Deep learning-based classroom behavior analysis provides new avenues for monitoring teaching quality in higher education. However, it faces challenges such as low detection accuracy, difficulty in recognizing small objects and handling occlusions as well as the difficulty in balancing real-time performance with accuracy.
This paper proposes an improved YOLOv11 method for classroom state recognition, achieving precise classification and behavior detection through the integration of AFGCAttention, SPDConv and RCSOSA modules. AFGCAttention optimizes feature weight allocation through an adaptive fine-grained channel attention mechanism, SPDConv enhances the processing capabilities for small objects and low-resolution images by converting spatial information into depth information and RCSOSA reduces channel redundancy while improving spatial object attention.
Experiments demonstrate that the YOLO-ASR model excels in precision, recall and mAP50. Compared to other You Only Look Once versions, it shows significantly improved detection accuracy and robustness in complex classroom environments, achieving an mAP50 of 93.8% and an mAP50-95 of 73.1%. Time-series analysis reveals dynamic changes in student behavior across teaching phases, including attention fluctuations, mobile phone use and signs of fatigue.
By analyzing student behavior across different classroom phases, patterns in mobile phone use and signs of fatigue were identified. These insights help teachers adjust their strategies, highlighting the method’s significance in monitoring teaching quality.
This study optimizes the YOLOv11 model for classroom behavior detection by integrating effective modules to enhance performance. It offers a novel approach for quantitatively assessing teaching effectiveness, providing data support for educational reform and advancing intelligent classroom management and innovative teaching models.
