Study On The Classroom Concentration Analysis Based On Facial Expression Recognition

To improve and maintain students' concentration in the classroom can effectively improve the learning effect and help students review and understand after class.Therefore,it is necessary to timely analyze and make statistics on students' performance in class.This thesis designs and implements a classroom concentration analysis system based on facial expression recognition.The classroom real images obtained through shooting can be used to achieve classroom attendance statistics through face detection and face recognition methods.Combining with the expression recognition method to analyze the students' concentration degree,and finally push the data visualization to students and parents,can effectively improve the teaching quality.However,due to irresistible factors such as the distance and large number of students in class,the students' face in the image is generally clear in the front row,while the face size of the students in the back row is small,and the occlusion problem also brings difficulties to face detection and subsequent focus analysis.To solve these problems,this thesis studies the face detection and expression recognition methods based on deep learning,introduces the principle and basic modules of convolutional neural network in detail,enumerates the problems and shortcomings in the current face detection and expression recognition methods,analyzes the reasons that affect the small face detection and necessity of model compression.A Multi-scale face detection method based on convolutional neural network and a face recognition method based on lightweight convolutional neural network are proposed.The main work of this thesis is as follows:1.In the shallow feature map of convolutional neural network,the small local perceptive field is helpful for localization but lacks global feature,while the high level feature has rich semantic information,but the low resolution of feature map is not conducive to the detection of small targets.Aiming at the face detection problem mentioned above,this thesis firstly extracts the output features of different layers in resnet101 basic network,uses deconvolution from top to bottom to enlarge the resolution of high-level feature map,and connects with its adjacent low-level feature side to get the fused feature map with rich semantic information and spatial details,which mainly solves the problem that the face detection method using only the highest-level feature map ignores the low-level spatial location information.A hybrid attention mechanism that combines channel attention and spatial attention is proposed to strengthen the role of important feature channels and feature map regions,while weakening the interference of irrelevant information.A comparative experiment on the WIDER FACE dataset verifies the effectiveness of the method presented in this thesis.2.The deeper the number of layers of the convolutional neural network makes the model parameters continuously increase,causing additional calculation costs and expenses,and even leading to overfitting and network degradation.In order to reduce the amount of redundant parameters,this thesis proposes a lightweight convolutional neural network to compress the existing neural network.This thesis uses 3 x 3 depthwise separable convolution and continuous asymmetric convolution to improve the Inception structure,and introduces a residual network to achieve skip connect.It effectively reduces the complexity of the connections between neurons,widens the network and obtains features of different scales.This thesis uses the center loss function and softmaxloss loss function to reduce the distance between classes.The SRGAN method based on generative adversarial network was used to super-resolution the FER2013 training set to improve the accuracy of the model.A comparative experiment was conducted in the FER2013 dataset,and the final confusion matrix and experimental results verified the improvement of this method in expression recognition.3.Based on the above face detection and expression recognition algorithms,this thesis designs a classroom concentration analysis system based on expression recognition to achieve real-time collection and processing of classroom images.The system uses face detection and face recognition methods to complete the recording of student attendance data,and combines facial expression recognition to analyze student's focus from the completeness of the face and facial expressions.The visualization of students' classroom data provides more data support and help for teachers,students and parents.