Micro-expression Recognition Based on Multi-scale Attention

The main challenges in micro-expression recognition are unbalanced sample data and subtle changes in the facial area. To tackle these problems, we propose a discriminative method, which combines spatio-temporal features and global-local information. Initially, motion amplification is employed to obtain an enlarged frame feature map between key frames. Subsequently, the TV-L1 optical flow method is utilized to extract optical flow calculation features and optical strain characteristics in the key frames, which are then combined through channel superposition. Lastly, the pre-trained network ResNet18 is enhanced with a global-local information module, enabling comprehensive utilization of global spatio-temporal features and local spatio-temporal features for micro-expression recognition. Experimental results demonstrate that the proposed method effectively identifies and categorizes micro-expressions, achieving accuracies of 81.4% and 75.7% in five-class tasks on the CASME II and SAMM datasets, respectively. Additionally, accuracies of 90.3% and 84.2% were attained in three-class tasks, surpassing existing recognition techniques and reaching advanced levels.