Research And Implementation Of Face Recognition Algorithm Based On Deep Learning

Human face is the most obvious biological feature of human.Compared with other biological information(iris,fingerprint,etc.),human face is more convenient,safe,noncontact,and will not cause any harm to human body.Therefore,human face Features are widely used as personal identification.In actual application scenarios,due to the influence of the environment,lighting,posture and other factors,the trained face recognition model often has a relatively high generalization error.For the above situation,the face recognition algorithm needs to be improved and optimized for specific application scenarios to improve the robustness of the algorithm.In this paper,the first three face data sets are preprocessed,then the face recognition algorithm is improved in both the network structure and the loss function,and finally based on the data preprocessed face data set and improved descendants Face recognition algorithm,designed and implemented a face recognition system.Data preprocessing: Deep learning is a hungry learning process.When training model parameters,a large amount of data needs to be given to the neural network.If the training data is too small,the trained samples often appear to be over-fitting,which directly leads to the training.Models cannot be used,so data is often the most critical part of deep learning.In order to ensure the diversity of the data and improve the generalization ability of the model,the three face datasets used in the paper CASIAWebFace,LFW and the constructed face datasets were expanded by using mirroring and cropping methods.,While increasing the number of data samples,to ensure the diversity of the data set.Then the face data set is cleaned,and the wrong samples in the data set are removed,which ensures the quality of the face data set and improves the accuracy of the model.Face recognition algorithm improvement: In the improvement of the loss function,considering that the Softmax Loss + Center Loss loss function can effectively reduce the intra-class distance while expanding the interval between classes,the Focal Loss loss function can strengthen the model during training.The ability of sample mining can effectively improve the robustness of the model,thereby ensuring the accuracy of the model.Therefore,this paper designs a joint loss function based on Softmax Loss,Focal Loss and Center Loss to improve the performance of the original VGGNet in complex face recognition scenarios.In the network structure,FSC-VGG adds a BN layer to each convolutional layer and a DropOut regularization method to the fully connected layer,which improves the convergence speed and generalization ability of the model during the training process.The CASIA-WebFace dataset is used as the training set,and the LFW dataset is used as the test set to train and test the FSC-VGG.After testing,FSC-VGG has improved the accuracy of face recognition by 1.14% compared to the original VGGNet.Face Recognition System Design: Based on the instructor's "Reliable Big Data Driven Student Physical Health Monitoring and Management Platform Key Technology Research" actual project,a face recognition system was designed.The system software can be divided into two parts: face detection and face recognition from the algorithm.The face detection algorithm is used to detect the face from the input picture.After the face detection algorithm will get the specific position of the face in the picture.The face recognition algorithm will judge the identity of the face detected by the face detection algorithm.Finally,the test verifies that the designed face recognition system has good stability.