Texture Classification Using Multi-dimensional LBP Variance
Niraj P. Doshi, Gerald Schaefer · 2013
Texture classification is an important task for a variety of computer vision applications. A successful group of texture algorithms based on local neighbourhood descriptors and known as LBP (local binary patterns) has been shown to provide good and robust discriminative power, and is typically applied in a rotation invariant form and calculated at multiple resolutions. Local contrast information can be integrated into the LBP histogram generation by using the variance as weights for LBP, leading to LBP variance (LBPV) texture features. Multi-scale LBPV histograms are obtained by concatenating the individual one-dimensional histograms derived from each scale. In this paper, we show that by calculating a multi-dimensional LBP variance (MD-LBPV) histogram improved texture classification can be achieved. We confirm this based on extensive experiments on several Outex benchmark datasets.