RealSR

What is RealSR?

The proposed methodology aims to address the limitations of existing super-resolution techniques in handling real-world images. Unlike previous approaches that rely on simplistic bicubic downsampling to generate low-resolution (LR) and high-resolution (HR) training pairs, the authors introduce a novel degradation framework that accounts for real-world blur kernels and noise distributions. This framework enables the acquisition of LR images that share a common domain with actual real-world inputs, thereby overcoming the gap between synthetic and real-world data encountered in prior work Building upon this improved degradation model, the authors develop a real-world super-resolution model focused on achieving better perceptual quality. Extensive experiments on both synthetic noise data and real-world images demonstrate the superiority of the proposed method, as it outperforms state-of-the-art techniques in terms of lower noise levels and enhanced visual quality. Furthermore, the authors' solution was the winner of the NTIRE 2020 Challenge on Real-World Super-Resolution, significantly outperforming other competitors by a wide margin

Highlights

• Novel degradation framework that estimates real-world blur kernels and noise distributions to generate LR images aligned with the target domain
• Real-world super-resolution model designed for improved perceptual quality
• Outperforms state-of-the-art methods on both synthetic and real-world datasets
• Winner of the NTIRE 2020 Challenge on Real-World Super-Resolution