Research On Multi-angle Face Frontalization Based On Generative Adversarial Network

Nowadays,face recognition technology has been widely used in production and life,while ensuring information security,it also brings many conveniences to users.However,in the surveillance scene,due to the installation height of the camera,the face is subject to the multi angle pose changes of yaw and pitch direction at the same time,and it is difficult to obtain a frontal face image,resulting in a serious degradation of the performance of the existing face recognition model.The face frontalization method based on the generative adversarial network can generate a realistic virtual frontal face that retains the identity information of the original input profile image,which is nested into the existing face recognition model as a preprocessing process to alleviate the influence of pose change on the performance of the recognition model.At present,there are three difficulties when using the generative adversarial network to frontalize multi-angle side face images.The first is the problem of self-occlusion due to rigid rotation of the face,which makes the key identity features of the face missing;the second,the pitch changes,which causes the facial features to deform;the third is the problem of low retention of identity features in generating virtual front images.Therefore,the main work of this dissertation is as follows:(1)A generative adversarial network based on the symmetry of facial feature maps is proposed.Aiming at the problem of face self-occlusion and low retention of virtual frontal face identity,this dissertation proposes facial feature map symmetry and periocular features loss based on the priori of face symmetry and periocular recognition.In this dissertation,the key point detector is used to detect the nose tip position of the side face,and the feature map extracted by the encoder is mirror symmetrical according to the nose tip position,which alleviates the lack of facial information from the feature level.Secondly,on the basis of the existing global identity feature retention method,the periocular features loss is added to improve the identity information retention ability of the generated image globally and locally.(2)A generative adversarial network based on facial feature mapping is proposed.Aiming at the problems of self-occlusion,facial features deformation and low identity retention of virtual frontal face images,this dissertation proposes a face feature mapping module based on the inherent mapping relationship between the frontal and profile features of the same person.Based on the profile feature encoder,a branch encoder that maps profile features to front features is added to form a two-way encoder,so as to alleviate the problem of lack of facial identity information and changes in facial regions.In addition,the periocular features loss is improved,and the pixel level loss of eye area is designed to improve the local identity information retention ability of virtual frontal face without destroying eye features.(3)Designed and developed a multi-angle profile image correction system.The function design and implementation of each module in the system are introduced in detail,and the correction model proposed in this dissertation is applied to the relevant modules.Finally,the corresponding results are displayed on the front-end page.