Frequency Generation for Real-World Image Super-Resolution

Single-image super-resolution (SISR) is essential for improving the extraction of useful information from images captured in the real world. Most existing super-resolution methods generally assume that low-resolution (LR) images are generated from high-resolution (HR) images through a known degradation model, such as bicubic downsampling. As a result, these methods do not exhibit favorable performance on real-world images with complex authentic degradations, significantly limiting their practicality, especially in application scenarios where image authenticity is strictly enforced. In this paper, We design a frequency separation network (FSN) to separate low-frequency information and generate high-frequency information, which can reconstruct high-resolution real-world images quickly and accurately. We proposed the various Gaussian filters as the frequency separation (FS) module to gradually separate the frequencies and route them to their respective feature extraction modules. Subsequently, we aggregate all the different frequency features using the adaptive feature fusion (AFF) module to generate the HR image. Therefore, FSN can focus on high-frequency information to restore image details and ensure stable restoration of important information, such as object contours, without generating false texture details. Extensive experiments demonstrated that our FSN achieves consistently superior visual quality and generalization ability with more realistic and natural textures in various scenarios.