Research map / Image generation / Image restoration
Adaptive Denoising via GainTuning
Sreyas Mohan, Joshua L. Vincent, Ramon Manzorro, Peter Crozier, Carlos Fernandez-Granda, Eero P Simoncelli
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 |
|---|---|---|---|
| add GainTuning (per-channel gain adaptation) to pre-trained denoiser | BFCNN · Set12, sigma=30 | PSNR (dB): 29.52 -> 29.61 | Improved |
| add GainTuning to pre-trained denoiser on out-of-distribution noise | DnCNN · Set12, sigma=80 | PSNR (dB): 18.48 -> 24.57 | Improved |
| add GainTuning to pre-trained denoiser on out-of-distribution noise | DnCNN · BSD68, sigma=70 | PSNR (dB): 22.15 -> 24.89 | Improved |
| GainTuning instead of LIDIA adaptation | DnCNN · Set12, sigma=80 | PSNR (dB): 24.17 -> 24.57 | Improved |
| GainTuning versus Self2Self trained only on test image | DnCNN · Set12, sigma=70 | PSNR (dB): 24.61 -> 25.54 | Improved |
| GainTuning versus supervised model trained on wide noise range | DnCNN · Set12, sigma=70 | PSNR (dB): 25.59 vs 25.54 | No clear effect |
| use SURE loss instead of blind-spot loss in GainTuning (ablation) | GainTuning · Gaussian noise denoising | PSNR | Improved |
| tune all parameters instead of only channel gains (ablation) | pre-trained CNN denoiser · electron microscopy (TEM) data | denoising quality | Hurt |
| tune all parameters instead of only gains (ablation) | pre-trained CNN denoiser · Urban100 (high self-similarity) | PSNR | Improved |
| Self2Self trained only on test data instead of GainTuning | out-of-distribution image content · piecewise constant to natural images | PSNR | Improved |
| add GainTuning to well-optimized pre-trained denoiser on in-distribution images | DnCNN with SURE · BSD68 in-distribution | PSNR: max degradation about 0.02 dB in 3 of 408 evaluations | Mixed |
Closest papers by meaning
- Fully Unsupervised Diversity Denoising with Convolutional Variational Autoencoders (ICLR 2021 · similarity 0.88)
- Universal Denoising Networks : A Novel CNN Architecture for Image Denoising (CVPR 2018 · similarity 0.88)
- Denoising and Regularization via Exploiting the Structural Bias of Convolutional Generators (ICLR 2020 · similarity 0.87)
- LAN: Learning to Adapt Noise for Image Denoising (CVPR 2024 · similarity 0.86)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.86)
- Generalization in diffusion models arises from geometry-adaptive harmonic representations (ICLR 2024 · 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.
