Algorithm for designing multiple gabor filters for segmenting multi-textured images

We present an algorithm for the design of multiple Gabor filters for the segmentation of multi-textured images. We draw upon earlier results that provide a segmentation error measure based on the predicted vector output statistics of multiple filter channels. This segmentation error measure is used to design the filter channels for a particular segmentation task. In our approach, the filter parameters are free to vary from channel to channel and are not restricted to some predetermined decomposition of the frequency plane. Thus, our method can generate more effective filter designs and result in more effective features for image segmentation than prior methods. Finally, we present texture segmentation results that confirm the efficacy of the proposed procedure. These results show effective segmentation of 8 textures using as few as 2 filters, whereas earlier approaches required 13 to 40 filters to segment 5 textures.