DDSRNet

DDSRNet (Deep Joint Denoising and Super-Resolution Network)

A new two-stage learning architecture designed for simultaneous image denoising and super-resolution. By decoupling these complex tasks, DDSRNet enables specialized processing for each, leading to more effective noise removal and higher-quality upscaling.

Paper

This implementation is based on our paper published at the 2025 Seventeenth International Conference on Quality Control by Artificial Vision; 1373705 (2025): Read the Paper on arXiv

Citation

@inproceedings{messai2025enhancing,
  title={Enhancing image quality and anomaly detection for small and dense industrial objects in nuclear recycling},
  author={Messai, Oussama and Zein-Eddine, Abbass and Bentamou, Abdelouahid and Picq, Micka{\"e}l and Duquesne, Nicolas and Puydarrieux, St{\'e}phane and Gavet, Yann},
  booktitle={Seventeenth International Conference on Quality Control by Artificial Vision},
  volume={13737},
  pages={21--28},
  year={2025},
  organization={SPIE}
}

Features

• Two-Stage Architecture: Decoupled denoising and super-resolution stages for specialized processing
• Advanced Loss Function: Novel composite loss combining multiple image quality metrics for robust training
• Multi-Task Learning: Simultaneous optimization of denoising and super-resolution objectives
• High-Quality Output: Enhanced image quality through specialized loss components

Advanced Loss Function

DDSRNet employs a sophisticated composite loss function that combines multiple image quality metrics:

Composite Loss Components:

• L1 Loss: Pixel-wise absolute difference between reference and output images
• SSIM Loss: Structural similarity index measure for perceptual quality (weighted by 10)
• PSNR Loss: Peak signal-to-noise ratio optimization (weighted by 100/PSNR)

Training Loss:

Loss_training = λ × Loss_denoising + β × Loss_super-resolution

Where λ and β are weighting coefficients that balance the contribution of each task-specific loss component.

Requirements

Ensure you have the following dependencies installed:

• Python 3.6 or higher
• PyTorch
• TensorBoardX
• torchsummary
• SciPy
• NumPy
• Pillow
• Matplotlib
• YAML

You can install all the required packages using the provided requirements.txt file.

License

This project is licensed under the MIT License. See the LICENSE file for details.