Optimization of Spatial Filters for Image Texture Feature Separation by Simulated Annealing

Frode Nergård and Trygve Randen
Høgskolen i Stavanger
P.O. Box 2557 Ullandhaug, N-4004 Stavanger, Norway

In Proc. NORSIG, Helsinki, Finland, September 1996, pp. 251-254

Full paper

Abstract

In a work by Randen and Husøy, an approach to the design of filters for texture feature extraction for two textures has been developed. The feature extraction process were modelled as a filter with a subsequent local energy function. The filter was optimized with respect to a criterion of obtaining high distance between the modelled feature means and low modelled feature variances. However, the optimization was based on some simplifying assumptions. Using the same criterion, without making these simplifications, we propose three methods for finding the optimal filter. All methods have their basis in simulated annealing, and one of them, called annealing evolution, are also based on genetic algorithms. Experiments indicate that the suggested methods were able to find better filters than the methods proposed by Randen and Hus{\o}y.