Realistic Image-to-Image Translation with Enhanced Texture

Guozhen Duan, Yanxiang Gong, Huijie Zhao, Wen Ma, Dongxing Song, Zheng Ma, Mei Xie

Abstract: In the image-to-image translation field, most researchers tend to achieve the overall translation of images without paying too much attention to the texture details of images. However, it is also of great importance to have enhanced and more realistic textures for synthesized images, which could bring better impressions. Therefore, in this work, we propose a method based on CycleGAN and the texture of output images is highly improved. The presented generator involves dilated convolutions which are conducive to processing image texture details. Furthermore, an improved cycle consistency loss is proposed for stable and effective training. The experiments demonstrate that our proposed method is able to generate images that contain more details and better meet the visual perception of humans. Our code will be publicly available at GitHub soon.