Real-world super-resolution based on iterative frequency domain degradation model.

Single-image super-resolution (SR) methods often encounter difficulties when applied to real-world images because of deviations between the degradation model and real-world degradation distribution. Moreover, SR methods suffer from insufficient generalization when faced with diverse degradation distributions. Recent studies have attempted to adopt more complex degradation models to simulate real-world degradation distributions, thereby improving the reconstruction quality and generalization ability of SR methods. However, these methods tend to produce over-smoothed results lacking fine-grained details. Furthermore, these methods neglect the disparity between synthetic and real-world images in the frequency domain. To improve the generalization ability and reconstruction quality of real-world SR methods, this study proposes a more practical degradation model. The proposed model employs aliasing estimation and an iterative frequency domain degradation algorithm to narrow the frequency domain gap between synthetic and real-world images, thus expanding the latent degradation distribution. The proposed model considers the impact of various factors such as camera shake, defocus, and image compression, effectively simulating the complex and diverse degradation processes in the real world. Experiments demonstrate that combining the proposed degradation model with existing SR methods can improve the image perceptual quality on multiple test benchmarks with different degradation distributions. [ABSTRACT FROM AUTHOR]