Research On The Application Of Face Recognition For Minors Based On Deep Learning

In recent years,with the rapid development of science and technology,The addiction of minors to online games has gradually become a hot issue of social concern.Minors are overly addicted to online games in the form of using adult identity information or renting other people’s accounts,which seriously affects their learning life and physical and mental health.At present,the traditional real-name authentication method has been unable to prevent minors from using false identity information to log into online games.Therefore,it is necessary to use the real person authentication method to detect and identify minors in online games,so as to prevent minors from indulging in online games.Based on artificial intelligence deep learning technology,this thesis studies the face recognition of minors in online game login.On the one hand,in order to improve the security of face recognition applications and prevent minors from using photos,replay videos and 3D masks to attack face recognition,it is very important to introduce liveness detection technology into face recognition systems.On the other hand,the face comparison technology can be used to compare the collected face with the input real face to identify whether the user is a minor.Therefore,by combining face comparison and liveness detection technology,it can effectively solve the risks and challenges in the real person authentication of minors.In the practical application scenarios of face recognition,there may be a large number of side faces and cross-pose faces in the collected face images,which seriously affects the accuracy of face comparison.In addition,there is a lack of high-frequency detail information for the collected face images with uneven illumination,which reduces the specific information between living and non-living,thus affecting the effect of living detection.In view of the above problems and challenges,this thesis proposes corresponding detection and recognition algorithms from two aspects of liveness detection and face comparison.The work of this thesis is as follows :(1)Aiming at the lack of high-frequency information in RGB images under uneven illumination,a face liveness detection algorithm based on homomorphic filtering image and two-stream network is proposed.Due to the complex environment in the process of face RGB image acquisition,it is easy to have the problem of uneven illumination,and most of the current face liveness detection is based on deep learning methods of RGB images,resulting in low performance of face liveness detection.Therefore,this thesis constructs a two-stream network combining RGB and homomorphic filtering enhanced images.The RGB image and homomorphic filtering image are input into the branches of the convolutional neural network and the Transformer model respectively.By means of attention feature fusion,the feature information extracted from the two images is fused to obtain the specific information of living and non-living bodies,and the classification algorithm of living and non-living bodies is realized.This study uses the public Replay-Attack dataset and the CASIA-FASD dataset for verification experiments,which proves that this algorithm is effective in solving the problem of performance degradation of face liveness detection under uneven illumination.(2)Aiming at the problem of low recognition accuracy caused by side faces and crosspose faces in open scenes,an improved face comparison algorithm based on Insight Face is proposed.On the basis of the original network structure,the SE attention mechanism is added and the self-attention mechanism is used to replace some ordinary convolution operations in the network,so as to construct the network structure based on the mixed attention mechanism.Among them,the embedded SE attention mechanism can increase the proportion of information in the key attention area of the face,reduce the invalid face feature information,and help to obtain feature information with stronger robustness and higher recognition accuracy.At the same time,due to the unique structure of the SE module,the convergence speed of face comparison is also greatly improved.Using the self-attention mechanism to replace the ordinary convolution operation can not only learn the correlation information of similar positions in the feature map,but also make the feature map have different degrees of importance in different spatial positions,improve the accuracy and generalization ability of face recognition,and cannot significantly increase the complexity of the network structure.By embedding the SE module based on the channel domain in the original network structure and replacing the ordinary convolution operation with the self-attention module based on the spatial domain,the key face information can be effectively extracted and the accuracy of face recognition can be improved.This study conducts verification experiments on the public LFW dataset,CFP-FP dataset,CPLFW dataset and VGG2-FP dataset.The experimental results show that the algorithm has high recognition accuracy and robustness,and has achieved competitive results in the field of face recognition.