Multiresolution Design of
Multiple Gabor Filters
for Texture Segmentation

Abstract

Gabor filters have been successfully applied to the segmentation of textured images, but the design of these filters remains an open issue. Two main approaches to the selection of Gabor filters are to employ a large bank of filters or to design a small set of customized filters. In the former approach, the filter parameters are predetermined ad hoc and not necessarily optimized for a specific task. This thesis addresses the latter approach and considers the design of single Gabor filters and the design of multiple Gabor filters at multiple resolutions to segment multiple textures.

A multichannel paradigm is developed as a basis for the design of the filters. The paradigm provides a unified framework that analytically relates the texture power spectra, Gabor-filter parameters, nonlinear processing, postfiltering effects, and segmentation error. Multidimensional statistical estimates of the filtered textures are generated from representative texture samples using a Rician statistical model and two different scales of the Gabor-filter envelope. Probability density estimates for each filtered texture are efficiently generated for an extensive set of candidate filter frequencies. Postfiltering and the accompanying effect on postfilter output statistics are also included in the design procedure. Further, the development yields new insight into the design of Gabor filters for texture segmentation.

Algorithms are developed both for filter design and for texture segmentation. The designed filters are integrated with a vector classifier and additional postprocessing to form an overall texture segmentation algorithm. Experimental results with multiple natural and synthetic textures confirm the efficacy of the filter-design methods and support the multichannel paradigm.