Kernel Estimation for super-resolution with Flow-based Kernel Prior

Abstract

Single-Image Super-Resolution methods typically assume that a low-resolution image is degraded from a high-resolution one through "bicubic" kernel convolution followed by downscaling. However, this induces a domain gap between training image datasets and the real scenario's test images, which are down-sampled from the images that underwent convolution with arbitrary unknown kernels. Hence, correct kernel estimation for a given real-world image is necessary for its better super-resolution. One of the kernel estimation methods, KernelGAN(1) locates the input image in the same domain of high-resolution image for accurate estimation. However, using only a low-resolution image cannot fully utilize the high-frequency information in the original image. To increase the estimation accuracy, we adopt a superresolved image for kernel estimation. Also, we use a flow-based kernel prior to getting a reasonable super-resolved image and stabilize the whole estimation process.