| With the rapid development of science and technology,deep learning has been applied in more and more fields.As one of the biometric features that are easier to obtain and identify,facial features are widely used in personal identity authentication.The release of large-scale face data sets,the continuous iterative update of the network structure and loss function,and the improvement of hardware have greatly improved the performance and recognition effect of the face recognition system,thus promoting the popularization of face recognition technology.Facial recognition systems are being used in more and more fields,such as financial payment,online ticketing,smart home,etc.However,while face recognition technology brings convenience,there are also some problems related to privacy security and identity authentication.Therefore,verifying the authenticity of the face,that is,live detection,has become an important part of the face recognition system.ring.At the same time,with the continuous popularization of face recognition technology,more and more needs to be deployed to edge devices have emerged in the market.Therefore,while ensuring the accuracy of face recognition,improving the speed of the endto-end process of face recognition has become a a popular job.In view of the above background,based on the deep learning algorithm,this paper proposes a method to optimize the face detection algorithm through the research of face detection,face recognition,living body detection and other technologies and algorithms,and introduces the course learning into the face recognition algorithm,to improve the robustness of the algorithm,and combined with the live detection algorithm,an end-to-end face live detection and recognition system is realized.The main research contents and results of the paper are as follows:(1)A single-stage multi-scale face detection algorithm is designed,and combined with the idea of knowledge distillation,an algorithm to improve its efficiency,KDMF,is proposed.The core of KDMF is knowledge distillation,that is,through the method of knowledge distillation,the large model is used as the face detection model of the feature extraction network as a reference,and it acts as a teacher model in the training process of the small model.Two-layer optimization,training a small model with end-to-end accuracy close to the large model,while increasing the inference speed of the model by nearly 2 times.(2)The face recognition algorithm based on corner margin is implemented,and an improved face recognition algorithm named CurFace is proposed based on the idea of course learning.CurFace introduces the concept of curriculum learning,aiming at the simple samples in the early training stage and the difficult samples in the later training stage,by introducing the concept of curriculum,the model pays more attention to the simple samples in the early stage of training,and pays more attention to the difficult samples in the later stage of training.In order to learn the process from easy to difficult,the accuracy and robustness of the model are improved.(3)Aiming at printing attack and screen flipping attack,combined with rich data augmentation methods,a living body detection dataset Normal-Spoof is constructed.Based on the Normal-Spoof dataset,a silent face detection algorithm based on attention mechanism-SE-MultiLive is designed and implemented.Combined with the attention mechanism,SEMultiLive fully learns the difference between the feature maps of each stage on different channels,and learns the overall features through methods such as global pooling,so that the model can fully integrate the features of each dimension,and at the same time through the input Multi-scale images improve the generalization ability of the model.(4)Combining face detection,live detection,face alignment and face recognition algorithms,a general face live detection and recognition system is designed and implemented.The system supports users to build base library feature files,and supports users to output videos,pictures and The real-time video captured by the camera can realize face detection,liveness detection,face alignment and face recognition based on the user’s input,and finally return the corresponding results and display them to the user.Users can also extend the general face detection and recognition system designed in this paper to various scenarios based on their own needs. |
