Optimal Texture Filter Design using Feature Extraction Modeling

Trygve Randen and John Håkon Husøy
Høgskolen i Stavanger
P.O. Box 2557 Ullandhaug, N-4004 Stavanger, Norway

To appear in Working Papers from Høgskolen i Stavanger, 1997
Also submitted to IEEE Transactions on Image Processing

Full paper


This paper presents a new method for modeling the means and variances of texture features extracted by filtering and local energy computation. Having a model for the means and variances of the extracted features, we express feature goodness criteria, and develop these into optimal filter design approaches for both two- and multiple texture problems. Earlier the ratio of the feature means has been used as a criterion for two-texture filter design. Here we examine this criterion and develop closed form design techniques for another feature mean criterion. We furthermore develop new closed form design approaches for the Fisher criterion and a generalized criterion on both the feature means and variances, and the assumptions necessary for the solutions to be valid. For the Fisher criterion we also develop and test the necessary model for gradient search optimization. Finally, we develop a scheme for designing an optimal bank of filters for discriminating texture mosaics with more than two textures.