Loading...
Publication
Affine image registration using genetic algorithms and evolutionary strategies
Authors
Advisor(s)
Abstract(s)
This thesis investigates the application of evolutionary algorithms to align two or
more 2-D images by means of image registration. The proposed search strategy is a
transformation parameters-based approach involving the affine transform. A noisy objective
function is proposed and tested using two well-known evolutionary algorithms
(EAs), the genetic algorithm (GA) as well as the evolutionary strategies (ES) that are
suitable for this particular ill-posed problem. In contrast with GA, which was originally
designed to work on binary representation, ES was originally developed to work in continuous
search spaces. Surprisingly, results of the proposed real coded genetic algorithm are
far superior when compared to results obtained from evolutionary strategies’ framework
for the problem at hand. The real coded GA uses Simulated Binary Crossover (SBX), a
parent-centric recombination operator that has shown to deliver a good performance in
many optimization problems in the continuous domain. In addition, a new technique for
matching points, between a warped and static images by using a randomized ordering
when visiting the points during the matching procedure, is proposed. This new technique
makes the evaluation of the objective function somewhat noisy, but GAs and other
population-based search algorithms have been shown to cope well with noisy fitness evaluations.
The results obtained from GA formulation are competitive to those obtained
by the state-of-the-art classical methods in image registration, confirming the usefulness
of the proposed noisy objective function and the suitability of SBX as a recombination
operator for this type of problem.
Description
Keywords
Evolutionary algorithm (EA) Image registration (IR) Affine transform Point-pattern matching Genetic algorithm (GA) Evolutionary strategies (ES) Simulated binary crossover (SBX)