Precision Enhanced and Lightweight Accelerated Models for Student Learning Behavior Detection

Student learning behavior detection has gained much attention since it can efficiently understand their learning status and assist teaching. However, in practical application, complex environment leads to low detection precision, and detection equipment with limited computing resources requires a more lightweight model. To address these challenges, we propose two novel models based on YOLOv8n. Firstly, we propose the DyHead-YOLOv8n, which integrates a Dynamic Head Block (DyHead Block) into YOLOv8n's detection heads to enhance the model's detection precision. Secondly, we introduce the RepPSHead-YOLOv8n, which employs a RepConv parameter-sharing head to lighten YOLOv8n's detection heads. Thirdly, to compensate for the precision drop caused by the parameter-sharing head, we use Slide Loss instead of BCE loss. Experimental results show that the DyHead-YOLOv8n outperforms YOLOv8n in detection precision across three datasets. The RepPSHead-YOLOv8n reduces the parameter numbers and GFLOPs by 21.6% and 19.8% respectively, compared to YOLOv8n, while maintaining a faster efficiency and competitive detection precision.