Feature selection for texture analysis using genetic algorithms

Abstract

Texture analysis is a key issue in the artificial processing of the visual information. Most methods compute a set of numerical values, which characterize the textures and express, in some cases, physical properties. These features, assembled in a feature vector, are eventually used in texture recognition and image segmentation. The computed characteristics are inherently of different significance in terms of distinguishing textures from one another. Having as the final goal the optimal performance of the classification process (i.e., minimum classification error rate), one must select the most significant features for a number of given texture classes. The exact relationship between the error rate of the classification process and the feature significance cannot be expressed in a mathematical form. Due to their robustness, even in the case of multimodal, discontinuous and non-differentiable functions, genetic algorithms have been considered as a suitable instrument in this optimization problem. Implementation of such algorithms is simple and applicable for all features and classifier systems. Computer experiments on texture samples from the real world show that the proposed framework is effective, giving better results than in the case of equal significant textural features.