Rotation Invariant Co-occurrence Matrix Features

Grey level co-occurrence matrix (GLCM) has been one of the most used texture descriptor. GLCMs continue to be very common and extended in various directions, in order to find the best displacement for co-occurrence extraction and a way to describe this co-occurrence that takes into account variation in orientation. In this paper we present a method to improve accuracy for image classification. Rotation dependent features have been combined using various approaches in order to obtain rotation invariant ones. Then we evaluated different ways for co-occurrence extraction using displacements that try to simulate as much as possible the shape of a real circle. We tested our method on six different datasets of images. Experimental results show that our approach for features combination is more robust against rotation than the standard co-occurrence matrix features outperforming also the state-of-the-art. Moreover the proposed procedure for co-occurrence extraction performs better than the previous approaches present in literature, able to give a good approximation of real circles for different distance values.