Research map / Image generation / Image restoration
IDR: Self-Supervised Image Denoising via Iterative Data Refinement
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| iteratively refine noisy-image targets and retrain using the known noise model | IDR · Gaussian denoising on BSDS300, σ=50 | PSNR: 28.17 -> 28.25 | Helped |
| use single noisy images plus a noise model instead of requiring aligned noisy pairs | IDR · SID, ISO 250 | PSNR: 33.44 -> 33.88 | Helped |
| apply IDR to raw images at the highest reported SID ISO | IDR · SID, ISO 300 | PSNR: 31.61 -> 32.14 | Helped |
| apply IDR to extremely noisy smartphone scenes | IDR · SenseNoise-500, ISO > 12800 | PSNR: 40.17 vs 40.23 | No clear effect |
| reduce target data bias by using less-corrupted targets | denoisers trained on biased noisier-noisy data · Gaussian, raw, and correlated noise | PSNR: -0.355 vs -0.008; -3.100 vs -0.310; -0.172 vs -0.019 | Helped |
| add Gaussian noise to clean targets with higher corruption strength | biased noisier-noisy training data · Gaussian, raw, and correlated denoising | PSNR drop from ideal noisy-clean training: σ 1: -0.008, -0.310, -0.019; σ 5: -0.355, -3.100, -0.172 | Hurt |
| add Gaussian blur to clean targets with higher corruption strength | biased noisier-noisy training data · Gaussian, raw, and correlated denoising | PSNR drop from ideal noisy-clean training: σ 1: -0.067, -0.081, -0.150; σ 5: -0.074, -0.101, -0.944 | Hurt |
| replace U-Net with a larger CNN backbone (ablation) | IDR · Gaussian denoising, σ=25 | PSNR: 31.84 -> 32.23 (SCN); 31.84 -> 32.59 (MPRNet) | Helped |
| replace U-Net with Swin Transformer (ablation) | IDR · Gaussian denoising, σ=25 | PSNR: 31.84 -> 32.22 | Helped |
| increase full data-refinement iterations from one to five (ablation) | full IDR · Gaussian denoising | PSNR: 24.84 -> 31.49 | Helped |
| use fast one-epoch refinement with cumulative model initialization instead of full iterations (ablation) | IDR · Gaussian denoising | PSNR: 31.49 -> 31.50 | No clear effect |
| increase unlabeled training data scale from 0.5k to 50k (ablation) | IDR · Gaussian denoising, σ=25 | PSNR: 31.65 -> 31.84 | Helped |
Closest papers by meaning
- Convexity-Aware Noise Calibration: A Self-Supervised Framework for Noise-Level-Unknown Image Denoising (CVPR 2026 · similarity 0.93)
- Noise2Score: Tweedie’s Approach to Self-Supervised Image Denoising without Clean Images (NeurIPS 2021 · similarity 0.92)
- Masked and Shuffled Blind Spot Denoising for Real-World Images (CVPR 2024 · similarity 0.89)
- APR-RD: Complemental Two Steps for Self-Supervised Real Image Denoising (AAAI 2025 · similarity 0.89)
- DenoiseRep: Denoising Model for Representation Learning (NeurIPS 2024 · similarity 0.89)
- Positive2Negative: Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising (CVPR 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.