Research And Establishment Of Classroom Obedience System Based On Deep Learning

In recent years,students have low enthusiasm for learning,and it is common for them to be late for class,leave early,sleep,play mobile phones,and even absent from class.Scientifically improving the status of students in class is of great significance to comprehensively improve the quality of teaching.Based on deep learning,classroom adherence system provides classroom help during class and class through gesture recognition,in order to more accurately understand the state of students' classes,so as to promote the effect of teaching.In this paper,the deep learning algorithm is used to detect the target behavior of the students in the video in real time,and the trained deep learning model is deployed to the application website platform to complete the web-based classroom obedience system.Considering that video detection requires real-time performance,the existing target detection algorithm parameters are too large,and the computing power of the web terminal is limited,we adopt the SSD target detection algorithm with fast detection speed,combined with the lightweight model MobilenetV1 network,to achieve the purpose of real-time detection.At the same time,the offline video detection module is built for the system,and the high-precision deep learning model Faster R-CNN model is adopted.Finally,we uses the tensorflow deep learning framework to implement the algorithm for training,and deploys the trained model to the.NET-built classroom observing platform.This process includes using the classroom monitoring video of Zhejiang Agriculture and Forestry University to sample pictures,using the LabelImg tool to label the pictures,and performing data enhancement through cropping and brightness enhancement.The main work is as follows:A.In order to solve the the problem of excessive parameters of SSD based on deep learning,this paper considers MobilenentV1's deep separable convolution structure,which can convert the convolution operation into a deep convolution and a pointwise convolution,thereby reducing the amount of parameters.Therefore,this paper adopts the lightweight convolutional neural network model of MobilenentV1 composed of deep separable convolution,and through the MobilenetV1-SSD algorithm fusion migration learning method to reduce the requirement for the number of samples and the training time,so as to realize the real-time video detection of the system.B.Based on the Inception-ResNet-v2 network,we designed a structure that combines shallow features and obtained a Faster R-CNN detection model with multi-channel feature fusion,which improves the accuracy of detection and better realizes offline video detection.C.Design and implementation of the classroom obedience system.We have designed the system in detail,and built the platform through C# language and SQL server database.After experimental testing,the model based on the MobilenetV1-SSD algorithm training and the multi-channel feature fusion Faster R-CNN model are exported to the system website,which can complete the real-time detection of students' detecting the target behavior and off-line analysis on the web.Help teachers improve the effectiveness of teaching.Finally,to achieve the purpose of helping teachers improve teaching results.