Gabor Filter Analysis for Texture Segmentation

R. Sandler, M. Lindenbaum
2006 Conference on Computer Vision and Pattern Recognition Workshop (CVPRW'06)  
Gabor features are a common choice for texture analysis. There are several popular sets of Gabor filters. These sets are usually designed based on representation considerations. We propose here an alternative criterion for designing the filters set. We consider a set of filters and their responses to a pairs of harmonic signals. Two signals are considered separable if the corresponding two sets of responses are disjoint in at least one of the responses. We look for the set of Gabor filters
more » ... izing the fraction of separable harmonic signals. The proposed semi-analytical algorithm calculates filters parameters for the optimal set, given the desired number of filters and the frequency range of possible signals. The resulting filters are significantly different from those traditionally used. We tested the proposed filters both in texture segmentation and texture recognition aspects with commonly used discrimination algorithms for each of the tasks. We show that, as expected, the resulting filters perform better than the traditional ones in discriminating synthetic and real textures. An important side effect of using the proposed filters with the popular features distribution based methods, considering a feature vector composed of the filters' responses, is the possibility to use a more compact (a lower number of feature vector prototypes) representation of the texture classes than using the common filters.
doi:10.1109/cvprw.2006.86 dblp:conf/cvpr/SandlerL06 fatcat:sapqhzfwvzeazkiyd36tiizwnm