Research map / Image generation / Image restoration
Zero-Shot Blind-spot Image Denoising via Implicit Neural Sampling
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 zero-shot blind-spot denoiser that uses distant visible pixels plus INR-estimated values of neighboring invisible pixels, backed by a risk analysis of linear blind-spot denoisers under noise and pixel correlation.
Problem
Blind-spot denoising at test time fails on real-world noise with strong local spatial correlation, because reducing noise correlation (via large masks/strides) also weakens the pixel correlations needed to predict masked pixels.
Approach
Use a subset of the mask complement that excludes pixels adjacent to visible ones (via dilation), and fill in the excluded neighboring pixels with an implicit neural representation (SIREN) fitted on visible pixels, jointly trained with a denoising U-Net and a sub-pixel consistency loss.
Key results
- Achieves 35.05 dB / 0.922 SSIM on SIDD Benchmark, above MASH (34.78 dB) and ScoreDVI (34.60 dB) among zero-shot methods.
- Ablation on SIDD Validation: Ldn alone gives 29.13 dB; adding Linr and Lmc raises it to 35.31 dB.
- Best zero-shot results on FMDD (33.95 dB), PolyU (37.88 dB), and CC (37.20 dB) in the reported table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed method instead of MASH | Zero-shot denoising · SIDD Validation | PSNR: 35.06 -> 35.31 | Helped |
| Use the proposed method instead of MASH | Zero-shot denoising · SIDD Benchmark | PSNR: 34.78 -> 35.05 | Helped |
| Use the proposed method instead of MASH | Zero-shot denoising · FMDD | PSNR: 33.71 -> 33.95 | Helped |
| Use the proposed method instead of ScoreDVI | Zero-shot denoising · PolyU | PSNR: 37.77 -> 37.88 | Helped |
| Use the proposed method instead of ScoreDVI | Zero-shot denoising · CC | PSNR: 37.09 -> 37.20 | Helped |
| Lower masking probability from p0=0.50 to p0=0.10 (ablation) | Proposed method · SIDD-Validation | PSNR: 34.84 -> 35.31 | Helped |
| Lower masking probability from p0=0.10 to p0=0.05 (ablation) | Proposed method · SIDD-Validation | PSNR: 35.31 -> 35.25 | Hurt |
| Lower masking probability from p0=0.50 to p0=0.05 (ablation) | Proposed method · BSD68 (i.i.d AWGN, σ = 25) | PSNR: 27.51 -> 26.74 | Hurt |
| Add Linr to the denoising loss (ablation) | Proposed method · SIDD-Validation | PSNR: 29.13 -> 34.75 | Helped |
| Add Lmc to Ldn + Linr (ablation) | Proposed method · SIDD-Validation | PSNR: 34.75 -> 35.31 | Helped |
| Add Linr to Ldn + Lmc (ablation) | Proposed method · SIDD-Validation | PSNR: 35.14 -> 35.31 | Helped |
| Vary dilation, including removing it (ablation) | Proposed method · SIDD datasets | PSNR: 35.27, 35.29, 35.31, 35.25; 35.00, 35.02, 35.05, 34.96 | No clear effect |
Surprising findings
- Lower masking ratio p0 improves real-world denoising but worsens AWGN denoising on BSD68, indicating a trade-off between correlated and i.i.d. noise.
- Dilation radius has a negligible effect on performance because the INR captures neighboring-pixel correlation.
Limitations
- Risk analysis is based on a simplified linear model with i.i.d. assumptions on the visible-pixel terms.
- Performance on i.i.d. AWGN degrades at small masking ratios, so the method is tuned toward correlated real-world noise.
Experimental setup
Datasets: SIDD FMDD PolyU CC BSD68
Benchmarks: SIDD Validation SIDD Benchmark FMDD PolyU CC BSD68
Metrics: PSNR SSIM
Baselines: DIP Self2Self PD-denoising NN+denoising APBSN-single ScoreDVI MASH BM3D DnCNN CVF-SID LUD-VAE R2R
Builds on: Noise2Void Noise2Self Noise2Noise SIREN pixel-shuffle downsampling MASH Deep Image Prior
Closest papers by meaning
- APR-RD: Complemental Two Steps for Self-Supervised Real Image Denoising (AAAI 2025 · similarity 0.90)
- Masked and Shuffled Blind Spot Denoising for Real-World Images (CVPR 2024 · similarity 0.90)
- Zero-shot Denoising via Neural Compression: Theoretical and algorithmic framework (NeurIPS 2025 · similarity 0.88)
- Masked Pre-training Enables Universal Zero-shot Denoiser (NeurIPS 2024 · similarity 0.87)
- Zero-Shot Noise2Mean: Gap Minimization for Efficient Denoising from a Single Noisy Image (AAAI 2025 · similarity 0.87)
- Exploring Efficient Asymmetric Blind-Spots for Self-Supervised Denoising in Real-World Scenarios (CVPR 2024 · similarity 0.87)
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.