A novel feature extraction method using histogram-based sparsity

From a perspective of feature extraction, we present a histogram-based sparsity descriptor (HSD) which is derived from the robust principal component analysis (RPCA) and histogram technique. Given a test image, sparse error images with respect to each class can be obtained by using RPCA decomposition. In order to extract the facial features in terms of intensity distribution, a sparseness measure based on intensity histogram is then introduced by computing the histogram sparsity of those sparse error images. By doing this, we can firstly choose t candidates similar to the test face image. Finally, we use LRC to find the correct individual. The efficacy and robustness of the proposed method is verified on three popular face databases (i.e., ORL, CMU PIE and AR) with promising results. We show that if the sparsity hidden in error images can be properly harnessed, the proposed HSD is able to correctly recognize corrupted or occluded faces.