Towards Real-world Continuous Super-Resolution: Benchmark and Method

Continuous Super-Resolution (CSR) has garnered considerable popularity for its capability to reconstruct high-resolution (HR) images from low-resolution (LR) inputs at various scales, thereby holding significant practical value in real-world applications. However, the existing studies have relied solely on synthetic datasets due to the scarcity of real-world continuous datasets. In this paper, we establish a real-world continuous super-resolution dataset, i.e., RealCSR, aiming to explore the realistic continuous image representation in real-world scenarios. Additionally, we introduce a Codebook-embedded Attention Network (CANet) for Real-world CSR to investigate improved information gain and distribution in the presence of real-world image degradation. We employ a codebook-embedded attention module to effectively capture and aggregate local and global information combined with codebook information. Specifically, CANet utilizes a multi-scale SR codebook and a dilated neighborhood attention mechanism to effectively capture and aggregate both local and global information. This enriched information is then fed into a scale-aware implicit fusion module modulated by a scale-guided degradation map. Experimental results demonstrate that our CANet outperforms existing state-of-the-art CSR methods and also achieves visually pleasant SR image predictions with realistic, accurate details. Our dataset, codes, and models are publicly available at https://github.com/doooooithey/CANet.

I. Introduction

Single Image Super-Resolution (SISR) has made substantial progress in generating HR images from LR inputs [1]–[5]. Despite their remarkable performance, a key limitation characterizes these methods: they are designed to execute superresolution at a single, fixed scale factor. This constraint implies that a model’s inability to adjust to varying scale factors necessitates the training of individual models for each specific scale factor. This not only amplifies computational demands but also results in a considerable increase in device memory usage. Consequently, the fixed scale factor constraint presents a significant challenge to the wider applicability and efficiency of SR models.