ARCHIVES

With a rapid expansion of online education, Student engagement monitoring has emerged as a critical challenge in intelligent learning environments. This study investigates the effectiveness of a lightweight deep learning architecture, MobileNetV2, for real-time classification of student engagement states based on facial expressions. The main research issue investigates whether high classification accuracy across several engagement categories can be attained by a computationally efficient model without losing deployment capability on devices with limited resources.To train and evaluate the model, a multi-class dataset comprising six behavioural states - bored, confused, drowsy, frustrated, engaged, and looking away were used. The proposed approach employs transfer learning with fine-tuning of MobileNetV2, coupled with data preprocessing and augmentation strategies to enhance generalisation. Performance was assessed using standard evaluation metrics, including accuracy, confusion matrix, and ROC-AUC analysis.Empirical results demonstrate greater performance, achieving validation accuracy exceeding 94% with minimal overfitting, as indicated by closely aligned training and validation curves. The model attained near-perfect ROC-AUC scores (0.99–1.00) across all classes. The confusion matrix further confirms high classification precision, with only minor misclassifications observed between visually similar states such as boredom and drowsiness.The findings emphasize the significance of lightweight CNN for scalable educational applications. From a critical scholarly perception, this work offers a meaningful contribution by balancing accuracy and efficiency, making it suitable for real-time deployment in mobile and edge-based learning systems. Future suggestions include integration into adaptive learning platforms and expansion toward multimodal engagement detection to enhance pedagogical responsiveness.