Research map / Label-efficient and robust learning / Self-supervised and contrastive
Exploring Efficient Asymmetric Blind-Spots for Self-Supervised Denoising in Real-World Scenarios
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 AT-BSN, a shifted-convolution blind-spot network with a tunable, asymmetric blind-spot at original resolution, and distills a lightweight non-blind-spot student from multiple blind-spot teachers for real-world self-supervised denoising.
Problem
Self-supervised denoising for real-world images with spatially correlated noise, where existing downsampling or neighborhood-masking methods lose texture detail or are computationally expensive.
Approach
Use a blind-spot network whose blind-spot size is freely tunable via feature shifting, training with a large blind-spot (k=9) and inferring with a small one (k=3) at original resolution; then treat the trained network as a meta-teacher that spawns multiple teachers across blind-spot sizes for multi-teacher distillation into a lightweight non-BSN student.
Key results
- AT-BSN (D) reaches 37.77 dB PSNR on the SIDD Benchmark and 38.29 dB on DND, outperforming other self-supervised methods such as SDAP (E) (37.24 dB SIDD) and LG-BPN (37.28 dB SIDD) among non-fully-self-supervised results.
- On the SIDD validation set, AT-BSN (D) achieves 37.88 dB / 0.946 SSIM, with 1.02M parameters and 48.92 GMACs on 512x512 patches, far fewer than AP-BSN + R3 (7653.97 GMACs) and LG-BPN (12168.22 GMACs).
- Best asymmetric setting found was training blind-spot k=9 and inference blind-spot k=3 on the SIDD validation set.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| distill a lightweight student from multiple blind-spot teachers instead of using AT-BSN alone | AT-BSN · SIDD Validation | PSNR: 36.80 -> 37.88 | Helped |
| distill using multiple teacher outputs rather than the mean teacher output (ablation) | Student A · SIDD Validation | PSNR: 36.79 -> 36.98 | Helped |
| distill using multiple teacher outputs rather than the mean teacher output (ablation) | Student B · SIDD Validation | PSNR: 37.60 -> 37.76 | Helped |
| distill using multiple teacher outputs rather than the mean teacher output (ablation) | Student C · SIDD Validation | PSNR: 37.72 -> 37.88 | Helped |
| use training blind-spot size 9 and inference blind-spot size 3 (ablation) | AT-BSN · SIDD Validation | PSNR | Helped |
| increase the teacher ensemble to blind-spots {0,1,3,5,7,9,11} (ablation) | AT-BSN teacher ensemble · SIDD Validation | PSNR: 37.41 -> 37.47 | Helped |
| add blind-spot 13 to the teacher ensemble (ablation) | AT-BSN teacher ensemble · SIDD Validation | PSNR: 37.47 -> 37.40 | Hurt |
| train at original resolution and use asymmetric blind-spots | AT-BSN · SIDD Validation | texture preservation | Helped |
Surprising findings
- Ensembling teachers by averaging outputs at test time gives less benefit than multi-teacher distillation, and mean-teacher distillation performs worse than multi-teacher distillation.
- Excessively large sets of blind-spot sizes for teacher ensembling can lead to overly smooth outputs and decreased performance.
Limitations
- Training and inference blind-spot sizes must be chosen via ablation (k=9 train, k=3 inference), and performance is sensitive to the combination, with degradation at k_a=7.
- Some reported results use training directly on the benchmark (†) which is not the main generalization setting; the default AT-BSN is below some baselines on SIDD benchmark (36.73 dB vs 37.41 dB for Spatially-Adaptive).
Experimental setup
Datasets: SIDD-Medium SIDD validation DND
Benchmarks: SIDD Benchmark SIDD Validation DND Benchmark
Metrics: PSNR SSIM MACs parameters
Baselines: AP-BSN + R3 C-BSN SDAP (E) LG-BPN Spatially-Adaptive (UNet) CVF-SID Neighbor2Neighbor R2R Noise2Void Noise2Self
Builds on: Laine et al. shifted-convolution BSN AP-BSN Noise2Void Noise2Self LG-BPN
Closest papers by meaning
- APR-RD: Complemental Two Steps for Self-Supervised Real Image Denoising (AAAI 2025 · similarity 0.91)
- Rethinking Transformer-Based Blind-Spot Network for Self-Supervised Image Denoising (AAAI 2025 · similarity 0.90)
- Pseudo-Siamese Blind-spot Transformers for Self-Supervised Real-World Denoising (NeurIPS 2024 · similarity 0.90)
- Masked and Shuffled Blind Spot Denoising for Real-World Images (CVPR 2024 · similarity 0.89)
- Zero-Shot Noise2Mean: Gap Minimization for Efficient Denoising from a Single Noisy Image (AAAI 2025 · similarity 0.88)
- Zero-Shot Blind-spot Image Denoising via Implicit Neural Sampling (CVPR 2025 · similarity 0.87)
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.