Unsupervised Texture Segmentation Algorithm Based on Novel Scale Exponent Features

Abstract

Texture segmentation is a preliminary step in a wide spectrum of computer vision applications. Although the search for robust texture descriptors has been going for decades, there is still lack of texture features providing theoretical as well as practical evidence for successful segmentation. In this paper a novel algorithm for estimating scale exponents is described and applied in texture segmentation task. The estimated scale exponents are linearly dependent on generalized fractal dimensions. It has been proven that fractal dimensions are invariant under bi-Lipshitz transforms [1], which are general smooth transforms including perspective transforms. We estimate scale exponents in blocks around feature points, which allow us to characterize local regions and further segment them. In the case of finite resolution images, the proposed estimation algorithm produces robust to rotations and illumination changes features as predicted theoretically. The extracted features are applied for unsupervised segmentation using c-means fuzzy clustering by estimating spatial membership functions for each texture cluster. We experimented with textures from well-known Brodatz[2] and Vistex[3] texture database.