Oriented Gradient Maps based automatic asymmetric 3D-2D face recognition

Recently, asymmetric 3D-2D face recognition has been paid increasing attention. It enrolls in textured 3D faces and performs identification using only 2D facial images, therefore it generally achieves a better result than 2D algorithms do, and avoids inconvenience of data acquisition and computation of 3D methods as well. In this paper, a biological vision-based facial representation, namely Oriented Gradient Maps (OGMs), is introduced for such an application. It simulates the response of complex neurons to gradient information within a pre-defined neighborhood, and thus can describe local texture changes of 2D faces and local geometry variations of 3D faces at the same time. Due to its property of being highly distinctive, these OGMs improve accuracies of both matching steps of asymmetric face recognition, i.e. (1) 3D-2D matching using Canonical Correlation Analysis (CCA); (2) 2D-2D matching using LBP histogram based features and Sparse Representation Classifier (SRC). Some comparative experiments are carried out on the complete FRGC v2.0 database, and the achieved results clearly highlight the effectiveness of the biological vision-based facial description and its successful application to asymmetric face recognition.