Micro-expression recognition through feature enhancement and region weighted fusion based on supervised contrastive learning

Micro-expression recognition has aroused active research interest due to its extensive applications in various fields including public security, human-computer interaction, medical care etc. However, micro-expression suffers from extremely low intensity and short duration, resulting in enormous difficulty in its accurate identification. In this article, to improve the feature correlation of homogeneous samples and enhance the ability of local detailed feature extraction, a feature enhancement and regional weighted fusion method for micro-expression recognition based on supervised contrast learning has been proposed. Specifically, using ResNet as backbone, a powerful dual branch network under supervised contrast learning is designed, which on the one hand extracts the features of the eye and mouth regions respectively, and on the other hand improves the feature correlation of the homogeneous sample pair. Among of them, motion amplification and optical flow are used to amplify the subtle facial features to improve their discrimination. To effectively perceive the vital fine-grained feature information, a SE-Conv refinement attention mechanism is proposed to suppress background interference and a region weighted fusion strategy is adopted to combine features from different facial regions. The proposed method has been evaluated extensively on three public datasets, outperforming most of state-of-the-art methods.