Single Frame Super resolution with Deep Residual Network - Generative Adversarial Networks

Abstract

Develop and evaluate a deep learning-based method to enhance satellite image resolution, addressing challenges posed by motion, imaging blur, and noise without modifying existing optical systems. The study utilized an enhanced super-resolution generative adversarial network (SRGAN) with ResNet-50 as the generator and a modified VGG-19 in the discriminator. The model was trained on remote sensing images from LISS imagery and compared with VDSR, SRGAN, and ESRGAN methods using SSIM and PSNR as evaluation metrics. Utilizing an enhanced SRGAN with ResNet-50 and modified VGG-19 significantly improves satellite image resolution. The proposed method consistently outperformed traditional CNN and GAN-based super-resolution techniques. Across three test datasets, the method achieved SSIM scores as high as 0.862 and PSNR scores of 33.256, 32.886, and 34.885, demonstrating its superior ability to preserve image properties and enhance resolution. The incorporation of perceptual loss alongside pixel loss contributed to improved visual quality, making the approach particularly effective in maintaining fine details and naturalistic high-frequency characteristics.