggestions for an alternative to the
71
logarithmic function in 1728 by Gabriel Cramer (Stevens, 1975, p.3) and by Plateau in 1872
(Laming & Laming, 1996), the Weber-Fechner law continued to be used for the next
hundred years (Wixted, 2002).
Stanley Smith Stevens (19061973) was the first to successfully challenge Fechners
law, and had a large impact on experimental psychology. Whereas Fechners measure was
based on the units of the JND, which were defined by the experimenter, Stevens approach
was to ask the subject to quantify their own sensations. Instead of trying to construct a
scale, he reasoned, let us use the one that the subjects bring with them the concept of
numbers that we use so frequently in our daily transactions. In the first applications of this
method during the 1950s, the experimenter provided a value for the first stimulus, and
subjects were asked to apply successive numbers in relation to this value. Stevens soon
realised that this was unnecessary. Subsequent tests abandoned the fixing of the first value,
leaving it to the discretion of each subject, and later equalised their various scales to a
common modulus. When Stevens method called magnitude estimation scaling (MES)
was applied to psychophysical experiments, the results did not accord with Fechners
function (Stevens, 1957). In the influential paper To Honour Fechner and Repeal his Law,
Stevens (1961) argued that a power function, not a log function, describes the operating
characteristic of a sensory system. (p.80). Stevens power law, as it is known, is written:
= k
Again, is the perceived sensation, the stimulus intensity, and as Stevens said: The
constant k depends on the unit of measurement and is not very interesting; but the value of
exponent serves as a kind of signature that may differ from one sensory continuum to
another. (1975, p.13).
When the results of magnitude estimation are plotted, the graphs show logarithmic
curves like those in Figure 21a. If the function is instead plotted on logarithmic axes
(Figure 21b) the results produce straight lines, the slope of which can be used to determine
72
the value of the exponent , by dividing the y value by the x value. For example, the slope
of the dashed line in Figure 21b can be calculated by reading its end point, which has x and
y values of 6 and 2 respectively (if we count the notches on the axes), and so the value of
is 2/6 0.3.
a) Linear
b) Logarithmic
=3
=1
=3
= 0.3
=1
= 0.3
Figure 21: Linear and logarithmic plots of idealised psychophysical functions: how
perceived sensation () varies with stimulus intensity () for different values.
Exponents with values less than 1 indicate sensory continua such as loudness ( =
0.67) for which sensation increases more slowly than the stimulus, which enables us to hear
a very wide range of volumes. Conversely, continua that increase more rapidly than the
stimulus value have an exponent greater than 1, which is the case for electric shock ( =
3.5). For continua with an exponent of around 1, perceptions are proportional to the
stimulus, as in estimations of coldness and visual length (Stevens, 1975, p.15). The
differences in sensitivity to stimulus intensity, which are characterised by Stevens power
law exponent , often reflect evolutionary adaptations. It is not hard to see why being more
or less sensitive to particular situations is useful, for example, the survival value of visual
accuracy in estimating distance. The results of Stevens method of magnitude estimation
demonstrated that many sensory continua are better described by a power law relation than
by Fechners logarithmic function. Stevens power law is an example of the occasionally
73
surprising effectiveness of empiricism and mathematics to describe patterns in our world, a

matter of which he was well aware:
Regardless of how successful quantification may prove to be in all the rest of science,
in psychophysics it has yet to shake off all suspicion, for there lingers in many of us a
feeling, not only that human experience is somehow inscrutable, but also that
measurement, because of some flinty rigor, may lacerate the human spirit if we probe
too deeply with the aid of scales and numbers. (Stevens, 1975 p.51)
Stevens method demonstrates that it is possible to measure what was previously

considered un-measurable, and the results of this method fit the power law relation for
which he argued. Although these are valuable contributions to the field of psychophysics,
there have been a few issues with his work. One of these concerns the way in which
Stevens treats the measures derived from his experiments as ratio scales. Some have
objected to the assumption of a ratio scale, and say that such methods first require
empirical proof that perceptions actually behave this way (e.g. Narens, 1996). Such
arguments are matters of representational measurement theory, and are too complicated to
describe here (see chapter 1 of Wixted, 2002 for a concise overview). Despite these issues,
Stevens ideas about the different types of scales used in measurement remain amongst his
most influential contributions to psychology (Teghtsoonian, 2001).
To complete this summary of the history of psychophysical measurement, let us look
briefly at a less significant but interesting contribution which relates to the properties of
various sensory modes: Stevens (1975) makes a distinction between two types of perceptual
continua, which are named prothetic (quantitative) and metathetic (qualitative). In hearing,
loudness forms a prothetic continuum because it varies as a magnitude or scalar quantity.
On the other hand, the pitch of a sound is metathetic because it depends not only on its
absolute pitch but also on its relation to other pitches. Another way to clarify the difference
between the two is to imagine what happens when stimuli are combined. Adding two
74
sounds that differ in volume and pitch will result in an increase of loudness 5 (which is a
prothetic quantity), but not an increase in pitch (which is a metathetic quality). This
suggests that it may be more accurate to think of continua as varying in dimensions, or
degrees of freedom, instead of a cruder categorisation by quality and quantity. Loudness is
quantitative because it has only one degree of freedom, as does the linear scale of real
numbers. Pitch has at least two degrees of freedom; its value is described not only by its
absolute frequency (middle C = 261.63 Hz) but also by the frequency of the note relative
to others (C is a semitone above B and a whole tone below D), which is itself an intricate
affair. Taste has at least five dimensions (salt, sweet, sour, bitter, and umami) and is further
complicated by its close relation to smell, which has many more.
These sorts of complex relationships characterise the objects of study in empirical
aesthetics. The way in which we categorise the elements of such relationships influences
the type of questions we can ask and the type of answers we may find. The distinction
between metathetic and prothetic percepts is not widely used in the psychological literature,
yet these kinds of characteristics are clearly of great concern to those who deal with
aesthetic matters. Whether the aesthetic perception of visual complexity is best described as
a quantitative or a qualitative continuum is not yet clear.
The Measurement of Perceived Complexity
Fechner was among the first to consider complexity in terms of aesthetic perception.
In the Vorschule der Aesthetik (1876), he proposed the principle of the aesthetic middle,
which stated that too much or too little complexity detracts from aesthetic appreciation.
This proposition was one of sixteen aesthetic principles published in the Vorschule which
were empirically unverified at the time, but which were later supported by Berlynes (1971)
findings. In the landmark publication Principles of Physiological Psychology (1874), Wundt
considered the effect of novelty on aesthetic experience, and observed that the pleasantness
5
Loudness is the name given to the subjectively perceived intensity of sound; volume is the name of the
corresponding objectively measured intensity: If the volume of sound is too high, it may be too loud for us.
75
of perceptions varies with stimulus intensity in terms of an aesthetic middle. Daniel Berlyne
(19221976) describes the principle in terms of a measurable aesthetic dimension:
If, as one proceeds along a certain dimension, beauty or pleasure or aesthetic value
rises for a while, reaches a maximum in some intermediate state, and then falls as the
opposite extreme from the starting point is approached, we have an example of the
inverted U-shaped curves that are quite often encountered in psychology. (Berlyne,
1971 p.124)
+
HEDONIC VALUE
(PLEASANTNESS)
STIMULUS
INTENSITY
Figure 22 The Wundt curve (adapted from Berlyne, 1971 p.89).
Wundt was the first to provide empirical evidence for this psychophysical
relationship, and the resulting inverted U-shaped curve is named after him (Figure 22). We
can understand aesthetic complexity in these terms, too. If we can quantify and manipulate
the complexity of stimulus images, we may be able to carry out an empirical investigation
to test the hypothesis of the aesthetic middle for visual complexity.
Measuring subjective perceptions of aesthetic properties can tell us much about
individual perceptions and individual works of art, but it is difficult to compare and infer
from these studies due to the variability and specificity of their results. An objective
measure, on the other hand, can be used to construct a scale to which subjective data can
be compared. If we can determine the (mathematical) relationship between the two scales,
we could use the objective method to predict perceptions of other measured artworks. This
was among the methods adopted from experimental psychology that Berlyne (1971)
76
applied to aesthetic questions. Berlynes results led to his theory of aesthetic preference in
terms of arousal, which can be regarded as a measure of alertness or excitation (1971, p.64).
His theory, which proposed an inverted U function of preference for arousal values of
aesthetic variables, dominated the field of experimental aesthetics for the past several
decades (Martindale, Moore & Borkum, 1990, p.53). Seven experiments by Martindale and
colleagues (1990) provided enough empirical evidence to disprove this theory, they claim.
Their methods, however, were the same that Berlyne had pioneered and which continue to
be used (the authors say that the experiments were ones that Berlyne was likely to have
carried out himself, had he not died suddenly).
To quantify the intensity of aesthetic stimuli, George D. Birkhoff (18841944)
proposed an Aesthetic Measure in a book of the same name (1932). Birkhoff studied the
forms of 2D abstract shapes and silhouettes of objects, such as vases, and he came up with
a formula which attempted to capture the unity in diversity of beauty (Barrow, 2003, p.2).
His method was to measure aesthetic value in relation to geometrical properties. The
aesthetic measure (M) varied in proportion to the amount of order (O), and inversely in
relation to complexity (C):
M=
O
C
Birkhoff defined complexity as the number of straight lines that make up a shape,
while order is derived from a complicated summation of symmetries (Barrow, 2003, p.3).
With this equation, simple figures such as a square have a high M because they have more
order than complexity. One problem with this aesthetic measure is that it is impractical to
apply to most artworks, whose complexity and ambiguity make such measurement virtually
impossible. Another problem is that it does not account for the complexity of art, nor why
we tend to value complex images. Contrary to Birkhoffs results, psychological experiments
with geometric shapes by H.J. Eysenck (19161997) revealed preferences for high order
and high complexity. Examples of the stimuli are illustrated in Figure 23. The results of
77
Eysencks experiments (1941a; 1941b) prompted him to re-formulate the aesthetic measure
(1971, p.164) as follows:
M=OC
When put this way, M increases with complexity, as opposed to Birkhoffs formula in
which M decreases as complexity increases. Yet either theory seems counter-intuitive; it
seems more likely that the highest preference would be found between the extremes of
complexity, as Fechners principle of the aesthetic middle and the Wundt curve would
suggest. Birkhoff and Eysenck had empirical support for their respective theories, however,
and Fechner had not. The difference in both the stimuli and the measure of complexity
used by each of the experimenters confounds a comparison of their results. But what we
can say is that Birkhoff merely assumed that complexity was a necessary element of an
aesthetic measure, whereas Eysenck had demonstrated that complexity was a factor in the
determination of aesthetic preference. The implication is that Eysencks formula was
supported by his experimental results because he only looked at a section at the low end of
the complexity spectrum, for which our preference does appear to increase with
complexity: It is in effect just the left-most section of the Wundt curve (Figure 22), if we
replace stimulus intensity with complexity. It suggests that if Eysenck had examined a
wider range of complexity that is, had he included much more complex stimuli then it
is possible that the results would have fitted the Wundt curve, and so the rather simplistic
equation M = OC would have had to be revised.
Eysencks (1941b) experiments with geometric figures (Figure 23) led to the proposal
of what he called T and K factors of aesthetic preference (1941b). The T factor describes a
general tendency, similar to Birkhoffs aesthetic measure M, which Eysenck conceived as
an analogue of (psychologist and statistician) Charles Spearmans g factor, which is an
established measure of general intelligence (McWhinnie, 1965, p.35). Within the general T
factor, i.e. when its effects have been eliminated from the results, another factor is revealed.
The K factor describes a bipolar division of preference for simple polygons, simple
78
rhythms, and highly unified pictures as compared with a preference for complex polygons,
loose rhythms, and diversified pictures. (McWhinnie, 1971, p.116). Eysencks theory was
based on the psychology of personality types, and although it did not determine what
caused these preferences, a study by Harold McWhinnie (1966) suggests that the K factor
might be attributed to learning as well as personality.
Figure 23 The twelve most-liked (left) and least-liked (right) of Eysencks (1941b)
experiments.
Eysenck (1941c) also performed tests on colour preference using ten cards based on
Ostwalds (1969) colour system. As with his earlier experiments, he found a general
aesthetic factor (the K factor, in which subjects tended to agree on colour preference
ranking) and a second bipolar (T) factor in this case a preference for either saturated or
unsaturated colours. In Eysencks experiments, single groups were tested which were then
categorised according to the results. A few years later, Frank Barron took the opposite
approach by starting with different groups of people and comparing their preferences.
Barrons (1952) experiments investigated aesthetic preference between artists and nonartists. His method was to use images from the Welsh Figure Preference Test 6 (Figure 24),
testing preferences for either simple-symmetrical or complex-asymmetrical figures.
A non-verbal personality test comprising a set of 400 black and white images and a response choice of
like/dislike.
79
Figure 24 Example images from the Welsh Figure Preference Test
Barron concluded that artists had a higher preference for the complex-asymmetrical
figures. A later experiment using reproductions of artworks (Barron. 1963) found that the
artist groups also preferred art that was modern, radical, experimental, primitive, and
sensual (McWhinnie, 1971, p.116). Barrons tests failed to differentiate between the effects
of training and personality, since the artists may have had a predisposition for liking
complex-asymmetrical images before the influence of their art training. Nonetheless, his
correlation of artistic creativity with preference for visual complexity fits with Eysencks
findings about the K factor, and has been corroborated by later experiments (e.g.
Eisenman, 1967).
Berlyne (1971) identifies four variables that contribute to visual complexity:
Irregularity of arrangement, amount of material, heterogeneity of elements, and irregularity
of shape (of elements). Berlyne also identified four classes of complexity, compiled from
his own experiments and those of others, which are comparable to Wolframs (2002)
classification: I. Regular arrangement, small amount of material, homogeneous, regular
shapes. II. Irregular arrangement, more material, heterogeneous, irregular shapes. III.
Fewer independent units, symmetry, non-randomised. IV. More independent units,
asymmetry, randomised (1971, p.198). If there is a problem with this classification, it is that
it appears a little too complicated or even arbitrary; there are many factors involved in the
categories which might be better understood in isolation. Nevertheless, Berlyne identifies
80
some of the significant visual properties that contribute to the perception of visual
complexity.
In summary, there are two critical factors in aesthetic experiments of visual
complexity: the way in which complexity is defined and measured, and the range of
complexity present in the stimuli. There have been many attempts to define complexity in
terms of aesthetic properties, and many different approaches to its classification and
measurement. The lack of common measures of visual complexity hinders comparison
between studies. In addition, the complicated nature of complexity measurement often
restricts studies to relatively simple geometric shapes which are more amenable to analysis.
In the next section of this chapter, we examine the influence of information theory on the
field of empirical aesthetics. We discuss the ways in which it surmounts the limitations of
previous measures and provides a common ground for the measurement of aesthetic
properties, particularly for visual complexity.
The Influence of Information Theory

In this age of computing technology, information has become a familiar concept,
though perhaps one that is not always well-understood. It is most frequently used to
describe abstract quantities of data, particularly in computing, but the origin of the word
relates to the formation of materials or concepts. To the ancient Greeks, the perception of
beauty was an experience based on the sensory in-formation of the Platonic forms
embodied in material objects. The way in which the concept of information became useful
to contemporary empirical research rests on its modern meaning, with its emphasis on
quantification. The birth of information theory in the 1940s coincided with the first digital
computers, which had resulted from a confluence of mathematics, physics and electronic
engineering. In effect, computing is applied information theory. Yet the very first digital
computer the Z3, built in 1941 by Konrad Zuse preceded by seven years the paper that
provided the foundation of information theory Claude Shannons A Mathematical Theory of
Communication (1948). The significance of Shannons paper is that it formalises a theory of
81
communication by quantifying information through the concepts of order, probability and

redundancy. The quantification of information provides a potentially useful tool for
scientific experiments because it allows for the measurement of aesthetic properties. In
addition, information theory offers to art theory and practice a new way of understanding
aesthetic properties and perceptions.
Concepts from Information Theory
Fred Attneave (19191991) was amongst the first psychologists to apply information
theory to aesthetic problems. In Applications of Information Theory to Psychology, Attneave
credits the idea that information is precisely measurable to Shannon (1948) and Norbert
Wiener (Cybernetics, 1948), whose work offered a shiny new tool kit to which more than
a few psychologists reacted with an excess of enthusiasm (1959a, p.v). Attneave identifies
the significant element of this new field:
Perhaps the most fundamental concept of information theory is that of a
continuum extending from extreme lawfulness, or redundancy, or regularity on one
hand, to extreme disorder, or unpredictability, or uncertainty on the other. One end of
this continuum is homogeneity, the other chaos. (Attneave, 1959b, p.503)
This is the idea, introduced in Chapter 1, upon which various informational measures
of complexity are based. We noted then that an informational measure is inconsistent with
the intuitive notion of complexity. The former is a scale of orderrandomness, and the
latter is understood as a scale of simplicitycomplexity. Whilst there is some agreement
between the concepts of order and simplicity (the lower ends of the two scales), there is a
notable difference between randomness and complexity (the upper ends). The difference is
not just conceptual, but also perceptual: Attneave (1954, p.188) notes the point that
random images can give the subjective impression of homogeneity, which is the opposite
end of the informational scale. An image made of 50% random black and white pixels has a
maximum measure of information content, but it is homogeneous in texture and may be
perceived as a simple uniform grey (if the pixels are small enough).
82
One of the earliest writers on information theory and visual aesthetics was Max
Bense (19101990), a scientist-turned-art-theorist who published Aesthetic Information in
1957. Benses information aesthetic treats art as a form of communication, modelled on
the information-theory approach. Bense adapts Birkhoffs (1932) aesthetic measure to the
informational framework by substituting elements of the formula M=O/C, replacing order
with redundancy, and complexity with statistical information (Walther, 2000). In so doing,
Bense incorporates the central concept of information theory, the scale of orderto
randomness, into an aesthetic measure. Bense makes a distinction between semiotic and
numeric aesthetic information; only the latter is open to analysis with information theory,
which deals principally with quantitative data.
Like Bense, the communication theorist Abraham Moles (19201992) distinguished
two types of aesthetics information semantic and aesthetic. Whereas semantic
information is expressible in symbols and is translatable, aesthetic information is
untranslatable it is not the message (its meaning) but the way that the message is
conveyed (such as its medium, syntax and context). Moles used information theory to focus
on the aesthetics of music. By its temporal nature, music fits naturally into an information
theory framework of communication which is based on linear, time-dependent sequences.
The aesthetic information of an object is a quantitative measure of the order/randomness
of its internal structure which provides a scale of reference for the measurement of
perception. Semantic information refers to things external to a sound or an image, whilst
aesthetic information has to do with the internal set of relations in the structure of pictorial
or musical elements. Therefore, semantic information is generally beyond the scope of
analysis for an information-based empirical investigation and the application of information
theory to aesthetics is restricted to the formal properties of a work of art.
In his information aesthetic, Bense introduced the idea of macro- and microaesthetics: Macro-aesthetics is concerned with the evident realms of perceptions of the
aesthetic object, micro-aesthetics with the not-evident realms of the aesthetic object
83
(Walther, 2000). The micro-aesthetic includes not only the un-perceivable structural
relations of elements, but also the a-perceptual elements such as the relationship between
theory and the work (Gianetti, 2004). As such, Benses theory deals with pure information
as well as with aesthetic information of a physical nature, but whereas Moles and Attneave
were writing principally for scientists, Bense was writing for artists. Accordingly, the
concept of micro-aesthetics has greater potential application in art practice than for an
empirical investigation, and it is discussed further in the next part of this chapter on visual
complexity in aesthetic practice.
Attneave (1954, p.189) employs information theory concepts to describe perception
as a process of economical description, analogous to data compression, in which
redundant visual information is discarded. In the case of random and chaotic images, we
reduce their structural detail (what Bense might call micro-aesthetic information) and
perceive a kind of overall texture (macro-aesthetic) as an economical description of sensory
visual information:
Perception might be conceived as a set of preliminary data-reduction operations,
whereby sensory information is described, or encoded, in a form more economical
than that in which it impinges on the receptors. (Attneave, 1959a, p.82)
This is an explicit application of concepts from information theory to psychological

phenomena, but at this stage in the field these are hypotheses that lack empirical support.
The value of the ideas is that they not only suggest mechanisms of perception but also the
types of measurement which are likely to prove appropriate in the quantitative
psychophysical study of complex perceptual processes (Attneave, 1954, p.192). The type
of measurement Attneave proposes is based on an informational scale.
Visual Order and Randomness
To explore the visual scale of orderrandomness, Attneave (1959b) created patterns

with stochastic (non-deterministic, i.e. random) processes: Starting with a blank grid, a coin is
flipped to decide whether to generate an upward- or downward-sloping diagonal line to fill
84
each square in the grid. This random image is then used as a basic repeating unit to form a
regular pattern. The repeating unit can be altered by randomly changing some of its
elements as it is repeated, which generates an overall pattern with random deviations from
the regular repetitive pattern. Information theory allows us to quantify the regularity or
randomness of the resulting patterns. The question Attneave was exploring was how much
randomness has to be introduced into the repeating pattern before it is no longer perceived
as regular. Figure 25 illustrates four images generated with this method, and we can see that
with only 15 per cent randomness it is already difficult to identify the underlying regular
pattern. Attneaves hypothesis was that most people would prefer slightly irregular patterns
to completely regular ones (1959b, p.505), but unfortunately this was not tested
experimentally.
Original
5%
10%
15%
Figure 25 Stochastic pattern composition (Attneave, 1959b). A repeating unit is

composed of a 3x3 grid of random diagonal lines, which is itself repeated 3x3 to form
a regular pattern (top left). Parts of the unit are altered in each repeat to give patterns
with 5, 10 and 15 % random deviation from the original pattern.
85
Another early contribution to the field of informational aesthetics was made by

Wilhelm Fucks, who had conducted research on the mathematical analysis of style (1952)
which predated the advent of information theory. Unfortunately, most of the relevant
publications by Bense and Fucks are unavailable in English translation, and so they must be
omitted from this contextual review. Nevertheless, we may at least glimpse something of
Fucks approach with two images from his book Exaktwissenschaftliche Musikanalyse (1965),
in which he applied information theory to the analysis of music: Figure 26 shows his
mapping of patterns of musical tones from various pieces of orchestral music onto a grid,
the dots presumably identifying repeating patterns and correspondences between notes or
phrases. What is interesting about these graphics is that although they are the result of
analysis rather than synthesis, they display elements of order and randomness comparable
to those present in Attneaves stochastic composition processes (Figure 25) and Bridget
Rileys dot paintings (Figure 28). If Fucks analytical images represent a valid mapping of
musical structure, then we may infer that a similar aesthetic of order and randomness is
involved in musical composition as it is in visual art. From the standpoint of visual art, it is
intriguing that what may sound like rather dry scientific and information-based analyses
furnish such appealing images as those produced by Attneave and Fucks. The
correspondence of musical and visual aesthetic structure also suggests the possibility of a
universal information aesthetic that may reach across creative disciplines and artistic media.
86
Figure 26 Wilhelm Fucks. Images from Exaktwissenschaftliche Musikanalyse (1965).
The Influence of the Computer
The advantage of the informational approach is that it allows for precise

measurement. The problem with the approach is that the amount of visual information in
artworks makes this measurement difficult. Arnheim put it like this:
The problem at hand can be approached also from the other end. Instead of analyzing
simple groups of elements, one can start with the peculiar patterns we call works of art
and ask what kind of information they furnish. [] Can a theory fitted to the
respectable objectives of engineering cope with such mavericks? (1959b, p.503)
This problem is not surprising, since we have learnt that complexity is associated
with difficulty of description and that artworks are visually complex. In this situation, it is
therefore natural to approach the problem by dealing with relatively simple stimuli which
are easy to make and measure. Before the advent of computing, this seemed to be the only
option for empirical investigations of visual complexity. The computer solves the problem
by speeding-up the process of making and measuring stimuli. For example, the images in
Figure 25 that Attneave said were extremely laborious to create by hand through
stochastic composition processes (1959b, p.506) I managed to reproduce in around ten
minutes by writing a Mathematica program that took only a few seconds to generate images.
87
Figure 27 shows an image I made based on Attneaves stochastic procedure, with a

thousand diagonal elements and 100 per cent randomness. The image also demonstrates
the property, noted by Attneave, that randomness is perceived as an overall texture.
Figure 27 A 100x100 grid of randomly-selected diagonal elements.
A more recent investigation into the role of order and randomness was undertaken
by Neil Dodgson (2008) who, unlike Attneave, had the benefit of using a computer.
Dodgson investigated the visual structure of Bridget Rileys early op-art. Unlike some of
Rileys more familiar regular patterns, these pictures contain apparently random elements as
well as order (Figure 28). Without knowing the particular technique employed by Riley,
88
Dodgson attempted to recreate her work by using computer algorithms to generate similar
types of images. The algorithms produced regular patterns with random deviations, and the
proportion of this randomness was manipulated.
Figure 28 Bridget Riley (1965) Fragment 6/9. Screenprint on Perspex, 625 721 mm.
By combining order and randomness in a computational process, Dodgson was able

to recreate the aesthetic properties of Rileys artwork, and to generate appealing visual
complexity. The images that were closest to the original artworks resulted from an
algorithm in which about one third of the picture elements in a regular pattern were
randomly added, moved or deleted, such that the images contained two thirds order and
one third randomness. I agree with Dodgsons conclusion:
The particular examples examined here have introduced very limited randomness and
yet even this limited randomness provides a great deal of complexity. This is a tribute
to the pattern-detection systems in the human brain. (2008, p.122)
Dodgson attributes the perception of complexity to a perceptual process of patterndetection. Pattern-detection is how we recognise order, complexity, and randomness all
of which are kinds of pattern. The formal constraints with which Dodgson structures his
89
images also limit the perceptual grouping of picture elements and thus the type of pattern
we might perceive. For example, in a regular chequered pattern of black and white squares,
we would likely perceive only regular perpendicular and diagonal structures. Disrupting this
regularity by introducing randomness opens up a pattern to alternative perceptual
groupings. The implication from Attneave (1959b) is that introducing more randomness
increases perceived complexity only up to a point, beyond which images just appear
random. In other words, the findings of Attneave and Dodgson imply the existence of a
threshold to the perception of complexity in relation to an informational measure of
complexity. This threshold we identified in Chapter 1 (p.45) as also being proposed by
Wolfram (2002, p.559) and Donderi (2006b, pp.8788).
Possible explanations for visual pattern-detection include the Gestalt principles of
grouping such as those of figure/ground, similarity, proximity, and symmetry (Bruce,
Green & Georgeson, 2003) and Richard Gregorys (1980) theory of perceptions as
hypotheses. The Gestalt approach could be called bottom-up because it describes
perceived forms as emerging from constituent units, whereas Gregorys approach is more
top-down since it depends on cognitive hypotheses based on prior experience. The first
significant point following from Attneaves results, however, is that an image becomes
more interesting when the number of possible groupings or hypotheses is increased. The
second point is that this visual interestingness can be promoted by the combination of
order and randomness, which gives rise to the perception of visual complexity. The most
complex image in Dodgsons paper (also the image that I find most appealing) is the one
that illustrates the transformations required to get from the regular underlying pattern to
the final version (Figure 29). The added colours and shapes increase the variety of elements
and increase the visual complexity of the image. The arrows and crosses invite a visual
exploration, especially when seen in the context of the original and underlying patterns.
90
Figure 29 How to get from simple order to complexity by introducing randomness.

The larger image shows how to transform the underlying regular pattern (top left) to a
representation of Rileys Fragments 6/4 (bottom left), by moving, adding and deleting
picture elements. (Dodgson, 2008)
The Computational Approach to Visual Perception
After the breakthrough of understanding visual properties in terms of quantifiable

information, the next logical step is to understand the process of perception in terms of the
computation of this information. David Marr (1982) describes vision as a complex
information-processing system, whose purpose is to reveal what is present in the world,
and where it is (p.3). We know that the input to this process is a two-dimensional visual
array (light entering the retina) and that the output is perception of the three-dimensional
world. The problem for the science of perception is: how does it work? In Marrs
approach, this formidable problem is divided into separate problems of understanding the
computation (what the system does with the array of visual information, and why), the
algorithm (how those computations are tackled) and the implementation (the mechanism or
hardware) involved in visual processing. The theory attempts to explain how a visual array
is processed to give a two-dimensional primal sketch of a scene composed of visual
91
primitives such as edges, lines and blobs, and how these primitives are further processed
into a 2D-sketch which includes an indication of depth relative to the observer, and
then into recognizable 3D objects and scenes.
Support for Marrs theory comes from neurophysiology and psychophysics, as
follows: The primitives that compose the 2D-sketch are mainly processed in the retina.
David Hubel and Torsten Wiesel (1962) identified that the outputs of photoreceptors (rods
and cones) in the retina are linked together to form receptive fields. In the visual system, a
receptive field is a group of photoreceptor cells in the retina that acts as an informationprocessing unit. The way in which the receptors are connected defines a small cluster
which is divided into a centre and surround. One type of receptive field responds to
stimulation in the centre, whilst in another type the same stimulus inhibits its response. In
general, receptive fields give a lower response for overall stimulation, and respond most
strongly when only one part (either the centre or the surround) is stimulated. This happens
when an edge is perceived, because a perceived edge is a relative difference in light intensity
(Marr reserves the word edge for perceptual phenomena and intensity gradient for the
physical equivalent). Because this process occurs in the retina, the recognition of edges
(i.e. adjacent light and dark areas) functions at a pre-conscious level of perception
(Snowden, Thompson, and Troscianko, 2006). The operation of receptive fields has a
direct bearing on the amount of detail we can perceive: smaller and more densely-packed
receptive fields enable the perception of finer detail. As with Marrs distinction between
edges and intensity gradients, detail is the perceptual equivalent of what is known as
spatial frequency, which is defined as the number of cycles (e.g. light and dark stripes) per
degree of visual angle (Bruce, Green & Georgeson, 2003).
Naturally, the computational approach involves a mathematical representation of
these processes and the use of computers to carry out the kind of laborious tasks that
visual processing requires, but what allows for this is the separability of the levels of
description. For example: if, instead of perceiving light intensity, the computational task is
92
to add numbers together, it makes no difference to the result whether we use decimal or
binary systems to represent the numbers, whether the choice of algorithm is efficient, nor
whether the calculation is implemented mentally or with an abacus or a computer.
Similarly, light levels and colours can be represented by neurochemical signals or binary
digits. Separating the problems frees up investigation to explore the most salient features
because at certain levels of description some computations are equivalent. For example, 2
+ 2 = 4 is equally true whether performed in the brain or on a computer. In this way, the
computational approach justifies the use of information technology as an investigative tool
for visual perception. Computational models based on this approach can be tested
empirically against human biology and psychology to determine how accurate and useful
they are. For instance, Marr & Hildreths (1980) algorithm, which was designed to model
the edge-detection process of receptive fields in the retina, has been found wanting as a
biological theory, but is now used in machine vision and in image-processing software such
as Photoshop, as illustrated in Figure 30.
Figure 30 Edge-detection in image processing: before (left) and after (right).
Marrs separation of the visual processes and subsequent identification of the edgedetection algorithm means that the computation of these receptive fields can be modelled
on computers. Its effect on digital images illustrates the identification of edges and lines
from a visual array; a stronger edge is more visible when contrast is greatest, as seen in
Figure 30, which was created by using a convolution algorithm similar to Marrs (the
Laplacian-of-Gaussian operator). The strongest edges appear to lie on the borders of the
93
lightest and darkest colour patches, and the weakest edges are next to the colours that are
closest in tone to the mid-grey background. The resulting greyscale image also represents
the fact that edge-detection is a process independent of colour perception it only depends
on light intensity.
If we accept Marrs sketch theory, the implication is that the perception of visual
complexity is a function of these visual primitives. More primitives and more types of
primitives make for greater visual complexity. We have already identified the quantity,
variety and order of picture elements as significant objective properties of visual
complexity; now with Marrs visual primitives we have identified significant subjective
picture elements: Edges, lines and blobs are the basic perceivable forms, which are further
distinguished by tone and colour.
Zeki and Lamb (1994) describe the perception of motion in terms of receptive fields
in the visual cortex area of the brain. Like those in the retina, these receptive fields are also
areas in which stimulation of multiple receptors combine to affect the response of a single
neuron. According to Zeki and Lamb (1994, p.613), artworks stimulate motion-sensitive
receptive fields in various ways. For example, the static representations of motion in the
Futurist paintings of Duchamp and Jean Tinguely stimulate different regions to those
activated by Alexander Calders mobiles. Zekis neuroaesthetics research explores the
psychophysical relation of visual art to neural activity. By understanding the perceptual
effects of art in terms of brain function, Zeki postulates that the artist is, in a sense, a
neuroscientist, exploring the potentials and capacities of the brain, though with different
tools (2009). The conclusion is that because kinetic art activates certain neural responses,
the artist who makes this artwork exploits the response and intuitively explores the
structural function of the brain. Whether this neuroaesthetic approach is the only way in
which we can fully understand how artworks give rise to aesthetic experiences, as Zeki
maintains, appears to be open to question. The focus and methods of Zekis research are
beyond the scope of the current project, but the modelling of perceptual processes enabled
94
by Marrs computational approach to vision has potential application for empirical studies
of visual complexity, as we see in the following section. The strength of Marrs theory is
that its separation of the problems allows for the modelling and analysis of perception.
Furthermore, its separation of the processes justifies the computational modelling and
analysis of aesthetics by incorporating the concepts of computation and information into a
theoretical framework of visual perception.
Aesthetic Measures
We noted previously that perceptions can be measured with various psychological
techniques such as magnitude estimation scaling (MES). MES provides a general method
that can be applied to a variety of subjective perceptions, including visual complexity. In
the following section we examine some of the contemporary techniques employed in the
measurement of objective and subjective visual complexity, with a view to identifying the
appropriate research methods for the current project. The potential of these methods for
the project is evaluated principally in terms of their ability to cope with images of art and
design and the strength of their correlation with human perception.
Eye-Tracking
Miall & Tchalenko studied artists movements during portrait drawing (2001) and
eyehand coordination strategies in copying complex lines (2008). Equipment tracked the
direction of the participants gaze whilst drawing, and revealed different patterns of
drawing strategies in the temporal cycles of looking in turn at the original and the copied
drawing. Besides eye-tracking, Paul Locher (2003) explored various experimental methods
to assess the contribution of pictorial balance in aesthetic perception of visual art. Lochers
results show that balanced pictorial arrangements attract more fixation on the surface than
unbalanced arrangements, and confirm that pictorial balance is a significant factor in the
creation and perception of visual art.
The eye-tracking technique does not constitute a measure of complexity as such, but
the quantitative nature of the data it furnishes suggests that some form of informational
95
measure could in theory be applied. In general, it seems that a more complex image invites
a more complex perceptual exploration as described by the patterns of gaze-fixation
identified by eye-tracking. It would be interesting to determine the correlation between the
complexity of the sequence of actions and the visual complexity of the resulting images.
Certainly, it takes more time to create a complex oil painting than a simpler charcoal
sketch, and a greater variety of movements are performed in the process of painting with
oils. Overall, though, despite its relative simplicity and practicality, the eye-tracking method
is only an indirect measure of visual complexity, and may not be the most appropriate
technique for the present study.
Aesthetic Preference
There have been numerous studies of preference for various types of visual stimuli,
including many based on the experimental framework of Berlyne (1971) discussed earlier.
Prominent researchers in the field include Colin Martindale and Paul Locher, each of
whom has served as president of the International Association for Empirical Aesthetics.
Both authors have employed an information-based approach to aesthetic measurement in
relation to the preference for complex images.
Martindale, Moore & Borkum (1990) carried out experiments based on Berlynes
theory that aesthetic preference is a function of hedonic arousal. The authors identify the
influence of Berlynes aesthetic theory in the number of citations of his work from three
international conferences on experimental aesthetics: Berlyne is the most-cited author,
appearing in 39% of the conference papers, followed by Arnheim and Gibson with 15%
each (Martindale, Moore & Borkum, 1990, p.53). Martindale and colleagues experiments
aimed to test the predictions of Berlynes theory, such as the prediction that the collative
variables including visual complexity are amongst the most significant determinants of
aesthetic preference. The first of these experiments focused on random polygons as visual
stimuli, and complexity was measured in terms of the number of turns (corners) present
in the shapes. Their results were overwhelmingly negative for Berlynes predictions,
96
showing that semantic information is more influential than the collative properties in
predicting aesthetic preference. These results were validated in further experiments which
used representational drawings and paintings as stimuli. However, when the experiments
included images of both representative and abstract artworks in addition to random
polygons, the result is that complexity is a better predictor of preference (Martindale et al.,
1990, p.73).
Martindale and colleagues (1990) found that their results were not affected by
knowledge of art. Participants art knowledge was measured by indicating how much they
knew about art in general on a numerical scale ranging from a lot (1) to nothing (7).
Locher, Smith and Smith (2001) also investigated the role of art experience in the aesthetic
judgement of visual stimuli. The stimuli included original paintings by Rembrandt, Breugel
and Vermeer from the New York Metropolitan Museum of Art, and reproductions of
those artworks in the format of digital images and projected slides. The results show that
ratings of aesthetic qualities were similar for both art-trained and untrained viewers, and
that the ratings were comparable for both original images and reproductions. The authors
name this indifference to stimulus format facsimile accommodation, and their results
provide experimental evidence of the comparability of aesthetic evaluation between these
diverse conditions. It suggests that although the originals were judged to be slightly more
interesting and pleasing, the effect of reproduction may not be as significant as expected on
the judgement of aesthetic properties, including visual complexity.
Marcos Roberts (2007) investigated the divergent experimental evidence for
Berlynes (1971) inverted U shape correlation between complexity and preference,
following on from Martindales (1990) challenge to the theory. Unlike the experiments of
Martindale and Locher, Roberts methods were based solely on subjective estimates of
complexity rather than objective measures. Digital reproductions of artworks of a variety of
styles including abstract and representational artwork were used as stimuli. From the
results, Roberts identified that the most significant factors of visual complexity are: a) the
97
amount and variety of picture elements, b) the organization of elements, and c) the
asymmetry of the elements.
Edge Detection
Experimental psychologists have employed Marrs theoretical framework and used

edge-detection algorithms to measure visual complexity. Forsythe, Sheehy and Sawey
(2003) investigate various measures of graphic icon complexity as a way of finding an easier
method than survey and questionnaire techniques. Of six tested aesthetic measures, the
strongest correlations are with structural variability (r = 0.65, p < 0.008) and edge
information (r = 0.64). The conclusion is that these computational measures match
perceived complexity closely enough to offer a workable alternative to the costly and timeconsuming surveys of icon design (2003, p.341). Forsythe (2008) uses various imageprocessing methods, including edge-detection algorithms, to examine the influence of
image familiarity on judgements of visual complexity. The analysis reveals a familiarity bias
in estimates of visual complexity: participants tested over a week already showed that their
familiarity with images affected their reported perceptions of visual complexity. The
explanation for this bias is that when observers are asked to consider the complexity of an
image, they also process task-irrelevant information, such as its familiarity and
meaningfulness (Forsythe 2008, p.124). Forsythe proposes the computational edge
detection technique as an alternative to perceived complexity that is less susceptible to bias.
The findings support the idea that a complex image is one with many edges,
Processing edges takes time, and mainly happens prior to semantic processing: Regardless
of experience, a more detailed object will take longer to reach semantic processing, because
the preprocessing of many edges is perceptually demanding (Forsythe et al., 2003, p.341).
Perceptually demanding means that it is a difficult and complicated process which takes
time and effort to achieve. This substantiates the connection between the number of edges
in an image and its visual complexity, which we have understood as being related to the
difficulty of description or execution. Weaknesses of this measure in regard to the aesthetic
98
perception of art include the fact that it disregards colour information and focuses only on
tonal variation. However, since perceptual edge-detection also operates before colour
processing and has been shown to be a significant factor in the construction of sensory
percepts (Marr, 1982), this criticism of the measure may be offset with evidence of its
correlation with human perception.
Fractal Measures
As well as a noun, the word fractal is an adjective which describes the intricacy of
these self-similar shapes. This intricacy is known as its fractal dimension. For example, a
fractal line that almost fills a two-dimensional surface has a fractal dimension greater than
one but less than two. Another fractal line that is even more intricate covers more of the
surface and has a higher fractal dimension. Ordinary Euclidean shapes such as squares
and circles have a fractal dimension equal to their topological dimension, which is always
a whole number (a line = 1, square = 2, cube = 3, etc.). Mandelbrot (1982) defined fractals
as the shapes for which their fractal dimension is not equal to their topological dimension
but is a fraction of that number (technically, when the Hausdorff-Besicovitch dimension
exceeds the topological dimension). Therefore, we can understand the meaning of the
adjective fractal as fractional dimension, and we can use it as a measure of complexity.
There are a variety of measures that capture this property, and a number of them
have been used to quantify the visual complexity of stimuli in aesthetic experiments. For
example, Julien Sprott (1994) employed two measures, fractal correlation dimension and
the Lyapunov exponent, to quantify the visual complexity of images. Sprott used iterated
function systems (IFS) to generate stimulus images of strange attractors, which are fractal
shapes in phase space (such as the Lorenz attractor in Figure 13). The Lyapunov exponent
measures the unpredictability of the dynamic process that generates the image (in this case,
an IFS). It measures the rate at which two adjacent points diverge in phase-space
(somewhat like measuring divergence of points on the Lorenz attractor, or the divergent
paths of the leaves that I dropped in a stream to demonstrate the effects of chaos). Sprott
99
generated 7500 fractal images and graded them according to his own aesthetic evaluation
on a five-point scale. He found a correlation between fractal measures and aesthetic rating:
the highest-rated images had also had the largest fractal values (Lyapunov exponent = -0.24
and a fractal dimension = 1.5).
Aks and Sprott (1996) performed a similar investigation, this time with a group of
test participants instead of the authors aesthetic evaluation. These tests measured the
fractal dimension (F) space-filling of the shapes, and the Lyapunov exponent (L). With 24
undergraduate students from art and science courses, the result was that the most preferred
images had values of F = 1.26, and L = 0.37. The authors claim that these values are close
to those found in natural objects, such as the coastline of Britain, which has a fractal
dimension of 1.25 (Aks & Sprott, 1996, p.2). A similar preference for images with fractal
values close to natural forms is reported by Taylor et al. (2003), who investigated the
aesthetic response to natural, artificial and hand-made fractals. Using a variety of these
stimuli with a range of fractal dimension between 1.1 and 1.9, the participants task was to
state which of a pair of images of differing fractal values was the most visually appealing
(2003, p.95). The results indicate a grouping of fractal values across all three types of image:
low preference for the lowest (1.11.3) and highest (1.51.9) fractal values, and high
preference for middle range of values (1.31.5). Further support for this preference comes
from Hagerhall et al. (2008), who studied the neurological response to viewing a range of
fractal images. With electroencephalogram (EEG) monitoring, Hagerhall found that
Patterns with a fractal dimension of 1.3 elicited the most interesting EEG, with the
highest alpha in the frontal lobes but also the highest beta in the parietal area, pointing
to a complicated interplay between different parts of the brain when experiencing this
pattern. (Hagerhall et al., 2008, p.1488).
While the results of these three experiments support the idea that medium
complexity fractal images are the most aesthetically pleasing, it also appears to question the
suggestion by Aks and Sprott (1996) that the explanation is a connection to natural fractal
forms. The problem is that natural fractals can be found over the entire range of fractal
100
dimensions. For example, Taylor et al. (2003) collated measures of natural fractals from a
variety of sources that evidence a range from as low as 1.1 in coastlines up to 1.9 in plants
(2003, p.92), and their experiments included samples covering this range. 7 Overall, the
findings of these experiments accord with Fechners principle of the aesthetic middle and
Berlynes (1971) framework of aesthetic preference for stimuli of medium intensity.
In a follow-up to their earlier experiment, Aks and Sprott (1996) investigate whether
the creative ability of participants affects preference for fractal values of the images. They
suggest that scientists may tend to see beauty in simplicity whereas artists may be more
tolerant of complexity (1996, p.9). Fifteen of the participants from the previous
experiment were assessed for creative ability with a questionnaire and grouped according to
whether their creativity was above or below the mean. Apart from a marginal preference
for detail amongst self-rated divergent thinkers, the analysis shows no significant
difference in preference between the groups, possibly due to the small sample size of
participants. These results suggest that self-reported creativity is a poor indicator of
aesthetic preference.
Few empirical studies of fractal aesthetics have utilised artworks as stimuli. The
reason is likely that fractal measures are of little use for non-fractal images, and that fractal
artworks are relatively scarce. However, there is one artwork that offers a large collection
of fractal material for investigation: Scott Draves (2009) Electric Sheep project (which is
discussed further in the section on art practice). Electric Sheep is an interactive online
screensaver in which users can vote for their preference of animated fractals (called sheep)
by pressing the up or down arrow key to indicate like or dislike of the current image. Votes
are collated in a central server which deletes the least popular sheep and picks out pairs of
the most popular ones to produce a new animation by combining the parameters of the
parents using a genetic algorithm. Users can also design their own sheep and upload them
7
Since these experiments examine two-dimensional images, the range of fractal dimension is restricted to
between 1 and 2.
101
to the server, so a new flock consists of a mixture of computer-generated and humandesigned animations, all of which are then subjected to the voting and culling/breeding
process. In its operation, therefore, Electric Sheep accumulates aesthetic preference into its
generative processes and provides a large resource for empirical investigation. A study
conducted by Draves et al. (2008) is based on the aesthetic judgements of 20,000
participants on 6,400 images. In the study, preference displayed an inverted U correlation
with fractal dimension, with a peak around 1.5. Once again, this reinforces the finding of
preference for mid-value fractals, though somewhat higher than previous results.
Taylor, Micolich and Jonas (2002) find that Jackson Pollocks drip paintings are
fractal. This may seem surprising when we consider the chaos in the drip technique and the
resulting appearance of disorder in the paintings, but measurements indicate fractal
dimensions in the range 1.31.9. Images of the drip paintings are used in Taylor et al.s
(2003) study to represent the category of human-made fractals. In a study of abstract
expressionist art, Mureika, Cupchik and Dyer (2004) suggest that the fractal dimension
measure is unable to distinguish between different styles. They propose a multifractal
measure, which offers a more refined statistical measure of self-similarity that is able to
make distinctions between styles. The study shows that among expressionist paintings, the
work of Pollock is unique in its degree of irregularity and roughness of edges (2004, p.56).
The authors suggest that the visual systems propensity for detecting edges which are so
abundant in Pollocks work allows for the recognition of order in the chaotic paint drips,
and that this is a possible source of their aesthetic appeal.
Data Compression
Beginning with a research project for the Canadian Ministry of Defence into the
readability of radar displays and naval charts, Donderi (2000) had the idea of using the
data-compression of digital image files as a measure of complexity. In short, the idea is that
simple images make small files when compressed, and more complex images make larger
compressed files. The argument for this method is that information theory (Shannon &
102
Weaver, 1948) provides a theoretical basis for the justification of using compression
techniques as a measure of visual complexity. As noted in the previous chapter, data
compression is a form of the minimum description length (MDL) measure of complexity,
and MDL is equivalent to the (technically incomputable) algorithmic information content
(AIC) or Kolmogorov complexity measure. Of the various measures of visual complexity
discussed thus far, Donderis method is the most direct application of an informational
scale of measurement. Because data compression algorithms are incorporated into many
common digital image file types, this method involves little more than creating an image
file with maximum (near-optimum) compression. Therefore data compression offers a
simple and efficient objective measure of informational complexity, and it follows that the
size of compressed image files is a measure of visual complexity.
The experiments of Donderi & McFadden (2005) provide empirical support for the
measure. They show that the difficulty of chart and radar display tasks increases with the
size of compressed image files. Since we have understood complexity as being related to
difficulty, these results support the method as a measure of visual complexity. Further
evidence for the validity of this relationship comes from a second experiment that involved
the task of identifying which of two small geometric shapes were present in aerial images of
the earth. The logarithm of compressed JPEG file size corresponds with the difficulty of
the task as measured by the success in identifying the correct shape. The idea behind the
experiment is that a more complex image makes it more difficult to identify the correct
shape placed within it. To corroborate this idea, Donderi and McFadden (2005) also
measured the subjective complexity of the images with Stevens (1975) magnitude
estimation method. Eleven participants judged visual complexity by rating 52 images with a
number which was in proportion to its perceived complexity. The analysis shows a weak
but significant correlation between file size and task accuracy (r = 0.14, p < 0.005) and a
strong correlation between file size and MES ratings (r = 0.77, p < 0.0001). The two
103
experiments confirm that JPEG file compression corresponds with task difficulty and
perceived complexity.
Despite the encouraging results of these experiments, it appears that there may be a
problem with the file-compression measure of complexity, which stems from the disparity
between information-based measures and perceived or intuitive complexity. In Chapter 1,
we noted that randomness is rated highest by informational measures, and that subjective
complexity lies somewhere in the middle of an informational scale, between order and
randomness (see Figure 16). It implies that the correlation between file size and subjective
perceptions of visual complexity may break down for random or near-random images. This
problem is summed up as: Can a display become so complex that appears simple?
(Donderi & McFadden, 2005, p.78). Donderi makes reference to Wolframs (2002)
observation that the perceived complexity of cellular automata images depends on the
identification of patterns. A random image, whilst being rated highly in terms of objective
complexity (as measured by file compression or AIC), has no patterns (no redundancy) and
so it may appear to be subjectively simple:
We do not know whether, if we study displays with even higher levels of JPEG file
complexity, we will begin to see a reduction of perceived complexity and a
concomitant change in performance. In short, we do not know where to locate the
inflection point that is implied by the Wolframs observation about the perceptual
simplicity of even more complex visual displays. (Donderi & McFadden, 2005, p.78)
The inflection point that Donderi mentions is the peak of the hypothetical inverted
U shape correlation between objective and subjective complexity. Before this peak the
two measures correlate positively, and beyond it perceptions of complexity decrease as
objective complexity increases. Thus the inflection point represents a perceptual threshold
of complexity on an informational scale; we are unable to perceive any more complexity.
The problem was investigated further by one of Donderis research students, Christina
Besner (2007), who studied whether the correlation of objective (file size) and subjective
(MES) complexity measures would still hold for random images. The stimuli comprised
104
images of nature scenes, camouflage patterns, and images made by combining two
together. Besner conducted two experiments with ten participants each, one with a set of
138 images and the other using 16 from that set. The result is a moderate correlation
between file size and estimated complexity (r = 0.553, p < 0.001) which becomes stronger
when the larger, more random files are ignored (r = 0.827, p < 0.001). From this result,
Besner draws the conclusion that Donderis measure remains valid for the images used; the
images were not so objectively complex that they appeared to be simple. The result can be
interpreted differently, however: As more random images are included, a weaker linear
correlation is actually what one might expect to find if the relationship between objective
complexity (based on file compression) and perceived complexity (MES) has an inverted
U shape. Besner acknowledges that the images were possibly just not random enough to
see this effect, which seems a fair assessment given that the stimuli were not generated
from random processes.
We have seen that this informational measure of visual complexity does indeed
correspond with perceived complexity for images such as radar displays and images of
natural scenes. But on the other hand it appears that there are limitations to this method
when it comes to images of greater objective complexity, that is, as images approach
randomness. Given that we understand images from art and design as being amongst the
most complex type of image in subjective terms, we face the following question: Will this
objective measure of visual complexity also work for images from art and design?
Remarkably, Arnheim had identified much earlier this problem of informational
measurement in the arts:
From this reasoning follows what must seem an appalling paradox to any friend of the
arts, namely, that complete disorder or chaos provides a maximum of information
whereas a completely organized pattern yields no information at all. The friend of the
arts, of course, takes the opposite for granted. (Arnheim, 1959, p.501)
With Donderis file-compression method, we have identified a potentially useful and

practicable measure of visual complexity for a study of art and design. We have also
105
identified the following gaps in knowledge: Firstly, is the relationship between compressed
file size and perceived complexity as suggested by Donderi and Wolfram that is, can we
find empirical evidence for a perceptual threshold of visual complexity where the
correlation between information content and subjective complexity has an inverted U
shape? Secondly, will the file-size measure of complexity and its correlation with subjective
complexity also work with images from art and design?
Summary
In this review of empirical aesthetics, we have encountered a variety of approaches.
Researchers have variously employed cognitive (Gibson), motivational (Berlyne), Gestalt
(Arnheim), and computational (Marr) approaches to visual perception and aesthetics.
Amongst these approaches, opinion is divided on whether aesthetic perception differs
from everyday perception. Similarly, there are many possible routes to understanding and
measuring visual complexity some intuitive and others more abstract and informational.
Empirical studies of aesthetics can be divided into two categories. What might be
called the analytic (or holistic) approach employs real artworks as stimuli, whereas the
synthetic (or elementary) approach creates stimuli for testing, such as geometric diagrams,
which serve as abstractions of certain features of visual art. Experiments involving genuine
artworks as stimuli can be traced back to the work of Fechner, who in 1871 attempted an
experiment to record the responses of an audience to two versions of the younger Hans
Holbeins Madonna with Burgomaster Meyer (Berlyne, 1971, p.10). Either alternative has
problems: The strength of the analytic approach is that it is based on observations of actual
artefacts, but these stimuli also have features that might unduly influence the properties
under investigation. The selection of stimuli from extant artefacts is thus fraught with
difficulties, especially because artworks tend to be held in diverse locations, which may
hamper the collection of a representative sample.
106
The synthetic approach has the benefit of circumventing these practical difficulties in
experimentally measuring or controlling the properties of real artworks as stimuli, since the
properties of the stimuli can be fully controlled and features of interest can be isolated. The
disadvantage of this approach is that it weakens the validity of generalizations that can be
made about actual artworks, since the data is based on observations of artificial stimuli in
test conditions which are often far removed from the typical gallery experience of visual
art. There is reason to believe that Donderis file-compression measure of visual complexity
is able to cope with both artificial stimuli and images of artworks since there is no
fundamental difference between the two types of image when reduced to the basic
properties of digital image files. The difference between artificial stimuli and actual
artworks is, after all, a matter of context, as discussed in the previous section on the
aesthetic theories of art. For that reason, the file-compression measure is a strong candidate
to be used in the following tests.
Aesthetic Practice
The aim of this section is to identify and examine creative practices that engage with
the aesthetics of visual complexity. The scope includes not only practices which manifest
complex visual material, but also practices that aim for simplicity of expression. This
material includes examples of digital and generative art which explicitly deal with concepts,
models and techniques from the sciences of complexity (those which use cellular automata,
for example) but these criteria do not constrain the scope of the contextual review.
Although the focus is mainly on two-dimensional visual material, some examples of threedimensional art and sound art are also considered. The choice of material in this section
cannot be expected to represent an accurate picture of the full range of todays art
practices; rather, the selection is biased towards artworks and artists I have encountered
personally and which stimulate my own practice. The examples of art practice are roughly
grouped by the complexity of the artwork, and are arranged in order from the simple and
regular to the chaotic and random in the following sections.
Aesthetic Practice
107
Simplicity and Minimalism

It may seem counter-intuitive, but I suggest that minimalism is the art movement
that focuses most explicitly on the concept of aesthetic complexity. Donald Judd is
regarded as a key figure in the genre, but minimalism is a term he rejected, preferring to
describe his work as the simple expression of a complex thought. Other artists associated
with minimalism rejected the term as a description of their work, despite the outward
appearance of reductive strategies in their artwork. Like other minimalists, Judds work
utilizes simple geometric forms which are often repeated in two- or three-dimensional
space. The simplest two-dimensional artwork is the monochrome a format adopted by
artists in a variety of art movements other than minimalism, including Kasimir Malevich,
Alexander Rodchenko, Ad Reinhardt, Robert Rauschenberg, Robert Ryman, Yves Klein
and Gerhard Richter (Figure 31).
Figure 31 Gerhard Richter, Grauer Spiegel (1991), coloured glass, 3000 1750 mm.
These artists did not only produce monochromatic works, however, and they were
not only interested in what we understand to be the central concerns of minimalism. These
Aesthetic Practice
108
concerns include a truth to materials and a restriction of focus to the internal structure of
the artwork without imitation or representation of external reality. In this way, minimalist
art claims to be more honest in its lack of illusion and its use of materials that simply
speak for themselves and are not subjugated to a role of signification. Reducing illusory
distraction and pictorial clutter allows the materials to speak and provides space for us to
reflect on the act of perceiving them. So even though its aim is to reduce formal
complexity and increase order, minimalism puts the concept of visual complexity at the
forefront of its theoretical and aesthetic considerations.
There is an interesting tension between simplicity (as a measure of complexity) and
minimalism (as an aesthetic approach), which suggests that we should not assume that the
one necessarily follows from the other. Jrgen Schmidhubers low complexity art uses
concepts from algorithmic information theory to create short and simple computer
programs. Schmidhubers stated goal is to create minimally complex programs that
generate images that look right (1997, p.97). Creating low-complexity algorithmic art is
difficult, he says, because we can never be certain that we have found the shortest
algorithm for the job. This uncertainty results from the fact that algorithmic complexity
(the length of the shortest program that will generate a given pattern) is incomputable. The
resulting images are mainly composed of circle segments which combine to create outlines
of recognisable shapes, such as butterflies and flowers. Obviously, looking right is highly
subjective, and to my eyes the style of Schmidhubers low-complexity images which are
far from minimalist is quite repellent. The differences between these examples of lowcomplexity art demonstrate that aesthetic quality is dependent on more than just the
complexity of the artwork.
Minimalist construction is not restricted to visual art and architecture; the movement
also flourished in music. One of the simplest musical scores, John Cages 433 consists of
three parts, each with the single instruction tacet (rest). Despite the simplicity of the piece,
the perceived work can be complex since it consists of the ambient sounds present at that
Aesthetic Practice
109
time and location. Through the sounds of the environment, Cage directs perception away
from the distractions of music towards the act of perception itself, and in doing so he
radically reconfigures the possibilities of sound art. Richters grey mirrors (Figure 31)
perform the same function; their colourless space literally reflects the act of perception, like
a visual analogue of the empty music in Cages composition. 433 and Grauer Spiegel may
be perceptually complex, but both are minimal in terms of their composition we might
say that they have low algorithmic complexity. With such minimalist artworks, the audience
is expected to contribute their share towards the aesthetic experience. In this sense, though
they may be quiet (literally or figuratively speaking), minimalist artworks are rarely passive
experiences.
Some of the most stimulating work for my own visual art practice discovered during
the course of this research has come from music and sound art. In particular, this includes
music from the independent record labels Touch (UK), 12k and LINE (USA), and RasterNoton (Germany). With widely differing styles of music, these labels share a concern for
visual and tactile aesthetic quality evident in their minimalist graphic design and packaging.
The owners of these labels are active as visual artists Jon Wozencroft (Touch), Taylor
Deupree (12k) and Richard Chartier (LINE) in photography and design; Olaf Bender and
Carsten Nicolai (Raster-Noton) in design and art. Design for music packaging has an
established history, and these contemporary graphic artists acknowledge the influence of
earlier designers such as Peter Saville (Figure 32). Savilles design for the cover of Unknown
Pleasures makes it one of the few albums of the time not to feature the artists name an
exercise in minimalism that echoes Joy Divisions bleak sound and Martin Hannetts crisp,
sparse production.
Aesthetic Practice
110
Figure 32 Peter Saville and Joy Division, Unknown Pleasures (1979).
Figure 33 Alva Noto (Carsten Nicolai), Xerrox Vol.2 (2009).
Aesthetic Practice
111
Raster-Noton has a graphic style of minimal digital futurism which represents its
experiments in the overlapping border areas of pop, art and science, such as the design
of the latest release by Alva Noto (Carsten Nicolai) in Figure 33. As the second volume in a
series of five called Xerrox, its minimal design represents the letterforms in the
corresponding part of its title (ER). In contrast, Touch presents an earthy realism in
Wozencrofts photographs of the British landscape (Figure 34) used, for example, for the
environmental recordings of Chris Watson. Both labels share the minimalist concern with
honesty in materials Touch with the images and sounds of the natural world and RasterNoton with the digital. The 12k aesthetic sits between these two extremes with a
minimalist hybrid of electronic and acoustic music, incorporating techniques such as field
recording and processing of recorded instruments. Figure 35 illustrates Deuprees album
design and the graphic design of its press release. The music by Solo Andata, an Australian
duo that uses instruments and field recordings, is typical of the 12k style soft, melodic
drones built up from looped guitar samples and repeating fragments of sound.
Figure 34 Jon Wozencroft, Willow (2009).
Aesthetic Practice
112
Figure 35 Taylor Deupree, press release for Solo Andata, Solo Andata (2009).
A number of these musicians are also involved in visual art practice and research. For
example, Jon Wozencroft currently teaches at the Royal College of Art. The University of
Yorks research group New Aesthetics in Computer Music has Mark Fell amongst its staff, and
resident artists involved in the project have included Russel Haswell, Olaf Bender, Peter
Rehberg and most recently Taylor Deupree. Together with Mat Steel, Mark Fell creates
music under the name SND on the Raster-Noton label, although they released their earliest
work via the now-defunct record label Mille Plateaux, named after the publication by
Deleuze and Guattari (1980). SNDs minimal electronic music uses a restricted palette of
sounds to create shifting patterns that occupy a strange place between dance music and
abstract sound art, popular culture and academicism.
Like some of their contemporaries, SND are often labelled as minimalist, and
although it is not always a term with which they are comfortable they do share its concerns.
For example, Fell says that he is not interested in simplicity or minimalism [] for me
this point is more to do with integrity (Fell, 2009). Deupree does accept the minimalist
tag, although more in the sense of economy than simplicity (Deupree, 2009). The
compositional techniques employed by these contemporary musicians can be traced back
Aesthetic Practice
113
to the work of earlier classical minimalists. For example, the tintinnabuli (Latin for little
bell) method created by Arvo Prt, generates permutations of simple triadic chords, often
with accompanying drones. Prt used the technique to achieve simplicity in his minimalist
sacred music, such as Fr Alina (1975): The complex and many-faceted only confuses me,
and I must search for unity (Morton and Collins, 1992, p.729). Steve Reich used a similar
method of combinations of simple repeating rhythms to generate complex musical
patterns. Another technique frequently used today is based on the granular synthesis
concept of Iannis Xenakis, which transforms sound into small fragments or grains that are
processed and re-arranged to form new sonic textures. Working with micro-sounds at very
small time scales, Xenakis used stochastic processes to create new forms of music that
explore the concepts of order and disorder (Matossian, 1986, p.239).
The visual element of SNDs performance at the Transmediale festival 2009 involved
a very simple system that can be described in a couple of sentences, but [which] generates
quite complex patterns (Fell, 2009). As the two musicians operated audio controls for the
music, the changing parameters were recorded as coloured lines that built up on a screen
behind them, creating an increasingly complex stratified vision of an element of the musical
process. Fell said of the projects computer code that what we wanted from the system we
wrote in two lines it was really that simple (Rousset 2009). Robert Henke, developer of
the Ableton Live music software, and a musician under the name Monolake, says that sonic
complexity does not need lots of distinct elements, but that it can also occur with a
seemingly simple texture that reveals more detail as the listener pays more attention:
Reduction for me is a trick to guide the listeners focus to the background where the
exciting things are hidden (Henke, 2009). Henke explains that his background in film
music means that he thinks in terms of layers of sound, spatial placement and slight
variations of seemingly repetitive elements. This hidden detail, he says, is the type of
complexity he finds beautiful:
Aesthetic Practice
114
I would describe it as inherent functional complexity: It is not the starting point but
the result of applying simple processes either on a compositional level or, on a smaller
timescale, to the sound creation. Or both of course. The beauty of such a complexity
is the amazing discrepancy between the complexity of the result and the []
simplicity of the process. (Henke, 2009)
Carsten Nicolai describes a similar process in his approach to music creation: At

first you deal with total regularities. I work with measuring systems, with quite a logical
background, very simple programs to create distances and really precise sine waves
(Borthwick, 2007). As Alva Noto, Nicolai produces music with pure tones, white noise,
and sampled music. What I do is based on classical compositions and architectural ideas,
so for me it is very important to see the sound and to work on it. The influence of Benses
concept of information-aesthetics is evident in Nicolais precision sculpting of sound in
Mikro Makro (1996), a collaboration between Nicolai and musician Mika Vainio. The
installation comprised two work tables on which were a few objects lenses, magnifying
glasses and electric light bulbs. Nicolai and Vainio composed the music that accompanied
the piece, which they later released under the names Noto and respectively. One track
contains extreme low and high frequencies combined with pure sine wave tones and bursts
of white noise. Despite the obvious digital origin of the music, the auditory experience
brings to mind the sounds of bats, whales, birds and waterfalls. Another is based on noise
sourced from recordings of pulsars and magnetic resonance imaging. In accord with
Benses conception of micro-aesthetics as involving elements not directly evident to the
senses (Walther, 2000), Nicolai (2009a) says that these micro-elements point to an
existence beyond our realm of perception. In the same way that the micro-detail of visual
randomness or chaos can be perceived as an overall texture, the micro-elements of sound
put together in Nicolais music are perceived as a texture of drones, hisses and hums. These
musical practices are examples of artistic creation through the generation, transformation
and combination of simple parts, which demonstrate the minimalist engagement with
aesthetic complexity.
Aesthetic Practice
115
The Grid
A grid is a visual device associated with regularity a repetition of simple elements.
In visual art, the grid is usually presented in two spatial dimensions. A one-dimensional grid
in time provides the structure of music and moving image, a three-dimensional grid is a
possibility in sculpture, and mathematical or computer modelling can incorporate grids of
many more dimensions. In two dimensions, squares, triangles and hexagons appear to be
the most common visual elements in human artefacts. Natural grids are formed by
collections of similar objects in proximity and are most commonly hexagonal: examples
include honeycomb, cells in an onion skin, Rayleigh-Bernard convection cells in heated
fluids, and the columnar basalt rock in the Giants Causeway, Northern Ireland (originally
molten but now solidified convection cells). A recent publication by Carsten Nicolai is a
catalogue of grids for graphic design, organised as a spectrum of complexity (Figure 36).
Grid Index (Nicolai, 2009b) is presented as an
indexation system that goes from the most basic regular and uniform geometrical
tilings to the most complex and irregular ones; a growing array of shapes, vertex
transitivity, and symmetry axis present us with an index of plane subdivision
possibilities (Nicolai 2009b, no page numbering).
Figure 36 Carsten Nicolai, selection of pages from Grid Index (2009).

Aesthetic Practice
116
The content of Grid Index is derived from many sources, with a notable amount from
Grunbaum and Shephards mathematical compendium Tilings and Patterns (1987) which was
one of the first books to publish non-periodic patterns made of regular shapes. Not only
does the grid provide a tool for the graphic design of image and text, but also for fine
artists and craftspeople. The grid has been the subject of aesthetic investigation and
discussion since at least the first mosaics, up to the present discourse by Rosalind Krauss
(1985) and James Elkins (2008). Krauss theorizes the grid as a kind of Platonic form
relative to nature and art:
Insofar as its order is that of pure relationship, the grid is a way of abrogating the
claims of natural objects to have an order particular to themselves; the relationships in
the aesthetic field are shown by the grid to be in a world apart and, with respect to
natural objects, to be both prior and final. (Krauss, 1985, p.11)
In this sense the grid is also part of the minimalist aesthetic a simple, self-contained
visual structure. Prominent examples of the grid in art include work by the Russian
constructivists, Piet Mondrian, Sol LeWitt, Agnes Martin, James Hugonin, Gerhard Richter
(Figure 37), and Chuck Close.
Figure 37 Gerhard Richter, 4900 Colours: Version II (2007), enamel paint on

Aludibond, 970 970 mm.
Aesthetic Practice
117
Richters output is interesting in terms of visual complexity because his practice

includes work that covers a wide range of complexity. For example, there are works that
are simple (the grey series of paintings), regular (colour charts), representational (painted
photographic images) and chaotic (abstract multi-coloured paintings). The colour chart
series of paintings are the ones that display regular, ordered grid arrangements, but they
could also be classified as random or chaotic because there is rarely any apparent pattern in
the arrangement of colours. In a variation on the strictly regular grid, Joseph Albers painted
the series Homage to the Square. Using a limited geometrical vocabulary, Albers explores
chromatic and formal relationships to experiment with perceptual effects. Besides the
square canvas and the flat shapes, there are also perceivable squares in the lines between
areas of colour, present in the same way that a square lined grid is implied between the
colours in Richters colour chart paintings (Figure 37). With the Homage to the Square series,
Albers demonstrates how visual complexity can arise from a seemingly simple combination
of elements.
Figure 38 Joseph Albers, Homage to the Square: Yellow Resonance, (1957), oil on
hardboard, 40 40 inches.
Aesthetic Practice
118
One of the most complex grids in visual art can be seen in two sculptures by Gabriel
Orozco. The concept was to suspend a whale skeleton in the space and to draw in graphite
a three-dimensional grid onto its two-dimensional surface. In 2006, Orozco produced the
first of these Mobile Matrix, a commission for the Jos Vasconcelos Library in Mexico
City. A second version of the work inverts the pattern of the grid, resulting in the much
heavier-looking Dark Wave (Figure 39). The third dimension and the complexity of the
surface make drawing this grid a formidable task. Of its construction, Orozco said, Its so
complex that if you have a bi-dimensional graphic or drawing of a skeleton and then try to
map out what you intend to do itits impossible. (Miller, 2008).
Figure 39 Gabriel Orozco, Dark Wave (2006). Whale skeleton and graphite.
Computer Art
The first computer art exhibitions happened around the same time in Germany and
the U.S. In 1965, Bela Julesz and Michael Noll showed Computer-Generated Pictures at the
Howard Wise Gallery, New York (Dietrich, 1986), and in the same year Frieder Nake,
George Nees and A. Michael Noll exhibited their work at the Studiengalerie of Stuttgart
University (Nake, 2005). A difference between the two exhibitions is that whilst the
German exhibiters were practising artists, the two Americans were undertaking scientific
research at Bell Laboratories. Juleszs research for Bell Labs included testing the
Aesthetic Practice
119
randomness of computer-generated number sequences, a job for which he used not a

mathematical procedure, but the natural pattern-detection of the human visual system;
deviations from randomness (i.e., patches of regular patterns) could be spotted easily when
the digit sequences were visualized (Siegel, 2004, p.721).
Frieder Nake describes the computer as a universal picture generator (Dietrich,
1986, p.166). Many of the scenes we are able to perceive visually can be captured on digital
camera. The image can be stored on computer and shown on screen or printed to give a
fair representation of a scene. Images can also be generated by the computer, and the
number of potential images is enormous. The amount of possible different images with a
standard screen resolution of 1280 1024 (1,310,720) pixels, using the RGB colour system
(16,777,216 colours), is a number with almost ten million digits. To put this number into
context, the estimated minimum number of atoms in the observable universe is around
1080, which has only 81 digits. Such a vast space of visual possibilities cannot be explored
methodically because of time and computing constraints. What is needed is a way of
navigating this space towards the aesthetic hot-spots in other words, an aesthetic rule or
algorithm. This is, in effect, the function of algorithmic art like that of the early computer
artists to find rules of composition that model those incorporated into our own aesthetic
sensibilities. Just as Dodgson (2008) recreated Rileys paintings with a computer, Noll also
recreated Rileys artwork. It suggests that Riley was considering these computational
elements of information, order and randomness, though without the aid of a computer.
Nake and Nees were taught at Stuttgart by Max Bense, who, together with Abraham
Moles, influenced the direction of their work and their acceptance of the computer as a
creative tool (Gianetti, 2004). The three Ns of computer art Nake, Nees and Noll
showed their work at Jasia Reichardts groundbreaking Cybernetic Serendipity exhibition at the
Institute of Contemporary Arts, London in 1968. The exhibition and its subsequent
publication (Cybernetics, Art and Ideas, 1971) dealt with the relationship of the computer and
the arts. Figure 40 shows a picture of Bela Julesz, published with his obituary in 2003,
Aesthetic Practice
120
which was taken at his and Nolls exhibition, Computer-Generated Pictures. It depicts Julesz in
front of one of the artworks, which is unmistakably Attneaves (1959b) stochastic
composition process (Figure 25). The evidence in the picture suggests that the early
computer artists were aware of the current scientific research. It seems also that Attneave
was equally attuned to the arts, since he had foreseen creative potential in the computer for
manipulating aesthetic rules rather than just aesthetic properties:
In some respects more appealing than either the imitative or the autonomous
computer would be one which allowed the rules of composition to be varied
easily and precisely. [] Such a device would offer extraordinary opportunities for
artistic exploration and experimentation. (Attneave, 1959b, p.510)
Attneave also noted that the benefits of this tool would throw a correspondingly
greater burden upon his evaluative or critical capacities. Moreover, the connection
between computer and empirical aesthetics also suggests that we could re-interpret the
scientific graphics as a pathfinder in the development of visual culture. James Elkins (2008)
sees this kind of value in certain types of scientific visualization, particularly those that are
difficult to achieve and which push the boundaries of visualization. To Elkins, this
difficulty in making and the resulting improbability of the images lend them aesthetic value.
Figure 40 Bela Julesz in 1965, with Attneaves (1959b) Stochastic Composition Process.
Aesthetic Practice
121
Fractal Art
A fractal artwork is one which has a pattern within a pattern, which is nested or selfsimilar, like the examples illustrated in Chapter 1 (Figures 24). These types of pattern can
be constructed in a variety of ways, which generally fall into one of two categories: topdown or bottom-up. For example, with a cellular automaton a fractal image is built up
from the bottom. A pattern emerges from small picture elements (cells/pixels), which form
the base level of detail in the image, as the image is generated row by row somewhat like
knitting. Alternatively, a similar image can be formed top-down by an iterative process of
subdivision. Figure 41 shows a fractal image created with these two different methods:
Figure 41 Fractal construction methods.
The upper image is created bottom-up from 64 steps of elementary CA rule 60,
starting from initial conditions of a single black cell. The lower series of images is made by
a top-down iterative process of copying the image, reducing in size, and pasting the copies
onto three quarters of the original image. Both the large CA image and the last image in the
Aesthetic Practice
122
series below it have the same amount of detail. Each is effectively based on a 6464 grid
(the grids represented here differ for clarity), and each contains six levels of nesting.
Mandelbrot (2004a) cites Hokusai as one of the earliest visual artists to create images of
fractals, in pictures such as The Great Wave (Figure 4). A contemporary example of natural
fractals in art is the tree-like sculpture of Jorge Mayet (Figure 42) which was presumably
made with a top-down method, starting with the central trunk and adding successively
smaller branches and roots.
Figure 42 Jorge Mayet, De Mis Vivos y Mis Muertos (2008). Electrical wire, paper,
acrylics, fabric; 143 84 84cm.
Aesthetic Practice
123
Taylor, Micolich and Jonas (2002) demonstrate that Jackson Pollocks drip paintings
(Figure 43) are fractal, by using a fractal measuring technique known as box-counting.
Taylors research found that the fractal dimension of the drip paintings increase over the
course of their construction, as one would expect in a top-down fractal construction, which
becomes more intricate as layers of paint are built up. The research also shows that during
the years of their execution, the drip paintings became increasingly complex. Over the
decade from 1943 to 1952, the fractal dimension of the paintings increased from around
1.1 to 1.7 (Taylor, Micolich and Jonas, 2002, p.206).
Figure 43 Jackson Pollock, Lavender Mist: Number 1 (1950), oil on canvas, 221 x 300cm.
These findings suggest that the fractal nature of Pollocks work was no accident, and
that even though the term fractal had yet to be coined and the images reach a wider public
awareness, the creation of these self-similar forms was a significant component of Pollocks
art. It begs the question why this was not recognised before this computational analysis. It
seems that the irregularity in the dripped paint masks their fractal qualities from our
perception, and the high fractal dimensions of the later work in particular makes it difficult
Aesthetic Practice
124
to perceive their inherent order. A possible explanation for this is the idea of a perceptual
threshold of visual complexity. The objective or informational complexity of the drip
paintings is just beyond the threshold, past the peak of the inverted U correlation with
perceived complexity. Their randomness makes it difficult to perceive the inherent order,
and together they contribute to the perception of visual complexity. There is something
aesthetically pleasing about the irregular fractals in Pollocks configuration of order and
chaos, just as there is in the shapes generated by Scott Draves Fractal Flames (Figure 44).
The generating algorithm is an iterated function system, and Draves contribution is to
have formulated a new colouring method that reveals the fine structure of these forms
(Draves and Reckase, 2007). By manipulating its parameters over time, it is possible to
animate the fractals. These animations form the basis of Draves Electric Sheep screensaver
project (Draves 2009), in which users can influence the selection and generation of sheep
or create their own and subject them to the same selection and generation process.
Figure 44 Scott Draves, aka Spot (2007). Two images generated by the Fractal Flames
algorithm for the Electric Sheep project.
Aesthetic Practice
125
Visualizing Complex Information

Unlike fine art, graphic design involves the manipulation of visual elements for a
practical purpose it is about more than just making pleasing visual forms, and so its
aesthetic is always secondary to its utilitarian demands. The purpose of such graphics is to
convey useful information. Edward Tufte formulates principles of design for information
graphics by analysing previous instances and from his own experience in design. Tufte
argues for making the most of established print and paper technology in place of poorerresolution electronic displays. One of the most cited and celebrated graphic designs in
Tuftes book, The Visual Display of Quantitative Information (2001), is a map of Napoleons
Russian campaign by Charles Joseph Minard (Figure 45). In only two dimensions on paper,
this chart elegantly and efficiently represents at least six dimensions of statistical data,
including the size of the army, its geographical position, elevation, and the local
temperature.
Figure 45 Charles Joseph Minard (1869). Chart depicting Napoleons 1812 Russian
campaign. Lithograph, 62 x 30 cm.
Tufte argues convincingly that by using low-quality displays (low-resolution, poor

colour reproduction, bad tone/contrast) and formulaic designs (such as the preset styles in
Microsoft PowerPoint), we deprive ourselves of a rich visual experience that can be both
illuminating and aesthetically pleasing at the same time. In fact, Tuftes guidelines barely
Aesthetic Practice
126
make the distinction between good form and good function, so closely are they tied
together in his principles. Above all, Tuftes graphics are economical there is no visual
redundancy, thus the data represented are about as compressed as is possible in graphical
form. Experiments by Donderi and McFadden (2004) provided empirical evidence that
more information-dense images can be easier to process: When university students and
professional mariners were tasked with reading naval radar and chart images, both groups
showed improved efficiency in answering questions about the displays when the images
were superimposed instead of presented separately. Donderi agrees that Tuftes aesthetic
parallels the computational approach: the information-theory context of Tuftes
analysis was implicit in the terms he used and the principles he advocated (2006b, p.83).
Like Tufte, Colin Ware argues for the visualisation of information, but whereas Tufte
wields authority by his demonstrable graphical skill and aesthetic sensitivity, Ware utilises
the body of psychological research on visual perception to strengthen his argument. Ware
et al. (2006) produce novel methods of visualizing marine data, in this case (Figure 46), a
visualisation of humpback whale movements.
Figure 46 Visualization of whale behaviour by Ware et al. (2006).
Aesthetic Practice
127
In his book on information visualization, Ware (2004) not only gives a

comprehensive account of the visual system and perceptual psychology, but also lists some
of the psychophysical principles and methods that are used in the work he cites and in this
research project. It is evident in his writing that Ware advocates a theory of active
perception the orthodox view that largely rejects Gibsons theory of direct perception:
At higher levels of processing, perception and cognition are closely interrelated, which
is the reason why the words understanding and seeing are synonymous. (Ware
2004, p. xxi)
Ware employs David Marrs computational approach to vision to justify a set of

principles for information graphic design. In particular, Ware uses Marrs idea of the 2D
sketch to implement better graphic design. The perception of the third dimension
proximity to the observer does not conform exactly to the line of sight; we experience
this dimension in degrees, he argues. Images of occluding shapes, for example, are strictly
two-dimensional, but are perceived as having some kind of depth. Joseph Albers painting
Homage to the Square (Figure 38) demonstrates this effect; it appears to be composed of
many squares superimposed on each other, when in fact there is only one square on the
face of the canvas (the other squares are actually shapes like an empty picture frame). The
perception of depth depends on certain visual cues, and graphic techniques such as
occlusion and shading can strengthen or lessen the perceptual effect.
Ben Fry explores ways of visualising large sets of complex data, and like Ware
acknowledges the perceptual problems this brings. For his PhD thesis, Fry (2004) worked
with one of the largest and most complex data sets the human genome. His statement of
the graphic design problem is that The ability to collect, store, and manage data is
increasing quickly, but our ability to understand it remains constant. (2004). Fry (1997)
explores methods of navigating data sets. His challenge is to come up with ways of
visualizing the kind of complex and dynamic data sets that fail to be served by Tuftes
guidelines for more traditional graphic data. This kind of data changes over time, and
Aesthetic Practice
128
constitutes a new type of complexity to which the visualisations have to adapt. Another
project that visualizes complex dynamic data is Thinking Machine (Figure 47) by Martin
Wattenberg. Thinking Machine visualizes the moves considered by a computer chess
program lines of thought are traced in coloured arcs that reveals the electronic
opponents strategy, the more likely moves being traced more times. The tactical strength
of each side is visualized in ripples that propagate from the pieces on the squares over
which they have control.
Figure 47 Martin Wattenberg, Thinking Machine (2002).
The website Visual Complexity (2009) is a resource for the visualization of complex
networks, which includes the work of Fry and Wattenberg amongst topics such as art,
biology, social networks and the World Wide Web. It is run by designer Manuel de Lima,
whose aim is to provide a resource for the visualization of complex networks and to
develop a critical understanding of visualization methods between various disciplines. After
knowledge networks and social networks, art is the third most popular subject for
visualizations on the site. One example is the work of Lee Byron, who wrote the
StreamGraph software (Byron and Wattenberg, 2008). The software visualizes information
Aesthetic Practice
129
gathered through the LastFM music website which records patterns of music listening
history (Figure 48). The horizontal axis represents time in weekly snapshots, and each artist
is represented by a labelled, coloured stripe, the thickness of which is determined by the
number of tracks played.
Figure 48 Visualization from LastGraph of my listening history (one year and detail).
What is interesting and valuable about the work of Tufte, Ware and Fry is their
willingness to exploit our considerable powers of perception rather than pander to illinformed principles of simplicity and ease of understanding. That is not to say that these
researchers promote visual complexity for its own sake, or that they aim to produce
material difficult to digest, but that in different ways they attempt to enrich our world of
visual information and reveal its complexity through clarification rather than obfuscation.
Ware describes information visualization as external cognition, a tool for augmenting the
cognitive powers of the mind (2004, p.xvii). If graphic visualization is externalized
cognition, then to perceive and understand information graphics is to re-internalize and
incorporate visual complexity, and the task of the information graphic designer is as Tufte
describes elegantly:
Aesthetic Practice
130
What is to be sought in designs for the display of information is the clear portrayal of
complexity. Not the complication of the simple; rather, the task of the designer is to
give visual access to the subtle and the difficult that is, the revelation of the complex.
(Tufte 2001, p.191, my emphasis)
These principles are followed throughout the thesis, since they seem appropriate for
both the art and design context of the investigation and for the complex quantitative nature
of the data that is collected and analysed. Benoit Mandelbrot wrote of the illustrations in
his influential book, The Fractal Geometry of Nature, that they were
designed to help make its contents accessible in various degrees to a wide range of
readers, and to try and convince even the purest among mathematicians that the
understanding of known concepts and the search for new concepts and conjectures
are both helped by fine graphics. However, showing pretty pictures is not the main purpose in
this Essay; they are an essential tool, but only a tool. (Mandelbrot, 1982, pp.2122)
The same can be said of my intentions with this thesis; the illustrations are an
essential tool for the description of the materials, methods and results of this study. In a
review of the use of graphics in psychological journals, Best, Smith and Stubbs (2001) find
that graphical presentation of data predominates in the hard sciences: On average, 1 out
of 10 pages contain graphs in journals of the hardest disciplines (e.g. physics, chemistry),
and only 1 in 100 for the softer sciences (such as psychology), with the number of tables
used in inverse frequency to graphs. Best et al. conclude that
If the softer sciences are to participate in the emergence of another golden age of
graphics, its practitioners will do well to cultivate visual thinking in their research
circles, take advantage of novel technologies in computing and graphic design, and
even contribute actively to the evolution of graphical methods. (2001, p.164)
The cultivation of visual thinking may well increase the respectability of the softer
disciplines in the scientific domain, but it seems plausible that it might also increase the
accessibility of this literature to non-scientific areas. This applies particularly for the
visually-literate cultures of art and design, which often comprise the objects of study for the
soft psychological research of empirical aesthetics.
Aesthetic Practice
131
Audio-Visual Art
In this section, we look at art with potential for great complexity that combines visual
material with sound and music. The first time I encountered Scott Draves fractal
animations they were projected alongside dance music. The movement of the animations
seemed to follow the changes in the music, but this was a perceptual illusion because there
was no physical link between the audio and the visual. The perception of disparate sounds
and visuals working together aesthetically has been known as the Cage effect, named after
John Cage 8, and is called synchresis by the film theorist Michel Chion (1994). Chion devised
the term from a combination of synchronised and synthesis, and he documented the use
of the effect in film, where the conjunction of sight and sound can be used to direct the
attention of the viewer. In psychology, this effect is classified as a form of multi-modal
perception, but Chions synchresis is not the same thing as the more well-known
psychological phenomenon synaesthesia. Whereas synaesthesia involves a confusion or
translation of sensory modes (such as tasting the colour blue), synchresis takes the form of
a multi-modal enrichment. We can use these two perceptual phenomena to consider audiovisual art by identifying to what extent an audio-visual artwork involves synchresis or
synaesthesia in its construction and its perception.
In digital art, it is relatively easy to perform a synaesthetic transformation from one
medium to another, since digital information is essentially of the same kind, whether
encoded as an image file or as a piece of music. A synaesthetic process forms the basis of
Ryoji Ikedas music and art installation Test Pattern (2008). Ikedas artwork involves a
system that converts any type of data (text, sounds, photos and movies) into barcode
patterns and binary patterns of 0s and 1s (Ikeda, 2009). The result is a dense field of
inhuman sound (the extreme nature of which may potentially damage listening equipment,
according to a warning sticker on the CD!) which Ikeda breaks down and weaves into
My research has yet to identify a reference to this effect in Cages own writing. Usage of the phrase Cage
effect appears to be restricted to the field of amateur computer graphics.

Aesthetic Practice
132
detailed rhythms. Ikeda (2009) says the aim is to examine the relationship between the
critical points of device performance and the threshold of human perception. Jack Ox
employs synaesthetic transformation of image and sound with her Twenty-First Century Color
Organ (2002). Oxs system uses MIDI music files to create immersive performances of
sound and 3D images. Various musical parameters are treated as data and the system
creates a corresponding data set in a visual vocabulary based on Oxs landscape drawings
and colour charts. If the system could work with live data from improvising musicians, as
intended, then the piece may function as a complex adaptive system (Ox and Britton,
2000, p.9). Another approach to visualizing sound is that of Martin Wattenberg (Figure 49),
whose The Shape of Song (Wattenberg, 2002) also works on MIDI music files. The system is
quite simple repeating patterns are joined by coloured arcs but the design is pleasing
and informative. The approach is comparable to the computational analysis of music
complexity by Wilhelm Fucks (Figure 26) which also relies on pattern-recognition. Instead
of aiming for a more interpretive translation of music in projects such as Oxs Colour Organ,
The Shape of Song sits somewhere between analytical and creative approaches to aesthetics.
Figure 49 Martin Wattenberg, The Shape of Song (2001). Visualizations of Koyaanisqatsi

by Phillip Glass (left) and Moonlight Sonata by Beethoven (right).
Aesthetic Practice
133
Evelina Domnitch and Dmitry Gelfand worked with scientists and musicians in the
project Camera Lucida (2004), which exploits the phenomenon of sonoluminescence: Highvolume ultrasound waves are propagated in a liquid, which induces the formation of
microscopic vacuums known as cavitation bubbles. As the bubbles rapidly collapse,
temperatures as high as that of the Sun are reached, and light is emitted. In this case, the
liquid is 97% sulphuric acid doped with xenon gas, and the apparatus produces ghostly blue
light that dances and flickers around within the tank. Domnitch and Gelfand collaborated
with musicians on the record label LINE (a subsidiary of 12k), who used hydrophonic
recordings to create music for the video. These musicians, including Taylor Deupree and
Alva Noto (Carsten Nicolai), translated the ordinarily inaudible frequencies of the
ultrasound and the bubble implosions into perceptible ranges of sound, creating
synchronised musical soundtracks for the video sequences. The relationship of sound and
image is one of cause and effect, so it is difficult to say whether its construction could be
described as a synaesthetic transformation, but the perceived effect is certainly one of
synchresis in which the two sensory modes appear unrelated but enhance each other.
Figure 50 Domnitch and Gelfand (2004) Camera Lucida. Sulphuric acid and xenon gas
in glass tank, with ultra-high frequency sound. Four images extracted from video.
Aesthetic Practice
134
Under the name Semiconductor, Ruth Jarman and Joe Gerhardt specialise in working
with digital animation to transcend the constraints of time, scale and natural forces
(Semiconductor, 2009). Magnetic Movie (2007) uses recordings of VLF radio (naturallyoccurring electromagnetic signals from the earths ionosphere and magnetosphere) to
accompany an animated visualisation of magnetic fields. Some of the movements are
animated by the sounds, producing an enchanting effect of synchresis. For example, the
white lines in Figure 51 grow in spasmodic bursts in proportion to the loudness of each
audible event. On Semiconductors website, Douglas Kahn describes the film:
What we hear is underscored with complex and supple orders, in fact, too complex
and supple to be ordered. We already have experience of them in the tangible
turbulence of water and the crazy convection of fluids combining, tongues of fire and
the thermal afterthought of smoke, the ribbons of clouds stiffly blown twisted up a
hill. (Kahn, 2007)
What Kahn is describing, in so many words, are chaotic systems. Turbulence in

fluids, for example, is what was originally represented in the dynamics of the Lorenz
attractor now an archetypal image of chaos. The earths magnetosphere is also turbulent
because its molten iron core is in flux and because it is affected by fluctuating radiation
emitted from the sun. The sound of this magnetic turbulence, recorded by Stephen P.
McGreevy, provides a naturally chaotic soundtrack for the film. That the animation is not a
direct synaesthetic visualisation of any particular field pattern is announced with the
opening words of the film, when the scientist Janet Luhman says, Magnetic fields, by their
nature, are invisible. Nevertheless, the way in which the sounds control the visual
elements makes for a viewing experience that invites the suspension of disbelief. Without
knowing the technical details of the animation, it is impossible to know exactly how much
of the experience is due to the effect of synchresis and how much this audio-visual
interaction is designed in its making, and this is part of what makes it fascinating.
Aesthetic Practice
135
Figure 51 Semiconductor, Magnetic Movie (2007), digital animation.
We end this discussion of contemporary art practice with two examples of audiovisual art, whose minimal simplicity brings us full circle back to where we started. Broadway
One by Ernest Edmonds and Mark Fell generates flat fields of hot colours with occasional
blips and clicks as the image suddenly changes. Apart from their synchronicity, the
relationship between image and sound is unclear. The work has a delicate balance and
intriguing quietness that draws attention and questions what we are actually perceiving.
Similarly, Colour Projections by Theo Burt is equally spare in its visuals, containing outlines of
simple geometric forms that rotate slowly accompanied by slowly-changing tones (Figure
52). Like Broadway One, its synchronisation of image and sound suggests an underlying
mechanism between the two, although the rules behind this relationship are not readily
apparent. Nevertheless, I managed to learn more about the mechanism by analysing the
sound, which revealed that the audio waveforms have a similar shape to the geometric
graphics. At the point of the animation shown in Figure 52, we can hear a low tone
corresponding to the large circle and a higher tone which appears as the smaller circle
begins to cut into the larger one. The sound analysis shows that the waveform has a related
shape which introduces more harmonic frequencies as the waveform is altered from a pure
Aesthetic Practice
136
sine wave. Thus the complexity of the graphics corresponds with the complexity of the
waveform, resulting in a matching perceived complexity of image and sound.
Figure 52 Top: Theo Burt, Colour Projections (2009), digital animation. Bottom:
visualization of the audio waveform taken at the corresponding time.
Rather than translating from one medium to another, Colour Projections and Broadway
One generate image and sound from internal computations as separate but related
instrumental expressions. Although Edmonds and Fell describe their artwork as
synaesthesic (2005), it could be argued that the process is more accurately characterised
as synchresis, with an integration of image and sound in terms of both its production and
its perception. Fell says of the work, The emphasis is on correlations that are purely
aesthetic. There is no innate or mathematical relationship between sound and colour.
Anything one does is purely invented (Whitelaw, 2007). The point is that any digital
synaesthetic translation is ultimately arbitrary, with the implication that the result actually
reveals more about the system and its creator than it does about the nature of the stuff that
is translated from one medium to another. This notion is supported by Moles (1966)
description of aesthetic information as being untranslatable: As opposed to semantic
Aesthetic Practice
137
information, which is based on a shared symbolic language, aesthetic information refers to

the repertoire of knowledge common to the particular transmitter and particular receptor
(Moles, 1966, p.129). The fact that any correspondence between media is ultimately
arbitrary undermines the idea of synaesthesia as a process capable of meaningful
translation.
What we are left with and what really matters is the perceptual effect of the
artwork and the quality of the aesthetic experience. It seems more difficult to create the
synchresis effect, but it may offer a more rewarding aesthetic experience. The most
successful of the audio-visual artworks, to my mind, are the ones that achieve aesthetic
synchresis, whether through an explicit connection or a more arbitrary translation process.
In either case the quality of the experience appears to be dependent on the synchresis
effect the mind has to work a little to make the connection, but the effort pays off in
aesthetic satisfaction. It is possible that the propensity for images to work with sounds in a
synchretic experience is sensitive to their levels of complexity. In my own experience, the
most rewarding aesthetic experiences of synchresis, and thus the most successful synchretic
artworks, appear to have similar levels of complexity in image and sound, either at a static
level or via synchronised changes. This is a potential subject for further investigation.
My Art Practice
My practice has contributed to this research by generating visual material for use as
test stimuli and by developing the skills and tools needed to manipulate complex imagery
using computer-based programs. Figure 53 shows the last large-scale hand-made piece,
which was produced for a friend and ex-colleague in return for lifts to work. The drawing
is a representation of the rule number 110 elementary CA on an A1 sheet of 2mm square
graph paper. The initial conditions (the top row) were produced with the rule 30 random
number generator in Mathematica and the rest of the CA was calculated and drawn by hand,
which meant that the resulting pattern was unforeseeable at the start. By forgoing the
computer I discovered a way of reducing the calculation of rule 110 without having to
Aesthetic Practice
138
consider each cell in turn, which speeded up the process considerably (around three or four
times faster). Nevertheless, I calculated that it took roughly the same amount of time to
make the drawing as I had spent in my colleagues car on the journey to work and back
over a year around a hundred hours. The top left-hand image in Figure 53 is
approximately life-size.
Figure 53 Untitled (2006), ink on paper (detail; in progress; finished).
Artwork produced before the start of this research was mainly based on craft
techniques combined with computational processes such as cellular automata. In 2005 I
attended the 5th. Creativity & Cognition conference in which I exhibited two paintings and
one needlepoint piece, all based on a simple kind of block CA. I had the chance to talk
Aesthetic Practice
139
with Frieder Nake, initially about a line from Thomas Pynchons novel Vineland, which I
was reading at the time:
If patterns of ones and zeroes were like patterns of human lives and deaths, if
everything about an individual could be represented in a computer record by a long
string of ones and zeroes, then what kind of creature could be represented by a long
string of lives and deaths? (Pynchon, 1998, p.90)
In conversation, Frieder said that he had learnt to not get annoyed when artists try to
use computers to do things like recreating natural scenes and images, which is so difficult
to achieve digitally that the results are often poor representations of what we usually
perceive. Computer graphics of natural forms and textures, for example, may seem
incongruous when juxtaposed with photographic images and they often date rather quickly.
Frieder saw more aesthetic value in work that makes use of the computers digital nature,
rather than work that fights against its visual characteristics. Use the computer for what it
gives and what it is good for; work with it, not against it was his message. This is exactly
the point of Frieders own work, and with these comments his aesthetic of simple visual
elements is seen in a new light that makes it seem less ascetic and more open. In one sense,
it is an approach similar to that of the minimalists a truth to materials. Using the
computer to recreate natural scenes is thus more artificial than making use of its own visual
language. Frieders conversation had a profound effect on my own practice, giving a new
direction to the production of computational artefacts. It did not negate my view that in
general the hand-made offered a more complex and appealing aesthetic than the digital, but
it showed a way in which the inherent (and sometimes off-putting) visual character of
digital art could be turned on itself to produce a more honest kind of artwork. Now, the
digital could be used without regret or apology for its visual nature.
A defining feature of digital art is that it is based on fundamental discrete units
numerical digits that are represented internally as binary magnetic or electrical charges and
externally as dots in print or pixels on screen. If an artist is trying to recreate a natural visual
scene, then these pixels should be imperceptible and their tell-tale characteristics are
Aesthetic Practice
140
avoided (such as aliasing the appearance of discontinuities in what should be smooth

lines or gradients). But if one accepts this characteristic, as Frieder Nake suggested, then
the idiomorphism of digits and pixels becomes a creative tool and a visual palette. The
result on my practice was to explore these features, principally by working with raster
(pixel-based) images, manipulating and building images a pixel at a time. Figure 54, for
example, is an unashamedly digital image which explores the perception of threedimensional form as represented in two dimensions and the impossible figures that can be
created between these modes of representation.
Figure 54 Iso C (2006), digital image.
To exploit the properties of digital media as Frieder Nake had described, I began to
use the kind of software that has the opposite aim of recreating natural effects and realistic
images. Ray-tracing is one of the techniques of computer graphics used in this regard,
based on the modelling of light and its interaction with surfaces and translucent media. The
path of a light particle is modelled from source to surface to eye, enabling the recreation of
Aesthetic Practice
141
realistic three-dimensional illumination in two dimensions. This modelling is fairly

complicated for a single ray of light, but to generate an entire image requires many such
rays to be traced, which requires a lot of processing power and time. The default settings
for many ray-tracing programs is to model the physical behaviour of light, but these rules
can be manipulated and the resulting images modified to create novel visual effects that
would be impossible in the real world. For instance, the amount of light reflected and
absorbed by a surface normally adds up to the amount transmitted to it, but by tweaking
the rules we can produce effects whereby the amount of light reflected exceeds these limits
(Figure 55). For me, this was using the computer as Nake described, freeing it from the
attachment to our normal perceptual world and opening it up to aesthetic exploration. The
same can be said for Semiconductors use of digital animation to escape the limitations of
the visible physical world. In essence, the process is analogous to the inventive use of
photographic technology in Man Rays rayographs and Marco Breuers photographs
made without a camera.
Figure 55 P32b inv greens (2007), digital image.

Aesthetic Practice
142
In my previous job as a biology technician, one of my responsibilities was preparing

visual material for teaching. This included tasks such as preparing projections from a videoequipped microscope as well as creating slide presentations for teachers or printed
handouts for students. Amongst the visual resources I used was a collection of highresolution scanned images from Ernst Haeckels Kunstformen der Natur, which was originally
published between 1899 and 1904 in volumes of ten prints. Haeckels illustrations show the
marvellous diversity and commonality of forms in the natural world, grouping images of
parts or whole organisms in symmetrical arrangements on the page to illustrate a particular
morphological theme. I obtained permission to use the images from Prof. Kurt Stber,
who had scanned an original copy of Haeckels prints and hosted them on his website
(Stber, 1999). I used these digital images in combination with cellular automata patterns
produced in Mathematica, firstly by overlaying the CA patterns on the Haeckel images, and
later by using the CA pattern to invert areas of the image beneath. Figure 56 shows a detail
of this process, in which the individual pixels are visible. Elementary CA rule 110 has been
used as the overlaying pattern, and its triangular shapes are the areas in which the Haeckel
image has been inverted (white becomes black, red becomes green, etc.). Figure 57 shows
an entire image made using this process, a copy of which is used in Test 3. These digital
prints were produced such that the individual pixels remained visible. Ordinarily, this
would have made the original images quite unpleasant as compression artefacts could be
revealed, but the added complexity of the CA pattern obscured this. To my mind, this
process produced some of the most complex images, which were just on the right side of
the edge of chaos and the threshold of perceivable visual complexity. The overall effect of
this technique is to reduce the contrast and colour saturation at a distance, whilst a closer
inspection reveals diverse tones and adjacent complementary colours. Because the colourinverted areas occupy just less than half the total area of the eCA rule 110 images, the
resulting inverted image still retains some of its original hues and tones when seen at a
distance (conversely, inverting the other area of the pattern produces the appearance of a
negative).
Aesthetic Practice
143
Figure 56 Tafel_008_300 + 110 D3 (2007), digital image (detail).
Figure 57 Tafel_008_300 + 110 D3 (2007), digital image.
In 2007, I entered a competition for a commission for the Minster School in

Southwell, Nottinghamshire. The brief was to design and produce translucent sliding doors
and a skylight for the new school building. The doors were to separate the main hall from a
small chapel, and the skylight was for the chapels ceiling. My entry was based on a block
cellular automata pattern in a spectrum of colours. The underlying pattern is based on a
rectangular repeating black and white pattern, but the colouring scheme I developed made
it difficult to perceive its underlying simplicity. The image used in the final design (Figure
58) repeats horizontally, such that the left edge would match up with the right if put
together (if the doors ran past each other and swapped places). I was quite pleased with the
design, but unfortunately it was not chosen for production.
Aesthetic Practice
144
Figure 58 Architectural design of translucent sliding doors for Southwell Minster

School commission (2007), digital image.
Shortly after the Southwell competition, I accepted a commission to create wall

designs for the refurbishment of the Deutsche Bank headquarters in London. A friend and
former colleague had contacted me as he was leading the team of architects that were doing
the work. The job had to be done with very short notice, and so I used some of the
patterns I had been working on for the Southwell commission. In this brief, the colours
were specified by the client. I produced about twenty design variations within a couple of
days, and the client selected two of these for production (Figure 59). The designs were for
decorative tiled walls within four employee shower blocks. A total of twelve prints were
produced on adhesive vinyl, which were adhesive on the printed side and mounted onto
glass panels fixed to the walls to protect them from moisture and abrasion (Figure 60). I
think the low complexity of the images work well with the relatively plain surroundings,
and the pattern echoes the shape of water dripping down glass and puddles on the floor.
Aesthetic Practice
145
Figure 59 Architectural designs for the Deutsche Bank shower blocks.
Figure 60 Photograph of installed wall decorations, Deutsche Bank, London.

Aesthetic Practice
146
Summary of Context
Discussions of aesthetic problems can be traced back to Plato and Aristotle, when
the ancient Greek word aestheta referred to things perceptible by the senses. When
Baumgarten coined the word aesthetics in the eighteenth century, it meant a science of
sensory perception; however, the term gained more currency in the art world, as ideas of
beauty were replaced by theories of taste. The scientific study of aesthetic perception did
not begin until a hundred years after Baumgarten defined the term, mainly due to the
difficulty of measuring perceptions. This field is known as psychophysics, so named
because its aim is to establish relationships between processes of perception and events in
the physical world.
What were previously regarded as scientifically unsurpassable divisions between the
mental and the material are now providing fertile grounds for research. Empirical studies of
the arts fall into this category, and are currently experiencing a burst of growth. However,
objections that the complex nature of creative cognition cannot or should not be examined
empirically are still heard in both the artistic and scientific communities: Art and its
experience is complex, but scientific study is selective and its results are limited. In
consequence, much of the scientific study of art has concentrated on solving simple,
manageable problems, or has been content to make only modest claims.
Perhaps surprisingly, even studies of visual complexity have mostly concentrated on
relatively simple geometric images, and studies that use actual works of art in empirical
research are scarce. It is almost as if real artwork is too complex to be examined empirically.
The lack of engagement with the complexity of visual art provides the context for the
current enquiry. The empirical part of this research project aims to counter this trend by
conducting tests of visual complexity using real artwork, and to answer the question: How
complex is contemporary visual art, in terms of objective properties and aesthetic
perception, and how does it relate to aesthetic value?
Summary of Context
147
Emergent themes include the various ways of approaching visual complexity in terms
of understanding, visualizing, or analyzing. Even within the relatively small field of
empirical aesthetics, the topic of visual complexity has been approached from a number of
disciplines, including art practice, art theory, perceptual and cognitive psychology, complex
systems theory, and computer graphics. These studies generally rely on conceptual
frameworks from either information theory and/or psychology, depending on the extent to
which the focus is on objective or subjective complexity. Although there is evidence of the
influence of scientific research in the arts, there is very little flow in the opposite direction,
which is perhaps explained by the move away from perception of contemporary art theory
and its rejection of empirical aesthetics (e.g. Dickie, 1961). In one sense, psychology is
central to any scientific enterprise because everything we know is based in the mind.
Similarly, the philosophy of the mind stands at the centre of all philosophy. In addition,
since everything we know comes to us via the senses, we ultimately arrive at perceptual
psychology in asking questions about aesthetics. The difference between psychology and
the philosophy of mind is that the former deals with specific scientific questions, and the
latter with a more general kind of question which cannot be answered empirically.
Empirical studies of aesthetics can be divided into two categories: What might be
called the analytic (or elementary) approach employs real artworks as stimuli, whereas
synthetic (or holistic) approaches use abstractions of artworks such as simple geometric
diagrams. Each comes with a drawback: The synthetic approach aims to circumvent the
practical difficulties in experimentally measuring or controlling the properties of real
artworks as stimuli, but this move simultaneously weakens the validity of generalizations
that can be made about actual artworks and perceptions of their properties.
The problem with objective measures of artworks is twofold, and is not a fault of the
measures themselves: The difficulty lies in the fact that artworks are likely to be more
visually complex than the geometric figures of psychological tests, and that it requires a
complex system (whether a method, mechanism or organism) to recognise this visual
Summary of Context
148
complexity. In theory, then, the accuracy of an aesthetic measure can be improved by more
closely imitating the functioning of our visual system (by taking into account the same
visual properties that we perceive), but the greater intricacy of the measure makes it more
difficult to apply in practice. Another problem is that the range of visual complexity in
many experiments has been limited to rather narrow bands, such that it is difficult to
envision the response across the entire spectrum.
Berlyne (1971) identified visual complexity as one of the collative variables, together
with novelty, ambiguity, uncertainty and surprise. These are the aesthetic properties that
involve comparison. As such, they cannot be measured as directly as the psychophysical
variables like brightness or loudness that correspond to simple physical phenomena,
because comparisons require something capable of comparing, whether it is an analytic
method or a sentient being, and which therefore influences the results of the measurement
because it must select which features (of those that are perceptible to it) to include in the
comparison. As aesthetic properties, therefore, the collative variables are a function of
perception and/or cognition, and not solely properties of a thing itself. Due to this
context-dependency, there can be no truly objective aesthetic measure in the sense of it
being more true than any other 9. However, that is not to say that a particular metric
cannot be objective insofar it can be replicated and verified, nor that one measure may be
more appropriate or more useful than another.
Measuring only subjective perceptions of aesthetic properties can tell us much about
individual perceptions and individual works of art, but it is difficult to compare and infer
from these studies due to the variability and specificity of their results. An objective
measure, on the other hand, can be used to construct a scale to which subjective data can
be compared. If we can determine the (mathematical) relationship between the two scales,
we can use the objective method to predict perceptions of other measured artworks.
9
A similar context-dependency also appears in many definitions of complexity, a feature that sets complexity
theory apart from other areas of scientific study.

Summary of Context
149
The experiments of Berlyne (1971) and Martindale, Moore and Borkum (1990) fall
into the category of the synthetic approach when they used geometric stimuli such as
random polygons. These methods are not applicable to artworks because objective visual
complexity is measured as a function of the number of turns in a polygon, and it causes
potential difficulties. Martindale et al. showed this to correlate poorly with subjective ratings
of complexity, even for these polygons. We could employ a more elaborate definition of
objective complexity, which would take account of such variables as the number, regularity,
homogeneity, density, and redundancy of picture elements, but this would make for an
even more unwieldy tool or measurement, even if it could be automated. Aesthetic
measures devised by George Birkhoff (1932) and Hans Eysenck (1941b) took just such an
approach to arrive at conflicting theories. Birkhoffs (1932) aesthetic measure was a
product of order and complexity: M = O C, whereas Eysenck (1971) proposed a
function of order divided by complexity: M = O / C. Both these measures have been
shown to be too simplistic, and their methods of enumerating properties, although
reasonable, are too complicated to use on real artworks (not least because of the ambiguity
of even the formal let alone semantic content of a picture).
An information-based measure of visual complexity, such as that of Donderi and
McFadden (2004), provides a more plausible approximation to perceived complexity, but
with knowledge from complex systems (Gell-Mann, Wolfram, Langton) we now
understand why this too has its limitations. Measures such as AIC deviate from our
intuitive notion of complexity, but they are useful in illuminating the relationship between
objective and subjective aspects of complexity. Further, whereas some measures are too
elaborate to cope with the variety and complexity of all but simple geometric figures, image
file compression provides a practicable method of measuring the visual complexity of a
wide range of visual material including images of contemporary artworks. It constitutes
perhaps the most practicable objective measure of visual complexity because it requires no
special computing capabilities or software beyond those of an ordinary PC, and takes
Summary of Context
150
almost no time to determine, unlike the fractal measures which involve a more detailed
computational analysis. Another reason for using this method is that it is more widely
applicable than fractal measures, such as that of Taylor, Micolich and Jonas (2002), which
are restricted to fractal images. Donderis file-compression measure provides an efficient
and reproducible measure of image complexity.
Using file size as an indicator of complexity makes it equivalent to Kolmogorov
complexity or algorithmic information complexity (AIC), in that both can be considered a
measure of information. This suggests that a measure of visual complexity based on such a
scale means that it may diverge from perceived complexity for random images. This
hypothesis predicts an inverted U shape correlation between objective (file-size) and
subjective (MES) measures of complexity. Here is one area in which we may contribute to
knowledge: by providing empirical data to test this hypothesis.
Identifying the Components of Visual Complexity

We have identified a number of different schemes for classifying and understanding
visual complexity, based on a variety of aesthetic properties. The informational scale of
complexity has order and randomness at opposite ends of its scale, but understanding of
visual complexity may be based on other conceptual pairs, as Berlyne observed:
The idea that aesthetic value requires a combination of two partly opposite and partly
complementary factors has cropped up repeatedly over the centuries. They have been
labelled uniformity and variety by Hutcheson (1725) and many later writers,
order and complexity by Birkhoff (1933), subjective redundancy and statistical
information by information-theoretic aestheticians (Gunzenhauser, 1962),
concinnity and empathy by Coates (1972), [and] coherence and mystery by
Kaplan (1973). (Berlyne, 1971, p.9)
It should be noted that these complementary pairs are not always conceived as
diametrically-opposed properties (like hot and cold), but can also be understood as being at
right angles to each other. In other words, we should imagine that together they form two
orthogonal axes instead of a single linear scale. (We will leave to one side the question of
Summary of Context
151
whether these conceptual pairs actually correspond to hedonic functions of the nervous
system, as Berlyne suggested, since it is beyond the scope of this investigation.) In this
project, we have two such orthogonal factors: complexity and order. Perceived complexity
is the principal object of investigation, understood as a scale of simplecomplex, which
constitutes the first factor. On the other hand, the objective measure of complexity by
compressed file size gives a scale of orderrandomness, and this constitutes the second
factor. The projects contribution to knowledge is based on the measurement and
correlation of these two aesthetic factors: The complexity factor is based on the reported
perceptions of test participants viewing a series of images. The order factor is based on the
formal properties of the images digital files, quantified through data compression. The
project aims to contribute to knowledge by extending the range of these factors under
investigation: Firstly, by sampling the entire scale of orderrandomness using computergenerated images to gain a wider picture of the subjective response to this range; Secondly,
by testing samples from contemporary art and design practice to see whether we can relate
these measurements back to the overall view developed in the first tests.
Here we present a model of the most pertinent variables that contribute to visual
complexity, developed from the material reviewed earlier in this chapter. Using the general
notion of complexity as difficulty-of-description, the images in Figure 61 represent a scale
of increasing complexity. If we describe each image completely and succinctly, the size of
the descriptions increases from left to right:
A
Quantity
Variety
Order
Figure 61 Three principal components of visual complexity.
Summary of Context
152
The differences between pairs of adjacent images illustrate the main components of
visual complexity. The difference between images A and B is one of quantity of picture
elements; the greater quantity in image B makes it more complex than A. Image C increases
the variety of elements by increasing the number of colours from two to nine, and so C is
more complex than B. Image D contains the same elements as C, but their order of
arrangement is different C is more regular and so D is more complex. In summary, the
main properties that contribute to visual complexity are the quantity, variety and order of
picture elements. The three components of visual complexity identified here are more than
theoretical; they share a correspondence with the most significant perceptual elements and
also with the principle components of digital image files. For the purpose of the projects
empirical tests, we need to constrain the quantity and variety of picture elements, while
manipulating their order. Each test must use a set of images with identical pixel dimensions
and the same colour system. In summary, then, visual complexity can be measured in many
different ways, based on various elemental properties. These elements of visual complexity
can be represented by digital images, thus they can be reduced to the common elements of
digital image files, namely: the dimensions of the pixel array (quantity), the colour values of
the pixels (variety) and the locations of those colour values in the array (order).
Visual Complexity and Image File Compression

One problem with trying to compare empirical studies of visual complexity is the
lack of a standard metric. Forsythe et al. (2003) argue that an objective measure of visual
primitives (in their case, edges) may provide a valid index of complexity for all 2-D stimuli.
Their perimeter measure of edges is based on Attneaves (1954) informational approach,
but with the benefit of using computer technology to implement the aesthetic measures
and analyse the results. Donderis file-size measure is also reliable and convenient, and also
has the potential to become a standard metric which may be favourable to edge-detecting
measures for its simplicity and its direct correspondence with the informational content of
a digital image. Donderis measure is applicable only to images of the same dimensions,
Summary of Context
153
however, because the size of a compressed file is less dependent on the randomness of its
data than on its dimensions. On its own, a file-size value tells us nothing about the
complexity or dimensions of an image; it only becomes meaningful in comparison with
other files of the same type. There is a need, then, for a more informative metric. One
potential measure is the compression ratio, which expresses how much a files size changes
after compression. There are reasons to believe that 1) not only would this give a relative
measure of randomness but also that 2) it might allow for a comparison across different
image dimensions. These two points are explained in more detail below:
1) Using the compression ratio gives a clearer idea of randomness than does the filesize alone. A totally random image, which cannot be compressed at all, will give the same
file size as any uncompressed image with equal dimensions. Therefore, the relative size of a
compressed file tells us how close it is to being purely random, and so the compression
ratio (compressed size divided by uncompressed) gives a quantitative measure of how
random an image file is. A compression ratio value near to 1 indicates incompressibility,
and therefore randomness, whilst a value near to 0 indicates high redundancy and order.
This relative measure is more useful than the absolute measure of file size because the latter
measure only becomes meaningful in comparison with other measured values.
2) It seems intuitively correct to assume that a given pattern will have a compression
ratio that remains roughly constant across a range of image dimensions, such that a small
check pattern will have a similar compression ratio to a larger version. To find out, we
could use a large image from which reduced-size copies are made, and calculate the
compression ratio for each size. If the ratios stay the same for a pattern or type of image,
then we can be confident that this measure is unaffected by image dimensions, unlike the
absolute file-size. We may expect that ratios will be fairly consistent for most image types,
but also that they will drift as the images become smaller and the distinction between
randomness and complexity becomes increasingly uncertain. It may be possible to test this
Summary of Context
154
technique more thoroughly in future, but in the current project we evaluate its effectiveness
compared to the method established by Donderi based on the absolute file size.
Locating Visual Complexity

With an understanding of informational measures of visual complexity and the
results of empirical tests of its perception, we may be able to determine more precisely
where the most complex and interesting images are to be found on the informational scale
of orderrandomness. Since Donderi & McFadden (2005) and Besner (2007) used file-size
and not compression ratio, we cannot make any quantifiable statements about how
perception varies with the randomness of those image files, because we cannot determine
quantitatively how random those images were. We may, however, learn something from
Dodgsons findings:
humans can easily detect patterns when up to about 25% of the pattern is removed
or disturbed, that removal of over about 50% of the pattern destroys it, that there is
an sthetically interesting region between these two values, and that a good, artistic,
balance between regularity and randomness is achieved by retaining about two-thirds
of the pattern, while manipulating the other one-third in some way. (2008, p.112)
Dodgsons results shed light on the question of where perceived complexity is to be

found on an informational scale of orderrandomness. If we fail to perceive patterns
beyond about 50% randomness then this is a threshold of visual complexity, since it is
effectively the edge of chaos in perceptual terms the point just before randomness. The
images that we perceive as most complex will be found near this threshold, and a region of
visual appeal is circumscribed by this threshold and a lower bound of 25% randomness,
centred roughly at one third. Because the file-compression measure of visual complexity is
technically a measure of randomness, we can translate Dodgsons estimates to a scale based
on the compression ratio of image files: We can expect the most appealing images to be
found in files with one third randomness, that is, with a compression ratio of
approximately one third (Figure 62).
Summary of Context
155
[Complex]
Ordered
Random
Region of
visual appeal
0
0.25
0.5
0.75
Figure 62 A scale of orderrandomness (top) corresponding with compression ratio

(bottom).
We should, however, be careful about extrapolating from Dodgsons experiment

because it was limited to images composed of only a few simple elements black disks of
various sizes on a white background, and also because it is not a formal empirical
investigation. We may well find that these estimates of the region of visual appeal are
somewhat different for image files in colour and with other types of pattern. Nevertheless,
the framing of complexity in terms of order and randomness is a useful descriptive tool,
and one that is particularly appropriate for an objective measure of visual complexity which
is based on image file compression. Despite the fact that this measure actually quantifies
randomness instead of what we intuitively think of as complexity, it remains a practicable
method for the investigation of visual complexity in contemporary art and design. In terms
of the information scale of visual complexity, what we have learned may be summed up
with Flakes (1990) statement that the interesting stuff is in the middle:
Beauty i.e., that which makes something interesting, is related to a mixture of
regularity and irregularity. When things are too regular, we usually find them to be
uninteresting because they yield no surprises for us. Complementary to this, highly
irregular things are often uninteresting because they make no sense. In the middle,
between regularity and irregularity, lies a place where things can be understood, but
not completely. (Flake 1990, p.411)
We have identified suggestions that randomness or chaos are perceptually less

complex and less interesting than a mixture of order and randomness from contributions
by Attneave (1954), Arnheim (1959), Berlyne (1971), Wolfram (2002), Gell-Mann (2003),
Summary of Context
156
and Donderi (2006). However, empirical research into perceptions of highly complex and
random images began only recently (Besner, 2007), and the validity of these claims is yet to
be established. Thus there remains a gap in knowledge of the relationship between
informational measurement and the perception of visual complexity.
Summary of Context
157
Chapter 3
Approaching Visual Complexity in Art &
Design
Research Focus
The principal focus of this project is the aesthetic property of visual complexity. For
practicality, the study is limited to the complexity of two-dimensional images in
contemporary art and design practice. The project is largely situated within the discipline of
fine art, but the scope of artwork under investigation includes two-dimensional visual
material from a variety of creative fields, namely: photography, architecture, graphic design,
illustration, decorative arts and textiles. For all these types of artefact, the research scope is
limited to their aesthetic attributes only; it excludes an appraisal of the utilitarian functions
of design and the interpretative aspects of art. In terms of Moles (1966) distinction
between semantic and aesthetic information, this project concentrates on aesthetic
information, but the project also deals to some extent with semantic information via
interviews with test participants to identify the criteria behind their aesthetic judgements.
The aim is to explore this subject via an empirical investigation into the aesthetic
property of visual complexity. Justification for this research focus comes from the
argument that an empirical study of aesthetics has the potential to complement the nonperceptual classificatory aesthetic definitions of art (see p.67). It can do this by providing an
explanatory theory based on the measurement of perceived complexity and the information
analysis of aesthetic images. Contributions to knowledge are made by extending
computational aesthetic measures to the field of art and design, and by providing empirical
evidence for a perceptual threshold in the informational spectrum of visual complexity.
Research Focus
158
The literature reviewed in the previous chapter serves to identify gaps in knowledge
and to select the research methods appropriate for this investigation. The present chapter
describes the methods and approach of this research project, beginning with the following
section which outlines the research methodology that guides the choice of methods.
Methodology
Ontology and Epistemology
In the investigation of visual complexity, this project focuses on the relationship of
objective properties to the subjective properties of perception. This constitutes a position
of realist ontology, which can be expressed as follows: The belief in the existence of things
external to our consciousness, and that the physical properties of those things are
independent of how we perceive them to be. Those properties are also independent of the
concepts that we employ in understanding them and the language we use to describe them.
This project does not subscribe to the view that everything we know is based on sociallyconstructed language. Empirical evidence for the perception of visual complexity
functioning partly at a pre-conscious level of perception (e.g. in the retinal processing of
visual primitives such as edges, lines and blobs) is enough to reject that view. The socialconstructivist position is a form of anti-realism (or relativism), which accepts the existence
of external objects and events, but denies the independence of realist claims to knowledge
about those objects. This project rejects this position and adopts a realist view.
The realist ontology of this project comprises artist and audience (subject), artwork
(object) and artworld (context). These basic ontological categories provide a foundation for
the projects working model (Figure 6), in which the aesthetic is identified as the interaction
of object and subject. Therefore, the objects of study in the current investigation are the
physical properties that give rise to this interaction and the perceptual phenomena that
result from it. The working model of aesthetics incorporates the ontology into the context
of the practice and perception of art and design. Although the model was derived
Methodology
159
independently of Dickies art circle theory, Chapter 2 analysed the structural similarity of
the two models. Their concurrence supports the validity of the ontological model in terms
of its connection to contemporary understanding of art practice.
In the working model, the aesthetic is understood as the interaction between subject
and object. The subjectivity of perception means that each individual has a unique
perception of a particular object, and as such it is only indirectly accessible to knowledge.
To say, that is beautiful is to ascribe a subjective value to an object. In this projects
conception of aesthetics, such a statement provides a route to understanding the aesthetic
interaction of subject and object. In other words, aesthetic interactions are phenomena
which are subjective (i.e. private and individual), and therefore difficult to access, but which
can be glimpsed through personal statements (that express ones perceptions) and artistic
acts (that express artistic aims). Beyond the analysis of art and design objects themselves, it
is these personal statements about subjective perceptions that form the objects of study in
this research, since they are the most direct and accessible forms of aesthetic judgement
available to an observer. Therefore, the epistemology of this project is not strictly
empiricist in its philosophical position (i.e. it does not accept that knowledge is only derived
empirically), but it is methodologically empirical because it is based on observation.
Problems with Empirical Aesthetics

Philosophical Problems
When Baumgarten coined the term aesthetics, it meant a science of aesthetic

perception. Wittgenstein said that a science of aesthetics is almost too ridiculous for
words (1966, p.11), because aesthetic questions have nothing to do with psychological
experiments, but are answered in an entirely different way (1966, p.17). Wittgensteins
argument is acceptable only if we admit that aesthetic questions are of a different kind to
psychological questions. The opposing view is expressed by Daniel Berlyne, founder of the
International Association for Empirical Aesthetics:
Methodology
160
Whether a branch of study can be called scientific does not depend on whether it has
yet answered its questions. It depends on what kind of questions it is asking and what
methods it adopts in seeking answers to them. (Berlyne, 1971, p.2)
Berlynes position is typical in modern psychology: what defines a field of science is

not the type of question, but the way in which the question is answered namely by the
empirical method of accountable observations. A psychological question, therefore, is one
that can be answered through the process of scientific method. For example, does God
exist? is not a scientific question because it cannot be answered empirically that is, we
cannot reach a satisfactory answer by thoroughly exhausting all possible experimental
attempts to provide evidence to the contrary. This is Karl Poppers (2002) concept of
science, which says that only falsifiable hypotheses belong to the legitimate scientific
domain. What this means is that science does not work by positive proof, but through lack
of evidence to the contrary. A single counter-example is enough to rule out an
experimental hypothesis, but unless and until one is found a hypothesis can become an
accepted theory. Hypotheses are generated and then pruned away by systematically
showing them to be wrong, eventually leaving only a few plausible explanations. By these
criteria, God exists is not a scientific hypothesis because it is not falsifiable; we cannot
provide evidence that God does not exist, and so the question remains beyond the scope of
scientific method. A successful theory is thus falsifiable in principle, but not in practice.
The question how long will I live? could be answered with methods from
astrology, religion, medicine, or statistics. The question is not scientific per se, but the
scientific method is one among many possible approaches to an answer. From the set of all
possible questions, therefore, scientific questions are the subset than can be answered
empirically, and some of these may be answered in other ways. Insofar as the classification
of a question as scientific depends on the nature of scientific method, it could be argued
Methodology
161
that a science is characterised more by the methods it uses in answering those questions
than by the type of questions themselves 10.
Traditional aesthetic questions, whose answers remain in contention, include what is
art?, why do we value art? and how does art acquire its value? These big questions of
aesthetics do not appear to be of the form that can be answered with scientific method, but
other aesthetic questions are open to more than one form of methodical enquiry. Which
of these two paintings is better? is an aesthetic question that can be answered empirically,
by asking a number of people to state their opinion and analysing the results. Whether we
are satisfied with such an answer, though, depends on our epistemological position. This
projects epistemology is outlined in the previous section, and does accept this form of
answer as a route to developing understanding of aesthetic perception. The aim is not to
reduce the aesthetic to a simple mathematical formula, but to enrich creative practice by
understanding the causes and effects of visual complexity in art and design.
Dickie (1961) opposed empirical aesthetics, but I believe that his institutional theory
of art can be supplemented with perceptual aesthetics. Dickies theory (1997) only
describes the mechanism of the artworld, in an attempt to avoid the pitfalls of essentialist
definitions of art which became problematic for philosophical aesthetics as art practices
changed. The problem with the institutional theory is that it does nothing to explain why
we value art. In an article in Leonardo, the psychologist Rolf Reber (2008) argues
convincingly that empirical enquiry has the opportunity to add to understanding the
10
It should be noted that there are also scientific questions that cannot be answered by science. For example,
radioactive decay can be predicted accurately for a large number of atoms because collectively they obey
statistical laws: For any radioactive material, the length of time it takes for exactly half of the atoms to decay is
a constant period known as its half-life. But the question when will a particular radioactive atom decay?
cannot be answered with anything other than a statistical probability. Similarly, there appear to be limits to the
measurement of both the momentum and position of quantum particles, which is embodied in Heisenbergs
famous uncertainty principle. These examples reveal interesting limits to our knowledge as well as intriguing
properties of our world.
Methodology
162
perception of art, specifically that psychology can help to assess artistic value. Reber says
that art theory and scientific method are able to integrate by empirically testing the criteria
specified by art theories (e.g. whether the golden ratio is more beautiful), but warns that
science is unable to evaluate the theories themselves: psychology can determine artistic
value, given a criterion, but not whether the criterion itself is a good or a bad one (Reber,
2008 p.371). Ken Friedman also makes the case for an empirical approach to design, based
on an understanding of aesthetics in the traditional meaning of sensory perception:
the scientific approach to design does not contradict the artistic aspect of design.
Successful design artifacts have aesthetic values and qualities, sensual and engaged. All
designed objects, tactile, mechanical, visual, auditory, are mediated through the
physical senses. Sensory quality is a central issue for articulate objects that work in a
physical world. (Friedman, 1997, p.58)
Friedmans argument for design science can also support this projects empirical
approach to visual complexity, since visual art is also mediated via the senses. Despite the
evidence of a move away from perceptual aesthetics in the institutional theory of art and in
the practice of relational aesthetics, there is no escaping the sensory foundation of aesthetic
experience. Therefore, an empirical approach to aesthetics has the potential to inform
understanding of contemporary aesthetic theory and practice and to fill the gaps in
contemporary knowledge of the visual arts. In this context, the aim of the current project is
to investigate just one element of perceptual aesthetics visual complexity.
Practical Problems
If we accept the empirical approach to aesthetics in principle, then we face problems

with the approach in practice. Visual complexity is difficult to measure directly, since it is a
property based on more fundamental attributes (quantity, variety and order). In addition,
the subjectivity of perceptions causes difficulties for measurement because they are highly
variable and inaccessible to direct observation. This is as much a problem for any other
psychophysical study as it is in the case of visual complexity. For example, the estimation
of hearing thresholds (equal loudness contours) suffers the same difficulties. The solution
Methodology
163
to the problem of variability is to repeat tests and to average the results, whilst the problem
of accessibility is solved using a method in which participants self-report their perceptions,
namely Stevens (1975) magnitude estimation scaling technique.
As an illustrative example of the current practice of standardisation in empirical
aesthetics, I take from the literature reviewed earlier the PhD thesis Complexity and Aesthetic
Preference for Diverse Visual Stimuli by Marcos Roberts (2007), who made the following
adjustments to a set of images of artwork:
To avoid the undesired influence of psychophysical variables, all stimuli were adjusted
to the same resolution of 150 ppi and set to the same size of 9 by 12 cm. Additionally,
the colour spectrum was adjusted in all images. Values of extreme illumination and
shadow in each picture were adjusted to reach a global tone range allowing the best
detail. Stimuli were classified according to their dominant tone (dark, medium, or
light), and those with a mean distribution of pixels concentrated in both the left (dark)
and right (light) extremes of the histogram were discarded. Thereafter the luminance
of the remaining stimuli was adjusted to between 370 and 390 lx. Stimuli that could
not be reasonably modified in this sense were discarded. Finally, the signature was
removed from all signed pictures. (Roberts, 2007, p.159)
Those digital images of artworks have been altered in size, proportion, content,
colour, brightness and tonal balance. These are all fundamental aesthetic properties, and
from an artistic point of view their alteration constitutes a mistreatment of the artwork. As
a consequence, the validity of this approach is undermined. The attraction of ironing-out
these confounding variables is understandable from the empirical point of view because it
makes the effects easier to detect and makes the analysis simpler. There are two arguments
for preserving the integrity of stimulus images, however one scientific (technical) and one
artistic (moral), as follows.
The first argument is that in the case of experiments on aesthetic complexity, the
treatment of artworks is a particularly sensitive issue, because by all accounts visual
complexity is a function of a number of more basic visual properties, including those
altered in the example above namely size, proportion, resolution, colour and tone. The
Methodology
164
literature review identified the quantity, variety and order of picture elements as the most
fundamental attributes of visual complexity, and in the case of Roberts (2007) experiments
these have all been altered. In trying to measure visual complexity, it makes sense to
minimise the alteration of properties to which objective measures and subjective
perceptions of complexity are sensitive, because these alterations confound the
experimental variables and weakens the ability to make inferences from the results of the
test stimuli back to the original artworks.
The second argument for preserving the integrity of an artwork as a stimulus image
is based on the moral principle that an artwork deserves respect in the same way as does its
creator. The Copyright, Designs and Patent Act 1998 (CDPA) was introduced to enable
artists and designers to control how their material is used. The fact that these rights are
retained by the author even when the copyright of the work is transferred illustrates their
significance. Provided that an experimenter has obtained licence to use the works in
question, there should be no issue of authorship involving the right of paternity (CDPA
77), but the treatment of an artwork for experimental purposes should respect the right of
integrity (CDPA 80, Appendix A). This is the right to object to derogatory treatment of
the work, which may include addition, deletion, alteration or adaptation (DACS, 2008).
To neglect this principle is to run the risk of undermining the value of the research, not
only from the art worlds point of view for whom such research is potentially rewarding
but also in terms of experimental validity. 11 In conclusion, this project aims to minimise the
alteration of artwork images in testing.
Methodological Approach
The central aesthetic question of this project is: Does visual complexity contribute to
the aesthetic value of art and design? The question cannot be answered directly through
empirical methods, but its component parts can be tackled in this way with techniques
11
I should add that I am not implying that Roberts (2007) has infringed copyright; I only mean to add weight
to my argument by emphasising the seriousness of these principles for artists and designers.
Methodology
165
from experimental psychology. The first step is to measure visual complexity and the
aesthetic judgement of visual stimuli. The objective measure, based on image file
compression, provides an informational scale of order to randomness that corresponds
with classification schemes of complexity from systems theory (Wolfram, 2002; Langton
1990). This suggests research questions that can be answered empirically: How complex are
examples of contemporary visual art and design? Is there a region of the informational
spectrum of complexity in which these artworks tend to be found? What is the relationship
between measures of objective informational complexity and subjective perceived
complexity? Artworks appear to share a level of complexity with living things both
display a balance of order and chaos. Might we find that, like the preference for natural
fractal values (Taylor, Micolich & Jonas, 2002), there is a preference for levels of visual
complexity similar to those in nature?
To answer the research questions, this project employs a methodology with an
integration of methods, both creative and empirical. The creative work explores methods
of generating visual complexity, which has applications for both art practice and the
empirical tests. One aim is to make artwork aiming for a particular level of complexity
the perceptual threshold of complexity hypothesized by Wolfram (2002) and Donderi
(2006) and then to test this experimentally by including this material in the test stimuli.
The art practice also helps to guide the creation of visual stimuli for testing by exploring
the computer as a universal picture generator with programs such as cellular automata.
After generating and collecting visual material, we can then perform empirical tests to
measure the property of visual complexity and its perception in art.
The working model of aesthetics provides a structure for this projects methodology.
Since the model identifies the aesthetic in the interactions between an artwork and its artist
or audience, the methodology focuses on these two sites of aesthetic production and
aesthetic perception, and also on the artworld which forms the context of this research
project. The plan is to approach these three areas of study aesthetic production, aesthetic
Methodology
166
perception and the artworld by respectively making, measuring and mapping: Firstly,
mapping aesthetic discourse with a contextual review of theory and practice identifies gaps in
knowledge and guides the choice of methods. The contextual knowledge informs the
making of artwork and visual stimuli for use in empirical tests. Finally, measuring the
perception of visual material provides an empirical route to understanding the aesthetic
judgement of visual complexity in art and design. The relationship of this methodology to
the working model of aesthetics is illustrated in Figure 63.
Make
Measure
artwork for
perception of
testing and as
complexity and
experiments in
aesthetic
aesthetic complexity
judgements of art
Map
aesthetic
discourse
Figure 63 The methodology and the working model of aesthetics.
Methods
The principal methods employed in this project are quantitative in nature. Their
purpose is to quantify visual complexity through the measurement of objective aesthetic
properties and subjective aesthetic perceptions. The aim is to investigate the correlation
between these two measures over a range of visual complexity. To support this quantitative
approach, however, the project also employs qualitative data collection and analysis
methods. Through interviews with test participants we aim to establish, if possible, the
criteria employed in making the aesthetic judgements in these tests to shed light on the
how that lies behind the how much. The following sections detail each of these methods.
Methods
167
Quantitative Methods
Measuring Objective Complexity
Donderis (2006a) file-compression technique offers a practical solution to the

problem of measuring visual complexity based on the identification of visual elements and
the patterns they form: Pixels are the elements and data compression does the patternrecognition. Digitisation provides a standard approach that can be applied to all 2D images.
Making a digital version of an image involves a quantization of its visual information. In
general, this means that information is transformed from continuous values to discrete
values. In the case of digitization of images, the quantization process transforms the
perceivable picture elements (e.g. patches of paint) into an array of pixels. The pixel thus
provides the basic visual element in digital images, and it is the statistical properties of these
elements that form the basis of the computational measure of visual complexity. We
identified the fundamental properties of visual complexity as quantity, variety and order.
These are also the fundamental attributes of digital image files; the quantity, variety and
order of pixels are the properties that determine the size of the file after data compression.
Therefore, in theory Donderis file-compression method provides a theoretically-sound
measure of visual complexity which also has empirical evidence for its success with radar
and chart images (Donderi & McFadden, 2004) scenes of nature (Donderi & McFadden,
2005) and camouflage images (Besner, 2007). The current project applies this method to
the aesthetics of art and design.
One advantage of the method is that in generating the digital image files to use as the
basis of physical stimuli, we simultaneously implement part of its analysis. By generating
compressed image files to start with, we have an initial indication of the objective
complexity of the images. Once we have a set of compressed images, we can simply sort
them in order of file size to get an idea of which are the most complex and which are the
simplest. Most of the source images used in these tests are generated primarily as PNG
files, whose compression algorithm gives the smallest files without data loss. This near-
Methods
168
optimum file compression means that the initial file size is close to the intended complexity
measure, and saves space on the computer system. A problem is that in making digital
reproductions of visual material, we introduce visual artefacts in the process by creating a
discrete-valued representation of a perceptually continuous visual array. By using highresolution scanning and printing, we can minimise the unwanted effects of this process.
Measuring Subjective Complexity
The chosen method for measuring perceived complexity and aesthetic judgements is
the psychophysical magnitude estimation scaling (MES) technique derived by Stevens
(1975) and employed by Donderi & McFadden (2005). The method requires the
participation of volunteers to engage in tasks, who for this project are recruited from staff
and students at Nottingham Trent University, principally from the schools of psychology
and art and design. By employing different groups of participants we are able to examine
the effect of art experience on aesthetic perception and judgement of visual complexity.
Conducting interviews with participants after the tasks also allows for qualitative data to be
gathered which can add to the interpretation of the quantitative results.
The primary subjective aesthetic measure is perceived complexity. The other
subjective measurements in this project are aesthetic preference (how much an image is
liked) and aesthetic judgement of artistic quality (an evaluation of the artistic merit of a
work). In this thesis, we abbreviate these two measures with the terms preference and quality.
In effect, by investigating aesthetic preference and judgement of quality, we are getting
close to what has traditionally been called beauty. Philosophical aesthetics classifies
aesthetic preference as a judgement of agreeableness, and aesthetic quality as a judgement
of taste. This aesthetic conception places aesthetic quality alongside the concept of beauty,
whereas for those outside the field of aesthetics beauty would be usually associated with
aesthetic preference rather than artistic quality. To avoid confusion in the tests, we use the
terms preference and quality.
Methods
169
Since the nature of data to be collected is quantitative, it is appropriate to use

statistical data analysis methods. Because we are looking for a correlation between objective
complexity and subjective aesthetic judgements, we need to use statistical tools such as
regression analysis which will tell us the strength and direction of the correlation. The
details of these statistical methods are discussed in the following chapters as they are dealt
with in the description of the test procedures and results.
Qualitative Methods
Semi-Structured Interviews
In order to see whether we can establish which criteria are employed in making the
aesthetic judgements in the tests, the plan is to conduct semi-structured interviews with
participants. Because the principal focus of this study is quantitative in nature, and also due
to constraints in time and resources, these interviews are not to be conducted for every
test, but only for the first part of the final test (Test 3a). The main objective is to determine
the criteria in judging aesthetic complexity, preference and quality. The interviews also
allow us to see how the participants coped with the task and offer them an opportunity to
ask questions and comment on the experience. This data is recorded as audio files, and
later transcribed for qualitative analysis, as described in the following section.
Transcription Coding and Analysis
A principal difference between the quantitative data gathered in the main tests and
the qualitative data gathered in interviews is that the former relates to objects seen whilst the
latter concerns events heard. Another difference is that the quantitative data relates to the
perception of aesthetic information, whilst the qualitative data deals with semantic
information. This requires a fundamentally different approach to analysis, because we are
dealing with semantic information rather than aesthetic. Content analysis is the broad
heading under which this type of analysis falls, and in this study we focus on transcription
coding as the principal method. Krippendorf (2004) describes three approaches to content
analysis, based on three different understandings of how the content (meaning) is identified
Methods
170
in the process of analysis. The first of these positions comprises the belief that the content
is inherent in the transcribed text the content is simply there on the page to be
discovered. The second regards the content as a property of the source of the text, namely
the interviewee, and accepts the fact that there is an element of uncertainty in the
transcription process. The third approach goes yet further by acknowledging the role of the
researcher in actively constructing the content of the interviewees words through the
process of transcription and analysis. In terms of the diagrammatic representation of
qualitative analysis in Figure 64, the first approach focuses only on the text and disregards
the other elements; the second also includes the interviewee as the source of the text; the
third approach incorporates all three by recognising the role of the researcher in
transcribing and analysing. This project adopts the third approach to content analysis.
Interviewees
Text
Researcher
Interview,
Coding,
Transcription
Analysis
Figure 64 Representation of the qualitative analysis process.
Amongst the various methods of content analysis, this project focuses on a form of
transcription coding as the main method of analysing the qualitative data gathered via the
semi-structured interviews. According to Saldana (2009), the process of coding identifies a
data sets primary content by reading and interpreting (de-coding) as well as identifying and
labelling (en-coding): coding is the transitional process between data collection and more
extensive data analysis (Saldana, 2009, p.4). The main concerns of this analysis are to
enumerate and categorise the various responses to the interview questions and to
determine through a basic statistical analysis their frequency and significance.
Methods
171
Plan
The literature reviewed in the previous chapter reveals that although visual
complexity has been a subject that dates back to the beginnings of experimental
psychology, there have been few attempts to study the entire range of visual complexity.
The reason for this may be that it is difficult to achieve, because it is not easy to create such
a range of stimuli and also because the analysis required gets more complicated with more
complex stimuli. Looking at the whole spectrum allows for a useful preliminary overview
of the phenomenon and provides a frame of reference for further tests. Therefore, this is
the aim of the initial tests in this project. The proposal for test 1 is to generate a series of
images that represent equidistant samples along a scale of objective complexity as measured
by file compression. In this way, we are able to generate a series of images that cover the
four classes of complexity identified by Wolfram (2002), namely: uniform, repetitive,
complex and chaotic.
The plan is to start with limited variables and increase them gradually in successive
tests. Test 1 begins with low-resolution black and white images (that is, with only 2 varieties
of picture element) in order to be able to focus on order/arrangement of elements. The
stimuli are generated from cellular automata and random programs. In terms of the
quantity, variety and order of picture elements: the quantity is fixed at a fairly low value
(one million pixels) and the variety is low (just two different types black and white) while
we manipulate the order of pixels. Test 2 increases the overall complexity of the stimuli by
increasing the quantity and variety of the picture elements: The CA and random images
have up to four colours with a resolution of just over 4 million pixels (with dimensions
2048 2048 pixels). Finally, test 3 uses images of art and design artefacts scanned at high
resolution (around 35 million pixels) and with full colour (using RGB colour space, up to
16 million colours). Test 3 takes place in two locations the psychology laboratory and the
art gallery which are presented as tests 3a and 3b respectively.
Methods
172
The relationship of the chosen methods to the structure of this project can be
illustrated by reference to the working model of aesthetics. Figure 65 shows how the
methods relate to the model. On the side of aesthetic production (between artist and
artwork), the method is to make and collect visual material for use in the tests. The test
design is informed by the mapping of current practice (in the artworld) via the contextual
review. The tests use two methods to measure visual complexity: Objective complexity (of
the artwork) is measured with the file-compression method, and subjective complexity (the
perception of the audience) is measured with the MES method. The perceptions of the
participants are further elaborated through interviews and qualitative analysis. In this way,
the methodology approaches both sides of the aesthetic in the model aesthetic
production and aesthetic perception, whilst the methods investigate both objective and
subjective aesthetic properties of visual complexity.
Collecting
artwork
Objective
Conducting
tests
Subjective
Conducting
interviews
Making
Quantitative
Qualitative
artwork
analysis
analysis
Contextual
review
Test design
Thesis
Figure 65 Diagram to show how the methods relate to the working model of aesthetics.
Methods
173
Chapter 4
Investigating the Spectrum of Visual
Complexity
Trial 1: Objective Measure
Phase one of the method trial examines the objective measure of visual complexity.
We use Donderis (2006a) measure of image complexity on images of Wolframs
elementary cellular automata (eCA). Using CA images has a few benefits: Firstly, it offers a
convenient method of generating manipulable stimuli; it is relatively easy to generate these
images in any colours or dimensions. Secondly, we can be confident that the images
represent a wide range of complexity (as mapped out by Langton and Wolfram), and the
relatively small number of eCA rules means that we can produce images of all of the 256
patterns in this type of CA. Lastly, it makes for an easy comparison of results against the
findings of Wolfram (1984) and Langton (1990). Using CA as test stimuli also has the
advantage of providing images that avoid the confounding factor of familiarity (Forsythe,
2008), since these types of pattern are not widely known to the general public and are
scarcely seen in contemporary visual art.
A set of all 256 elementary cellular automata is generated by writing a Mathematica
program. The eCA images have just two colours (black and white) and each has dimensions
of 100100 pixels. The compressed PNG image files range in size from 96 to 1,500 bytes.
Figure 66 shows the set of images arranged in order of their rule number, from 0 at the top
left to 255 at bottom right. With this type of CA, there are many more regular patterns than
random ones, therefore some pattern types are over-represented in this set. In subsequent
tests, a selection of CA stimuli is made by sampling from equidistant points along the range
of file sizes, avoiding the duplication of similar images.

174
Figure 66 Images of all 256 elementary cellular automata, sorted by rule number.
When the set of compressed image files is sorted by size, the result is the
arrangement shown in Figure 67. It is readily apparent that the clustering of image types
mirrors Langtons arrangement of Wolframs four complexity classes: The simple, uniform
images come first, 12 followed by increasingly complex repetitive patterns, and finally the
most random images are found at the end. Between the regions of order and chaos are a
few complex images that have elements of order and randomness.
12
It seems that the reason for black images coming before white is a quirk of the sorting system because both
black and white uniform images are 96 bytes (768 bits) in size. These are all bi-level images, which use only
one bit per pixel to store colour information, and so it takes no more information to store white than black.

175
Figure 67 Images of all 256 elementary cellular automata, sorted by file size.
Note that the images with the largest size are fairly smooth in texture, whereas the
more lumpy-looking images are those between the regions of order and chaos. The lumpy
ones look like that because they have a greater number of different patterns than the other
images. If we judge the complexity of images by the size of their description, as we said
earlier, then these lumpy-looking images are the most complex. Because these images
contain some regularity as well as randomness, they do not produce the largest files, since a
regular pattern is compressible. The file-size measure is in agreement with Wolframs
classification and Langtons ordering, but because of this we also see how the measure is
similarly inconsistent with our intuitive understanding of visual complexity.

176
In one sense this result is trivial, since it could have been predicted by anyone who
understands image file compression. On the other hand, it is surprising to see such a close
match between different schemas from experimental psychology and complexity science.
As far as the literature review reveals, this trial is the first time that data compression has
been used as a measure of cellular automata complexity, and provides the first direct link
between the information-theory measure of visual complexity and Wolframs (2002) and
Langtons (1991) complexity schemas. As such, it may constitute a contribution to
knowledge, though the potential application of this finding is as yet uncertain.
What links the complexity schemas with the file-compression measure of complexity,
and explains the result of this trial, is the concept of order as developed in information
theory (Shannon, 1948). What this means is that the objective measure of visual complexity
is actually a measure of randomness (or entropy as Shannon called it), and is not
necessarily a measure of complexity as we perceive it. This might not be a problem for the
current investigation, though, since it still provides a practicable method which not only
accords with models from complexity theory but which is also robust and clearly
understandable. It also supports the idea that complexity can be understood as a mixture of
order and randomness, since we find the most complex images between these regions.
Nevertheless, the implication is that the correlation which Donderi found between file size
and perceived complexity may break down for the most random images. This suggests two
things: Firstly, that Donderi & McFadden (2004) only looked at relatively simple images,
for which the correlation is strong because there is agreement between informational and
intuitive measures of complexity for the lower part of these scales. Secondly it provides a
reason for investigating a wider spectrum of visual complexity in order to resolve this
concern and to contribute to the field by extending research into this area. In conclusion,
we have validated the objective measure of visual complexity and discovered significant
issues which will bear on the interpretation of subsequent test results. Next, we look at the
use of cellular automata images in the measurement of perceived complexity.

177
Trial 2: Subjective Measure

To check that cellular automata images are feasible to use as test stimuli for
subjectively perceived complexity we performed an informal test. Sixteen black and white
images of cellular automata patterns, chosen from the set of 256 used in the previous trial,
are selected to represent a range of complexity. The images are printed on A4 paper and
laid out randomly on a table top. Five people (three PhD students and two members of
staff from the School of Art & Design) were tasked with arranging the images in order of
perceived complexity by physically laying-out the images in sequence. Figure 68 shows the
averaged results.
Figure 68 Image set used for the trial of the subjective measure, sorted in order of
increasing perceived complexity from top left to bottom right.

178
The results are encouraging because in general the images were sorted in an order
that approximately matches the sequence of complexity classes found in the previous trial.
The biggest disagreement between individuals seems to be with the chaotic images some
thought that those were the most complex and others thought they were much simpler
but in general the ordering of the images was fairly consistent in terms of the placement of
various complexity classes and their types of pattern.
One drawback of measuring perceived complexity by rank order is that the type of
data it provides only tells us which images are perceived to be more complex than others; it
cannot tell us whether any two images are perceived to be near or far apart on a scale of
visual complexity (for example, images ranked in first and second place may actually be
farther apart on a scale of complexity than their rank order might suggest). To address this
issue, we need an alternative method which provides a ratio scale in place of an ordinal
scale of numbers, which allows for a more sophisticated statistical analysis. For example,
with ordinal scales, certain useful statistical tests are prohibited such as the mean and
standard deviation. These tests are required in order to be able to calculate the statistical
correlation between the objective and subjective measures of complexity, as in Donderi and
McFadden (2005). Although this methods trial is an informal test procedure and lacks a
thorough statistical analysis, the results are promising and they provide a basis for
continuing the investigation towards a more formal analysis. The following test uses the
same type of CA image with greater resolution, and uses a technique for the subjective
measure of complexity that provides a more sophisticated ratio scale of measurement.

179
Test 1: Monochrome CA and

Random Images
Donderi and McFadden (2005) found that the compressed file size of digital images
is an indicator of subjective ratings of complexity for certain images. The previous trial of
measures suggests that the correlation may apply to non-random images only. The present
test aims to explore whether this correlation holds over a wide range of objective
complexity, from simple order to random patterns. The contextual review in Chapter 2
reveals that studies in empirical aesthetics rarely attempt to investigate the full spectrum of
complexity, but that in fact this might be a useful exercise. Looking at the whole range
provides a valuable overview, especially for mapping the subjectively perceived complexity
of CA images in relation to established schemas of objective complexity. It also enables us
to see whether or not the correlation between objective and subjective complexity breaks
down for random images, as suggested. If it does break down, then Donderis claim must
be qualified with a statement of the range of visual complexity to which it applies, but it is
only by investigating the whole spectrum of patterns that we can establish with confidence
which regions of complexity we are dealing with. If we know where the end-points
(minima and maxima) of the scale are, then we can measure more precisely the subjective
perception of visual complexity in relation to the schemas of Wolfram and Langton, which
are based on the objective properties of CA images.
Method
In this test, we use Donderis file compression and MES technique to gather
estimates of objective and subjective complexity. By sampling from the whole spectrum of
complexity, we are able to build up a picture of the perceptual response to an objective
range of visual complexity based on an informational scale of measurement. Lastly, we use
statistical analysis to determine the correlation between the two measures. These
quantitative results are presented as graphical representations of how perceptions of
complexity vary across a wide range of image complexity.
Test 1: Monochrome CA and Random Images

180
Participants
For convenience, the subjects in this test are Masters students in the College of Art
& Design and the Built Environment at NTU. It was decided to select half from Fine Art
and half from Construction Management courses, as those groups represent the most and
least trained in art at that level of study. The selection is designed to reveal the influence of
art training upon the aesthetic judgements being measured. Eight subjects are selected by
canvassing the students and accepting the first four volunteers from each course. One
student from Construction Management dropped out, and an art-trained member of staff
performed the eighth test. All but one of the subjects were women.
Materials
The test stimuli comprise two sets of 25 black and white images printed at
265265mm on A3 cartridge paper. All images are composed of one million (10001000)
black and white pixels and surrounded by a thin mid-grey border to delineate the edges of
each image against the white paper. At the printed resolution (approximately 96ppi),
individual pixels are just recognisable seen at close range, being just over a quarter of a
millimetre square. The image sets were created with Mathematica programs, one set being
made with cellular automata (Figure 69), and the other from random processes (Figure 70).

181
Figure 69 Set of 25 cellular automata images, sorted in order of increasing file size
(from top left to bottom right).
The random images are designed to show a similarly wide range of detail as the CA
images, but with different shapes. They were made by starting with images in which each
cell is randomly assigned the colour black or white. Next, the image is blurred by a small
amount with a Gaussian function, which introduces shades of grey, and then a threshold
filter applied, which changes each pixel to black or white again based on whether its grey
level is above or below the midpoint. This has the effect of pooling black and white areas,
depending on their average tone. Each random image has different amounts of blurring
applied, with the same threshold filter. The visual effect of increased blurring in this

182
process is similar to that of water condensing on the side of a cold glass tiny droplets
form initially which gradually draw together to form increasingly large beads (Figure 70).
Figure 70 Set of 25 random blob images, sorted by increasing file size (top left to
bottom right).
A large number of each type of image was generated with programs written in
Mathematica and then converted to maximally-compressed PNG files using the freeware
graphics software IrfanView (Skiljan, 2009) together with Ken Silvermans (2006)
PNGOUT plug-in, which enable batch-processing of image files and near-optimum
compression, respectively. PNG was chosen as the file type for measuring objective
complexity as it produces the smallest files for this type of bi-level image (Salomon, 2004).

183
Pairs of similarly-sized files of cellular automata and random images were selected to range
from the smallest to the largest, to make two sets of 25 images that covered a similarly wide
range of objective complexity.
Procedure
Participants rated images for preference and complexity using the method of
magnitude estimation scaling, whereby numbers are assigned to reflect the strength of
perception. Tests were conducted in a well-lit room with only the investigator present,
except in one case where two participants were tested concurrently. Verbal and written
instructions informed the subjects of their task (see appendix A), and a practice run with a
small set of different images was conducted before the tests began in order to familiarise
the participants with the procedure. Participants were given the opportunity to ask
questions and testing began only when they reported they were comfortable.
Each participant rated each of the two sets of images twice over once for
preference and once for complexity, so a test comprised four sets of 25 ratings. The order
of images presented in each run was randomised. The small number of participants meant
that not all of the possible 24 sequence permutations could be used, so the order of tasks
was balanced between subjects, with each task being performed in equal rank frequency.
Rating types were grouped together, with half the subjects rating complexity first, and the
other half preference first. The printed images were presented in sets stacked in
randomised order, and the participants turned over the top sheet to reveal the next image
once they had recorded their rating by writing on a numbered score sheet. It took on
average half an hour to rate the four sets of 25 images. After the task, participants were
invited to comment on the tests.
Results
Because this is the first of three empirical studies in this project, the present section
focuses more on evaluating the viability of methods and determining the subsequent
direction of later tests. The most important issues are to correctly interpret the results, to

184
validate the suitability of research methods, and to identify how to proceed for the
following tests.
The data for all eight participants sets of four ratings are shown in Figure 71. This
figure provides a general illustration of the kinds of differences between individual ratings
rather than a precise statistical analysis. Nevertheless, the numerical scales upon which the
data are represented are the same for each plot, so the display accurately reflects the relative
values collected. Each graph is drawn to the same scale, showing increasing rank order of
image PNG file size (3 to 123 kilobytes) from left to right on the horizontal axis, and
subjective ratings (of the range 0 to 35) on the vertical axis. The top 3 rows are responses
from Construction Management students and the rest from Fine Art, with all except the
top row being women. The diagram is an attempt to follow Tuftes (2001) principles of
information visualisation, with an economical, accurate and elegant presentation of the
salient features of the data set without clutter and extraneous detail.

185
Complexity CA
Complexity R
Preference CA
Preference R
Figure 71 Complete set of un-adjusted subjective ratings (in labelled columns) for
participants 1 to 8 (coloured rows, in test sequence from top to bottom) in Test 1.

186
It is testament to the effectiveness of our visual system that an illustration such as

Figure 71, which contains a total of 800 data points, provides a more comprehensible
statistical description than an equivalent table of values. Provided we know what it is the
diagrams illustrate, this most basic of statistical descriptions, the scatter plot, shows with
clarity any trends in the patterns of data, such as their dispersion, slope or range.
From these scatter plots, we can already see that there does appear to be a positive
correlation between objective and subjective measures of complexity (as one increases, so
does the other), and that the correlation with preference is more variable (there are some
upward and some downward slopes). This variability in individual trends means that we
should be cautious about statistical analyses based on averages of these distributions. For
example, if half the subjects showed a positive correlation and half negative, the averages
would produce a result of zero correlation, which is statistically correct but less useful than
the more detailed overview in Figure 71.
Data Transformation
There is more than one type of variance in this data, however, and depending on the
type of procedure employed, some of them are trivial and some are not. Non-trivial
variance is expressed in the difference between an individuals ratings: Larger numbers
mean greater complexity, and we should not alter the data in a way that affects this
property. But the difference in the range of numbers between subjects is less important. To
explain, imagine a situation where ratings by two participants differ only in that in one set
each value is exactly twice as large as in the other. It might seem appropriate to transform
the data to make it comparable. In this test, the lowest and highest values used across all
subjects are 0 and 35, while the smallest and largest ranges used are 3 (1 to 3) and 30 (1 to
30). What we want is to make the ranges closer together so they can be compared. The
assumption behind these data transformations is that the range of subjective impressions is
approximately equal, even if subjects use different ranges of numbers to represent their
perceptions. However, Stevens (1975) method of magnitude estimation also assumes that

187
the subjects scale of numbers is a ratio scale (i.e. a scale incorporating a zero point, where
numbers express ratios, so that 3 means three times larger than 1). Because of this
assumption, the only permissible transformations are those that preserve the property of
rationality. In this test, we use the mean logarithm of the scores, and in subsequent tests we
use the geometric mean, which is a similar type of transformation.
Figure 72 is a graphical representation of the effect of various types of
transformation on the subjective ratings (note the varying scales of the y-axes). If the raw
data (Figure 72a) were transformed by treating each subjects lowest and highest values as
equal making each maximum 1 and each minimum 0, we would destroy the property of
rationality, because the zero points are moved: Figure 72b shows that each subjects graph
line is stretched vertically to make it touch the top (the value 1) and the bottom (0). In
experimental psychology, a common way of dealing with this type variance in scales is to
use the logarithm of the values a transformation which retains the rationality of the scales
relative to zero, and does not alter the relative positions of the data. The logarithm
transformation (Figure 72c) does not make each scale exactly the same, but it transforms
them to similar ranges and enables averaging of data. Because the logarithm of zero is
infinity, data containing zeroes must be altered before the log transform by adding a small
number. Figure 72d shows an alternative the square root, which seems to offer a viable
alternative to the logarithm transform. For this test and subsequent tests, the logarithm
transform is selected, because it fulfils the necessary criteria and is the same technique
employed by Donderi (2006a).

188
a. Original data
b. Min-Max 0-1
c. Logarithm
d. Square root
Figure 72 Visualising the effect of various data transformations. All graphs represent
the data for subjective ratings of complexity (y-axis) for CA images against rank order
PNG file size (x-axis). (Cf. first column of Figure 71, above, and Table 1, below).

189
Analysis
After the data has been suitably transformed, we can perform calculations that tell us
how closely the trends of the subjective data match the trends of the objective measures,
that is, the correlation between them. According to a standard text on statistics for
psychologists (Howell, 2007), the most common form of correlation, and the appropriate
method to use in this case, is Pearsons product-moment correlation, which is expressed as
an r value a measure without units that can range from -1 to +1. A positive value
indicates a correlation in which one variable increases with the other, a negative value
indicates a variable that decreases as the other rises, and a value close to zero means that
there is little correspondence between the two. Table 1 shows how the data
transformations in Figure 72 affect these r values (the first column corresponds to the
graphs in Figure 4).
Mean Complexity
Mean Preference
Data
Transformation
CA
Random
CA
Random
None
0.700
0.827
0.571
-0.086
Min-Max 0-1
0.717
0.778
0.621
-0.056
Logarithm
0.680
0.753
0.515
-0.034
Square root
0.651
0.781
0.468
0.028
Table 1 Influence of various data transformations of subjective ratings on correlations

with PNG file size.
It looks like the worst transformation (MinMax 01) gives the best correlation,
but this should not mislead us into thinking that it is the correct one to use. The most
important aspect of the r values in Table 1 is not the minor differences between data
transformations, but their similarity. It tells us that there is a strong correlation between
subjective and objective complexity for both image sets, and that there is a weak correlation
between file size and preference for the CA images. There is almost no correlation with
preference for the random images, but we noted earlier that actually there are almost equal

190
numbers of negative and positive correlations amongst the subjects, which therefore cancel
each other out in these statistics.
The most important aspect of the r values in Table 1 is not the minor differences
between data transformations, but their similarity. It tells us that there is a strong
correlation between subjective and objective complexity for both image sets, and that there
is a weak correlation between file size and preference for the CA images. There is almost
no correlation with preference for the random images, but because there are almost equal
numbers of negative and positive correlations amongst the subjects, they cancel each other
out in these statistics.
File Types
Image file types differ in the way they handle pictorial data, and use various methods
of encoding and compression, so the file size of a given image will vary from one type to
the next. It is likely that nearly-optimal compression is a favourable condition for
correlation with subjective complexity ratings, but no one has yet undertaken an exhaustive
trial of the different image file types and compression methods for this purpose, so it
remains to be seen which is best for this job and why. This matter cannot be fully explored
in the present study, but the results should be able to give an indication of the relative
performance of various file types. The correlation of five different image file types with
subjective ratings is shown in Table 2:
File Type
mean(log) Complexity
mean(log) Preference
CA
Random
CA
Random
PNG
0.608
0.740
0.404
-0.045
JPEG
0.547
0.813
0.118
-0.155
TIFF (LZW)
0.607
0.717
0.294
-0.038
ZIP (PNG)
0.607
0.741
0.404
-0.060
ZIP (JPEG)
0.644
0.806
0.415
-0.152
Table 2 Correlation (r values) of subjective ratings and file size for various file types.

191
What is most apparent in Table 2 is the general agreement amongst file types, with
much stronger correlations for complexity ratings than for preference. Donderi and
colleagues found the best correlations with lossless JPEG images compressed as ZIP files,
and that is also the case here (N.B. ZIP is not an image file type, so it cannot be used to
display graphics it is used by Donderi to provide a second process of compression.) PNG
was chosen as the working file type for creating the stimulus images because they gave the
smallest files, and JPEG and TIFF were selected for the test because they are commonly
used in graphics applications. ZIP files were made of uncompressed PNG images and
lossless JPEGs.
Lossless JPEG had to be used because lossy compression was found to degrade the
images by producing greyscale artefacts which blurred their crispness, and this would have
made for an unfair comparison with the other image types. JPEG was expected to perform
poorly in this study as an indicator of subjective complexity because it is known to give
poor compression ratios for bi-level (black and white only) images. Actually, it shows the
weakest correlation with the CA images, but the strongest for the random images. This
could be explained by the fact that JPEG encodes two-dimensional regions of image data,
unlike the other file types here, and so the blobby nature of the random images is
captured more effectively than the intricate CA images. This is also reflected in the
difference between the total sizes of image sets using different file types (Table 3).
File type
Size of image folder (MB)

CA
Random
PNG
1.26
1.36
JPEG
24.9
14.0
TIFF (LZW)
2.07
1.72
ZIP (PNG)
1.18
1.21
ZIP (JPEG)
16.5
13.6
Table 3 Total sizes of all 25 images in each set for various file types.

192
The difference between the folder sizes of PNG and ZIP (PNG), together with the
similarity of their r values in Table 2, indicates that the type of image encoding, rather than
the compression method, which most affects the rank-ordering of images and their
subsequent correlation with subjective ratings (these two file types share image-encoding
methods, but use different methods of compression).
Preference and Complexity Ratings
Table 4 shows the relevant correlations between subjective ratings. There are weak
correlations between preference and ratings of complexity for both the image sets, but they
differ in direction: For the CA images, preference increases with complexity, but for the
random images it decreases, meaning that the participants preferred the more complex CA
images and the simpler random images. There is a strong correlation for ratings of
complexity between the two sets, which is encouraging because it suggests that there is a
relatively stable relationship between objective and subjective measures even with visually
dissimilar types of image. The statistical significance of these correlations has been avoided
thus far, but will be taken up in the discussion of results.
Mean (log) Rating
Correlation
Mean (log) Rating
Complexity CA
0.34
Preference CA
Complexity Random
-0.27
Preference Random
Complexity CA
0.80
Complexity Random
Preference CA
-0.26
Preference Random
Table 4 Correlation between subjective ratings.
Average Responses
Figure 73 shows the averages (mean) across all participants for the four subjective
ratings compared to file size. At this relatively crude stage of analysis, we can see that
subjective ratings of visual complexity generally correspond with the file size of the images.
This is true of complexity ratings for both sets of images, whose results are fairly similar,

193
whereas preference ratings of the two sets are quite distinct in shape but indicate a weaker
correlation.
Average ratings, mean logarithm
Complexity CA
Complexity R
Preference CA
Preference R
Rank order, file size

Figure 73 Mean log ratings plotted against rank order PNG file size.
Influence of Art Training
Table 5 shows how the correlations with file size differ between non-art-trained (MA
Construction Management) and art-trained (MA Fine Art) subjects. Ratings of complexity
for both image sets do not differ greatly between the two sets of subjects, but their
preference does. The non-art-trained subjects show a fairly strong preference for the more
complex CA images and the simpler random images. In contrast, the art-trained subjects

194
have only a slight preference for the more complex of both image sets. Figure 74 illustrates
these different ratings, in the same layout as Table 5. The greatest visible difference is that
for preference ratings of random images (right-most column), which slopes downwards for
the non-art-trained subjects and upwards for the other group.
Group
mean(log) Complexity
mean(log) Preference
CA
Random
CA
Random
Construction
Management
0.760
0.699
0.567
-0.790
Fine Art
0.439
0.705
0.159
0.328
Table 5 Influence of art training on judgements of complexity and preference.

Complexity CA
Complexity R
Preference CA
Preference R
Figure 74 Visualisation of the different responses between participants from

Construction Management (top row, red) and Fine Art (bottom, blue). x-axis: mean
log scores; y-axis: rank order file size.
The results show a greater consistency between the groups ratings of complexity
than amongst ratings of preference. The trends revealed by these analyses are comparable
to the findings of Locher, Smith and Smith (2001), who reported a similar tendency for art
training to affect preference of images more than perception of visual properties such as
complexity.

195
Inferential Statistics
The following statistical analyses differ from those of the previous section in that
they allow us to make inferences from the sample data back to the population at large.
These statistical methods can also tell us the degree to which the inferences are valid, that
is, the proportion of the correlation that can be accounted for by the particular factors
under scrutiny. Regression analysis allows us to find the mathematical relationship between
variables, so that we may make predictions of one based on the values of the other. We
find the mathematical relationship by attempting to find the line of best fit for the points of
data. Figure 75 shows some examples of regression analysis for ratings of complexity of the
CA images. Each graph uses a different type of equation as its model, and the regression
analysis provides estimates of their parameters, which are plotted as red lines through the
data points. The measure of how closely the estimated line fits the data is called the
coefficient of determination, or the r squared value. The r2 values in these graphs are
adjusted to compensate for the number of parameters used. Given a measure of the
objective complexity of an image, these functions can be used to estimate its perceived
complexity or preference.

196
Data points
Linear
(Complexity CA)
r2 =
0.34
y = 0.75 + 1.610-6x
2nd order
3rd order
r2 = 0.45
r2 = 0.72
y = 0.70 + 5.710-6x 3.310-11x2
y = 0.57 + 0.00002x 3.710-10x2 + 3.710-15x3
Figure 75 Examples of curve-fitting data by regression analysis, using successively

larger-order polynomials. Plots show mean log complexity ratings (y -axis, no units)
against PNG file size (x-axis, in bytes).
Figure 75 shows the lines of best fit when the data is arranged by absolute file size,
but there are alternative ways of arranging this data: It can be sorted into rank order of file
size (as in Figure 73) or we can use the logarithm of the file sizes. The latter method is the
one adopted by Donderi as a measure of image information content (in line with
Shannons quantitative definition of information), and the lines of best fit for this method
are illustrated in Figure 76.

197
Data points
Linear
r2 =
(Complexity CA)
0.56
y = 0.065 + 0.174x
2nd order
3rd order
r2 =
r2 = 0.68
0.69
y = 4.76 + 2.40x 0.253x2
y = 9.94 + 6.02x 1.09x2 + 0.064x3
Figure 76 Lines of best fit for mean log complexity ratings (y-axis, no units) against log
PNG file size (x-axis, in bytes). R squared values are adjusted.
The logarithm transform of file sizes produces more consistent results than the untransformed data. The third-order polynomial used in the regression analysis produces
quite different curves for the data depending on whether a transformation is used. The
logarithm has the effect of squeezing the points towards the right, which irons out the dip
in the line. Figure 77 shows the lines of best fit for each of the four different subjective
ratings, and provides a succinct description of the results of this test.

198
Complexity CA
Complexity R
2nd order
2nd order
r2 = 0.69
r2 = 0.88
y = 4.76 + 2.40x 0.253x2
y = 5.21 + 2.54x 0.260x2
Preference CA
Preference R
Linear
r2 =
2nd order
0.18
r2 = 0.35
y = 0.593 + 0.065x
y = 6.23 2.49x + 0.278x2
Figure 77 Lines of best fit for the four ratings of test images. Mean log ratings (y-axis,
no units) vs. log file size (x-axis, in bytes).
Statistical significance
Correlation coefficients (r values) are usually given with reference to their statistical
significance (p value), which tells us the probability that the result could have occurred by
chance alone. The smaller p is, the more confident we can be that the results are due to the
factors under study. The necessary degree of certainty varies from study to study,
depending on their aims and purposes, but in general the more data we can acquire, the
more we can be certain of the results. In the present study, we have found a quite strong
correlation between objective and subjective measures of visual complexity, but the small

199
number of participants means that we cannot be very certain about its significance. One
way to increase the significance of the results, then, is to repeat the tests with more
participants. This is the plan for the next set of tests, which introduce colour to the stimuli.
Summary
Using the logarithm of both objective and subjective measures of complexity, the
lines of best fit are consistent with one expected outcome of this pilot test. It was predicted
that Donderis linear correlation between the two measures of complexity would break
down for the most complex images (i.e. that the line would begin to curve downwards on
the right), and that this might not be revealed unless a wide range of visual complexity is
tested. The results of this test cautiously confirm this hypothesis there is distinct inverted
U shape to the graphs of perceived complexity vs. measured complexity in Figure 77.
Perhaps surprisingly, the results are similar for the two different sets of images that vary
greatly in appearance. Preference, on the other hand, shows much more variation, which is
not surprising given that the workings of aesthetic taste are dependent on many more
variables than were accounted for in the present study. Whereas the complexity ratings are
similar for both image sets, there is a marked difference in preference between the two.
Also, between the two groups of participants there is greater disparity in preference than
there is for complexity. A tentative conclusion is that art training does not affect perception
of complexity, but it does alter preference for levels of complexity.
The results of Test 1 support this hypothesis: Ratings of perceived complexity form
an inverted U shape when correlated with the objective measure of complexity based on
information content. However, the graphs of complexity ratings and file size in Figure 77
describe mainly the left portion of this inverted U shape. There are two alternative
explanations for this result; either the perception of random images does not function as
predicted by Donderi (2006) or the stimulus images are not random enough to see the
predicted effect. Given that the stimuli are fairly low-resolution black and white images,
which therefore fall short of the perceived complexity of artworks such as Pollocks drip

200
paintings, the latter alternative is the more likely. The implication for the following test is
that the objective complexity of stimuli should be increased and that we should include
images that are more random. In other words, we should raise the level of overall
complexity in the image set, and widen the spectrum of complexity that it represents. In
terms of the fundamental variables of visual complexity identified in Chapter 2, this means
that we increase the quantity and the variety of picture elements and widen the scale of
orderrandomness.
Discussion: Visual Resolution
The results of these tests indicate that the objective and subjective measures of visual
complexity correlate linearly only for the least complex images (the upward-sloping left
portion of the curve on the complexity graph). Beyond a certain threshold (the peak of the
curve), as images become more random they are rated as less complex (the right-hand side
of the curve). The stimuli in this study had quite a low printed resolution; individual pixels
were clearly visible. It is thought that by increasing the resolution, the most random images
will appear even less complex, because as the image resolution approaches the limit of our
visual acuity, random images appear smoother or more uniform, and would appear to be
more simple. Figure 78 illustrates this phenomenon. Each image is composed of random
blobs (created using the method in Test 1 described above), but they vary in terms of scale:
the images increase in resolution towards the right. What is noticeable is that the images
appear smoother with increased resolution. We can imagine the series progressing towards
a uniform grey that would appear to be a very simple image, even though we know it is
actually composed of a complicated arrangement of different elements.
Figure 78 The effect of increasing resolution of random blob images.

201
The point at which the increased resolution appears uniform is dependent on the
eyesight of the participant, principally on the property of visual acuity (the ability to resolve
detail). Since individual acuity will vary we cannot specify exactly where this threshold lies,
but given the common standard measure of 20/20 acuity we can calculate the necessary
printed resolution for creating stimuli that approach this threshold. In effect, the relatively
low resolution of stimuli in Test 1 limited the approach towards the perception of
uniformity. This provides another explanation for the result that only the left portion of the
predicted inverted-U-shape correlation is observed. We propose to use a higher resolution
in subsequent tests, as this should lead to a more pronounced effect on the correlation
between the two measures. The smoothing effect means that the most random images will
be rated less complex than the corresponding objective measure as resolution is increased.
In theory, then, a random image whose resolution (spatial frequency) is greater than our
visual acuity should be indistinguishable from the simplest possible uniform image. We can
infer that subjective complexity forms a continuous spectrum whose end-points meet,
unlike the objective measure of complexity which exhibits the clearly opposing poles of
order (or uniformity) and randomness. In this sense, we could describe the objective scale
of visual complexity as linear, and the subjective scale as circular like a colour wheel. This
idea is taken up in the concluding chapter.
Test 2: CA and Random Images in

Colour
Test 2 uses random and cellular automata images again, but this time with greater
resolution and in colour instead of black and white. In addition to increasing the quantity
of picture elements by increasing the resolution of the stimuli, the introduction of colour
marks a qualitative difference it increases the variety of picture elements and raises the
overall level of complexity of the image set compared to Test 1. The present test has
images that are more random, but the range of patterns is approximately the same as
before, representing all four classes of complexity. In this way, we also widen the range of
Test 2: CA and Random Images in Colour

202
visual complexity present in the stimuli. The benefit of this action is threefold: It allows for
a corroboration of the initial findings, it further tests the validity of the method to cope
with more complex images, and it enables a more detailed examination of the hypothesis
concerning the perception of randomness.
Method
Materials
To create the image set for this test, another program was written in Mathematica. The
program generated random images and cellular automata images not the elementary CA
with two colours or states that were used before, but CA with four colours. The number
of rules in this four-colour CA is vastly larger than the elementary CA, and so it is not
possible to generate images for each rule. Instead, the rules are chosen at random by
generating a number between zero and the number of possible rules of this CA
(approximately 3 1038). Normally, different CA rules are represented with the same
colouring, but for the purpose of this test we want to avoid a bias towards a particular
colour palette. Therefore, the program that generated the stimulus images also assigned
colours randomly to each image. The nature of the CA patterns means that they have
between two and four colours each.
After generating a few hundred images, a set is created by sorting the compressed
image files in order of size and then making selections spaced roughly equally across the
range. The set selected for this test is illustrated in Figure 79. Two sets of the stimulus
images were printed at a size of 20cm square on medium-weight white card. Each image
was labelled on the reverse side with a reference number for identifying images in the tests.
These labels were assigned randomly, and the same labels applied to both image sets.

203
Figure 79 Set of stimulus images in Test 2, arranged in order of increasing file size
from top left to bottom right

204
Participants
Volunteers to participate in the tasks were recruited from postgraduate students in

the College of Art & Design and the Built Environment. 20 people took part in the tests,
including art-trained and un-trained students.
Procedure
The participants task is the same as the previous test: Each participant rated a set of
images for complexity and preference by writing scores on a sheet. The tasks took place at
desks in the postgraduate research office with plenty of room and illumination. The
sequence was alternated between successive participants, so half rated complexity first and
vice versa. Scores were written in a box labelled with a number corresponding to that on
the back of each printed image. The sets of images were shuffled for each participant to
obtain a random presentation sequence.
Results
The test results are presented graphically in Figure 80. The top two images show the
complete sets of ratings for complexity and preference, visualized with higher scores as
hotter colours. The average (geometric mean) ratings are illustrated in the central pair of
images, and at the bottom are plotted in graphs against rank order file size. The pattern of
colours and the shape of the graphs indicate a degree of correspondence between
subjective ratings and image file size, with a stronger correlation for complexity ratings than
for preference.

205
Preference
Complexity
Rank order, log JPEG file size
Rank order, log JPEG file size
Figure 80 Test 2 results for complexity (left) and preference (right). Top: Complete
data colours represent rating values (blue = low, red = high), participants scores are
in rows and stimulus images are represented in columns in order of file size. Middle:
Geometric mean ratings. Bottom: Plots of geometric mean ratings against rank order
JPEG file size (no units).
Analysis
The statistical analysis reveals a moderate correlation between average (geometric
mean) subjective ratings of complexity and preference (r = 0.486). The correlation of
preference and objective complexity (file size) is slightly weaker, varying with the file type
(see Table 6). Similarly, the correlation with complexity ratings varies with file type. TIFF
and JPEG file types show a statistically significant correlation with complexity ratings
(TIFF: r = 0.547; JPEG: r = 0.542, d.f. = 18, p < 0.05) and only JPEG shows a significant
correlation with preference ratings (r = 0.494, d.f. = 18, p < 0.05) for this sample size.
d.f. = 18, p < 0.05
PNG
TIFF
JPEG
Complexity
0.267
0.547
0.542
Preference
0.279
0.389
0.494
Table 6 Correlations for Test 2.


206
If, like Besner (2007), we split the set of images into the smallest and the largest files,
we find that the correlation with first half of files (smallest files with simpler images) is
much stronger (r = 0.814) than that with the second half (r = -0.078). This shows a
dropping-off of the correlation curve for the largest files (most random images). Figure 81
shows the results of regression analysis for complexity and preference ratings with all three
file types. As in Test 1, we have evidence of the inverted U shape correlation, which is
Preference
Complexity
stronger between perceived complexity and file size than for preference.
TIFF
PNG
JPEG
Figure 81 Regression analysis (2nd order polynomial), mean scores (no units) vs. log
file size (bytes).
Instead of expressing image file complexity using the logarithm of the file size, as
Donderi & McFadden (2005), we can use a measure based on file compression. The
compression ratio of the image files is the ratio of the size of the compressed file to the
original uncompressed file, giving a relative value between 0 and 1 (with no units of
measurement, because they cancel out). The value of using the compression ratio is that it
provides a scale from a maximum of order (maximum compressibility) to a maximum of
randomness (minimum compressibility). The results are similar to the other plots, again
showing the same shape correlation (Figure 82). The compression ratio scale reveals more

207
clearly which areas of the informational scale of complexity are represented by the stimuli
in this test and which are omitted; clearly, we have few representative samples of maximum
randomness, probably because of the limited number of colours used in the stimuli (a
maximum of four). The graphs seem to suggest that the threshold of perceived complexity
(the peak of the curve) is quite low a compression ratio of approximately 0.15 but it is
possible that the images are artificially low on this scale: The limited number of colours
restricted the file size of even the most random images, which meant that the range of sizes
did not approach the upper limit of the compression ratio.
Complexity
Preference
Figure 82 Regression analysis (2 poly) for ratings of complexity and preference (y-axis,
mean scores, no units) against compression ratio of PNG images (x-axis, no units).
Effect of Art Experience
Comparing ratings between art-trained and untrained groups of participants, we find

no significant difference in complexity ratings, but there is a difference in preference
(Figure 83). A between-subject T-test (two-tailed) for art-trained and untrained participants
shows no significant difference in complexity ratings between the two groups (t = 4.21, d.f.
= 96, p > 0.1). However, there is a significant difference in preference ratings between the
groups (t = 7.21, d.f. = 96, p < 0.01), with art-trained participants giving higher scores on
average (M = 36.07, SD = 16.80) than untrained participants (M = 31.86, SD = 10.82).
Despite there being no significant difference in average scores for perceived complexity,

208
art-trained participants show a stronger correlation between complexity ratings and log
PNG file size (r = 0.75) than untrained participants (r = 0.44).
Figure 83 Comparison of geometric mean ratings between art-trained (blue) and

untrained (red) participants, against rank order file size (x-axis, no units).
Discussion
In this study, class 4 CA images are found to be among the highest-rated as
subjectively complex. We also find an inverted U correlation between objective and
subjective complexity. This finding supports Wolframs suggestion (2002, p.559) that there
is a threshold of visual complexity beyond which we fail to perceive underlying order or
pattern. The suggestion is also supported by the result that the class 3 chaotic CA images
(which are actually deterministic but which appear random) are rated lower in subjective
complexity than the class 4 patterns, even though they are larger in compressed file size.
These findings constitute empirical evidence for the hypothetical threshold of perceived
complexity, and thus make a contribution to knowledge in the field of empirical aesthetics.
The finding that between the two groups of participants there was only a significant
difference in preference suggests that art-training has a greater effect on preferential
judgement than on judgements of perceived complexity. If we understand art education
and practice to involve the cultivation of aesthetic judgement, we might reasonably expect
art-trained people to be more critical, and to rate artworks lower on average than
inexperienced viewers. So, why art-training should give rise to higher ratings of preference

209
overall, and why it does not appear to affect perceptions of visual complexity, is unclear.
We may gain more insight into these questions in the following tests, where we repeat the
participant grouping for images of art and design. Beside ratings of complexity and
preference, a further rating task is added in which participants score the perceived artistic
quality of the images. It is hoped that introducing this task may clarify the issue by
separating judgements of preference what Kant called a judgement of agreeableness,
which is an interested judgement from judgements of quality, which in Kantian terms is a
disinterested aesthetic judgement. This is a common distinction in art; we may have high
regard for an artwork in the canon, but we may at the same time dislike its appearance (for
example, the work of George Grosz, Philip Guston, and Jake and Dinos Chapman).
In tests 1 and 2, PNG was chosen as the working file type because it produced the
smallest files for the type of two-colour image used as stimuli, whereas JPEG files of the
same quality were 10 to 200 times larger. Given the fact that JPEG does not generally cope
as well as PNG with bi-level (black and white) images (Salomon, 2004), it is perhaps
surprising that on average the strongest correlation in Test 1 is given by using ZIP
compression on lossless JPEG files. Similarly, JPEG compression performed comparably
with the other file types in Test 2. These results are in agreement with Donderi and
McFadden (2005) and Besner (2007). It is possible that the reason for the success of JPEG
results lies in a combination of near-optimum (ZIP) compression together with operations
that are similar to our own visual processing, namely JPEG encoding, which takes
advantage of human perception of colour and tone. From these results, we may make the
tentative suggestion that JPEG encoding may parallel the low-level visual processing of the
retina, while ZIP compression corresponds to cognition taking place in areas of the visual
cortex. Obviously, to corroborate this suggestion would require further research beyond
the scope of the current investigation.

210
Summary
The results indicate that preference is highly variable, and cannot be reduced to a
simple function of image complexity. It would appear that preference has something to do
with the feel of the images in Test 1, because the most preferred tended to be those with
smoother edges, suggesting a possible link to tactile preference. In test 2, with a different
set of images, it is difficult to see any strong factor that influences preference, but informal
interviews with participants suggests that colour preference was amongst the main
influences. In the next tests, the motivations behind individual preferences are elaborated
with the use of more formal interviews and questions after testing.
Unlike the conclusion of Besners (2007) similar investigation into the correlation
between objective and subjective measures of visual complexity for fairly complex images,
some of the results of this study do appear to support Wolframs idea of a threshold of
visual complexity: The statistical correlations and graphical presentations reveal that
subjective ratings peak in the centre of the scale of objective image complexity, and begin
to fall as images approach randomness. Even the file-size measure of visual complexity is
equivalent to Kolmogorov complexity or AIC, which as we know does not quite follow our
intuitive understanding of complexity, it nevertheless offers a convenient and robust
method of dealing with the problem of aesthetic measurement. The results of tests 1 and 2
support the proposal to use the file-compression measure of visual complexity with images
from art and design.

211
Chapter 5
Visual Complexity and Aesthetic
Judgement
Introduction
The previous tests showed that the chosen methods are feasible to measure visual
complexity in general, using artificial stimuli. The present test finds out whether these
methods work with images of artefacts from art and design. Using real work is problematic
because its complexity and variety demands a sophisticated yet robust method of analysis.
The method must negotiate the dual constraints of artistic integrity and empirical validity.
Is it really possible to measure the complexity of art with this technique? This chapter
documents the design, implementation and analysis of the solutions to these problems.
Primarily, this test is concerned with visual complexity at the computational level of
description. It is beyond the present scope to explore how the different stages of visual
processing contribute to the perception of complexity in biological or psychological terms.
Rather, the plan is to measure visual complexity and to see how it corresponds with
aesthetic perceptions of artworks. The central question is: Does the file compression
measure of complexity correlate with subjective aesthetic judgements of contemporary
visual art and design? Aesthetic judgements are measured by asking participants to quantify
their perceptions of a variety of images, and a statistical analysis determines how these
ratings correlate with each other. Three common image file types and various compression
algorithms are evaluated as measures of visual complexity, together with subjective
judgements of complexity, preference and quality. In order to determine the criteria that lie
behind these aesthetic judgements, we also conduct interviews with participants and
perform a qualitative analysis of this data.
Introduction
212
Test 3: Images from Art and Design

The previous tests used synthetic stimuli based on cellular automata and random
images, firstly in black and white and later in colour. The black and white image sets have a
limited variety of elements (only black or white pixels), whereas the later tests have greater
variety with up to four colours in each image. The present test uses samples from a wide
range of art and design practices. To allow for a fair reproduction of the artefacts, the
digital image files use the RGB colour system, which means that the variety of pixels is
vastly greater than the previous tests (the RGB system has around 16 million colours).
Together with the increased dimensions of the images in test 3, this increased variety
means that the number of possible images in the current setup is much greater than before,
and so our stimuli represent a much smaller sample of the spectrum of complexity. As long
as we try to sample as wide a range as possible, this should not present a problem.
These tests are designed to progress towards greater complexity of stimuli by
gradually increasing the parameters of the variables that contribute to visual complexity.
Once again, the technique of magnitude estimation scaling (MES) is used to gather
subjective ratings of complexity and preference for images, which are then compared to an
objective measure of complexity based on the compression of the image files. Mapping the
responses to a wide range of complexity in Tests 1 and 2 provides a useful reference for the
present test. In Test 3, the aim is to find out: 1) how different types of visual art on the
spectrum of visual complexity; 2) how this complexity relates to aesthetic judgements of
preference and quality; and 3) whether those judgements can be related to and predicted by
measures based on image file properties. This test has two parts, which use the same
stimuli in two different locations, in order to find out whether this has any effect. The first
test (3a) takes place in a psychology laboratory, and the second (3b) in an art gallery.
Hypotheses
Many empirical studies of visual complexity have focused on non-art images. In
addition, works of visual art tend to be more complex than the sorts of stimulus images
Introduction
213
used in those studies, for example Snodgrass & Vanderwart (1980). These two
circumstances guide the design of this test, since they point to gaps in knowledge about the
aesthetics of high complexity in general (especially when this complexity approaches
randomness) and about the perception of complexity in art in particular. This test aims to
fill these gaps in knowledge. The hypotheses are:
1. Image file compression will correlate with judgements of complexity.
2. Familiarity with art and design will correspond with lower ratings of complexity.
3. The test environment will not affect aesthetic judgements.
Besides these central questions, this test also aims to evaluate measurements of visual
complexity in art: Will the method that worked for Donderis radar charts and the cellular
automata patterns in the previous tests still be effective for images of contemporary art and
design? We also evaluate the comparative performance of data compression algorithms in
three common image file types (and compression algorithms) as follows: TIFF (lossless
LZW), PNG (lossless Deflate) and JPEG (lossy JPEG). One feature of the test is that once
the subjective data has been collected it possible to carry out retrospective analysis on the
image files. We also have the opportunity to apply image transformations, such as edgedetection algorithms, and to see how this affects file compression and its correlation with
subjective data.
Test 3a: Psychology Laboratory

Method
Materials
To gather visual material besides my own artwork, invitations were sent to artists and
designers which asked if they would lend a piece of work for use in these tests. Fifteen
people responded positively to the invitations, with a few respondents providing more than
one piece of work. Because this test uses as stimuli artefacts from art and design, we will

214
have to use scanned digital images for the objective measure, but for the subjective ratings
we could use either reproductions or the originals. Experiments by Locher, Smith & Smith
(2001) reported little difference for ratings of complexity across different formats of
reproduction, and almost no effect of art training either: Locher et al called this effect
facsimile accommodation, and it suggests that perception of collative aesthetic properties
such as complexity is less affected by the format and location of an artwork than we might
have thought. These findings support the decision to use digital reproductions of artworks
for both the objective and subjective measures in the next tests. Using the same digital files
as the basis of the objective and subjective tests also makes for a fairer comparison
between the two measures (as opposed to using the original artefacts for subjective ratings
and digital reproductions for objective complexity measure).
There are a few important criteria for the representation of this material as stimuli:
Firstly, they must represent a wide range of contemporary work, so that the test results are
relevant to todays visual practices. This was fulfilled by using painting, drawing, woven and
printed textiles, photography, and digital prints loaned from contemporary practitioners,
including university students and staff, professionals, and two children. There are equal
proportions of images in landscape and portrait (horizontal and vertical) alignment.
Secondly, the images must show the artefacts without cropping or re-sizing. This
requirement imposes a constraint on the size of work that may be used, which is
determined by the size of the scanner in this case, A2 paper size (420 594 mm).
Thirdly, the resolution of the images is an important factor, because it affects the measure
of complexity and because it also influences the quality of the visual array perceived.
300ppi is used for scanning, and 150dpi for printing, since the average human visual acuity
is also around 150dpi (20:20 vision means the ability to resolve two points separated by one
minute of arc, which is approximately equal to perceiving one dot at 150dpi when viewed
at the closest focusable range of about 6 inches).

215
Stimuli
The stimulus images in this test comprise fifty colour images of art and design work,
scanned at 300dpi and printed life-size on A2 paper. The decision to use printed images,
rather than displaying the original works, was made in order to allow for a fair comparison
with the file-size measure of image complexity. Given that one aim of the test is to
correlate objective (file-compression) and subjective (MES) measures of complexity, it is
important that the two measures are based on the same thing in this case, digital image
files are the basis of both the file-size measure and the printed stimuli. For some pieces
such as the resin-based paintings, this could have been a particular difficulty since their
appearance changed dramatically with the lighting conditions. The decision to use prints as
opposed to original artefacts is also supported by Marrs account of the principal physical
factors in the visual perception of an image: Marr names geometry, reflectance, illumination
and viewpoint as factors that determine the intensity values of an image (1982, p.41). Using
prints effectively limits the potential variation in reflectance that would be caused by the
variations in viewpoint and illumination that naturally occur in the context of viewing.
Admittedly, this variation cannot be eliminated entirely, but it is constrained to an
acceptable level primarily by using prints, and secondarily by printing onto paper with a
semi-gloss finish. The minimisation of variation in reflectance levels is confirmed by
viewing the prints from different angles and in different locations relative to the sources of
illumination.
There was a problem with scanner used (which was the only one of that size
available), which introduced a colour cast and produced scan lines caused by variations in
the scanning heads. To correct this problem, I created a blank scan of the white scanner
lid to produce an image of the imperfections only, and then subtracted it from each image
by processing in Photoshop. This seemed to solve the problem the printed colours looked
more like the artefacts and it removed the scan lines.

216
Most of the artefacts have a specific orientation they have a right way up, but
some, such as the textiles, do not. Of the stimulus images with a particular orientation,
there is an approximately equal number of horizontal and vertical. The remaining images
were laid out so that the total numbers of landscape and portrait orientations were equal.
The scans and prints have fixed dimensions, but the size of the art and design objects
varies some of them fill the area and some do not. The images that do not are printed
with white borders. If we are to measure the images objective complexity by file size, then
we face a decision about whether or not to take account of these borders, and this depends
on which type of complexity measure is chosen. On the one hand, because the borders are
composed only of uniformly coloured pixels, these areas are highly compressible, which
means that they would be expected to add little to the overall file size and so might be
safely disregarded for the absolute file size measure. But on the other hand, if we are to use
the compression ratio instead of the absolute file size, then these borders may affect the
result. This problem and its solutions are discussed in the presentation of results.
Participants
Thirty-one people participated in these tests, 19 male and 12 female. Staff and
students from the schools of Art & Design and from Psychology at NTU were recruited by
email or by personal invitation, and some personal friends also volunteered to take part. Of
the 31 participants, 6 were people who had contributed artwork for the tests. From a
scientific point of view, this situation may appear to be a potential confounding variable. It
is justified in this case with the argument that it reflects the everyday processes of art and
design: art practitioners commonly evaluate our own work against that of others.
Setting
The setting for Test 3a was a laboratory used for teaching psychology. It contained
ten large hexagonal tables, which provided plenty of room to lay out five prints on each
table, and enough space for participants to walk around and view the images. There was a

217
computer on which data was entered, and a desk and seating for administrative work.
Interviews with participants were held at this desk in the room.
Because many of the stimulus images are printed with white borders, the tables
darker surface made the borders an intrusive visible feature of the stimuli. The problem is
that this noticeable white border added to the visual array present in the stimuli. To get
around the problem we could have cut out the images of the artefacts, but it would have
resulted in stimuli of different sizes, reducing their consistency and making the analysis
more difficult. The solution was to cover the tables in white paper to provide a background
of the same colour as the borders, which made them almost unnoticeable.
Variables and Units of Measurement
Of the many variables in this test, some need to be controlled whilst others are
recorded. The principal dependent variable under measurement is visual complexity, and
since this property can be affected by more primitive attributes such as image size and
resolution, these need to be made consistent. Here is a summary of the test variables:
Dependent: Objective: complexity (measured by image file compression, as file size
and compression ratio). Subjective: complexity, preference and quality (measured by MES).
Independent: Image file type/compression algorithm (TIFF/LZW,
PNG/DEFLATE, JPEG/JPEG); test environment (psychology lab, art gallery); art

experience (trained, untrained).
Controlled: Digital image dimensions (7016x4961 pixels) and scan resolution (300
ppi); colour model (24-bit sRGB); image print size (594 420 mm) and print resolution
(150 dpi).
The file size measure of complexity is expressed in bytes, but compression ratio has
no units because they cancel out. The subjective MES data are similarly dimensionless

218
because there are no standard units for magnitude estimation of aesthetic properties (unless
we convert the data to z-scores, which are units of standard deviation from the mean).
Procedure
Participants were given written and verbal instructions before conducting the test.
Each person was informed that their task was to rate the images for the various aesthetic
properties by quantifying their perceptions (MES). Usually with magnitude estimation
scaling, a standard is provided and a score assigned as a reference. In this test, participants
were instructed that the minimum score of zero corresponds to a complete lack of
perceived complexity, preference or quality, such that for the complexity measure a score
of zero would apply to a blank image (i.e. the simplest possible image). Participants
recorded their ratings for the images by writing numbers on score sheets. Explicit
definition of the properties was not given, only clarification of the sense in which the word
was meant. For example, the most frequent request was for clarification of quality; it was
explained that it meant how good work of art it was perceived to be (i.e. quality of the
artwork, not the print quality). Participants were free to choose a rating scale, and were
allowed as much time as needed to evaluate each image. Including interviews with
participants after the task, the procedure took from half an hour to an hour per person.
Results
A graphical representation of the average ratings of complexity, preference and
quality is presented in Figure 84, in which the images are arranged in order of geometric
mean scores. With this arrangement, it is difficult to see any patterns in the data, but it
allows us to identify particular artworks. In regard to my creative experiments with images
approaching the threshold of visual complexity, the majority of these images in this test are
amongst the highest rated for perceived complexity. As a comparison to these subjective
ratings, the same images are arranged in order of JPEG file size in Figure 85.

219
Figure 84 Test 3A: Images arranged by average ratings of complexity (top), preference
(middle) and quality (bottom).

220
Figure 85. Images arranged in order of JPEG file size (from largest at top left to
smallest at bottom right).
From this arrangement, it is apparent that the smallest files are those with the largest
areas of homogeneous colour, and also that these tend to be the ones with large white
borders. The image in Figure 86 is the fourth-smallest compressed file because of its large
areas of flat colours mainly blue and black. The image three places to its left in Figure 85
is made of multiple smaller versions of the same picture, and even though this one has a
larger white border, its greater detail makes it a bigger file. It would be useful to find an
objective measure of image complexity that ignored the white border in the same way that
the subjective ratings appear to do. A possible solution is to use a measure which is not
based on the entire image but only on the non-border areas, and for this we need to use the
compression ratio of the file. This method is examined further in the discussion section
that follows these results.
Figure 86. P32b inv grey cell green gappy close5 big c (2007), digital print.

221
Table 7 shows the result of Pearsons product-moment correlation for pairs of the
three subjective ratings, using geometric mean scores. There is a stronger correlation with
image quality ratings than that between ratings of complexity and preference. In other
words, the results suggest that visual complexity is more closely related to how good an
artwork is perceived to be than to how much it is liked as an image.
Complexity
0.54
0.81
0.81
Preference
Quality
Table 7 Correlations between subjective ratings (geometric mean scores).
Correlations between subjective ratings and file compression are compiled in Table 8.
Subjective ratings
(geometric mean)
Log file size
Compression ratio (virtual)
TIFF
PNG
JPEG
TIFF
PNG
JPEG
Complexity
-0.27
-0.25
-0.19
-0.15
-0.18
0.03
Preference
-0.33
-0.31
-0.42
-0.15
-0.15
-0.15
Quality
-0.32
-0.33
-0.23
-0.05
-0.07
0.07
Table 8 Correlations between subjective ratings and objective measures of complexity.
The results show almost no relation, with a tendency for a negative correlation. This
means that the effectiveness of the file-size technique as a measure of visual complexity
does not work for the images in Test 3a. An explanation for the poor result compared to
that in test 1 and 2 might be that the wide variety of artefacts represented in the stimulus
images confounded judgements of complexity. In other words, there were many other
factors at work in the style, size, colour and composition of the stimuli which may have
influenced the reported aesthetic judgements. This interpretation would be in line with
Birkhoffs restriction of his aesthetic measure to similar objects and images. Given that
there are significant correlations between the subjective ratings (Table 7), however, it is

222
reasonable to assume that the variety of images also causes difficulties for the filecompression measure, and therefore that the main reason for the poor correlation between
objective and subjective measures lies with the performance of the file-compression
Quality
Preference
Complexity
measure on these images.
TIFF
PNG
JPEG
Figure 87 Test 3a: plots of log file size (x-axis, bytes) vs. subjective scores (y-axis), with
regression analysis (2nd order polynomial).
Discussion
Compression Ratio
Instead of using the absolute file size, another way of measuring the data
compression of image files (and therefore the complexity of the images) is to use the
compression ratio. This is a measure calculated by dividing the size of the compressed file
by the size of the uncompressed file. It is more informative because it indicates not only
the size of an image file but also how much it has changed through compression. There are
two conventions of writing compression ratios: One is such that for a compressed file

223
which is a quarter of its original size, we would write 4:1 or 4. The alternative is to write
0.25, and this is the convention adopted here, since it is the direct result of the calculation.
One way of deriving the compression ratio of an image is to make two versions of
the file one compressed and the other uncompressed and to read their actual file sizes.
A problem with this is that the size of the files may vary with different computers and also
that lossy file types, such as JPEG, cannot be saved without compression. A more reliable
figure can be obtained by calculating the amount of memory (RAM) required to display the
image on a computer. The size in memory is very close to the uncompressed size, and it is
more consistent because it depends only on the image area (number of pixels) and the bit
depth (amount of memory allocated per pixel) of the chosen colour system (Salomon,
2004). In this test, all the images share the same number of pixels (4961 7016 =
34,806,376) and the same colour system and bit-depth (sRGB, 24 bits or 3 bytes per pixel),
and so they all have the same size in memory (34,806,376 3 = 104,419,128 bytes or about
a hundred megabytes). Therefore, calculating the compression ratio (file size divided by
size in memory) gives us a relative measure rather than an absolute one, which is more
informative.
The Problem with Borders
Because all the test images have the same size in memory, the rank order based on
compression ratio is no different to that based on the absolute file size, and so this measure
of image complexity is still affected by the border areas. To take account of the effect of
borders on the compression ratio, we need to use a value for the size in memory which is
based only on the number of pixels that make up the main image by subtracting the
number of border pixels from the total. In practice, this is achieved by image analysis in
Photoshop of files with transparent border areas, since this clearly distinguished the border
pixels from the rest of the image (otherwise, white pixels in the image may have been
counted as white border pixels). So, for images that have a border, we calculate the
compression ratio based on the size of a virtual image which is equal in area to the whole

224
minus the number of border pixels. Figure 88 illustrates how file size and compression
ratio vary with different images. The first image is large, the second is the same size
including a border, and the third is the same size as the central image without the border.
File size
File size
Size in memory (actual)
Size in memory (virtual)
Compression ratio (actual)
Figure 88 Two types of compression ratio with images of different sizes and borders.
In this hypothetical example (Figure 88) each image contains the same kind of
pattern, such that they would be rated subjectively as equally complex. We want the
compression ratio measure to do the same. The file sizes are represented in the top graphs,
which show that the second and third images have almost equal sizes because the plain
border is highly compressible and adds little to the size of the middle image. The second
row of graphs shows the size in memory required to display the images. The actual size in
memory is proportional to the total image area, whereas the virtual size in memory is
proportional only to the non-border area. The compression ratio (bottom row of graphs) is

225
calculated by dividing the file size (top row) by the size in memory (middle). We can see
that the compression ratios in the bottom-right graph are almost equal, which reflects the
compressibility of the pattern that all the images share, without being unduly influenced by
the dimensions of the images or the presence of borders. This method, which I will call the
virtual compression ratio, is the proposed solution to the unwanted effect of border areas.
When we arrange the set of images in order of virtual compression ratio of JPEG files
(Figure 89) it appears not to be biased by the border areas, unlike the absolute file size
measure (Figure 85).
Figure 89 Images from test 3 arranged by JPEG virtual compression ratio

(compressed file size relative to size in memory of image area minus border pixels).
The blue and black image (Figure 86) now has the highest virtual compression ratio
it is the one rated simplest by this measure of visual complexity, probably because it mainly
comprises large areas of a few flat colours. In contrast, the same image was subjectively
rated 8th most complex on average, perhaps because it suggests more complex shapes than
it is composed of, since it depicts a complicated three-dimensional structure in simple tonal
form. However well the correlation with the subjective measure turns out, we have now
found a way of taking account of the unwanted influence of the border areas on the
objective measure of image complexity.

226
In summary, the compression ratio is useful as a measure of visual complexity

because it tells us how random an image file is, and also because it allows for comparison
of images with different dimensions. It is affected by the presence of borders, however, but
by using the virtual size in memory to calculate compression ratio we neutralise this
unwanted effect by basing the measure on the significant (non-border) area of the image.
Since the printed stimuli depict the artefacts at actual size, this makes the measure relative
to the actual size of the original artefact, which may also bring this computational measure
closer to perceived aesthetic properties.
Interviews with Participants

After the rating task described in the previous section, we interviewed the
participants of Test 3a. The purpose of these interviews is to see whether we can establish
which criteria are employed in the perception of visual complexity and the judgement of
preference and quality. The aim is to see whether we can identify the aesthetic properties
that constitute the perception of visual complexity, and to see what we can find out from
these results, in order to support the quantitative measurement of objective aesthetic
properties and their subjective perceptions. A qualitative analysis in the form of
transcription coding is employed in order to identify and categorise the responses, and a
basic quantitative analysis establishes their frequency and significance. In performing an
analysis of qualitative data in this case, interview transcripts we are dealing with data
based on semantic information, which contrasts with the quantitative analysis of data based
on aesthetic information that forms the principal research methods in this project. This
qualitative analysis thus provides an opportunity to support the results of the quantitative
analysis: We may be able to find out which aesthetic properties have an effect on the
perception of visual complexity. We may also establish the criteria behind judgements of
preference and quality, and whether there any relationships between these criteria. The
following sections document the process of interview, transcription, coding, and analysis.

227
Interview Procedure
In all but one case, interviews were conducted in the test location directly after the
participants had completed the rating tasks. These were one-to-one interviews with the
researcher, although on occasion other test participants who were engaged in the task were
present in the same room. Interviews were recorded as audio files, with written notes taken
at the same time, and were later transcribed. The 31 participants comprised staff and
students from two departments within the University psychology and art & design. Some
of these were people that I know well, but most were volunteers who I had not met before.
A semi-structured approach is adopted, in which a series of seven questions were put
to the participants, with opportunity for the conversations to diverge from this structure
and to allow lines of questioning to develop. Interviews began by giving the participants an
opportunity to ask questions and by asking how comfortable they were with the task. The
main focus of the questions was to identify the criteria employed in the rating task by
establishing how the participants had made these judgements. Questions generally took the
form What made one image appear more complex (or more preferred, or higher quality)
than another?, and follow-up questions were frequently employed to elaborate
participants responses. After responses to these questions had been gathered, participants
were asked about their involvement with art and design and to what extent they were
trained in this discipline, the purpose of which was to define the grouping for analysis into
art-trained and untrained participants. Finally, participants were given an opportunity to
add further comments. In summary, the structure of the interview is as follows (Table 9):
1
Do you have any questions about the tests?
Were you comfortable with the procedure?
What were the criteria you used to rate complexity?
What factors influenced your most and least preferred images?
What criteria did you use to judge quality?
Would you say that art and design is part of your day-to-day life?
Is there anything else you would like to comment on?
Table 9 Interview structure.


228
Transcription
Before we can begin the analysis, the recorded interviews need to be transcribed.
This was performed by the researcher, since I had also conducted the interviews and taken
notes. A total of around 6 hours of recording were transcribed to approximately 50,000
words, it taking on average five or six times the length of each interview to transcribe into
text. Participant identities are anonymised, and we refer to individuals by their number (P1,
P2, etc.). When, on occasion, a word or phrase could not be identified, this was marked in
the transcription with the symbol [?], sometimes with the addition of a comment
explaining whether it was inaudibly quiet, or with an alternative interpretation for a word if
there was some ambiguity. Where possible, participants were contacted to clarify such
issues with interpreting the recordings and the transcripts. Punctuation of these
transcriptions was a matter with some degree of freedom in its interpretation, but an
attempt was made to apply consistency to the texts, for example, using commas (,) for brief
pauses and ellipses () for longer ones. The interviews were transcribed into a tabular
format, with each alternate contribution from participant and interviewer being placed in a
separate box within the table. The timings of these responses were labelled in a column
every so often throughout the transcribed text, in order to facilitate finding a portion of a
text in an audio file and to provide a record of the interview process. An example of these
transcribed texts can be found in Appendix C.
Coding
After the transcription process, we may begin the qualitative analysis. In this project,
we adapt methods detailed by Saldana (2009) The Coding Manual for Qualitative Researchers,
which provides a down-to-earth account of qualitative coding methods with an emphasis
on their practical application. Saldana describes textual coding as an exploratory and
iterative process, recommending that the coding goes through first-cycle and then secondcycle coding stages. These methods analyse the transcriptions by identifying and
enumerating answers to the interview questions, enabling a basic statistical analysis to
establish the frequency and the significance of these responses.

229
Structural Coding
The first-cycle coding method applied in this project is what Saldana categorises as
an elemental method called structural coding, which is applicable to studies with multiple
participants, standardised or semi-structured data gathering protocols, hypothesis testing or
exploratory investigations to gather topics lists or indexes of major categories or themes
(2009, p.66). Structural coding is therefore a suitable method for the interview transcripts in
this project. We employ this method in order to identify the major themes of the
participants responses as led by the questions in the semi-structured interview. It works by
applying a content-based or conceptual phrase representing a topic of enquiry to a
segment of data that relates to a specific research question used to frame the interview
(Saldana, 2009, p.66). The structural coding method breaks down each transcript into
sections according to these codes by grouping together individual statements. In practice, it
was applied by colour-coding sections of the transcribed texts (cf. Table 10 and Table 12).
This process forms structural units of analysis for second-cycle coding. The seven
structural codes are (Table 10):
1
Questions about the task
Comfort with the procedure
Complexity criteria
Preference criteria
Quality criteria
Art training
Comments about the task
Table 10 List of structural codes and their colour codes.
Pattern / Focused Coding
The second cycle coding method employed in this project is based on a mix of
methods which Saldana (2009) identifies as pattern coding and focused coding. Pattern coding
generates inferential codes that identify emergent themes it identifies the what and the
how in order to answer the why questions. This is appropriate to our study, in which we
are looking for the reasons why participants rated images as they did by identifying which

230
aesthetic properties influenced their judgements. Focused coding searches for the most
frequent or significant initial codes to develop salient categories, and it requires decisions
about what makes the most analytic sense (Saldana, 2009, p.155).
The second-cycle coding process begins by generating a tentative list of codes
focused on answers to the interview questions. The structural codes provide a means of
identifying the relevant portions of the transcribed texts. For each topic (i.e. each of the
seven structural codes), a set of initial codes is generated by listing each answer and
grouping these into pattern codes. Pattern codes are refined by iterating the method
through a cycle of generating, pruning, grouping, splitting and sorting. The list of codes for
complexity criteria (i.e. for structural code 3) is reproduced in Table 11, and the full list of
codes can be found in Appendix D.
01. Quantity of picture elements
02. Variety of picture elements
a. Number of varieties
b. Amount of variation
03. Level of detail (amount of detail, busy-ness, intricacy)
04. Number of colours
05. Amount of colour (vibrancy)
06. Amount of time
a. In making the work
b. Looking at / understanding the work
07. Redundancy (if the details repeated, its sort of less complex, [randomness] would make it, er, more
visually complex but less informationally complex, if youve got too many colours theyll almost become
like one colour, then then it negates it)
08. Amount of work or difficulty
a. in making
b. in perceiving (cognitive load, how much time or how much effort it takes to unpick whats
actually going on within the image so that you can understand the image Its like a difficulty to
reading the images.)
09. Amount of information
a. Aesthetic (visual)
b. Semantic
10. Depth (Three-dimensionality, layers)
11. Ambiguity
12. Size
13. Flaws
Table 11 List of codes for complexity criteria, with examples of responses coded.

231
Some of the codes in Table 11 have explanatory examples of the types of response
that are encoded. These are the codes that group together different responses, or in which
there is some variation in the phrases used. Those codes without explanatory examples did
not have these issues, generally being stated in an unambiguous form using words similar
to the descriptive codes themselves. Where codes have parts a and b (codes 02, 06, 08 and
09 in Table 11), these allow for various levels of response. For example, one interviewee
mentioned amount of information as a criterion for rating complexity, and this is coded
as 09. If an interviewee makes a more specific statement, such as amount of visual
information then this would be coded as 09a. By grouping or splitting these multiple-part
codes, we can make the analysis more specific or more detailed as required.
Once a stable set of codes has been established by iterating the code-generating
process, the next stage is to apply these codes to the texts, so that we may identify the
presence and frequency of interviewee responses. It is important to note that although the
pattern codes are generated from each of the individual structural codes, the answers to
which the codes refer may be found in parts of the text other than its own topic, and so the
whole of each text must be carefully scanned in order to identify responses and apply codes
for each of the topics. A topic section may contain more than a single code, depending on
its content. An example from a coded transcription is shown in Table 12.
The codes applied to the transcripts are constructed from the name and number of
the pattern code, for example Complexity01 or Quality07, rather than using a more
descriptive code such as level of detail. The reason for using this format is that it allows
for a computer-based quantitative analysis by counting the occurrences of these unique
numeric codes, which are less problematic to process than descriptive codes because they
are composed of single words as opposed to strings of text. Unfortunately, this format
also makes it difficult to read the coded transcripts and identify the meaning of the codes
without referring to the complete list (included in Appendix D), but since a quantitative
analysis is our goal, we retain this format.

232
00:00
GB
P2
GB
00:28
P2
GB
00:36
P2
GB
P2
00:48
GB
P2
GB
00:56
1:02
1:06
P2
GB
P2
GB
P2
Have you got any questions for me?

No.
Thats easy. Were you comfortable with the
procedure?
Yeah, fairly straightforward.
Good. Right, can I ask what were the criteria you
used to rate complexity of the images, you know,
what made one image more complex than another?
I think just the level of detail in the image.
OK
Some just have a lot more detail in them than others
and some are very simplistic.
Mm-hm. Not the size?
No, not the size at all, no.
OK, um What factors influenced your most or least
preferred images?
Umm
What made you like something?
I dont know. I think its just kind of instinctive really
OK
I like art but I think I know very much what I like and
what I dont like. And quite often its, um, quite
dichotomous. I either very much like it or dont
Questions01
Comfort01
Complexity03
Preference02
Table 12 A portion of a transcribed and coded text.
Results
After collecting the list of codes from the right-most columns of the coded
transcriptions such as Table 12, we are able to perform a statistical analysis. First the data
has to be cleaned by removing duplicates in each participants data, because we want to
avoid counting the same answer more than once per person. The reason for this is that we
are trying to find out how many participants gave each response, and not how many times each
participant gave a particular response. After data cleaning we perform a count of the
pattern codes within each structural code and rank them accordingly. The following
sections include selected results of this statistical analysis and qualitative overviews of the
coding for each of the interview topics (the complete tables of these results can be found in
Appendix E). Where applicable, we group the codes and perform another statistical count
as a second stage of analysis. Following these analyses, we see what patterns we can find in
the distribution and type of responses by categorising and classifying the grouped codes.

233
Questions about the Task
Half of the participants (18 out of 31) stated that they had no questions at the
beginning of the interview. The most popular questions were about the purpose of the
tests (5 participants) and how many of the images were my artworks (4 participants). Some
said that until this task, they had not considered visual complexity as an aesthetic property
of artworks, but that they found the idea quite interesting, and that it either made them reappraise artwork they had already seen, or would make them think about visual complexity
the next time they encountered art. Several people wanted to know what my findings were,
and a few were particularly interested in seeing their own results. Confidentiality was
maintained with this data, in accordance with the ethical guidelines of the University, but I
obliged a few of the participants with graphical representations of their individual results
(unfortunately, the poor statistical correlation of this data meant that the graphs were not
very appealing). More than once I was asked What is the most complex artwork you have
seen? I found it hard to answer, so I turned the question around. The most common
answer was Jackson Pollocks drip paintings.
Comfort with the Procedure
24 of the 31 participants said that they were comfortable with the task. Surprisingly,
given their experience with test procedures and rating scales, it was the psychology-trained
individuals who found the rating procedure more difficult than the art-trained individuals.
A possible interpretation, which was suggested by a few interviewees, is that the
psychologists familiarity with percentage and Likert scales of rating made the less-common
MES technique seem more unusual than it did to the artists, who were generally unfamiliar
with any such test procedures. Some of the participants revealed that they had used a
different rating system to that specified in the instructions (see Appendix A). This meant
that whereas the MES method specifies only a lower bound (zero) to rating values, some
participants admitted using a scale out of ten or a percentage scale. Since these
participants said that they had understood that they could have exceeded the limits of these
scales (because this had been explained verbally beforehand, being the essential principle of

234
the MES rating method), the data sets remain valid and are not discarded from the current
analysis.
Complexity Criteria
Table 13 shows the results of rank-order analysis of the criteria for judging visual
complexity. It includes a description of each code, its frequency count, percentage (of the
total number of responses in that structural code group) and rank order. We can see that
the top three criteria are the level of detail, the number of colours and the redundancy of
the images. Level of detail is the descriptive code for a variety of responses, most of which
actually contained the word detail, but also included in this code were such phrases as
busy-ness (P7) and intricacy (P30). In contrast, number of colours is a relatively
unambiguous concept that was not expressed in a greatly different kind of wording.
Criteria for judging Complexity (18)
Code
Count
% (73)
Rank
Level of detail
03
10
13.70
Number of colours
04
12.33
Redundancy
07
12.33
Amount of work or difficulty: In perceiving
08b
9.59
Depth
10
9.59
Quantity of picture elements
01
6.85
Variety of picture elements: Number of varieties
02a
5.48
Amount of work or difficulty: In making
08a
5.48
Amount of colour
05
4.11
Amount of information: Aesthetic (visual)
09a
4.11
Variety of picture elements: Amount of variation
02b
2.74
11
Amount of time: In making the work
06a
2.74
11
Amount of time: Looking at / understanding the work
06b
2.74
11
Size
12
2.74
11
Amount of information
09
1.37
15
Amount of information: Semantic
09b
1.37
15
Ambiguity
11
1.37
15
Flaws
13
1.37
15
Table 13 Results of quantitative analysis of complexity codes.

235
The most frequently counted code for complexity criteria is the level of detail. If we
understand the meaning of this term as relating to the constituent elements of an image,
then it refers to the amount of visual elements present and, at the same time, to the size of
those elements. Thus, the rendering of an image in greater detail but at the same size
means that it is composed of a greater number of smaller elements. Conversely, with a
greater number of smaller elements at our disposal, we are able to compose an image with
more detail. So, we can understand this criterion as relating to one of the fundamental
attributes of visual complexity (identified in Chapter 2, pp.151153) namely, the quantity
of picture elements. Greater detail means more elements and greater visual complexity.
The second-most frequent responses identified in the results are the number of
colours and the redundancy of the images. In relation to the fundamental attributes of
complexity, the criterion number of colours corresponds to the variety of picture elements.
A greater variety of elements contributes to greater complexity in terms of an objective
measure because it requires a longer and more complicated description. As in the case of
detail (quantity of picture elements), this criterion also has a positive correlation with
complexity; more perceived colours means more perceived complexity.
In contrast, redundancy has a negative correlation with complexity, and it concerns
the last of the three fundamental attributes of visual complexity the order of picture
elements. The descriptive code redundancy is identical with the concept from information
theory which has been discussed throughout this thesis and which forms the basis of the
file-compression measure. This is the idea that repetitive patterns offer no new information
and thus do not contribute much to estimates of complexity further than that of the
repeating unit itself. Participants expressed this concept in various ways, such as P7, who
initially described their perception of complexity as being proportional to the amount of
perceived detail, but who later qualified that statement with the concept of redundancy:

236
at first I was thinking just busy-ness, as, you know, how busy it is, how many
different things are going on, basically. [] But then if the details repeated, its sort of
less complex somehow.
Similarly, P16 expressed a comparable idea with reference to the number of colours, rather
than to the repeating unit of patterns:
if its black and white, it would be more simplistic than if it had more colours in it.
But then, that said, if youve got too many colours theyll almost become like one
colour, then then it negates it.
This statement also expresses the idea (discussed in Chapter 4, pp.201202) of the way in
which image resolution and its perceived detail relates to complexity. A high-resolution
image composed of randomly-coloured pixels may be perceived as a uniform hue if its
resolution exceeds the resolving power (visual acuity) of the viewer. The responses of these
interview participants confirm that such an effect leads to a reduction in their perception of
complexity. In addition, this result offers an explanation for the correlation between
perceived complexity and the file-compression measure.
The next-most common responses after the top three, with 7 counts each, are depth
and amount of work or difficulty in perceiving. The idea of depth as an aesthetic property
of two-dimensional images was expressed in a few different ways by the participants, from
the straightforward its more to do with depth rather than anything (P18) to more
ambiguous phrases such as amount of layering (P26) and whether it was twodimensional or three-dimensional (P29). It is interesting to note that the aesthetic property
of depth in a two-dimensional image is restricted to subjective perceptions, and cannot be
found in the objective properties of the image (because it has no actual depth). Presumably,
the perception of three dimensions forms a more complex percept than that of only two
dimensions, because to perceive visually is, in a sense, to re-construct a visible scene in our
minds. In Marrs (1982) computational theory of perception, depth perception is built up in
stages from a 2D primal sketch, based on a two-dimensional array of edges and lines,
through a 2 D sketch to a full 3D model representation, which is based on cues such

237
as occluding edges, perspective and shading. Thus, the perception of three dimensions
involves a building-up of visual elements and is understandably a more complex
psychological phenomenon than the perception of two dimensions. This offers an
explanation why depth should be found amongst the most frequently reported criteria for
judging visual complexity. In addition, we can understand the perception of depth and its
contribution to the perception of visual complexity in terms of the other criteria that rank
alongside depth, namely the amount of work or difficulty in perceiving an image: An
image that looks more three-dimensional than another is likely to require more work in
terms of processing visual elements and constructing a perception.
Most of the complexity criteria identified in these results are aesthetic properties, but
some are not: Amount of work or difficulty (code 06) and amount of time (code 08) are
not properties of the aesthetic image itself but factors relating to the production and the
perception of an image. Like the perception of depth, the amount of work or difficulty in
perceiving an image is a subjective aesthetic property which cannot be measured directly in
terms of objective aesthetic properties. These criteria appear to be based on the amount of
physical or mental work required in making or perceiving an artwork. Presumably, they are
based on the perception of particular aesthetic properties that give rise to the impression of
work or effort on the part of the maker, but without performing another interview to
follow up this idea, the current analysis is unable reveal to what extent this may be true.
Nevertheless, the results indicate that these criteria are in fact related to the objective
complexity of images in that a greater quantity and variety of picture elements, which make
an image more objectively complex, requires greater effort in perceiving and thus
contributes to the perception of greater complexity.
If we perform a second analysis based on the next level of code grouping that is, by
counting together the two-part codes (with parts a and b), the results are as shown in Table
14. Now, the most frequent coded response is the amount of work or difficulty that goes
into making and perceiving the work. This result reflects our understanding of measuring

238
visual complexity in terms of the difficulty in describing or reproducing an image, which

underlies the file-compression measure employed in this study. The criteria previously
counted as the top three (detail, redundancy and number of colours) are now each shifted
one place lower in the rankings below this new grouping.
Criteria for judging Complexity
Code
Count
% (73)
Rank
Amount of work or difficulty (perceiving/making)
08
11
15.07
Level of detail
03
10
13.70
Redundancy
07
12.33
Number of colours
04
12.33
Depth
10
9.59
Variety of picture elements (number/variation)
02
8.22
Amount of information (aesthetic/semantic)
09
6.85
01
6.85
Amount of time (making/perceiving)
06
5.48
Amount of colour
05
4.11
10
Size
12
2.74
11
Table 14 Second-level analysis of complexity criteria (codes with more than 1 count).
Preference Criteria
In the coding process, there was often uncertainty about the direction of criteria,
that is, whether a particular criterion was employed positively, such that it increased
preference, or whether it was employed as a negative trait that reduced preference. This
issue arose when participants listed criteria without specifying whether it added to or
detracted from their preference for an image. For example, preference code 12 may include
statements of preference for and against computer-generated images. A more detailed
interview process than that of the present study could pursue this issue and resolve the
ambiguity about negative and positive uses of criteria. In the few cases for which the
direction of the criteria were actually specified, we are able to include this in the codes,
such as preference code 15, which expressed a dislike of pretension. In short, this analysis
reveals which criteria are employed, but not how they were employed. Despite this limitation,

239
however, the responses generally indicate positive criteria, which do contribute to greater
preference.
Another point to be made about the coding process is that in the cases where there is
clearly an oppositional pair of criteria, such as abstract and figurative, a two-part code is
created. Parts a and b represent the two opposing criteria, and a descriptive phrase is
used to identify both together. For example, codes 05a (representational) and 05b
(abstract) could be grouped as level of figuration. As in the case of the two-part criteria
in the list for complexity, those for preference could be counted together in a further
analysis (this applies to only two of the preference codes: level of figuration and quality of
line). Since these potential groupings bring together opposing preferences, and the validity
of this grouping is questionable, we present only the first level of analysis here (Table 15).
Criteria for judging Preference (25)
Code
Count
% (50)
Rank
Instinctive judgement
02
12.00
Strong colours (bold, saturated)
03
12.00
Quality of line: Swirly, curved
08b
12.00
How much time would like to spend looking at it
09
8.00
Level of figuration: Abstract
05b
6.00
Level of figuration: Representational
05a
4.00
Computer-generated
12
4.00
Complexity
13
4.00
Pretension (dislike)
15
4.00
Complementary colours
17
4.00
Table 15 Top results of quantitative analysis of preference codes (codes with more
than one count).
In contrast to the criteria for complexity, these responses are much more varied, and
produced the longest list of codes (25 in all). Unlike the complexity criteria, which appear
to form natural groupings, the criteria for preference are much less homogeneous and are
thus more difficult to make sense of. Some participants said that they liked many of the
images very much (for example one participant wanted to buy one of the images) whereas

240
others were far more critical, admitting to liking only a handful of the images. The variety
of responses recorded from interviews is reflected in the poor statistical correlation of the
participants aesthetic judgements. Amongst the top responses we find that preference is an
instinctive judgement, which lacks objective criteria, and which we are thus unable to
further analyse. The other two top criteria, also with 6 counts each, reveal preferences for
strong colours and curvilinear images. The amount of time spent looking at an image came
as the fourth most popular criteria, but it is difficult to read much into this because we
cannot be sure what motivates this response; saying that we like to spend time looking at
an image does not explain why we like to look at it. Complexity appears twice amongst the
total of 50 counted codes, but we suspect that if the tests had not focused on this property,
these responses may not have appeared as frequently.
There is little overlap between the criteria employed for judging complexity and
preference. This supports the findings of the quantitative analysis, which indicates that
there is little correlation between these two judgements. A possible explanation for this
result is that a judgement of preference is not an aesthetic judgement, as defined by Kant,
but a judgement of agreeableness (see Chapter 1, p.19). Preference is subjective, as are the
aesthetic judgements of the beautiful and the sublime, but unlike those a judgement of
preference is not normative (or universal, in Kants terminology). Therefore, judgements
of preference are not of the same class as aesthetic judgements proper, which appeal to
objective criteria and make a claim to truthfulness that can be argued against.
The contextual review of material in Chapter 2 identified empirical research on the
perception of visual complexity, and noted that a number of these focused on preference
for images in relation to the judgement of perceived complexity. Evidence for an inverted
U correlation between preference and complexity (i.e. a preference for mid-range values
of complexity), as proposed by Berlyne (1971), has been supplied by some of this material,
whilst others offer evidence to the contrary. Roberts (2007) suggests that the use of
different kinds of visual stimuli and different definitions of complexity in such tests are

241
likely causes for their divergent results, and also that individual differences in personality
and cognitive style of participants may have a greater influence upon results than has been
assumed. The results of the present study offer little clarification of the issue, other than to
support the idea that visual complexity and preference are of two different classes of
aesthetic judgement, which provides a hint as to why there may be so much uncertainty
and disagreement about their relationship.
Quality Criteria
There is more agreement amongst the criteria for judging artistic quality than those
for preference, and there are fewer codes altogether. Only five of the thirteen criteria have
more than one count amongst the total of 40 coded responses (Table 16). The top criterion
is the level of skill perceived to have been required in making the work. Skill is not an
objective aesthetic property; it is a subjective impression which results from the perception
of aesthetic properties, but at this stage it is difficult to say which and why.
Criteria for judging Quality (13)
Code
Count
% (40)
Rank
Skill
03
20.00
Amount of work: In making
01a
15.00
Suitability for exhibition
05
15.00
Technical accomplishment (level of finish)
07
15.00
Difficulty in making the work
04
10.00
Same as for preference
10
7.50
Table 16 Top results of quantitative analysis of quality codes (codes with more than
one count).
Three participants said that their rating of quality was the same as their preference.
We can understand why participants might have a preference for artworks as a result of
their perceived quality, but not why some should reverse this relationship and rate as high
quality the things that are most preferred. More interestingly, six participants rated quality
in terms of the suitability of the artefacts for exhibition. Some of these participants
reported that this type of rating consisted of judging whether an image would hang in an

242
international exhibition, or a local gallery, or on the fridge door at home. This constitutes a
different category of response to those criteria based on skill and effort; it is a criterion
with reference to the context of the artworld an a-perceptual aesthetic criterion, yet based
on aesthetic perception. With the current form of data and the present analytical method,
we are unable to say which aesthetic properties might be relevant to this form of
judgement, other than to note that, like the criterion skill, it is a high-level concept which
depends on a contextual knowledge of the artworld and its artworks.
Some of the most frequent responses are related to the difficulty of making a work in
terms of the level of skill required and the amount of time or effort that it requires. Thus,
we could group together the codes 01a (amount of work in making) and 04 (difficulty in
making) because of their correspondence in meaning. The result is that amount/difficulty
of making the work is the second-most frequent of the grouped responses, with 10 counts
in total (25%). Similarly, we could group the codes 03 (skill) and 07 (technical
accomplishment), because these are similar concepts that differ mainly in being expressed
in relation to either the artwork itself or to the maker of the work. This places the group
skill/technical accomplishment at the top of the frequency counts with a total of 14
(35%).
Criteria for judging Quality
Code
Count
% (40)
Rank
Skill/Technical accomplishment
03/07
14
35.00
Amount/Difficulty of making the work
01a/04
10
25.00
05
15.00
10
7.50
Table 17 Second-level analysis of quality criteria (codes with more than 1 count).
Art Training
There is little to be said about the qualitative analysis of responses to the question of
art training and experience, because its purpose is to divide the participants into those with
and without art training and experience. This division serves to group the participants in
order to perform a quantitative analysis and find out whether this has any effect on the

243
pattern of aesthetic judgements. Of the 31 participants in Test 3a, 11 are found to have
formal art training and are actively engaged in art and design. In comparison, 18 of the 27
participants in Test 3b have art training. This means that in Test 3 overall there are 29 arttrained participants out of a total of 58, which is exactly half. The results of the quantitative
analysis of differences between the two groups in Test 3 are presented in the section on
pages 261262.
Comments about the Task
The top results (those with more than a single count) for the coding of participant
comments are shown in Table 18. Almost half the participants had no comments to make.
The majority of the remainder said that they found the task to be an interesting experience,
and a few of these reported that they had enjoyed it as well. The two other comments in
the table are significant criticisms of the task procedure, with two counts each. The first of
these concerns the fact that the interviews were conducted in the same room as the rating
task, which two participants objected to because they could overhear the conversations,
which may have affected the results of their rating. The interviews were held in the same
room to allow for the management of both interview and rating task by the researcher, who
was working alone, and to allow the participants to complete the required tasks in the
shortest possible time. We accept that these conditions were not ideal, and would make
provisions for a better system if the tests are repeated.
Comments (12)
Code
Count
% (35)
Rank
No comments
01
16
45.71
Interesting experience
02e
14.29
Enjoyed the experience
02a
8.57
Shouldnt be using the same room for interviews and task
02c
5.71
Too many images
02h
5.71
Table 18 Top results of participant comments (codes with more than one count).
The other significant comment is that there were too many images to rate. 50 images
had to be rated 3 times (for complexity, preference and quality), but a consequence of

244
reducing this number is that it weakens the statistical significance of the quantitative
analysis. A more even balance could perhaps be achieved with a slight reduction in the
number of stimuli, but since most of the participants were happy with the procedure, we
do not consider this issue to be a major concern for the validity of the results. Amongst the
remaining comments (detailed in Appendix E), we find a query about whether the artefacts
represented in the image set are actually comparable. This is a legitimate concern, and may
explain the poor correlation in the quantitative analysis.
Analysis: Categorizing the Codes

In the previous analysis of complexity criteria, we grouped similar codes together and
re-counted the codes for these new groups, which offers a broader analytical view of the
results. We can go further than this, as suggested by Saldana (2009), by categorizing the
codes in order to reveal further patterns in the data. In this stage of analysis, we define a
category as a group of criteria, and a class as a group of categories. We do not take this
stage very far, nor apply it to all of the coding results, because the analysis undertaken thus
far is sufficient to answer the principal research question of identifying the most significant
criteria. Therefore, we restrict the categorisation of criteria to focus on the judgements of
complexity and quality in order to further explore the relationship between the two.
The majority of criteria for judging complexity are aesthetic properties, the top three
including detail, number of colours and redundancy. In relation to the objective
properties of visual complexity identified in Chapter 2 (pp.151153), we may infer that
detail in this context can be understood as relating to the quantity of picture elements.
Similarly, the number of colours can be seen to correspond to the variety of picture
elements, and the third criterion redundancy equates to the order of these elements.
Thus, in the top three responses to the question of which subjective properties constitute
perceived complexity, we find a reflection of the most fundamental objective properties of
visual complexity. This finding is further supported by the presence of other criteria that
also relate to the properties of quantity, variety and order. Some of these are explicit in the

245
participants responses and the descriptive codes (such as quantity of picture elements),
whilst others are implied or deduced from the coded responses (Table 19).
Criteria for judging Complexity
Category
Class
Level of detail
Quantity
Size
Aesthetic
properties
Variety
Number of colours
Amount of colour
Redundancy
Order
Depth
Production

Flaws
Difficulty
Amount of time: Looking at/understanding the work

Perception

Ambiguity
Table 19 Categorisation of criteria for judging complexity.
By identifying the codes with these three principal components of visual complexity,
we can see that the remaining criteria are of a different kind. In general, these other criteria
relate to practical matters of making and perceiving and artwork, usually in terms of either
the difficulty or amount of work involved. This finding also supports our previous
discussion on the measurement of complexity (Chapter 2) in terms of the difficulty of
description or reproduction. One category of criteria (production) is based on the
practical difficulty of making an artwork. A second category (perception) relates to
perceptual difficulties in making sense of a work. Greater difficulty in production or
perception leads to higher ratings in complexity. As a result, we group these two categories
into a class which signifies difficulty in both aesthetic production and aesthetic perception.

246
The same analysis can be performed with the criteria for judging quality (Table 20).
The majority of the top criteria can be categorised under the same label as is used for the
complexity criteria, namely production. It makes sense to include this category in the same
class as before (difficulty), because the criteria also reflect the difficulty of producing a
work: Skill, for example, is a measure of the capacity to perform tasks, and a more difficult
task requires greater skill. For the remaining criteria, it makes less sense to classify the
categories that have only a single code in each because no grouping can be undertaken (we
define a class as a group of categories). Suitability for exhibition is categorised as relating to
contextual knowledge, whilst same as for preference is a matter of personal inclination.
Criteria for judging Quality
Category
Class
Production
Difficulty
Context
N/A
Personal
N/A
Skill
Table 20 Categorisation of criteria for judging quality.
From this analysis of complexity and quality results, there are two main findings.
Firstly, we find some similarities between their criteria. Whilst the majority of complexity
criteria are classified as aesthetic properties, some of the remainder share the same
categorisation as quality criteria insofar as they relate to the difficulty of production
involved in making a work of art. This finding corroborates our understanding of visual
complexity in terms of difficulty in description or reproduction. Perception can be
understood as a description or a reproduction of the perceptible world. The perception of
complexity depends largely on aesthetic properties that can be categorised in terms of the
quantity, variety and order or picture elements. This second finding supports our
description of visual complexitys principal attributes. It also provides a link between
objective and subjective aesthetic properties which may explain why we find a correlation
between the file-compression measure and the perception of complexity.

247
Summary
The analysis of participant transcriptions reveals that the most frequent criteria
employed in judgements of visual complexity are related to its principle objective aesthetic
properties. Perceptions of the amount of detail, the number of colours, and the redundancy
of the images correspond, respectively, to the quantity, variety and order of picture
elements. The next-most significant criterion relates to the difficulty in perceiving an image.
When grouped with the difficulty of making, it gains significance and is the most frequent
criterion.
Judgements based on preference are found to be much more variable and depend to
a greater extent on subjective criteria than on aesthetic properties. Some said it was an
instinctive judgement made without reference to objective criteria, and equal numbers
show a preference for strong colours and for curvilinear images. There is little
correspondence between preference criteria and those for the other two judgements.
The criteria for quality are related largely to the amount of skill or the amount of
work involved in making, the suitability of a work for exhibition, and the level of technical
accomplishment of the work. None of these are purely aesthetic properties; they relate to a
contextual knowledge of the works production and presentation, although they must be
derived from such properties as are perceived in an image.
In the later stages of analysis, we find similarities and differences in criteria for
complexity and quality. The ranking of individual criteria differ between these two
judgements, and whilst the complexity criteria are mainly based on aesthetic properties,
those for quality generally correspond to factors involved in the making of a work. Yet
both are understood as relating to the difficulty of making and/or perceiving an aesthetic
artefact, which relates to our objective measurement of complexity based on image file
compression. We may infer that the properties that constitute the perception of visual
complexity also contribute to the perception of difficulty in the production of an artefact.
The aesthetic properties that are considered relevant to the perception of complexity are

248
interpreted as skill and effort when it comes to judging quality. A greater degree of those
properties of complexity leads to a perception of greater skill or effort on the part of the
maker. In other words, the things that contribute to the perception of skill or effort in the
making of an image make it more difficult to perceive, and are thus also the things that
make an image look complex. The findings of this qualitative analysis support the evidence
of a link between visual complexity and aesthetic quality which is found in the correlation
of the quantitative analysis.
Test 3b: Art Gallery

The aim of this study is to replicate the previous test in an alternative environment.
The present test measures aesthetic judgements of the same art and design images, but in
the context of an art gallery instead of a psychology lab. The advantage of this experiment
design is that it allows us to see what effect there is upon aesthetic judgements between the
two environments one more controlled and the other more traditional. Notwithstanding
these differences, the repeat test also allows for a pooling of results. The test compares
objective measures of image complexity with subjective perceptions of complexity,
preference and quality, to find out how these aesthetic judgements correspond to each
other, and how they vary with art training and in different environments.
Method
Participants
University staff helped to recruit participants for the tests by inviting their students
to take part. Unfortunately, there was a high drop-out rate amongst those who responded,
meaning that only 27 people out of a target figure of 30 actually participated in the tests.
There were 16 female and 11 male participants, and the proportion of art/design-trained to
non-art-trained individuals was also 16 to 11. Eight of the participants had also taken part
in the previous test. This was originally thought to present a problem, but actually it allows

249
for an enquiry into the perceived effect of the different settings by those who have
experienced both tests.
Setting
To conduct the test, it was necessary to find a gallery that could display the 50
printed images. Since it is located in the centre of town as part of the university campus,
the 1851 Gallery was chosen as a suitable site (Figure 90). Being part of the universitys
Waverley building, the gallery is frequently used by staff and students as an informal
teaching and working space, but this did not significantly affect the tests and in fact it
allowed for the participation of passers-by. The exhibition space was booked a few months
before the test took place, and a small social event was scheduled for the opening. The
gallery room has a high ceiling and plenty of natural illumination from five large windows.
Figure 90 Composite view of Test 3b setting in the 1851 Gallery.
The gallery walls are painted a neutral cream colour in the places where the work was
to hang. The colour of the background to the stimulus images is a significant concern,
because many of them have a white border. Whatever is chosen as a background will form
part of the visual array presented to the test participants, and therefore it has the potential
to affect aesthetic judgements. In the previous test, a white background was prepared by
covering the tables on which the images were displayed in white paper. Neither re-painting
the gallery white nor re-colouring the image borders to match the gallery walls were
practicable options, however (the latter option being incompatible with the requirement of
using the same stimuli as the previous test). As a result, this is one potential confounding

250
factor that may have to be accounted for in the analysis of results, although in practice it
appeared to have little effect according to the participants who were interviewed.
Procedure
For this test, the procedure was the same as before, except for the layout of the score
sheets, which were re-designed to reflect the arrangement of the images in the gallery by
depicting thumbnail images of the stimuli in the same layout. This design meant that
participants no longer had to write each score by a numerical index, making the rating task
slightly simpler than the previous test, and also allowing for easier comparisons between
images and their scores. On average, it took around half an hour to complete the task.
Results
Ranked Average Scores
To illustrate the ranked averages, thumbnail images of the test stimuli arranged by
geometric mean score are shown in Figure 91 for each of the three ratings complexity,
preference and quality. These figures tell us that photographs seem to have been rated
medium in complexity but fairly high in preference and quality. Almost ironically, the image
rated highest for artistic quality was a photocopy of a crumpled coloured-pencil scribble,
which doesnt sound very impressive on paper, but which is actually a piece by the
experienced artist Jonathan Willett. Textiles are generally rated in the bottom half of the
ranks, except for two silk prints that came 15th and 18th for preference. Two woven
flooring fabrics were rated lower for preference, and two cellular automata pattern
furnishing fabrics came out lowest. The images at the bottom of all three rating categories
are the two paintings done by my 3 year old twin nephews, of a zebra and a dinosaur. They
are noticeably higher in preference ratings; a few participants commented that they were
charming pictures, whilst one or two participants thought that they were imitations of
childrens drawings done by more competent artists.

251
Figure 91. Images arranged by geometric mean scores of complexity (top), preference
(middle) and quality (bottom) for Test 3b.

252
Correlations
The correlations between objective measures of complexity and subjective aesthetic

judgments are shown in Table 21. The table includes data for the two different complexity
measures one based on actual file size and the other on the virtual compression ratio.
Absolute file size measure appears to perform slightly better than the virtual compression
ratio, which goes against our expectations. Overall, however, there is practically no linear
correlation between these scores, and this too counters what we might have expected from
Donderis results. The correlations with compression ratio are also illustrated in Figure 92.
Subjective ratings
(geometric mean)
File size (actual)
TIFF
PNG
JPEG
TIFF
PNG
JPEG
Complexity
-0.21
-0.20
-0.09
-0.15
-0.18
0.03
Preference
-0.27
-0.26
-0.30
-0.15
-0.15
-0.15
Quality
-0.20
-0.19
-0.19
-0.05
-0.07
0.07
Quality
Preference
Complexity
Table 21 Correlations between subjective ratings and objective image complexity.
TIFF
PNG
JPEG
Figure 92 Test 3b: plots of log file size (x-axis, bytes) vs. subjective scores (y-axis).

253
Figure 93 illustrates regression analysis for subjective ratings plotted against

compression ratio. From these graphs, we can see that JPEG file compression gives a
narrower range of compression ratios than the other two file types, most likely because it is
a lossy algorithm. JPEG converts RGB colour channels into luminance (Y) and
chrominance (Cb and Cr) channels, and then discards more of the colour data to which we
are less visually sensitive than luminance (Salomon 2004, p.329334). Even though the
JPEG algorithm was designed to take advantage of the peculiarities of the human visual
system, it remains to be seen how this should affect its correlation with subjective
judgements of complexity since its performance is not significantly better or worse than the
Quality
Preference
Complexity
other two file-types thus far.
TIFF
PNG
JPEG
Figure 93 Test 3b: plots of compression ratio (x-axis, no units) vs. subjective scores (yaxis).
The weak negative correlations are what we might expect to find if we knew that the
stimulus images were generally approaching randomness, such that the images were rated

254
high in objective complexity, and lower in subjective judgements. If we were correct in

assuming a threshold of perceptual complexity at a compression ratio of around 0.5, the
fact that many of the image files are above this value might go some way to explaining the
result. This is certainly the case for the TIFF and PNG files, but not for JPEG, whose files
have compression ratios no larger than 0.4. This effect is more noticeable for the virtual
compression ratio than it is for the file-size measure.
The virtual compression ratio did not perform as hoped. Against our expectation,
this supposed solution to the problem of borders actually gives a weaker correlation than
the absolute file-size measure. There is still reason to believe that the principle is sound, but
in practice, with the weak correlation in the present results, it offers no benefit for the set
of images used in this test.
Test 3 Findings
Comparing Results between Tests 3a and 3b
We can compare the subjective results from tests A and B by plotting the data
together. There are a couple of ways to do this, illustrated in Figure 94. The upper graphs
show the geometric mean scores for the two tests. We can see that the scores follow each
other quite closely and also that the complexity scores are slightly higher for test A. The
lower set of graphs shows the correlation between the two tests, and the red lines show the
best fit. If two ratings are of equal ranges, then the line will fit a 45 degree slope, whereas if
the scores differ significantly the line will deviate from this angle. The slope of the graphs
shows that only complexity ratings differ appreciably between the two groups of
participants.
Test 3 Findings
255
Complexity
Preference
Quality
A
B
Figure 94 Top row: Comparison of test 3a (blue) and 3b (red) results for complexity,
preference and quality (x-axis: rank order file size, no units; y-axis: GM scores, no
units). Bottom row: Correlations of geometric mean scores between tests A and B.
The two tests correspond with each other insofar as they tend to agree on the relative
complexity of images (i.e. participants agree which are the most and least complex images),
but the scales of numbers used to represent perceived complexity are higher in test 3a. In
other words, participants perceived more complexity in test 3a. One reason for this
difference may be that the images were viewed more closely in test 3a where images were
presented on tables than in 3b where the images hung on walls. This result discounts the
hypothesis that the test setting would not affect aesthetic judgements. It also suggests that
perceived complexity is more sensitive to environmental variables, such as setting and
viewing conditions, than the judgement of preference and quality which is likely to be more
dependent on subjective variables amongst the participants.
Table 22 compares correlations of subjective ratings for the two tests. It reveals a
similar pattern of results for each, with the strongest correlation between complexity and
quality and the weakest between complexity and preference. The consistency between these
Test 3 Findings
256
results from tests in different locations implies that we can be confident in the collected
data, whilst the strength of the correlations indicates that there are shared criteria being
employed in these aesthetic judgements. Interviews with test participants elaborate these
findings by examining the criteria employed in the tests, which is presented in next section.
Complexity
0.27
Preference
Complexity
0.73
0.66
0.54
Quality
0.81
0.81
Preference
Quality
Table 22 Correlation between subjective ratings for tests 3A (left) and 3B (right).
To find out whether there is a significant difference between the two tests, we
perform a t-test. The results of this and other statistical tests are included in Table 23.
Subjective ratings
(geometric mean)
Correlation
Complexity
0.94
Preference
Quality
Mean
difference
T-test statistics
t
4.54
3.95
0.0001
0.66
-0.49
-0.78
0.44
0.80
-0.84
-1.12
0.26
Table 23 Statistics for data of tests 3A and 3B. T-test is two-sided, p = 0.01.
Considering that these two tests took place at different times and locations with
many different participants, there is a surprising consistency in the subjective responses.
The strongest correlation between the two tests is found in the complexity ratings.
Nevertheless, there is a significant difference in complexity ratings between the tests (r =
0.94, d.f. = 56, p < 0.01), but not for the other two ratings. The r value describes the
consistency in the pattern of the rating, whereas the t value describes the difference
between the ranges of numbers between the two groups. This means that both groups see
the same images as being about the same relative complexity, but the scales used are quite
different (t = 3.95, d.f. = 56, p < 0.01). Irrespective of the location, the two groups are
Test 3 Findings
257
rating in a consistent way. It seems that there is a reduction in perceived complexity in the
gallery context, where the viewing distance was on average greater than in the psychology
laboratory. The relative differences in complexity perceived between the set of images in
each location was about the same, but the results suggest that the overall level of detail or
high spatial frequency information is higher when viewing nearer. Preference remains
constant because moving further away reduces the complexity consistently for all the
images. The inference is that visual complexity is a function of the amount of visual
information available.
Combined Results from Tests 3a and 3b

Visualizing Data
Plotting the graphs in a grid (Figure 92) allows us to see the difference between file
types and between ratings, but because it is based on averages it fails to reveal any detailed
patterns within the data. More detailed patterns can be found by representing the full data
set in another way, using colour. Figure 95 represents the entire data set for Tests 3a and
3b combined. There is one coloured square plotted for each image (50 columns) and each
participant (58 rows) for each of the three rating scales. This makes a total of 8,700 data
points, in a data set with 5 dimensions (participant, image, rating type, file compression,
and rating score). In each plot the image columns have been arranged in order of increasing
file size from left to right. The colours represent the magnitude of the scores red for high
and blue for low.
The value of this basic form of data presentation is that it allows us to use our
powers of visual perception to look for patterns in the data (order in the chaos). If
participants scores corresponded with file size, we would see the colours arranged in a
spectrum from cool on the left to warm on the right. Here, there is only a slight tendency
in the opposite direction noticeable for complexity, which reflects its weak negative
correlation, and there is no obvious pattern in either of the other images. The visible
structures in Figure 95 reflect quantitative patterns in the data. Horizontal regions of colour
Test 3 Findings
258
indicate similarly-rated images for a single participant. Vertical regions of similar colours
indicate agreement between participants ratings, and these patterns are the more important
for the present study, since differences between individuals are later averaged out in
statistical analysis. However, there is actually very little structure in these images, which is a
direct reflection of the lack of order in the data; compared with the more obvious patterns
of data from test 2 (Figure 80) these images are quite disordered.
Complexity
Preference
Quality
Figure 95 Representation of combined data from test 3A and 3B.
Even though the lack of patterns in the data means that unfortunately we can tell
very little from the statistical results, we can visualize the rank order of the stimulus images
as arranged by the average subjective ratings of complexity, preference and quality (Figure
95). This allows us to see what the quantitative data cannot tell us namely, the type of
image as well as its placing in the rankings. The paintings of a dinosaur and a zebra by my
three-year old nephews, Oliver and Isaac Avison, are placed low in ratings of complexity
and quality, but Olivers dinosaur image is notably higher in preference a detail supported
by a few of the test participants who commented in interviews that they had rated it highly
because it was charming. Two other notably childish or amateurish-looking pieces (an
image of Spiderman and an abstract painting) are also scored low for all three ratings.
Test 3 Findings
259
Figure 96 Images arranged by geometric mean scores of complexity (top), preference

(middle) and quality (bottom) for combined results of tests A and B.
Test 3 Findings
260
The most complex rated image is a drawing by Bommsoon Lee, and the majority of
the images in the top twenty rated as most complex are my own which aimed for the
threshold of visual complexity. The image rated highest for both preference and quality is
one I created with Draves Fractal Flames algorithm, which did rather resemble a flame with
its curling red and orange irregular fractal forms. It is interesting to note that the repetitive
patterned textiles were not rated amongst the most simple. A blue and white non-repeating
pattern textile is rated as medium complexity, whereas a similar orange textile with less
contrast is rated far lower, suggesting the perception of an overall texture in the latter case
which lowers its perceived complexity.
When the data from test3a and 3b are combined, the strongest correlation between
subjective ratings is found between complexity and quality (Table 24). The correlation
between quality and preference is very similar, which is what we have come to expect,
because people tend to like the things that they perceive to be good. That between
preference and complexity is around half as strong, in line with the results of previous tests.
Complexity
0.42
Preference
0.80
0.78
Quality
Table 24 Correlations between subjective ratings (geometric mean scores) for tests 3a
and 3b combined.
Regression analysis for the combined results of Tests 3a and 3b are illustrated in
Figure 97. The coefficient of determination (R2) is a value between 0 and 1 that expresses
how well the predicted values match the actual data. Given that there is almost no
correlation between objective and subjective measures of complexity in our data, we are
unlikely to find a good fit to the data. Although graphs in Figure 97 appear to show an
inverted U shape, the poor correlation and weak R squared value means that there is
Test 3 Findings
261
practically no relationship between the file compression measure of complexity and these
Quality
Preference
Complexity
aesthetic judgements. This is also clearly seen in the distribution of data points.
TIFF
PNG
JPEG
Figure 97 Plots of compression ratio vs. geometric mean scores (no units).
The analysis in Table 25 shows what we expected: Each pair of objective and
subjective scores has an R2 value less than 0.1, which means that the best-fitting equations
fail to estimate our data, probably because it is so poorly correlated. Similar results are
found for first, second and third-order polynomials. This means that the file compression
measure of visual complexity has failed to work with these images from art and design,
unlike its success with the sets of cellular automata and random images in the previous
tests. The likely reason for the poor correlation with this sample of images is due to their
great variety. Whilst this variety was one of the test objectives, with the aim of representing
a wide range of visual complexity, the difference between types of media and styles of
artwork appears to have provided extra variables that have masked the relationship
between file size and aesthetic judgements that we found in the earlier tests on more
homogeneous visual material.
Test 3 Findings
262
Subjective ratings
(geometric mean)
File size
TIFF
PNG
JPEG
TIFF
PNG
JPEG
Complexity
0.051
0.053
0.012
0.072
0.075
0.080
Preference
0.084
0.083
0.096
0.023
0.023
0.021
Quality
0.061
0.065
0.045
0.024
0.027
0.051
Table 25 Coefficients of determination for subjective ratings and objective measures

of image complexity.
Artists vs. Non-Artists
Comparative graphs of the combined Test 3 data for art-trained and untrained
groups of participants are presented in Figure 98. The slopes indicate a significant
difference for all three ratings, with art-trained participants rating consistently higher.
Complexity
Preference
Quality
A-T
U-T
Figure 98 Art-trained (blue / A-T) and untrained (red / U-T) ratings for tests 3A and
3B combined. Top: rank order file size (x-axis, bytes) vs GM scores (y-axis, no units).
Bottom: Correlations between art-trained and untrained participants.
The results of statistical tests are presented in Table 23, which confirms that all three
ratings are positively correlated between art-trained and untrained participants, with a
Test 3 Findings
263
strong correlation for complexity ratings (r = 0.90) and a weaker correlation for preference
(r = 0.59). There is a significant difference between the groups: art-trained participants gave
consistently higher scores for all three ratings (e.g. for complexity t = 5.66, d.f. = 56, p <
0.01). This finding is unexpected. We found in Test 2 that artists rated higher for
preference alone, and we find it hard to explain this result when it might seem reasonable
to think that the artists greater exposure to visual material and familiarity with aesthetic
judgement would result in more critical, lower ratings overall. Test 2 findings also
suggested that perceived complexity is less variable than judgements of preference, but this
is cast into doubt with the results of Test 3. We are thus unable to account for the finding
that art-training appears to correspond with higher ratings for all three aesthetic
judgements complexity, preference and quality. It is possible that art training equips an
art audience with more criteria with which to find virtue in images, or that it relates to a
greater capacity to perceive aesthetic information and a greater aesthetic sensitivity.
Subjective ratings
(geometric mean)
Correlation
Complexity
0.90
Preference
Quality
Mean
difference
T-test statistics
t
6.65
5.66
2 10 -7
0.59
5.05
7.81
8 10 -12
0.72
6.15
7.91
6 10 -12
Table 26 Statistics for data from the art-trained and untrained participants. T-test is
two-sided, p = 0.01.
Hypothesis Evaluation
We return to the test hypotheses to see how they fared in light of the test 3 results:
1. Image file compression will correlate with judgements of complexity.
We failed to find any notable correlation between file compression and judgements
of complexity. It seems that the sample of visual material in this test was so varied that it
masked the correlation with file compression that was found in earlier test results.
Test 3 Findings
264
2. Familiarity with art and design will correspond with lower ratings of complexity.
We found the reverse in Test 3, in which ratings by art-trained participants were
consistently higher on average than untrained participants, not just for complexity but for
judgements of preference and quality also.
3. The test environment will not affect aesthetic judgements.
The results show that the test location significantly affected only perceived
complexity, and not judgements of preference and quality.
Summary
Overall, the results of Test 3 are quite negative: they failed to support any of the
three hypotheses or show a significant correlation of compressed file size with aesthetic
judgements of complexity, preference and quality. The poor correlation is not what we
expected, given the results of the earlier tests, but it does make some sense if we take into
account the variety of the visual material used for the test stimuli which appears to have
masked the previously found correlation. We cannot say at this time whether this means
that the file compression measure is unsuitable for art and design images in general or
whether it failed only for the sample used in the current project. The relative success of the
methods with CA and random images in Tests 1 and 2 suggests that the problem lies with
the choice of stimuli in Test 3, and that we should not rule out the file compression
method for further testing in future.
Art-experience appears to have increased the overall ratings of aesthetic judgements,
a finding we are unable to explain at present, especially as it conflicts with the results of
tests 1 and 2 for which art-experience only influenced judgements of preference. The test
location significantly affected judgements of complexity, probably due to the difference in
viewing conditions: Ratings in the art gallery, where the participants could more easily
determine their own viewing distance from the gallery walls, were lower than those of the
Test 3 Findings
265
closer and less-variable viewing range in the psychology lab. The difference in rating scales
is perhaps due to the increased detail that becomes perceivable at closer ranges.
The use of compression ratio as an alternative to the file-size measure of image
complexity is demonstrated to be potentially viable, although its effectiveness could be
more reliably established with further tests. The advantage of the compression ratio is that
it is more informative (it indicates the amount of randomness) and more comparable
between tests. In regard to the proposed virtual compression ratio as a potential solution to
the problems with image borders, it appears to be sound in theory and it did re-arrange the
files as intended, but made surprisingly little difference to results in practice. This is
probably due to the variation in stimuli and the low statistical correlation in the present
data which likely obscured its potential value in this application.
Despite the poor correlation with file compression, there is a stronger correlation
between subjective ratings, particularly between complexity and quality. This pattern is
fairly consistent between different test locations and between groups of subjects. The
statistical relationship is supported by the interviews with participants who reported the use
of similar criteria in evaluating these two aesthetic properties. Both visual complexity and
artistic quality are judged in relation to the difficulty of making and understanding an
artwork in terms of the time and effort required. This is a significant finding because it
constitutes empirical support for the central thesis of this research project, namely that
visual complexity is a significant factor in the aesthetic value of visual art and design. The
finding therefore makes a contribution to knowledge.
Test 3 Findings
266
Chapter 6
Conclusions
Summary of Findings
At the beginning of this thesis, we considered the variety of patterns in our local
environment. It seems that the size of the area occupied by a particular type of pattern is
inversely proportional to its visual complexity. The largest areas, such as sky and walls, have
the simplest patterns. Next-most common are likely to be repetitive patterns either
regular (clothing and furnishings), or irregular and chaotic (grass, rock, foliage). The most
visually complex artefacts are likely to occupy the smallest areas. This relates to the
information theory conception of complexity in terms of probability and predictability.
Information is quantified in terms of its predictability, and complexity is less predictable
than order. By this conception, however, randomness is less predictable and so it has
greater statistical information than complexity. Therefore an information-based measure
fails to describe our intuitive understanding of complexity, which to us has greater
meaningful information than randomness. This problem has been central to the current
project, and forms the basis of its contribution to knowledge. The following sections
provide an account of the ways in which it has dealt with this issue. Firstly, we summarize
the findings by answering the research questions, and identify the contribution to
knowledge. Then we discuss particular implications of the contribution in further depth,
and outline the impact of the research on my art practice. Finally, we identify potential
opportunities for future research in development from the current project.
Answers to Research Questions

In this section, we collate the findings of this project by answering the research
questions which were formulated at the end of the first chapter (discussed pp.1013 and
Conclusions
Summary of Findings
267
listed p.46). The following sections are grouped as were the research questions, by the
research stages adapted from the model of Phillips and Pugh (2000).
Focal Theory
Visual complexity appears to be interesting because the perception of complexity is a

demanding perceptual task. Faced with a complex visual array, we engage in perceptual
processes such as edge-detection and pattern-detection, and since visual complexity is
related to the quantity, variety and order of perceived edges and patterns, a more complex
image may provide a more satisfying aesthetic experience. The reason for the high
complexity amongst artworks may be that artists made this discovery intuitively and
worked to take advantage of successful experiments in visual complexity. Certainly, the
evidence of Pollocks gradually-increasing fractal complexity in his drip paintings would
support this idea. Another reason for complexity in art may stem from the representation
of the natural world, which contains complex patterns of many varieties. On the other
hand, non-representational art tends to be less visually complex, and is none the less
aesthetically valuable for it. We identified minimalism as the movement that most closely
associates its theory and practice with the concept of aesthetic complexity, albeit with the
result of a reduction towards the simpler end of the spectrum.
The study of contemporary art practice indicates that minimalism is manifest at
various stages of the creative process, and also that simple processes can produce complex
aesthetic perceptions. These finding appears to have been made almost concurrently in
empirical aesthetics and in the arts shortly after the introduction of information theory, as
evidenced in the work of Attneave, Moles, Bense and Nake. From information aesthetics
also came the complementary realization that detailed randomness or chaos can produce
perceptual effects of simple textures, an idea which is incorporated in Benses microaesthetics and is currently employed by contemporary musicians and visual artists such as
Carsten Nicolai and Mark Fell. The concept of aesthetic information thus provides as
useful a framework for creative approaches to the production and perception of visual
Conclusions
Summary of Findings
268
complexity as it does for the scientific approach, for which it offers a scale of measurement
than can be applied to aesthetic properties and perceptions.
Contextual Theory
Complex systems theory generally understands complexity as the difficulty of making

or describing something, and it measures this difficulty in terms of time, effort or cost. A
more complex object takes more time or effort to describe, and so the size of a description
is an indication of complexity. Donderis (2000) file compression of digital image files
constitutes a basic measure of complexity which equates to the shortest description of an
image based on an array of coloured pixels. We identified the fundamental components of
visual complexity as being the quantity, variety and order of visual elements, and these are
exactly the properties encoded by digital image files, which describe the number, colour
and arrangement of pixels. Quantity and variety of elements have a simple proportional
relationship to complexity, whereas order is a more subtle matter. Because the filecompression measure is based on the information theory concept of redundancy, the scale
of order it furnishes is a measure of randomness (entropy, redundancy, and
incompressibility). This means that the file compression measure does not accord with our
intuitive notion of complexity, because we perceive randomness as a simple texture, for
example. The hypothesized relationship between the compression of image files and the
perception of visual complexity, made by Wolfram (2002) and Donderi (2006a), suggests
that we should find an inverted U correlation. Donderi and McFaddens (2004) and
Besners (2007) experiments failed to provide evidence for this hypothesis and left a gap in
knowledge.
Whereas empirical aesthetics is a small but flourishing area of current research in
perception, contemporary aesthetic discourse has disengaged with perception in its theory
and to some extent in its practice, such as in Bourriauds relational aesthetics (2002). The
model with currency in todays artworld is the institutional theory, which successfully defies
the neo-Wittgensteinian challenge of defining art posed by Weitz (1956). Dickies art circle
Conclusions
Summary of Findings
269
(1997) does this by describing the mechanism of the artworld rather than the properties of
artworks, but it does so at the expense of providing a satisfying explanation to the
significant aesthetic questions of why we value art and what counts as aesthetic value. Part
of the justification for the focus of the current project is to re-assess perceptual aesthetics,
concentrating on the aesthetic property of visual complexity. The justification for
investigating complexity is that it captures much of the essential information in pictures
because it is based on the fundamental attributes of quantity, variety and order of picture
elements.
Research in vision shows that the perception of complexity largely depends on preconscious processing in the retina, where the edge-detection mechanism of receptive fields
of photoreceptors ultimately gives rise to the perception of recognizable shapes. Marr
(1982) formulated an understanding of vision in terms of the information-processing of
edge-detection and the formation of visual percepts, which allows for a computational
approach to aesthetic perception. Marrs perceptual theory provides a framework that
justifies the computational approach of Donderi and McFadden (2004), and thus supports
this projects methodology which employs the file-compression method pioneered by
Donderi.
Data Theory
The methodology aims to approach both sides of the aesthetic identified in the
working model the sites of aesthetic production and perception. Aesthetic production is
approached through creative experimentation in art practice and by collecting samples of
contemporary work from artists and designers. Aesthetic perception is approached
empirically by using the visual material created and collected as the basis of stimuli in
aesthetic tests. With this material, we can measure both objective and subjective
complexity, and test the hypothesis of the inverted U correlation between informationbased and perceived complexity. We aim to identify whether there is a perceptual threshold
to visual complexity by mapping the perception to a wide spectrum of complexity using
Conclusions
Summary of Findings
270
cellular automata and random images. Later tests use images of the collected artwork to
establish whether the measure would work for the aesthetics of art and design. The design
and execution of the tests gain validity from the success of trials with the aesthetic
measures. Although the ability to transfer conclusions back to the larger population is
restricted by the small sample sizes in these tests, they are sufficient enough to be confident
in the results obtained, as evidenced in the statistical analyses, and are comparable to a
number of the professional empirical studies cited in this thesis.
The results of the first two tests with CA and random images show strong evidence
for the hypothetical inverted U correlation between image file compression and perceived
complexity. In Tests 1 and 2 there is little sign of a unified response to preference for levels
of complexity, with only a slight preference for higher complexity amongst art-trained
participants. There is a significant difference in ratings between art-trained and untrained
participants, with artists rating higher for preference in Test 2. In Test 3 artists rated higher
for all three ratings.
The two locations of Test 3 appear to make a significant difference to complexity
ratings but not to preference or quality. Complexity ratings are highly correlated in the
ranking of images between tests 3a and 3b, but the scales used by participants are
significantly lower in the context of the art gallery (Test 3b), which implies that more visual
complexity is perceived at the closer viewing distance in the psychology lab (Test 3a).
Overall, the file-compression measure fails to reveal a significant correlation with any of
the three aesthetic ratings in Test 3. Nevertheless, the correlation between subjective
ratings indicates a positive connection between visual complexity and aesthetic value,
whether or not it involves conscious perception. This finding is supported by the
qualitative analysis, which identifies the criteria behind the aesthetic judgements of
complexity and quality. The interview analysis also supports the idea that complex images
engage us because we have to work at understanding them and because we find pleasure in
this task we are rewarded aesthetically for the perceptual effort. The results of the
Conclusions
Summary of Findings
271
qualitative analysis suggest that we find complex images rewarding because it is related to
the understanding that visual complexity is evidence of skill and effort on the part of the
maker.
Contribution to Knowledge
The material critically reviewed in this study constitutes a mapping of contemporary
aesthetic practice relating specifically to the area of visual complexity. The review presents
an organised account of theory and practice across the boundaries of three fields:
philosophical aesthetics and art discourse, empirical aesthetics, and art practice. A
contribution to knowledge is made with this up-to-date mapping of visual complexity
which integrates knowledge from separate disciplines and applies it to current
understanding of art and design practice. An argument is made for the re-assessment of
perceptual aesthetics, with the view that it has the potential to supplement the weak
explanatory power of contemporary aesthetic theories of art. The argument is supported
with empirical evidence, which takes the form of the following contributions:
The results of the first experimental trial with the file-compression measure of
complexity on images of cellular automata reveal a direct connection between this
information-based measure and schemas of complexity from systems theory. The reasons
for the correspondence lie in the equivalence of the underlying scales in terms of measuring
complexity by difficulty of description. Although this finding seems almost trivial given the
wide understanding of these principles, it nevertheless makes a small but original
contribution to knowledge by providing clear empirical evidence for the relationship
between these two specific concepts of complexity, and possibly constitutes the first
measurement of cellular automata programs by the file-compression method.
Test 1 and Test 2 provide empirical evidence for the hypothetical inverted U
correlation between file compression and perceived complexity. Against the findings of
Besner (2007), which were interpreted as inconclusive, these results constitute an original
Conclusions
Summary of Findings
272
contribution to knowledge, as being the first hard evidence for this correlation. We found
no strong evidence for a similar correlation between visual complexity and aesthetic
preference with the images in these tests.
It was assumed that Test 3 would similarly prove the effectiveness of the filecompression measure for images from art and design, as it had in the case of CA and
random images, but the results fail to support this idea. There was no evidence of the
correlation of objective visual complexity and subjective aesthetic judgements, which
indicates that either the measure may not work for art and design, or simply that it failed to
work with the sample in this test. Given the success in earlier tests, the latter is deemed to
be the more likely explanation. Nevertheless, we did find evidence of a relationship
between visual complexity and aesthetic value, based primarily on the correlation of
subjective aesthetic ratings. This is supported by evidence of shared evaluative criteria
being employed for judgements of both complexity and quality, based on the analysis of
participant interviews. Against the failure to find meaningful results from the filecompression measure, the findings of this relationship between visual complexity and
aesthetic value constitute another contribution to knowledge.
Through the quantitative analysis, we find a positive correlation between the
objective measure of visual complexity (based on the data compression of image files) and
the subjective perception of complexity (measured with MES). We also find a correlation
between subjective ratings of visual complexity and aesthetic value (quality). Although we
can determine fairly precisely the strength and direction of these correlations, the statistics
tell us nothing about the reasons why. The qualitative analysis sheds light on the matter by
indicating that there are shared criteria between complexity and quality. It suggests that the
aesthetic properties that constitute the perception of complexity are interpreted as signs of
skill, effort or difficulty in the production of an artwork and, further, that visual complexity
also involves skill, effort or difficulty in its perception. The understanding of visual
complexity in terms of the difficulty of its production and perception also relates to the
Conclusions
Summary of Findings
273
file-compression measure, which we can understand as a measure of the difficulty of

description. This is because it is a measure based on the size of the shortest possible
description which is in turn based on the quantity, variety and order of pixels (how many,
which colours, and where). The qualitative analysis reveals that a majority of the criteria
involved in the perception of complexity can be classified in the same terms, namely the
quantity, variety and order of picture elements.
In summary, the knowledge we have gained about visual complexity in this project
can be characterised as a theory which is both descriptive and explanatory. The theory
allows us to organise our understanding of visual complexity and to use this knowledge in
predicting how a given image will be perceived, which has applications for both empirical
and creative approaches to complexity. In this way, the studys contribution to knowledge
serves as the basis for action in both the scientific analysis and the creative exploration of
the aesthetics of visual complexity.
Implications
Measures of Complexity
With our current understanding of visual complexity, we can try to integrate some of
the key theoretical models cited in this thesis, in particular the various measures of
complexity and their relation to aesthetic perception. From the combination of Dodgsons
(2008) and Attneaves (1959b) findings, we have the suggestion that in terms of the
spectrum of complexity measured by file compression ratio, there is a region of visual
appeal between the values of 0.20.5, or around one fifth to one half randomness. In
addition, this region appears to include a perceptual threshold beyond which we cannot
identify structural regularity. The region of visual appeal is illustrated in Figure 99, in
conjunction with Langtons (1990) lambda parameter graph with Wolframs (2002) four
complexity classes mapped out. The figure also shows black and white images of
elementary cellular automata arranged in order of increasing file size from left to right. The
Conclusions
Implications
274
leftmost image is a uniform black, the images in the centre are cellular automata patterns,
and the image on the right is random. Arrows from these images indicate their approximate
distribution on the scale of complexity measured by compression ratio and also their
position relative to Langton and Wolframs mapping of complexity classes.
Compression Ratio
Order
Randomness
Region of
visual appeal
0
0.25
0.5
0.75
Figure 99 An informational scale of orderrandomness and compression ratio (top),

Langtons lambda parameter with Wolframs complexity classes (middle), and
corresponding examples of elementary CA images (bottom).
Conclusions
Implications
275
One of the findings of this project is that by simply sorting these images by file size,
we end up with an arrangement identical to Langtons and Wolframs maps of complexity.
Figure 99 shows that the positions of the images correspond to the levels of complexity in
Langtons graph and to the compression ratio scale above it. This is especially evident
when using the same cellular automata that Wolfram used the relatively simple
elementary CA. The results of the trial with the file-compression measure provide empirical
evidence of the fundamental equivalence of this information-based scale of visual
complexity with measures established in the field of complex systems theory.
If we compare the complexity spectra represented in Figure 99 with the results of
this project, there is no such peak in preference for complexity in any of the three tests, but
we do find a similar peaked correlation between objective (file compression) and subjective
(MES) complexity for the sets of cellular automata and random images. Using the
compression ratio rather than the file size allows us to map the results back onto these
models. Comparison with the test 2 results (Figure 100, red), which gave one of the best
correlations between file compression and perceived complexity, shows a grouping of data
points near the lower end of the compression scale and the correlation curve is pushed to
the left. The reason for this is probably that these four-colour images produced much
smaller files than the uncompressed file size which is used in the calculation of
compression ratio. Therefore the range is artificially compressed because in fact the set did
include completely random images but they were relatively small (in contrast, manycoloured random images do approach this theoretical maximum file size). The range of file
compression in test 3 is much greater (Figure 100, green), but the result is a negligible
correlation with perceived complexity, and the illustrated curve fits the data so poorly that
it should be disregarded. Nevertheless, the results of tests 1 and 2 still provide empirical
evidence for the U-shaped correlation between perceived and information-based visual
complexity.
Conclusions
Implications
276
Compression Ratio
Figure 100 Test 2 (red) and test 3 (green) results complexity ratings vs. TIFF
compression ratio, compared to Langton and Wolframs mapping.
Visual Information
There is a close connection between our understanding of complexity, beauty and
consciousness. All three have been described in terms of unity and diversity of visual
information: Complexity is a structural unity of many different parts; consciousness the
basis of perception is a state of unification amongst the diverse areas of the brain
(Massimi et al., 2005); and beauty is found in the unified perception of visual elements in a
complex image.
In visual complexity, there is also a connection between its perception and its
measurement, since both involve pattern recognition. Data compression, which provides
the basis of our objective measure of complexity, also involves a form of pattern
recognition; repeating patterns of data are redundancies that can be compressed. Not only
does this explain why digital image file compression can be used as a measure of visual
complexity, but it also suggests that some form of data compression is involved in visual
perception. Both Attneaves perception as economical description (1954, p.189) and
Wolframs account of visual complexity as failure of perception to extract a short
Conclusions
Implications
277
description (2003, p.559) support this idea. An explanation along these lines is required to
account for the biology of the visual system: The retina processes visual information from
130 million photoreceptors and sends it to the brain along only 1.2 million ganglion cells in
the optic nerve. Somehow, this visual information undergoes a hundred-fold reduction in
the process of perception. David Marrs computational model of receptive fields explains
how information from groups of sensory elements is processed to higher-order perceptual
constructs such as edges and blobs. So it seems likely that this hundred-fold decrease in
visual information somehow undergoes a data-compression operation. Whether these
processes actually occur in perception, or whether that is merely our best hope of
understanding the phenomenon, is a question that may soon be answered as the science of
perception advances.
We have seen how information theory offers a crude but effective measurement of
complexity. The information-theoretical approach to aesthetics treats visual complexity as
an aspect of communication in which the complexity of a message is related to the
difficulty of encoding, transmitting and receiving it. This difficulty is quantified in terms of
the probability of producing or guessing the message. A statistically unlikely or unknown
message has high information content, whereas a probable or predictable message has low
information content. Message is an abstract term that can stand for any transmitted
information, whether a spoken language or visual material. In this sense, what information is
is unspecified, yet we are still able to say how much information is present in a message:
Information theory measures the amount of information in an observation as the
negative logarithm of its probability. Information itself is never rigorously defined; it is
only quantified. (Crutchfield, 1990)
The nearest thing to a definition of information is perhaps the one offered by

Gregory Bateson: Information is a difference which makes a difference (2002, p.459).
Only things that make a meaningful difference contribute new information, and this is
always relative to the receiver of the information. Bateson considers this understanding of
Conclusions
Implications
278
information to bridge the gap between map and territory, or between representation and
reality. Bateson says that perceived differences are the sort of thing that gets onto the
map from the territory (2002, p. 458). We might also say that these mapped features are
also affordances (Gibson, 1986) because they are the perceived differences which can be
acted upon. In terms of visual complexity, the features that are mapped in empirical
investigation are those perceivable features of artworks that make a meaningful difference
and which can be acted upon in terms of aesthetic judgement. Attneave (1971) understood
visual complexity as a property involving comparison (a collative variable), which
therefore depends on the perception of differences: How complex a pattern is judged to
be depends on the existence, extent and direction of differences along its simultaneously
presented elements (1971, pp.106107). Through the computational approach, we have
identified quantity, variety and order of elements as the essential variables in the perception
and measurement of visual complexity. These are the critical properties that constitute
digital image files, but they are also the most pertinent perceptual features. The relationship
between a work of visual art (the territory) and a digital representation of the artwork (the
map) is bridged by Batesons understanding of information and Gibsons concept of
affordances. The features common to both map and territory are those that are most
significant to the map-maker and map-reader.
The Spectrum of Complexity and the Window of

Visibility
The computational model of visual complexity developed in this thesis allows for
predictions to be generated and tested. It was predicted that with the objective measure of
visual complexity (based on image file compression), images at opposite ends of the scale
would be perceptually similar, i.e. that the objectively simplest and the most complex
images would be perceived subjectively to be equally complex (both perceived as simple
images). The relationship between the objective measure and the subjective perception of
visual complexity is expressed in graphical form as a curve an inverted U shape. Findings
Conclusions
Implications
279
in perception research offer an explanation for the U-shape correlation between objective
aesthetic properties and subjectively perceived visual complexity:
Campbell & Robson (1968) plotted the contrast sensitivity function (CSF), which is a
perceptual threshold based on contrast and spatial frequency (detail). The CSF describes
the limits of perception the point at which details become too small and/or have too
little contrast to be perceived. Campbell & Robson used sinusoidal grating images, which
vary from black to white in smooth gradations of contrast and spatial frequency. Figure 101
illustrates such an image, with contrast increasing from top to bottom, and spatial
frequency increasing from left to right. We perceive contrast only above a certain
threshold, and we are more sensitive to middle values of spatial frequency (peaking at
around 4 cycles per degree), with the result that the CSF produces a curve which describes
an inverted U shape. This curve is not illustrated explicitly in Figure 101 because it varies
for each individual, but it can be perceived as the border between the visibly stripy areas
and the plain grey areas. We can notice this border changing as we alter the viewing
distance; it appears lower at a greater distance. Other animals have greater sensitivity to
contrast (for example, cats can see faint shadows that we cannot, so their CSF curve is
higher than ours) or have greater resolving power (an eagles eyes perceive greater detail,
which extends the CSF curve to the right), but somewhat surprisingly humans have the
best overall CSF with the largest area under the curve. The area under the CSF curve
describes the ranges of contrast and detail that we are able to perceive, and it is known as
the window of visibility.
In this projects tests, the correlation of objective and subjective visual complexity
measures also has an inverted U shape. This result is explained partly by the aspects of
information theory and data compression discussed earlier, and partly by the CSF curve.
The test stimuli sampled the objective spectrum of complexity from uniform to random.
We find that images in the middle of this objective scale are perceived as the most
complex. The simplest (uniform) and most complex (random) images, as measured by fileConclusions
Implications
280
compression, correlate with the lowest values of perceived complexity. We can now
consider these images in terms of the contrast sensitivity function. The simplest possible
greyscale image (e.g. made of identical mid-grey pixels), has minimum spatial frequency and
minimum contrast, and so it would be located at the top-left corner of the CSF diagram in
Figure 101. Conversely, the most objectively complex greyscale image (composed of many
different pixels, each of which is a random shade of grey from black to white) has high
spatial frequency and high contrast, and so it would be located at the bottom-right corner
of the CSF diagram. Therefore, both images lie just outside our window of visibility, which
leads to a somewhat surprising conclusion: The simplest image and the most complex
image (as defined by an informational measure of complexity such as file-compression) are
perceptually indistinguishable from each other.
Figure 101 Image showing variation in spatial frequency (increasing from left to right)
and contrast (increasing from top to bottom). The areas that appear plain grey are
patterns beyond the range of our perception in terms of these two attributes.
Generally, the visibly stripy parts appear to form an inverted U shape (the CSF curve),
the area beneath which equates to the window of visibility.
Conclusions
Implications
281
In this thesis, visual complexity is described as a variety of patterns at different

levels. If we understand each possible pattern as having a particular spatial frequency and
contrast (that is, as occupying a point under the CSF curve), then the window of visibility
describes the range of patterns that are detectable by the visual system and hence
potentially available to perception. In the results of our correlation between objective and
subjective visual complexity, the inverted U-shape curve of the spectrum of complexity
reflects the biological limitations of what is detectable and hence perceivable.
The Spectrum of Complexity and Colour Spectra

With the knowledge that the opposite ends of the informational scale of visual
complexity are perceptually identical, we can understand visual complexity in a similar way
to the understanding of colour. Like colour, visual complexity is an aesthetic property that
has both objective and subjective aspects. Light is really out there in an objective sense,
and at the same time its phenomenal reality is tied to subjective perception. What we
perceive in vision is a range of colours which can be grouped perceptually into reds,
yellows and blues etc., but visible light is just one part of a much larger spectrum of
electromagnetic radiation. Similarly, we perceive a spectrum of visual complexity that has
qualitatively distinct phases order, repetition, complexity and chaos. In both cases, it is
possible to measure the spectrum quantitatively from infra-red to ultra-violet and from
simple order to chaotic randomness. The objective colour spectrum can be observed in
rainbows and in white light refracted by a prism. The analogous spectrum of complexity is
represented by the schemas of Wolfram and Langton, and also by arranging image files in
order of their data compression (p.176, Figure 67), all of which have at their core a concept
of information. Data compression is actually quite a crude measure, but it is nevertheless
particularly useful in organising visual material and in understanding visual complexity.
In both colour and complexity, their objective scales of measurement form
continuous ranges with distinct end-points, and yet in perceptual terms these end-points
meet: The simplest image is a single block of colour, for example a uniform grey. The most
Conclusions
Implications
282
complex image would be composed of a great number of elements, each of a different

colour, ordered randomly. As we discussed above, such an image is perceived as a uniform
grey identical to the simplest possible image. With this kind of perceptual scale, it makes
sense to map it out as we perceive it. This is why we have colour-wheels for use in art and
design; the circular format of the colour spectrum corresponds to our perception, and it
makes more sense to present it this way than in the format of the objective colour
spectrum, which is laid out linearly. The implication of this understanding of complexity is
that the relationship of objective and subjective complexity is analogous to the relationship
of objective and subjective properties of colour. Therefore, colour models are potentially
useful in the understanding of complexity, and may perhaps be utilised in creative art
practice and in the graphical representation of empirical complexity data.
Maps, Measures and Models of Complexity

To return to measures of complexity, we have found a distinction between
information-based measures such as AIC, MDL, and data compression and intuitive
concepts of complexity such Gell-Manns effective complexity (EC). This distinction is
most clearly expressed with the example of randomness, which is rated high by AIC and
low by EC. The results of this project offer empirical evidence for the idea that the
perception of visual complexity is more closely related to EC than to AIC: The filecompression measure has been shown to be related to AIC, and performs as we expected,
showing a correlation with perceived complexity of ordered images which diverges as
images approach randomness. The images with the largest file size and lowest compression
ratio appear to be almost random. Although they are rated highest by the objective
measure, they are perceived to be fairly simple.
We can understand this situation in terms of Batesons definition of information as
a difference that makes a difference (2002 p.459). A random pattern provides an overall
texture to an image that varies greatly from one element to the next, but which is generally
quite consistent overall. This relates back to the description in Chapter 1 of randomness as
Conclusions
Implications
283
smooth variation and chaos as lumpy. If we were presented with two or three different
random images made of black and white pixels, we would probably find it fairly difficult to
distinguish between them. From the results we can also conclude that the perceived
consistency of random images is dependent on their resolution; greater visible resolution
makes a more consistent texture overall. Batesons definition offers an explanation for this:
The smooth differences in random patterns may be perceptually indistinguishable because
these consistent differences (between adjacent picture elements) make little difference
(overall) to us perceptually. In contrast, what we perceive as a complex image has lots of
differences that do make a difference. Therefore, a complex image has lots of differentiable
patterns at various scales, which are perceived as many interconnected levels of pictorial
organisation.
In constructing models of complexity, we must not forget that the map is not the
territory. Batesons definition of information acts as a bridging statement between map
and territory, between the model and that which is modelled. The things that we map are
the perceivable features that make a difference to us. What makes it onto a map is not
arbitrary it is the stuff that actually matters. This is why medical artists are sometimes
employed to draw instead of photograph features of illness and disease, because they are
able to filter the available visual information and record what is most pertinent and useful
for identification and treatment. The significance of Batesons bridging statement for this
thesis is that the file-compression measure of visual complexity is based on an image file
which is in effect a map of a visual territory. The features encoded in this map are the
number, colour and location of pixels, which are also the fundamental attributes of visual
complexity as identified in this thesis (quantity, variety and order of picture elements), but
more importantly they are precisely those features that make a difference to us. The
implication is that the current projects methods are supported with this understanding of
visual complexity and information by providing a theoretical bridge between the modelling
or measuring of perception and its action in the real world.
Conclusions
Implications
284
Impact on My Art Practice

This research project has had a mutually beneficial relationship with my art practice.
While the practice supports the empirical research by providing visual material for testing,
the research findings have also informed the practice and influenced its direction. In terms
of disciplinary areas, my work has grown beyond fine art and towards design for products
and decorative arts, while retaining mix of computational designs and craft techniques. My
experience with the commissions for the Minster School and the Deutsche Bank, together
with the findings of Taylor (2006) that certain fractal values can reduce physiological stress,
are stimulating explorations of decorative pattern using cellular automata programs for
potential application in workplace environments or even hospitals. The CA images would
make a good basis for textile design, since the patterns can be made to repeat, or not, as the
case requires. For 3D products, I have plans for a lampshade design based on a CA pattern
to be made in translucent coloured acrylic, and a prototype is under construction in Lego
bricks which is starting to look more interesting than the original design for laser-cut acrylic
(Figure 102).
Figure 102 Design sketch for lampshade based on Lego block cellular automata.
Conclusions

285
Even though I have been an amateur musician for some time, until recently music
had never entered my art practice, except indirectly as a soundtrack to its making. In 2008 I
showed around twelve artworks in the Gold Soundz exhibition curated by Geoff Diego
Litherland which was themed around artwork inspired by music, especially music played
with guitars often quite loudly (Figure 103). In addition to paintings, drawings and prints,
which were mostly based on CA programs, my work included a digital animation. The
video was based on the Electric Sheep fractals made with Draves Fractal Flames
algorithm, and the soundtrack was commissioned from Dr. Torben Smith, a chemistry
teacher and amateur electronic musician with experience in music for film. Since the
animation formed a visual loop composed of three shorter cycles, the requirement for the
music was that it should also loop around in time with the visuals. Torbens music was
based on the experience of a cycle of waking and sleeping when hitting the snooze button
on an alarm clock and the accompanying perceptual distortion of time. The aim of the
animation was to explore the perceptual effect of Chions (1994) concept of synchresis,
which had come to my attention through the contextual mapping in this project. The
individual pieces were fine, but in combination the result lacked the synchresis effect.
Figure 103 Gold Soundz exhibition (September, 2008), Southwell Artspace.
Before working on these animations, my practice had not involved time-based media.
The cellular automaton programs employed in the practice do involve an element of time
in their computation, but in the artwork up until now this temporal dimension had been
transformed into a spatial dimension. This sounds quite abstract, but actually all the images
Conclusions

286
of cellular automata in this thesis have this property the vertical spatial dimension
represents the temporal sequence of the pattern evolution. In the same way that the
evolution of 1D CA is represented in two dimensions, some of my artwork involved
representations of 2D CA in three dimensions. In contrast, this latest artwork explores the
three dimensions of moving image, which involves 2D space plus time.
Figure 104 S[4]n, (2009), digital animation, frame 278/9600.
Some of the animations are created with the Fractal Flames algorithm have led to the
conclusion that these forms may be fractal in time as well as fractal in space. As part of the
Electric Sheep project, these animations are normally restricted to 128 frames and last just a
few seconds. For my creative experiments I stretched out the animations by generating
more frames, up to almost ten thousand in one case (Figure 104). As fractals, the shapes
generated by this algorithm have a potentially infinite depth; one can zoom into an area to
find more detail, and so on ad infinitum. I extended the duration of these animated images,
which in effect is a magnification in time analogous to the magnification of spatial detail.
The magnification in time of this fractal reveals greater detail that is otherwise un-
Conclusions

287
perceivable; as the animations duration is extended, previously fuzzy areas are resolved
into fast-moving particles or vibrating strings. The implication of this observation is that
the animated forms produced by Draves algorithm are fractal in time as well as in space.
This is not just a fractal distribution of events in time, as in the random pattern of noise
bursts in audio circuits, but a kind of fractal motion. It means that whatever frame rate we
choose to use with this algorithm, there will always be visible some parts moving quickly
and some moving slowly in fact, there will always be parts moving at all speeds, but we
can only ever perceive part of this range. I intend to explore this further, to research the
concept of time-fractals and to understand the algorithm in order to determine whether
this is indeed the case and, if so, how it works.
Figure 105 P32 t+r sb3, (2009), animation frame 17/2400.
The current work is exploring techniques of sound design to generate audio for the
animations. The idea is to control audio parameters with measurements of image properties
derived from the analytical methods of this project. By measuring the file size of each
frame of animation (such as Figure 104 and Figure 105), we have a value that can be used
to modulate the pitch or tone of a sound at a time corresponding to that frame in the
animation sequence. The result is a synchronisation of visual and musical elements.
Similarly, the number of unique colours could be used, or any other aspect of digital image
files that is able to be quantified can be mapped to any audio control parameter. In light of
the discussion in Chapter 2 about synchresis and synaesthesia, we understand that any
transformation between auditory and visual stimuli is ultimately arbitrary, but in this case
the connection is grounded on a relation between objective properties, and the point is that
we can use these processes to experiment with the subjective effects of synchresis. This
Conclusions

288
work is evidence of the ways in which the empirical research techniques and findings of
this project contribute to the development of aesthetic practice.
Future Directions
Since the results of this project were inconclusive about the potential for the filecompression measure with images of art and design, we cannot rule out the potential value
of further investigations along these lines. In fact, the success of the measure up to the
introduction of art and design images suggests that we should refine the tests to establish
what led to the masking of the correlation in this case. Future refinements for tests 1 and 2
would include the inclusion of greater randomness and finer detail in test stimuli in the
perception of CA and random images, as well as the introduction of more numbers of
colours. A repeat of test 3 would reduce the variety of styles and types of visual artefact
whilst maintaining a wide spectrum of complexity. The lack of a wide objective spectrum
of complexity in this sample is evidenced in the almost-vertical clustering of data points in
the results of test 3, so we should address the issue in future.
It would be interesting to explore more about the aesthetic production of complexity
than could be accommodated in the current project. For instance, one idea for my own art
practice is for a series of paintings to begin with a blank canvas and then to make each
successive painting more complex than the last. This idea could be used to explore
aesthetic production by setting it as a task to art students, perhaps creating one drawing per
day in a sketchbook. The collected sketchbooks would offer a fascinating insight into ways
of understanding aesthetic complexity. As a basic visual analysis, we could lay out the pages
in sequence from beginning to end, that is, from the simple to the complex, with each
students work in rows. Would we see the appearance of randomness anywhere in these
complexity spectra, and might we find anything like the complexity classes of cellular
automata? These appear to be potentially rewarding areas for future research.
Conclusions
Future Directions
289
It seems that the same methods and approach employed in the current project could
be extended from visual art to music and be applied to an investigation of the aesthetic
complexity of sound. Precisely the same psychophysical techniques would be applicable to
the perception and measurement of audio stimuli as they were applied to visual stimuli.
Similarly, the same technique of data compression could be used as the basis of a measure
of audio file complexity. There are analogies between types of digital file for audio and
visual applications: both come in lossy and non-lossy formats. Lossy compression discards
perceptually insignificant information, and so there is a fundamental similarity between
lossy JPEG image file compression and lossy MP3 audio file compression. A significant
difference between audio complexity and visual complexity is that the former is time-based.
In this project, we disregarded the temporal dimension in our study of visual aesthetics,
though it was noted that complexity is related to difficulty of description in terms of the
time it takes to create or understand it. In one sense, musical analysis might be simpler
because it is oriented mainly along this single temporal dimension, which simply encodes a
series of positive and negative values which express the amplitude of a waveform (and the
movement of a speaker cone). This might make it seem less complex than encoding twodimensional images, but the variety of perceived pitches, timbres and rhythms make sound
a highly complex perceptual phenomenon. Whether a simple file-compression measure
could effectively capture this audio complexity, and whether it might correlate with
perceptions of musical complexity, are identified as potential areas for further investigation.
With the experience and techniques developed in the current project, it would be a natural
progression and a relatively small step to change focus from the aesthetics of visual
complexity to the complexity of music and sound.
With a refinement of the current findings on visual complexity, and potential work
on audio complexity, we would have a platform to investigate more fully the perception of
aesthetic complexity in audio-visual media. It is possible that the effect of synchresis is
sensitive to complexity. My hypothesis is that the perceptual effect is strongest when the
Conclusions
Future Directions
290
levels of complexity in the image and sound are approximately the same, whether at a static
level or varying synchronously. This hypothesis is based on my experience and on the
analysis and evaluation of the examples of art practice evidenced in this thesis. If we could
establish the measurement of audio-visual complexity with file-compression and measure
the aesthetic perception of synchresis with MES, we would be able to investigate the
relationship between the two.
As a final word to this thesis, we can say that the visual complexity of artwork
affords the opportunity of aesthetic exploration, but too little can be dull, and too much
can be confusing. Like life on earth, the practice of art is an endlessly creative and
increasingly complex activity. Art offers an opportunity for both artists and audiences to
reflect on the process of perception itself. We now understand visual complexity as a
demanding and rewarding perceptual task. Through its difficulty in making and
understanding, visual complexity provides both a rich aesthetic experience and an
opportunity for reflection and insight. Perception involves making sense of the world, in
terms of both understanding and forming sensory percepts. Seeing is understanding.
Therefore, understanding how we see that is, understanding the perception and aesthetics
of visual art allows for insights into ourselves and the world around us.
To see clearly is poetry, prophecy, and religion, all in one.
John Ruskin (1856) Modern Painters, vol. III, part IV, chapter XVI.
Conclusions
Future Directions
291
References
Adams, D.N. (1998). Is there an Artificial God? Transcription of speech given at Digital Biota
2, Oxford, UK, 2nd. September 1998 [online].
http://www.biota.org/people/douglasadams/index.html [Accessed 3 July 2008].
Aks, D.J. and Sprott, J.C. (1996). Quantifying Aesthetic Preference for Chaotic Patterns.
Empirical Studies of the Arts, 14 (1): 116.
Arnheim, R. (1959). Information Theory: An Introductory Note. The Journal of Aesthetics and
Art Criticism, 17 (4): 501503.
Arnheim, R. (1974). Art and Visual Perception. Berkley: University of California Press.
Attneave, F. (1954). Some informational aspects of visual perception. Psychological Review,
61(3): 183193.
Attneave, F. (1959a). Applications of Information-Theory to Psychology: A Summary of Basic
Concepts, Methods, and Results. New York: Henry Holt.
Attneave, F. (1959b). Stochastic Composition Processes. The Journal of Aesthetics and Art
Criticism, 17 (4): 503510.
Barron, F.B. (1952). Artistic Preferences as a Factor in Personality Style. Journal of Psychology,
33: 199203.
Barron, F.B. (1963). Creativity & Psychological Health: Origins of Personal Vitality and Creative
Freedom. New York: Van Nostrand.
Barrow, J.D. (2003). Art and Science Les Liaisons Dangereuses? In: Casti, J. L. and
Karlqvist, A. (eds.), 2003. Art and complexity. Amsterdam; London: Elsevier.
Bateson, G. (2002). Mind and nature: A necessary unity. Cresskill, N.J.: Hampton Press.
Beardsley, M.C. (1982). The Aesthetic Point of View: Selected Essays. Ithaca, New York;
London: Cornell University Press.
Benjamin, W. (1936). The Work of Art in the Age of Mechanical Reproduction. In:
Benjamin, W. and Arendt, H. (1992). Illuminations. Translated by Harry Zohn.
London: Fontana.
References
292
Bennett, T. (1996). Really useless knowledge: A political critique of aesthetics. In: Palmer,
J. & Dodson, M. (eds.) Design and Aesthetics: A Reader. London: Routledge. pp.3348.
Berleant, A. (1970). The Aesthetic Field: A Phenomenology of Aesthetic Experience. Springfield, Ill.:
Thomas.
Berlyne, D.E. (1971). Aesthetics and Psychobiology. New York: Appleton-Century-Crofts.
Besner, C. (2007). Studies of Visual Complexity. M.Sc. thesis (unpublished): McGill University,
Canada.
Best, L.A., Smith, L.D., & Stubbs, A.D. (2001). Graph Use in Psychology and other
Sciences. Behavioural Processes, 54: 155165.
Birkhoff, G.D. (1932). Aesthetic Measure. Harvard University Press.
Bishop, C. (2004). Antagonism and Relational Aesthetics. October. 110: 5179.
Borg, G. & Borg, E. (2001). A New Generation of Scaling Methods: Level-Anchored Ratio
Scaling. Psychologica, 28: 1545.
Borthwick, B. (2007). The Perfect Storm. Interview with Olaf Bender, Frank Bretschneider,
Carsten Nicolai [online]. The Wire, 238. Available at:
http://www.thewire.co.uk/articles/185/ [accessed 27 June 2009].
Bourriaud, N. (2002). Relational Aesthetics. Dijon: Les Presses du Rel.
Bourriaud, N. (2005). Postproduction: Culture as Screenplay: How Art Reprograms the World.
Second edition. New York: Lukas & Sternberg.
Bruce, V., Green, P.R. and Georgeson, M.A. (2003). Visual Perception: Physiology, Psychology
and Ecology. Fourth edition. New York: Psychology Press.
Byron, L. and Wattenberg, M. (2008). Stacked Graphs Geometry & Aesthetics. IEEE
Transactions on Visualization and Computer Graphics, 14 (6): 12451252.
Carroll, N. (1995). Philosophy of Art: A Contemporary Introduction. London: Routledge.
Carroll, N. (2000). Theories of Art Today. Madison; London: University of Wisconsin Press.
Chaitin, G.J. (1966). On the Length of Programs for Computing Finite Binary Sequences. J.
Assoc. Computing Machinery. 13: 547569.
Chalmers, D.A. (1995). Facing Up to the Problem of Consciousness. Journal of Consciousness
Studies, 2(3): 200219.
References
293
Chion, M. (1994). Audio-vision: sound on screen. Edited and translated by Claudia Gorgman.
New York, Chichester: Columbia University Press.
Cornock, S. and Edmonds, E. (1973). The Creative Process where the Artist is Amplified
or Superseded by the Computer. Leonardo. 6: 1116.
Copyright, Designs and Patents Act (1998) [online]. Available at:
1988.http://www.opsi.gov.uk/acts/acts1988/ukpga_19880048_en_1 [Accessed 4th.
September, 2008].
Crutchfield, J.P. (1990). Chaos and Complexity. In: Burkhardt, H. and Smith, B. (eds.),
1991. Handbook of Metaphysics and Ontology. Mnchen: Philosophia Verlag.
Cupchik, G.C. (1983). The Scientific Study of Artistic Creativity. Leonardo, Special Issue:
Psychology and the Arts, 16 (3): 193195.
DACS (Design and Artists Copyright Society), (2008). [online] Available at:
http://www.dacs.org.uk/index.php?m=15 [accessed 4th. September, 2008].
Danto, A.C. (1964). The Artworld. The Journal of Philosophy. 61(19): 571584.
Danto, A.C. (1981). The Transfiguration of the Commonplace. Cambridge, MA; London:
Harvard University Press.
Danto, A.C. (1993). Responses and Replies. In: Rollins, M. (ed.), Danto and his Critics.
London: Blackwell, pp.193216.
Design Museum (2007). Dieter Rams. Selector for 25/25 - Celebrating 25 Years of Design. [online]
Available at: http://designmuseum.org/design/dieter-rams [accessed 8 August
2009].
Deupree, T. (2009). The 12k aesthetic. In: The 12k Forum [online]. Thread started: June 30.
Available at: http://12k.com/forum/index.php [accessed 23 July 2009].
Dickie, G. (1961). Is Psychology Relevant to Aesthetics? Philosophical Review. 71 (3): 285
302.
Dickie, G. (1971). Aesthetics: An Introduction. USA: Pegasus.
Dickie, G. (1974). Art and the Aesthetic: An Institutional Analysis. Ithaca; London: Cornell
University Press.
Dickie, G. (1997). The Art Circle: A Theory of Art. Evanston, Ill.: Chicago Spectrum Press.
References
294
Dietrich, F. (1986). Visual Intelligence: The First Decade of Computer Art (19651975).
Leonardo. 19 (2): 159169.
Dodgson, N.A. (2008) Regularity and randomness in Bridget Riley's early Op-art. In:
Brown, Cunningham, Interrant, McCormack, eds., Proceedings of Eurographics Workshop
on Computational Aesthetics in Graphics, Visualization, and Imaging. pp.107114.
Domnitch, E. and Gelfand, D. (2004). Camera Lucida: A Three-Dimensional
Sonochemical Observatory. Leonardo. 37 (5): 391396.
Donderi, D.C. (2000). A complexity measure for visual graphic displays. Paper presented at the
Experimental Psychology Society/Canadian Society for Brain, Behavior and
Cognitive Science Joint Meeting, Cambridge, U.K.
Donderi, D.C. (2006a). An information theory analysis of visual complexity and
dissimilarity. Perception, 35 (6): 823835.
Donderi, D.C. (2006b). Visual Complexity: A Review. Psychological Bulletin, 132 (1): 7397.
Donderi, D.C. (2007). Re: Aesthetics & Complexity, 26 June. (don.donderi@mcgill.ca) Email
to author (guy.birkin@ntu.ac.uk).
Donderi, D.C. and McFadden, S. (2004). A single marine overlay is more efficient than
separate chart and radar displays. Displays. 24 (45): 147155.
Donderi, D.C. and McFadden, S. (2005). Compressed file length predicts search time and
errors on visual displays. Displays. 26 (22): 7178.
Draves, S. (2009). Electric Sheep [online]. Available at: http://community.electricsheep.org/
[Accessed 3 June 2009].
Draves, S. and Reckase, E. (2007). The Fractal Flames Algorithm. [online] Available at:
http://flam3.com/flame_draves.pdf [accessed 13 August 2009]
Draves, S., Abraham, R., Viotti, P., Abraham, F.D. and Sprott, J.C. (2008). International
Journal of Bifurcation and Chaos. 18 (4): 12431248.
Dunne, A. (1999). Hertzian Tales: electronic products, aesthetic experience and critical design.
London: RCA CRD Research Publications.
Edmonds, E.A. and Fell, M. (2005). Kyoto to Sheffield. In: Proceedings of the 5th Conference on
Creativity and Cognition. Goldsmiths College, London, 1215 April, 2005.
References
295
Eisenman, R. (1967). Complexity-Simplicity Creativity, Intelligence and Other Correlates.

Journal of Psychology. 67 (33): 134.
Elkins, J. (2008). Six Stories from the End of Representation: images in painting, photography,
astronomy, microscopy, particle physics, and quantum mechanics, 1980-2000. Stanford,
Calif.: Stanford University Press.
Eysenck, H.J. (1941a). General Factor in Aesthetic Judgments. British Journal of Psychology.
31: 94102.
Eysenck, H.J. (1941b). The Empirical Determination of an Aesthetic Formula. Psychological
Review. 48: 8392.
Eysenck, H.J. (1941c). A Critical and Experimental Study of Colour Preferences. The
American Journal of Psychology. 54 (3): 385394.
Eysenck, H.J. (1971). Factors Determining Aesthetic Preferences for Geometrical Designs
and Devices. British Journal of Aesthetics, 11 (2): 154166.
Fechner, G. T. (1966). Elements of Psychophysics, Vol. 1. Translated by H. E. Adler; Edited by
D. H. Howes & E. G. Boring. New York: Holt, Rinehart and Winston. (Original
work published 1860.)
Fell, M. (2009). Re: Art and Complexity. 12th June. From: mark@markfell.com Personal
communication to: guy.birkin@ntu.ac.uk
Flake, G.W. (1990). The Computational Beauty of Nature: computer explorations of fractals, chaos,
complex systems, and adaptation. Cambridge, Mass.; London: MIT Press.
Forsythe, A. (2008). Confounds in Pictorial Sets: The role of complexity and familiarity in
basic-level picture processing. Behavior Research Methods. 40 (1): 116129.
Forsythe, A. Sheehy, N. and Sawey, M. (2003). Measuring Icon Complexity: An Automated
Analysis. Behavior Research Methods, Instruments, & Computers. 35 (2): 334342.
Friedman, K. (1997). Design Science and Design Education. In: McGrory, P., ed. The
Challenge of Complexity. Helsinki: University of Art and Design Helsinki UIAH, 5472.
Frye, N. (1969). Fearful Symmetry: A Study of William Blake. Princeton, N.J.: Princeton
University Press.
Fucks, W. (1952). On Mathematical Analysis of Style. Biometrica. 39 (12): 122129.
References
296
Gell-Mann, M. (1995). What is Complexity? Complexity. 1 (1): 1619.

Gell-Mann, M. (2003). Regularities and Randomness: Evolving Schemata in Science and
the Arts. In: Casti, J. L. and Karlqvist, A. (eds.), 2003. Art and complexity. Amsterdam;
London: Elsevier.
Gianetti, C. (2004). Cybernetic Aesthetics and Communication [online]. Available at:
http://www.mediaartnet.org/themes/aesthetics_of_the_digital/cybernetic_aesthetic
s/ [accessed 24 June 2008].
Gibson, J.J. (1950). The Perception of the Visual World. Boston, Massachusetts: HoughtonMifflin.
Gibson, J.J. (1978). The Ecological Approach to the Visual Perception of Pictures.
Leonardo, 11: 227235.
Gibson, J.J. (1971). The Information Available in Pictures. Leonardo. 4: 2735.
Gibson, J.J. (1986). The Ecological Approach to Visual Perception. Dallas; London: Houghton
Mifflin.
Gleick, J. (1993). Chaos: Making a New Science. New edition. London: Abacus.
Gombrich, E.H. (1977). Art and Illusion: A study in the psychology of pictorial representation. Fifth
edition. London: Phaidon.
Grunbaum, B. and Shephard, G.C. (1987). Tilings and Patterns. New York: W.H. Freeman.
Hagerhall, C.M., Laike, T., Taylor, R.P., Kller, M., Kller, R., and Martin, T.P. (2008).
Investigations of human EEG response to viewing fractal patterns. Perception. 37 (10):
14881494.
Henke, R. (2009). Re: Aesthetics and Complexity. 17th December. From: henke@monolake.de
Personal communication to: guy.birkin@ntu.ac.uk
Howell, D.C. (2007). Statistical Methods for Psychology. 6th. ed. Belmont, Ca., USA:
Thomson/Wadsworth.
Hubel, D.H. and Wiesel, T.N. (1962). Receptive Fields, Binocular Interaction and
Functional Architecture in the Cats Visual Cortex. J. Phyisiol. 160: 106154.
Ikeda, R. (2009). Test Pattern [online]. Available at:
http://www.ryojiikeda.com/test+pattern/ [accessed 30 September 2009]
References
297
Kahn, D. (2007). Semiconductors Magnetic Movie [online]. Available at:

http://www.semiconductorfilms.com/root/Magnetic_Movie/Magnetic.htm
[accessed 3 January 2008].
Kant, I. (2007). Critique of Judgement. Translated by James Creed Meredith. Oxford: Oxford
University Press.
Kauffman, S.A. (2003). The Emergence of Autonomous Agents. In: Gregerson, N.H. (ed.)
From Complexity to Life. New York: Oxford University Press, pp. 4771.
Kauffman, S.A. (2008). Reinventing the Sacred: A New View of Science, Reason and Religion. New
York: Basic Books.
Kolmogorov, A.N. (1965). Three Approaches to the Quantitative definition of
Information. Problems of Information Transmission. 1: 17.
Krauss, R. (1985). The Grid. In: The Originality of the Avant-Garde and Other Modernist Myths.
Cambridge, MA: MIT Press. 922.
Krippendorf, K. (2004). Content Analysis: An Introduction to Its Methodology. 2nd. ed. Thousand
Oaks, CA, USA: Sage Publications.
Kuhn, T.S. (1996). The Structure of Scientific Revolutions. 3rd edition. Chicago; London:
University of Chicago Press.
Laming, J. and Laming, D. (1996). J. Plateau: On the measurement of physical sensation
and on the law which links the intensity of these sensations to the intensity of the
source. / J. Plateau: Report on Psychophysical study: Theoretical and experimental
research on the measurement of sensations, particularly sensations of light and of
fatigue by Mr. Delboeuf. (originally published 1872). Psychological Research. 59 (2):
134144.
Langton, C.G. (1990). Computation at the edge of Chaos: Phase-Transitions and Emergent
Computation. Physica D. 42: 1237.
Langton, C.G. (1991). Computation at the edge of Chaos: Phase-Transitions and Emergent
Computation. Ph.D. thesis, University of Michigan.
Levinson, J. (2003). Philosophical Aesthetics: An overview. In: Levinson, J. (ed.), The
Oxford Handbook of Aesthetics, pp.324. Oxford; New York: Oxford University Press.
References
298
Li, W. and Packard, N. (1990). The Structure of the Elementary Cellular Automata Rule
Space. Complex Systems. 4: 281297.
Lind, R. (1995). The Aesthetic Essence of Art. The Journal of Aesthetics and Art Criticism. 50
(2): 117129.
Locher, P. (1995). A Measure of the Information Content of Visual Art Stimuli for Studies
in Experimental Aesthetics. Empirical Studies of the Arts, 13 (2): 183191.
Locher, P. and Dolese, M. (2004). A Comparison of the Perceived Pictorial and Aesthetic
Qualities of Original Paintings and their Postcard Images. Empirical Studies of the Arts.
22 (2): 129142.
Locher, P., Smith, L., and Smith, J. (2001). The influence of presentation format and viewer
training in the visual arts on the perception of pictorial and aesthetic qualities of
paintings. Perception. 30 (4): 449465.
Luhmann, N. (2000). Art as a Social System. Translated by Eva M. Knodt. Stanford, Calif:
Stanford University Press.
Mandelbaum, M. (1965). Family Resemblances and Generalization Concerning the Arts.
American Philosophical Quarterly. 2(3): 219228.
Mandelbrot, B.B. (1982). The Fractal Geometry of Nature. Oxford: Freeman.
Mandelbrot, B.B. (2004a). Fractals and Chaos: The Mandelbrot Set and Beyond. New York:
Springer.
Mandelbrot, B.B. (2004b). A Theory of Roughness. Edge [online]. Available at:
http://www.edge.org/3rd_culture/mandelbrot04/mandelbrot04_index.html
[accessed 16 February 2009]
Manovich, L. (2008). Software Takes Command [online]. (s.l.): (s.n.) Available at:
www.softwarestudies.com/softbook [accessed 28 June 2009].
Martindale, C., Moore, K., & Borkum, J. (1990). Aesthetic Preference: Anomalous Findings
for Berlynes Psychobiological Theory. The American Journal of Psychology. 103 (1): 53
80.
Massimi, M. et al. (2005). Breakdown of Cortical Effective Connectivity During Sleep.
Science. 309 (30): 22282232.
Matossian, N. (1986). Xenakis. London: Kahn & Averill.
References
299
Maturana, H.R. & Varela, F.J. (1980). Autopoiesis and cognition: the realization of the living.
London: Dordrecht.
McWhinnie, H.J. (1965). A Review of Some Research on Aesthetic Measure and Perceptual
Choice. Studies in Art Education. 6 (2): 3441.
McWhinnie, H.J. (1966). Effects of Learning on Preference for Complexity. Perceptual and
Motor Skills. 23: 11922.
McWhinnie, H.J. (1971). A Review of Selected Aspects of Empirical Aesthetics III. Journal
of Aesthetic Education. 5 (4): 115126.
Miall, R.C. and Tchalenko, J. (2001). A Painter's Eye Movements: A Study of Eye and
Hand Movement during Portrait Drawing. Leonardo. 34 (1): 3540.
Miall, R.C. and Tchalenko, J. (2009). Eye-hand strategies in copying complex lines. Cortex.
45 (3): 368376.
Miller, W. (2008). Gabriel Orozco: Mobile Matix. [online]. Available at:
http://blog.art21.org/2008/09/18/gabriel-orozco-mobile-matrix/ [Accessed 5
September 2008].
Moles, A. (1966). Information Theory and Aesthetic Perception. London: University of Illinois
Press.
Morton, B. and Collins, P., eds. (1992). Contemporary Composers. Chicago; London: St. James.
Mureika, J.R., Cupchik, G.C. & Dyer, C.C. (2004). Multifractal fingerprints in the visual
arts. Leonardo. 37(1): 5356.
Nake, F. (2005). Computer Art. A Personal Recollection. In: C&C 05: Proceedings of the 5th
Conference on Creativity & Cognition. London, 2005. 5462.
Narens, L. (1996). A theory of ratio magnitude estimation. Journal of Mathematical Psychology.
40 (2): 109209.
Nicolai, C. (2009a). Artists website [online]. Available at: http://www.carstennicolai.de/
[accessed 14 August 2009].
Nicolai, C. (2009b). Grid Index. Berlin: Die Gestalten Verlag.
References
300
Nicolis, G. and Rouvas-Nicolis, C. (2007). Complex Systems [online]. Scholarpedia, 2

(11):1473. http://www.scholarpedia.org/article/Complex_systems [accessed 12 June
2008].
Norman, D.A. (1988). The Psychology of Everyday Things. New York: Basic Books.
Norman, D.A. (1998). The Design of Everyday Things. London: MIT Press.
Norman, D.A. (2009). Affordances and Design [online]. Available at:
http://www.jnd.org/dn.mss/affordances_and.html [accessed 23 October 2009].
OED (2009). Oxford English Dictionary [online]. http://dictionary.oed.com/
Ostwald, W. (1969). The Color Primer: a basic treatise on the color system of Wilhelm Ostwald. New
York; London: Van Nostrand Reinhold.
Ox, J. (2002). 2 Performances in the 21st Century Virtual Color Organ. Proceedings of the 4th
conference on Creativity & Cognition. Loughborough, UK. 2024.
Ox, J. and Britton, D. (2000). The 21st Century Virtual Reality Color Organ. IEEE
MultiMedia. 7(3): 69.
Persaud, N. (2005). Humans can consciously generate random number sequences: A
possible test for artificial intelligence. Medical Hypotheses. 65 (2): 211214.
Philips, E.M., and Pugh, D.S. (2000). How to Get a PhD: A Handbook for Students and their
Supervisors. Buckingham: Open University Press.
Popper, K. (2002). The Logic of Scientific Discovery. London: Routledge.
Pynchon, T.R. (1998). Vineland. London: Vintage.
Reber, R. (2008). Art in Its Experience: Can Empirical Psychology Help Assess Artistic
Value? Leonardo. 41(4): 367372.
Reichardt, J. (1971). Cybernetics, Art and Ideas. London: Studio Vista.
Roberts, M.N. (2007). Complexity and Aesthetic Preference for Diverse Visual Stimuli. PhD thesis.
Universitat de les Illes Balears.
Rousset, I. (2009). In the Room with SND, Transmediale 09. Berlin. [online] Available at:
http://www.derivative.ca/Events/SNDTransmediale/ [accessed 23 June 2009].
Saldaa, J. (2009). The Coding Manual for Qualitative Researchers. London: Sage Publications.
References
301
Salomon, D. (2004). Data Compression: The Complete Reference. London: Springer.

Schmidhuber, J. (1997). Low Complexity Art. Leonardo. 30 (2): 97103.
Semiconductor (2009). Artists website [online]. Available at:
http://www.semiconductorfilms.com/index.html [accessed 3 December 2009].
Shannon, C.E. (1948). A Mathematical Theory of Communication. Bell System Technical
Journal, 27: 379423; 623656.
Shannon, C.E. and Weaver, W. (1948). The Mathematical Theory of Communication. Urbana:
University of Illinois Press.
Siegel, R.M. (2004). Choices: The Science of Bela Julesz. PLoS Biol [online] 2(6): e172.
Available at:
http://www.ncbi.nlm.nih.gov/pmc/articles/PMC423145/pdf/pbio.0020172.pdf
[accessed 10 June 2008].
Silverman, K. (2006). Ken Silvermans Utility Page [online]. Available at:
http://advsys.net/ken/utils.htm [accessed 27 November 2009].
Simon, H.A. (1996). The Sciences of the Artificial. 3rd. edition. Cambridge, Mass.: MIT Press.
Skiljan, I. (2009). IrfanView [online]. Available at: http://www.irfanview.com/ [accessed 27
November 2007].
Snowden, R., Thompson, P., and Troscianko, T. (2006). Basic Vision: An introduction to visual
perception. New York: Oxford University Press.
Solomonoff, R. (1964). A Formal Theory of Inductive Inference, Part 1 and Part 2.
Information and Control. 7: 122, 224254.
Sprott, J.C. (1994). Automatic Generation of Iterated Function Systems. Comput. &
Graphics. 18 (3): 417425.
Stevens, S.S. (1957). On the psychophysical law. Psychological Review, 64 (3): 153181.
Stevens, S.S. (1961). To Honour Fechner and Repeal His Law. Science, New Series, 133
(3446): 8086.
Stevens, S.S. (1975). Psychophysics. Introduction to its perceptual, neural, and social prospects. New
York: Wiley-Interscience.
References
302
Stiles, K. and Selz, P., eds. (1996). Theories and Documents of Contemporary Art: A Sourcebook of
Artists Writings. Berkeley, Calif.; London: University of California Press.
Stber, K. (1999). Ernst Haeckel: Kunstformen der Natur 18991904 [online]. Available at:
http://caliban.mpiz-koeln.mpg.de/haeckel/kunstformen/natur.html [accessed 24
September 2006]
Taylor, R.P. (2006). Reduction of physiological stress using fractal art and architecture.
Leonardo. 39 (3): 245251.
Taylor, R.P., Micolich, A.P., and Jonas, D. (2002). The Construction of Jackson Pollocks
Drip Paintings. Leonardo. 35 (2): 203207.
Taylor, R.P., Spehar, B., Wise, J.A., Clifford, C.W.G., Newell, B.R. and Martin, T.P. (2003).
Perceptual and Physiological Responses to the Visual Complexity of Pollocks
Dripped Fractal Patterns. Journal of Nonlinear Dynamics, Psychology, and Life Sciences. 9
(1): 89114.
Teghtsoonian, R. (2001). S.S. Stevens. In: Smelser, N.J. & Baltes, P.B. (Eds.), International
Encyclopedia of the Social and Behavioral Sciences. Elsevier.
Tufte, E.R. (2001). The Visual Display of Quantitative Information. Second edition. Cheshire,
Conn.: Graphics Press.
Visual Complexity (2009). [online] Available at: http://www.visualcomplexity.com/vc/
[accessed 24 June 2009].
Walther, E. (2000). Max Benses Informational and Semiotical Aesthetics. [online] Available at:
http://www.stuttgarter-schule.de/bense.html [Accessed 4 June 2008]
Ware, C. (2004). Information Visualization: Perception for Design. Second edition. San Francisco,
CA: Morgan Kaufmann.
Ware, C., Arsenault, R., Plumlee, M., and Wiley, D. (2006). Visualizing the underwater
behavior of humpback whales. Computer Graphics and Applications, IEEE. 26 (4): 14
18.
Wattenberg, M. (2002). Arc Diagrams: Visualizing Structure in Strings. Information
Visualization, 2002. INFOVIS 2002. IEEE Symposium on Visualizing Information.
pp.110116. [online] Available at: http://www.research.ibm.com/visual/papers/arcdiagrams.pdf [accessed 12 June 2008].
References
303
Weisstein, Eric W. (2009). Cellular Automaton [online]. Available at:

http://mathworld.wolfram.com/CellularAutomaton.html [accessed 30 May 2009]
Weitz, M. (1956). The Role of Theory in Aesthetics. The Journal of Aesthetics and Art Criticism,
15 (1): 2735.
Whitelaw, M. (2007). A Conversation with Mark Fell. Thursday, November 08, 2007. [online].
Available at: http://teemingvoid.blogspot.com/2007/11/conversation-with-markfell.html [accessed 12 June 2009].
Wiener, N. (1948). Cybernetics. New York: Wiley.
Wilson, E.O. (1999). Consilience: the unity of knowledge. London: Abacus.
Wittgenstein, L. (1966). Lectures and Conversations on Aesthetics, Psychology, and
Religious Belief. Ed. Cyril Barrett. Oxford: Basil Blackwell.
Wittgenstein, L. (1972). The Blue and Brown Books (Preliminary studies for the Philosophical
Investigations). Oxford: Blackwell.
Wittgenstein, L. (2001). Philosophical Investigations: The German Text, with a Revised English
Translation. 3rd Edition. Translated by G.E.M. Anscombe. Oxford: Blackwell.
Wixted, J., ed. (2002). Stevens Handbook of Experimental Psychology. 3rd. ed., Vol. 4:
Methodology in Experimental Psychology. New York; Chichester: Wiley.
Wolfram, S. (1984). Cellular Automata as Models of Complexity. Nature, 311: 419424.
Wolfram, S. (2002). A New Kind of Science. Champaign, Illinois: Wolfram Media, Inc.
Wundt, W.M. (1874). Grundzge der physiologischen Psychologie. Leipzig: Engelmann.
Zangwill, Nick (2008). Aesthetic Judgment. The Stanford Encyclopedia of Philosophy (Fall 2008
Edition) [online], Edward N. Zalta, ed.
http://plato.stanford.edu/archives/fall2008/entries/aesthetic-judgment/ [accessed
16 April 2008].
Zeki, S. (2009). Statement on Neuroaesthetics [online]. Available at:
http://www.neuroesthetics.org/ [accessed 23 June 2009]
Zeki, S. and Lamb, M. (1994). The Neurology of Kinetic Art. Brain. 117: 607636.
References
304
Appendices
Appendix A: Test 1 Instructions
The following text comprises the written instructions for test participants:
You will be presented with a series of images in a random order. Your task is to rate
the complexity of the images by assigning a number to each one to show how
complex you think it is. There will be a practice run to begin with.
Call the first image any number that seems appropriate to you. Then assign successive
numbers in such a way that they reflect your subjective impression. There is no limit
to the range of numbers that you may use. You may use whole numbers, decimals, or
fractions. Try to make each number match the property as you perceive it.
Appendix B: Copyright Law

Following are extracts from the Copyright, Designs and Patents Act 1988 (c. 48)
Chapter II Rights of Copyright Owner
16 The acts restricted by copyright in a work
(1) The owner of the copyright in a work has, in accordance with the following provisions of this
Chapter, the exclusive right to do the following acts in the United Kingdom
(a) to copy the work;
(b) to issue copies of the work to the public;
(c) to perform, show or play the work in public;
(d) to broadcast the work or include it in a cable programme service;
(e) to make an adaptation of the work or do any of the above in relation to an adaptation;
and those acts are referred to in this Part as the acts restricted by the copyright.
Chapter IV Moral Rights
77 Right to be identified as author or director (right of paternity)
(1) The author of a copyright literary, dramatic, musical or artistic work, and the director of a
copyright film, has the right to be identified as the author or director of the work in the
circumstances mentioned in this section; but the right is not infringed unless it has been asserted in
accordance with section 78.
Appendices
305
80 Right to object to derogatory treatment of work (right to integrity)

(1) The author of a copyright literary, dramatic, musical or artistic work, and the director of a
copyright film, has the right in the circumstances mentioned in this section not to have his work
subjected to derogatory treatment.
(2) For the purposes of this section
(a) treatment of a work means any addition to, deletion from or alteration to or adaptation
of the work, other than
(i) a translation of a literary or dramatic work, or
(ii) an arrangement or transcription of a musical work involving no more than a
change of key or register; and
(b) the treatment of a work is derogatory if it amounts to distortion or mutilation of the
work or is otherwise prejudicial to the honour or reputation of the author or director;
and in the following provisions of this section references to a derogatory treatment of a work shall
be construed accordingly.
(4) In the case of an artistic work the right is infringed by a person who
(a) publishes commercially or exhibits in public a derogatory treatment of the work, or
broadcasts or includes in a cable programme service a visual image of a derogatory treatment
of the work,
(b) shows in public a film including a visual image of a derogatory treatment of the work or
issues to the public copies of such a film, or
(c) in the case of
(i) a work of architecture in the form of a model for a building,
(ii) a sculpture, or
(iii) a work of artistic craftsmanship,
issues to the public copies of a graphic work representing, or of a photograph of, a
derogatory treatment of the work.
Appendix C: Interview Transcript

Test 3a Interviews: Participant 25
Time
Speaker
GB
P25
GB
Transcription
Have you got any questions?
Um only that one that I had about complexity
with words if youre looking at the complexity of
the image, how does that correlate?
Ive no idea.
Appendices
306
Code
Questions 2h
P25
GB
1:00
P25
GB
P25
GB
P25
GB
P25
GB
P25
3:00
GB
3:07
P25
GB
Fair enough.
Were you comfortable with the thing in general and
with the rating system?
Yeah, um. It was nice actually not having an upper
limit, although I found in sometimes I was limiting
myself and getting quite similar numbers, but
yeah, no I was comfortable going round and doing
it.
But you only got some numbers because those
Yep. The only thing would it be good to say
now? about seeing the numbers above the
paintings.
Yeah
Above each picture. Yeah, seeing the numbers
there, sometimes I would come to a picture and the
number that instinctively came into my head was
then the number I saw above the picture, so that
made me think am I being truthful to what I
actually think or have I been influenced by catching
sight of the numbers beforehand?
Mm, yeah. Let me ask about the criteria that you
used to rate things. What made one image more
complex than another?
Um I suppose it was the amount of information
that I was having to decipher from an image, er
Like, colour I found colour did definitely make a
difference as well. When youre looking at the
geometric the pattern ones, the sort of those
ones I definitely felt were more complex due to the
colours that were in them. If the colours were more
similar they seemed less complex. If the colours
contrasted more, they seemed more complex.
Although I was aware that they probably werent
both as complex really, but the colours seemed to
change it for me.
Well, there is no really
OK, yep, um
Thats why Im doing this test. There is no, er
really how complex they are at all, honestly.
I suppose another way that I rated the complexity
was how long how long I wanted to stand there
and really look at it to try and get information out
of it, um The ones the sketchy ones that you
can very quickly identify whats going on, it can be
pretty to look at and look at the the way that it
has been sketched, but the ones where you had to
really stand and, um, be taken in by the picture I
would probably have rated as a lot more complex.
Mm-hm. So how did you cope with the text, the
Appendices
307
Comfort 1
Comments 10
Complexity 9
Complexity 4
Complexity 2b
Complexity 8a
Complexity 9a
P25
GB
4:09
P25
GB
P25
GB
P25
GB
5:30
P25
GB
P25
GB
P25
GB
P25
ones that had words in it?

Um well, cause it was different as well cause
some of text was actually telling you bits and other
bits of the text was just sort of like story time, if
you know what I mean. It wasnt so much giving
off, er, so much information.
So did what the words say add complexity or is
it?
In some ways I sort of tried to ignore the words,
and look more like a picture, um what was being
shown.
Yeah, right. What about the images that you
preferred the most, did they have anything in
common, or perhaps did you were you aware
Um, the ones I preferred the most were abstract.
By far, they were definitely the ones that I preferred
the most. And probably the computer-generated
ones as well.
What kind of qualities did they have in common?
Um well, colour was definitely one thing that
drew me in, but I couldnt say definite type of
colour, but I do know the colours, er, influencing
my preference. And shape. The shapes that were in
them Colour and shape. But they I know that the
highest marks I gave were to very some of them
were very different in their colour and their shape,
but it was just what grabbed me and took me. Tried
to be as instinctive as possible.
I was going to ask about that. It was an instinctive
decision?
Yep. It was definitely instinctive. One thing that I
did think about when I was going round, though, is
because I had seen some of your artwork before,
therefore it was more familiar to me than some of
the other ones, and I did wonder of that would be
playing a part.
Oh, Im sure. But then again, this is similar to what
would happen in the real world and youre going
round a gallery, some of those are more familiar.
Yeah, I know, I did think about it and I did think,
well, still instinctively out of the ones Im familiar
with I still know which ones I prefer. So it didnt
worry me too much.
OK. What about quality, did you find that easy to
do?
Quality is
And how did you make those decisions?
How did I judge quality? Actually, I did find that
the hardest, cause I was very aware of it, how your
Appendices
308
Preference 5b
Preference 2
GB
P25
GB
P25
GB
P25
8:10
GB
P25
GB
P25
GB
P25
GB
P25
GB
P25
GB
personal preference of what you think is good

affects how you rate the quality.
Yeah. Im assuming that for most people the ease
of judging the quality of these things is related to
their general experience and familiarity with doing
this kind of thing, so you know, if youre in art, if
youre formally trained in art, its probably an easier
job to rate quality. Would you agree with that?
I would agree with that.
So, how much is art and design part of your day-today activities?
Um
Any formal training?
Er not, well, in certain aspects of it such as, er,
set design and looking at colour and Ive done
some art history at university, so but never to full
degree level. I would say I have a basic persons
knowledge of, well, probably slightly higher than
the average laymans knowledge of art, but
But you dont make stuff yourself?
Not really, no.
So how did you go about judging it then?
From my experience throughout life
What factors what aspects of the images did you
try and base it on?
Different images I suppose how how trained I
thought the artist would have to have been to have
produced that piece of work. Use of colours, I
suppose complexity in some ways did come into it
as well, if youre looking at it that way. Do you
need more?
We can wrap it up. Is there any other comments
you want to make?
Um, not
Was it easy or difficult overall, then? Enjoyable?
Interesting?
I found it very interesting, I really did. Thinking
having to think of Actually one thing I really
liked was that you had to round the three times and
look at them all individually. On I know that
some of my thoughts changed the third time I was
going round, so what Id done the first time, but
sort of thinking of oh no, would I have changed
that. I didnt go back and change anything, but I
was constantly thinking about how I look at these
pictures in three different ways.
OK. Thats it then. Thank you.
Appendices
309
Training 2a
Quality 4a
Quality 11
Comments 6
Appendix D: List of Codes

Questions about the Task
01. No
02. Yes
a.
b.
c.
d.
e.
f.
g.
h.
i.
j.
k.
What is the purpose of the test?

How many of the images are yours?
Are some images drawings of other images?
Where can I buy this image?
How are you going to level off the different ratings?
Are you an artist?
Is there supposed to be any pattern to how the pictures are laid out?
How does the complexity of words correlate with the complexity of the images?
Do you have a way of determining differences in preference between genders?
What would you say complexity is?
Where do the child-like images come from?
Comfort with the Procedure

01.
02.
03.
04.
Yes
No, it was difficult
Complexity was most difficult
Quality was most difficult
Complexity Criteria
01. Quantity of picture elements
02. Variety of picture elements
a. Number of varieties
b. Amount of variation
03. Level of detail
04. Number of colours
05. Amount of colour
06. Amount of time
a. In making the work
b. Looking at / understanding the work
07. Redundancy
08. Amount of work or difficulty
a. in making
b. in perceiving
09. Amount of information
a. Aesthetic
b. Semantic
10. Depth
11. Ambiguity
12. Size
Appendices
310
13. Flaws
Preference Criteria
01.
02.
03.
04.
05.
06.
07.
08.
09.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
Harmony of picture elements

Strong colours
Warm colours
Level of figuration
a. Representational
b. Abstract
Detail
Repetition / pattern
Quality of line
a. Geometric, straight
b. Swirly, curved
Clean and tidy
Would I have it at home?
Computer-generated
Complexity
Emotional response
Pretension
Images covering the entire surface
Relation to physical context
Relates to quality
Mundane
Schematic images
Textiles
Interest
Quality Criteria
01. Amount of work
a. In making
b. In perceiving and understanding
02. Amount of time to make the work
03. Skill
04. Difficulty in making the work
05. Suitability for exhibition
06. Coherence
07. Technical accomplishment
08. Creativity
09. Concept
10. Same as for preference
11. Complexity
Appendices
311
12. Varied depending on medium
Training in Art & Design

01. Formal training, actively engaged
02. Some formal training, not engaged
03. No formal training
Comments about the Task

01. None
02. Yes
a.
b.
c.
d.

Unusual rating procedure, being allowed to define own scale
Shouldnt be using the same room for interviews as for people doing the task
Different to an exhibition because of the viewing angle looking down at images
on tables.
e. Interesting experience
f. Question whether objects represented are comparable
g. Would be good to hang the images
h. Too many images
i. Rating scores may have been influenced by the sight of images index numbers
j. Art and design training helps make judgements of complexity
k. Images of artefacts are harder to rate than actual artefacts
Appendix E: Coding Results

The tables below show the complete set of results for the analysis of participant
interviews conducted in Test 3a (see pp.227249).
Questions (12)
No questions
What is the purpose of the test?
How many of the images are yours?
Are some images drawings of other images?
Where can I buy this image?
How are you going to level off the different ratings?
Are you an artist?
Is there a pattern to how the pictures are laid out?
How does the complexity of words correlate with the
complexity of the images?
Do you have a way of determining differences in preference
between genders?
What would you say complexity is?
Where do the child-like images come from?
Appendices
312
Code
01
02a
02b
02c
02d
02e
02f
02g
Count
18
5
4
2
1
1
1
1
%(37) Rank
48.65
1
13.51
2
10.81
3
5.41
4
2.70
5
2.70
5
2.70
5
2.70
5
02h
2.70
02i
2.70
02j
02k
1
1
2.70
2.70
5
5
Comfort with the procedure (4)

Comfortable
Uncomfortable
Quality was the most difficult
Complexity was the most difficult
Code
01
02
04
03
Count
24
7
6
2
%(39) Rank
61.54
1
17.95
2
15.38
4
5.12
4
Art training / experience (3)

No formal training
Formal training, actively engaged
Some formal training, not engaged
Code
03
01
02
Count
16
11
4
%(31) Rank
51.61
1
35.48
2
12.90
3
Criteria for judging Complexity (18)

Level of detail
Number of colours
Redundancy
Depth
Amount of colour
Amount of information: Aesthetic (visual)
Amount of time: Looking at / understanding the work
Size
Amount of information: Semantic
Ambiguity
Flaws
Code
03
04
07
08b
10
01
02a
08a
05
09a
02b
06a
06b
12
09
09b
11
13
Count
10
9
9
7
7
5
4
4
3
3
2
2
2
2
1
1
1
1
%(73) Rank
13.70
1
12.33
2
12.33
2
9.59
4
9.59
4
6.85
6
5.48
7
5.48
7
4.11
9
4.11
9
2.74
11
2.74
11
2.74
11
2.74
11
1.37
15
1.37
15
1.37
15
1.37
15
Criteria for judging Preference (25)

Strong colours (bold, saturated)
Quality of line: Swirly, curved
Level of figuration: Abstract
Level of figuration: Representational
Computer-generated
Complexity
Pretension (dislike)
Harmony of picture elements
Warm colours
Detail
Repetition / pattern
Quality of line: Geometric, straight
Clean and tidy
Code
02
03
08b
09
05b
05a
12
13
15
17
01
04
06
07
08a
10
Count
6
6
6
4
3
2
2
2
2
2
1
1
1
1
1
1
%(50) Rank
12.00
1
12.00
1
12.00
1
8.00
4
6.00
5
4.00
6
4.00
6
4.00
6
4.00
6
4.00
6
2.00
11
2.00
11
2.00
11
2.00
11
2.00
11
2.00
11
Appendices
313
Would I have it at home?

Emotional response
Images covering the entire surface
Relation to physical context (how the image worked in the
space)
Relates to quality
Mundane (dislike)
Schematic images
Textiles (dislike)
Interest
11
14
16
1
1
1
2.00
2.00
2.00
11
11
11
18
2.00
11
19
20
21
22
23
1
1
1
1
1
2.00
2.00
2.00
2.00
2.00
11
11
11
11
11
Criteria for judging Quality (13)

Skill
Amount of work: In perceiving and understanding
Amount of time to make the work
Coherence (meaningful arrangement of parts)
Creativity
Concept (the idea behind the work)
Complexity
Varied depending on medium
Code
03
01a
05
07
04
10
01b
02
06
08
09
11
12
Count
8
6
6
6
4
3
1
1
1
1
1
1
1
%(40) Rank
20.00
1
15.00
2
15.00
2
15.00
2
10.00
5
7.50
6
2.50
7
2.50
7
2.50
7
2.50
7
2.50
7
2.50
7
2.50
7
Comments (12)
No comments
Interesting experience
Shouldnt be using the same room for interviews as for
people doing the task
Too many images
Unusual rating procedure, being allowed to define own scale
Different to an exhibition because of the viewing angle
Question whether objects represented are comparable
Would be good to hang the images
Rating scores may have been influenced by index numbers
Art and design training helps make judgements of
complexity
Images of artefacts are harder to rate than actual artefacts
Code
01
02e
02a
Count
16
5
3
%(35) Rank
45.71
1
14.29
2
8.57
3
02c
5.71
02h
02b
02d
02f
02g
02i
2
1
1
1
1
1
5.71
2.86
2.86
2.86
2.86
2.86
4
6
6
6
6
6
02j
2.86
02k
2.86
Appendices
314

Aesthetic Complexity Practice and Percep

Uploaded by

Copyright:

Available Formats

Aesthetic Complexity Practice and Percep

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Aesthetic Complexity Practice and Percep

Uploaded by

Copyright:

Available Formats

Aesthetic Complexity:

Practice and Perception

A thesis submitted in partial fulfilment of the requirements of

To the following people I am particularly grateful:

Planned Contribution to Knowledge...................................................................... 47

Computer Art ...........................................................................................................119

Combined Results from Tests 3a and 3b .............................................................258

Aesthetics & Complexity

Aesthetics & Complexity

On the surface of a painting we perceive forms, which may be abstract or

Aesthetics & Complexity

Background: Art Practice and Research

Aesthetics & Complexity

evolutionary history. In fact, the needlepoint pattern is based on a cellular automaton

Figure 5 Elementary Cellular Automata #110 (2003), needlepoint (detail, work in

Aesthetics & Complexity

To support this line of enquiry, we need to develop a theoretical base by reviewing

Aesthetics & Complexity

Aesthetics & Complexity

Art and Design

Here, the concept of beauty is employed by Kant to distinguish judgements of taste

Aesthetics & Complexity

Nevertheless, it is plausible that aesthetic concepts are more fundamental, because if

Aesthetics & Complexity

Aesthetics & Complexity

Working Model of the Aesthetic

Aesthetics & Complexity

Figure 6 The working model of the aesthetic.

In this model, the aesthetic is identified as residing in the interactions with an

Aesthetics & Complexity

Aesthetics & Complexity

The same could be said of consciousness of sensory capabilities and mental

peer-reviewed publication system, choosing instead to self-publish funded by a successful

Figure 7 Visualization of the 1D CA rule If neighbouring cells have different states,

Figure 8 Evolution of elementary CA rule 150.

Figure 9 Examples of elementary cellular automata in Wolframs four classes of

Aesthetics & Complexity

Wolframs classification constitutes a qualitative analysis of complexity, in which the

Figure 10 Proportions of complexity classes in various types of CA. Wolfram, 2002,

Christopher Langtons Ph.D. thesis (1991) comprised a quantitative examination of

Aesthetics & Complexity

Figure 11 Location of the Wolfram classes in Langtons lambda space. (Langton,

Wolframs qualitative mapping is thus augmented by Langtons quantitative measurement,

Aesthetics & Complexity

Aesthetics & Complexity

Aesthetics & Complexity

Chaos and Randomness

Aesthetics & Complexity

Figure 13 The Lorenz attractor (top, front and side views).

Lorenz modelled the evolution of simplified weather systems by reducing the

Aesthetics & Complexity

Gell-Mann (2004) proposes a refinement of this crude measure, which he calls

Low information content

High information content

Figure 16 Effective complexity and algorithmic information complexity. Adapted

Technically, AIC is incomputable because we can never be certain that we have

Measuring Visual Complexity

gap in knowledge of visual complexity which has the potential to contribute to

Aesthetics & Complexity