Moment Features in Directional Subband Domain for Rotation Invariant Texture Classification

Abstract

This paper presents a study on moment features in directional subband domain for rotation invariant texture image classification. The directional subband decomposition is obtained through a biorthogonal angular filter bank. Moment features are extracted from each directional subband. Two rotation invariant feature generation techniques are examined, including eigenanalysis of covariance matrix and DFT encoding. Feature vectors are further classified by multi-class linear discriminant analysis (LDA). LDA training is based on feature vectors collected from non-rotated training images, and test is performed on images rotated at various angles. Experimental results are provided to demonstrate the effectiveness of directional subband domain feature extraction method for rotation invariant classification. Performance of various feature sets are compared, and the best feature combination is presented