metadata
license: mit
language:
- en
tags:
- computer-vision
- object-detection
- yolov8
- document-analysis
- historical-documents
- heritage-ai
- pytorch
- ultralytics
pipeline_tag: object-detection
🏛️ YOLOv8 — Historical Document Ornament Detector
Automatic detection of typographic ornaments in 16th–18th century printed documents. Developed as part of the TypoRef project at PolyTech Tours.
📌 Model Summary
| Property | Details |
|---|---|
| 🏗️ Architecture | YOLOv8 (Ultralytics) |
| 🎯 Task | Object Detection |
| 📊 mAP@50 | 95% |
| 🗂️ Dataset | 50+ expert-annotated historical document pages |
| 📅 Document period | 16th – 18th century printed books |
| ⚙️ Framework | PyTorch + Ultralytics |
| 📉 Processing speedup | 20% faster than manual workflow |
| 📜 License | MIT |
🧠 What This Model Does
This model detects and localizes typographic ornaments and decorative graphic elements in scanned pages of early modern European printed books.
It was built to replace a slow, fully manual cataloguing process for the TypoRef digital humanities project, enabling automated analysis of thousands of document pages that would otherwise require extensive expert annotation.
Detected classes: typographic ornaments, decorative initials, vignettes, and other graphic elements typical of 16th–18th century printing.
📈 Performance
| Metric | Score |
|---|---|
| mAP@50 | 95% |
| Training duration | 6 months iterative refinement |
| Annotations integrated | 50+ pages in 2 months |
| Processing time reduction | 20% vs previous pipeline |
🚀 How to Use
from ultralytics import YOLO
# Load the model
model = YOLO("best.pt")
# Run inference on a document scan
results = model("your_document_scan.jpg", conf=0.35)
# Show results
results[0].show()
# Save annotated image
results[0].save("output.jpg")
🗂️ Training Data
- Source: Historical printed books from the TypoRef corpus (16th–18th century)
- Annotations: Expert-annotated by digital humanities researchers at PolyTech Tours
- Volume: 50+ annotated document pages
- Augmentation: Standard YOLOv8 augmentation pipeline
⚠️ Limitations
- Optimized for black-and-white or greyscale document scans
- Performance may degrade on very low-resolution scans (< 150 DPI)
- Trained on Western European printing conventions — may generalize poorly to other traditions
🔗 Related Resources
👤 Author
Martin Badrous — Computer Vision & Deep Learning Engineer
