PME-MER: Efficient Position and Motion Encoder for Micro-Expression Recognition

Facial micro-expressions (MEs) are caused by brief and subtle movements of facial muscles, revealing a person's genuine emotions and offering valuable insights into lie detection, criminal analysis, and numerous human-computer interaction systems. Although deep learning-based micro-expression recognition (MER) approaches have achieved significant success, these methods often incorporate identity information that adversely impacts facial expression recognition and ignore capturing the synergistic interactions among different motion regions. To overcome these limitations, this paper proposes PME-MER, a novel MER method capable of effectively encoding facial position information and capturing muscle motion patterns. Specifically, we introduce Adaptive Positional and Local Motion Encoding (APME), a single-stream framework using cross-attention to align motion features with facial structure, facilitating accurate position calibration and minimizing the impact of identity information. Furthermore, we develop Collaborative Motion Encoding (CME), which utilizes self-attention and the top-k mechanism to explore the synergistic relationships among various facial regions, capturing comprehensive facial muscle movement information while reducing the impact of irrelevant regions on model predictions. Experimental results on four benchmark datasets demonstrate that our method significantly outperforms state-of-the-art methods, and ablation studies further demonstrate the effectiveness of our method.