Key Technology Researches On Multi-pose Face Recognition

Face recognition has extremely important academic research value and very broad market application prospects.The research on face recognition in real scenes is of great significance to public security and national defense construction.After long-term research and development,face recognition has achieved rich results and progress.However,face recognition technology faces many challenges in real scenes and it is urgent to continue in-depth research,excavation,and solutions.In the wild,multi head pose changes have a significant impact on the performance of the face recognition system.Improving the face recognition rate,in this case,is a challenging task.To this end,this article proposes a series of methods to deal with multi head pose changes,which are specifically introduced as follows:In Chapter 2,the Discrete Gaussian Distribution Label(DGDL)for head pose modeling is proposed.Based on DGDL,a novel head pose estimation method is proposed.Existing meth-ods mainly use three-dimensional face models,facial landmark points and image classification methods to model the head pose.Unlike them,converting the head pose angle into the dis-crete Gaussian distribution label,which can reflect the arbitrariness and continuity of the head pose more accurately.This chapter also proposes a Spatial Channel-aware Residual Attention(SCR-AT)to solve the problem that the current head pose estimation methods are not robust to different ROIs.Finally,this chapter proposes an end-to-end head pose estimation framework.Such a framework uses an head pose estimation loss function based on DGDL for supervised training and uses SCR-AT to further optimize the robustness of head pose estimation to different areas of interest on faces,with high accuracy and quickness.In Chapter 3,based on the DGDL proposed in Chapter 2,a multi-pose face synthesis method is proposed,namely Identity-and-Pose-Guided Generative Adversarial Network(IPG-GAN).Specifically,IPG-GAN uses a dual training strategy that uses the identity and head pose information of the training image at the same time in each iteration of training.Unlike most current methods that can only synthesize frontal faces or limited profile faces,IPG-GAN pro-poses a new adversarial head pose estimation loss function based on DGDL,which is used to guide the generation network to synthesize photorealistic face images with original identity in-formation.According to the target DGDL code,IPG-GAN can synthesize a face image with the specified head pose.Finally,this chapter discusses and verifies that synthesizing the profile face from the front face can better retain the identity information of the face,synthesizing faces from different head poses into the same pose and then comparing them will help improve face recognition rate under multi head pose changes.In Chapter 4,based on the DGDL proposed in Chapter 2,a multi-pose face classification loss function and a feature transformation method are proposed,namely Pose-Guided Margin Loss(PGML)and Pose-Guided Feature Transformation Network(PGFT-Net).Usually,most of the current face representation learning methods do not consider the head pose while train-ing.Unlike them,PGML uses DGDL to perform independent supervision processes for each class of the face images based on an improved Softmax loss function during training,and the learned features are then soft clustered about the head pose.That is,in each class of the feature subspaces,the smaller the difference in the head poses,the smaller the feature distance of the face features,and there is no clear cluster center.Therefore,when a head pose is specified,there are clear boundaries among the features of different identities,which greatly improves the separability.In addition,the proposed PGFT-Net is trained under the supervision of the PGML along with the feature extraction network,and it can transform the deep face features from dif-ferent poses into the same according to the target DGDL encoding so that the compared features correspond to the same head pose,leading to the improvement of the recognition rate.In Chapter 5,a 3D-2D Unconstrained Multi-Pose Face Database(3D-2D-UMPFD)is es-tablished.Based on the PGML proposed in Chapter 4 and other classification loss functions,3D-assisted 2D face recognition is researched on this database.Different from most of the current public face databases,3D-2D-UMPFD contains real collected 3D faces,corresponding multi-pose high-definition faces,and video faces.The two-dimensional face images include rich changes in head pose,illumination,expression,resolution,motion blur,and other condi-tions.In this chapter,3D faces are used for registration.Then a number of 2D face images with specific head poses are obtained through projection so that the video face can be compared with the projection face images with similar head poses like itself.Compared with the traditional 2D face recognition,the 3D-assisted 2D recognition can effectively improve the face recognition rate under multi head pose changes.Additionally,based on the 3D-assisted 2D face recogni-tion,using the PGML proposed in Chapter 4 can further improve the recognition rate compared to the traditional face classification losses.