Non-frontal face recognition method with a side-face-correction generative adversarial networks

Frontal face image recognition is the main target of traditional face recognition.The deflection of the human face often causes the dislocation of the facial features,which leads to the reduction of the recognition accuracy of the non-frontal face.To solve the above problems,a non-frontal face recognition model based on generative adversarial network is proposed.In this model,the angle information is encoded separately by using a two-channel generator and auto-coding network,and the non-frontal face image in natural environment is corrected to obtain the frontal face image.Through the multi-discriminator mechanism of facial attention,we set discriminators in the eye, eyebrow, nose, mouth and the whole area of the face image so as to retain the details of the face to the maximum extent while ensuring the clarity of image.Then the corrected face features are extracted by Facenet and MTCNN to obtain the non-frontal face recognition results.The model is validated on multi-PIE dataset and CFP dataset.The results show that the accuracy of non-frontal face recognition is improved by 1% in CFP dataset compared with VGG-FACE, TP- CNN and HPN.