Research map / Image generation / Image restoration
Zero-Shot Noise2Mean: Gap Minimization for Efficient Denoising from a Single Noisy Image
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-N2M, a zero-shot single-image denoiser that uses a random mask-based downsampler and two lightweight UNet-free networks (MBD-Net and GRD-Net) to shrink the clean-image gap between input and training target.
Problem
Self-supervised zero-shot denoising from a single noisy image (e.g., ZS-N2N) suffers from a gap between the clean images of the network's input and target, which degrades denoising quality.
Approach
A random mask-based downsampler creates multiple similar-but-distinct downsampled pairs whose mean image is the target (reducing the gap to one quarter); a second network then uses the mean image as input and the frozen first network's output as target, eliminating the gap.
Key results
- On synthetic Gaussian sigma=25 on Kodak24, ZS-N2M reaches 29.56 dB PSNR versus 28.27 dB for ZS-N2N.
- On real-world SIDD, ZS-N2M achieves 36.89 dB / 0.927 SSIM versus 33.28 dB for ZS-N2N among zero-shot methods.
- ZS-N2M* (MBD-Net only, ~20K parameters) matches ZS-N2N's speed (20 sec GPU) while improving PSNR over it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ZS-N2M instead of the strongest listed baseline | ZS-N2M · Kodak24, Gaussian σ = 25 | PSNR (dB): 29.56 vs 30.95 | Hurt |
| Use ZS-N2M instead of the strongest listed baseline | ZS-N2M · Set14, Gaussian σ = 25 | PSNR (dB): 29.05 vs 30.94 | Hurt |
| Use ZS-N2M instead of the strongest listed baseline | ZS-N2M · Kodak24, Gaussian σ ∈ [5, 50] | PSNR (dB): 29.90 vs 28.58 | Helped |
| Use ZS-N2M instead of the strongest listed baseline | ZS-N2M · Set14, Gaussian σ ∈ [5, 50] | PSNR (dB): 29.23 vs 27.88 | Helped |
| Use ZS-N2M instead of the strongest listed baseline | ZS-N2M · Kodak24, Poisson λ = 25 | PSNR (dB): 27.93 vs 28.02 | Hurt |
| Use ZS-N2M instead of the strongest listed baseline | ZS-N2M · Set14, Poisson λ = 25 | PSNR (dB): 27.43 vs 28.13 | Hurt |
| Use ZS-N2M instead of the strongest listed baseline | ZS-N2M · Kodak24, Poisson λ ∈ [5, 50] | PSNR (dB): 27.17 vs 26.56 | Helped |
| Use ZS-N2M instead of the strongest listed baseline | ZS-N2M · Set14, Poisson λ ∈ [5, 50] | PSNR (dB): 27.55 vs 27.42 | Helped |
| Use ZS-N2M instead of the strongest listed baseline | ZS-N2M · Microscopy | PSNR (dB): 34.53 vs 34.58 | No clear effect |
| Use ZS-N2M instead of the strongest listed baseline | ZS-N2M · SIDD | PSNR (dB): 36.89 vs 37.23 | Hurt |
| Use ZS-N2M instead of the strongest listed baseline | ZS-N2M · PolyU | PSNR (dB): 37.65 vs 38.13 | Hurt |
| Add GRD-Net after MBD-Net (ablation) | ZS-N2M on Kodak24 · Gaussian σ = 25 | PSNR (dB): 29.39 -> 29.56 | Helped |
Surprising findings
- Directly setting the mean image as the target hurts performance because it increases noise similarity between input and target, so the method uses an indirect loss instead.
Limitations
- Full ZS-N2M is slower than ZS-N2N and ZS-N2M* because the GRD-Net is a larger UNet.
- Best results on real data are obtained with alpha_ie tuned per setting (1 synthetic, 10 real), requiring hyperparameter choice by noise level.
Experimental setup
Datasets: Kodak24 Set14 ImageNet PolyU SIDD Fluorescence Microscopy fastMRI
Benchmarks: Kodak24 Set14 SIDD PolyU Fluorescence Microscopy fastMRI
Metrics: PSNR SSIM GPU time CPU time parameter count
Baselines: N2C N2N N2V NB2NB DIP Self2Self BM3D ZS-N2N Noise2Fast MASH
Builds on: ZS-N2N Noise2Noise Neighbor2Neighbor Self2Self
Closest papers by meaning
- Zero-Shot Noise2Noise: Efficient Image Denoising Without Any Data (CVPR 2023 · similarity 0.91)
- Positive2Negative: Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising (CVPR 2025 · similarity 0.89)
- APR-RD: Complemental Two Steps for Self-Supervised Real Image Denoising (AAAI 2025 · similarity 0.89)
- Masked Pre-training Enables Universal Zero-shot Denoiser (NeurIPS 2024 · similarity 0.88)
- Zero-shot Denoising via Neural Compression: Theoretical and algorithmic framework (NeurIPS 2025 · similarity 0.88)
- Exploring Efficient Asymmetric Blind-Spots for Self-Supervised Denoising in Real-World Scenarios (CVPR 2024 · similarity 0.88)
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.