Sliding Distribution Learning for Facial Expression Recognition with Noisy Labels
Hanqi Gan, Yihua Tan · 2025
Due to the high cost of annotation and the ambiguity between classes, many facial expression datasets rely on noisy one-hot labels. This introduces two significant challenges for facial expression recognition (FER): the one-hot labels are limited in their representation of the underlying semantics, and the noise in these labels can misguide the model. In this work, we introduce a novel approach that integrates both Label Distribution Mining and Noise Correction (SDLNL) for FER datasets. Specifically, our method begins with the original one-hot labels and progressively refines them into distributional representations that better reflect real-world expression patterns, all while mitigating the influence of noisy samples. SDLNL leverages the network's forgetting mechanism to identify and filter out noisy instances. A feature space representing facial expressions is then learned through contrastive learning, based on the clean data. Subsequently, the labels for both clean and noisy samples are refined in a divide-and-conquer fashion, utilizing the learned feature space. Experimental results demonstrate that our approach outperforms existing methods on both real-world FER benchmarks and synthetic noisy datasets.