Micro‐Expression Recognition Under Occlusion Based on Expression Semantic Surface Texture

Micro-expression recognition is a fine-grained task aimed at capturing subtle, brief facial movements. However, facial occlusions in real-world scenarios significantly challenge existing models, often leading to feature sparsity and disconnection. This arises from both the loss of spatial and temporal information, further disrupting feature dependencies across facial regions. To address these issues, we introduce the concept of expression semantic surface texture, designed to separate and reconstruct expression-aware and expression-irrelevant features. We propose a dual-branch collaborative network: one branch extracts spatial features using optical flow and frame differencing, while the second branch filters out expression-irrelevant features using occlusion position embedding and reconstructs expression-aware features in occluded regions. Experiments show that our network effectively restores disrupted features and outperforms recent state-of-the-art methods in occluded micro-expression recognition.