Research map / Image generation / Image restoration
PEIE: Physics Embedded Illumination Estimation for Adaptive Dehazing
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 PEIE, a physics-embedded illumination estimation method for real-world image dehazing that introduces the Illumination-Adaptive Scattering Model (IASM), a synthesis pipeline built on it, and an Illumination-Adaptive Dehazing Unit inside a U-Net (IADNet).
Problem
Single image dehazing networks trained on ASM-based synthetic data generalize poorly to real-world hazy images because the Atmospheric Scattering Model cannot represent non-uniform illumination and glow, and paired real data is scarce.
Approach
Replaces the global atmospheric light in the ASM with local incident illumination and local environmental illumination (IASM), uses it to synthesize more realistic hazy pairs from clean images, and designs a dehazing unit whose latent-space operations mirror the IASM with multi-scale pooling-based illumination estimates.
Key results
- On RTTS, PEIE achieves NIQE 3.753 and BRISQUE 11.762, the best among compared methods (RIDCP BRISQUE 15.137).
- In the human subjective study on HSTS, PEIE scores 4.36, the highest among methods (RIDCP 4.02, PSD 3.96).
- NIMA on RTTS is 4.856, second only to RIDCP (4.919); FADE is 0.950, ranked third.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PEIE data synthesis instead of OTS (ablation) | IADNet · RTTS | NIQE: 4.962 -> 3.753 | Helped |
| use PEIE data synthesis instead of RIDCP synthesis (ablation) | IADNet · RTTS | NIQE: 4.020 -> 3.753 | Helped |
| add environmental illumination simulation to incident illumination adjustment (ablation) | PEIE synthesis pipeline · RTTS | NIQE: 3.706 -> 3.753 | Hurt |
| add incident illumination adjustment to environmental illumination simulation (ablation) | PEIE synthesis pipeline · RTTS | NIQE: 4.333 -> 3.753 | Helped |
| replace FDU with IDU (ablation) | IADNet · RTTS | NIQE: 4.504 -> 3.753 | Helped |
| replace PDU with IDU (ablation) | IADNet · RTTS | NIQE: 3.796 -> 3.753 | Helped |
| use local rather than global features for LI (ablation) | IADNet IDU · RTTS | NIQE: 3.894 -> 3.753 | Helped |
| use local rather than global features for LE (ablation) | IADNet IDU · RTTS | NIQE: 4.087 -> 3.753 | Helped |
| use different-scale rather than same-scale local features for the illumination terms (ablation) | IADNet IDU · RTTS | NIQE: 3.928 -> 3.753 | Helped |
| use finer-scale features for LI rather than LE (ablation) | IADNet IDU · RTTS | NIQE: 3.838 -> 3.753 | Helped |
Surprising findings
- FADE favors darker images, so methods like D4 and SLP score better on it despite visually darker outputs.
Limitations
- The synthesis pipeline cannot cover the full real-world degradation space, as acknowledged by the authors.
- Evaluation relies on no-reference metrics and a small 10-image human study since real hazy ground truth is unavailable.
- Environmental illumination is approximated with random values and a blur-based glow term rather than physically derived.
Experimental setup
Datasets: RIDCP clean images (500) RESIDE RTTS Fattal's dataset RESIDE HSTS RESIDE OTS
Benchmarks: RTTS Fattal's dataset HSTS
Metrics: FADE NIQE BRISQUE NIMA human subjective score (1-5)
Baselines: IDE SLP DAD PSD D4 RIDCP FFANet C2PNet DehazeFormer OTS
Builds on: Atmospheric Scattering Model RIDCP FDU PDU Retinexformer
Closest papers by meaning
- Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · similarity 0.87)
- PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing (NeurIPS 2026 · similarity 0.86)
- Depth-Centric Dehazing and Depth-Estimation from Real-World Hazy Driving Video (AAAI 2025 · similarity 0.86)
- PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.85)
- Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images (CVPR 2025 · similarity 0.85)
- Dehaze-RetinexGAN: Real-World Image Dehazing via Retinex-based Generative Adversarial Network (AAAI 2025 · 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.