Kgsr: A Kernel Guided Network for Real-World Blind Super-Resolution

Abstract

Deep learning has dominated super resolution (SR) field due to its remarkable reconstruction performance in the past few years. However, the existing methods typically assume the acquisition of the low-resolution (LR) image from the unknown high-resolution (HR) image by a predetermined kernel (\eg, Bicubic downscaling). This is rarely a special case in real-world LR images, the degradation processes in real applications are various,  complicated and unknown. When the assumed downscaling kernel deviates from the true one, the results of the state-of-the-art methods significantly deteriorate. In this paper, we propose a kernel guided network for real-world blind SR, named KGSR, which cast the original unsupervised and blind problem into a supervised learning and non-blind problem.  Specifically, KGSR learns two networks (\ie, Upscaling and Downscaling) using only patches of the input test image. On the one hand, since the SR-kernel has the cross-scale recurrence property inside a single image, the Downscaling network learns the image-specific degradation process based on a generative adversarial network. Thus, the Downscaling network can obtain a produce a downsampled version of the LR test image where the acquisition process is unknown or non-ideal. In addition, we use a specific discriminator to force the Downscaling network to focus on the orientation of kernels. On the other hand, the accurate blur kernel could bring better  performance. Based on the guidance of the correct image-specific SR-kernel learned from the Downscaling network, and the downsampled version of the LR input, the Upscaling network can generate a high-quality HR image from LR image. In Upscaling network, we further propose an effective module to use the learned image-specific SR-kernel. KGSR is fully unsupervised, but can generate the image-specific SR-kernel and high-quality HR image, simultaneously. Extensive experiments on standard benchmarks demonstrate the effectiveness of the proposed method against the state-of-the-art methods. Furthermore, the proposed method can provide visually favorable SR results with a shorter runtime on real-world LR images.

Keywords: Blind Super-Resolution, Kernel Estimation, Discriminator, Unsupervised Learning, Non-ideal Degradation