FSRNet: Facial Structural-aware Representation Learning for Facial Expression Recognition via Swin Transformers

Dynamic facial expression recognition (DFER) is an essential upstream task in the areas of human-machine interaction, driving safety, education and health aids. However, the inherent uncertainty between variations significantly impacts the effectiveness of DFER in practical applications, particularly due to the similar facial changes associated with different facial expressions. To address this challenge, we introduced a novel relationship-driven approach based on the Transformer architecture that can learn the relationships between facial expressions and facial variations. Specifically, we designed several facial variation tokens to explicitly encode the variation status of key facial regions. In addition, we developed a novel module to model the interactions between facial variation tokens and help the model better focus on the specific facial variations of facial expressions. Experimental results on two challenging benchmark DFER datasets show that our proposed FSRNet achieves state-of-the-art performance.