A Network Combining CNN and Transformer for Blind Image Super-Resolution

Blind super-resolution (SR) requires not only estimating blur kernel, but also super-resolving low-resolution image based on estimated blur kernel. Most blind SR methods use convolutional neural networks (CNNs) for kernel estimation, which cannot exploit long-range dependency within image domain, thus failing to predict blur kernel accurately. To address this issue, we propose a network combining CNN and transformer named NCCT for kernel estimation. By modeling local and non-local image priors simultaneously, NCCT outperforms other blind SR methods in terms of kernel estimation accuracy. Moreover, we design a network module named RRFDB for constructing lightweight blind SR network, which runs faster and achieves comparative SR performance with fewer parameters compared with other state-of-the-art blind SR methods.

I. Introduction

Single image super-resolution (SISR) aims at restoring a high-resolution (HR) image from a given low-resolution (LR) counterpart, which has attracted more and more attention due to its wide application. Since the first SR method SRCNN [1] was proposed, various deep SR networks have been developed to achieve outstanding performances. However, most SR methods assume that the degradation process from HR image to LR image is determined or predefined (e.g. bicubicly downsampling), which does not hold true in real situations. The SR performances drop severely when the image degradation is different from the hypothetical one. Blind SR methods are specifically introduced to upscale LR images when the degradation process is unknown.