Evolving descriptors for texture segmentation

Abstract

A new method for texture segmentation is proposed. In this supervised method, texture descriptors are based on grey level co-occurrences. A texture descriptor corresponds to an individual in a population, and a population corresponds to a given texture class. A genetic algorithm generates several distinct and efficient individuals adapted to discriminate proposed textures. Then, the individuals compete for territories in an image composed of samples derived from the learned textures. At the end of the simulated evolution, population territories match the texture regions of the image. The method provides excellent results on various examples of synthetic and natural textures.