Texture images have abnormal, microscopic characteristics, but some parts of the image maintain statistical regularity from a macroscopic point of view. In order to capture these characteristics that improve image segmentation results, a new wavelet-based, multi-scale Bayesian texture image segmentation method, based on EHMM-HMT (enhanced hidden Markov model-hidden Markov tree) and MSWHMT (multi-states weighted hidden Markov tree) modes, is proposed. The image blocks’ relative interactions are described through the EHMM model effectively, and the homogenous raw segmentation, propitious to final fusion, is obtained on the coarsest scale. Subsequently, in order to reduce mislabeling the boundaries of raw segmentation and to decrease the computing complexity of the model, the MSWHMT model is proposed with better raw segmentations of high accurate boundary detection put on finer scales. Finally, a pixel level segmentation is reached through a multi-scale Bayesian fusing strategy that combines with the boundaries. The method is compared to HMTseg, HMT (boundar based+MAP), and EHMM-HMT (MAP) algorithm through several micro-texture images to demonstrate its competitive performance. It has also been found to improve the accuracy of macro-texture image segmentations.