Texture image segmentation using complex wavelet transform and Hidden Markov models

Abstract

Hidden Markov tree (HMT) is a tree-structure statistical model, which is used to capture the statistical structure information of smooth and singular regions. It works by modeling the relationship between the wavelet coefficients interscales. For the discrete wavelet transform (DWT) has its own drawbacks inherently, such as shift variance, lack of directionality, etc. The traditional HMT model based on DWT often leads to an unideal segmentation result. Because of the near shift-variance and good directional-selectivity of complex wavelet transforms, here the authors proposed a complex wavelet domain HMT model (C-HMT) to improve the accuracy of multiscale classification results. To get an accurate final segmentation, labeling tree model was used to fuse the interscale classification results. In the experiment, the classification and segmentation results of the proposed method are found to be better than the traditional wavelet-based models.