Research map / Architectures and efficiency / CNN design
Parameterized Blur Kernel Prior Learning for Local Motion Deblurring
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 parameterized motion blur kernels (length, angle, curvature) estimated by a memory-bank network, and a dual-branch Mamba/CNN deblurring network (PGDN) guided by the estimated kernels and blur masks for local motion deblurring.
Problem
Local motion deblurring requires restoring only blurry regions while keeping sharp background, but existing methods depend on manually annotated coarse blur masks and ignore blur kernel characteristics.
Approach
Model motion blur kernels with three interpretable parameters (length, angle, curvature), estimate them per pixel with a memory-bank network trained via a reblurring loss, derive a fine-grained blur mask from kernel length, and feed kernel parameters and mask into a dual-branch network (U-shape Mamba global branch, mask-guided CNN local branch).
Key results
- On ReLoBlur+GoPro training, PGDN reaches 36.64 PSNR / 0.9351 SSIM, over 1 dB above LMD-ViT (35.42 PSNR) with fewer parameters (29.21M vs 54.50M) and FLOPs (0.613T vs 1.485T).
- Kernel-estimation ablation: full MPENet reblur PSNR 41.76 vs 36.69 for Whyte and 40.16 for UFPNet.
- On GoPro global deblurring, PGDN achieves 34.17 PSNR, comparable to the best baselines (FFTformer 34.21).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PGDN instead of Uformer-B | Local motion deblurring · ReLoBlur testset (trained on ReLoBlur) | PSNR: 35.03 -> 36.53 | Helped |
| Train on ReLoBlur+GoPro instead of ReLoBlur alone | PGDN · ReLoBlur testset | PSNR: 36.53 -> 36.64 | Helped |
| Use PGDN instead of LMD-ViT | Local motion deblurring · ReLoBlur testset (trained on ReLoBlur+GoPro) | PSNR: 35.42 -> 36.64 | Helped |
| Use PGDN instead of NAFNet | Global motion deblurring · GoPro | PSNR: 33.71 -> 34.17 | Helped |
| Use PGDN instead of Restormer | Global motion deblurring · GoPro | SSIM: 0.9611 -> 0.9501 | Hurt |
| Add kernel modeling to direct kernel estimation (ablation) | MPENet · Kernel reblurring fidelity | PSNR: 39.56 -> 41.37 | Helped |
| Add the blur-prior memory bank (ablation) | MPENet with kernel modeling · Kernel reblurring fidelity | PSNR: 41.37 -> 41.53 | Helped |
| Add the auxiliary reblurring network (ablation) | MPENet with kernel modeling and memory bank · Kernel reblurring fidelity | PSNR: 41.53 -> 41.76 | Helped |
| Add estimated kernel parameters (ablation) | PGDN dual-branch baseline · ReLoBlur testset | PSNR: 35.47 -> 35.83 | Helped |
| Add the estimated blur mask (ablation) | PGDN dual-branch baseline · ReLoBlur testset | PSNR: 35.47 -> 36.09 | Helped |
| Replace SSM with CNN backbone (ablation) | PGDN · ReLoBlur testset | PSNR: 36.53 -> 34.97 | Hurt |
| Replace SSM with Transformer backbone (ablation) | PGDN · ReLoBlur testset | PSNR: 36.53 -> 35.14 | Hurt |
Surprising findings
- On GoPro global deblurring, PGDN has higher PSNR than most baselines but lower SSIM (0.9501) than many of them, e.g., NAFNet 0.9668.
Limitations
- Kernel estimation accuracy is evaluated only via reblur fidelity since no ground-truth kernels exist.
- Planar motion model with three parameters may not capture all real-world blur patterns.
- Global deblurring SSIM lags several baselines despite comparable PSNR.
Experimental setup
Datasets: ReLoBlur GoPro
Benchmarks: ReLoBlur GoPro
Metrics: PSNR SSIM weighted PSNR weighted SSIM Params (M) FLOPs (T)
Baselines: LBAG+ LMD-ViT MUGNet MIMO-Unet Restormer Uformer-B NAFNet HINet DeblurGAN-v2 DeepDeblur Whyte UFPNet
Builds on: Mamba NAFNet Restormer LBAG UFPNet
Closest papers by meaning
- Non-Blind Deblurring: Handling Kernel Uncertainty With CNNs (CVPR 2018 · similarity 0.93)
- Adaptive Window Pruning for Efficient Local Motion Deblurring (ICLR 2024 · similarity 0.90)
- DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Prior and Prediction Inverse Kernel Transformer for Single Image Defocus Deblurring (AAAI 2024 · similarity 0.88)
- Real-World Efficient Blind Motion Deblurring via Blur Pixel Discretization (CVPR 2024 · similarity 0.87)
- Dynamic Gaussian Splatting from Defocused and Motion-blurred Monocular Videos (NeurIPS 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.