Evolutionary Segmentation of Texture Image Using Perturbation Optimization

Abstract

Texture segmentation methods using split and merge, pyramid node linking, or active contour model, et al. has been proposed. However, they need to tune parameters for the optimum segmentation of every image querying. It is difficult to fix a parameter set for all the query images. To solve this problem, a powerful method using simple genetic algorithm (SGA) is proposed. Segmentation results do not vary according to a tuned parameter set for the evolutionary algorithm. But, since it takes the global and parallel optimization approach on an image plane, it cannot search the optimum solution efficiently. In this paper, for efficient optimization, we propose a new evolutionary segmentation method of texture image using perturbation optimization. After dividing the original image into small rectangular regions, we extract feature vectors from small regions. We search the candidate of cluster numbers for each small region by perturbating the cluster number into another one and by inspecting the change of the fitness for each perturbed number. The combination of cluster numbers is optimized locally and successively on an image plane by the perturbation optimization. This brings us the enhancement of segmentation accuracy and the curtailment of processing time compared with the method using SGA.