Research On Multi-pose Face Recognition Algorithm Based On Deep Learning

In recent years,with the rapid development of computer technology and the application of Internet technology,face recognition technology has been widely used in attendance attendance,intelligent monitoring,identity verification and face payment.With the introduction of deep learning,its recognition rate has also achieved good results.However,in an unrestricted scene,the recognition performance is poor due to uncontrollable factors such as gesture,illumination,angle of view,expression,and occlusion.Among them,attitude change is one of the most important factors and the difficulty of current research.In this paper,the problem of pose change in face recognition is studied.From the perspective of pose transformation,the generation of multi-pose virtual face and pose estimation are proposed into the multi-pose face recognition algorithm.The main research work of this paper is as follows:(1)Introduce the basic idea principle from the two perspectives of attitude correction and transformation,and compare and summarize the advantages and disadvantages of each.It is pointed out that the virtual sample obtained by pose transformation can be better in the multi-pose face recognition problem.Identify the effect.(2)Research on the generation of virtual samples,and propose a multi-pose virtual view algorithm based on VAE/ACGAN.With the idea of VAE/GAN,the combination of ACGAN's auxiliary conditional tag generation type anti-network and variational self-encoder VAE is used to generate multi-pose virtual samples and increase training samples to improve the recognition rate.(3)Studying the pose estimation problem,by analyzing the advantages and disadvantages of the existing methods,based on the geometric relationship of the key points of the face,the pose estimation based on the MTCNN key point detection is proposed.Firstly,the MTCNN is used to detect the key points of the face,and then the geometrical relationship of the key points is used to obtain the deflection angle of the face to realize the attitude estimation.The algorithm proposed in this paper firstly uses the VAE/ACGAN network model to generate corresponding multi-pose virtual samples from all faces in the frontal face database to construct a new training sample library.At the time of identification,the object to be identified is firstly estimated,and then the FaceNet network is used to extract features from the face in the same posture to realize face recognition.Finally,the comparison experiments between MultiPIE and CelebA datasets show that the average face recognition accuracy of the algorithm is 87.98% and the robustness of the attitude changes is good.