SAR-to-Optical Image Translation Using SSIM and Perceptual Loss Based Cycle-Consistent GAN
Jieon Hwang, Chushi Yu, Yoan Shin
Abstract
Synthetic aperture radar (SAR) and optical sensing are different earth observation methods. Compared to the optical sensors, the SAR has the imaging advantages such as all-weather, all-time, ability to traverse clouds and vegetation, etc. We propose the application of the cycle-consistent generative adversarial network (CycleGAN) for translating the SAR images and the optical images with each other by training two generator networks and two discriminator networks simultaneously. In this paper, we add the structure similarity index measure (SSIM) loss factor and perceptual loss into the basis CycleGAN's loss function for keeping rich structural information. The tentative study demonstrated the potential of the SAR and the optical image translation based on the CycleGAN. The proposed method was verified and the experimental results showed the superiority of the proposed method with the combined loss function.