High-resolution Optical Image Cloud Removal and Image Restoration Using StyleGAN2

In order to solve the problem of information loss and image restoration caused by cloud occlusion in high-resolution optical images, this paper proposes an image restoration method based on StyleGAN2. Through the preprocessing step, the image containing clouds is input into the StyleGAN2 model. In StyleGAN2, the mapping network converts the random noise vector in the latent space into the style vector. StyleGAN2 is improved by removing the AdaIN layer and introducing the weight modulation and demodulation steps. Next, the improved StyleGAN2 model is used to extract features and control the style of images containing clouds. By adjusting the style vector, the model generates an image that is similar to the input image but with the clouds removed. This paper also uses the Style-mixing technology in StyleGAN2 to simulate the image features after the clouds are removed and through continuous iteration and optimization, the generated image gradually approaches the real scene. Finally, the generated images are further processed through the post-processing step to improve the quality and clarity of the image. Compared with the WGAN baseline model, StyleGAN2 shows advantages in indicators such as PSNR, SSIM and FID, with an average FID value of 11.276. This paper provides new solutions and ideas for the field of optical image processing.