Research On Dense Face Recognition Algorithm For Classroom Scene

The flexible and changeable teaching methods in today’s universities have put forward stricter requirements on classroom attendance,and the traditional manual attendance method has gradually become weak.At present,the excellent performance of deep learning in the field of image detection provides a new idea for the development of attendance methods.However,the current face recognition algorithms based on deep learning mostly focus on the situation where there is no occlusion of the front face in restricted scenes,and their actual application in multi-person attendance scenes has poor detection effect.Therefore,this thesis carries out research on the algorithm of face recognition class attendance system,and combines face recognition with class attendance to provide theoretical support and technical support for the building of smart campus.Face detection in the classroom environment has problems such as scale change,pose change and occlusion.For this reason,this thesis optimizes the existing multi-task convolutional neural network MTCNN face detection algorithm.First of all,small target face oversampling is introduced in the data set processing to increase the proportion of small target face samples and enhance the robustness of the algorithm to small face detection.Then,a scale adaptive network is proposed to predict the potential face scale coefficient of classroom to reduce the number of image pyramid operations in face detection algorithm and improve detection efficiency.Finally,the body-assisted detection module was used to determine whether the face frame with low confidence should be continuously detected according to the score of the detection module,which improved the accuracy of face detection under occlusion.In the case of low pixels and occlusion,it is difficult to obtain the feature information of face recognition algorithm,resulting in low recognition accuracy.Therefore,this thesis proposes a face recognition algorithm in classroom environment based on FaceNet.First,face alignment based on facial boundary lines is introduced to improve the ability of face feature extraction across datasets.Then,in the face alignment operation,the key point offset regression is used to correct the face boundary to improve the prediction accuracy.Finally,neighborhood key point regression is used to explore the potential geometric relations between the facial features,so as to improve the accuracy of face recognition under occlusion.In order to solve the problem of failure of identity information matching in blind spot of the camera vision,an identity prediction algorithm is proposed in this thesis.Firstly,the classroom image is perspective transformed to obtain the student seat distribution table.Then the multiple distribution table is counted to calculate each student’s personal regional preference attribute and adjacent personnel preference attribute.Finally,the information of the person with the highest matching degree is calculated for the missed personnel through preference attribute and historical attendance rate,and the information is pushed intelligently.Based on the existing hardware equipment in the classroom,the face recognition class attendance system is built based on the improved MTCNN algorithm and FaceNet algorithm.By testing the detection and recognition algorithms on the fused data sets,the results show that the accuracy of the improved detection and recognition algorithms is improved by 8.53%and 10.18% respectively.Finally,the actual application test of the attendance system shows that the average recognition accuracy of the algorithm is 99.97% for 5-second surveillance video,and the detection speed is 2.53 frames per second,which meets the demand of attendance accuracy and real-time.