Bayesian texture segmentation based on wavelet domain hidden markov tree and the SMAP rule

According to the sequential maximum a posteriori probability (SMAP) rule, this paper proposes a novel multi-scale Bayesian texture segmentation algorithm based on the wavelet domain Hidden Markov Tree (HMT) model. In the proposed scheme, interscale label transition probability is directly defined and resoled by an EM algorithm. In order to smooth out the variations in the homogeneous regions, intrascale context information is considered. A Gaussian mixture model (GMM) in the redundant wavelet domain is also exploited to formulate the pixel-level statistical features of texture pattern so as to avoid the influence of the variance of pixel brightness. The performance of the proposed method is compared with the state-of-the-art HMTSeg method and evaluated by the experiment results.