| In recent years,artificial intelligence technology has changed people’s lives dramatically.Intelligence plays a high-quality auxiliary role in all walks of life.Universities have also started to use the monitoring video intelligent analysis system to help teachers know students’ classroom status.However,there are a large number of students’ facial images in the classroom videos,once they are used for other purposes,the identity information of many students will not be protected well.Therefore,while using classroom video analysis to help improve the quality of education,we should not ignore the privacy protection measures for students.This paper studies the classroom student head detection network,head state classification network and face privacy protection algorithm.On the basis of these studies,an intelligent statistics system for classroom student raising rate is designed and implemented,which combined with privacy protection.This system can help teachers optimize teaching,but also protect students’ personal privacy.The main research work of this paper is as follows:(1)To improve the low students’ heads detection rate caused by the low head resolution of in the classroom videos,this paper presents a YOLOv5＿improved students’ head detection algorithm for classroom video,called YOLO-Head: By reducing the number of down-sampling layers in the original network,the model structure of the original network is adjusted so that more feature information is retained when the input image passes through the main network into the detection network;At the same time,Ghost Module is introduced to replace the convolution layer in the original network to compress the network model parameters and to speed up network detection.Compared with YOLO v5,the accuracy of YOLO-Head on SCU-HEAD Part A dataset and self-made SCU-HEAD dataset is improved by 1.3%and 0.9%,respectively,reaching 94.3% and 98.2%,and the detection speed in GPU environment is increased from 62 FPS and 59 FPS to 76 FPS and 72 FPS,respectively.(2)In view of the poor head state classification results of existing network,this paper puts forward an improved network named Efficient Net-Class based on Efficient Net.By simplifying the network and introducing a selective kernel module,the feature information is enhanced and the reliability of feature extraction is enhanced to cope with the low resolution of the input image.In addition,the random erase method is introduced during the network training phase to enhance the overall similarity of the head area image of students,which will improve the training effect of the network and the classification ability of the network to enhance the performance of the network.Efficient Net-Class’ s head state classification accuracy on the self-made dataset is 94.8%,which is 1.0% higher than the original network.(3)In order to protect students’ facial privacy while preserving facial non-private information for future research,this paper combines Deep Privacy with Fuzzy to propose a DP-Blur algorithm: For the front-row students in the classrooms,their facial images,which are with high resolutions,are replaced by passing through Deep Privacy face-changing algorithm;and for the students in the back rows of the classrooms,their facial images with low resolutions are blured using a fuzzification method.With the combination of these two methods,all students’ private information can be effectively protected.(4)This paper designs and implements an intelligent statistics system for students’ head raising rate in college classes which combines with privacy protection.The whole system consists of a PC video analysis platform,a database server and a mobile display platform.The system has such functions as students’ head raising rate statistics,student video privacy protection,class history data viewing and so on.The system can assist teachers conveniently and effectively to make a comprehensive analysis and reflection on students’ learning situation. |
PDF Full Text Request