Difference theoretic feature set for scale-, illumination- and rotation-invariant texture classification

Texture identification and classification under varying scale, rotation and illumination conditions is a challenging task in pattern recognition and grey level difference statistics have been extensively used for this purpose. This study presents a new set of features for scale-, rotation- and illumination-invariant texture classification derived from the correlated distributions of local and global grey level differences of intensities in the texture image. The authors analyse the terms in the correlation formula for determining the difference-based feature set that is invariant and unique for a texture class. A comprehensive evaluation is performed on a huge database of digitally created texture samples of varying scale, orientation and brightness. The one-nearest neighbour classifier is used in the authors´ experiments and the results indicate high classification accuracy for the proposed feature vector under varying scale, rotation and brightness conditions. The proposed method is compared with the highly efficient rotation- and illumination-invariant local binary pattern (LBP) and LBP variance techniques and the scale- and rotation-invariant MRS4 technique and is found superior in performance with an additional advantage of reduced feature dimension.