Reviving Medieval Byzantine Seals:
A Synthetic-to-Real Approach to Character Recognition

(IJDAR 2026)

This repository contains the official implementation of the paper:

“Reviving Medieval Byzantine Seals: A Synthetic-to-Real Approach to Character Recognition”
published in International Journal on Document Analysis and Recognition (IJDAR), 2026.

DOI: https://doi.org/10.1007/s10032-026-00579-5

---

Overview

Byzantine seals represent a rich but challenging historical source due to erosion, partial visibility, and complex script variations. In this work, we propose a synthetic-to-real pipeline for character detection and recognition on medieval Byzantine seals.

The repository includes:

• Synthetic data generation using Blender
• Training pipelines for character classification and detection
• Configuration files for reproducibility
• Evaluation and plotting utilities
• Provided datasets (labels only)

---

Repository Structure

.
├── datasets/
├── src/
│   ├── data_analysis/
│   └── data_synthesis/
├── config/
├── requirements.txt
└── README.md

datasets/

Contains datasets provided by the authors. In some cases, this may include label files only, without the original images (see licensing note below).

src/

Main source code of the project.

src/data_analysis/

Includes code for:

• Character classification (e.g., ResNet-based models)
• Character detection (e.g., YOLO-based models)
• Training and evaluation pipelines
• Additional experimental models
• Plotting utilities used to generate figures shown in the paper

src/data_synthesis/

Contains Python scripts designed to be used with Blender for generating synthetic seal images.

This module implements:

• Procedural 3D generation
• Noise modeling
• Domain randomization
• Rendering pipeline for synthetic dataset creation

config/

Contains configuration files used for:

• Synthetic data generation
• Model training
• Experimental setups

These files can be used either to reproduce our experiments or as templates to understand required parameters.

requirements.txt

Lists the Python dependencies used in our experiments.

We recommend creating a dedicated virtual environment before installation:

pip install -r requirements.txt

---

Data and Licensing Notice

This repository may include annotation files only for some datasets.

Seal images referenced in the annotations originate from the Dumbarton Oaks Research Library and Collection:

https://www.doaks.org/resources/seals/byzantine-seals/

We do not own, host, or redistribute these images. Users must access them directly through the official Dumbarton Oaks website and comply with their licensing terms.

---

Reproducibility

To reproduce the main experiments:

1. Install dependencies:

   pip install -r requirements.txt
   

2. Inspect the appropriate configuration file inside config/.

3. Run training scripts from src/data_analysis/.

4. For synthetic data generation, execute the Blender-based scripts in src/data_synthesis/.

---

Citation

If you use this work, please cite:

@article{dalmassoseals2026,
  author = {
    Dalmasso, Gianluca and
    Reineri, Patric and
    Noel, Mathieu Pscherer and
    Achard, Ninon and
    Caseau, Beatrice and
    Likforman Sulem, Laurence and
    Cavagnino, Davide and
    Lucenteforte, Maurizio and
    Fiandrotti, Attilio and
    Eyharabide, Victoria
  },
  title = {Reviving Medieval Byzantine Seals: A Synthetic-to-Real Approach to Character Recognition},
  journal = {International Journal on Document Analysis and Recognition (IJDAR)},
  year = {2026},
  month = may,
  doi = {10.1007/s10032-026-00579-5},
  url = {https://doi.org/10.1007/s10032-026-00579-5}
}

---

Contact

For questions regarding the implementation or datasets, please open an issue on this repository.