Micro-expression recognition with supervised contrastive learning

Facial micro-expressions are involuntary movements of facial muscles that expose individual underlying emotions. Because of the subtle and diverse facial muscles change, extracting effective features to recognize micro-expressions is challenging. In this paper, a framework for micro-expression recognition with supervised contrastive learning (MER-Supcon) is proposed, and the primary purpose is to extract crucial features of micro-expressions and overcome the noise caused by irrelevant facial movements. First, a novel dual-terminal micro-expression acquisition strategy is proposed and applied to obtain optical flow maps, which aims to expand the datasets and reduce the adverse impact of micro-expression spotting. Then, supervised contrastive learning is introduced to learn the key representation of micro-expressions for classification. The results on CASME II and SAMM datasets show that the approach is effective and competitive compared with the state-of-the-art methods both on three-classes and five-classes evaluations.