Face Frontalization And Recognition Method Of Generative Adversarial Network Based On Identity Constraint And Two-stage Mask Guidance

With the rapid development of deep learning and the continuous growth of public face datasets,the accuracy of face recognition on standard datasets has already surpassed that of humans.However,face recognition is essentially a passive biometric technology that recognizes non-cooperative objects.Faces acquired in an unconstrained environment usually have large changes in posture.When the yaw angle of the face exceeds 45°,the accuracy of face recognition drops rapidly.Face frontalization and re-recognition based on the Generative Adversarial Network can improve the accuracy of face recognition under large yaw angles.This article is based on the Generative Adversarial Network face frontalization framework to improve,to generate as realistic as possible while maintaining identity information consistency of frontal images.The main research content of this paper includes the following aspects:(1)A face frontalization and recognition method based on multi-task learning and identity constraints based on Generative Adversarial Networks is proposed.Based on face frontalization framework of Generative Adversarial Networks,this method draws on the multi-task learning mechanism,designs an pose classification module and an identity constraint recognition module,and embedded them in the encoder to obtain the pose coding feature and the identity coding feature respectively.At the same time,the generator also inputs the frontal face image,and the goal is to generate its own frontal face image to obtain the identity coding feature of the frontal face,and by designing a feature loss function to constrain the identity coding feature of the profile face to approximate the identity coding feature of its frontal face.The experimental results on the M~2FPA dataset show this method can generate a frontal face image that maintains identity consistency,and improves the accuracy of face recognition under a large yaw angle.(2)A two-stage Generative Adversarial Networks face frontalization and recognition method based on mask guidance is proposed.The method divides the face frontalization into two stages.The first stage is based on Generative Adversarial Networks from the facial features mask image of the profile face to generate the facial features mask image of the frontal face,and the second stage uses the generated facial features mask image to guide the face frontalization framework based on Generative Adversarial Networks to generate the frontal image.The experimental results show that the frontal image generated by this method has better consistency with the real facial image in the local area of the facial features.(3)A face frontalization and recognition method based on symmetry prior and local discrimination based on multi-path Generative Adversarial Networks is proposed.This method first flips the face image with a yaw angle of-90°?0°horizontally as a face image from 0°?+90°,and uses a local generator to correct the left eye area,nose area,and mouth area of the profile face image.According to the inherent biological characteristics of the symmetry of the front face,the frontal left eye area is horizontally flipped as the frontal right eye area,and finally the global generator merges the results of the local frontalization to generate the frontal face image.At the same time,in order to allow the global generator to better fuse the local frontalization results,the corresponding local area of the generated frontal image is extracted for image discrimination and identity discrimination.The experimental results show that the frontal image generated by this method can better restore the texture details of the local area of the facial features,and the idea of local matching can further improve the accuracy of face recognition.