Generative Adversarial Networks Based Face Frontalization

Face recognition is one of the most widely studied topics in computer vision.There are wide applications of face recognition,such as access con-trol,identity verification and video surveillance.Dealing with variations of face images is the key challenge in many face-related applications.Most re-search efforts have focus on how to distinguish intra-personal variations of pose,lighting,expression,occlusion and age from inter-personal variation which distinguishes face identities.Among them,pose and expression have always been important challenges because they can dramatically increase intra-person variances,sometimes even exceeding inter-person variances.Traditional face recognition techniques have made great success under well-controlled condi-tions.However,in unconstrained environment,it is hard to perform stable face recognition under the influence of pose,illumination,and change of facial ex-pression.The focus face recognition research has gradually changed from face recognition in constrained environment to face recognition in unconstrained en-vironment.Face variation research attaches great significance for practical ap-plication.The main content of this paper is face synthesis based face recognition in unconstrained environment,including face identification(1:N)and face veri-fication(1:1).The key idea is to eliminate effects of face variations and im-prove face recognition by face synthesis.This paper focuses on face multi-pose synthesis and face normalization.These two methods present the key topic in different ways,and study face recognition both in constrained environment and unconstrained environment.For the pose problem,this paper considers face identity representation and face variations representation,based on single task multi-pose face generation model.For the weakness of single task model,the proposed task specific multi-pose face generation model construct two network for two different subtask.The two subtask are interacted on each other,and promote performance of face generation and face recognition.Extensive experiments on MultiPIE database demonstrate the effectiveness of task specific networks.To further eliminate variations for face recognition in unconstrained envi-ronment,this paper propose Generative Adversarial Networks based face nor-malization model to distil face identity and dispel face variations.FNM apply GAN to eliminate face variations,and apply face expert network to keep face identity.With the reconstruction of normal face,pixel-wise loss is applied to stabilize optimization process.Local attention mechanism is proposed to re-fine local facial texture.Extensive experiments on both constrained and un-constrained databases demonstrate the ability of proposed FNM in generating high-fidelity normalized face and improving face recognition.