Synthetic Minority with CutMix for Imbalanced Image Classification

Chenghua Zeng, Huijuan Lu, Kanghao Chen, Ruixuan Wang, Jun Tao

Abstract: The class-imbalanced distributions between different classes in the visual world pose great challenges for deep learning-based classification models particularly on correct prediction of minority classes. In this study, different from existing strategies to alleviate the data imbalance issue, a novel mechanism based on the CutMix regularization technique is proposed for imbalanced image classification. The novelty is from two aspects. First, a novel sampling strategy is proposed to create the synthetic training data with a more balanced distribution. Second, labels of synthetic images were assigned with a bias toward minority classes. With the novel sampling and label assignment, more synthetic images of minority classes can be obtained to balance the class distribution of training data. Experiments on three benchmark datasets justified that the proposed method consistently outperforms commonly used strategies to alleviate the class imbalance issue.