Research On Multi-angle Face Recognition Of Video Sequences Based On Face Synthesis

Face recognition is one of the important methods in human identification tasks based on computer vision.The current face recognition methods are mostly aimed at solving the recognition problem of frontal or near frontal faces,so most of them require the system to capture the frontal image of the human for recognition.However,video sequences often contain complex environmental backgrounds,with human faces showing multiple poses,or even largerangle profiles.In this case,most of the existing methods have poor recognition results.So how to solve the multi-angle face recognition problem based on video sequence is one of the huge challenges facing the current human-machine interaction system.In order to solve the problem of poor face recognition results due to the severe lack of facial texture in profile images.This paper proposes a multi-angle face recognition system for video sequences based on face synthesis.The recognition system improves the detection and alignment module in the traditional face recognition system.The specific research work is as follows:Firstly,in order to eliminate the influence of the large inclination of the face area in the horizontal direction on the subsequent module,two facial feature point positioning models,Open Pose and Face Alignment Network,are used to improve the detection module in a typical face recognition system.A face leveling method based on feature point positioning is proposed,and face preprocessing is performed to make the system suitable for face recognition of various horizontal poses in video sequences.Secondly,for the facial area where there is still a large side face posture after leveling,TPGAN(Two-Pathway Generative Adversarial Network)is used to generate the front face to realize accurate front face recognition.Furthermore,a facial scaling method based on feature point positioning is proposed,and the input end of TP-GAN model is adjusted to handle any size of video sequence.Thirdly,for the frontal synthesis face recognition results from the video sequence,a side face angle of 50° is set as the system threshold.At the same time,according to the output characteristics of facial feature point location model,a data pre-screening optimization algorithm is proposed to eliminate the interruption of the subsequent recognition caused by the error frame of the positioning failure,and optimize the performance of multi-angle face recognition.Finally,the proposed system(OP-TPGAN,FAN-TPGAN)is compared with the traditional face recognition system on the multi-angle face video dataset and XM2 VTS dataset,including the stability of multi-angle face positioning,the accuracy of large pose profiles recognition and the accuracy of multi-angle face recognition in complex scenes.The experimental results show that the proposed method achieves good results in multi-angle face location and recognition.