Orthonormal Convolutions For The Rotation Based Iterative Gaussianization

Over the past few years, blind image super-resolution kernel estimation has become an emerging research topic. Recent SR kernel estimation works KernelGAN and FKP, as well as KernelNet, have shown promising results when applied to real-world SR problems. Among them, KernelNet improves the accuracy of SR kernel estimation while reducing computation time. Despite this, none of these previous studies addressed the estimation of SR kernel for scenarios that incorporate reference higher quality images. Our study presents KernelNet-R, a reference-based SR kernel estimator network. in both synthetically-generated pairs, as well as real-world pairs of images from wide-angle and telephoto camera images, the proposed solution reaches state-of-the-art (SoTa) kernel estimation performance.