The island model parallel GA and uncertainty reasoning in the correspondence problem

This work presents a pointwise approach to the correspondence problem in computer vision using contextual and structural features of a point. Multiple points are simultaneously considered, under the structural coherence constraint related to the similarity of the geometric features of emerging polygonal regions. Perceptron neural networks compute structural features. The correspondence is achieved by optimizing similarity measurements among the points features and by satisfying the structural coherence constraint. The island model parallel genetic algorithm (GA) searches a very large space by evolving several populations separately, using the Dempster-Shafer calculus for fitness evaluation. Changes were made in the parallel GA model in an attempt to introduce a higher level of diversity in the process. The occlusion problem is approached by iterating the whole process, alternating the reference image. Experimental results using a pair of real world indoor images demonstrate the usefulness of the approach for the correspondence problem. The reported simulations were conducted using 9 virtual machines. Comparisons made with previous work show a higher accuracy for the maximization process. A disparity map is constructed by considering the set of corresponding points as control points and by minimizing the differences in similarity among the image points