Research map / Image generation / Image restoration
Efficient Scene Recovery Using Luminous Flux Prior
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 Luminous Flux Prior (LFP), a training-free scene recovery method that estimates transmission from the rate of change of luminous flux, handling haze, sandstorm, and underwater degradation.
Problem
Restoring images degraded by adverse weather (haze, sandstorms, underwater) where physical priors fail under violated assumptions and learning-based dehazing generalizes poorly and is computationally costly.
Approach
Derives an analytical transmission formula from the Bouguer-Lambert law and Retinex-style ratios so that transmission depends only on luminous flux F and its local rate of change ΔF (eliminating depth and attenuation coefficient), computed with sliding windows and then used in the standard scattering inversion with guided filtering.
Key results
- On RESIDE-IN, LFP reaches 19.76 PSNR / 0.869 SSIM versus 16.88 / 0.791 for the best prior baseline BCCR, and it leads on all four dehazing test sets in Table 2.
- LFP runs in 0.863 s on CPU at 4K resolution versus 15.11 s for DCP, and 0.019 s on GPU at 4K, reported as up to 34x faster than the compared GPU models.
- The transmission-estimation KLD for ΔF ranges 0.12-0.29, versus 0.43-1.90 for uniform or truncated-normal baselines, which the authors use to argue ΔF strongly correlates with transmission.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LFP instead of BCCR | Scene recovery · RESIDE-IN | PSNR: 16.88 -> 19.76 | Helped |
| Use LFP instead of DCP | Scene recovery · RESIDE-IN | SSIM: 0.818 -> 0.869 | Helped |
| Use LFP instead of DCP | Scene recovery · RESIDE-OUT | PSNR: 19.13 -> 21.05 | Helped |
| Use LFP instead of DCP | Scene recovery · RESIDE-OUT | SSIM: 0.815 -> 0.861 | Helped |
| Use LFP instead of DCP | Scene recovery · Haze-4K | PSNR: 14.01 -> 18.29 | Helped |
| Use LFP instead of DCP | Scene recovery · Haze-4K | SSIM: 0.760 -> 0.873 | Helped |
| Use LFP instead of BCCR | Scene recovery · RS-Haze | PSNR: 18.84 -> 18.24 | Hurt |
| Use LFP instead of BCCR | Scene recovery · RS-Haze | SSIM: 0.747 -> 0.786 | Helped |
| Use LFP instead of ROP | Scene recovery · 4k CPU | runtime (seconds): 2.515 -> 0.863 | Helped |
| Use LFP instead of ROP | Scene recovery · 4k GPU | runtime (seconds): 0.894 -> 0.019 | Helped |
| Increase the sliding-window size Ω (ablation) | LFP · Dense haze or scenes with greater depth of view | image quality | Helped |
| Increase parameter k (ablation) | LFP · Scenes with greater depth of view or denser haze | image quality | Helped |
Limitations
- Relies on empirical choice of k=0.5 and window size 7, which the paper tunes by experience and shows sensitivity to in ablation.
- Physical derivation uses approximations (e.g., simplifying the exponential to a linear term and treating incident flux as constant), and the correlation is validated via KLD on a heuristic transmission map rather than ground-truth depth.
Experimental setup
Datasets: RESIDE SIRB UIEBD Haze4K RS-Haze
Benchmarks: RESIDE Haze4K RS-Haze SIRB UIEBD
Metrics: PSNR SSIM run-time (seconds)
Baselines: DCP BCCR FVR HRDCP ROP CAP LDCP MSCNN AOD-Net gUNet-T FFA MSBDN
Builds on: Dark Channel Prior Retinex Bouguer-Lambert law guided filter
Closest papers by meaning
- ESR-NeRF: Emissive Source Reconstruction Using LDR Multi-view Images (CVPR 2024 · similarity 0.84)
- Acquire and then Adapt: Squeezing out Text-to-Image Model for Image Restoration (CVPR 2025 · similarity 0.84)
- UP-Restorer: When Unrolling Meets Prompts for Unified Image Restoration (AAAI 2025 · similarity 0.83)
- Real-World Adverse Weather Image Restoration via Dual-Level Reinforcement Learning with High-Quality Cold Start (NeurIPS 2025 · similarity 0.83)
- Robust Inverse Graphics via Probabilistic Inference (ICML 2024 · similarity 0.83)
- Continuous Adverse Weather Removal via Degradation-Aware Distillation (CVPR 2025 · similarity 0.83)
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.