ECycleGAN

Abstract

Unsupervised image-to-image translation, which aims in translating two irrelevant domains of images, has increased substantially in recent years with the success of Generative Adversarial Networks (GANs) based on the cycle-consistency assumption. Especially, the Cycle-Consistent Generative Adversarial Network (CycleGAN) has shown remarkable success for two domains translation. However, the details about texture and style are often accompanied with unpleasant artifacts. To further enhance the translational quality, we thoroughly study the key components of CycleGAN - network architecture and adversarial loss, and improve each of them to derive an Enhanced CycleGAN (ECycleGAN). In particular, we propose a perceptual loss function which motivated by perceptual similarity instead of similarity in pixel space. Moreover, we introduce the Residual Dense Normalization Block (RDNB) to replace the residual blocks as the basic network building unit. Finally, we borrow the idea from WGAN-GP as the adversarial loss functions. The ECycleGAN, thanks to these changes, demonstrates appealing visual quality with more realistic and natural textures than any state-of-the-art methods.