A face recognition method based on residual image representation and feature extraction

To address the problem of non-well controlled face recognition, such as illumination changes, pose variation and random pixel corruption, we propose a robust face recognition method based on representation and feature extraction of residual images. Represented by sparse representation and linear regression, linear representation methods typically use training samples to represent and reconstruct test samples, and determine classification results according to the distance between test samples and reconstruction samples. In this paper, we consider to use linear regression to obtain reconstruction samples of the test sample with respect to each subject, and compute residual images by the difference between test sample and reconstruction samples. Then we analyze intensity distribution of residual images between the correct subject and other subjects, and adopt intensity transform to surpass the intra-class difference and strengthen the inter-class difference. Finally, we use wavelet decomposition to extract global intensity distribution of residual images, and introduce information entropy to illustrate the uncertainty of intensity distribution, which are extracted as discriminating features. Compared with several popular face recognition methods, the efficacy of the proposed method is verified on four popular face databases (i.e., ORL, Extended Yale B, Georgia Tech and AR) with promising results.