ChannelMix: A Mixed Sample Data Augmentation Strategy for Image Classification

Recent emerging convolutional neural networks (CNNs) have powerful feature extraction ability and can achieve the state-of-the-art classification performance. However, there exists a problem that CNNs have the memorization to training samples and sensitivity to adversarial examples, resulting in overfitting and generalization decline. To solve the problem, this paper proposes a novel mixed sample data augmentation (MSDA) strategy named ChannelMix to improve network performance. Specifically, ChannelMix uses the multi-channel information and labels of paired samples to regularize the training process through convex combination, which can guide the network to pay more attention to the less discriminative parts. Extensive experiments on the CIFAR-10, CIFAR-100 and WHU-RS19 datasets demonstrate that ChannelMix can significantly improve the generalization and the classification performance of CNNs and stabilize the training process simultaneously.