Classroom behavior recognition supports instructional assessment but faces challenges such as feature ambiguity due to frequent occlusions, multi-scale target coexistence, and complex background interference. We propose an enhanced EduFocus-YOLO based on YOLOv11: first, a Dynamic Multi-scale Fusion Pyramid (DMFP) is constructed to enable cross-layer feature interaction and preserve small-target information; second, a Task Align Dynamic Detection Head (TADDH) decouples classification and regression features and uses occlusion-aware attention to improve feature discrimination while reducing parameters; third, a WIoU loss with dynamic gradient adjustment is used to mitigate low-quality sample interference. Experiments on a self-built classroom behavior dataset (SCBA) show that the proposed method achieves 93.2% mAP50, 1.2% higher than the baseline, and meets real-time detection requirements.