Dynamic Temporal Attention Enhanced Multi-scale Hierarchical Network for Micro-expression Recognition

Despite significant advances in micro-expression recognition, challenges persist due to the subtle and rapid nature of facial micro-movements. Conventional self-attention mechanisms, while effective in capturing the feature similarity, lack targeted guidance for the temporal pattern of micro-expressions: rapid onset, instantaneous peak and gradual offset, hindering the model's ability to localize apex frame regions with the highest information density. To address this limitation, we propose a dynamic temporal attention enhanced multi-scale hierarchical network (DTA-EMHNet), which rethinks attention computation in micro-expression analysis. Our key innovation is to introduce instantaneous intensity and evolution velocity as the primary basis for generating attention weights, enabling adaptive adjustment of attention across temporal steps based on motion intensity and change velocity. Experiments on benchmark datasets demonstrate that DTA-EMHNet significantly improves apex frame localization accuracy and overall recognition performance, establishing new state-of-the-art results in micro-expression recognition.