Research map / Image generation / Image restoration
PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world 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 PRISM, a physically unfolded proximal dehazing network (PSAR) with online non-uniform haze synthesis and a quality-gated, teacher-audited self-distillation adaptation (SSDA) for real-world dehazing.
Problem
Real-world image dehazing under non-uniform haze, spatially varying color shifts, and scarcity of paired real hazy-clean data, which creates a synthetic-to-real domain gap.
Approach
Unfolds closed-form proximal updates of the atmospheric scattering model with spatially varying atmospheric light and learnable transmission/radiance refinements, combined with online non-uniform haze synthesis and self-distillation that only accepts teacher pseudo-labels passing an NR-IQA quality gate, plus a teacher-based residual-haze audit using estimated transmission.
Key results
- On RTTS, PRISM reports FADE 0.470 (vs. HazeFlow 0.583), PAQ2PIQ 74.05, and CLIPIQA 0.456, the best in all three metrics.
- On Fattal's dataset, PRISM reports FADE 0.225, PAQ2PIQ 76.86, and CLIPIQA 0.590, best among compared methods.
- On paired NH-HAZE, PSAR reports PSNR 20.10 and SSIM 0.688, above DehazeDDPM (19.44 / 0.627).
Surprising findings
- Removing the PGSR term raises CLIPIQA slightly (0.463 vs 0.456) while FADE worsens notably, illustrating a trade-off between texture naturalness and haze removal.
- Adding learnable refinement to the atmospheric light branch (w/ f_A) degrades FADE and CLIPIQA, supporting keeping A as a closed-form proximal update.
Limitations
- Formulation is tied to haze-type atmospheric scattering and not studied for snow, rain, underwater or low-light scenes.
- Failure cases remain near object boundaries with extremely dense haze.
- Existing NR-IQA metrics may not fully reflect color fidelity, halo suppression, and structural faithfulness.
Experimental setup
Datasets: RIDCP RESIDE-URHI RTTS Fattal's dataset NH-HAZE
Benchmarks: RTTS Fattal's dataset NH-HAZE
Metrics: FADE PAQ2PIQ CLIPIQA PSNR SSIM NIMA MUSIQ ManIQA LIQE
Baselines: PDN MSBDN Dehamer DAD PSD D4 RIDCP CORUN CoA HazeFlow DGUNet
Builds on: atmospheric scattering model RIDCP CORUN PDN LPIPS
Closest papers by meaning
- Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · similarity 0.89)
- Real-world Image Dehazing with Coherence-based Pseudo Labeling and Cooperative Unfolding Network (NeurIPS 2024 · similarity 0.87)
- PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.87)
- Depth-Centric Dehazing and Depth-Estimation from Real-World Hazy Driving Video (AAAI 2025 · similarity 0.87)
- Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images (CVPR 2025 · similarity 0.87)
- PEIE: Physics Embedded Illumination Estimation for Adaptive Dehazing (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.