Model-based region growing segmentation of textured images

An approach to the use of a region-growing technique for segmentation of textured images is presented. The algorithm is model-based, with each mixture region in the image modeled by a noncausal Gaussian Markov random field (GMRF). No a priori knowledge about the different texture regions, their associated texture parameters, or the available number of texture regions is required. The algorithm first partitions the image into small disjointed square windows. The texture within each window is modeled by a noncausal GMRF. Most of the windows are homogeneous. A hierarchical merge-split region-growing process is then employed to reconstruct most of the homogeneous regions that are presented in the image. The growth of various homogeneous regions is directed by a texture distance defined by a likelihood ratio test statistic based on the underlying GMRF model assumptions. The algorithm was tested on real textured images and proved to be robust and effective