Face verification based on gabor region covariance matrices

Abstract

This paper introduces a novel face verification approach using the Gabor Region Covariance Matrices (GRCM). First, we represent the face images with d dimensional Gabor images. Then, we divide these images into overlapping regions. From each region, we compute a d × d covariance matrix. Inspired by the GMM-UBM speaker verification framework, we propose a new decision rule based on the Riemannian mean of the Gabor region covariance matrices computed from background faces. Finally, score normalization techniques are incorporated in the proposed framework to enhance the verification performance. Extensive experiments on two benchmark databases, namely Banca and SCface showed very interesting results which compare favorably against many state-of-the-art methods.