Performing Effective Generative Learning from a Single Image Only

Abstract

Generative adversarial networks (GANs) can be well used for image generation. Yet their training typically requires large amounts of data, which may not be available. This paper proposes a new algorithm for effective generative learning given a single image only. The proposed method involves building GAN models with a hierarchical pyramid structure and a parallel-branch design that enables independent learning of the foreground and background areas. This work conducts a set of well-designed experiments. The results well demonstrate that the proposed method produces the images of higher quality and better diversity than existing methods do. Thus, this work advances the field of generative learning for image generation.