Research map / Image generation / Image restoration
Blur2Blur: Blur Conversion for Unsupervised Image Deblurring on Unknown Domains
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.
Blur2Blur trains a blur-to-blur translator from unpaired blurry and sharp images of a target camera to convert unknown blur into a known blur that a pretrained supervised deblurring network (NAFNet/Restormer trained on GoPro) can handle.
Problem
Deblurring images from a specific, unknown camera when no paired blurry-sharp training data is available and pretrained supervised deblurring networks generalize poorly to unseen blur.
Approach
Instead of learning unknown-blur to sharp directly, learn a GAN-based blur-to-blur translator (MIMO-UNet generator with perceptual reconstruction loss and gradient-penalty discriminator) that maps the camera's blur into a known blur domain built from the camera's own sharp images augmented with blur kernels transferred from a known dataset, then deblur with a pretrained supervised network.
Key results
- With NAFNet, Blur2Blur (GoPro known blur) improves PSNR by 2.20 dB on RB2V Street, 2.31 dB on REDS, and 2.67 dB on RSBlur over the GoPro-trained NAFNet baseline's unsupervised counterpart setting.
- Blur2Blur with Restormer increases PSNR by 2.63 dB on RB2V Street, 2.12 dB on REDS, and 2.91 dB on RSBlur.
- Unpaired baselines CycleGAN and DualGAN score around 21-24 dB PSNR, far below Blur2Blur-enhanced NAFNet (about 27-29 dB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GoPro-targeted Blur2Blur translation before NAFNet | NAFNet · RB2V Street | PSNR: 24.78 -> 26.98 | Helped |
| Add GoPro-targeted Blur2Blur translation before NAFNet | NAFNet · REDS | PSNR: 25.80 -> 28.11 | Helped |
| Add GoPro-targeted Blur2Blur translation before NAFNet | NAFNet · RSBlur | PSNR: 26.33 -> 29.00 | Helped |
| Use a 6:4 rather than 5:5 unknown-blurry-to-sharp data ratio (ablation) | Blur2Blur with NAFNet · GoPro–RB2V Street | PSNR: 26.02 -> 26.98 | Helped |
| Use a 7:3 rather than 5:5 unknown-blurry-to-sharp data ratio (ablation) | Blur2Blur with NAFNet · GoPro–RB2V Street | PSNR: 26.02 -> 26.92 | Helped |
| Use an 8:2 rather than 5:5 unknown-blurry-to-sharp data ratio (ablation) | Blur2Blur with NAFNet · GoPro–RB2V Street | PSNR: 26.02 -> 25.98 | No clear effect |
| Use a 9:1 rather than 5:5 unknown-blurry-to-sharp data ratio (ablation) | Blur2Blur with NAFNet · GoPro–RB2V Street | PSNR: 26.02 -> 24.32 | Hurt |
| Use a 6:4 rather than 5:5 unknown-blurry-to-sharp data ratio (ablation) | Blur2Blur with NAFNet · GoPro–REDS | PSNR: 27.53 -> 28.11 | Helped |
| Use a 7:3 rather than 5:5 unknown-blurry-to-sharp data ratio (ablation) | Blur2Blur with NAFNet · GoPro–REDS | PSNR: 27.53 -> 28.10 | Helped |
| Use an 8:2 rather than 5:5 unknown-blurry-to-sharp data ratio (ablation) | Blur2Blur with NAFNet · GoPro–REDS | PSNR: 27.53 -> 27.00 | Hurt |
| Use a 9:1 rather than 5:5 unknown-blurry-to-sharp data ratio (ablation) | Blur2Blur with NAFNet · GoPro–REDS | PSNR: 27.53 -> 26.43 | Hurt |
Surprising findings
- Generalized deblurring methods (BSRGAN, RSBlur) and unpaired translation methods perform poorly despite expected generalization, sometimes below the GoPro-trained supervised baseline.
- Using a blur-to-sharp ratio that is not balanced (9:1) hurts performance, indicating too few sharp images causes overfitting.
Limitations
- Requires access to the target camera to collect unpaired blurry and sharp images.
- Performance depends on the choice of known blur domain and blur-to-sharp data ratio.
Experimental setup
Datasets: REDS GoPro RSBlur RB2V Street PhoneCraft WritingHands
Benchmarks: RB2V Street REDS RSBlur GoPro PhoneCraft
Metrics: PSNR SSIM NIQE
Baselines: NAFNet Restormer CycleGAN DualGAN BSRGAN+NAFNet RSBlur+NAFNet
Builds on: CycleGAN Pix2Pix MIMO-UNet blur kernel extractor (blur transfer) WGAN-GP gradient penalty
Closest papers by meaning
- Real-World Efficient Blind Motion Deblurring via Blur Pixel Discretization (CVPR 2024 · similarity 0.89)
- Unsupervised Blind Image Deblurring Based on Self-Enhancement (CVPR 2024 · similarity 0.88)
- BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.88)
- GS-Blur: A 3D Scene-Based Dataset for Realistic Image Deblurring (NeurIPS 2024 · similarity 0.88)
- LDP: Language-driven Dual-Pixel Image Defocus Deblurring Network (CVPR 2024 · similarity 0.87)
- Prior and Prediction Inverse Kernel Transformer for Single Image Defocus Deblurring (AAAI 2024 · 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.