Diffusion-augmented learning for long-tail recognition

In general, natural data are long-tail distributed over semantic classes. Existing recognition methods tackle the imbalanced classification problem by designing architectures that focus more on the tail classes without altering the imbalance in the original dataset. In this paper, we propose an augmentation learning strategy, which not only alleviates the bias of the classifier towards head classes but also generates a balanced augmented dataset. Our method is called Diffusion-Augmented Learning (DA), which can be easily applied to any long-tail recognition model. Its core idea is to leverage the generative capabilities of diffusion models to generate a balanced augmented dataset with high confidence and rich feature information. Subsequently, through a specific sampling strategy, we mix the original dataset with the augmented dataset to tackle long-tail recognition. The purpose of our method is twofold: 1) to utilize the generative capability of diffusion models to add sufficient and reliable augmented data for training, which improves the robustness of the model, and 2). to introduce more augmented data for tail classes through both probabilistically replacing strategy and an improved diffusion model, which alleviates the model's bias towards head classes. We carried out a series of contrast experiments on the CIFAR10-LT, CIFAR100-LT, and ImageNet-LT datasets, and the experimental results demonstrate the superiority of our proposed method and that the augmented dataset is versatile.