基于非高斯分布的ICA纹理分类方法研究

Non-Gauss distribution's ICA texture classification

纹理特征的提取是纹理分析的重要研究内容之一。基于独立成分分析的纹理分类方法中,目前主要选取独立成分分析系数的二阶统计量或频谱直方图作为纹理特征描述。本文根据独立成分分析系数满足非高斯分布的特点,提出了一种新的对野值鲁棒的纹理特征提取方法,该方法较好地描述了独立分量分析系数分布的反对称性和稀疏性,在仿真实验中取得了较好的分类效果。同时针对不同的特征提取方法,利用独立性不同的滤波器研究了滤波器独立性对分类性能的影响,进一步揭示了独立成分分析系数的非高斯性与特征辨识能力之间的关系。

Extracting effective features remains to be an important research problem in texture analysis. Based on independent component analysis （ICA）, second-order statistical features and spectral histogram of ICA coefficients have been used to characterize texture properties for classification purpose. In this paper, according that the ICA coefficients satisfy non-gauss distribution in texture classification based on independent component analysis （ICA）, new features extracting method is put forward, which is robust to outliers. The method measures ICA coefficients＇asymmetry and sparsity, which are shown to yield better classification performance than the previously reported ICA features. Furthermore, the influence of different texture feature effected by independent of ICA filters are investigated. The non-gauss property of ICA coefficients and the discriminating power of features are further revealed.