Research And Implementation Of Face Correction Based On Generative Adversarial Networks

Nowadays,face recognition technology has been widely applied in the field of public security and plays an important role.However,the location of surveillance cameras is fixed,it is difficult to ensure the acquisition of frontal face.As a result,the recognition accuracy is too low to meet the actual business needs.For this reason,this thesis carries out the research of face correction technology.By extracting the local semantic information of face and combining with the theory of generative adversarial networks,the angle correction of face image is achieved,and the accuracy of face recognition is improved.The contents of this thesis are as follows:(1)A method of extracting facial local semantic information based on facial feature points is proposed.Aiming at the problem of correcting image distortion caused by the lack of semantic information in most face correction technology,this thesis proposes a local semantic information extraction method based on facial feature points.The affine transformation is made through the facial feature points,and the face regions are quickly divided to extract high-level semantic information.Experiments show that this method can accurately and efficiently acquire local semantic information of human face.(2)A new face correction technique based on improved generative adversarial networks(GAN)is proposed.Aiming at the problem that current face correction technology can't maintain the identity features of the face,a new face correction technique based on improved generative adversarial networks(GAN)is proposed in this thesis.Local semantic information and global information are combined to constrain the generator,and the encoder-decoder structure is used to learn the potential relationship between the non-frontal face and the front face in the generator.At the same time,the recognition module is added to retain the non-frontal face identity information.Structural loss,semantic region loss and identity preservation loss are introduced to maintain the face structure information and identity features,so as to make the generated face more realistic.Finally,the experimental comparison verifies the corrective effect of the proposed method,and verifies the improvement of the face recognition by the proposed method in this thesis on the face data set.(3)A face correction and recognition system is designed and implemented.The overall architecture of the system is designed,and the design and implementation of each functional module of the system are introduced in detail.The face correction technology proposed in this thesis is applied to the relevant modules.Finally,the system is demonstrated.