FKD-YOLO: a Lightweight Student Classroom Behavior Recognition Algorithm

Recognizing student classroom behavior is pivotal for gaining insight into students’ learning status and enhancing teaching effectiveness. However, existing recognition algorithms are plagued by excessive parameters, high computational demands, and suboptimal accuracy. To tackle these challenges, this paper proposes a lightweight student classroom behavior recognition algorithm called FKD-YOLO. First, the PCBM module is designed to reduce the number of model parameters and computational complexity while improving feature extraction efficiency. Second, the Mish activation function is employed to enhance the accuracy and generalization of the model. Third, the MPDIoU loss function is introduced to improve bounding box convergence in complex classroom scenarios and enhance detection capability. Additionally, the Same-Stage Semantic Distillation is applied to mitigate semantic mismatches between teacher and student features during distillation, thereby enhancing the student model’s ability to detect subtle behavioral changes. Experimental results on two student classroom behavior datasets and the PASCAL VOC 2007+2012 dataset demonstrate that the algorithm achieves excellent detection performance and generalization while maintaining lightweight.