Abstract
Data augmentation is beneficial for improving robustness of deep meta-learning. However, data augmentation methods for the recent deep meta-learning are still based on photometric or geometric manipulations or combinations of images. This paper proposes a generative adversarial autoaugment network (GA3N) for enlarging the augmentation search space and improving classification accuracy. To achieve, we first extend the search space of image augmentation by using GANs. However, the main challenge is to generate images suitable for the task. For solution, we find the best policy by optimizing a target and GAN losses alternatively. We then use the manipulated and generated samples determined by the policy network as augmented samples for improving the target tasks. To show the effects of our method, we implement classification networks by combining our GA3N and evaluate them on CIFAR-100 and Tiny-ImageNet datasets. As a result, we achieve better accuracy than the recent AutoAugment methods on each dataset.
Original language | English |
---|---|
Article number | 108637 |
Journal | Pattern Recognition |
Volume | 127 |
DOIs | |
State | Published - Jul 2022 |
Bibliographical note
Publisher Copyright:© 2022 Elsevier Ltd
Keywords
- Adversarial learning
- AutoAugment
- Classification
- Data augmentation
- Deep learning
- Generative adversarial network