Research map / Image generation / Image restoration
SIDL: A Real-World Dataset for Restoring Smartphone Images with Dirty Lenses
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.
Introduces SIDL, a real-world paired dataset of smartphone photos taken through physically contaminated PVC film lenses (dust, fingerprint, water, scratch, mixed), and benchmarks seven restoration models on it.
Problem
Image restoration research assumes clean lenses, but smartphone lens contamination (water drops, fingerprints, dust, scratches) causes degraded images with no suitable real paired benchmark.
Approach
Captures real degradations by placing custom-made physical dirty films (PVC with simulated contaminants) in front of an iPhone lens with a clean reference per scene, rather than synthesizing degradations, and stratifies difficulty by PSNR into easy/medium/hard levels.
Key results
- DiffUIR achieved the best average restoration across degradation types and levels (e.g., 29.34 dB / 0.9335 on Easy, 21.71 dB on Hard), though all models remained limited, with Hard-level averages around 17.5 to 22.5 dB.
- Pretrained models underperformed models trained on SIDL (e.g., MambaIR Dust 19.33 vs 23.42 dB), indicating the dataset's distinct degradation patterns.
- Restoration performance improved with more training scenes up to about 100, then plateaued.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune the pretrained DiffUIR on SIDL | DiffUIR · SIDL Dust test set | PSNR: +2.77 | Helped |
| Fine-tune the pretrained DiffUIR on SIDL | DiffUIR · SIDL Water test set | PSNR: +2.11 | Helped |
| Fine-tune the pretrained FFTformer on SIDL | FFTformer · SIDL Dust test set | PSNR: +1.34 | Helped |
| Fine-tune the pretrained FFTformer on SIDL | FFTformer · SIDL Water test set | PSNR: +0.89 | Helped |
| Fine-tune the pretrained MambaIR on SIDL | MambaIR · SIDL Dust test set | PSNR: +4.09 | Helped |
| Fine-tune the pretrained MambaIR on SIDL | MambaIR · SIDL Water test set | PSNR: +2.85 | Helped |
| Increase the number of training scenes from 10 to 240 (ablation) | Restormer · SIDL test set (15 scenes) | PSNR: 20.18 / 0.75 -> 23.73 / 0.86 | Helped |
| Increase the number of training scenes from 200 to 240 (ablation) | Restormer · SIDL test set (15 scenes) | PSNR: 23.55 / 0.85 -> 23.73 / 0.86 | No clear effect |
| Train on SIDL instead of the Wang et al. dirty-lens dataset | NAFNet · SIDL scratch test set | PSNR: 25.55 / 0.8239 -> 27.55 / 0.8829; 21.61 / 0.7235 -> 25.16 / 0.8282; 18.16 / 0.7598 -> 20.24 / 0.7734 | Helped |
| Use AutoDIR instead of DiffUIR | SIDL Water restoration · SIDL Water test set, Easy / Medium / Hard | PSNR: 26.83 vs 27.42; 24.67 vs 24.65; 19.19 vs 19.91 | Mixed |
Surprising findings
- AutoDIR, an unsupervised pretrained model, performed comparably to DiffUIR trained on SIDL on water-degraded test images, but was about 22x slower.
Limitations
- Covers only a limited set of common contaminant types; real-world lens contamination may be more varied.
- Contaminants are simulated with physical PVC films rather than captured on actual lenses across the full dataset, though a small real-lens check was performed for fingerprint and water.
- Benchmark models were trained with a single run per setting; no broader hyperparameter or seed analysis is reported.
Experimental setup
Datasets: SIDL
Benchmarks: SIDL
Metrics: PSNR SSIM
Baselines: AirNet NAFNet Restormer FFTformer DiffUIR MambaIR AutoDIR
Builds on: Let's see clearly DiffUIR MambaIR
Closest papers by meaning
- DreamClear: High-Capacity Real-World Image Restoration with Privacy-Safe Dataset Curation (NeurIPS 2024 · similarity 0.81)
- RestoreAgent: Autonomous Image Restoration Agent via Multimodal Large Language Models (NeurIPS 2024 · similarity 0.80)
- MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration (CVPR 2025 · similarity 0.80)
- AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation (ICLR 2025 · similarity 0.80)
- AWRaCLe: All-Weather Image Restoration Using Visual In-Context Learning (AAAI 2025 · similarity 0.80)
- DiffRAW: Leveraging Diffusion Model to Generate DSLR-Comparable Perceptual Quality sRGB from Smartphone RAW Images (AAAI 2024 · similarity 0.80)
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.