Designing multiple Gabor filters for multitexture image segmentation

Abstract

We consider the problem of segmenting multitextured images using multiple Gabor lters. In particular, we present a mathematical framework for a multichannel texture-segmentation system consisting of a parallel bank of lter channels, a vector classier stage, and a postprocessing stage. The framework establishes mathematical relationships between the predicted texture-segmentation error, the frequency spectra of constituent textures, and the parameters of the lter channels. The framework also permits the systematic formulation of lter-design procedures and provides predicted vector output statistics that are useful for classier design. This paper focuses on the mathematical framework and provides experimental results that conrm the utility of the framework in the design of a complete image-segmentation system. The results demonstrate eective segmentation using a straightforward classier and fewer than half the number of lters needed in previously proposed approaches. Subject terms: Gabor prelter, Gabor lter, Gabor function, texture segmentation, statistical image analysis, texture analysis, computer vision, image segmentation.