Texture image segmentation using Vonn mixtures-based hidden Markov tree model and relative phase

作     者：Niu, Pan-pan Wang, Li Shen, Xin Wang, Qian Wang, Xiang-yang 

作者机构：Liaoning Normal Univ Sch Comp & Informat Technol Dalian 116029 Peoples R China 

出 版 物：《MULTIMEDIA TOOLS AND APPLICATIONS》 (多媒体工具和应用)

年 卷 期：2020年第79卷第39-40期

页      面：29799-29824页

核心收录：

学科分类：0808[工学-电气工程] 08[工学] 0835[工学-软件工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National Natural Science Foundation of China [61472171, 61701212] Key Scientific Research Project of Liaoning Provincial Education Department [LZ2019001] Natural Science Foundation of Liaoning Province [2019-ZD-0468] 

主　　题：Texture image segmentation Undecimated dual tree complex wavelet transform Relative phase Vonn probability density function Hidden Markov tree 

摘      要：Texture segmentation is a frequently occurring and challenging problem in many computer vision and pattern recognition applications. The importance of phase information for texture analysis has been earlier established for many image processing. Undecimated dual tree complex wavelet transform (UDTCWT) is a new image decomposition. It not only provides exact translational invariance and rich directional selectivity, but also offers perfect consistent relative phase relationships across scales. In this paper, we propose a novel texture image segmentation framework using Vonn mixtures-based hidden Markov trees (HMT) and UDTCWT domain relative phase. Firstly, we analyze the robustness and marginal distribution of UDTCWT relative phases, and various strong dependencies between UDTCWT relative phases. Then, we propose a new HMT statistical model in UDTCWT domain, namely Vonn mixtures-based HMT, by describing the UDTCWT relative phases statistical distribution with Vonn mixtures (VM), which can capture both the subband marginal distributions and the strong dependencies across scales of the UDTCWT relative phases. Finally, we develop a texture image segmentation framework using the Vonn mixtures-based HMT model of UDTCWT domain relative phases, in which expectation-maximization (EM) parameter estimation, Bayesian multiscale raw segmentation, and context based multiscale fusion are used. Comparing to the state-of-the-art techniques, the proposed method can not only produce high-quality segmentation results in a more efficient way, but also keep a lot of boundary details in the segmentation results.