DLW-YOLO: Improved YOLO for Student Behaviour Recognition

The intricate backgrounds and multi-scale feature distributions in classroom behavior images present a significant challenge for accurately detecting student behavior. To address this challenge, we propose a DCNv2-LSKA-WloU-YOLO (DLW-YOLO) network model based on the YOLOv8 network model. The model incorporates Deformable ConvNets v2 module(DCNv2) into the C2f module and introduces a flexible C2f-DCN sampling module, enhancing the network's feature extraction capability for irregular targets by better characterizing images of various sizes. Additionally, by leveraging the concept of Large Separable Kernel Attention (LSKA), the model introduces an adaptive capability with long-range dependency and an SPPF-LSKA module to effectively reduce background interference in behavior detection tasks. Optimizing the bounding box loss function using Wise-IoU (WloU) speeds up the network's convergence. Testing was conducted using a real dataset of student behavior in class-room settings. The experimental results demonstrated a 5.1% increase in the mAP@0.5 of the improved network model compared to the YOLOv8n model. Finally, comparing it with other classical object detection algorithms confirmed that the model offers advantages in terms of accuracy. The detection results confirmed that the DLW-YOLO network model accurately detects student behavior across various levels of occlusion in complex classroom environments, providing a theoretical foundation for the implementation of smart classrooms.