基于对偶树复小波变换的纹理特征提取及分割

Feature Extraction and Segmentation for Texture Based on Dual-tree Complex Wavelet Transform

对偶树复小波变换具有良好的方向选择性和平移不变性。该文在分析对偶树复小波分解后的6个高频子带所对应的模值子带直方图的基础上,提取一种新的纹理特征——Gamma分布参数与Lognormal分布参数的组合特征。基于该特征进行纹理图像分割,分割过程中使用了边缘保持平滑技术对特征进行平滑,并使用K均值聚类实现无监督分割。实验表明,文中所使用的特征提取方法新颖,分割结果的边缘准确性与区域一致性具有抗噪性,是一种有效的纹理分割方法。

Dual-Tree Complex Wavelet Transform（DT-CWT） can provide shift invariance and directional selectivity. In this paper, the histograms of the magnitude sub-bands produced by the DT-CWT six high-frequency sub-bands are analyzed. A novel texture feature based on the Gamma and Lognormal probability density is proposed, the Gamma and Lognormal parameters as the feature vector are chosen and the features using the technique of Edge Preserving Noise Smoothing Quadrant（EPNSQ） is smoothed. Furthermore, a texture segmentation algorithm is implemented by K-means clustering using the smoothed features. The experiments demonstrate the novelty and effectiveness of the proposed feature and segmentation algorithm with better boundary precision and region consistency.