DBRA-SCSA-YOLO: a multi-attention network for image-based student classroom behavior detection

Vision-based classroom behavior detection is challenged by dense student distributions, frequent occlusions, background clutter, and large-scale variations. To address these issues, this study proposes DBRA-SCSA-YOLO, an enhanced YOLOv8-based detector integrating a Dual-Branch Receptive-field Attention (DBRA) module and a Spatial-Channel Separable Attention (SCSA) mechanism. DBRA improves receptive-field adaptability and multi-scale feature representation, while SCSA performs staged spatial localization refinement and channel-wise semantic recalibration after feature fusion. Experiments were conducted using a three-category subset of SCB-Dataset5, with YOLOv8 variants and transformer-based detectors included as comparison baselines. The model was evaluated using standard object-detection metrics, including Precision, Recall, and mAP. Across five random seeds, DBRA-SCSA-YOLO achieved mean Precision, Recall, mAP@50, and mAP@50–95 values of 0.695, 0.685, 0.720, and 0.536, respectively. Compared with YOLOv8n, the proposed model improved mean mAP@50 and mAP@50–95 by 9.5 and 9.0% points. Ablation and architectural sensitivity analyses further supported the complementary roles of DBRA and SCSA under the reported experimental conditions. The framework operates on static classroom images and does not incorporate temporal information. Therefore, the findings should be interpreted as within-dataset evidence, and independent classroom datasets are needed to establish cross-environment generalization and practical deployment capability.