Detecting cheating attempts in classroom videos by removing background

This study focuses on the analysis of video recordings captured by digital cameras during face-to-face exams in a classroom setting. It aims to detect abnormal behaviors and improve performance by removing background objects. The goal is to develop a new model to accurately identify actions during classroom exams. The study begins with the creation of the COPYNet dataset, which comprises a substantial collection of approximately 30,000 images. This dataset is used to develop and validate algorithms for detecting abnormal behaviors and to evaluate the performance improvements achieved after background object removal. The dataset is divided into five different groups based on behavior categories. To achieve high performance in solving the image classification problem, transfer learning is employed using the pre-trained ResNet model, which is hybridized separately with the Faster R-CNN and YOLOv5 algorithms. A deep neural network framework (COPYNet) is created to model normal behavior and generate an abnormal behavior score. The COPYNet framework demonstrates a precision of %90, a recall of %88, and an accuracy of %88. These figures represent a significant leap in anomaly detection compared to the existing literature. The presented results highlight the model's ability to accurately distinguish between different behavior classes, making it a valuable tool for detecting suspicious behaviors during face-to-face exams. Consequently, when the model detects abnormal activity, it aims to trigger an alert that can be sent to the proctor, serving as a decision support mechanism for exam invigilators. Based on the achieved success rates, our study shows better results in capturing suspicious movements during classroom exams compared to previous studies in the literature.