Staged Multi-Scale Processing for Motion Blur Removal
Removes motion blur from images using multi-scale-stage architecture with progressive refinement. Three model variants available: S-GoPro (lightweight/fast), GoPro (standard/balanced), L-GoPro (high-quality). Processes arbitrary image sizes with configurable speed/quality trade-off.
What it does
MSSNet is a deep learning model that removes motion blur from photographs using staged multi-scale processing. Architecture progressively refines deblurred output through multiple scales, achieving state-of-the-art performance. Supports efficient inference with multiple model size options.
Problem it solves
- Remove motion blur from photographs
- Restore image sharpness from blurry camera captures
- Batch processing of blurred image collections
- Efficient deblurring with multiple model size options
- Real-time deblurring applications
- Flexible speed/quality trade-off for different deployment scenarios
Input/Output
Dataset versions:
S-GoPro | GoPro | L-GoPro |
: Lightweight model (wf=20), fast inference, lower parameter count | : Standard model (wf=54), balanced speed and quality | : High-quality model (wf=80), highest quality output |
Accuracy & Speed
- Performance:
- Accuracy: State-of-the-art on standard deblurring benchmarks
- Training dataset: GOPRO_Large dataset
- Variants: Three options for speed/quality/parameter trade-off
- Metrics:
- Qualitative restoration performance (benchmark comparisons available)
- Suitable for batch processing and real-time applications
- Parameter efficiency across model variants
Technical Details
- Architecture: Multi-Scale-Stage Network
- Core Innovation:
- Multi-scale processing organized in stages
- Progressive refinement of deblurred output
- Each stage processes image at different scales
- Cumulative refinement improves sharpness quality
- Efficient parameter utilization through staged design
- Design Benefits:
- Staged architecture enables flexible computation allocation
- Multi-scale processing captures blur patterns at multiple resolutions
- Progressive refinement reduces compounding errors
- Configurable width factor (wf) controls model size/capacity
- Framework & Dependencies:
- Framework: PyTorch (v1.4 or v1.7)
- Input resolution: Variable (processes arbitrary image sizes)
- GPU: CUDA support (GPU required for real-time inference)
- Training:
- Dataset: GOPRO_Large (standard deblurring benchmark)
- Approach: Supervised learning with sharp ground truth images
- Learning strategy: Multi-scale-stage optimization
- Custom datasets: Supported with proper setup
Technical Details
Provider | Open-source |
Provider type | Specialized |
License | Provenance uncertain — good-faith deployment based on public availability |
EU AI Act risk class | Minimal Risk |
Art. 50 transparency | Not applicable |
Region availability | Available globally |
Training data summary | Pending — provider has not yet published per Art. 53(d) |
For more on how we classify models and mark outputs, see our AI Policy.
Model Source
- GitHub Repository: https://github.com/kky7/MSSNet
- Paper: "MSSNet: Multi-Scale-Stage Network for Single Image Deblurring" https://cg.postech.ac.kr/Research/MSSNet/MSSNet.pdf
- Project Page: https://cg.postech.ac.kr/Research/MSSNet/