Pose Variable Face Recognition Based On Deep Learning

In recent years,with the improvement of computer hardware conditions and the rapid development of computer vision technology,face recognition technology has been widely used in many occasions in social life,such as check-in attendance,identity verification,and face-to-face payment.With the support of deep learning technology,the accuracy of face recognition has been greatly improved.However,in unrestricted scenes,due to uncontrollable factors such as head posture,lighting changes,facial expressions,occlusion,and makeup changes,it has led The recognition effect has dropped significantly.In this paper,the research on the problem of reduced recognition performance caused by changes in face poses has been carried out.With the theme of face pose changes,experiments have been conducted around face detection,positive face synthesis,and extraction of face-independent face features.The purpose is to improve the accuracy of the multi-pose face recognition algorithm.The main research work of this paper is summarized as follows:(1)MTCNN is the current popular face detection and face alignment algorithm,but when the face image has a large-angle posture change,the network's missed detection rate will also increase,and MTCNN is based on the multi-task cascade structure of the volume With integrated networks,the speed of face detection needs to be improved.In response to the above problems,we use data enhancement operations to flip the training samples,rotate the operation to obtain more face images under a greater angle of view,increase the posture diversity of the training sample data,improve the detection ability of the model,and introduce Octconv convolution to replace R-Net,The traditional convolution operation in O-Net network to reduce the network parameter redundancy and improve the network running speed.Finally,according to the position of the face detection frame of MTCNN and the coordinate information of the left and right eyes,the face pose estimation is performed by measuring the angular relationship between the left and right eyes and the geometric center of the face area,and the face recognition and face pose are summarized for the subsequent chapters.Yihua provides preparation work.(2)Considering that the change of the face pose will cause the intra-class gap of facial features to become larger and the inter-class gap to narrow,this paper proposes a multi-pose face recognition algorithm based on the feature adaptive mapping method of face pose.First add the face detection module(MTCNN),and use the angle relationship between the left and right eyes and the geometric center of the face area to estimate the face head pose,and then use the deflection angle of the face in the yaw direction to construct the face pose constraint The parameters are used for network training.Secondly,use the residual network to build a feature mapping module based on face pose.This module can adaptively add face pose constraints to the face depth feature representation to map the face features with changed poses to the front face.The face feature space is as close as possible.(3)In order to further study the impact of face pose changes,this paper proposes a face synthesis method based on WGAN-gp generation adversarial network.Compared with the traditional face synthesis method based on GAN framework,this model integrates an identity recognition module,Can extract effective facial identity features and used to supervise the training of the face generation module.In addition,this method makes full use of the a priori knowledge of the symmetry features of the human face,and designs a human face symmetry feature extraction module.The role of this module is to improve the visual quality of the synthesized face on the one hand,and to speed up the network on the other.Convergence speed of training.We conducted verification experiments on multiple face databases,and showed that this method improves the graphic quality of positive face synthesis.