In education, monitoring learning behaviors and exam cheating is crucial for understanding students' learning status and ensuring fair exams. It's a key factor in achieving high-quality management. Deep learning technology can improve the efficienc...
In education, monitoring learning behaviors and exam cheating is crucial for understanding students' learning status and ensuring fair exams. It's a key factor in achieving high-quality management. Deep learning technology can improve the efficiency and quality of educational supervision by applying object detection techniques to recognize student behavior. However, student learning behavior detection faces challenges in practice. First, detection scenarios in actual classrooms are complex. Challenges such as occlusion, low image resolution, large variations in target size, and ambiguous feature distinctions lead to decreased detection accuracy. Second, there's an increasing demand for lightweight models in environments with limited hardware resources. Third, in exam settings, detection models must operate in real-time and efficiently while identifying covert cheating behaviors involving subtle features and small targets, resulting in high miss rates and low detection accuracy. This study makes the following key innovative contributions. Firstly, this paper addresses the issue of low detection accuracy in complex classroom environments by proposing DyHead-YOLOv8n. The model integrates the DyHead Block into YOLOv8n's detection head, enhancing its scale, spatial, and task awareness. This strengthens the model’s core feature learning capability and improves detection accuracy. Secondly, to address hardware limitations in scenarios requiring lighter detection models, RepPSHead was proposed. This shared detection head employs a single module for classification and regression for each scale. It replaces the YOLOv8n's original detection head, reducing parameters and GFLOPs by 20%. The model achieves a faster speed while maintaining comparable accuracy. Thirdly, this study proposes a downsampling module called SpAttDown that uses channel attention to address the loss of small targets and subtle cheating behavior features during exam behavior detection. The module uses average and max pooling techniques to extract key spatial features and generates attention weights based on these features. This integrates channel attention with the downsampling process, minimizing spatial information loss and enhancing the extraction of subtle features. The SpAttDown-YOLOv11n model was designed to improve the learning of cheating behavior features. Experimental results on three exam datasets demonstrate that the proposed model achieves higher detection accuracy than existing state-ofthe-art models without increasing latency. In addition, to seek the optimal balance between detection accuracy and speed of the model in different scenarios and to evaluate the model’s generalization ability in diverse detection environments, the study integrated the lightweight detection head RepPSHead and the downsampling module SpAttDown into YOLOv8n and YOLOv11n, respectively. The proposed combinations were evaluated on both the learning behavior dataset and the exam behavior dataset. Experimental results show that in the learning behavior detection scenario, the SpAttDown-RepPSHeadYOLOv8n model demonstrates the best balance between processing speed and accuracy. In the exam behavior detection scenario, the SpAttDown-YOLOv11n model achieves the optimal balance between speed and accuracy. This contributes to the field of cross-scenario student behavior recognition.