Pet Hair Color Transfer Based On CycleGAN

Generative adversarial networks (GANs) have shown great performance on image-to-image translation tasks. Many approaches have been proposed for translation of human face images, scene pictures and artful paintings, but few works considered about translating a pet image. In this paper, we propose a method based on cycle-consistent adversarial network (CycleGAN) to solve pet hair color transfer problem. Given a pet image, our model can translate its hair color into a desired one while keeping its other features unchanged, which makes our generated images seem quite realistic. We do several improvements on CycleGAN including doing segmentation to avoid the influence of background, and using spectral normalization to improve the quality of generated images. We build a large pet image dataset consisting of a total number of 7. 5K images, categorized by different hair colors. Our proposed method is trained and tested on this data set and the results show the promising performance on translating between white and orange hair color of dog images.