Head-Movement-Robust Micro-Expression Detection Method via 3D Motion Correction and Transformers

In micro-expression detection, head movements may seriously interfere with subtle and transient micro-expression signals, significantly limiting detection performance. Therefore, we propose a robust detection framework that integrates 3D motion correction and a Transformer network. This framework adopts a two-stage design. In the first stage, a depth-weighted optical flow method is proposed to decompose and suppress head motion in three-dimensional space and extract anti-interference temporal optical flow features. In the second stage, a Transformer-based encoder is used to model and classify feature sequences by leveraging its capability for global feature modeling integrated with positional information. The experimental results on the CASME3-PartC and MEVIEW datasets show that the proposed method achieves F1-scores of 0.151 and 0.326, respectively, outperforming existing methods and achieving the current optimal performance. Cross-dataset experiments further validated the good generalization ability of this method.