Research map / Image generation / Image restoration
Zero-shot Denoising via Neural Compression: Theoretical and algorithmic framework
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 ZS-NCD, a zero-shot denoiser that trains a neural compression network with an entropy-regularized objective on overlapping patches of a single noisy image and averages patch reconstructions, plus finite-sample upper bounds for compression-based ML denoising under AWGN and Poisson noise.
Problem
Denoising a single noisy image without clean references or training data (zero-shot denoising), where learning-based zero-shot methods often overfit or need early stopping, and where compression-based denoisers have historically underperformed on images.
Approach
Treat an untrained neural compression network (encoder, entropy model, decoder) as the denoiser: train it on overlapping patches of one noisy image with a rate-distortion-style objective, so the built-in entropy constraint prevents overfitting, then aggregate overlapping patch outputs; the fidelity weight lambda is chosen so the residual MSE matches the known noise variance (or a Poisson-derived t
Key results
- ZS-NCD reaches 31.35 dB on Set11 with AWGN sigma=15 and 33.18 dB on Kodak24 at sigma=15, outperforming the zero-shot baselines in the table.
- Under Poisson noise with alpha=25, ZS-NCD achieves 27.10 dB on Set11 versus 26.01 dB for the next-best zero-shot method (ZS-N2S).
- On PolyU real camera denoising and Mouse Nuclei microscopy, ZS-NCD is reported as outperforming other learning-based zero-shot denoisers, with 35.84 dB on PolyU (Unknown noise).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ZS-NCD instead of the strongest listed baseline | Kodak24 AWGN denoising · Kodak24 AWGN, σ=15, 25, 50 | PSNR (dB): 32.37 -> 33.18; 29.98 -> 30.60; 27.06 -> 27.89 | Helped |
| Use ZS-NCD instead of the strongest listed baseline | Kodak24 Poisson denoising · Kodak24 Poisson, α=15, 25, 50 | PSNR (dB): 28.22 -> 27.64; 29.26 -> 28.77; 30.41 -> 30.60 | Mixed |
| Use ZS-NCD instead of the strongest listed baseline | Mouse Nuclei microscopy denoising · Mouse Nuclei AWGN, σ=10 and 20 | PSNR (dB): 38.65 -> 38.23; 34.96 -> 34.71 | Hurt |
| Use ZS-NCD instead of the strongest listed baseline | PolyU real-camera denoising · PolyU camera images, unknown noise | PSNR (dB): 35.71 -> 35.84 | Helped |
| Replace the convolutional encoder-decoder with an MLP (ablation) | ZS-NCD on Set11 · Set11 AWGN, σ=25 and 50 | PSNR (dB): 28.93 -> 29.52; 25.58 -> 25.89 | Helped |
| Average predictions from overlapping patches instead of retaining one fixed patch location (ablation) | ZS-NCD on Parrot · Parrot, Set11 AWGN σ=25 | PSNR (dB): 25.90 -> 28.14 | Helped |
| Use Poisson negative log-likelihood instead of MSE (ablation) | ZS-NCD with estimated α on Cameraman · Cameraman, Set11 Poisson α=15 and 50 | PSNR (dB): 23.41 -> 23.13; 25.22 -> 24.88 | Hurt |
| Use Poisson negative log-likelihood instead of MSE (ablation) | ZS-NCD with estimated α on Cameraman · Cameraman, Set11 Poisson α=15 | SSIM: 0.7554 -> 0.7567 | Helped |
Surprising findings
- Replacing the convolutional encoder-decoder with a fully connected MLP (about 2.3M params) improved Set11 denoising performance rather than degrading it, which the authors attribute to the entropy constraint preventing overfitting.
- BM3D remains the strongest method on grayscale Set11 at sigma=15 and 25 and on Mouse Nuclei, with ZS-NCD second best in several settings.
Limitations
- Theoretical results assume a fixed compression code and do not cover zero-shot learned codes.
- Lambda selection relies on knowing the noise variance (or estimating alpha for Poisson noise), and Poisson alpha is estimated heuristically assuming mean intensity near 0.5.
- On some grayscale images BM3D outperforms ZS-NCD, and performance is inconsistent for some baselines, so comparisons rely on best-achieved baseline numbers.
Experimental setup
Datasets: Set11 Set13 Kodak24 Mouse Nuclei PolyU
Benchmarks: Set11 Set13 Kodak24 Mouse Nuclei PolyU
Metrics: PSNR SSIM
Baselines: JPEG2K BM3D DIP Deep Decoder ZS-N2S Self2Self ZS-N2N DS-N2N Pixel2Pixel
Builds on: Deep Image Prior Zero-shot Noise2Noise Deep Decoder BM3D
Closest papers by meaning
- Masked Pre-training Enables Universal Zero-shot Denoiser (NeurIPS 2024 · similarity 0.90)
- Zero-Shot Blind-spot Image Denoising via Implicit Neural Sampling (CVPR 2025 · similarity 0.88)
- Zero-Shot Noise2Mean: Gap Minimization for Efficient Denoising from a Single Noisy Image (AAAI 2025 · similarity 0.88)
- Compressed Image Generation with Denoising Diffusion Codebook Models (ICML 2025 · similarity 0.87)
- Efficient Online Training for Zero-Shot Time-Lapse Microscopy Denoising and Super-Resolution (AAAI 2025 · similarity 0.87)
- Zero-Shot Image Restoration Using Few-Step Guidance of Consistency Models (and Beyond) (CVPR 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.