Robust consistency learning for facial expression recognition under label noise

Label noise is inevitable in facial expression recognition (FER) datasets, especially for datasets that collected by web crawling, crowd sourcing in in-the-wild scenarios, which makes FER task more challenging. Recent advances tackle label noise by leveraging sample selection or constructing label distribution. However, they rely heavily on labels, which can result in confirmation bias issues. In this paper, we present RCL-Net, a simple yet effective robust consistency learning network, which combats label noise by learning robust representations and robust losses. RCL-Net can efficiently tackle facial samples with noisy labels commonly found in real-world datasets. Specifically, we first use a two-view-based backbone to embed facial images into high- and low-dimensional subspaces and then regularize the geometric structure of the high- and low-dimensional subspaces using an unsupervised dual-consistency learning strategy. Benefiting from the unsupervised dual-consistency learning strategy, we can obtain robust representations to combat label noise. Further, we impose a robust consistency regularization technique on the predictions of the classifiers to improve the whole network’s robustness. Comprehensive evaluations on three popular real-world FER datasets demonstrate that RCL-Net can effectively mitigate the impact of label noise, which significantly outperforms state-of-the-art noisy label FER methods. RCL-Net also shows better generalization capability to other tasks like CIFAR100 and Tiny-ImageNet. Our code and models will be available at this https https://github.com/myt889/RCL-Net.