An Efficient Model For Student Behavior Recognition in Classroom

AI and big data analysis for student classroom behavior recognition can be used as auxiliary means to improve teaching quality. Recognition in classroom scenarios suffers from issues such as tiny targets and complex environmental interference. To tackle these problems, an efficient model based on YOLOv4-tiny is proposed in this paper. Specifically, we design a new module named ResBlock-S to reduce the floating point operations (FLOPs) of the model to improve the speed. Then, the introduction of the Convolutional Block Attention Module (CBAM) mechanism to obtain extra local information of images during the training process, which can ensure the recognition accuracy. As most available public datasets are not applicable to this work, we construct a classroom behavior dataset. Experiments were conducted on the public dataset and our self-built dataset to verify the performance of our model in general scenarios and classroom scenarios, respectively. Compared with YOLOv4-tiny and other lightweight CNN models such as MobileNetv2, MobileNetv3 and ShuffleNetv2, the mean Average Precision (mAP) of our approach on the self-built dataset is higher and up to 89.9%. Additionally, the detection speed of our approach is faster than the aforementioned methods, which is up to 167 fps.