TEXTURE CLASSIFICATION BY CENTER-SYMMETRICAL AUTOCORRELATION, USING KULLBACK DISCRIMINATION OF DISTRIBUTIONS

We propose a new method of texture analysis and classification based on a local center-symmetric covariance analysis, using Kullback (log-likelihood) discrimination of sample and prototype distributions. Features of our analysis are generalized, invariant, local measures of texture having center-symmetric patterns, which is characteristic of many natural and artificial textures. We introduce two local center-symmetric auto-correlations, with linear and rank-order versions (SAG and SRAC), together with a related covariance measure (SCOV) and variance ratio (SVR). All of these are rotation-invariant, and three are locally grey-scale invariant, robust measures. In classification experiments, we compare their discriminant information to that of Laws' well-known convolutions, which have specific center-symmetric masks. We find that our new covariance measures, which can be regarded as generalizations of Laws' measures, perform better than Laws' approach despite their measure of texture pattern and grey-seal.