DUAL: A Dual-Stage Approach for Facial Expression Recognition Based on Contrastive Learning

<p dir="ltr">Facial Expression Recognition (FER) remains a challenging task in computer vision. Recent works</p><p dir="ltr">have shown excellent performance in overall recognition accuracy, but its accuracy significantly</p><p dir="ltr">decreases when recognizing similar expressions. This is due to inter-class homogeneity and intra-</p><p dir="ltr">class heterogeneity. To address these issues, we propose a novel dual-stage network called DUAL,</p><p dir="ltr">inspired by contrastive learning. First, we increase the distance between negative samples while</p><p dir="ltr">reducing the distance between positive ones. This is achieved by dynamically updating pairs of</p><p dir="ltr">comparison samples. Second, we introduce a two-stage network architecture. The first stage uses</p><p dir="ltr">two branches to extract image features and facial keypoint features. These branches interact to learn</p><p dir="ltr">coarse-grained features through mutual guidance. The second stage focuses on fine-grained features</p><p dir="ltr">using scale-specific residual blocks. This allows the model to identify facial regions that are critical</p><p dir="ltr">for recognizing expressions. We conducted extensive experiments on multiple datasets. The results</p><p dir="ltr">show that DUAL surpasses state-of-the-art models in items of performance. Additionally, the model</p><p dir="ltr">shows high accuracy even in noisy conditions, highlighting its robustness.</p>