Pose-invariant Face Recognition

Face recognition is an important research topic in the field of biometrics,and it is widely used in social life, including video surveillance, security, con-venient payment and so on. Traditional face recognition techniques have made great success under well-controlled conditions. However, in the unconstrained environment, it is hard to perform stable face recognition under the influence of pose, illumination, and change of facial expression.The focus face recognition research has gradually changed from face recognition in constrained environ-ment to face recognition in unconstrained environment. The pose problem is one of the most important ones in the mentioned factors. It is of great signifi-cance to study the pose problem of face recognition for the practical application.The main contents of this paper are face identification (1: N) and face verification (1: 1). It is mainly concerned with two kinds of methods, face pose synthesis and face pose normalization. In this paper, the pose problem in face recognition is described first, and the related algorithms are introduced. And then the face pose synthesis and face pose normalization are presented as two kinds of method.For the face pose synthesis algorithm, the method Generic Elastic Model(GEM), which reconstructs three-dimensional face model from a single input face image is extended to, Multi-Depth Generic Elastic Model (MD-GEM),which makes use of the hypothesis that the depth of the human face varies lin-early across individuals. It makes up for the deficiencies of the GEM using a single depth map model, and the pose-invariant face recognition experiment on the MultiPIE database proves that the MD-GEM outperforms GEM. Then,GEM is combined with Quotient Image (QI) method as extension in terms of illumination. Idea of generate samples virtually under various poses and illumi-nation conditions is carried out, based on single training sample. The pose and illumination-invariant face recognition experiments on the MultiPIE database achieves recognition rate of 91.1%, as good as state-of-the-art, and has the char-acteristics of requiring less training samples and less parameters to be tuned.For the face pose normalization algorithm, this paper presents an algo-rithm based on 3D model which can keep the consistency of human face illu-mination. A generic 3D face model is aligned to the input face image based on the detected five landmarks. The face contour is detected for the accurate estimation of the self-occlusion region. We apply the Quotient Image, as a face symmetrical feature which is robust to illumination to fill the self-occlusion re-gion so that the output can keep the light condition of input face. This method also provides the idea of simultaneously lighting normalization. State-of-the-art performance is achieved in LFW face verification experiment, 91.5% and Mul-tiPIE pose-invariant face recognition experiment, 99.5%,with the advantages of suitable for real-world scenario applications. Then, the pose and illumination normalization based on deep learning is explored. The pose and illumination normalization is considered as a nonlinear transformation problem, which can be solved by deep convolution neural network. By using the method of sample synthesis to achieve data augmentation, the number of training data is much less while the performance is as good as other similar methods.At the end of this paper, we analyze and compare the advantages and disad-vantages of relevant algorithms in this paper, and propose some future research directions.