Research map / Image generation / Image restoration
Efficient Multi-scale Network with Learnable Discrete Wavelet Transform 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 MLWNet, a single-input multiple-output multi-scale deblurring network with a learnable 2D discrete wavelet transform module and wavelet/multi-scale losses, achieving state-of-the-art PSNR on RealBlur and efficient runtime.
Problem
Single-image blind motion deblurring: existing multi-scale methods need complex fusion modules and hand-made low-resolution image pairs, and lose high-frequency detail during coarse-to-fine restoration.
Approach
A SIMO multi-scale baseline that takes only the full-resolution blurred image as input and outputs restored images at each scale, combined with a learnable 2D discrete wavelet transform (group-convolution wavelet kernels constrained by perfect-reconstruction-based wavelet loss) to improve detail recovery.
Key results
- MLWNet-B reaches 33.84 PSNR / 0.941 SSIM on RealBlur-J and 40.69 PSNR on RealBlur-R, 0.91 dB and 0.49 dB above GRL-B.
- On RSBlur, MLWNet-B reaches 34.94 PSNR / 0.880 SSIM, the highest among compared methods.
- On GoPro, MLWNet-B gets 33.83 PSNR / 0.968 SSIM, 0.001 SSIM behind FFTformer while running about 60% faster.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SIMO instead of SISO (ablation) | MLWNet baseline · RealBlur-J ablation | PSNR: 32.29 -> 32.37 | Helped |
| Use SIMO instead of MIMO (ablation) | MLWNet baseline · RealBlur-J ablation | PSNR: 32.19 -> 32.37 | Helped |
| Add WFB and WHB to the SIMO baseline (ablation) | MLWNet baseline · RealBlur-J | PSNR: 32.37 -> 32.40 | Helped |
| Add learnable wavelet transform modules (WFB and WHB) (ablation) | MLWNet baseline · RealBlur-J | PSNR: +0.25dB | Helped |
| Use wavelet loss with the wavelet modules (ablation) | MLWNet baseline · RealBlur-J | PSNR: 32.57 -> 32.62 | Helped |
| Omit wavelet loss while retaining WFB and WHB (ablation) | MLWNet baseline · RealBlur-J | PSNR | Hurt |
| Increase noise-difference level (ablation) | MLWNet-B · GoPro | PSNR: 34.81 -> 34.66 -> 33.76 -> 33.19 -> 32.63 | Hurt |
| Increase noise-difference level (ablation) | MLWNet-B · RealBlur-J | PSNR: 33.92 -> 33.81 -> 33.97 -> 33.93 -> 33.54 | No clear effect |
Surprising findings
- Without the wavelet loss, adding WFB and WHB gives minimal gains, suggesting the wavelet convolution degenerates into a general group convolution.
- Performance on synthetic GoPro blur lags behind real blur, which the authors attribute to unnatural discontinuous trajectories and noise-level differences.
Limitations
- Does not achieve the same top performance on synthetic blur (GoPro) as on real blur.
- Performance drops faster with increasing noise difference on GoPro than on RealBlur-J.
- Ablations run with a reduced width-32 model for 200K iterations on RealBlur-J.
Experimental setup
Datasets: RealBlur RSBlur GoPro
Benchmarks: RealBlur-J RealBlur-R RSBlur GoPro
Metrics: PSNR SSIM MACs runtime
Baselines: DeblurGAN-v2 SRN MPRNet SDWNet MIMO-UNet+ DeepRFT+ BANet Stripformer MSSNet MSDI-Net MAXIM-3S FFTformer
Builds on: NAFNet MIMO-UNet perfect reconstruction filter banks auxiliary head
Closest papers by meaning
- Deep Multi-Scale Convolutional Neural Network for Dynamic Scene Deblurring (CVPR 2017 · similarity 0.91)
- Scale-Recurrent Network for Deep Image Deblurring (CVPR 2018 · similarity 0.89)
- Real-World Efficient Blind Motion Deblurring via Blur Pixel Discretization (CVPR 2024 · similarity 0.87)
- Motion-adaptive Separable Collaborative Filters for Blind Motion Deblurring (CVPR 2024 · similarity 0.86)
- BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.86)
- Exploiting Deblurring Networks for Radiance Fields (CVPR 2025 · similarity 0.85)
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.