This  paper  analyzed  the  effect  of  the  difference  in  tilt  angle  between  the  original  image  and  the  reference  image on the simulation results when deep learning-based COSMO-SkyMed SAR-to-Optical image translation. For 'Image-to-Image Translation' of two satellite images with different tilt angles, the UNSB (Unpaired Image-to-Image Translation via Neural Schrödinger Bridge) model with FFC (Fast Fourier Convolution) blocks was used. As a reference image, the results were compared and analyzed by conducting an experiment with an optical image with a difference in tilt angle of about 19.8° from the SAR (Synthetic Aperture Radar) image and an optical image with a difference in tilt angle of about 43.7°. As a result of the experiment, when referring to an optical image with a small difference in tilt angle from the SAR image, the slope of the building in the SAR image was relatively well corrected while maintaining the edge of the terrain features. Additionally, the quantitative metric SSIM achieved a value of approximately 0.8782, indicating relatively strong performance.



Keywords :   Image Restoration, COSMO-SkyMed, UNSB, FFC, High-resolution SAR Image