Student Behavior Monitoring in Experimental Exams: Anti-Exam Model and STEAM-Exam Dataset

This paper introduces the Anti-Exam model and the STEAM-Exam dataset to enhance behavior recognition during AI-assisted scoring of junior high school practical examinations, promoting exam integrity, fairness, and student's personal development. The Anti-Exam model combines MODNet portrait matting with the YOLOv9 object detection algorithm, leveraging images from a real-time AI-assisted scoring device equipped with three cameras to simulate a realistic examination environment. The STEAM-Exam dataset includes both MODNet-processed images and those with natural backgrounds, demonstrating varying processing speeds based on background complexity. Evaluation results reveal that the YOLOv9 model achieves the highest precision (98.7%), recall (98.7%), mAP50 (99.4%), and mAP50:95 (98.6%) among YOLOv7 and YOLOv8 models. Processing speeds differ between images with backgrounds (37.4ms per image for inference) and plain background images (35.1ms per image for inference). This research underscores the importance of AI in ensuring exam fairness and transparency, addressing challenges such as occlusion and cheating detection. The STEAM-Exam dataset is a critical resource for advancing AI-assisted educational assessments, aligning with China's policy to include practical examination results in entrance scores. Our findings suggest significant potential for AI to revolutionize educational assessments, ensuring fairness and integrity while supporting student development.