Rotation-invariant texture classification

Abstract

In this paper we propose a method for rotation-invariant 2D texture classification. Energy-normalized texture features are obtained by multiscale and multichannel decomposition using Gabor and Gaussian filters. Rotation invariance is achieved by the Fourier expansion of these features with respect to orientation. Unlike most previously reported methods, the textures are modeled with nonparametric feature distributions. In the experiments involving two standard datasets, with the classifier trained on samples of only one rotation and tested for all the others, high recognition rates were obtained.