Improved Residual Dense Network for Large Scale Super-Resolution via Generative Adversarial Network

Recent single image super resolution (SISR) studies were conducted extensively on small upscaling factors such as x2 and x4 on remote sensing images, while less work was conducted on large factors such as the factor x8 and x16. Owing to the high performance of the generative adversarial networks (GANs), in this paper, two GAN’s frameworks are implemented to study the SISR on the residual remote sensing image with large magnification under x8 scale factor, which is still lacking acceptable results. This work proposes a modified version of the residual dense network (RDN) and then it been implemented within GAN framework which named RDGAN. The second GAN framework has been built based on the densely sampled super resolution network (DSSR) and we named DSGAN. The used loss function for the training employs the adversarial, mean squared error (MSE) and the perceptual loss derived from the VGG19 model. We optimize the training by using Adam for number of epochs then switching to the SGD optimizer. We validate the frameworks on the proposed dataset of this work and other three remote sensing datasets: the UC Merced, WHU-RS19 and RSSCN7. To validate the frameworks, we use the following image quality assessment metrics: the PSNR and the SSIM on the RGB and the Y channel and the MSE. The RDGAN evaluation values on the proposed dataset were 26.02, 0.704, and 257.70 for PSNR, SSIM and the MSE, respectively, and the DSGAN evaluation on the same dataset yielded 26.13, 0.708 and 251.89 for the PSNR, the SSIM, and the MSE.