Research map / Image generation / Image restoration
Unsupervised Blind Image Deblurring Based on Self-Enhancement
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 SEMGUD, an unsupervised blind image deblurring framework that generates pseudo-paired blurry/sharp data with a GAN-based multi-generator setup, a re-degradation principal component consistency loss, and an iterative self-enhancement strategy that improves the NAFNet reconstructor without extra inference cost.
Problem
Real-world image deblurring lacks paired blurry/sharp training data, and unsupervised methods lag behind supervised ones; existing unsupervised methods also do not improve pseudo-pair quality or performance without increasing inference cost.
Approach
Progressively regenerate pseudo-paired blurry/sharp data using the previously trained reconstructor as a guide (self-enhancement), while a GAN-based multi-generator framework with a kernel-guided re-degradation principal component (Gaussian multi-scale) consistency loss makes synthetic blur closer to real blur.
Key results
- On GoPro, SEMGUD reaches 29.06 PSNR / 0.927 SSIM, about 2.94 dB above the best unsupervised baseline in the table (UAUD, 26.12 PSNR).
- The self-enhancement strategy raises NAFNet-based reconstructor PSNR on GoPro from 27.68 to 29.06 (+1.38 dB), versus +0.11 dB for geometric augmentation.
- On RealBlur-R, SEMGUD (35.51 PSNR) is reported to exceed NAFNet (35.50 PSNR) despite training only on GoPro.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SEMGUD instead of UAUD | unsupervised deblurring methods · GoPro | PSNR: 26.12 -> 29.06 | Helped |
| Use SEMGUD instead of UAUD | unsupervised deblurring methods · HIDE | PSNR: 24.37 -> 27.64 | Helped |
| Use SEMGUD instead of UAUD | unsupervised deblurring methods · RealBlur-R | PSNR: 32.91 -> 35.51 | Helped |
| Use SEMGUD instead of UAUD | unsupervised deblurring methods · RealBlur-J | PSNR: 26.84 -> 28.01 | Helped |
| Add self-enhancement (ablation) | NAFNet deblurring reconstructor · GoPro | PSNR: 27.68 -> 29.06 | Helped |
| Add self-enhancement (ablation) | FCLGAN · GoPro | PSNR: 24.70 -> 25.11 | Helped |
| Add self-enhancement (ablation) | FCLGAN · HIDE | PSNR: 23.49 -> 23.79 | Helped |
| Add self-enhancement (ablation) | UFPNet · GoPro | PSNR: 27.37 -> 28.41 | Helped |
| Add self-enhancement (ablation) | UFPNet · HIDE | PSNR: 25.91 -> 26.67 | Helped |
| Add blur kernel estimation to the RPC² loss (ablation) | SEMGUD generator · GoPro | PSNR | Helped |
Surprising findings
- On RealBlur-R, the unsupervised SEMGUD trained only on GoPro slightly outperforms the supervised NAFNet in PSNR.
- Supervised NAFNet and UFPNet trained on GoPro exhibit anomalous 'mode collapse' artifacts when applied to RealBlur-J.
Limitations
- Comparison set of unsupervised deblurring methods is small, as acknowledged by the authors.
- Method relies on a blur kernel estimation network; performance drops substantially without it.
- Training data is GoPro-derived unpaired blurry and sharp images, so generalization is evaluated only from GoPro to other datasets.
Experimental setup
Datasets: GoPro RealBlur HIDE
Benchmarks: GoPro HIDE RealBlur-R RealBlur-J ReloBlur
Metrics: PSNR SSIM
Baselines: CycleGAN DualGAN UIDGAN FCLGAN USR-DA USDF UAUD DeepDeblur SRN HINet Stripformer MSDI-Net
Builds on: CycleGAN NAFNet UAUD FCLGAN UFPNet
Closest papers by meaning
- Cross-Scale Self-Supervised Blind Image Deblurring via Implicit Neural Representation (NeurIPS 2024 · similarity 0.90)
- BluRef: Unsupervised Image Deblurring with Dense-Matching References (CVPR 2026 · similarity 0.89)
- A Comparative Study for Single Image Blind Deblurring (CVPR 2016 · similarity 0.89)
- Blur2Blur: Blur Conversion for Unsupervised Image Deblurring on Unknown Domains (CVPR 2024 · similarity 0.88)
- SelfPromer: Self-Prompt Dehazing Transformers with Depth-Consistency (AAAI 2024 · similarity 0.87)
- DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 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.