Research On Classroom Student Identity And Behavior Status Recognition Based On Deep Learning

With the deep integration of education,teaching and information technology,more and more schools have carried out digital upgrading in classroom teaching and student management.Among them,the identification of students’ identity and behavior state in classroom teaching can provide effective data support and data analysis for teaching quality assessment and improvement of teaching methods.Based on deep learning,this paper established the identification network of classroom student identity and behavior state,optimized and improved the accuracy and efficiency of identification,identified the status of students in class and student identity,and accurately matched information,so as to achieve the purpose of intelligent analysis and management of education and teaching.The main contents are as follows:First of all,this paper expounds the background of the integration of artificial intelligence and education teaching,analyzes the algorithm of artificial intelligence and class students identify monitoring research present situation,the basic principle of convolution is given network,network structure,focus on the most widely used several target detection algorithm network and face recognition algorithms,and analyzes the structure and performance of the algorithm.Secondly,aiming at the problem of dense seats and low face recognition accuracy of students in the back row of the classroom,a classroom face recognition algorithm based on attention mechanism,CBAM-Insightface,was proposed.This algorithm uses CBAM-Mobile Net backbone network,which is a convolutional network with attention mechanism.This NETWORK ADOPTS LIGHTWEIGHT DESIGN,WHICH REDUCES THE number OF parameters and COMPUTATION REQUIRED FOR model calculation,filters out the features that may lead to inconspicuous classification and false detection,increases the weight of important features,and improves the speed and accuracy of recognition.The LFW dataset was used to test it,and the average accuracy reached 98.75%,and the average monitoring speed reached 10 frames per second.Thirdly,in order to construct a behavior recognition method suitable for classroom environment,a class behavior recognition algorithm based on Shuffle-YOLO is proposed.On the basis of YOLOv4 network,the feature extraction network is combined with deep separation convolution and group convolution,and the Shuffle-YOLO network is designed by using the channel shuffle method.Aiming at the actual teaching scene in the classroom,the detection accuracy of the existing algorithm backbone network is improved and optimized,which makes the network more efficient.The average accuracy(m AP)of Shuffle-YOLO algorithm in Smart-Classroom student behavior dataset was up to 92.6%.The performance of the Shuffle-YOLO algorithm in recognition of each category on VOC 2007 dataset was improved compared with the original algorithm,and the average accuracy m AP of all categories was 79.2%.It is 3.4%higher than the original algorithmFourthly,in order to carry out targeted personality analysis on students’ classroom behaviors and effectively integrate the information identified by face recognition and behavior recognition algorithms,a student identity and action recognition method based on deep learning is proposed.In this algorithm,object detection and face recognition algorithms are cascaded and spatial attention module is applied to the network.In order to solve the problem of high false detection rate,the spatial attention module is used to identify the face position matching with the action,and the face is screened.The CBAM-Mobile Net backbone network is trained with arcface loss function to extract the face feature code,and the identity information is obtained by comparing with the data in the database.In the experiment,the efficiency and accuracy of different backbone networks in actual detection and recognition are compared.In the 2K resolution classroom scene image recognition experiment,the detection accuracy is85.2%,the average detection time per frame is 97 milliseconds,and the output frame per second(FPS)is 10.3.Finally,it summarizes the work content and innovation of the full text,and prospects the future development direction of the application of intelligent detection algorithm in deep learning in education and teaching research..