Research map / Image generation / Super-resolution
Unsupervised Degradation Representation Aware Transform for Real-World Blind Image Super-Resolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DRAT, a blind super-resolution network whose dual-encoder (DEDR) subtracts content features from LR features to isolate degradation representations, which then modulate features via a DRT module inside global degradation aware blocks.
Problem
Blind image super-resolution with unknown degradations: existing degradation estimation methods (supervised kernel estimation or unsupervised degradation representation) are confounded by image content, leading to inaccurate degradation representations.
Approach
A dual-encoder trained with contrastive learning (auxiliary encoder on HR patches for content, blur encoder on LR images for mixed content+degradation) whose difference yields content-free degradation representations; these are used via SFT-style affine transforms inside global degradation aware blocks that apply spatial and channel attention.
Key results
- On DIV2KRK x2, DRAT reaches 33.31 dB PSNR vs 32.96 dB for DAN-Real, and 32.46 dB for DSSR.
- On DRealSR x4, DRAT reaches 31.00 dB / 0.8737 SSIM, the best among compared methods, with 9.59M parameters versus 13.66M for DCLS.
- Ablation on DIV2KRK x2: removing DEDR drops PSNR from 33.31 to 32.82 dB; removing DRT drops it to 32.88 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GDA blocks (ablation) | DRAT · DIV2KRK, ×2 | PSNR: 32.69 -> 33.31 | Helped |
| Use DEDR instead of a conventional blur encoder (ablation) | DRAT · DIV2KRK, ×2 | PSNR: 32.82 -> 33.31 | Helped |
| Add DRT module (ablation) | DRAT · DIV2KRK, ×2 | PSNR: 32.88 -> 33.31 | Helped |
| Use DRAT instead of DAN-Real | blind SR models · DIV2KRK, anisotropic Gaussian, ×2 | PSNR: 32.96 -> 33.31 | Helped |
| Use DRAT instead of DCLS | blind SR models · DIV2KRK, anisotropic Gaussian, ×4 | PSNR: 28.99 -> 29.21 | Helped |
Surprising findings
- Table 3 reports ResShift at 25.45 dB on RealSR-Nikon, far below its diffusion-based reputation, and DAN-Real trails DRAT on several datasets while being smaller.
Limitations
- Degradation representation is trained separately from SR network via contrastive learning, and the auxiliary encoder assumes content can be captured from HR patches.
- Evaluation relies on synthetic Gaussian kernel degradations for most benchmarks; real-world evaluation is limited to a few datasets.
- Authors note the approach has not been extended to complex scenarios such as video.
Experimental setup
Datasets: DIV2K Flickr2K Set5 Set14 BSD100 Urban100 Manga109 DIV2KRK RealSR DRealSR ImageNet-Test
Benchmarks: Set5 Set14 BSD100 Urban100 Manga109 DIV2KRK RealSR DRealSR ImageNet-Test
Metrics: PSNR SSIM
Baselines: IKC DANv1 DANv2 DCLS DSSR CDCN MRDA DASR DAN-Real HAT ResShift
Builds on: MoCo SFT DASR MRDA DCLS
Closest papers by meaning
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.91)
- DORNet: A Degradation Oriented and Regularized Network for Blind Depth Super-Resolution (CVPR 2025 · similarity 0.89)
- DDR: Exploiting Deep Degradation Response as Flexible Image Descriptor (NeurIPS 2024 · similarity 0.89)
- CDFormer: When Degradation Prediction Embraces Diffusion Model for Blind Image Super-Resolution (CVPR 2024 · similarity 0.89)
- Low-Res Leads the Way: Improving Generalization for Super-Resolution by Self-Supervised Learning (CVPR 2024 · similarity 0.89)
- SEAL: A Framework for Systematic Evaluation of Real-World Super-Resolution (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.