Deep Learning Based Student Engagement Detection: Comparison of Model Architectures and Computational Time Efficiency

Detecting student engagement during online learning is a significant challenge in educational technology, especially with varying lighting conditions and devices. This study develops a deep learning-based system using a custom dataset gathered from video recordings of Zoom lecture sessions. The recordings were processed at an extraction rate of 25 fps and anonymised to maintain data privacy. The videos collected in the Zoom gallery view were processed into individual clips before experts performed annotation into two classes: engaged and disengaged. A baseline model and three pre-trained models (MobileNetV2, ResNet50, EfficientNetB0) are compared to measure the effectiveness of the data augmentation approach in improving performance. For real-time implementation, a web-based user interface was developed using Streamlit, which allows users to upload videos or use a live webcam to detect student engagement. In addition, this study also evaluates the computational time efficiency for model training on three computing platforms: CPU, GPU, and DGX. The results show that the baseline model without augmentation exhibits high accuracy (98%) but raises concerns about overfitting. Data augmentation is applied to address this, and the pre-trained models are used for retraining. MobileNetV2 with augmentation achieves the highest accuracy of 92.96%, followed by EfficientNetB0 and ResNet50, with accuracies of 81.8% each. In the computational efficiency analysis, GPU consistently delivers faster training time than CPU for all tested models. Despite having a significant advantage in training time for simple models, the DGX platform performs similarly to GPUs on more complex models.