Research map / Image generation / Image restoration
Robust Test-Time Adaptation for Single Image Denoising Using Deep Gaussian Prior
Qing Ma, Pengwei Liang, Xiong Zhou, Jiayi Ma, Junjun Jiang, Zhe Peng
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 |
|---|---|---|---|
| initialize ZS-N2N from a pre-trained Gaussian denoiser (10 iterations vs 2000) (ablation) | ZS-N2N · Kodak24 Gaussian sigma=50 | PSNR (dB): 24.81 -> 26.07; time 9.04 s -> 0.07 s | Improved |
| fine-tune pre-trained Gaussian denoiser with ZS-N2N pair construction (ablation) | Pre-trained Neighbor2Neighbor denoiser · Kodak24 Gaussian sigma=25 (in-distribution) | PSNR (dB): 29.70 -> 29.33 | Hurt |
| fine-tune pre-trained denoiser with test-time adaptation on out-of-distribution noise | Pre-trained Neighbor2Neighbor denoiser · Kodak24 Gaussian sigma=50 | PSNR (dB): 20.94 -> 27.11 | Improved |
| fine-tune pre-trained denoiser with test-time adaptation on Poisson noise | Pre-trained Neighbor2Neighbor denoiser · Kodak24 Poisson lambda=10 | PSNR (dB): 19.94 -> 26.64 | Improved |
| add test-time adaptation on in-distribution Gaussian noise | Pre-trained Neighbor2Neighbor denoiser · Kodak24 Gaussian sigma=25 | PSNR (dB): 29.70 -> 30.75 (+1.05 dB) | Improved |
| add Gaussian pre-training before adaptation with pixel-bank pairs (ablation) | Ours · Kodak24 Gaussian sigma=25 | PSNR (dB): 29.65 -> 30.75 | Improved |
| construct pairs by non-local pixel search (pixel bank) instead of downsampling (ablation) | test-time fine-tuning · Kodak24 fine-tuning over iterations | PSNR (dB): peak 27.14 vs NB2NB 26.36 and ZS-N2N 26.40 | Improved |
| run more fine-tuning iterations with ZS-N2N pair construction (ablation) | ZS-N2N pairs · Kodak24 fine-tuning over iterations | PSNR (dB): 26.40 -> 16.42 over iterations | Hurt |
| apply Ours to real-world PolyU noise without pre-training (vs with pre-trained model) (ablation) | Ours · PolyU real-world | PSNR (dB): 36.12 (without pre-train) vs 36.10 (pre-trained only) | No clear effect |
Closest papers by meaning
- Convexity-Aware Noise Calibration: A Self-Supervised Framework for Noise-Level-Unknown Image Denoising (CVPR 2026 · similarity 0.90)
- Self-Supervised Image Denoising Using Implicit Deep Denoiser Prior (AAAI 2023 · similarity 0.90)
- High-Quality Self-Supervised Deep Image Denoising (NeurIPS 2019 · similarity 0.90)
- FBI-Denoiser: Fast Blind Image Denoiser for Poisson-Gaussian Noise (CVPR 2021 · similarity 0.89)
- IDR: Self-Supervised Image Denoising via Iterative Data Refinement (CVPR 2022 · similarity 0.89)
- Self-Calibrated Variance-Stabilizing Transformations for Real-World Image Denoising (ICCV 2025 · similarity 0.89)
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.
