Texture classification framework using gabor filters and local binary patterns

Abstract

In this paper, a novel method towards rotation, scale and illumination invariant texture image classification is introduced. We exploit the useful rotation and scale characteristics of Gabor filters and illumination invariance characteristics of LBP, proposing image features which are invariant to the above mentioned imaging dynamics. The images are first filtered using Gabor filters followed by a summation of filter responses across scales. An LBP of the resulting features is calculated followed by an integral histogram of LBPs across various orientations of the Gabor filters. An experimental validation of the invariance of the descriptor is shown on a texture classification problem using two publicly available datasets: the USC-SIPI, and CUReT texture datasets. Our experiments show that the proposed descriptor outperforms the other methods that have been considered in this paper.