ViT-RF: A Two-Stage Transformer-Based Framework for Robust Classroom Learning Emotion Recognition

In contemporary education paradigms, enhancing teaching quality has become a focal point, particularly by capturing real-time student engagement and comprehension. Traditional classroom assessments based mainly on postclass surveys and teacher observations often fail to reflect the dynamic nature of learning and suffer from data limitations, which hinders effective instructional interventions. To address these challenges, this study proposes an innovative ViT-RF Classroom Learning Emotion Analysis (CLEA) framework. This framework integrates RetinaFace-based face detection with Vision Transformer (ViT) technology, leveraging deep feature extraction, multiscale detection, global feature capture, and self-attention mechanisms to perform detailed face detection, nuanced feature modeling, and precise state classification. A novel data collection strategy is also introduced to overcome the constraints of conventional methods. Experimental results indicate that in the mid-to-late training phases, training accuracy converges to nearly 100 percent, whereas validation accuracy stabilizes between 89 percent and 93 percent; the early-stopping mechanism ensures termination at the point of optimal generalization. Compared with classical baseline models, ViT-RF exhibits a marked enhancement in accuracy.it lays a technological foundation for the advancement of intelligent education and serves as a valuable resource for optimizing classroom instruction in the context of modern educational informatization.