Research map / Image generation / Image restoration
APR-RD: Complemental Two Steps for Self-Supervised Real 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 APR (Adjacent Pixel Replacer), a downsampling-free pair sampler for blind-spot networks, plus Recharged Distillation (RD) to refine high-frequency detail, achieving state-of-the-art self-supervised real sRGB denoising on SIDD and DND.
Problem
Self-supervised blind-spot networks degrade on real sRGB noise because real noise is spatially correlated, and prior sub-sampling/downscaling pair generators cause aliasing that loses high-frequency texture.
Approach
APR replaces each pixel with a random adjacent pixel within a 3x3 window to create decorrelated training pairs without downscaling, combined with the Neighbor2Neighbor regularization; RD then distills from a set of recharged BSN outputs in which each target mixes noisy and denoised pixels, so the distilled network refines clean details without extra architecture changes.
Key results
- APR (RD) reaches 38.00 dB / 0.947 on SIDD validation, above AT-BSN (D) at 37.80 dB / 0.944.
- APR (RD) reaches 38.57 dB on DND benchmark and 38.83 dB when fully self-supervised (†), versus 38.34 dB for AT-BSN (D).
- Ablation shows lambda=4 is optimal for the regularization factor (37.231 dB on SIDD validation vs 37.091 dB at lambda=0).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add APR sampling to AT-BSN training (ablation) | AT-BSN · SIDD Validation | PSNR (dB): 37.80 -> 37.23 | Hurt |
| Add APR sampling to AT-BSN training (ablation) | AT-BSN · DND Benchmark | PSNR (dB): 38.34 -> 38.10 | No clear effect |
| Add Recharged Distillation to APR (ablation) | APR · SIDD Validation | PSNR (dB): 37.23 -> 38.00 | Helped |
| Add Recharged Distillation to APR (ablation) | APR · SIDD Benchmark | PSNR (dB): 37.62 -> 38.26 | Helped |
| Add Recharged Distillation to APR (ablation) | APR · DND Benchmark | PSNR (dB): 38.10 -> 38.57 | Helped |
| Set regularization factor λ to 4 rather than 0 (ablation) | APR · SIDD Validation | PSNR: 37.091 -> 37.231 | Helped |
| Increase regularization factor λ from 4 to 16 (ablation) | APR · SIDD Validation | PSNR: 37.231 -> 37.119 | Hurt |
| Use six RD targets, TRD={x̃0, x̃1, .., x̃5} (ablation) | APR (RD) · SIDD Validation | PSNR: 37.77 -> 38.00 | Helped |
| Use six RD targets, TRD={x̃0, x̃1, .., x̃5} (ablation) | APR (RD) · DND Benchmark | PSNR: 38.55 -> 38.57 | No clear effect |
Surprising findings
- APR alone (37.23 dB on SIDD validation) underperforms AT-BSN (D) at 37.80 dB, with gains only appearing after RD distillation.
Limitations
- Optimal inference shift factor varies by test dataset (s=1 best for DND, larger s for SIDD), which RD only partially unifies.
- APR distorts GT correlation, an acknowledged trade-off.
- SIDD benchmark numbers for self-supervised baselines are taken from a Kaggle competition rather than the official server because the server was unavailable.
Experimental setup
Datasets: SIDD medium
Benchmarks: SIDD validation SIDD benchmark DND benchmark
Metrics: PSNR SSIM
Baselines: AT-BSN (D) AP-BSN (R3) C-BSN SDAP (E) SA (UNet) CVF-SID Neighbor2Neighbor Restormer VDN CBDNet DnCNN
Builds on: AT-BSN Neighbor2Neighbor Noise2Void Laine-BSN
Closest papers by meaning
- Exploring Efficient Asymmetric Blind-Spots for Self-Supervised Denoising in Real-World Scenarios (CVPR 2024 · similarity 0.91)
- Zero-Shot Blind-spot Image Denoising via Implicit Neural Sampling (CVPR 2025 · similarity 0.90)
- Pseudo-Siamese Blind-spot Transformers for Self-Supervised Real-World Denoising (NeurIPS 2024 · similarity 0.89)
- Zero-Shot Noise2Mean: Gap Minimization for Efficient Denoising from a Single Noisy Image (AAAI 2025 · similarity 0.89)
- Rethinking Transformer-Based Blind-Spot Network for Self-Supervised Image Denoising (AAAI 2025 · similarity 0.88)
- Masked and Shuffled Blind Spot Denoising for Real-World Images (CVPR 2024 · similarity 0.88)
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.