Research map / Image generation / Image restoration
Motion-adaptive Separable Collaborative Filters for Blind 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 the MISC Filter, a motion-adaptive separable collaborative filter that aligns blur along predicted motion flow and filters the image directly in image space for blind motion deblurring, combined with a residual reconstruction network.
Problem
Blind motion deblurring with spatially variant, complex real-world motion, where prior methods reconstruct residuals in feature space and do not handle motion directly in image space.
Approach
A motion-guided alignment module warps blur toward the motion middle using predicted flow and mask, followed by a separable collaborative filter whose kernels, weights, and offsets are all predicted adaptively per pixel; the motion estimation network is shared with residual reconstruction (shared encoder-decoder, filtering first).
Key results
- Achieves 41.23 dB PSNR on RealBlur-R versus 40.11 dB for FFTformer, and 33.88 dB on RealBlur-J versus 32.62 dB for FFTformer.
- Ablation on GoPro: the full MISC Filter raises PSNR from 32.40 dB (base) to 32.83 dB, and the shared-encoder filter-first coupling gives the best result (32.83 dB).
- On hardware-induced UDC blur (T-OLED), reaches 38.42 dB PSNR, comparable to specialized methods, with 0.05 s runtime.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MISC Filter instead of the strongest listed baseline | Blind motion deblurring · RealBlur-R | PSNR: 40.61 -> 41.23 | Helped |
| use MISC Filter instead of the strongest listed baseline | Blind motion deblurring · RealBlur-J | PSNR: 33.35 -> 33.88 | Helped |
| use MISC Filter instead of the strongest listed baseline | Blind motion deblurring · GoPro | PSNR: 34.21 -> 34.10 | Hurt |
| use MISC Filter instead of the strongest listed baseline | Blind motion deblurring · HIDE | PSNR: 32.83 -> 31.66 | Hurt |
| add motion-guided alignment (ablation) | Base model on GoPro · GoPro | PSNR: 32.40 -> 32.51 | Helped |
| add the kernel estimator (ablation) | Base+MGA on GoPro · GoPro | PSNR: 32.51 -> 32.55 | Helped |
| add the weight estimator (ablation) | Base+MGA+kernel on GoPro · GoPro | PSNR: 32.55 -> 32.67 | Helped |
| add the offset estimator (ablation) | Base+MGA+kernel on GoPro · GoPro | PSNR: 32.55 -> 32.71 | Helped |
| use the shared-based coupling strategy with filtering first (ablation) | MISC Filter network on GoPro · GoPro | PSNR: 32.83 vs. 32.38–32.63 | Helped |
| replace alternative filtering methods with MISC Filter (ablation) | Filtering module on GoPro · GoPro | PSNR: 32.21, 32.48, 32.51, 32.54 -> 32.83 | Helped |
| increase kernel size from n=3 to n=7 (ablation) | MISC Filter on GoPro · GoPro | PSNR: 32.60 -> 32.83 | Helped |
| increase kernel size from n=7 to n=11 (ablation) | MISC Filter on GoPro · GoPro | PSNR: 32.83 -> 32.87 | Helped |
Surprising findings
- Sharing the encoder-decoder between the motion estimation and residual reconstruction networks, with filtering applied first, gave the best results among ten coupling strategies.
Limitations
- Does not handle the low-light problem in UDC hardware-induced blur, so it falls short on that task.
- Performance on GoPro and HIDE is comparable to FFTformer rather than clearly superior.
Experimental setup
Datasets: RealBlur-R RealBlur-J GoPro HIDE T-OLED
Benchmarks: RealBlur-R RealBlur-J GoPro HIDE T-OLED
Metrics: PSNR SSIM runtime parameters
Baselines: MPRNet MIMO-Unet+ MAXIM Uformer-B Restormer Stripformer NAFNet FFTformer UFPNet DeepRFT+
Builds on: deformable separable convolution FFTformer DeepRFT AdaCoF
Closest papers by meaning
- Region-Adaptive Dense Network for Efficient Motion Deblurring (AAAI 2020 · similarity 0.89)
- Real-World Efficient Blind Motion Deblurring via Blur Pixel Discretization (CVPR 2024 · similarity 0.88)
- Parameterized Blur Kernel Prior Learning for Local Motion Deblurring (CVPR 2025 · similarity 0.87)
- Adaptive Window Pruning for Efficient Local Motion Deblurring (ICLR 2024 · similarity 0.86)
- Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring (CVPR 2024 · similarity 0.86)
- BlurDM: A Blur Diffusion Model for Image Deblurring (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.