MERCoL: video-based facial micro-expression recognition via bimodal contrastive learning

Micro-expressions are brief, subtle, involuntary facial gestures revealing genuine mental activity, with numerous real-world applications. Owing to their transient nature, low intensity, and capture difficulty, many existing recognition algorithms based on handcrafted features and deep learning methods lack accuracy. We propose a micro-expression recognition framework called MERCoL, which utilises bimodal contrastive learning to extract common and distinctive features from limited dataset samples. The network comprises three modules: bimodal feature extraction, bimodal contrastive learning fusion, and classification. First, micro-expression sequences are divided into RGB and optical flow sequences, with a contrastive learning-based loss function for learning common bimodal features. Second, bimodal features are then fused and labelled data is used to optimise the network for distinctive feature learning. At last, experiments on CASME II, SAMM, and MMEW datasets demonstrate our algorithm's superiority compared to state-of-the-art methods.