Research map / Label-efficient and robust learning / Domain adaptation and generalization
Adaptive Domain Learning for Cross-domain 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 adaptive domain learning (ADL), which uses a small target-sensor set plus cross-sensor RAW data, filtering out harmful source batches via a dynamic validation set and soft PSNR criterion, together with a sensor-type/ISO channel-wise modulation network for cross-domain RAW denoising.
Problem
RAW image denoising models trained on one camera sensor generalize poorly to other sensors, and collecting large paired datasets for each sensor is costly; the goal is to adapt using existing multi-sensor data plus a few target-sensor samples.
Approach
Iteratively accept or reject source-domain batches by checking whether a one-step update improves PSNR on a randomly sampled target validation subset (with a priority-queue soft threshold), and condition the denoiser on sensor type and ISO via channel-wise FiLM-style modulation.
Key results
- On SIDD cross-sensor leave-one-out, ADL averages 47.68 dB PSNR versus 47.14 dB for Blind2Unblind and 47.03 dB for transfer learning.
- On ELD/SID DSLR sensors, ADL averages 37.01 dB PSNR (SSIM 0.868) versus 36.63 dB for transfer learning and 36.31 dB for MZSR.
- Applying ADL to the Zhang et al. calibration model with multi-sensor data improves SIDD GP PSNR to 45.83 dB versus 45.32 dB for naive fine-tuning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add source-domain adaptive learning to filter harmful data (ablation) | Cross-domain RAW denoising · ELD and SID (Sony) | PSNR: 35.88 -> 37.28 | Helped |
| add target-domain pretraining (ablation) | ADL · ELD and SID (Sony) | PSNR: 36.15 -> 37.28 | Helped |
| add ISO to sensor modulation (ablation) | ADL · ELD and SID (Sony) | PSNR: 37.13 -> 37.28 | Helped |
| add sensor type to modulation (ablation) | ADL · ELD and SID (Sony) | PSNR: 36.81 -> 37.28 | Helped |
| add dynamic validation (ablation) | ADL · ELD and SID (Sony) | PSNR: 36.89 -> 37.28 | Helped |
| apply PMN training strategy (ablation) | ADL · SIDD (Avg.) | PSNR: 47.68 -> 47.74 | Helped |
| apply PMN training strategy (ablation) | ADL · ELD and SID (Avg.) | PSNR: 37.01 -> 37.19 | Helped |
| use ADL rather than naive fine-tuning to select multiple-domain data | Zhang et al. noise calibration model · SID (Sony) | PSNR: 44.68 -> 45.17 | Helped |
| add ADL to fine-tuning with harmful synthetic data | Cross-domain denoising · SID (Sony) | PSNR: 34.59 -> 35.13 | Helped |
| add misaligned data during naive fine-tuning | Cross-domain denoising · SID (Sony) | PSNR: 35.01 -> 19.06 | Hurt |
Surprising findings
- Naive fine-tuning with multi-sensor source data can reduce performance relative to single-sensor fine-tuning, while ADL recovers gains.
- Adding harmful source data (misaligned pairs) drops naive fine-tuning PSNR to around 20 dB, while ADL stays near baseline.
Limitations
- PSNR is the only main metric since no perceptual metric is suited to RAW data.
- Gains over strongest baselines are modest (around 0.5 dB on average).
- Relies on a target validation subset that is sampled from the very small target set, with stability concerns when it is tiny.
Experimental setup
Datasets: SIDD ELD SID
Benchmarks: SIDD ELD SID
Metrics: PSNR SSIM
Baselines: Fine-tuning BM3D DIP ZS-N2N MZSR Transfer learning (Kim et al.) Blind2Unblind Prabhakar et al. Led Zhang et al. noise calibration PMN
Builds on: Noise2Noise MAML adaptive instance normalization Led PMN
Closest papers by meaning
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.88)
- MERIT: Multi-domain Efficient RAW Image Translation (CVPR 2026 · similarity 0.86)
- Cross-Level Distillation and Feature Denoising for Cross-Domain Few-Shot Classification (ICLR 2023 · similarity 0.85)
- Proxy Denoising for Source-Free Domain Adaptation (ICLR 2025 · similarity 0.85)
- LAN: Learning to Adapt Noise for Image Denoising (CVPR 2024 · similarity 0.85)
- Remix-DiT: Mixing Diffusion Transformers for Multi-Expert Denoising (NeurIPS 2024 · similarity 0.83)
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.