SPCL-MER: Supervised Prototypical Contrastive Learning for Micro-Expression Recognition

Micro-expressions serve as a crucial psychological stress response, which can reveal people’s genuine emotions. However, extracting recognizable micro-expression features is still a challenging task due to issues such as data scarcity and subtle motion variations. In this paper, we propose a novel two-stage micro-expression recognition framework, SPCL-MER. Firstly, we employ a COC structure as the backbone, which groups pixels by similarity to extract features and make the model more suitable for processing optical flow images. Secondly, to overcome the effects caused by irrelevant facial motion, we propose a new loss function based on Supervised Contrastive Loss and Prototypical Loss that allows the network to obtain more robust features from a limited number of samples. Our extensive experiments on two publicly available datasets demonstrate that our approach achieves state-of-the-art performance. The codes are available at https://github.com/fxq0216/SPCL-MER.