Research On Unconstrained Face Recognition Based On Deep Learning

With the rapid development of Internet technology,face recognition technology has been applied in every aspect of people's life.The emergence of face recognition application brings a lot of convenience,which correspondingly promotes the in-depth research of on face recognition technology.Nowadays,face recognition technology has been transformed from the traditional recognition methods such as template matching and geometric features to the face recognition method based on artificial neural network,and the recognition effect under the experimental environment has even reached a particularly high level.However,in real life based on non-ideal conditions,the effectiveness of face recognition is affected by many factors,such as lighting,occlusion and multiple poses.Among these interference factors,multi-pose problem is still the difficulty and focus of research.In real scenes,face Angle deflection will lead to the loss of feature information of face,resulting in low recognition effect.Therefore,how to improve the recognition rate in the case of multi-attitude Angle deflection is the focus of the research.In this paper,the main factors affecting face recognition under uncoordinated conditions are studied,and the multi-pose face recognition based on deep learning is mainly studied.First,the paper introduces the main flow of face recognition,the main methods of multi-pose recognition and the common multi-pose face data sets.Due to the Equivariance of feature can map profile feature to frontal feature and based on the deep learning of a neural network face recognition can be used to extract abundant feature information,this paper introduces the error brought by identifying the profile features in deep space to modify the transformation between profile and frontal based on the above advantages.Based on deep learning,this paper USES open source frameworks like Pytorch and Tensor Flow to build the recognition model,USES open data set VGGFace2 for training network,and then USES open data set CFP,IJB-A and LFW for testing.Experiments show that the multi-attitude recognition model presented in this paper has a better recognition effect for challenging data sets,and can achieve an accuracy rate of 96.83% in face recognition,which is better than the general multi-attitude recognition model.Given the current neural network layer deepening,more and more updated parameters have been proposed for neural network training.This article compares three neural networks,adopting two classic Res Net18 and Res Net50 and lightweight neural network model proposed by Google in 2019 Mobile Net V3 for training.From the training results,Res Net50 and Mobile Net V3 effect is good,but the deep Mobile Net V3 adopts the thought of separation of convolution,Compared with Res Net18,the network calculation parameters are less,and the training and parameter iteration update will be faster.For final identification tests,Mobil Net V3 had 92.9 percent better accuracy than Res Net50's 92.5 percent,which has more than three times the reduction in network parameters.For attitude Angle estimation in multi-pose,this paper uses 68 key points based on Dlib model to detect and obtain face features,and uses the nonlinear least square method to to obtain the corresponding rotation vector with solve Pn P function in Open CV,so as to obtain the deflection Angle of the face data set in three-dimensional space.Since multi-pose faces are mainly affected by yaw Angle,this paper uses EPn P algorithm to fit the attitude estimation model,which realizes the attitude reconstruction through nonlinear optimization.The experimental results show that the design of the model method also has A better performance of recognition under extreme attitude,the average recognition result on the LFW data set can reach 97%,and even around 95% on the more complex data set IJB-A.It is of certain significance to improve the recognition effect by means of attitude Angle deflection in deep space.