Learning the Degradation Distribution for Blind Image Super-Resolution

TL;DR

This work tackles blind image super-resolution by treating degradations as stochastic rather than fixed. It introduces a probabilistic degradation model (PDM) that learns the distribution of degradations through two modules: a kernel module that maps a latent variable to blur kernels and a noise module that maps a latent variable (and optionally image content) to noise. The degradation process is constrained as a linear operation, enabling end-to-end adversarial training to align generated LR images with real test data, while decoupling degradations from image content. The proposed PDM framework (and its SR-integrated variant PDM-SR/PDM-SRGAN) achieves state-of-the-art performance on mainstream NTIRE datasets, with strong improvements in perceptual and structural metrics and clear evidence that learned degradations cover a broader, more realistic range than deterministic counterparts. The approach provides a flexible, priors-friendly pathway to bridge training-test gaps in blind SR and can be extended to include additional corruptions such as JPEG.

Abstract

Synthetic high-resolution (HR) \& low-resolution (LR) pairs are widely used in existing super-resolution (SR) methods. To avoid the domain gap between synthetic and test images, most previous methods try to adaptively learn the synthesizing (degrading) process via a deterministic model. However, some degradations in real scenarios are stochastic and cannot be determined by the content of the image. These deterministic models may fail to model the random factors and content-independent parts of degradations, which will limit the performance of the following SR models. In this paper, we propose a probabilistic degradation model (PDM), which studies the degradation $\mathbf{D}$ as a random variable, and learns its distribution by modeling the mapping from a priori random variable $\mathbf{z}$ to $\mathbf{D}$. Compared with previous deterministic degradation models, PDM could model more diverse degradations and generate HR-LR pairs that may better cover the various degradations of test images, and thus prevent the SR model from over-fitting to specific ones. Extensive experiments have demonstrated that our degradation model can help the SR model achieve better performance on different datasets. The source codes are released at \url{git@github.com:greatlog/UnpairedSR.git}.

Figures (7)

• Figure 1: Degradation-learning-based methods usually use adversarial training to encourage a degradation model to produce synthetic LR images in the same domain with the test images.
• Figure 2: The details of our probabilistic degradation model (PDM)). The discriminator is left out for more intuitive description.
• Figure 3: Visual comparisons on image 0827x4 of 2017Track2 (top) and image 0860x4m of 2018Track2 (bottom). The PSNR / SSIM / NIQE scores are denoted below the SR results respectively. Red: the best. Blue: the second best. Zoom in for best view.
• Figure 4: Visual comparisons on image 0010 (top) and image 0097 (bottom) of 2020Track2. Zoom in for best view.
• Figure 5: Visual comparisons on image 0804 of 2020Track1. Red: the best. Blue: the second best. Zoom in for best view.