The Student Classroom Behavior Recognition Based on Feature Enhancement

Object detection and pose recognition technologies are advancing rapidly. However, in the task of recognizing student behaviors in class, there still exist issues such as imprecise recognition due to mutual occlusions among students and high computational costs resulting from complex models designed to improve accuracy. To address these problems and enhance recognition accuracy, we proposes a feature enhancement-based method for student classroom behavior recognition. We design a feature extraction module that incorporates the Squeeze-and-Excitation Networks (SENet) channel attention to enhance the model's ability to extract channel features. Then, introduce Switchable Atrous Convolution (SAC) in the output head module to further strengthen feature representation. Finally, adopt Weighted Intersection over Union version 3 (WIoU v3) based on a dynamic non-monotonic focusing mechanism as the loss function to better handle errors in bounding box center positions. Experimental results show that, on the public dataset SCBDataset3-S and a custom student classroom behavior dataset, the proposed model outperforms several YOLO series models while maintaining a relatively low model complexity.