Contextual classification and segmentation of textured images

An algorithm which combines the merits of statistical classification- and estimation-theory-based approaches is proposed for textured image segmentation. The texture regions are modeled by noncausal Gaussian Markov random fields (GMRF). The algorithm is comprised of two stages. In the first stage, the image is partitioned into small blocks of pixels. GMRF parameter estimates are extracted as the feature vector for each block. A maximum likelihood spatial classifier, which explores the class conditional correlation properties among neighboring feature vectors, is proposed for classifying each block into one of m-possible texture classes. The result of classification is a coarse segmented image. The locations of the edges are estimated in the second stage using a line-by-line maximum likelihood edge-estimation technique. Each detected edge sequence is further modeled as an autoregressive (AR) process and processed by a Kalman filter in order to smooth the detected boundary