Multi-source Unsupervised Domain Adaptation for Micro-expression Recognition

Micro-expressions are brief, involuntary facial expressions that reveal genuine emotions, making their accurate detection crucial in fields like security and psychology. Due to their subtle intensity and short duration, accurately capturing the nuanced movements of micro-expressions presents a significant challenge. While current micro-expression recognition methods have made great strides on individual ME datasets, practical applications face additional challenges. Differences in acquisition environments and equipment can create distributional discrepancies between the training and test sets, potentially affecting recognition results. Additionally, the small size of single micro-expression datasets can lead to inadequate model training when used alone. To address these issues, we propose a micro-expression recognition method based on a multi-source unsupervised domain adaptation technique (MSU-MER). This method incorporates multi-source domain adaptation into micro-expression feature extraction to reduce distributional differences between domains. Our experiments, conducted on three publicly available micro-expression datasets, demonstrate that the proposed method achieves competitive results.