1 INTRODUCTION
Image registration is a crucial step for integration
information from two or more images of the same target
area, these images are taken from different sensors
(multi-modal analysis), by the same sensor at different times
(multi-temporal analysis) or different viewpoints
(multi-view). It is one of the key technologies of image
analysis and understanding. As an important tool, image
registration has been widely used in the fields of biomedical
research [1, 2, 3], satellite imaging [4, 5, 6] and remote
sensing [7, 8].
There are two main categories in image registration
algorithm: gray level information based and feature based
approaches. Much work has recently focused on gray level
based method, in which the gray level information is used to
calculate similarity between two images. It doesn’t need
complicated preprocessing, such as feature extraction and
image segmentation. In gray level based image registration,
there are generally three important components [9].
z Searching space
The kernel of image registration is to find a best spatial
transformation which maps any point of the floating
image to the corresponding point of the reference image.
Thus the searching space is a set of potential
transformation parameters, including translation,
rotation, scaling parameters.
z Similarity metric
The similarity metric is a criterion to estimate how
closely the features and gray levels of the reference
image and the floating image match. The sum of squared
intensity difference [10], generalized correlation
coefficient [11], and information theoretic measures [3,
This work was supported in part by the National Nature Science
Foundation of China (No. 61271232, 61571236, 61502245), and Jiangsu
Province Postdoctoral Science Foundation (No. 1402018A).
4] are widely used similarity measures.
z Searching strategy
The searching strategy is used to find a class of best
parameters of spatial transformation to optimize the
similarity metric. Local and global method can be used.
For local method, such as steepest descent gradient,
Powell’s direction set, conjugate gradient. Examples for
global optimization method are particle swarm
optimization (PSO) [3, 9, 12], genetic algorithm (GA)
[13], and differential evolution algorithm (DE) [4, 14].
Mutual information, as an effective similarity metric, is used
in this paper. Searching strategy is used to find the best
transformation to maximize the mutual information.
Differential evolution, proposed by Storn and Price in 1997
[15], has been proven great exploration ability and can
achieve much reliable results in many fields. Therefore, we
apply DE algorithm to find the optimal combination of the
parameter values involved in the affine transformation.
However, as an Evolutionary Algorithm, differential
evolution still has the probability to be trapped into local
optima. This paper we propose an improved differential
evolution algorithm with replacement strategy (DERS) to
enhance its searching ability for image registration problem.
The following paper is organized as follows:
Section 2 presented the traditional DE and our proposed
DERS algorithm. Section 3 describes the image registration
problem and defines the affine transformation and the
mutual information. Experiments are stated and discussed in
Section 4 and Section 5 concludes the paper.
2 DIFFERENTIAL EVOLUTION WITH
REPLACEMENT STRATEGY
2.1 Differential Evolution
Differential evolution is a population based stochastic
meta-heuristic algorithm, especially suit for real parameter
-
temporal Image Registration Utilizing a Differential Evolution Algorithm
with Replacement Strategy
Feiyi Xu
1
, Haidong Hu
2
, Hao Gao
1*
, Member IEEE, Baoyun Wang
1
1. The Institute of Advanced Technology, Nanjing University
of Posts and Telecommunications, Nanjing 210023
E-mail: tsgaohao@gmail.com
2. Beijing Institute of Control Engineering, Beijing, 100190
Abstract: Image registration is effect by finding the best spatial transformation between the reference image and the
floating image in terms of optimizing a similarity metric. Mutual information, as an effective method, is a reliable
criterion for image registration. In this paper, we propose an improved differential evolution algorithm with replacement
strategy to maximize the mutual information for multi-temporal image registration. Experimental results show that our
new algorithm achieves more precise and robust results when compared with several other classical optimization
methods.
Key Words: Image registration, differential evolution, spatial transformation, mutual information
752
978-1-4673-9714-8/16/$31.00
c
2016 IEEE