[공지사항]을 빙자한 안부와 근황 
Show more

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.

Reviving Cultural Heritage: A Novel Approach for Comprehensive Historical Document Restoration

Created by
  • Haebom

Author

Yuyi Zhang, Peirong Zhang, Zhenhua Yang, Pengyu Yan, Yongxin Shi, Pengwei Liu, Fengjun Guo, Lianwen Jin

Outline

This paper presents a full-page HDR dataset (FPHDR) and an automated HDR solution (AutoHDR) to address the Limitations problem in the field of historical document restoration (HDR). FPHDR consists of 1,633 real images and 6,543 synthetic images, including character and line-level location information and character annotations for various damage levels. AutoHDR mimics the restoration process of historians through a three-step approach: OCR-based damage localization, visual-linguistic contextual text prediction, and patch autoregressive appearance restoration. The modular architecture enables flexible human-machine collaboration to support intervention and optimization at each restoration step. Experimental results show that AutoHDR improves the OCR accuracy from 46.83% to 84.05% when processing severely damaged documents, and up to 94.25% through human-machine collaboration. This study contributes significantly to the advancement of automated historical document restoration and cultural heritage preservation. The model and dataset are publicly available on GitHub.

Takeaways, Limitations

•
Takeaways:
◦
Overcoming the limitations of existing methods by providing a full-page-size HDR dataset (FPHDR) and an automated HDR solution (AutoHDR).
◦
Dramatically improve OCR accuracy, increasing the efficiency of historical document restoration.
◦
Modular architecture enables flexible human-machine collaboration.
◦
Contributes greatly to the preservation of cultural heritage.
◦
Enables follow-up research through publicly available datasets and models.
•
Limitations:
◦
The high proportion of synthetic images may limit the diversity of actual damage types.
◦
There is a possibility that performance for certain types of damage may be relatively lower than for others.
◦
Further research is needed to improve the efficiency and optimization of human-machine collaboration.
◦
Generalization performance needs to be evaluated across multiple languages and handwritings.
👍
Made with Slashpage