Compactifying multi-dimensional LBP variance texture descriptors based on DCT and feature reduction

Texture analysis and classification have received significant research interest and have been shown to be essential in many computer vision systems and applications. Local binary patterns (LBP) are powerful yet simple texture descriptors which describe the texture neighbourhood of a pixel using simple comparison operators, and are often calculated based on varying neighbourhood radii to provide multi-resolution texture description. Furthermore, local contrast information can be integrated into LBP leading to LBP variance (LBPV) features. In conventional LBP methods, the histograms corresponding to different radii are simply concatenated resulting in a loss of information between different resolutions and added ambiguity. Multi-dimensional LBPV (MD-LBPV) preserves the relationships between the scales by building a multi-dimensional histogram of LBPV patterns and can lead to improved texture classification. In this paper, we address the relatively large feature length of MD-LBPV descriptors, and show that feature reduction based on discrete cosine transform (DCT) combined with principal component analysis (PCA) can yield effective and compact texture descriptors with high classification accuracy.