Research map / Label-efficient and robust learning / Noisy labels
Evaluating Unsupervised Denoising Requires Unsupervised Metrics
Adria Marcos Morales, Matan Leibovich, Sreyas Mohan, Joshua Lawrence Vincent, Piyush Haluai, Mai Tan, Peter Crozier, Carlos Fernandez-Granda
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Estimate PSNR with uPSNR using independent noisy references from the same clean image | Denoisers on natural images with Gaussian noise · Natural images, Gaussian noise σ=25, 50, 75, and 100 | PSNR agreement: 26.19 -> 26.21; 23.95 -> 24.02; 22.72 -> 22.75; 21.84 -> 21.84 | No clear effect |
| Use spatially subsampled references from a single noisy image instead of references with the same underlying clean image (ablation) | uPSNR · Natural images, Gaussian noise | PSNR agreement: 24.20 -> 26.20; 26.54 -> 27.61; 26.19 -> 28.14; 27.22 -> 25.40 | Hurt |
| Use spatially subsampled references from a single noisy image (ablation) | uPSNR · Electron microscopy, Poisson noise | PSNR agreement: 20.18 -> 20.21; 24.86 -> 24.74; 25.74 -> 25.86; 24.65 -> 24.79 | No clear effect |
| Estimate PSNR using uPSNR from three noisy references rather than averaging-based PSNR from ten noisy references | Denoisers on RAW videos with real noise · RAW videos, all ISO levels | PSNR estimate: 32.18 -> 32.27; 42.99 -> 43.03; 30.47 -> 30.49; 44.72 -> 44.80 | No clear effect |
| Use spatial subsampling when the image content is not sufficiently smooth at the pixel resolution (ablation) | uMSE and uPSNR · Single-image evaluation of natural images | metric bias | Hurt |
| Increase image smoothness before deriving references through spatial subsampling (ablation) | uMSE and uPSNR · Controlled natural-image experiment | uMSE error | Improved |
| Use three independent noisy references and subtract the pairwise-reference noise correction | uMSE · Signals satisfying independence and centered-noise conditions | MSE estimation | Improved |
| Use uPSNR derived from uMSE rather than requiring clean ground truth | Denoising evaluation · Noisy-data evaluation under the theorem assumptions | PSNR estimation | Improved |
Closest papers by meaning
- Fully Unsupervised Diversity Denoising with Convolutional Variational Autoencoders (ICLR 2021 · similarity 0.89)
- Optimal Transport for Unsupervised Denoising Learning (TPAMI 2022 · similarity 0.89)
- Unsupervised Image Restoration Using Partially Linear Denoisers (TPAMI 2021 · similarity 0.88)
- IDR: Self-Supervised Image Denoising via Iterative Data Refinement (CVPR 2022 · similarity 0.88)
- Zero-Shot Blind-spot Image Denoising via Implicit Neural Sampling (CVPR 2025 · similarity 0.86)
- Efficient Online Training for Zero-Shot Time-Lapse Microscopy Denoising and Super-Resolution (AAAI 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.
