Texture Descriptors via Stable Distributions

Abstract

In this paper, we present a texture descriptor which hinges in the use of the local image statistics so as to recover a compact representation of the texture under study. To this end, here, we make use of stable distributions and their link to Fourier analysis so as to provide a means to compute in a computationally efficient manner a local texture descriptor. This link between stochastic processes and Fourier analysis provides an efficient means to compute texture spectra which can be interpreted as a probability distribution for purposes of recognition and analysis. Making use of our local descriptor, we provide a metric between texture pairs that can be made devoid of rotations on the texture plane by recovering the optimal linear transformation via procrustes analysis. We demonstrate the utility of our descriptor and its associated metric on a database of real-world textures.