Smart CCTV Exam Monitoring to Detect and Alert Suspicious Activities Using Deep Transfer Learning

Examinations are essential in assessing learning outcomes and evaluating students’ academic performance. Academic dishonesty, including cheating during exams, violates academic integrity and is a critical issue in educational settings. This study presents an approach to detecting exam cheating by developing an intelligent CCTV system for exam monitoring, capable of issuing alerts for suspicious actions indicative of cheating. The system can detect cheating behaviors, including referencing books/notes/papers, talking to others in the room, accessing the internet, seeking answers from friends via phone calls, and using mobile phones or other electronic devices. The dataset used in this experiment is imbalanced and consists of multiple classes. Deep transfer learning is applied as an advanced feature extraction technique, leveraging prior learning rather than starting from scratch. This study integrates ResNet50V2 with three deep learning architectures, Gate Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN). These integrated models were then compared with other transfer learning models, including VGG16, VGG19, ResNet50V2, Inception, and EfficientNetB7. The experimental results reveal that the ResNet50V2 setup yields the best overall performance, with the GRU and LSTM variants attaining the highest F1 scores and accuracies. Both scores and accuracies of ResNet50V2-GRU and ResNet50V2-LSTM models were exactly 1.0 contrasted with the CNN variant which obtained about 0.4151 for F1 score and 0.7232 for accuracy. Compared to other models such as VGG16-LSTM, VGG16-GRU, and VGG16-CNN, the F1 scores achieved only in that order as low as 0.141, 0.3255, and 0.141. Also, the accuracies of VGG16 models are 0.5446, 0.7679, and 0.5446. The examination of model training and testing duration indicates that ResNet50V2-GRU model performs faster than ResNet50V2-LSTM model. Overall, the ResNet50V2-GRU model exceed all the tested model in the paper in term of the classification effectiveness, model training time, and model testing time. Additionally, the study found that deep transfer learning can effectively handle imbalanced datasets without requiring data resampling during preprocessing.