Engagement Prediction Method Based on Spatiotemporal Facial Feature Redistribution

Automated detection of learners' engagement levels enhances online teaching by enabling timely feedback and personalized adjustments based on student needs, thereby improving teaching effectiveness. Traditional methods rely heavily on single-frame facial spatial data, often neglecting temporal emotional and behavioral features and being sensitive to pose variations. Additionally, convolutional padding can degrade feature maps, impacting feature extraction quality. To tackle these challenges, we propose the Spatiotemporal Facial Feature Redistribution and Temporal Convolutional Network (SFRNet). This hybrid neural network architecture incorporates three key components: utilizing the spatial attention mechanism LKA to capture local patches and mitigate pose effects, employing the FOWD module to redistribute feature weights, eliminate irrelevant features, and enhance facial representation, and analyzing temporal changes using the ModernTCN module to detect engagement levels. We curated a near-infrared engagement video dataset (NEVD) to validate SFRNet's efficiency, demonstrating superior performance in engagement video analysis on both NEVD and publicly available DAiSEE datasets through extensive experimentation and comprehensive studies.