Melissa21/Historical-Document-Stain-Compositing-Benchmark

Biological Degradation Datasets for Historical Documents

These datasets were created to support research on the automatic localization of biological degradation in historical documents and were used in the experiments presented in the accompanying publication.

The repository contains two complementary groups of datasets:

• Foxing and insect-induced degradation
• Fungal and bacterial degradation

The synthetic datasets were generated using multiple compositing strategies, including direct copy-paste, alpha blending, and CMYK-based approaches.

---

Dataset Structure

1. Foxing_feces

This directory contains datasets derived from naturally occurring foxing stains and insect-induced degradation observed in historical documents.

dataset_combo_alpha

Combined dataset containing real historical document images and synthetic images generated using the alpha-based blending approach.

dataset_combo_cmyk

Combined dataset containing real historical document images and synthetic images generated using the CMYK-based compositing approach.

syn_dataset_alpha

Dataset containing only synthetic images generated using the alpha-based blending approach.

The corresponding detector was evaluated on real historical document images.

syn_dataset_cmyk

Dataset containing only synthetic images generated using the CMYK-based compositing approach.

The corresponding detector was evaluated on real historical document images.

real_dataset

Dataset containing only real annotated historical document images with foxing and insect-induced degradation.

---

2. Fungi_bacteria

This directory contains synthetic datasets generated from laboratory-prepared microbial degradation samples.

The source collection contains 12 microbial classes, consisting of 6 bacterial and 6 fungal classes:

• Epicoccum layuense — fungus; 16 samples
• Exiguobacterium indicum — bacterium; 20 samples
• Alternaria alternata — fungus; 15 samples
• Aspergillus ochraceus — fungus; 15 samples
• Talaromyces purpureogenus — fungus; 8 samples
• Micrococcus sp. — bacterium; 8 samples
• Sphingomonas aquatilis — bacterium; 8 samples
• Curtobacterium flaccumfaciens — bacterium; 7 samples
• Rhodococcus degradans — bacterium; 7 samples
• Penicillium manginii — fungus; 8 samples
• Arthrobacter sp. — bacterium; 8 samples
• Cladosporium pseudocladosporioides — fungus; 7 samples

The same source stains, historical document backgrounds, deterministic generation plans, positions, scales, and rotations were used across all synthetic variants. The datasets therefore differ primarily in the compositing method applied.

syn_copy_paste

Synthetic dataset generated using direct copy-paste without additional blending.

syn_alpha_blend

Synthetic dataset generated using the alpha-blending approach.

syn_cmyk_blend

Synthetic dataset generated using the CMYK-based blending approach.

syn_cmyk_texture

Synthetic dataset generated using CMYK-based blending with texture preservation.

syn_cmyk_texture_feather_noise

Synthetic dataset generated using CMYK-based blending with texture preservation, feathering, and noise-based transition masks.

---

Annotation Format

All datasets are annotated using the YOLO object detection format.

Each annotation file contains one object per line in the following format:

<class_id> <x_center> <y_center> <width> <height>

Bounding-box coordinates are normalized to the range [0, 1].

---

Class Mapping

Foxing_feces

Class ID	Class
0	Foxing
1	Insect-induced degradation

Fungi_bacteria

Class ID	Species
0	Sphingomonas aquatilis
1	Rhodococcus degradans
2	Micrococcus sp.
3	Curtobacterium flaccumfaciens
4	Epicoccum layuense
5	Cladosporium pseudocladosporioides
6	Exiguobacterium indicum
7	Arthrobacter sp.
8	Talaromyces purpureogenus
9	Penicillium manginii
10	Aspergillus ochraceus
11	Alternaria alternata

---

Experimental Use

Dataset A – Foxing and Insect-Induced Degradation

Models were trained using:

• synthetic-only data
• real-only data
• combined real and synthetic data

All models were evaluated on the same independent set of real historical document images.

Two synthetic compositing approaches were compared:

1. Alpha-based blending
2. CMYK-based compositing

---

Dataset B – Fungi and Bacteria

Five synthetic compositing strategies were compared:

1. Direct copy-paste
2. Alpha blending
3. CMYK blending
4. CMYK blending with texture preservation
5. CMYK blending with texture preservation, feathering, and noise

All models were trained and evaluated under the same deterministic experimental configuration.

---

Intended Use

The datasets are intended primarily for research in:

• cultural heritage preservation
• historical document analysis
• biological degradation detection
• object detection
• synthetic data generation
• synthetic-to-real evaluation
• domain adaptation

---

Limitations

The datasets are based on a finite collection of extracted degradation samples.

Although geometric transformations, different historical document backgrounds, and multiple compositing strategies increase visual variability, the generated samples remain derived from the available source stains.

For the microbial benchmark, the reported test data are synthetic. Therefore, performance on these datasets should not be interpreted as direct evidence of generalization to real historical documents containing microbial degradation.

An independent real-world microbial degradation dataset is required for synthetic-to-real evaluation.

---

License

This dataset is distributed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

The dataset may be used for non-commercial research and academic purposes with appropriate attribution.

---

Associated Publication

This dataset accompanies the paper:

M. Grznárová et al.
“Synthetic Dataset Generation and Object Detection Benchmarking for Biological Stain Analysis in Historical Documents”
IEEE ICCP 2026.

---

Citation

If you use this dataset in your research, please cite the dataset and the associated publication.

Dataset citation

Grznárová, M. (2026).
Biological Degradation Datasets for Historical Documents.
Hugging Face Datasets.

Dataset repository:

https://huggingface.co/datasets/YOUR_USERNAME/YOUR_DATASET_NAME

BibTeX

@dataset{grznarova2026biologicaldegradation,
  author    = {Grznárová, Miriam},
  title     = {Biological Degradation Datasets for Historical Documents},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Melissa21/Historical-Document-Stain-Compositing-Benchmark}
}

---

Contact

For questions related to the dataset, please contact:

Miriam Grznárová
Faculty of Mathematics, Physics and Informatics
Comenius University Bratislava, Slovakia