Unsupervised texture segmentation/classification using 2-D autoregressive modeling and the stochastic expectation-maximization algorithm

The problem of textured image segmentation upon an unsupervised scheme is addressed. In the past two decades, there has been much interest in segmenting images involving complex random or structural texture patterns. However, most unsupervised segmentation techniques generally suffer from the lack of information about the correct number of texture classes. Therefore, this number is often assumed known or given a priori. On the basis of the stochastic expectation-maximization (SEM) algorithm, we try to perform a reliable segmentation without such prior information, starting from an upper bound of the number of texture classes. At a low resolution level, the image model assumes an autoregressive (AR) structure for the class-conditional random field. The SEM procedure is then applied to the set of AR features, yielding an estimate of the true number of texture classes, as well as estimates of the class-conditional AR parameters, and a coarse pre-segmentation. In a final stage, a regularization process is introduced for region formation by the way of a simple pairwise interaction model, and a finer segmentation is obtained through the maximization of posterior marginals. Some experimental results obtained by applying this method to synthetic textured and remote sensing images are presented. We also provide a comparison of our approach with some previously published methods using the same textured image database.