Local Binary Patterns Based on Subspace Representation of Image Patch for Face Recognition

摘要

In this paper, we propose a new local descriptor named as PCA-LBP for face recognition. In contrast to classical LBP methods, which compare pixels about single value of intensity, our proposed method considers that comparison among image patches about their multi-dimensional subspace representations. Such a representation of a given image patch can be defined as a set of coordinates by its projection into a subspace, whose basis vectors are learned in selective facial image patches of the training set by Principal Component Analysis. Based on that, PCA-LBP descriptor can be computed by applying several LBP operators between the central image patch and its 8 neighbors considering their representations along each discretized subspace basis. In addition, we propose PCA-CoALBP by introducing co-occurrence of adjacent patterns, aiming to incorporate more spatial information. The effectiveness of our proposed two methods is accessed through evaluation experiments on two public face databases.