Image Registration Based On Multi-Objective Evolutionary Algorithm

Image registration is a vital task in image processing, it is the process of geometrically aligning two or more images, which are taken at different times, from different viewpoints or by different sensors of the same scene. Image registration has been widely used in image change detection, image mosaicking, medical image processing and pattern recognition.Firstly, we introduce the theoretical knowledge about image registration, include the definition of image registration, the basic process of image registration and image resampling methods. Next is the detailed introduction of rigid body transformation, affine transformation, nonlinear transformation and project transformation. The basic framework is vital for the understanding of image registration, so we analyze it in four aspects: feature space, search space, similarity metrics and optimization methods. Usual y, there are three types of image registration methods: feature-based methods, area-based methods and transform domain-based methods. Feature-based methods only use the points or edge features extracted from image, so they are fast and have high accuracy. At the same time, we show out a brief comparison for the classic Harris corner detector, SFIT detector, and SUSAN corner detector. The area-based methods use the statistics information of gray value as the similarity, use certain optimization methods to find the best transform parameters. The area-based methods are widely used because they are robust to noises and make no assumptions for the image model.In this dissertation, we combine area-based methods and feature-based methods, and use the multi-objective evolution algorithm based on decomposition to optimize the spatial information and mutual information, phase congruency is used for feature extraction because it is invariant to the change of il umination. Finally we select some points on the Pareto front for the image registration. In the fourth part of the dissertation, we construct three objectives, namely angle criterion, distance criterion, and point matching criterion. We combine the phase congruency model and multi-objective genetic algorithm for image registration, a rank is given to every solution when it is evaluated, and genetic methods like crossover and mutation are executed based on their rank. The experimental results show that the method is feasible and effective.