Research On Cross-race Face Recognition Based On Edge-cloud Collaboration

As an important personal identification method,face recognition is widely used in daily life such as shopping,payment,security check,travel and attendance.And its application prospects are very broad.Because face recognition systems require large computing power and storage space,the face images to be recognized are often transmitted to the cloud platform through the network.However,problems such as congestion,failure and delay may occur during network transmission.These problems cause the face recognition system to have poor real-time performance.It is difficult to meet the needs of practical applications and is poor to have user experience.This paper studies the face recognition of different races,and proposes a cross-race face recognition method based on edge-cloud collaboration.This method combines the processing power of cloud computing and the real-time nature of edge computing,so that the face recognition system is not constrained by the network state.And its application is more extensive and the user experience is better.The main research contents are as follows:(1)Research on cross-race face recognition method based on cloud computing.Aiming at the problem that the uneven distribution of races in the face dataset leads to the bias of the recognition accuracy of different races.On the basis of the Race Net single-race face recognition method,a race-based face recognition method based on Race Net-MDM is proposed to improve the accuracy of the network in the target race.First,the maximum mean discrepancy(MMD)is used to minimize the difference between the source race and the target race,and initially improve the robustness of the network in the target race.Then use the DBSCAN clustering method to generate some pseudo-labels according to the target face features,and the pseudo-label sample training enhances the network’s ability to discriminate the target race.Considering that some target images do not generate pseudo-labels,a Mutual Information(MI)loss is proposed to make full use of all target images to further improve the performance of the network.Finally,the effectiveness of the method in this paper is verified by comparative experiments on the test dataset.(2)Research on lightweight cross-race face recognition method based on edge computing.Aiming at the problem that the cloud-based cross-ethnic face recognition model is deployed on edge devices with limited computing resources,the reasoning time is too long.First,a lightweight method of Race Net single-race face recognition model based on knowledge distillation and pruning is proposed,and then an MDM method is proposed to transfer the lightweight model across races to improve its accuracy in the target race.The model lightweight method uses the Mobile Net V2 network to perform knowledge distillation on Race Net,and uses the local L2 distance distillation loss function to supervise the learning process.In order to further reduce the redundancy in the distillation model,the scale factor of the Batch Normalization(BN)layer is trained with regularization to identify unimportant channels for clipping and fine-tuning.The feasibility of this method is proved by comparative experiments on edge devices.(3)Design and implementation of a cross-race face recognition prototype system based on edge-cloud collaboration.According to the requirements of independent deployment and collaborative recognition of cloud and edge face recognition models,the edge-cloud collaborative system is designed,which can be divided into cloud computing platforms and edge computing platforms.Firstly,the overall framework of the edge-cloud collaboration system is studied,and then the specific modules of the cloud and edge platforms are implemented,and the scheduling strategy of face recognition tasks is studied.Finally,face recognition tests are carried out in different network states to verify the effectiveness of the edge-cloud collaboration system.