CRmix: A regularization by clipping images and replacing mixed samples for imbalanced classification

Deep convolutional neural networks (CNNs) often show inferior classification performance when training samples are seriously class-imbalanced. In this paper, we present a new Mixup-based regularization method, which loosens the constraint of only using synthetic samples to train the model. Specifically, we choose to cut and paste patches between the training images to generate samples, and restore corresponding minority image samples in the synthetic samples which contain a large proportion of minority image patches. At the same time, we restrict the location of patch generation, so as to intercept more image areas of training samples. By doing so, we increase the number of minority samples for the training set. The proposed method performs good generalization on minority classes by training the classifier to push the decision boundary toward majority classes. Besides, our approach integrates well with conventional rebalancing techniques such as re-weighting or re-sampling. We validate our method on the imbalanced datasets created by CIFAR-10, CIFAR-100, CINIC-10 and Tiny-ImageNet. The results demonstrate that our method is superior to some other Mixup-based regularization techniques (e.g., Mixup, Remix and Cutmix). For example, on the long-tailed version of CIFAR-10 (CIFAR-100) with an imbalance ratio of 100, our proposed method achieves a 4.90% (4.06%) improvement in classification accuracy compared with the classical Remix method.