A Survey on Data Augmentation Methods Based on GAN in Computer Vision

Data augmentation is a significant technique in deep learning. As Generative Adversarial Networks (GAN) have gained wide attention, data augmentation methods based on GAN are utilized in many applications. This paper is aimed to provide a survey on various GAN-based data augmentation methods in computer vision. These methods are divided into three types according to their function. The first is generating samples from latent space directly. The second is image translation. The last produces augmentation operations utilizing adversarial strategy. The theories and applications of these methods are introduced in detail. Their defects and future research directions are discussed at the end.