🔧

Image Restoration - Motion Blur Removal (MIMO-UNet)

🛠

MIMO-UNet (Multi-Input Multi-Output U-Net) for Motion Blur Removal

Removes motion blur from images using coarse-to-fine multi-scale architecture with single encoder/decoder. Three variants available: MIMO-UNet (fast baseline), MIMO-UNet-Plus (balanced), MIMO-UNet-RealBlur (highest quality). Processes arbitrary image sizes with low computational cost.

-

What it does

MIMO-UNet is a deep learning model that removes motion blur from photographs and restores sharpness. Uses a multi-scale coarse-to-fine strategy with a single encoder (multi-scale inputs) and single decoder (multi-scale outputs). Asymmetric feature fusion efficiently processes multi-scale features without cascaded sub-networks.

Problem it solves

  • Remove motion blur from photographs
  • Restore sharpness to blurred camera images
  • Batch deblurring of image collections
  • Real-time image deblurring on GPU
  • Fast alternative to cascaded deblurring networks
  • Enhance low-quality/blurred video frames

Input/Output

Input:
Input: Blurred RGB image (any size)
Output:
Output: Sharp/deblurred RGB image (same size as input)
image
icon

Available models:

MIMO-UNet
MIMO-UNet-Plus
MIMO-UNet-RealBlur
: Fast baseline
: Balanced accuracy/speed
: Highest quality

Accuracy & Speed

  • Performance:
    • Accuracy: Qualitative visual restoration (no standard quantitative benchmark)
    • Training datasets: GoPro (2,103 blur/sharp pairs), RealBlur dataset
    • Variants: Three options for speed/quality trade-off
  • Metrics:
    • No quantitative accuracy reported (image restoration is qualitative)
    • Fast inference relative to cascaded deblurring methods
    • Suitable for batch processing and real-time applications
image

Technical Details

  • Architecture: Multi-Input Multi-Output U-Net
  • Core Innovation:
    • Single encoder processes multi-scale input images (coarse to fine)
    • Single decoder outputs deblurred images at multiple scales
    • Asymmetric feature fusion for efficient multi-scale feature processing
    • No cascaded sub-networks (unlike traditional coarse-to-fine approaches)
  • Design Benefits:
    • Reduced computational cost vs. cascaded networks
    • Leverages multi-scale information in single forward pass
    • Flexible input/output resolution handling
    • Efficient parameter utilization
  • Framework & Dependencies:
    • Framework: PyTorch (v1.4+)
    • Input resolution: Variable (processes arbitrary image sizes)
    • GPU recommended for real-time inference
  • Training:
    • Datasets: GoPro (2,103 blur/sharp image pairs), RealBlur
    • Approach: Supervised learning with sharp ground truth images
    • Learning strategy: Coarse-to-fine multi-scale optimization
    • Parameters: Multi-scale architecture (exact count varies by variant)

Technical Details

Provider
Open-source
Provider type
Specialized
License
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