Research map / Image generation / Image restoration
LAN: Learning to Adapt Noise for Image Denoising
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 LAN, which keeps a SIDD-pretrained denoiser frozen and learns a per-pixel offset added to the noisy input, optimized with a self-supervised loss, to handle unseen real-world noise.
Problem
Image denoising networks trained on one noise distribution generalize poorly to unseen noise at test time; existing test-time adaptation fine-tunes the network, which can be suboptimal or degrade performance.
Approach
Instead of adapting the denoising network, freeze it and learn a pixel-wise offset added to the noisy input so that the input noise is shifted toward the distribution the network was trained on, guided by a self-supervised loss (ZS-N2N or Nbr2Nbr).
Key results
- On SIDD→Nam with Restormer and Nbr2Nbr loss, LAN improves PSNR by 0.35 dB over pretraining while full-trainable adaptation improves only 0.05 dB.
- LAN reaches 38.86 dB on SIDD→Nam with Restormer and ZS-N2N after 20 iterations, versus 38.35 dB for full-trainable adaptation.
- LAN is more efficient in runtime and memory than full-trainable adaptation for Restormer and Uformer at 256×256 (Table 2 ratios, e.g. 74.10% runtime for Uformer ZS-N2N).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LAN instead of full-trainable adaptation | Restormer with Nbr2Nbr · Nam | PSNR (dB): 38.05 -> 38.38 | Helped |
| Use LAN instead of full-trainable adaptation | Restormer · Nam (ZS-N2N) | PSNR (dB): 38.35 -> 38.86 | Helped |
| Use LAN instead of meta-learning adaptation | Uformer · Nam (ZS-N2N) | PSNR (dB): 38.52 -> 38.36 | Hurt |
| Use LAN instead of last-layer adaptation | DnCNN · Nam (Nbr2Nbr) | PSNR (dB): 36.81 -> 36.80 | Hurt |
| Use LAN instead of first-layer adaptation | denoising networks · PolyU and Nam | PSNR (dB) | Helped |
| Adapt all network parameters rather than use LAN | Uformer with ZS-N2N · Nam | PSNR (dB): 38.36 -> 38.07 | Hurt |
| Train a randomly initialized network on each noisy image rather than use pretrained denoising knowledge | DnCNN · PolyU and Nam | PSNR (dB): 31.47; 32.93; 34.47; 35.39 | Hurt |
Surprising findings
- Full-trainable test-time adaptation often degrades performance relative to the pretrained denoiser.
- Zero-shot denoisers trained from scratch on a single noisy image perform much worse than the SIDD-pretrained model with adaptation.
Limitations
- Number of trainable offset parameters scales with input image size, reducing efficiency for large images.
- Gains on some settings are small (e.g., DnCNN on PolyU), and results are reported only for SIDD→PolyU and SIDD→Nam.
Experimental setup
Datasets: SIDD PolyU Nam BSDS500
Benchmarks: PolyU Nam
Metrics: PSNR SSIM
Baselines: full-trainable fine-tuning first-layer fine-tuning last-layer fine-tuning meta-learning (first-order MAML) zero-shot denoising (ZS-N2N, Nbr2Nbr)
Builds on: ZS-N2N Neighbor2Neighbor Noise2Noise MAML
Closest papers by meaning
- Toward Convolutional Blind Denoising of Real Photographs (CVPR 2019 · similarity 0.89)
- Universal Denoising Networks : A Novel CNN Architecture for Image Denoising (CVPR 2018 · similarity 0.89)
- GAN2GAN: Generative Noise Learning for Blind Denoising with Single Noisy Images (ICLR 2021 · similarity 0.88)
- Transfer Learning From Synthetic to Real-Noise Denoising With Adaptive Instance Normalization (CVPR 2020 · similarity 0.88)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.87)
- APR-RD: Complemental Two Steps for Self-Supervised Real Image Denoising (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.