Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Neural Restoration of Greening Defects in Historical Autochrome Photographs Based on Purely Synthetic Data

Created by
  • Haebom

Author

Saptarshi Neil Sinha, P. Julius Kuehn, Johannes Koppe, Arjan Kuijper, Michael Weinmann

Outline

This paper presents a first-of-its-kind approach to automatically remove green discoloration defects from digitized autochrome photographs. To address the challenges of restoring defects such as blurring, scratches, color bleeding, and fading caused by aging and improper storage in autochrome photographs, we present a method to accurately simulate defects and train a generative AI model using synthetic data and ground truth defect annotations. Specifically, we design a loss function that considers color imbalances between defect and non-defect regions, enabling efficient and effective restoration that accurately reproduces the original colors and minimizes manual effort. Our focus is on addressing systematic defects that are difficult to restore using existing software (e.g., Adobe Photoshop).

Takeaways, Limitations

•
Takeaways:
◦
Presenting the first automated approach for removing green discoloration defects in autochrome photographs.
◦
Presenting an effective model training method using defect simulation and synthetic data.
◦
Overcoming the limitations of existing methods, enabling primary color reproduction and reducing manual work.
◦
Solving color imbalance problems using generative AI models and special loss functions.
•
Limitations:
◦
Currently, we are focusing only on green discoloration defects. It may be difficult to apply to other types of defects.
◦
Since this is a simulation and model training for defects specific to autochrome photos, it may be difficult to generalize to other types of photos.
◦
The lack of a publicly available autochromic photo defect annotation dataset may limit the evaluation of the model's generalization performance.
👍
Made with Slashpage