Image Augmentation Strategies to Train GANs with Limited Data

Abstract

Training modern generative adversarial networks (GANs) to produce high-quality images requires massive datasets, which are challenging to obtain in many real-world scenarios, like healthcare. Training GANs on a limited dataset overfits the discriminator on the data to the extent that it cannot correctly distinguish between real and fake images. This paper proposes an augmentation mechanism to improve the dataset’s size, quality, and diversity using a set of different augmentations, namely flipping of images, rotations, shear, affine transformations, translations, and a combination of these to form some hybrid augmentation. Fretchet distance has been used as the evaluation metric to analyze the performance of different augmentations on the dataset. It is observed that as the number of augmentations increase, the quality of generated images improves, and the Fretchet distance reduces. The proposed augmentations successfully improve the quality of generated images by the GAN when trained with limited data.