Augmenting Colonoscopy Using Extended and Directional CycleGAN for Lossy Image Translation

CycleGAN [1] is one recent successful approach to learn a mapping from one image domain to another with unpaired data. We investigated CycleGAN as a solution to artistic style transfer, in particular, translating photographs to Chinese paintings. To improve the stability of training, we improved CycleGAN based on Wasserstein generative adversarial network (WGAN) and further improved WGAN with gradient penalty. The performance of CycleGAN, CycleWGAN, and Improved CycleWGAN are compared on a self-collected dataset CNPaintings both quantitatively and qualitatively.