基于方向多尺度变换和统计建模的纹理分类方法

Texture classification based on directional multiscale transform and statistical modeling

纹理是图像分析和识别中经常使用的关键特征,而小波变换则是图像纹理表示和分类中的常用工具.然而,基于小波变换的纹理分类方法常常忽略了小波低频子带信息,并且无法提取图像纹理的块状奇异信息.本文提出小波子带系数的局部能量直方图建模方法、轮廓波特征的Poisson混合模型建模方法和基于轮廓波子带系数聚类的特征提取方法,并将其应用于图像纹理分类上.基于局部能量直方图的纹理分类方法解决了小波低频子带的建模难题,基于Poisson混合模型的纹理分类方法则首次将Poisson混合模型用于轮廓子带特征的建模,而基于轮廓波域聚类的纹理分类方法是一种快速的分类方法.实验结果显示,本文所提出的三类方法都超过了当前典型的纹理分类方法.

Texture is a key feature frequently used for image analysis and recognition, and wavelet transforms are common tools for image texture analysis and classification. However, wavelet-based texture classification methods usually neglect the information in the low-pass subband, and cannot capture piece-wise singularities contained in image texture. In this paper, we propose local energy histograms （LEHs） for modeling wavelet sub- band coefficients, Poisson mixture models （PMMs） for modeling contourlet subband features, and clustering for extracting contourlet subband features. Then, these LEH-based texture classification method alleviates modeling methods are utilized to texture classification. The the difficulty of modeling wavelet subband coei^icients, the PMM-based texture classification method is the first to model contourlet subband features using Poisson mixture models, and the texture classification method based on clustering in contourelet subands is a fast classification ap- proach. Experimental results reveal that our proposed methods outperform some current state-of-the-art texture classification methods.