Towards Fast Gabor Wavelet Feature Extraction for Texture Segmentation by Filter Approximation

Abstract

Gabor wavelet transform is one of the most effective feature extraction techniques for textures. As the Gabor wavelets are believed to be rather consistent to the response of Human Vision System (HVS), and many successful examples are reported in the areas of texture analysis. However, computational complexity of the feature extraction is still high even for computers nowadays, especially large sized image is involved. This paper attempts to break through the bottle-neck in the whole extraction process, that is to accelerate the convolutions by approximating the originally non-separable Gabor filter kernels to separable ones. Although the final computed features are not exactly the same as original ones, we prove that acceptable results can be achieved for segmentation purpose. While the acceleration ratio is as satisfactory as a gain of about $30\%$ in time in the worst case with a MATLAB implementation.