Research map / Image generation / Image restoration
CoA: Towards Real Image Dehazing via Compression-and-Adaptation
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 CoA, a compression-and-adaptation scheme that distills a compact dehazing student in the synthetic domain (MoC) and then adapts it to real hazy images via a bilevel EMA-based scheme with a CLIP-based loss (BiA).
Problem
Real image dehazing needs both efficiency for resource-constrained devices and adaptability to diverse real-world haze scenes, which existing methods trade off.
Approach
Divide-and-conquer: first compress a teacher dehazing model into a small student via feature-alignment distillation on synthetic data (MoC), then adapt the student to unlabeled real hazy images using a bilevel formulation with EMA updates and a CLIP-based real-domain loss (BiA).
Key results
- On RTTS, CoA reduces FADE from 1.2117 to 0.9332 for DEA and cuts parameters by 94.06%, 86.74%, and 73.89% for MSBDN, DehazeFormer, and DEA respectively.
- CoA achieves 1.69M parameters and 2.67G FLOPs, the smallest among compared dehazing methods in Table 3.
- On the underwater benchmark, CoA reaches UIQM 4.1284 versus 4.0965 for the best listed baseline DM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add model compression (MoC) to the student model (ablation) | CoA student model · RTTS | FADE: 1.23 -> 1.13 | Helped |
| add model compression (MoC) to the student model (ablation) | CoA student model · RTTS | PM2.5: 135.76 -> 124.43 | Helped |
| add model compression (MoC) to the student model (ablation) | CoA student model · RTTS | BIQME: 0.57 -> 0.58 | Helped |
| use only lower-level modeling instead of bilevel modeling (ablation) | CoA · Real test set | dehazing performance | Hurt |
| use only upper-level modeling with CLIP fine-tuning instead of bilevel modeling (ablation) | CoA · Real test set | visual quality | Hurt |
| add the BiA phase (ablation) | CoA · Real images with complex backgrounds or high noise | FADE | Helped |
| add the BiA phase (ablation) | CoA · Real images with complex backgrounds or high noise | PM2.5 | Helped |
| set EMA coefficient α to 0.95 (ablation) | CoA BiA phase · RTTS | evaluation metrics | Helped |
Surprising findings
- Non-reference metrics such as FADE and BIQME are used as the main evaluation, without paired ground-truth real-image results.
- The method fails to fully remove non-uniform haze on remote sensing images, which the authors attribute to spectral-band effects.
Limitations
- Struggles with remote sensing haze images with non-uniform fog from multi-band imaging constraints.
- Real-domain evaluation relies on no-reference metrics rather than ground truth.
- Compressed student relies on a pre-trained teacher and a hybrid Res2Net/MSBDN architecture.
Experimental setup
Datasets: Haze4K RESIDE THaze RTTS URHI FATTAL NHRW
Benchmarks: RTTS URHI FATTAL NHRW
Metrics: FADE PM2.5 Entropy BIQME UIQM CCF model size (M) FLOPs (G) inference time (ms)
Baselines: SGID Dehamer C2P KANet DEA D4 D4+ RIDCP AirNet WeatherDiff DiffUIR GS
Builds on: MSBDN DehazeFormer DEA CLIP Res2Net
Closest papers by meaning
- Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · similarity 0.88)
- Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired Training (AAAI 2025 · similarity 0.88)
- Iterative Predictor-Critic Code Decoding for Real-World Image Dehazing (CVPR 2025 · similarity 0.87)
- DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 2025 · similarity 0.86)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.86)
- Adaptive Dynamic Dehazing via Instruction-Driven and Task-Feedback Closed-Loop Optimization for Diverse Downstream Task Adaptation (AAAI 2026 · similarity 0.85)
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.