Long‑Tailed Learning with In‑ and Out‑of‑Distribution Noisy Labels in the Open World

When training deep neural networks in practical application scenarios， the data used often have various biases， such as long-tailed category distributions， in-distribution noise， and out-of-distribution noise. Most existing methods focus on solving the problem of category imbalance or dealing with noisy labels， but rarely consider both aspects simultaneously， especially when in- and out-of-distribution noises exist at the same time. We propose an imbalanced noisy labels calibration （INLC） method to address this challenge. To handle out-of-distribution samples， we use the model’s consistent predictions to filter them out and assign uniform labels， thereby enhancing the model’s ability to detect out-of-distribution samples. For in-distribution samples， we use the Jensen-Shannon divergence to distinguish noise and reduce misclassification of clean samples， especially in tail categories. To address the problem of category imbalance， we introduce an additional semantic classifier to mitigate the bias of pseudo-labels towards majority categories. Finally， we adopt a consistency regularization method based on strong data augmentation to further improve the model’s generalization performance. We conducted extensive experiments on simulated and real-world datasets， covering different levels of category imbalance from low to high and different proportions of label noise. Experimental results show that INLC significantly alleviates the impact of label noise and category imbalance， and improves the classification accuracy by more than 2% compared with the previous state-of-the-art baseline methods.