Research On Face Frontalization Method Based On 3D Model Symmetry Constraints

In the current information age,identity authentication through biometrics has become the primary method in various scenarios.Due to the non-contact nature of face features,even if the person to be identified is in a long-distance state,his identity can be accurately and quickly confirmed,so face recognition technology is particularly important.Illumination changes,posture problems,and wide-range expression changes will all affect the results of face recognition.One of the important reasons that affect face recognition is the variety of pose changes.Pose changes will cause the face partition area to be occluded and cause information loss.The current solution is to rotate the face pose to the front,that is,face frontalization,and then supplement the missing area information.Therefore,the study of face frontalization is of great significance in face recognition.The research methods of face frontalization can be divided into two types according to whether the 3D model of the face is used,namely the method based on 2D and the method based on 3D model.The traditional 3D model-based method constructs a 3D model of the face,fits a2 D face image to the 3D model,and then rotates it to a frontal pose and maps it to a 2D image to achieve frontalization.Recently,Generative Adversarial Nets(GAN)have developed rapidly,the current mainstream research method is to combine 3D models with GAN,which effectively guarantees the quality of generated images while increasing the generation speed.The research content of this paper is as follows:(1)In order to solve the problem that the existing face frontalization of 3D model with GAN needs to be supervised by paired images,and the generated frontal face image lacks symmetry,this paper proposes a face frontalization method based on symmetry constraints.By performing multiple rotation and rendering operations on the acquired 3D face model,a self-supervised image pair is constructed,and then repaired using GAN to solve the problem of no paired images in an unrestricted environment.The facial geometric feature points in the face image are robust to factors such as illumination and posture.Many face attributes are symmetrical,such as eyebrows and eyes are completely symmetrical,and the nose and mouth can be considered to be left-right symmetrical about the midline of the face.The frontal face image generated by directly using the multiple rotations of the 3D model of the face to realize the frontalization of the face will lack symmetry and the result will not be realistic enough.By using the image segmentation network to obtain the face attributes with symmetry,the face with symmetric attributes add symmetry constraints for more realistic frontal face images.The method proposed in this paper is compared with the original method for two experiments.The qualitative experimental results show that adding symmetry constraints in the process of face frontalization can effectively improve the authenticity of the face frontalization results.At the same time,the face recognition network is used for quantitative experiments,the experimental results show that adding symmetry constraints increases the accuracy of face recognition by0.05% when using the LFW dataset for validation,which can be better applied in the field of face recognition.(2)Aiming at the lack of symmetry of the frontal face image generated by using the 3D face model,this paper proposes a face frontalization method based on flip fusion.Not only some geometric features in the face image have symmetry,but also the overall structure of the face image,such as the contour of the face.In the process of constructing paired images through multiple rotations of the 3D model of the face,the rotation principle is used to create a flipped image pair,and then a face image with complete symmetry is constructed through image fusion,and finally the GAN is used for repair generation to obtain the frontal face images.Experimental results indicate that the method of face frontalization based on flip fusion can improve the overall symmetry of the generated image and improve the authenticity.At the same time,experiments were carried out on the face recognition network,and the accuracy of face recognition increased by 0.035% using the LFW dataset for validation,which verified the effectiveness of the face frontalization method based on flip fusion,which can perform better face recognition.