Technology And Application Research Of High Precision And Lightweight Face Recognition Based On ZYNQ

Face recognition has been one of the important technical means of biometric information identification,and is now widely used in daily life,finance,and security.In recent years,face recognition technology has achieved great success with the development of deep learning.However,the large number of parameters and computation generated by increasingly complex convolutional neural networks make it difficult to deploy these algorithms in resourceconstrained embedded or mobile devices.To address this problem,this paper designs and implements a high-precision lightweight face recognition algorithm and a ZYNQ-based face recognition system.The main work of this paper is as follows.(1)A high-precision and lightweight DGFace Net face recognition model is developed to solve the problem of difficulty in deploying a high-precision and complex face recognition model for embedded and other resource-constrained devices.First,we use the idea of "linear operation" to generate similar features instead of a convolutional operation,and design an inverted pyramid strategy to dynamically adjust the number of similar features;second,based on this,an efficient dynamic ghost backbone network is built;finally,DGFace Net network is constructed by stacking dynamic ghost backbone networks.(2)The class-margin-linear softmax(CML-softmax)loss function is designed to address the problems of non-convergence of loss values,unstable training,or the need to pre-train the model when using margin-based loss functions for supervised training of lightweight models.CML-softmax designs a monadic quadratic function that makes the logit curve almost linear in[0,π],which makes the performance and convergence of face recognition in low-dimensional output better and effectively solves the above problem.The CML-softmax loss function has demonstrated excellent performance in many validation datasets and massively popular benchmark tests.(3)A joint margin Unified Face loss function is intended to explore a more efficient marginbased loss function.Unified Face,unlike other approaches,believes that additive and multiplicative margins ought to be employed together,which means that additive margins should be introduced in the target angle,and multiplicative margins should be added in the nontarget angle.Unified Face achieves excellent performance in both regular and lightweight model training,according to experimental results.(4)Implementing a ZYNQ-based face recognition system using the DGFace Net lightweight face recognition model.The face recognition algorithm is deployed on a real embedded platform,while the inference time of the model is verified.Speed tests on embedded devices show that the actual inference time of DGFace Net is 11.08 times,8.57 times,2.75 times,and 2.82 times faster than Res Net-50,Efficient Net,Mobile Net V2,and Mobile Face Net,respectively.DGFace Net can guarantee the model performance while DGFace Net can significantly improve the operation efficiency of the model in resource-constrained embedded devices,and is suitable for deployment in resource-constrained embedded devices,etc.