Action Decouple Multi-Tasking for Micro-Expression Recognition

Micro-expressions are brief, involuntary facial movements that reveal genuine emotions. However, extracting and learning features from micro-expressions poses challenges due to their short duration and low intensity. To address this problem, we propose the ADMME (Action Decouple Multi-tasking for Micro-Expression Recognition) method. In our model, we adopt a pseudo-siamese network architecture and leverage contrastive learning to obtain a better representation of micro-expression motion features. During model training, we utilize focal loss to handle the class imbalance issue in micro-expression datasets. Additionally, we introduce an AU (Action Unit) detection task, which provides a new inductive bias for micro-expression detection, enhancing the model’s generalization and robustness. Through five-class classification experiments conducted on the CASMEII and SAMM datasets, we achieve accuracy rates of 86.34% and 81.28%, with F1 scores of 0.8635 and 0.8168, respectively. These results validate the effectiveness of our method in micro-expression recognition tasks. Furthermore, we validate the effectiveness of each component of our approach through a series of ablation experiments.