Research On Pose Robust Face Recognition Algorithm Based On Deep Learning

In recent years,with the development of deep learning theory,artificial intelligence technology has been evolving and has made many achievements.As a contactless biometric recognition technology,face recognition has been widely used in financial security,public security,education and other fields.However,in unrestrictive real scenes,illumination,expression and pose change usually lead to a sharp decline in face recognition accuracy.Among them,the pose change directly affects the symmetry and integrity of the face image,and becomes main bottleneck of face recognition technology.Aiming at the problem that the performance of face recognition algorithm for multi-pose face recognition drops obviously,this thesis studies and implements a pose robust face recognition algorithm based on deep learning.First of all,this thesis studies facial pose estimation task,which is to predict three face deflection angles: yaw,pitch and roll.In order to realize the prediction of face deflection angles,this thesis models this problem as a classification-regression problem,and improves the performance of facial pose estimation algorithm from two aspects: feature extraction and feature mapping.We firstly build a facial pose feature learning network based on attention mechanism,which can promote the network to learn more beneficial features for angle classification by the combination of channel and spatial attention module.Secondly,according to the influence of different classification granularity on angle prediction,we design a facial pose estimation algorithm based on multi-granularity classificationregression to improve the accuracy of face deflection angle classification by combining different granularity classification.Finally,we construct a multi-granularity classificationregression network based on attention mechanism Mg Cr ANet to realize the task of facial pose estimation.Extensive experiments demonstrate that the mean absolute error of Mg Cr ANet model is 5.5263 for predicting three face deflection angles on AFLW2000 dataset,which reduces 0.6287 compared with the mainstream facial pose estimation algorithm Hope Net.Next,in view of that there is more pose information in low layer feature of face recognition network,we propose an algorithm of feature fusion of high and low layer to realize pose robust face recognition.In the process of feature fusion,the semantic information is introduced into low layer feature to realize feature enhancement,so as to improve the efficiency of feature fusion.Considering the specific conditions of different facial pose change,we propose a deep feature fusion algorithm based on face deflection angle and take the coefficient from the non-linear mapping of face deflection angle yaw,which is predicted by the Mg Cr ANet model,as the fusion factor of the low layer features.Then the fusion weight of low layer feature is automatically adjusted according to the degree of facial pose change.In this way,the ability of face recognition network to represent faces with different poses is promoted,and thus the performance of multi-pose face recognition is improved.Finally,based on the algorithm in this thesis,we construct a deep feature fusion network DFFNet to realize pose robust face recognition.Experimental results show that the algorithm proposed in this thesis can improve the face verification accuracy.Compared with DCNN and Triplet Embedding algorithms,the proposed algorithm has improved the accuracy by 5.01% and 0.74% respectively on the cross-pose face dataset CFP_FP.The pose robust face recognition algorithm that is researched and implemented in this thesis can be widely used in video surveillance and other unrestrictive environments,such as nonsensing face recognition in outdoor public places.