Selection of ICA Features for Texture Classification

In the literature of texture analysis, research has been focused on the issue of feature extraction. Much less attention has been given to the important issue of feature selection, however. Most of the methods rank the features by some criteria, for instance, the eigenvalues and the Fish Criterion, and select some percentage of the top features. In this paper, we propose a feature selection scheme for texture classification. We use the filter bank obtained by independent component analysis (ICA) of nature scenes for multichannel feature extraction and the least squares support vector machine (LS-SVM) for classification. The dimension of the ICA features is first reduced using principal component analysis (PCA). Recursive feature elimination (RFE) is then employed to select the relevant features for LS-SVM classification. Our experimental results show that the proposed method achieves better classification accuracy than the simple PCA and the Fisher Criterion methods.