SNEFER: Stopping the Negative Effect of Noisy Labels Adaptively in Facial Expression Recognition

Noisy labels make a profoundly negative impact on facial expression recognition (FER) due to interclass similarity and subjective annotation. Recent works mainly focus on distinguishing the clean samples or mining the latent truth, which not only need extra computational overhead but also have the error risk of relabeling. In this work, we propose an SNEFER model without elaborately discriminating noisy labels, which can adaptively stop the negative effect by a novel contrastive regularization term. Specifically, we establish two branches: class prediction and contrastive regularization, in a parallel manner. Though the class prediction branch learns facial expression representations over noisy labels, it keeps the ability to measure the similarity of two facial images but is not robust for FER. If the similarity exceeds a given threshold, the regularization branch pulls them tightly in the embedding space through a bilateral contrastive regularization (BCR) loss, which prevents noisy labels from degrading the FER model adaptively. Experimental results demonstrate that our SNEFER model outperforms the state-of-the-art methods on three FER benchmarks under different noisy levels.