Discrete MRF model parameters as features for texture classification

Texture classification systems are characterized, existing techniques for texture classification are reviewed, and a method for extracting textural features for classification is proposed. A second-order, four-color, auto-binomial Markov random field (MRF) with four parameters is fitted to given textured images, and the estimated parameters are used as features for classification. The MRF-based features are compared experimentally to the features derived from spatial gray-level dependence matrices (SGLDM) for synthetic and natural textures. The MRF-based features are generative, but the SGLDM features are only descriptive. MRF-based features can be extracted in one step, which means that the four features can be extracted by simply fitting the model to a texture pattern. Experiments show that the MRF-based features outperform the SGLDM-based features