Research map / Image generation / Image restoration
Specularity Factorization for Low-Light Enhancement
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 RSFNet, a model-driven recursive specularity factorization network unrolled from ADMM with ~200 learnable parameters, used for zero-reference low-light enhancement and as a prior for dehazing, deraining and deblurring.
Problem
Low-light image enhancement needs a way to model spatially varying illumination degradation; existing factorizations (Retinex, intensity, frequency) are either multiplicative, hard-segmented, or fixed in number, and zero-reference methods are often heavy or less interpretable.
Approach
Decompose an image into K additive soft specular factors by recursively solving a sparsity-regularized ADMM problem while relaxing the sparsity threshold per factor, and unroll the ADMM iterations into a network whose only learnable parameters are a few scalars (thresholds and step sizes), followed by a small UNet fusion network trained with zero-reference losses.
Key results
- On LOLv1, RSFNet reaches 22.17 PSNR-Y and 0.860 SSIM-Y, versus 19.85 for the best prior zero-reference method (PNet).
- Mean across four LOL datasets: 21.16 PSNR-Y and 0.854 SSIM-Y, the best among compared methods.
- Adding factors to AirNet improves dehazing PSNR from 23.18 to 24.96 (uni-task), deraining from 34.90 to 36.19, and deblurring from 26.42 to 27.29.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add RSF factors as additional input | AirNet · RESIDE dehazing | PSNR: 23.18 -> 24.96 | Helped |
| Add RSF factors as additional input | AirNet · Rain100L deraining | PSNR: 34.90 -> 36.19 | Helped |
| Add RSF factors as additional input | AirNet · GoPro deblurring | PSNR: 26.42 -> 27.29 | Helped |
| Add exposure loss (ablation) | RSFNet · LOLv1 | PSNRy: 8.12 -> 22.17 | Helped |
| Add color loss (ablation) | RSFNet · LOLv1 | PSNRy: 16.05 -> 22.17 | Helped |
| Add smoothing loss (ablation) | RSFNet · LOLv1 | PSNRy: 20.13 -> 22.17 | Helped |
| Add final denoising step (ablation) | RSFNet · LOLv1 | PSNRy: 19.51 -> 22.17 | Helped |
| Use the fusion network instead of weighted running-average fusion (ablation) | RSFNet · LOLv1 | PSNRy: 19.32 -> 22.17 | Helped |
| Vary the number of factors; choose K=5 (ablation) | RSFNet · LOLv1 | PSNR and SSIM | Mixed |
Surprising findings
- Removing the fusion network entirely and using a simple running average of factors still performs well (PSNR 19.32), with only 45 parameters used at inference.
Limitations
- Sensitive to initialization of the underlying unrolled algorithm; uses dataset mean as heuristic initialization.
- Does not adapt dynamically to each input in the current version.
Experimental setup
Datasets: LOLv1 LOLv2-real LOLv2-synthetic VE-LOL DICM LIME MEF NPE VV RESIDE Rain100L GoPro
Benchmarks: LOLv1 LOLv2-real LOLv2-synthetic VE-LOL DICM LIME MEF NPE VV RESIDE Rain100L GoPro
Metrics: PSNR SSIM LPIPS NIQE LOE
Baselines: LIME DUAL SDD ECNet ZeroDCE ZeroDCE++ RUAS SCI PNet GDP
Builds on: ALISTA Zero-DCE AirNet ADMM LISTA
Closest papers by meaning
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.85)
- Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (CVPR 2024 · similarity 0.85)
- Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration (CVPR 2025 · similarity 0.84)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.84)
- Dispel Darkness for Better Fusion: A Controllable Visual Enhancer based on Cross-modal Conditional Adversarial Learning (CVPR 2024 · similarity 0.84)
- SpecNeRF: Gaussian Directional Encoding for Specular Reflections (CVPR 2024 · 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.