Noise tolerant descriptor for texture classification

Among many texture descriptors, the LBP-based representation emerged as an attractive approach thanks to its low complexity and effectiveness. Many variants have been proposed to deal with several limitations of the basic approach like the small spatial support or the noise sensitivity. This paper presents a new method to construct an effective texture descriptor addressing those limitations by combining three features: (1) a circular average filter is applied before calculating the Complemented Local Binary Pattern (CLBP), (2) the histogram of CLBPs is calculated by weighting the contribution of every local pattern according to the gradient magnitude, and (3) the image features are calculated at different scales using a pyramidal framework. An efficient calculation of the pyramid using integral images, together with a simple construction of the multi-scale histogram based on concatenation, make the proposed approach both fast and memory efficient. Experimental results on different texture classification databases show the good results of the method, and its excellent noise robustness, compared to recent LBP-based methods.