Asymmetric CycleGAN for image-to-image translations with uneven complexities

CycleGAN is one of the famous and basic methods for unpaired image-to-image translation tasks. Inspired by the experiments of the NIR-RGB translation, which is a kind of translation where images are translated from simple to complex or vice versa, we concluded the definition of asymmetric translation task. Because of the complexity difference between two domains, the complexity inequality in bidirectional translations is significant. We analyzed and witnessed the limitation of the original CycleGAN in asymmetric translation tasks and proposed an Asymmetric CycleGAN model with generators of unequal sizes to adapt to the asymmetric need in asymmetric translations. An empirical metric was also given to determine the asymmetric task from the aspect of image entropy and could be treated as the auxiliary guidance to design the asymmetric generators. Besides, the edge-retain loss between the input and the generated images was introduced to enhance the structural visual quality. Residual-block-net based and U-net based generators were both applied here to verify the Asymmetric CycleGAN. The performance of different depth of generators for Asymmetric CycleGAN was also discussed on the basis of experiments. The qualitative visual evaluation demonstrated that our model had achieved great improvements compared to original CycleGAN. (C) 2020 Elsevier B.V. All rights reserved.