Go to ICLR 2024 Conference homepageA Novel Approach for Micro-Expression Recognition Incorporating Vertical Attention and Position Localization
Micro-expression (ME) is a kind of facial expression that is short-lived and difficult for ordinary people to detect. Micro-expression can reflect the real emotion that people try to hide. It is difficult to identify micro-expression due to the fact that the duration is short and it only involves partial muscle motions, which brings great challenges to the accurate identification of micro-expression. To address these issues, we propose a novel neural network for micro-expression recognition (MER), focusing on subtle changes in facial movements using a CVA (Continuously Vertical Attention) block, which models the local muscle changes with minimal identity information. Additionally, we propose a facial position localization module called FPF (Facial Position Focalizer) based on Swin Transformer, which incorporates spatial information into the facial muscle movement pattern features used for MER. We also proved that including AU (Action Units) can further enhance accuracy, and therefore we have incorporated AU information to assist in micro-expression recognition. The experimental results indicate that the model achieved an average recognition accuracy of 94.35% and 86.76% on the popular CASME II and SAMM micro-expression datasets, improved by 6% and 1.98% compared to state-of-the-art models, respectively.