Fine-Grained Data Augmentation using Generative Adversarial Networks

Se-Hun Kim, Chunmyung Park, Min-Ho Choi, Seung-Jin Yang, Kyujoong Lee, Hyuk-Jae Lee

Abstract: This paper presents fine-grained data augmentation, a data augmentation method for deep neural network training that can be applied to tasks with a small number of images, such as in the medical field or vision-inspection tasks. For small-datasets, the number of images per class is usually unbalanced and overfitting occurs when training small-datasets. In this paper, data augmentation skills using generative adversarial network for image super-resolution tasks is presented. Data augmentation with generative adversarial network for image super-resolution tasks retains the overall shape and form, but changes only the details of features. The proposed method achieves better performance when training CIFAR-100 and CUB-200-2011 datasets from scratch. The proposed method is being actively developed to further improve the performance of image classification and will be applicable to object detection.