TM-SGANet: Micro-expression recognition by integrating terrace feature extraction and spatiotemporal genetic attention
Ning Wang, Qi Hu, Jinshu Zhao, Shiro Kan, Jiajian Li, Mingxin Zhao · Applied Soft Computing · 2026
Micro-expressions are key cues that reveal underlying emotions. However, their extremely brief and subtle nature poses significant challenges for automatic recognition. Although various handcrafted and deep learning-based feature engineering methods have been proposed for micro-expression recognition, existing approaches often struggle to precisely capture the subtle features of micro-expressions and the minute motion variations within sequences. To better perceive the regions where micro-expressions occur and their fine-grained changes, this paper proposes a novel micro-expression recognition method that integrates the Terrace Module for Multi-Level and Multi-Scale Feature Extraction (TMMFE) and Spatiotemporal Genetic Attention (SGA). During the feature extraction stage, TMMFE enables the network to effectively capture both primary and secondary regions of interest, thereby reducing the loss of subtle micro-expression features. The SGA module leverages spatial and temporal characteristics in the data to capture fine motion differences across video frames. The proposed method was extensively evaluated on the SAMM, CASME II, SMIC, and CAS(ME)³ datasets. Experimental results demonstrate that our method achieves significant improvements in performance compared with state-of-the-art micro-expression recognition approaches.