MTFNet: Multi-Scale Transformer Framework for Robust Emotion Monitoring in Group Learning Settings

Identifying students' learning states in authentic classroom settings is a prominent topic in educational technology. This study addresses the challenges posed by complex facial environments and the scarcity of data in such settings. we propose a Multi-Scale Transformer with Frame Shuffled Order Predict Network (MTFNet), based on a spatial-temporal feature extraction structure, to perform effective learning-related facial expression recognition in a primary school classroom. Specifically, we combine a Multi-Scale Facial Feature Fusion Module (MFFF) based on Grouped Spatial Convolution(GS Conv) to effectively capture multi-level facial features and improve the model's robustness in complex environments. Additionally, a Frame-wise Shuffle Order Prediction Module (FSOP) is introduced to enhance the model's ability to understand the dynamic changes of emotional intensity by predicting the emotion expression sequence. Experiments on both the DFEW dataset and our dataset demonstrate excellent performance and generalization in real-world applications.