International Journal of Signal Processing, Image Processing and Pattern Recognition

Vol.7, No.4 (2014), pp.167-176

http://dx.doi.org/10.14257/ijsip.2014.7.4.16

A Survey of Recent and Classical Image Registration Methods

Siddharth Saxena1 and Rajeev Kumar Singh2
1
2

(Research Scholar), Department of CSE &IT, MITS, Gwalior

(Assistant Professor), Department of CSE & ITMITS, Gwalior
1
Siddharthsaxena48@gmail.com, 2rajeev.mits1@gmail.com
Abstract

Image Registration is the process of aligning two or more images of the same scene taken
from different viewpoints, at different times, from different sensors. Image registration aligns
two images, i.e., base image and reference image, geometrically. There are different
approaches of image registration and these approaches are categorized according to their
nature that is areas based and feature based. These approaches are also categorized
according to the four simple steps of image registration procedure: feature detection, feature
matching, function mapping and transformation and re-sampling. Advantages and
disadvantages of different methods are discussed in the paper. The main aim of this paper is
to provide the knowledge of different image registrations methods used in different
application area.
Keywords: Image registration, feature detection, feature matching, mapping function,
sampling

1. Introduction
Image registration is the process of aligning two or more images of the same scene taken
from different views, at different times and from different sensors. It geometrically aligns two
images- base and the reference image. Image registration is a very important step in the
process of analyzing an image through which necessary information is found by combining
information from different data sources. Generally image registration required in remote
sensing (for monitoring environment, weather forecasting), in medical image processing
(combining CT, monitoring tumor growth) and in computer vision.
The comprehensive study of image registration techniques was published in 1992 by
Brown. The goal of this paper is to introduce techniques came after that and maps the current
development of image registration. We did not go into the details of particular algorithm and
not shows the results of comparative experiments. Instead of it we summarize the useful
approaches and point out interesting part of image registration.
In section II both the concepts and problems of image registration are discussed. In section
III both area based and feature based techniques of feature detection is discussed. In section
IV various feature matching algorithm is discussed. In section V different methods for
mapping
Function is discussed. In section VI different methods for image sampling and
transformations are discussed. Section VII shows the future trends in image registration.

2. Image Registration Methodology

Before proceeding towards various image registration methods we have to understand the
terminology for better understanding [3]:

ISSN: 2005-4254 IJSIP

Copyright 2014 SERSC

International Journal of Signal Processing, Image Processing and Pattern Recognition

Vol.7, No.4 (2014)

Target image: The image which doesnt change and is used as the basis of other images.
Source image: The image which is geometrically aligned with the target image.
Transformation and Warping: The mapping function which is used to modify the source
towards the target image.

2.1 Classification of Image Registration Techniques

In general image registration techniques can be categorized according to the various
distinguish points. Based on image acquisition, Brown in 1992, divided image registration in
four classes [1]. Barbara Zitova and Jan Flusser have categorized the image registration
techniques in the area based methods and feature based method. Maintz proposed a ninedimensional method that provides a categorization, which basically used in medical imaging
application.
2.1.1

Classification Based on Application

Image registration can be categorized into four different classes according to the image
application [1].
Different viewpoints: Same scenes of the image are captured from different viewpoints.
The purpose behind this to achieve a two-dimensional image representation. Example is
remote sensing- For the mosaicking of image.
Different times: Same scenes of the image are acquired at different time depending on
different conditions. The purpose behind this is to find and evaluate changes occur in the
scene during different time period. Example is remote sensing and computer vision.
Different sensors: Same scenes of the image can be acquired from different sources. The
purpose behind this is to analyze the information from different sources for obtaining the
more detailed information. Example is remote sensing and medical imaging.
Scene of model registration: Image of the scene and model are to be registered. The
purpose behind this is to localize the acquired image in the model and compare them.
Example is computer vision, remote sensing and medical imaging.
2.1.2

Classification Based on Nine Dimensional Scheme

Maintz proposed a nine-dimensional scheme to characterize an image registration which is

basically used for medical imaging [3].
Dimensionality: it refers to the number of geometric dimensions which any image space
has. It can be two dimensional or three dimensional.
Nature of registration: it is the aspect of two views used to effect registration. It may be
intensity based or surface based methods.
Nature of transformation: it refers to the property which holds various transformations
like affine, projective, and rigid.
Domain of transformation: it tells whether transformation is calculated local or global.
Degree of interaction: it refers to the control given by human over the registration
algorithm.
Optimization procedure: it refers to the standard approach in which the quality of
registration is estimated continually during registration.
Modalities involved: it refers to the means by which the images to be registered are
acquired.
Subject: it refers the sensed image.
Object: it refers to the region to be registered.

168

Copyright 2014 SERSC

International Journal of Signal Processing, Image Processing and Pattern Recognition

Vol.7, No.4 (2014)

2.1.3

Classification Based on Essentials

Barbara and Jan classified the image registration techniques as follows [1, 3]:

Area Based Methods: These methods are applied information about image is absent and
distinctive information is provided by gray level or colors.
Feature Based Methods: These methods are applied when local structure information
about an image is given.

2.2 Steps Involved in Image Registration Process

Due to diversity of image to be registered and various degradation in images it is
impossible to define a universal method for image registration. That why each image
registration method has its own importance. Majority of image registration method have four
basic steps [1]:
Feature Detection: salient and different objects in an image (edges, lines, contours) are
detected manually or automatically. These feature points are represented by their
descriptors.
Feature Matching: In this step, the features detected in the reference image and those
detected in sensed image has been matched. Feature descriptor and similarity measure are
used for this purpose.
Transform Model Estimation: the type and parameter of mapping function is established
which is used to align the sensed image with reference image.
Image Re-sampling and Transformation: The sensed image is transformed by establish
mapping functions.
The implementation of each registration step has its own typical problems. First we have to
decide what kind of feature is appropriate for the given task. The features should be
frequently spread over the image and easily detectable. The detected features in both the
sense image and reference image must have enough common elements. The detection method
should have good spatial accuracy and has not been affected by assumed degradation.
In the feature matching step, the problem caused by incorrect feature detection and image
degradation can arise. Feature can be dissimilar due to different imaging conditions. We have
to choose feature descriptor and similarity measure carefully so that it doesnt affect the
matching. They have to be discriminable enough.
The type of mapping function should be chosen according to the prior information about
image degradation. If no information is available then it should flexible enough to handle all
kind of degradation.
Finally, the choice of the appropriate type of re-sampling technique based on the accuracy
of the interpolation and computational complexity. Some cases use bilinear and nearest
neighbor approach.

3. Feature Detection
Formerly, the features are objects detected by manually by expert. There are two main
approaches to feature understanding [1, 3].
3.1 Area-Based Methods
Area-Based methods put emphasis on feature matching rather than detection. No features
are detected in this approach. So the first image registration step is omitted in area-based
methods.

Copyright 2014 SERSC

169

International Journal of Signal Processing, Image Processing and Pattern Recognition

Vol.7, No.4 (2014)

3.2 Feature-Based Methods

Feature-based approach is based on the extraction of salient features in the images.
Regions, lines, points are understood as features here. They should be spread over the image
and efficiently detectable in both images. The number of common elements of the detected
set of features should be high, regardless of image geometry, additive noise. Feature based
methods do not work on image intensity values as opposed to area-based method.
3.2.1

Region Feature Detection

The region features can be the closed boundary of appropriate size, reservoirs, forest,
urban areas. Regions are generally represented by their center of gravity. Region feature are
detected by means of segmentation methods. The accuracy of segmentation significantly
influences the resulting registration. S.K. Pal [4] proposed a method as a refinement to
segmentation to produce better registration accuracy. In 2004, Harris-Laplace region detector
[3] locates potential relevant points with the Harris corner detector and then selects the point
with a characteristic scale. A new region feature descriptor based image registration proposed
in 2012.
3.2.2

Line Feature Detection

The line features can be represented by line segments, object contours, roads or anatomical
structure in medical imaging. Standard edge detection methods like canny detector or
Laplacian detector are used for line feature detection. The Marr-Hildreth edge detector [6]
has been a very popular edge detector before canny presented his paper. Canny edge detector
[5] is widely considered to be the standard edge detection algorithm. A recent detailed
comparative study of various edge detection algorithms can be found in [7]. Apart from it Li
proposed a new method to detect the lines in the reference and source image. Figure 1 shows
the artificial image and its corresponding line segments and edges.

Figure 1. Artificial Image Contains Thick Line Segments Running Horizontally,

Vertically and Obliquely Across the Image, in Second Image Detected Line
Segments are Present and in Third, Fourth Edge Detection is Figured
3.2.3

Interest Point and Corner Feature Detection

Corners are estimated as points of high curvature on the region boundaries. The first corner
detector has been in the late 1970s. In 1977, Moravec [8] defined the concept of point of
interest as different regions in images. Based on Moravec concept Harris developed the
algorithm known as Harris corner detector. Figure 2 shows the image containing different
types of corners.
A comparative study of interest point performance on a data set can be found in 2011 and
different detection method can be found in 2009.

170

Copyright 2014 SERSC

International Journal of Signal Processing, Image Processing and Pattern Recognition

Vol.7, No.4 (2014)

Figure 2. Artificial Image Containing Different Corner Types and Real Image of
Block and House Scene
3.2.4

Shape Feature Detection

Some objects may be recognized by their outline shape and it is a very powerful feature in
image processing. Yang Mingqiang and KpalmaKidiyo in 2008 discussed the essential
properties of shape to be feature that includes rotation, translation, scaling, noise resistance
and reliability [9]. Figure 3 shows the shapes of different objects like wing and palm tree.

Figure 3. Wings are Longish and Triangular, Palm Trees Consist of a Lengthy
Stem and a Bushy Head

4. Feature Matching
After detecting the features, we have to match them. We can say that we have to determine
which feature come from corresponding locations in images that are different. Again we have
to discuss two different aspect of feature matching. One is area based and other is feature
based.
4.1 Area Based Methods
All techniques in area based methods merge the feature detection step with feature
matching step and deal with the images without detecting the salient feature of object.
4.1.1

Correlation or Pixel Based Methods

Cross correlation is the first basic approach of registration process. It is generally used for
pattern matching. The classical method of area based method is cross correlation [3].
Cross correlation is a type of similarity measure or match metric C (u, v) of image I(x, y)
with displacement u in X direction and v in Y direction. Two dimensional cross correlation
functions is shown below.
(
) (
)
(
)
( )
)
(
This similarity measure is computed for window pairs from the sensed and target images.
The window pair for which the maximum is found should be set as the corresponding one.

Copyright 2014 SERSC

171

International Journal of Signal Processing, Image Processing and Pattern Recognition

Vol.7, No.4 (2014)

Two main disadvantages of the correlation like methods has the high computational
complexity.
4.1.2

Fourier Methods

Correlation theorem states that the Fourier transform of the correlation of two images is
the product of the Fourier transform of one image and complex conjugate of Fourier
transform of other [10].
The Fourier transform of image f(x, y) is a complex function in which each function has a
) and an imaginary part (
) at each frequency (
) of
real part ((
frequency spectrum.
(
)
( )
(
)
(
)
Where (
)is a magnitude and (
) is a phase angle.
The phase shift correlation method is based on the Fourier shift which is proposed for the
registration of translated image.
Recently, image mosaic based on phase correlation and Harris operator is proposed by
yang in which first the scaling and translation is performed then the unregistered image is
adjusted. After that feature point are detected and matched.
It is observed that if the computational speed is required or the images are corrupted by
noise then Fourier methods are preferred than correlation methods.
4.1.3

Mutual Information Methods

Mutual information based registration process begins with the estimation of the joint
probability of the intensities of corresponding pixels in the two images. Mutual information
between two random variables X and Y is given by the formula [3].
(
)
( )
( )
( )
Where ( ) = -Ex log P(X) represents entropy of random variable and is probability
distribution of X. this method is based on maximizing MI.
It is observed that MI gives accurate result than any other registration method. But when
images have low resolution or it contains little information then it gives worse results.
4.2 Feature Based Methods
Feature based method used image feature derived by feature extraction algorithm instead
of intensity values for matching purpose.
4.2.1

Methods using Spatial Relations

Methods based on spatial relations are usually applied if detected feature are not clear or
their neighbors are distorted.
Barrow in 1977 has introduced the chamfer matching for image registration. Line feature
detected in image are matched by the minimization of the distance between them. Recently,
Gongjian wen developed high performance feature matching method for image registration
by combining spatial and similarity information [12].
4.2.2

Relaxation Methods

In the relaxation method, one of the famous method is consistent labeling problem (CLP)
in which we label each feature from the sensed image with the label of a feature from the
targeted image. So it is consistent with other images [12].
Another solution to CLP problem and to the image registration is backtracking, where
consistent labeling is generated recursively.

172

Copyright 2014 SERSC

International Journal of Signal Processing, Image Processing and Pattern Recognition

Vol.7, No.4 (2014)

4.2.3

Pyramids and Wavelets

In 1977, when a sub-window was used to find out the probable candidates of the
corresponding window in reference image and then full size window is applied.After that a
rectangular grid of windows is taken on which cross correlation is performed for reducing the
computational load. All these techniques are just an example of early pyramidal methods [3].
Recently, wavelet decomposition of the images was proposed for pyramidal approach.
There are many comparison tests have been carried out to establish which wavelet family has
the best performance.
4.2.4

Methods using Invariant Descriptor

Another method for exploiting the spatial relations is the correspondence of features can be
estimated by their descriptor. Descriptor contains information about feature points detected in
both the reference image and source image. The most common method is to use closed
boundary regions as features. Theoretically any invariant and discriminative enough shape
descriptor cab ne employed in region matching such as shape vectors. There are various types
of descriptors are used but we will proceed with the examination of the SIFT, SIFT variants
and SURF feature matching.

SIFT Descriptor and Its Variants

A different algorithm of feature extraction is registration based on Scale Invariant Feature

Transform (SIFT) proposed by David G. Lowe in 1999[11]. Lowe in 2004 proposed SIFT for
extracting distinctive invariant features from images that can be invariant to rotation and
scaling. It was widely applied in mosaicking recognition, retrieval etc., There are various
variants of SIFT that proposed for different types of images which provide better result. Some
of them are listed below with comparison to SIFT on the basis of time and computational
complexity.
Table 1. Comparison of Different SIFT Image Registration Methods
S.
NO.
1.
2.

EVOLUTI-ON
YEAR
2004
2004

TYPE OF
IMAGE
All types of images
Gradient Images

2005

Gradient Images

4.

NAME OF THE
ALGORITHM
Distinctive SIFT
PCA SIFT ( Principal
Component
Analysis)
GLOH ( Gradient
and orientation
histogram)
CSIFT (Color)

2006

Colored images

5.

ASIFT (Affine)

2009

6.

OCT SIFT (Octave)

2010

7.

ROBUST SIFT

2010

Object Images
having smooth
boundary
Synthetic Aperture
Radar
Remote sensing
images

3.

Copyright 2014 SERSC

COMPARIS-ON
WITH SIFT
Same as SIFT
Fast descriptor but
less than Distinctive
SIFT
More robust than
SIFT and PCS SIFT
Performs better
with respect to the
number of the
detected features
Reduce complexity

Reduce processing
time
More accuracy and
correct match rate

173

International Journal of Signal Processing, Image Processing and Pattern Recognition

Vol.7, No.4 (2014)

8.

BF SIFT ( Bilateral
Filter)

2012

Synthetic Aperture
Radar

Non iterative and

fast

9.

Multilevel SIFT

2012

Large size VHR

images

Solves memory
problem and reduce
computational
complexity

SURF Descriptor

SURF (Speeded UP Robust Feature) is a local robust feature detector and firstly presented
by Herbert BAY in 2006 that can be used in various computer vision tasks like object
recognition, feature matching. SURF is inspired by SIFT descriptor. The standard version of
SURF is several times faster than SIFT and it is also more robust against SIFT [13].

5. Transform Model Estimation

After the feature detected and matched properly then the mapping function is constructed.
In this step we choose the type of mapping function and its estimation. We can divide the
mapping function into two broad categories according to the amount of image they used i.e.,
Global model and local model [1]. Global model uses all control points into consideration for
estimating one set of mapping function which is valid for an entire image. On the other hand
local mapping function takes the image as a combination of patches and mapping function
depends on the location of their support in the image. When sufficient matching points is
given then scaling parameter, rotation angle and translation parameter can be retrieved by
least square method [2].

6. Image Resampling and Transformation

After feature extraction and feature matching step, we have to perform transform model
which we have estimated in transform model estimation step. It is required to calculate
mapping function. Models of mapping function are of two different types [14].
6.1 Global Models
This model uses all matching points to calculate one mapping function parameter which is
applied on the whole image, also known as rigid transformation. Global model use a
polynomial with two variables and it is called bivariate polynomial of low degree. This
transformation consists of rotation, scaling and translation.
6.2 Local Models
This model emphasis on dividing the entire image into different regions and mapping
function parameters are calculated for each part and it is known as elastic transformation.
Image resampling is performed after transformation model estimation. In this step, each
pixel in the image is either increased or decreased.
Practically image resampling is carried out by convoluting image with interpolating
function such as bilinear and bi-cubic functions, cubic B-splines and Gaussian and sin
functions.

7. Conclusion
Image registration is one of the important tasks when we want to integrate and analyze
information from different sources to obtain the more accurate information. This paper gives
a survey and review of the classical and recent registration methods. The main aim of this

174

Copyright 2014 SERSC

International Journal of Signal Processing, Image Processing and Pattern Recognition

Vol.7, No.4 (2014)

survey is to present the major advances in the field of image registration. Inage registration
involves a vast problem set which includes different problems faced in image processing like
image fusion, object detection and others. This paper makesgenuine efforts to cover all
possible techniques and work done in the image registration field.

References
[1] B. Zitova and J. Flusser, Image registration methods: A survey, Image and Vision Computing, vol. 21,
(2003), pp. 977-1000.
[2] M. Deshmukh and U. Bhosle, A survey of image registration, International Journal of Image Processing
(IJIP), vol. 5, Issue 3, (2011).
[3] G. L. Brown, Survey of Image Registration Techniques, ACM Computing Surveys, vol. 24, no. 4, (1992),
pp. 325-376.
[4] N. R. Pal and S. K. Pal, A review on image segmentation techniques, Pattern Recognition, vol. 26, (1993),
pp. 1277 1294.
[5] J. Canny, A computational approach to edge detection, IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 8, (1986), pp. 679-698.
[6] R. Maini and Dr. H. Aggarwal, Study and Comparison of Various Image Edge Detection Techniques,
International Journal of Image Processing (IJIP), vol. 3, Issue 1, (2010).
[7] H. Vishwakarma and S. K. Katiyar, Comparative study of edge Detection algorithm on the remote sensing
images using matlab, International Journal of Advances in Engineering Research (IJAER), vol. 2, Issue VI,
(2011) December.
[8] J. Liu, A. Akas, Al-Obaidi and A. Moravec, A comparative study of different corner detection methods,
Computational Intelligence in Robotics and Automation (CIRA), IEEE International Symposium, (2009).
[9] Y. Mingqiang, K. Kidiy and R. Joseph, A survey of shape feature extraction techniques, published in
Pattern Recognition, Peng-Yeng Yin (Ed.), (2008), pp. 43-90.
[10] R. N. B. Well, The Fourier Transform and Its Applications, McGraw-Hill, New York, (1965).
[11] D. G. Lowe, Distinctive image features from scale-invariant key points, International Journal of Computer
Vision, SIFT., vol. 60, no. 2, (2004), pp. 91110.
[12] K. Sharma and A. Goyal, Classification Based Survey of Image Registration Methods, IEEE, 4th ICCCNT,
(2013) July 4-6, Tiruchengode, India.
[13] H. Bay, A. Ess, T. Tuytelaars and G. L. Van, Speeded-up robust features (SURF), Computer Vision and
Image Understanding, (2008).
[14] J. Sachs, Image Resampling, (2010) Available: www.dlc.com/resampling.pdf.

Authors
Siddharth Saxena, he has received the B.E. degree in Computer
Science and Engineering, from ITM University, Gwalior, India in
2011. Currently he is pursuing M. Tech in Information Technology
from MITS, Gwalior and it will be completed in 2014. His
researchincludeImage restoration and Image Restoration

Prof. Rajeev Kumar Singh, received the B. Tech degree in

Information Technology from Mahatma Gandhi Chitrakoot
Gramodayavishwavidyalaya Chitrakoot, Satna India in 2009,
M.Techin Software Engineering from LNCT, Bhopal, India
in2011.
He
is
currently
assistantprofessor
in
theDepartmentofComputerScienceandEngineering,
M I T S Gwalior.
His teachingand researchinclude Image Processing and software
engineering.

Copyright 2014 SERSC

175

International Journal of Signal Processing, Image Processing and Pattern Recognition

Vol.7, No.4 (2014)

176

Copyright 2014 SERSC