在本論文中，我們提出了關聯隱藏馬可夫樹模式 (contextual hidden Markov tree model, CHMT) 和邊界精細化 (boundary refinement) 方法來 做紋理影像分割。關聯隱藏馬可夫樹模式是由建立在小波轉換架構下的 隱藏馬可夫樹模式 (hidden Markov tree model, HMT) 改良而來的。隱藏 馬可夫樹模式是用來捕捉小波係數之統計特性的一種樹狀結構機率模 式； 隱藏馬可夫樹可以完整的描述小波係數的繼承性 (persistence property)，但不太具有聚集性 (clustering property)。而關聯隱藏馬可夫樹 模式則使用擴增點 (extended nodes) 的觀念，來加強隱藏馬可夫樹模式的 聚集性。 在影像分割的應用上，因為邊界精細化方法加入影像像素的位置資 訊，區分為同質區域及邊界區域，因此，我們使用邊界精細化方法來加 強粗分割的正確性。首先，對於每一種紋理影像，利用關聯隱藏馬可夫 樹模式訓練一組代表此紋理影像的關係參數；接著利用這些參數算出不 同解析度區塊的最大相似度函數值做第一次分割；但分割結果，解析度 愈高的影像正確率愈低。接著依照影像的區域性融合不同解析度的分割 結果以得到更精確的分割結果。 A multiscale texture image segmentation approach based on the contextual hidden Markov tree (CHMT) model and boundary refinement is proposed. The hidden Markov tree models (HMT) is a statistical model of tree structure for capturing properties of wavelet coefficients. The HMT model describes persistence property of wavelet coefficients, but loses clustering property. We have proposed the CHMT model which improved from the HMT model by enhancing the clustering property. The CHMT model reinforces clustering property by using extended coefficients without changing the wavelet tree structure; thus the HMT training scheme can be easily modified to estimate the parameters of the CHMT model. In this study, the CHMT model is applied for texture segmentation. For each texture, we use the CHMT model to train a set of parameters and then utilize these parameters compute likelihood functions for all mulitscale squares of a test image. At last, we segment the test image with the principle of maximum likelihood. Only based on the CHMT model, the segmentation results are not good enough when the size of dyadic square is small; thus the boundary refinement algorithm is adopted to fuse the multiscale square to get better-quality segmented results. The segmented results based on the HMT and CHMT models are compared to show the improvement of the CHMT model over the HMT model; moreover, the boundary refinement algorithm is also evaluated to show its ability.
