On unsupervised segmentation of color texture images

The authors propose an orthogonal polynomial based color texture model for color texture segmentation. The proposed model is based on a complete set of difference operators which are easily configurable from a set of orthogonal polynomials. The operators are employed to represent a color image region as a linear combination of the operator's responses towards texture and noise. A simple statistical design of experiments paradigm is used for separating out the responses towards color texture from the responses towards noise. A local descriptor for color texture called cpronum is computed by encoding only the significant responses towards texture. Finally a global descriptor for color texture called cprospectrum is obtained by measuring the frequency of occurrences of the cpronums in the color image. The usage of the proposed color texture representation scheme is highlighted in unsupervised color texture segmentation.