{Junmo Kim Yooshin Cho JEONGHYO HA Youngsoo Kim}

Abstract

Blind super-resolution (blind-SR) is an important task in the field of computer vision and hasvarious applications in real-world. Blur kernel estimation is the main element of blind-SR along with theadaptive SR networks and a more accurately estimated kernel guarantees a better performance. Recently,generative adversarial networks (GANs), comparing recurrence patches across scales, have been the mostsuccessful unsupervised kernel estimation methods. However, they still involve several problems. ① Theirsharpness discrimination ability has been noted as being too weak, causing them to focus more on patternshapes than sharpness. ② In some cases, kernel correction processes were omitted; however, these areessential because the optimally generated kernel may be narrower than a point spread function (PSF)except when the PSF is ideal low-pass filter. ③ Previous studies also did not consider that GANs areaffected by the thickness of edges as well as PSF. Thus, in this paper, 1) we propose a degradation andranking comparison process designed to induce GAN models to became sensitive to image sharpness,and 2) propose a scale-free kernel correction technique using Gaussian kernel approximation including athickness parameter. To improve the kernel accuracy further, we 3) propose a combination model of theproposed GAN and DIP(deep image prior) for more supervision, and designed a kernel correction networkto propagate gradients through developed correction method. Several experiments demonstrate that ourmethods enhanced the l2 error and the shape of the kernel significantly. In addition, by combining withordinary blind-SR algorithms, the best reconstruction accuracy was achieved among unsupervised blur kernelestimation methods.