Intra-Class Cutmix for Unbalanced Data Augmentation
Caidan Zhao, Lei Yang · 2021
In the case of the training dataset suffering from heavy class-imbalance, deep learning algorithms may perform poorly. Due to the data-poor, the neural network cannot fully learn the representation of minority classes. In this paper, we proposed a data augmentation strategy called Intra-Class Cutmix for unbalanced datasets. Our algorithm can enhance the learning ability of neural networks for minority classes by mixing the intra-class samples of minority classes, and correct the decision boundary affected by unbalanced datasets. Although the method is simple, for unbalanced datasets, our method can be used as a supplement to traditional data augmentation methods (such as Randomerasing, Cutmix, etc.) to further enhance the performance of the network. In addition, Intra-Class Cutmix is also suitable for advanced re-balancing strategies. We conducted experiments on the CIFAR-10, CIFAR-100 and Fashion-MNIST datasets. Our results proved the effectiveness and universality of our method.