Research map / Image generation / Image restoration
Noise Modeling in One Hour: Minimizing Preparation Efforts for Self-supervised Low-Light RAW Image Denoising
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 a simplified noise synthesis pipeline for self-supervised low-light RAW denoising that hypothesizes the system gain instead of calibrating it and uses directly sampled dark frames for signal-independent noise, cutting preparation from days to about an hour.
Problem
Noise synthesis for self-supervised low-light RAW denoising requires laborious system gain calibration and signal-independent noise profiling, making deployment at scale costly.
Approach
Hypothesize the system gain as QE x AG with an assumed QE (~40%) instead of calibrating it, since shot-noise variance scales with K without introducing intensity bias; sample signal-independent noise directly from collected dark frames rather than fitting statistical or network-based profiles, and omit high-bit-depth noise recovery.
Key results
- On ELD, the method reaches 46.43 dB PSNR average (SSIM 0.9834), slightly above PNNP's 46.39 dB and above PMN's 45.92 dB.
- On LRID, the method gives 45.08 dB overall PSNR, up to 0.54 dB above PNNP (44.54 dB) as stated.
- Varying system gain K over broad ranges changes PSNR by less than 0.1 dB, and using only 10 dark frames at inference drops overall PSNR by about 0.1 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace calibrated K with hypothesized K on SID and ELD (ablation) | Ours · SID and ELD | PSNR: 40.90 / 0.9487 vs 40.85 / 0.9478; 46.37 / 0.9832 vs 46.43 / 0.9834 | No clear effect |
| use broad-range hypothesized K instead of calibrated K (ablation) | Ours · SID and ELD | PSNR: 40.90 / 0.9487 vs 40.85 / 0.9484; 46.37 / 0.9832 vs 46.33 / 0.9826 | No clear effect |
| randomly sample K ∈(4, 17) instead of calibrated K ∈(8.7, 8.8) (ablation) | Ours · LRID | PSNR: 45.00 / 0.9773 -> 45.09 / 0.9778 | No clear effect |
| calibrate inference dark shading using fewer dark frames (ablation) | Ours · LRID | PSNR: 45.08 with 180 dark frames; 44.98 with 10 dark frames; 44.67 with 2 dark frames | No clear effect |
| reduce training dark frames per ISO from 400 to 10 (ablation) | Ours · SID and ELD | PSNR: SID: 40.85 -> 40.80; ELD: 46.43 -> 46.31 | No clear effect |
| replace direct sensor sampling with Tukey-Lambda noise profiling (ablation) | Ours · SID and ELD | PSNR: SID: 40.85 -> 40.56; ELD: 46.43 -> 45.63 | Hurt |
| replace direct sensor sampling with GMM noise profiling (ablation) | Ours · SID and ELD | PSNR: SID: 40.85 -> 40.54; ELD: 46.43 -> 45.92 | Hurt |
| add HBNR to the noise synthesis pipeline (ablation) | Ours · SID, ELD, and LRID | PSNR: SID: 40.85 -> 40.91; ELD: 46.43 -> 46.38; LRID: 45.08 -> 45.03 | Mixed |
| use Ours rather than PNNP | self-supervised RAW denoising · SID avg. | PSNR: 40.83 -> 40.85 | Helped |
| use Ours rather than PNNP | self-supervised RAW denoising · ELD avg. | PSNR: 46.39 -> 46.43 | Helped |
| use Ours rather than PMN | RAW denoising · LRID Overall avg. | PSNR: 44.97 -> 45.08 | Helped |
Surprising findings
- Self-supervised training with the proposed pipeline outperforms the supervised PMN method on SID, attributed to imperfect pixel alignment in SID noisy-clean pairs.
- High-bit-depth noise recovery (HBNR) gives no clear benefit, with less than 0.1 dB PSNR difference when omitted.
Limitations
- Requires known analog gain (AG); consumer cameras and smartphones hide AG behind ISO values.
- Evaluated only on three datasets from a few sensors (two Sony A7S2 and one Redmi K30 smartphone sensor).
Experimental setup
Datasets: SID ELD LRID LLD calibration dark frames
Benchmarks: SID ELD LRID
Metrics: PSNR SSIM
Baselines: Paired supervised training PMN P-G SFRN ELD NoiseFlow StarLight PNNP LED LLD
Builds on: ELD SFRN PMN LLD PNNP
Closest papers by meaning
- 2-Shots in the Dark: Low-Light Denoising with Minimal Data Acquisition (CVPR 2026 · similarity 0.91)
- Realistic Noise Synthesis with Diffusion Models (AAAI 2025 · similarity 0.87)
- Compress Clean Signal from Noisy Raw Image: A Self-Supervised Approach (ICML 2024 · similarity 0.87)
- APR-RD: Complemental Two Steps for Self-Supervised Real Image Denoising (AAAI 2025 · similarity 0.86)
- Positive2Negative: Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising (CVPR 2025 · similarity 0.86)
- Zero-Shot Noise2Mean: Gap Minimization for Efficient Denoising from a Single Noisy Image (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.