Selection of non-uniformly spaced orientations for Gabor filters using multiple kernel learning

The 2-D Gabor filters are useful for analyzing textured images. The tunable parameters for these filters are orientation, scale, and center frequency. An important issue in the texture analysis using Gabor filters is to select a set of filters such that the responses of selected filters contain most of the information in an image. In this paper, we present the multiple kernel learning (MKL) based approach to select Gabor filters with appropriate orientations, keeping the scale and center frequency fixed. A base kernel function is used for the response obtained from a filter with an orientation. A kernel obtained as a linear combination of base kernels, weighted according to the relevance of the orientations, is used. The weights determined using the MKL approach are used to select the relevant orientations. Effectiveness of the proposed approach is studied for an image categorization task.