Statistical Models for Texture Classification and Segmentation

Abstract

Texture, being surface property of every object, plays important role in human visual system for object identification and recognition. Texture classification and segmentation are the important operations towards recognition. Simultaneous Autoregressive (SAR) models had been successfully used in texture classification and segmentation but it has difficulty in selecting the appropriate neighborhood and window size used to estimate the model parameters. The Rotation-Invariant (RI) and Multi-resolution (MR) SAR models were found to produce acceptable classification and segmentation results. The suitability of these two models is verified in the work with comprehensive and variety of texture banks namely stochastic, periodic and mixed. The size of textures used to train the classifier also affects the classification accuracy to great extent. The comparison between Euclidean and Mahalanobis classifier is also provided in the paper. MR-RISAR with two or more resolution levels and model order of two or more gives acceptable classification results. The MR-SAR model parameters are used here to segment a multi-textured image. Preprocessing and feature weighting improves the segmentation quality except at the texture boundaries. The %ERROR parameter, defined as ratio of number of miss-labeled pixels to total number of pixels, is used to quantify the segmentation quality.