Two Heads Better than One: Dual Degradation Representation for Blind Super-Resolution

TL;DR

The paper tackles blind single-image super-resolution under unknown degradations by proposing a dual-branch degradation extractor that disentangles blur and noise in the wavelet high-frequency domain. It couples these embeddings with a conditional SR network using a normalization codebook purification to produce robust, degradation-aware restorations. The approach combines contrastive learning, SR reconstruction, and embedding regularization to achieve state-of-the-art results on synthetic blind-SR benchmarks and demonstrates strong generalization across datasets and degradation types. This disentangled degradation modeling offers practical benefits for real-world SR where degradations are diverse and unknown.

Abstract

Previous methods have demonstrated remarkable performance in single image super-resolution (SISR) tasks with known and fixed degradation (e.g., bicubic downsampling). However, when the actual degradation deviates from these assumptions, these methods may experience significant declines in performance. In this paper, we propose a Dual Branch Degradation Extractor Network to address the blind SR problem. While some blind SR methods assume noise-free degradation and others do not explicitly consider the presence of noise in the degradation model, our approach predicts two unsupervised degradation embeddings that represent blurry and noisy information. The SR network can then be adapted to blur embedding and noise embedding in distinct ways. Furthermore, we treat the degradation extractor as a regularizer to capitalize on differences between SR and HR images. Extensive experiments on several benchmarks demonstrate our method achieves SOTA performance in the blind SR problem.

Figures (6)

• Figure 1: (a) Analysis in the frequency domain among HR image, LR image, and LR image with noise. Noise degradation wrongly increases the high-frequency components, indicated by the blue arrow. Meanwhile, blur degradation leads to the loss of these high-frequency details, indicated by the orange arrow. (b) and (c) illustrate the concepts of previous UDP methods and our approach.
• Figure 2: An overview of our proposed method.
• Figure 3: Visualization of different degradations using colored dots in three settings. (a) w/ different blur kernels and w/o noise level. (b) w/ different blur kernels and noise level is 10. (c) w/ a fixed kernel and different noise levels. Yellow box means using a unified degradation representation, green box denotes blur representations, and blue box indicates noise representations. Compared with DASR and CDSR, our method can clearly distinguish different degradations in all the test settings (i.e., (a), (b), and (c)).
• Figure 4: Visual comparison on in-domain (first row) and out-domain (second and thrid row) synthetic test data for $\times4$ SR.
• Figure 5: Qualitative comparison on NTIRE2020Track1 (first row) and DIV2KRK (second row).