Comparative Analysis of Pix2Pix and CycleGAN for Image-to-Image Translation

Abstract

Image-to-Image translation technology is nowadays a prevailing research orientation in computer-vision domain, which aims to translate the styles and features in images from one image domain to another. With the rapid development of the convolutional neural networks, especially generative adversarial network technology, breakthroughs have been made in the performance of image translation, which has been widely used in many fields such as labeling photos to synthesize photos, reconstructing objects from line drawings, and coloring pictures. The Image-to-Image Translation problem is essentially a pixel-to-pixel mapping. Limited by specific task settings, the performance of different general frameworks often fluctuates greatly when dealing with different translation tasks. In this paper, focusing on the above problems, the advanced and commonly used Image-to-Image Translation frameworks such as Pix2Pix and CycleGAN, are selected to compare and analyze the advantages in detail from different dimensions including network structure, loss function, applications and model accuracy. Furthermore, for different practical application scenarios, we discuss solutions based on these two representative frameworks and show the results after image translation processing. Finally, the existing problems and future research directions are discussed and summarized.