Research map / Architectures and efficiency / Attention and transformers
Adaptive Window Pruning for Efficient 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 LMD-ViT, a U-shaped Transformer for local motion deblurring that adaptively prunes non-blurry windows via a learned blurriness confidence predictor (Gumbel-Softmax training, mask-supervised pruning loss), cutting FLOPs by 66% and roughly doubling speed over Transformer deblurring baselines.
Problem
Local motion blur (moving objects against static backgrounds) is handled poorly by global deblurring methods, which distort sharp regions and waste computation on sharp pixels, especially for high-resolution images.
Approach
An adaptive window pruning Transformer block (AdaWPT) predicts per-window blurriness confidence, samples keep/drop decisions (Gumbel-Softmax in training, thresholding at test), and runs window-based self-attention and feed-forward only on kept windows, with the confidence predictor supervised by newly annotated blur masks.
Key results
- LMD-ViT reaches 35.42 dB PSNR and 30.25 dB weighted PSNR on ReLoBlur, vs 34.92 / 29.30 for LBAG and 35.14 / 29.92 for Uformer-B.
- Compared with Transformer baselines, LMD-ViT reduces FLOPs to 1.485T from 4.375T (Uformer-B), about 66% lower, with inference time 0.56s vs 1.31s.
- Window pruning precision at the last pruning block is 96.8% on the ReLoBlur test set.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply adaptive window pruning | LMD-ViT · ReLoBlur test | FLOPs: 4.376T -> 1.485T | Helped |
| apply adaptive window pruning | LMD-ViT · ReLoBlur test | inference time: 1.30s -> 0.56s | Helped |
| remove window pruning (ablation) | LMD-ViT · ReLoBlur test | weighted PSNR: 30.25 -> 30.22 | Hurt |
| use our annotated blur masks instead of LBFMG masks (ablation) | LMD-ViT · ReLoBlur test | weighted PSNR: 30.14 -> 30.25 | Helped |
| use our annotated blur masks instead of LBFMG masks (ablation) | LBAG · ReLoBlur test | weighted PSNR: 28.31 -> 29.30 | Helped |
| replace LeFF with W-LeFF (ablation) | LMD-ViT · ReLoBlur test | inference time: 0.86s -> 0.56s | Helped |
| replace LeFF with W-LeFF (ablation) | LMD-ViT · ReLoBlur test | PSNR: 35.45 -> 35.42 | No clear effect |
| increase feature channels from 16-32-64-128-256 to 32-64-128-256-512 (ablation) | LMD-ViT · ReLoBlur test | PSNR: 34.98 -> 35.42 | Helped |
| set pruning threshold β to 0.7 instead of 0.5 (ablation) | LMD-ViT · ReLoBlur test | weighted PSNR: 30.25 -> 30.14 | Hurt |
| set pruning threshold β to 0.2 instead of 0.5 (ablation) | LMD-ViT · ReLoBlur test | weighted PSNR: 30.25 -> 30.21 | Hurt |
| use AdaWPT blocks 2–8 instead of 1–9 (ablation) | LMD-ViT · ReLoBlur test | weighted PSNR: 30.25 -> 30.34 | Helped |
| use AdaWPT blocks 3–7 instead of 1–9 (ablation) | LMD-ViT · ReLoBlur test | weighted PSNR: 30.25 -> 30.38 | Helped |
Surprising findings
- On globally blurred GoPro inputs, LMD-ViT's performance is slightly below Uformer and it does not prune windows, since blurriness confidence approaches 100% everywhere.
- Pruning the most blocks (all 9) gives the best weighted PSNR among pruning configurations while using fewer parameters, and pruning in all blocks still outperforms baselines.
Limitations
- Not capable of real-time deblurring.
- Pruning precision is not 100%, so some blurred tokens miss deblurring and global-blur quality is slightly below Uformer.
- Evaluation relies on the authors' own blur mask annotations on ReLoBlur, and real-world evaluation is a small user study of 18 images.
Experimental setup
Datasets: ReLoBlur GoPro
Benchmarks: ReLoBlur GoPro
Metrics: PSNR SSIM weighted PSNR weighted SSIM precision inference time FLOPs parameters
Baselines: DeepDeblur DeblurGAN-v2 HINet MIMO-UNet LBAG Restormer Uformer-B
Builds on: Restormer Uformer LBAG Gumbel-Softmax Swin Transformer
Closest papers by meaning
- Parameterized Blur Kernel Prior Learning for Local Motion Deblurring (CVPR 2025 · similarity 0.90)
- Real-World Efficient Blind Motion Deblurring via Blur Pixel Discretization (CVPR 2024 · similarity 0.87)
- Motion-adaptive Transformer for Event-based Image Deblurring (AAAI 2025 · similarity 0.87)
- ID-Blau: Image Deblurring by Implicit Diffusion-based reBLurring AUgmentation (CVPR 2024 · similarity 0.87)
- BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.87)
- Motion-adaptive Separable Collaborative Filters for Blind Motion Deblurring (CVPR 2024 · 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.