Generative Adversarial Network based Single Image Super Resolution using Deep Residual-in-Residual Dense Network

Abstract

The paper presents a Deep Residual-in-Residual Dense Network that introduces three principal refinements to further improve Single Image Super Resolution. The first refinement involves optimizing the Residual-in-Residual Dense Block to enhance hierarchical feature extraction in high-resolution images. The second advancement replaces the gradient operator with a learned edge extractor for better preservation of high-frequency details. The third improvement incorporates a suitable mid-level insight as a prior from a single layer to multiple layers of the network, ensuring the retention of vital structural intricacies in the output images. Collectively, these changes have enhanced the ability of the network for intricate detail extraction and preservation. Extensive simulations exhibit the efficacy of the presented network in terms of PSNR, SSIM, and LPIPS when compared with other state-of-the-art models.