Classroom Behavior Detection and Evaluation Method Based on Deep Learning Algorithms

With the rapid development of intelligent teaching technologies, automatic classroom behavior detection can provide objective visual evidence for classroom participation analysis and teaching feedback. However, real classroom scenes commonly involve dense student distributions, large scale variations between front and rear rows, frequent occlusion, and subtle visual differences between similar behaviors such as reading and writing. To address these challenges, this paper proposes SPD-Net, a classroom behavior detection network based on YOLO11m. The proposed method introduces three complementary improvements to the baseline: detail-enhanced upsampling for small-scale behavior regions, multi-scale and fine-grained feature enhancement for visually similar behaviors, and feature-conditioned spatial-prior attention for suppressing irrelevant background responses. Experiments are conducted on a three-class subset constructed from SCB-Dataset5, including reading, writing, and hand-raising. SPD-Net achieves a mean average precision (mAP) of 71.6% at mAP@0.5 and 54.7% at mAP@0.5:0.95, outperforming the YOLO11m baseline by 2.3 and 1.8 percentage points, respectively. In addition, the detected behavior instances are summarized into interpretable behavior-distribution information, providing support for classroom review and teaching-feedback analysis.