Preserving Full Degradation Details for Blind Image Super-Resolution

TL;DR

This work tackles blind image super-resolution by addressing unknown real-world degradations. It introduces ReDSR, a framework that learns degradation representations by re-degrading HR inputs to reproduce the LR, and couples this with an energy distance loss to bound the representations within a predefined distribution, thereby preserving full degradation details. The method includes an encoder, a training-time degrader, and a degradation-aware SR generator, optimized with a re-degradation loss, an energy-distance loss, and a modulated SR loss that weights samples by a confidence derived from reprojection quality. Extensive experiments on synthetic and real degradations show state-of-the-art performance, with ablations confirming the contribution of each component and the advantage of the bounded representation space for generalization.

Abstract

The performance of image super-resolution relies heavily on the accuracy of degradation information, especially under blind settings. Due to absence of true degradation models in real-world scenarios, previous methods learn distinct representations by distinguishing different degradations in a batch. However, the most significant degradation differences may provide shortcuts for the learning of representations such that subtle difference may be discarded. In this paper, we propose an alternative to learn degradation representations through reproducing degraded low-resolution (LR) images. By guiding the degrader to reconstruct input LR images, full degradation information can be encoded into the representations. In addition, we develop an energy distance loss to facilitate the learning of the degradation representations by introducing a bounded constraint. Experiments show that our representations can extract accurate and highly robust degradation information. Moreover, evaluations on both synthetic and real images demonstrate that our ReDSR achieves state-of-the-art performance for the blind SR tasks.

Figures (11)

• Figure 1: An illustration of degradation representation space. (a) Representations learned without a bounded constraint (e.g., DASR wang2021unsupervised). (b) Representations learned by our method with a bounded constraint.
• Figure 2: An overview of our ReDSR framework.
• Figure 3: Visualization results produced by different method on Urban100. Input LR images are produced with only isotropic Gaussian blur kernels.
• Figure 4: Visualization of representations extracted from LR images with different kernel widths $\sigma$. The settings of the model variants are shown in Table \ref{['tab0']}.
• Figure 5: Visualization of representations in different epochs.