This research is dedicated to the analysis and detection of anomalies within images captured by digital cameras during face-to-face examinations. The focal point is the development of a novel model designed to identify exam activities with a high degree of precision. Central to this work is the creation of the COPYNet Dataset, a substantial collection of approximately 30,000 images. This dataset is pivotal for the development, verification, and performance evaluation of anomaly detection algorithms. It is meticulously segmented into five distinct groups, each corresponding to a particular behavioral category crucial for anomaly detection. To achieve superior performance in image classification, the transfer learning method is independently hybridized with the Faster R-CNN and YOLOv5 algorithms using the pretrained ResNet model. This leads to the creation of a deep neural network framework, COPYNet, designed to generate an anomaly score by modeling typical behavior. Significantly, the COPYNet framework demonstrates remarkable precision (0.90), recall (0.88), and accuracy (0.88), marking a considerable advancement in anomaly detection compared to existing literature. The results underscore the model's capability to accurately categorize diverse activity classes, making it a promising instrument for addressing the challenge of identifying suspicious behaviors during face-to-face exams. Consequently, when the model identifies an unusual activity, it triggers an alert to be dispatched to the proctor, serving as a decision support mechanism for exam invigilators. Given the obtained success rates, our study proposes a promising solution for detecting suspicious behavior during face-to-face exams, surpassing previous studies in the field.
