摘 要
紋理是人們視覺系統(tǒng)對自然界物體表面現(xiàn)象的一種感知，它作為物體表面的一種基本屬性廣泛存在于自然界中，是人們描述與區(qū)分不同物體的重要特征之一。紋理分析技術(shù)是圖像處理領(lǐng)域一個經(jīng)久不衰的研究熱點，紋理特征提取作為紋理分析的首要問題，成為人們關(guān)注的焦點。本文在傳統(tǒng)的紋理特征提取方法的基礎(chǔ)上，提出了一種雙樹復(fù)小波域共生矩陣紋理特征提取方法，對提取的特征值用聚類的方法進行性能分析，并應(yīng)用于圖像檢索中。本文的主要工作如下：
1.復(fù)現(xiàn)了GLCM、DT-CWT紋理特征提取方法。通過實驗比較了特征值之間的相關(guān)性，選擇能量、熵、慣性矩和局部平穩(wěn)這4個非相關(guān)特征值。構(gòu)造共生矩陣參數(shù)，通過考察構(gòu)造參數(shù)對特征值的影響來確定構(gòu)造參數(shù)。該方法簡潔、計算量小。在DWT和DT-CWT比較分析之后，利用DT-CWT從多方向和多尺度兩個方面對圖像紋理分析。設(shè)計了濾波器，并驗證了DT-CWT的性質(zhì)。
2.在深入研究GLCM和DT-CWT紋理特征提取方法的基礎(chǔ)上，提出了一種雙樹復(fù)小波域的共生矩陣紋理特征提取方法。該方法利用雙樹復(fù)小波模型，構(gòu)造同時滿足正交和重構(gòu)的濾波器，對紋理圖像進行多層分解。通過計算多層低頻子帶圖像的共生矩陣，提取描述紋理圖像在不同尺度下的特征矢量；通過計算一層分解不同方向子帶圖像的共生矩陣，提取描述紋理圖像不同方向的特征矢量。該方法能有效地描述紋理的尺度特性和方向特性，而且算法直觀簡明。
3.利用聚類分析，對GLCM、DT-CWT和雙樹復(fù)小波域共生矩陣的紋理特征提取方法所提取的特征矢量進行性能分析。將每一幅紋理圖像所提取的特征矢量視為一個聚類的樣本，不同類型所有紋理圖像所提取的特征集合作為不同的聚類。通過聚類內(nèi)部距離、聚類間距離及其比值等指標，分析比較了上述3種方法的特征提取性能。實驗結(jié)果表明，本文提出的方法具有較好的聚類性能，所提取的紋理特征的聚類分離度優(yōu)于其它兩種方法，并能較好地保持聚類內(nèi)部樣本的差異性。
4.將GLCM、DT-CWT和雙樹復(fù)小波域共生矩陣的紋理特征提取方法應(yīng)用于圖像檢索。利用檢索圖像與圖像庫中圖像之間的紋理特征距離函數(shù)作為圖像相似性度量值來檢索圖像，分析比較了上述3種方法的圖像檢索平均查準率。實驗結(jié)果表明，本文提出的方法計算效率高、操作方便，有效地提高了圖像檢索的正確率。

關(guān)鍵詞 紋理特征提?。浑p樹復(fù)小波；灰度共生矩陣；聚類分析；圖像檢索






































Abstract
Texture is a perception of the natural phenomenon from the visual system. It is widespread in nature as one of the basic properties of the surface, which always be used as the improtant characteristics to describe and distinguish the different objects. Texture analysis is a hotspot in image processing, since texture feature extraction is the primary problem of it, has been the focus of attention. A new method of texture feature extraction based on co-occurrence matrix in dual-tree complex wavelet domain was proposed in this paper. The paper analyzed the image texture features with clustering, and apply into image retrieva l. The main work and innovations are as follows:
1.The paper reproduce the texture feature extraction methods of GLCM and DT-CWT. In GLCM, it chooses energy, entropy, inertia and local stationary as the values of the texture feature. The structural parameters are determined by examing the impact of the texture feature. The experimental results showed GLCM is simple and less computation. In DT-CWT, by comprised the CWT and DT-CWT, we verify DT-CWT as the best way to analysis the images texture through the multi-scale and multi-directions.
2.A new method of texture feature extraction based on co-occurrence matrix in dual-tree complex wavelet domain was proposed in this paper. It uses dual-tree complex wavelet to decomposed the image texture with the filters which satisfy both orthogonal and reconstruction. The low-frequency sub-images produced by the multilayered dual-tree complex wavelet decomposition of texture images were utilized to calculate the co-occurrence in different directions for extracting the image texture features. The high-frequency sub-images produced by the multilayered dual-tree complex wavelet decomposition of texture images were utilized to calculate the co-occurrence in different directions for extracting the image texture features. The experimental results showed that this method can effectively extract the texture features in the multi-scale and multi-directions.
3.The paper uses clustering to do the performance analysis for the feature vectors which extract from the GLCM, DT-CWT and the new method. Then make each texture feature vectors as a sample of a cluster, all the different types of feature vectors form the cluster. By comparison the three methods of image texture extraction with internal distance, the distance between the cluster and the ratio of them, the experimental results show that the extracted texture features had favorable cluster separability and kept otherness of samples in the same cluster.
4.The paper apply the GLCM, DT-CWT and the new method into image retrieva l. Comparised three methods with the average precision of image retrieva l, the experimental results show that the new method has efficiency calculation, easy operation and improve the accuracy of image retrieva l effectivel..