Directional statistical Gabor features for texture classification

Abstract

In texture classification, methods using multi-resolution directional (MRD) filters such as Gabor have not often shown significantly better performance than simple methods using local binary patterns, although they have a robust theoretical background and high computational complexity. We expect that this is because such methods usually make use of only the modulus parts of complex-valued MRD-filtered images and do not fully utilize their phase parts and other directional information. This letter presents a rotation-invariant feature using four types of directional statistics obtained from both the modulus and phase parts of Gabor-filtered images. First, modulus statistics, scale-shift cross-correlations, and orientation-shift cross-correlations are computed over all directions for each pixel of Gabor-filtered images, and global autocorrelations are computed over all pixels of each Gabor-filtered image. Global means and standard deviations for the three types of directional statistics and directional means and standard deviations for the global autocorrelations are then computed to form a feature vector. Experimental results with Brodatz, STex, CUReT, KTH-TIPS, UIUC, UMD, ALOT, and Kylberg databases show that the proposed method yields excellent performance compared with several conventional methods.