| With the continuous development of education intelligence,the combination of online and offline teaching mode has become the mainstream trend of modern education,but online education cannot grasp the concentration of each student in time,which will seriously affect the quality of students’ listening,so it is especially important to design an online classroom student concentration analysis model to monitor the concentration of each student in real time to improve the quality of teaching.In this paper,a concentration analysis model is developed to address the problem that the online classroom cannot monitor each student’s concentration in real time,combining students’ face position information,head posture information and facial expression information:(1)An improved Retinaface face detection algorithm is proposed to address the problem of low accuracy of face detection due to indoor environmental factors.The algorithm introduces the PfAAM module after the feature extraction network MobileNetV1-0.25,and additionally improves the more complex 3D dense point regression loss in the loss function.Experiments are conducted on the Wired Face dataset,and the results show that the improved algorithm improves the recognition accuracy on the Hard subset by 7.72%compared with that before the improvement.(2)The PnP algorithm is used to calculate the student head deflection angle.14 representative facial key points are selected,and the face key points in 2D images are matched with those in the 3D face model to calculate the rotation matrix and translation matrix of head deflection,and then the student head deflection Eulerian angle is obtained.Experimental validation is performed in a real online classroom scenario,and the results show that the algorithm can accurately identify the head deflection angle in each direction.(3)A multi-scale feature fusion algorithm DS-EfficientNet is proposed to address the misclassification problem caused by the incomplete extraction of facial features in the current expression classification task,which can extract both shallow detail features and deep global features of the face to make the network with stronger feature extraction capability,in addition to replacing the fully connected layer in the original network with global average pooling and adding batch normalization to reduce the number of model parameters and avoid overfitting the network.The final experiments were conducted on Fer2013 and CK+datasets,and the results showed that the recognition accuracy of DS-EfficientNet network reached 73.47%and 98.84%on the two datasets,respectively,which verified the effectiveness of the improved algorihm.(4)The concentration evaluation model was established by combining students’ facial behavior states using the fuzzy comprehensive evaluation method,and the concentration score of online classroom students was calculated.Students of different age groups were selected for the experiment,and the concentration change curves of students were plotted,and the system scores were compared with the teachers’ rating results.In order to verify the correlation between them,Pearson correlation coefficient was used as the evaluation index,and the correlation coefficient was 0.96,which proved that the concentration evaluation model established in this paper had a high correlation with the teacher’s evaluation results and verified the reliability of the algorithmic model in this paper. |
