Micro-Expression Recognition Method Based on Transformer with Separable Self-Attention

Micro-expressions, which are brief facial movements reflecting an individual's emotions and psychological state, are crucial in fields like psychological counseling and trust evaluation. However, they are challenging to capture, and traditional methods struggle with effective recognition. While neural networks are a primary tool for this task, they are limited by their large number of parameters and slow processing speeds. Transformer architectures have shown promise in micro-expression recognition but face similar challenges. To address these issues, an improved HTNet model, called Lightweight HTNet (LWHTNet), has been proposed. Specifically, by using a Separable Self-Attention mechanism to replace the multi-head attention mechanism in Transformer layers and incorporating depthwise separable convolutions in the block aggregation layers, the model reduces the number of training parameters. In experiments, LWHTNet demonstrated strong performance on composite datasets, achieving an Unweighted F1 Score (UF1) of 0.8498 and an Unweighted Average Recall (UAR) of 0.8307. Additionally, the optimized model showed significant improvements in inference speed and parameter efficiency.