Research map / Label-efficient and robust learning / Semi-supervised learning
Real-world Image Dehazing with Coherence-based Pseudo Labeling and Cooperative Unfolding Network
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 CORUN, a cooperative deep unfolding network modeling transmission and scene under the atmospheric scattering model, plus Colabator, an iterative mean-teacher pseudo-labeling framework with an optimal label pool for real-world image dehazing.
Problem
Real-world image dehazing is hard because real haze distributions are poorly modeled and paired real-world training data is scarce, so synthetic-to-real domain gaps limit generalization.
Approach
Unfold proximal gradient descent over both transmission map and scene radiance (cooperative T&S modules with learnable step sizes and global coherence loss), and generate pseudo labels with a mean-teacher whose best-ever outputs are kept in an optimal label pool, weighted by CLIP-based haze density and MUSIQ quality with local coherence certainty maps.
Key results
- CORUN+ with Colabator reaches FADE 0.824, BRISQUE 11.956, NIMA 5.342 on RTTS, beating the second-best RIDCP (BRISQUE 17.293, NIMA 4.965).
- Colabator improves DGUN on RTTS from FADE 1.111 to 0.857 and NIMA 4.813 to 5.190.
- Adding Colabator to C2PNet, FFA-Net, and GDN improves BRISQUE and FADE across all three.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CORUN+ instead of the strongest baseline for each metric | Real-world image dehazing methods · RTTS | FADE: 0.876 -> 0.824 | Helped |
| use CORUN+ instead of RIDCP | Real-world image dehazing methods · RTTS | BRISQUE: 17.293 -> 11.956 | Helped |
| use CORUN+ instead of RIDCP | Real-world image dehazing methods · RTTS | NIMA: 4.965 -> 5.342 | Helped |
| add Colabator to DGUN | DGUN · RTTS | BRISQUE: 25.085 -> 20.731 | Helped |
| add Colabator to CORUN (ablation) | CORUN · RTTS | BRISQUE: 16.541 -> 11.956 | Helped |
| add mean-teacher to Colabator (ablation) | CORUN+ · RTTS | BRISQUE: 15.728 -> 11.956 | Helped |
| add trusted weighting to Colabator (ablation) | CORUN+ · RTTS | BRISQUE: 16.606 -> 11.956 | Helped |
| add optimal label pool to Colabator (ablation) | CORUN+ · RTTS | NIMA: 5.285 -> 5.342 | No clear effect |
| add CPMM (ablation) | CORUN · RTTS | BRISQUE: 38.197 -> 11.956 | Helped |
| use partitioned and full trusted weights instead of only full weights (ablation) | CORUN+ · RTTS | BRISQUE: 13.099 -> 11.956 | Hurt |
| retain strong augmentation (ablation) | Colabator · RTTS | BRISQUE: 12.671 -> 11.956 | Helped |
| remove DA-CLIP (ablation) | Colabator · RTTS | NIMA: 5.358 -> 5.342 | Hurt |
Surprising findings
- Removing strong augmentation or DA-CLIP yields a higher NIMA (5.084 and 5.358) than the full model, though with worse BRISQUE or FADE.
- Increasing unfolding stages beyond 4 degrades results (6 stages has BRISQUE 16.227 versus 11.956 at 4).
Limitations
- Struggles with severely degraded inputs such as strong compression and extreme dense haze, leaving texture details low quality.
- Only addresses dehazing and not other degradations such as deblurring or low-light enhancement.
- Pseudo labels from the teacher can still be erroneous, and the paper reports no compute or runtime efficiency analysis.
Experimental setup
Datasets: RIDCP500 RESIDE URHI RTTS Fattal's dataset
Benchmarks: RTTS Fattal's dataset
Metrics: FADE BRISQUE NIMA user study score mAP
Baselines: PDN MBDN DH DAD PSD D4 RIDCP DGUN
Builds on: PDN DGUN RIDCP mean-teacher DA-CLIP
Closest papers by meaning
- Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · similarity 0.89)
- PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.87)
- PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing (NeurIPS 2026 · similarity 0.87)
- Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images (CVPR 2025 · similarity 0.87)
- Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired Training (AAAI 2025 · similarity 0.87)
- Dehaze-RetinexGAN: Real-World Image Dehazing via Retinex-based Generative Adversarial Network (AAAI 2025 · similarity 0.86)
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.