<p>Real-world data tend to follow a long-tailed distribution. Deep models trained on such datasets often exhibit bias toward head classes and perform poorly on tail classes. Existing methods, such as reweighting, have significantly improved the performance of the long-tailed recognition. However, they fail to address the problem of the great uncertainty in model prediction, which results in the weak generalization ability of deep networks. To tackle the issue, we propose a novel approach called consistency self-distillation with augmented mixture (CSAM), which consists of two core components: augmented mixture (AM) and consistency self-distillation (CSD). AM generates two different types of augmented samples to enrich the data, thus improving generalization and reducing the uncertainty in model predictions. Specifically, it employs a weak–strong augmentation strategy and a global–local mixture method to generate weakly augmented global and strongly augmented local samples. CSD further reduces the prediction uncertainty by distilling knowledge from predictions of the weakly augmented global data to regularize strongly augmented local images. Moreover, we propose a hybrid rebalancing strategy that integrates resampling and logit adjustment methods to handle the worse tail class compression problem in our method. Our approach achieves outstanding performance on CIFAR10-LT, CIFAR100-LT, ImageNet-LT, and Places-LT, demonstrating the effectiveness and superiority of our proposed CSAM.</p>