A novel algorithm for Bayesian document segmentation is proposed based on the wavelet domain hidden Markov tree (HMT) model. Once the parameters of model are known, according to the sequential maximum a posterior probability (SMAP) rule, firstly, the likelihood probability of HMT model for each pattern is computed from fine to coarse procedure. Then, the interscale state transition probability is solved using Expectation Maximum (EM) algorithm based on hybrid-quadtree and multiscale context information is fused from coarse to fine procedure. In order to get pixel-level segmentation, the redundant wavelet domain Gaussian mixture model (GMM) is employed to formulate pixel-level statistical property. The experiment results show that the proposed scheme is feasible and robust. A novel algorithm for Bayesian document segmentation is proposed based on the wavelet domain hidden Markov tree (HMT) model. Once the parameters of model are known, according to the sequential maximum a posterior probability (SMAP) rule, first, the likelihood probability of HMT Then, the interscale state transition probability is solved using Expectation Maximum (EM) algorithm based on hybrid-quadtree and multiscale context information is fused from coarse to fine procedure. In order to get pixel -level segmentation, the redundant wavelet domain Gaussian mixture model (GMM) is employed to formulate pixel-level statistical property. The experiment results show that the proposed scheme is feasible and robust.