利用多頻道分解以及隱藏式馬可夫模式做紋理狀影像的辨認

Abstract

本論文所研究的是紋理狀影像的辨認，我們提出了一個架構，結合多頻道 分解以及隱藏式馬可夫模式來解決問題，這個架構不受影像旋轉及灰度轉 換的影響。首先將二維的紋理狀影像，利用螺旋狀的取樣取得一維的訊號 ，然後利用正交對稱濾波器(QMF)將訊號分成數個子頻(subband)，在各個 子頻上取自相關函數(二階及高階)作為特徵值，將這一串特徵值訓練成一 個隱藏式馬可夫模式；在辨認時未知紋理狀影像的特徵值跟所有的隱藏式 馬可夫模式比對，找出機率最大的一個，則此架構辨認此未知影像是屬於 此隱藏式馬可夫模式。在此系統架構中，不受灰度轉換的影響是由去平均 值及正規化達成；而不受旋轉的影響則是藉由螺旋狀的取樣及隱藏式馬可 可夫模式達成。從實驗結果中，可看出此系統架構具有良好的辨識效果， 對16種紋理狀影像的最高辨識率可達到95.14%。 This thesis concerns the problem of texture recognition. We propose an algorithm to combine the multichannel wavelet decomposition and hidden markov model(HMM) to solve the problem. This algorithm is invariant to rotation and gray scale transformation. First we convert 2-D texture images to 1-D signals by spiral sampling and use the QMF to decompose signals into subbands signals. In each band, we take the autocorrelation functions (including 2nd and high orders) as features. And the sequence of features are modeled as an HMM. During recognition, the unknown texture is matched against all the models and the best matched one identifies the texture class. In our algorithm features are made invariant to gray scale transformation by mean removal and normalization. The rotation problem is solved by using the spiral sampling and HMM. From the experiments, we see that the algorithm shows good recognition result. The highest recognition rate for 16 kinds of textures can be up to 95.14%。