Improving the performance of Evolutionary algorithms in imaging optimization

This research applies Evolutionary algorithms (EAs) to the task of finding a proper mapping between geometrically distorted images, which arises in applications like image registration and object recognition. The task is formulated as an imaging optimization problem of minimizing the difference between the images. The objective of the research is to improve the computational performance of EAs in imaging optimization, by developing a fairly general approach based on a broader hybridization of EAs with the following techniques. (1) Augmenting EAs with local optimization technique, in the form of a two-phase (random/direct) cyclic search procedure reducing the excessive computational cost of local search. (2) Utilizing a frontal algorithm of forming new population assuring its diversity, and restoring the fair and effective usage of the search space disrupted by evolutionary operators. (3) Utilizing Image local response which directly extracts the main shape features; reduces the computational cost of fitness evaluation; and provides an efficient image model for adaptive local search, reduction of parameter space, and multi-sensor image fusion. (4) Introducing an advanced image model which reduces the amount of processed information, and includes the following steps: computing Image response, building response histogram, decomposing image into sections, decomposing sections into quadtrees, and defining the main shape feature, a hull. (5) Representing the sought image mapping as a piece-wise affine transformation allowing for significant mutual distortion of the compared images, so that different image sections have their respective local affine transformations. (6) Organizing image sections in a tree-structure which is processed in a hierarchical top-to-bottom manner, with local transformations of parental sections serving as initial approximations for local transformations of the offspring. (7) Utilizing multiple aligned populations aimed at increasing the coherence and robustness of the mapping during simultaneous processing of multiple views. (8) Utilizing multi-objective optimization, where different fitness functions are processed at the different computational stages, thus increasing the confidence of the search. (9) Implementing optimization search as two consecutive passes. During the first pass, global optimization seeks for a proper mapping of the image hull. During the second pass, the hull transformation is used as an initial solution for the final piece-wise optimization.