A Visual Mamba-Based Architecture for Student Engagement Recognition in Distance Learning Environments

With the rapid advancement of Artificial Intelligence (AI) and Computer Vision (CV), student engagement recognition has become a prominent research focus in the domain of intelligent education. However, most existing models for engagement recognition face difficulties in effectively capturing long-range dependency (LRD) features. Moreover, these models are often hindered by class imbalance issues in the training datasets, which negatively affect recognition performance. To address these challenges, this study proposes a novel VMamba-based model that incorporates a Distribution Alignment Optimization (DAO) loss function for enhanced student engagement recognition. The proposed approach not only effectively captures LRD features from facial images but also alleviates the detrimental impact of class imbalance in the dataset. According to the experimental results, conducted on the imbalanced SVD-OL dataset demonstrate that the proposed model achieves Recall, Precision, and F1-score values of 82.74%, 85.50%, and 84.00%, respectively. These results indicate that the proposed method outperforms existing state-of-the-art approaches in student engagement recognition.