Design And Implementation Of Lightweight Face Recognition System Based On MobileFaceNet

With the rapid development of artificial intelligence,face recognition has attracted much attention in the field of computer vision.At the same time,face recognition has also performed well in many fields such as intelligent security,public applications and financial security.In recent years,the rapid development and popularity of mobile devices,face recognition technology has been widely used on mobile devices.Thanks to the rapid development of face recognition technology,many researchers regard it as a research direction and have obtained excellent research results.With the further development of technology,face recognition will have greater application prospects and development space in the future.With the increase in the scale and complexity of the problem,in order to obtain higher recognition accuracy,the depth and volume of the network model have also increased exponentially.However,due to the constraints of mobile devices in many aspects such as computing power and storage space,it is difficult to transplant the network model with complex calculation and huge volume to mobile devices.Therefore,the main purpose of this paper is to design and implement a lightweight face recognition model with high recognition accuracy,small model size and fast running speed.The research contents and achievements of this paper are as follows:After a thorough research and understanding of deep learning theory and current excellent lightweight face recognition models,this paper proposes SE-MobileFaceNet based on the improvement of existing models.On the basis of inheriting the advantages of the original model,an SE module is embedded in the network,and the importance of each feature channel is automatically obtained through learning,improving useful channel features and suppressing unimportant channel features,thereby improving the face Recognition accuracy.In the training phase of the model,Softmax loss and InsightFace loss were used to train the network to optimize the network,and finally a high-precision lightweight face recognition model was obtained.The test result on the LFW dataset was 99.67%.The experimental results show that our proposed face recognition model can obtain higher accuracy than most current algorithms,which proves the feasibility of the proposed method.In order to make the parameters of the lightweight face recognition model smaller and the network inference speed faster,we performed network pruning on the model.This paper used the greedy criterion-based pruning method to evaluate and pruned the least important parameters in the network,and then fine-tuned the network to reduce the accuracy loss caused by network pruning.The pruning process used a pruning standard based on the Taylor expansion.We performed pruning experiments on both the face detection network MTCNN and the face recognition network SEMobileFaceNet.The optimal pruning result was obtained through the trade-off between the pruning rate and the accuracy rate.Experiments show that network pruning can effectively reduce the model size to 38%,and the network inference speed is increased by 1.8 times.Based on the above research results,this paper designs and implements a lightweight face recognition system suitable for mobile devices.The design and implementation of the face recognition system were introduced in detail,and the face recognition system application was tested for function and performance on the Android platform.The system has perfect functions and smooth running,and can complete the entire process from the collection of face information to the output of face recognition results.In the performance test,the accuracy rate of 1: 1 face recognition on the 10 types of face data collected in this paper is 95.37%,and the average recognition process takes 359 ms.The accuracy rate of 1: N face recognition is 96.49%,and the average recognition process takes 196 ms.The size of the face recognition application is 7.76 MB,which has a high recognition rate,low running time,small volume,and broad application prospects.