Texture classification using compact multi-dimensional local binary pattern descriptors

Texture analysis and classification plays an important role in many computer vision systems and applications. Local binary patterns (LBP) form a simple yet powerful texture descriptor characterising local neighbourhood properties, which, due to its effectiveness and robustness, is widely used. LBP descriptors can also be recorded at different radii leading to multiscale features. While in conventional LBP, this information is recorded, in form of a histogram, separately for each of the scales, it was shown that a multi-dimensional feature representation removes some ambiguity and leads to better texture classification. However, these multi-dimensional LBP (MD-LBP) histograms also make rather large feature descriptors which limits their practical use. In this paper, we show that we can effectively compactify the information contained in MD-LBP histograms, and show, through extensive experiments on seven Outex datasets, that with the same feature length as the original LBP method we are able to obtain clearly improved texture classification which closely matches that achieved by the full MD-LBP descriptor.