在本論文中，我們提出一個使用關聯隱藏馬可夫樹模式 (contextual hidden Markov tree model, CHMT) 來做多重解析度紋理影像分割 (multiscale texture image segmentation)。這個模式是由建立在小波轉換(wavelet transform) 架構下的隱藏馬可夫樹模式 (hidden Markov tree model, HMT) 改良而來的。兩個模式都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)。在本研究中，我們改進的地方就是使用虛擬係數 (virtual coefficient) 的觀念，來加強HMT模式的聚集性。在影像分割的應用上，我們使用貝式 (Bayesian) 判定的觀念來分割紋理影像。首先，對於每一種紋理，我們利用只含有此種紋理的影像訓練關聯隱藏馬可夫樹模式的參數；接著利用這些參數算出不同解析度區塊的相似度 (likelihood) 函數值；再比較函數值的大小就可以將不同解析度的區塊一一分類；最後，融合 (fuse) 大區塊跟小區塊的分割結果以得到更精確的分割結果。

A multiscale texture image segmentation approach using Contextual hidden Markov Tree (CHMT) model is proposed. In 1998, a tree-structure statistical model called hidden Markov tree (HMT) model that captures the interscale dependences of the wavelet coefficients was presented. The CHMT model was improved from the HMT model by enhancing the intrascale dependences between wavelet coefficients in a scale.
Based on the description of the CHMT model, we employ a Bayesian segmentation method to perform texture image segmentation. Initially, the model parameters are trained by the EM algorithm. For each texture type, we train parameters in wavelet domain from training images which is consisted of homogeneous regions. Based on the trained parameters, we can compute the likelihood function of blocks at different scales and make a multiscale classification. At last, we fuse the classification results of large blocks and small blocks to obtain a more accurate segmentation result.

[1] Bouman, C. and M. Shapiro, “A multiscale random field model for Bayesian image segmentation,” IEEE Trans. Image Processing, vol.38, no.2, pp.162-177, Mar. 1994.
[2] Brodatz, P., ed., Textures