Dynamic Graph-Transformer with Self-Supervised Learning for Micro-Expression Recognition

Abstract

Micro-expression recognition (MER) remains challenging due to extremely weak facial motions, strong locality, and the limited scale of available annotations. Existing Transformer-based methods are effective at modeling global dependencies, but they usually encode facial structure only implicitly. In contrast, graph-based approaches can introduce explicit region relations, yet many of them rely on fixed topologies and are not tightly integrated with modern self-supervised learning paradigms. To address these issues, we propose a Dynamic Graph-Transformer framework for MER (DGT-MER), which jointly models structured facial region interactions and global motion context. Specifically, we construct a multi-relation facial graph using geometric adjacency, motion similarity, and bilateral symmetry, and couple it with a Transformer branch through cross-branch attention. In addition, we introduce a graph-guided self-supervised pretraining strategy that combines node reconstruction, edge prediction, and cross-view contrastive learning to improve representation robustness under limited supervision. Experiments on CASME II, SAMM, and CASME3 show that the proposed framework achieves competitive performance and consistent improvements over strong baselines.