Robust facial expression recognition by simultaneously addressing hard and mislabeled samples

Facial Expression Recognition (FER) in the wild is a challenging task, especially when training data contains numerous mislabeled samples and hard samples. Typically, FER models either overfit to the mislabeled samples or underfit to the hard samples, resulting in degraded performance. However, most existing methods fail to address both issues simultaneously. To overcome this limitation, this paper introduces a novel FER method called Noise-Hard robust Graph (NHG), which dynamically supervises the updating of the adjacency matrix in the Graph Convolutional Networks, striking a balance between suppressing the impacts of noisy labels and encouraging learning from hard samples. First, we map high-dimensional facial expression features onto low-dimensional manifolds to initialize the topological relationships between the samples, thus measuring the hard sample relationships more accurately. Second, we design a Label Consistency Mask (LCM) strategy to retain potential connections for hard sample learning. LCM could also potentially preserve correct connections while noisy labels exist, supporting noise-robust learning. Third, based on differences in trends of ${\ell }_2$-norm variation between mislabeled samples and hard samples, the ${\ell }_2$-norm regularization (L2R) suppresses the learning of mislabeled samples while preserving the learning potential of hard samples and suppressing the propagation of their features within the graph. Experimental results demonstrate that our method achieves competitive performance compared to state-of-the-art methods in scenarios with noisy labels and hard samples.