Novel Multi-image Mixing for Long-tailed Image Classification

Long-tailed data often leads to a decrease in the generalization of deep neural networks. In this paper, we propose a data augmentation method called MosaicMix. Specifically, MosaicMix divides the image regions in a mosaic manner, and then cuts and pastes the randomly sampled images into the divided regions of the majority-class images to perform multi-image mixing. while the minority-class images remain unchanged. The mixing on majority-class images can suppress the model’s bias to majority classes. At the same time, mixing multiple images and keeping minority-class images unchanged help increase the chances of minority-class images appearing to appear in image mixing and improve the model’s attention to minority classes. We conducted experiments on CIFAR10-LT, CIFAR100-LT and ImageNet-LT. The results show that compared with other image mixing methods, MosaicMix significantly improves the accuracy of minority classes and performs well on long-tailed datasets.