Research map / Image generation / Image restoration
Masked and Shuffled Blind Spot Denoising for Real-World Images
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 MASH, a single-image blind-spot denoiser that adaptively selects the input masking ratio from an estimated noise-correlation gap and applies local pixel shuffling to decorrelate noise.
Problem
Single-image self-supervised denoising of real images whose noise is spatially correlated, which degrades blind-spot denoising (BSD) methods that assume independent noise.
Approach
Shows that the optimal masking ratio of BSD depends on noise correlation (high masking for correlated noise, low for iid), uses the gap between noise estimates at two masking ratios as a correlation proxy to choose the ratio, and shuffles pixels within small tiles of flat regions of a pseudo-clean image to weaken noise correlation.
Key results
- MASH reaches 35.06 dB PSNR on SIDD validation and 34.78 dB on SIDD benchmark, versus 34.75 and 34.60 for ScoreDVI, the best single-image competitor.
- On FMDD, MASH achieves 33.71 dB / 0.882 SSIM, outperforming all single-image and dataset-based methods listed.
- Adaptive masking plus local pixel shuffling improves over the baseline by about 2 dB on SIDD and 1.5 dB on FMDD; adaptive masking selects the optimal ratio in about 88.7% (SIDD) and 92.4% (FMDD) of images.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add adaptive masking (ablation) | Blind spot denoising without local pixel shuffling · SIDD | PSNR (dB): 33.12 -> 34.45 | Helped |
| Add adaptive masking (ablation) | Blind spot denoising with local pixel shuffling · SIDD | PSNR (dB): 33.86 -> 35.06 | Helped |
| Add local pixel shuffling (ablation) | Blind spot denoising without adaptive masking · SIDD | PSNR (dB): 33.12 -> 33.86 | Helped |
| Add local pixel shuffling (ablation) | Blind spot denoising with adaptive masking · SIDD | PSNR (dB): 34.45 -> 35.06 | Helped |
| Add adaptive masking (ablation) | Blind spot denoising without local pixel shuffling · FMDD | PSNR (dB): 32.25 -> 33.56 | Helped |
| Add adaptive masking (ablation) | Blind spot denoising with local pixel shuffling · FMDD | PSNR (dB): 32.92 -> 33.71 | Helped |
| Add local pixel shuffling (ablation) | Blind spot denoising without adaptive masking · FMDD | PSNR (dB): 32.25 -> 32.92 | Helped |
| Add local pixel shuffling (ablation) | Blind spot denoising with adaptive masking · FMDD | PSNR (dB): 33.56 -> 33.71 | Helped |
| Increase local-shuffling neighborhood size from s=2 to s=5 (ablation) | MASH · SIDD Validation | PSNR (dB): 34.65 -> 35.15 | Helped |
| Increase local-shuffling neighborhood size from s=5 to s=6 (ablation) | MASH · SIDD Validation | PSNR (dB): 35.15 -> 35.09 | Hurt |
Surprising findings
- Higher masking ratios help for highly correlated noise while low masking is best for iid noise, the reverse of the intuition that more blindness always regularizes more uniformly.
Limitations
- Adaptive masking selects the optimal ratio only in about 90% of cases, so selection errors remain.
- Training is sequential (masking estimation then shuffling), making it slower than the baseline; the authors note parallelization as future speedup.
- Results are limited to single-image test-time training; no comparison to dataset-based methods is claimed as superior except on FMDD.
Experimental setup
Datasets: SIDD FMDD PolyU Kodak (synthetic noise generation)
Benchmarks: SIDD validation SIDD benchmark FMDD PolyU
Metrics: PSNR SSIM inference time parameters FLOPs
Baselines: Self2Self NN+denoiser DIP BM3D PD-denoising APBSN-single ScoreDVI CVF-SID LUD-VAE DnCNN
Builds on: Noise2Void Self2Self Deep Image Prior Noise2Noise AP-BSN
Closest papers by meaning
- Blind2Unblind: Self-Supervised Image Denoising With Visible Blind Spots (CVPR 2022 · similarity 0.90)
- Zero-Shot Blind-spot Image Denoising via Implicit Neural Sampling (CVPR 2025 · similarity 0.90)
- Exploring Efficient Asymmetric Blind-Spots for Self-Supervised Denoising in Real-World Scenarios (CVPR 2024 · similarity 0.89)
- Pseudo-Siamese Blind-spot Transformers for Self-Supervised Real-World Denoising (NeurIPS 2024 · similarity 0.88)
- APR-RD: Complemental Two Steps for Self-Supervised Real Image Denoising (AAAI 2025 · similarity 0.88)
- Rethinking Transformer-Based Blind-Spot Network for Self-Supervised Image Denoising (AAAI 2025 · similarity 0.85)
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.