Research map / Image generation / Image restoration
Real-World Efficient Blind Motion Deblurring via Blur Pixel Discretization
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.
Decomposes blind motion deblurring into blur pixel discretization (a classifier producing a blur segmentation map via log-Fourier deconvolution) and a discrete-to-continuous converter, yielding an efficient deblurring model comparable to larger SOTA methods.
Problem
Blind motion deblurring for high-resolution images with large motion requires models that are both accurate and computationally efficient for resource-constrained devices; small kernel-free models degrade significantly on large motion.
Approach
Treat deblurring partly as pixel-level classification: a blur pixel discretizer predicts per-pixel blur classes (a discretized image residual error) using log-Fourier basis kernels and a latent sharp image, and a D2C converter (NAFNet-based) uses this class map as prior to regress the continuous residual at low cost.
Key results
- SegDeblur-S (14.44 GMACs) improves RealBlur-J total PSNR from 31.99 to 32.53 dB over NAFNet-32 at similar cost, and large-motion PSNR from 28.27 to 29.12 dB.
- SegDeblur-L reaches 33.51 dB on RealBlur-J with TLC, above UFPNet's 33.35 dB at 243 GMACs, using 62.68 GMACs.
- SegFFTformer improves FFTFormer on GoPro from 34.14 to 34.38 dB with about 4 extra GMACs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the full blur-segmentation-map, latent-sharp-loss, and logarithmic-Fourier components (ablation) | NAFNet-32 · RealBlur-J test | PSNR: 31.99 -> 32.53 | Helped |
| Add latent sharp loss to the blur-segmentation-map variant (ablation) | SegDeblur-S · RealBlur-J test | PSNR: 31.96 -> 32.17 | Helped |
| Add logarithmic Fourier transform to the variant with blur segmentation map and latent sharp loss (ablation) | SegDeblur-S · RealBlur-J test | PSNR: 32.17 -> 32.53 | Helped |
| Replace no prior with a latent sharp image prior (ablation) | NAFNet · RealBlur-J test | PSNR: 31.99 -> 32.18 | Helped |
| Replace latent sharp image prior with latent quantization map prior (ablation) | NAFNet · RealBlur-J test | PSNR: 32.18 -> 32.27 | Helped |
| Use deconvolved image prior instead of latent sharp image prior (ablation) | NAFNet · RealBlur-J test | PSNR: 32.18 -> 32.44 | Helped |
| Use blur segmentation map prior instead of deconvolved image prior (ablation) | NAFNet · RealBlur-J test | PSNR: 32.44 -> 32.53 | Helped |
| Increase the number of classes from 1 to 16 (ablation) | SegDeblur-S · RealBlur-J test | PSNR: 31.91 -> 32.53 | Helped |
| Set the number of classes to 32 instead of 16 (ablation) | SegDeblur-S · RealBlur-J test | PSNR: 32.53 -> 32.50 | Hurt |
| Set latent sharp loss weight λ to 1.0 instead of 0.1 (ablation) | SegDeblur-S · RealBlur-J test | PSNR: 32.14 -> 32.53 | Helped |
| Set latent sharp loss weight λ to 2.0 instead of 1.0 (ablation) | SegDeblur-S · RealBlur-J test | PSNR: 32.53 -> 32.36 | Hurt |
| Add SegDeblur-L instead of NAFNet-64 | NAFNet-64 · RealBlur-J large-motion test | PSNR: 28.27 -> 29.12 dB | Helped |
Surprising findings
- Simply estimating the blur segmentation map without latent sharp loss and log-Fourier space gives no gain over NAFNet-32 (31.96 vs 31.99 dB).
Limitations
- Efficient SegDeblur-S trails larger methods on some benchmarks (e.g., RealBlur-R 39.75 vs 40.20 dB for FFTFormer).
- On-chip runtime is slower than commercial applications (2.35s vs 1.64s and 2.03s) without NPU acceleration.
Experimental setup
Datasets: RealBlur RSBlur ReLoBlur GoPro
Benchmarks: RealBlur-J RealBlur-R RSBlur ReLoBlur GoPro
Metrics: PSNR SSIM parameters GMACs on-chip execution time
Baselines: NAFNet-32 NAFNet-64 MPRNet MIMO-UNet+ FFTFormer Stripformer MAXIM-3S UFPNet MSDI-Net Samsung EnhanceX Google Unblur
Builds on: NAFNet kernel-based deblurring with basis kernels logarithmic Fourier deconvolution
Closest papers by meaning
- Blur2Blur: Blur Conversion for Unsupervised Image Deblurring on Unknown Domains (CVPR 2024 · similarity 0.89)
- GS-Blur: A 3D Scene-Based Dataset for Realistic Image Deblurring (NeurIPS 2024 · similarity 0.88)
- BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.88)
- ID-Blau: Image Deblurring by Implicit Diffusion-based reBLurring AUgmentation (CVPR 2024 · similarity 0.88)
- Motion-adaptive Separable Collaborative Filters for Blind Motion Deblurring (CVPR 2024 · similarity 0.88)
- Parameterized Blur Kernel Prior Learning for Local Motion Deblurring (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.