Distribution Mining and Noise Correction for Facial Expression Recognition

Due to the high annotation cost and inter-class ambiguity, many facial expression datasets only offer noisy one-hot labels. This creates two challenges for facial expression recognition (FER) as the one-hot labels only represent limited single semantics, and the noisy labels will mislead the model. Our study proposes a novel method to perform both label Distribution Mining and Noise Correction (DMNC) on the FER dataset. That is, while eliminating the disturbance of noisy samples, we start from the original one-hot labels and gradually refine them to distributions of expressions close to the real world. Specifically, DMNC uses the network forgetting property to select noisy samples. Then based on clean samples, an expression feature space is learned by contrastive learning. Finally, the labels of clean and noisy samples are refined or relabelled in a divide-and-conquer manner using the learned feature space. Experiments show that our method exhibits leading performance on both real-world FER benchmarks and synthetic noisy datasets.