A progressive decay Mixup data augmentation for long-tailed image classification

Image classification with long-tailed distribution ( i.e ., a few classes occupy most of the training samples, while most classes have only a small number of representations) is a challenging visual task at present. In this article, we propose a novelty Mixup-based data augmentation technique, which uses two sampling mechanisms ( i.e ., instance-based sampling and class-based sampling) to independently sample and mix the training samples. Specifically, we obtain a conventional sample distribution from instance-based sampling and a balanced sample distribution from class-based sampling. The two sets of samples are then mixed to provide a more balanced sample distribution for model training and effectively reduce the under-fitting of tail classes. In addition, we also design a decay mixing strategy to dynamically adjust the sample mixing weights during the training process. By doing so, we can gradually guide the model training and prevent over-fitting. We validate the proposed method on the long-tailed version of datasets created by CIFAR-10, CIFAR-100, CINIC-10, Tiny-ImageNet, Fashion-Mnist, and ImageNet-LT. Experimental results demonstrate that our method achieves better classification performance than other advanced Mixup-based techniques in data imbalanced scenarios.