Research And Implementation Of Face Recognition Method In Classroom Enrollment System

With the concept of deep learning being put forward,face recognition has made outstanding breakthroughs and has been widely applied.Applying face recognition to classroom enrollment system can solve the shortcomings of traditional enrollment and has important practical value.This thesis focuses on the problem of face recognition in the enrollment system.The difficulty is that the running environment is mobile device with limited hardware resources and how to solve the influence of illumination and other factors on the collected face image.To solve the above problem,the main research work in this thesis is as follows:1.Aiming at the quality problems such as blurred face and bad posture in mobile face collection process,this thesis proposes a fast face image quality assessment algorithm based on deep learning.At present,most of the proposed image quality assessment algorithms are based on statistical learning methods.Due to the particularity of the face image,there are large errors between the results of current assessment algorithm and the actual situation of the face image,and it is difficult to meet the real-time requirements.In this thesis,image quality classifier based on lightweight network MobileNet is constructed through migration learning after the images in the dataset were preprocessed.The experimental results show that the accuracy of 97.4%is achieved when the face image quality is divided into two categories of qualified and unqualified,and the processing time of each frame is about 0.23s on the mobile phone.When we divide the face images into five categories,such as blurred,dark.sheltered,bad posture and qualified,we can get the accuracy of 97.2%,and the processing time of each frame can reach 0.25s,which meets the requirement of real-time.2.Aiming at the problem of high hardware requirements for face recognition on mobile device,this thesis proposes an improved face recognition algorithm based on FaceNet.Many deep learning models proposed so far have achieved very good recognition results on face recognition,such as FaceNet.However,these algorithms mostly run on the server with strong computing power,when run on a mobile device with limited computing power,the desired effect cannot be achieved.In this thesis,lightweight network SqueezeNet and MobileNet are used as the basic network of FaceNet,and center loss and softmax loss dual loss are used for training.These lightweight networks use point convolution,depthwise separatable convolution to speed up the computation,which greatly reduce the amount of calculation under the premise that the accuracy is almost not affected.Experimental results show that the SqueezeNet-based model can achieve an accuracy of 98.5%and the MobileNet-based model can achieve 96.4%accuracy.3.Using the above research results,the classroom enrollment system based on face recognition is designed and implemented to verify the application of this algorithm.