Decorrelated local binary patterns for efficient texture classification

Local binary patterns (LBP) has been successfully applied to several tasks in computer vision due to its efficacy and computational simplicity. LBP can be computed in different neighborhoods to derive multi-scale LBP (MS-LBP) for performance enhancement. In MS-LBP, different scales LBP histograms are either combined in a concatenate way assuming that different scales LBP are independent, or jointly combined to generate a multi-dimensional histogram. However, the independence assumption does not hold so that the cross-scale information is lost in concatenate MS-LBP, while joint MS-LBP suffers high feature dimension. Then, to deal with the independence assumption and better exploit multi-scale information, a new texture descriptor called decorrelated local binary patterns (dLBP) is proposed in this paper. Unlike traditional MS-LBP schemes, discrete cosine transform (DCT) is firstly applied as a decorrelation transform to different scales differences to derive independent patterns. Then, the histograms corresponding to each pattern are concatenated as a new texture descriptor. Besides, the decorrelated magnitude components are also utilized to further enhance the performance. Experimental results show that the proposed dLBP features outperform both the concatenate MS-LBP and some recent state-of-the-art schemes for texture classification.