Research map / Image generation / Image restoration
Positive2Negative: Breaking the Information-Lossy Barrier in Self-Supervised Single 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 Positive2Negative, a self-supervised single-image denoising paradigm that renoises a predicted denoised image with scaled positive and negative copies of the predicted noise and enforces consistency between the two denoised outputs, avoiding downsampling and masking.
Problem
Self-supervised single image denoising: existing Noise2Noise- and Noise2Void-style methods rely on information-lossy operations (downsampling, masking, extra noise), which degrade denoising quality (residual noise, aliasing, texture loss).
Approach
Use the predicted noise from a single noisy image to build positive and negative renoised images (exploiting the approximately zero-mean symmetric noise distribution), then train the network so the two denoised outputs agree, with a gradually varying norm (2 to 1.5); this preserves all image information instead of discarding pixels.
Key results
- Achieves 35.24 dB / 0.889 SSIM on SIDD validation, outperforming prior single-image methods such as MASH (35.06) and ScoreDVI (34.75).
- Reaches 38.20 dB on PolyU and 37.92 dB on CC, the best among self-supervised single-image methods in Table 1.
- Converges in under 100 iterations (P2N-Fixed under 20) versus 800x3 for MASH and 400 for ScoreDVI, with 0.99M parameters and 11.4 GFLOPs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Positive2Negative instead of MASH | Self-supervised single-image denoising · SIDD val | PSNR: 35.06 -> 35.24 | Helped |
| Use Positive2Negative instead of ScoreDVI | Self-supervised single-image denoising · SIDD ben | PSNR: 35.39 -> 35.80 | Helped |
| Use Positive2Negative instead of Self2Self | Self-supervised single-image denoising · CC | PSNR: 37.44 -> 37.92 | Helped |
| Use Positive2Negative instead of ScoreDVI | Self-supervised single-image denoising · PolyU | PSNR: 37.77 -> 38.20 | Helped |
| Use Positive2Negative instead of MASH | Self-supervised single-image denoising · FMDD | PSNR: 33.71 -> 34.00 | Helped |
| Replace DCS with Noise2Noise supervision (ablation) | Positive2Negative · CC and SIDD | PSNR: 37.92 -> 34.08; 35.24 -> 25.42 | Hurt |
| Replace DCS with Noise2Void supervision (ablation) | Positive2Negative · CC and SIDD | PSNR: 37.92 -> 33.79; 35.24 -> 24.51 | Hurt |
| Replace RDC with Noise2Noise data construction (ablation) | Positive2Negative · CC and SIDD | PSNR: 37.92 -> 17.55; 35.24 -> 24.20 | Hurt |
| Replace RDC with Noise2Void data construction (ablation) | Positive2Negative · CC and SIDD | PSNR: 37.92 -> 14.14; 35.24 -> 22.30 | Hurt |
| Vary scale parameter σ across 0.25, 0.50, 0.75, and 1.00 instead of fixed setting (ablation) | Positive2Negative · CC and SIDD | PSNR: CC: 37.33, 37.64, 37.84, 37.92, 37.86; SIDD: 35.24, 35.29, 35.30, 35.24, 35.11 | Mixed |
| Replace ∥· ∥2→1.5 with ∥· ∥2 (ablation) | Positive2Negative · PolyU | PSNR: 38.20 -> 38.20 | No clear effect |
| Replace ∥· ∥2→1.5 with ∥· ∥2 (ablation) | Positive2Negative · SIDD | PSNR: 35.24 -> 35.20 | No clear effect |
Surprising findings
- Constructed RDC data is unsuitable for Noise2Noise and Noise2Void: plugging RDC into them collapses training (e.g., RDC+N2N gets 25.42 dB on SIDD), because the constructed noise is correlated.
- Using the DCS consistency loss without RDC performs very poorly (N2N+DCS 17.55 dB on CC), showing the two components must be combined.
Limitations
- Only validated on single-image self-supervised denoising; not on noisy-dataset settings.
- Requires initialization from a pre-trained Gaussian denoiser to avoid trivial solutions (zero or identity mapping), unlike training from scratch.
- Relies on noise being approximately zero-mean and symmetric; the paper notes this assumption is not strictly required but is untested in strongly asymmetric cases.
Experimental setup
Datasets: DIV2K (pretraining with Gaussian noise) SIDD CC PolyU FMDD
Benchmarks: SIDD validation SIDD benchmark CC PolyU FMDD
Metrics: PSNR SSIM
Baselines: Noise2Void CVF-SID LUD-VAE DIP PD-denoising NN+denoiser Self2Self R2R APBSN-single ScoreDVI ZS-N2N MASH
Builds on: Noise2Noise Noise2Void R2R ZS-N2N MASH
Closest papers by meaning
- Noise2Same: Optimizing A Self-Supervised Bound for Image Denoising (NeurIPS 2020 · similarity 0.91)
- Noise2Score: Tweedie’s Approach to Self-Supervised Image Denoising without Clean Images (NeurIPS 2021 · similarity 0.89)
- Zero-Shot Noise2Mean: Gap Minimization for Efficient Denoising from a Single Noisy Image (AAAI 2025 · similarity 0.89)
- APR-RD: Complemental Two Steps for Self-Supervised Real Image Denoising (AAAI 2025 · similarity 0.87)
- Ditch the Denoiser: Emergence of Noise Robustness in Self-Supervised Learning from Data Curriculum (NeurIPS 2025 · similarity 0.87)
- UNSURE: self-supervised learning with Unknown Noise level and Stein's Unbiased Risk Estimate (ICLR 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.