SAR-to-Optical Image Translation Using SSIM Loss Based Unpaired GAN

The synthetic aperture radar (SAR) image can be acquired regardless of weather condition and time through radar reflection. However, visual effects and observations of the SAR image is difficult due to single channel information and speckle noise. In this paper, we present the disadvantage that cycle-consistency loss focuses only on texture based on the CycleGAN, and propose to apply the dual contrastive learning GAN (DCLGAN) that maximizes mutual information between SAR and optical images by learning the correlation of image patches. Also, we propose a method to apply the structural similarity index measure (SSIM) loss instead of cycle-consistency loss to address the lack of expressiveness by emphasizing structural features. The proposed DCLGAN eliminates cycle-consistency loss of CycleGAN and maximizes mutual information through contrastive learning to translates unpaired SAR-optical images. In addition, the proposed SSIM loss helps translate the SAR-optimal image by recognizing perceived changes in structural information. We confirm the performance through comparative experiments with existing models to verify unpaired SAR-to-optical image translation.