Research map / Image generation / Image restoration
Guided Real Image Dehazing Using YCbCr Color Space
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 SGDN, a dual-color-space (RGB guided by YCbCr) real-image dehazing network with a Bi-Color Guidance Bridge and Color Enhancement Module, plus a new 1,758-pair real-world well-aligned haze dataset (RW²AH).
Problem
Learning-based real-world image dehazing relies on RGB features that blur under haze and leaves residual haze, and real hazy/clean paired data is hard to collect due to background misalignment, so supervised training is limited.
Approach
Uses YCbCr features (less affected by haze in chrominance) to guide RGB feature recovery via phase integration in the frequency domain and cross-attention in the spatial domain (BGB), then aggregates YCbCr channel information to enhance RGB color perception (CEM), with a shared encoder and asymmetric encoder-decoder.
Key results
- On Real-World Smoke, SGDN reaches 23.41 PSNR and 0.790 SSIM, versus 22.59 PSNR for the next-best DAE-Net on PSNR.
- On the RW2AH dataset, SGDN achieves 22.26 PSNR and 0.668 SSIM, with FADE 0.400 and NIQE 5.008, the best among compared methods.
- Ablation on RW2AH shows BGB adds 2.73 dB PSNR and CEM adds 2.26 dB over the baseline (per the paper's text).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SGDN instead of the strongest listed baseline | real-world smoke dehazing · Real-world Smoke | PSNR: 22.59 -> 23.41 | Helped |
| use SGDN instead of the strongest listed baseline | real-world smoke dehazing · Real-world Smoke | SSIM: 0.780 -> 0.790 | Helped |
| use SGDN instead of the strongest listed baseline | RW²AH dehazing · RW²AH | PSNR: 21.39 -> 22.26 | Helped |
| use SGDN instead of the strongest listed baseline | RW²AH dehazing · RW²AH | SSIM: 0.627 -> 0.668 | Helped |
| use SGDN instead of the strongest listed baseline | RTTS dehazing · RTTS | NIQE: 5.0843 -> 5.2114 | Hurt |
| use both RGB and YCbCr rather than RGB only (ablation) | SGDN · RW²AH | PSNR: 18.44 -> 22.26 | Helped |
| use both RGB and YCbCr rather than YCbCr only (ablation) | SGDN · RW²AH | PSNR: 19.02 -> 22.26 | Helped |
| use YCbCr rather than HSV as the guidance color space (ablation) | RGB+color-space SGDN · RW²AH | PSNR: 21.17 -> 22.26 | Helped |
| use YCbCr rather than YUV as the guidance color space (ablation) | RGB+color-space SGDN · RW²AH | PSNR: 21.49 -> 22.26 | Helped |
| add BGB (ablation) | RGB+YCbCr baseline without BGB or CEM · RW²AH | PSNR: 19.10 -> 21.83 | Helped |
| add CEM (ablation) | RGB+YCbCr baseline without BGB or CEM · RW²AH | PSNR: 19.10 -> 21.36 | Helped |
| add IAM to the BGB baseline (ablation) | SGDN without CEM · RW²AH | PSNR: 19.10 -> 21.60 | Helped |
Surprising findings
- HSV and YUV combinations with RGB were compared, and the paper notes that single-color-space models underperform combined ones, with RGB+YUV gaining over RGB+HSV.
- On synthetic SOTS-Outdoor, SGDN's PSNR (36.22) trails several baselines such as C2PNet and DCMPNet, which the paper attributes to its real-world focus.
Limitations
- Struggles with dense, non-uniform fog patches, sometimes producing artifacts.
- Performance can be inconsistent across scenes with varying fog density and distribution.
- RW2AH pairs are not perfectly aligned, only sufficiently aligned for supervised learning.
Experimental setup
Datasets: RW2AH Real-World Smoke (I-HAZE, O-Haze, NH-Haze) BeDDE MRFID RESIDE/RTTS SOTS
Benchmarks: RW2AH Real-World Smoke RTTS SOTS-Indoor SOTS-Outdoor
Metrics: PSNR SSIM FADE NIQE mAP
Baselines: DCP RefineDNet CDD-GAN D4 PSD YOLY FSDGN DehazeFormer C2PNet RIDCP MB-Taylor NSDNet
Builds on: Atmospheric Scattering Model Large Kernel Attention MB-Taylor DehazeFormer
Closest papers by meaning
- Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images (CVPR 2025 · similarity 0.86)
- Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · similarity 0.86)
- Real-world Image Dehazing with Coherence-based Pseudo Labeling and Cooperative Unfolding Network (NeurIPS 2024 · similarity 0.85)
- PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.85)
- Iterative Predictor-Critic Code Decoding for Real-World Image Dehazing (CVPR 2025 · similarity 0.85)
- PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing (NeurIPS 2026 · similarity 0.84)
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.