Reference Based Image Super Resolution

Reference-based image super-resolution (Ref-SR) aims to enhance low-resolution images by leveraging information from a corresponding high-resolution reference image. Current research heavily utilizes deep learning models, particularly transformer networks and those incorporating attention mechanisms, to effectively match and transfer texture details from the reference to improve the low-resolution image. This approach addresses limitations of single-image super-resolution by providing additional contextual information, leading to improved visual quality and quantitative metrics (like PSNR and SSIM) across various applications, including remote sensing and electron microscopy. Furthermore, extensions to video super-resolution are actively being explored, utilizing multiple camera inputs for enhanced spatio-temporal resolution.