Dynamic Mutual Enhancement Network For Single Remote Sensing Image Dehazing

Adversarial examples are specialised inputs that are deliberately designed to cause neural networks to misclassify. interestingly, in this paper, we demonstrate that adversarial examples can also be utilized for good to improve the performance of imbalanced learning. We provide a new perspective to deal with imbalanced data: adjust the biased decision boundary by training with Guiding Adversarial Examples (GAEs). GAEs refer to adversarial examples in minority classes that can be transferred to majority classes within a few steps. These examples can further be utilized to guide the learning of the biased decision boundary. Note that our method does not require training from scratch, we can simply retrain the biased model on the original data for a few epochs. Extensive experiments show that our method outperforms other state-of-the-art methods on several benchmark datasets, and our method requires just a few training epochs (\eg on CIFAR10 dataset, the training epoch of our method is $0.1\times$ of other state-of-the-art methods).