Research On Face Frontalization And Multi-pose Face Generation Based On Generative Adversarial Networks

In recent years,as a biometric technology based on facial feature information for identity recognition,face recognition has been widely used in fields such as intelligent security,financial transactions,public transportation.With the improvement of hardware performance and the improvement of software algorithms,face recognition technology has become increasingly mature.However,due to the large discrepancy of face in different poses,there are still many challenges in recognizing face images with large angles.At the same time,due to the insufficient samples of current multi-pose face databases,it will cause certain difficulties in the subsequent training of the multi-pose face recognition models.In response to the above problems,this paper proposes a face frontalization algorithm and a multi-pose face generation algorithm based on generative adversarial networks.The research content of this paper can be mainly generalized as the following two points:(1)Aiming at the problem of low recognition rate of face images at large angles,this paper proposes a face frontalization algorithm based on two-pathway attention generative adversarial networks.This algorithm adopts a two-pathway generator structure to strengthen the extraction of local face information,making generated faces more realistic.In terms of the discriminator,this algorithm uses the discriminator structure based on attention mechanism.By adding non-local mean attention blocks to the discriminator,the generator can better learn structure information of human face.At the same time,the algorithm also introduces a symmetric loss function and an identity preserving loss function on the basis of the adversarial loss function,and jointly promotes the generator to generate a real frontal image that maintains the identity of human face.(2)Aiming at the problem of insufficient samples in current multi-pose face databases,this paper proposes a multi-pose face generation algorithm.This algorithm divides the multipose face generation into two subtasks: face frontalization and face rotation.Face rotation refers to mapping a profile face image to a frontal face image,and face rotation refers to mapping the frontal face image generated in the previous step to the required pose.This paper proposes a two-stage training method.In the early stage of model training,the two modules are trained separately.When the modules are all converged,the two modules are combined for training through the cycle consistency loss function,which further improves the ability to generate multi-pose human faces.Comparative experiments on the Multi-PIE,LFW and CFP datasets show that the face frontalization algorithm can not only generate visually realistic face images,but also effectively maintain the identity feature information of human face.Experiments on the MultiPIE dataset show that our multi-pose face generation algorithm can maintain the identity information of the face while generating multi-pose faces,and subsequent dataset expansion experiment also shows that the multi-pose face generation algorithm proposed in this paper can effectively expand the face database.