dnb-portrait-cropper

DNB Portrait Detector

A YOLO-based object detection model for identifying portrait engravings in digitized historical documents.

This model was developed as part of the HERMES-Forschungsstudienprogramm to support the automatic detection and extraction of portrait illustrations from historical collections.

You can find the result for the model inference here: https://huggingface.co/datasets/luanamcosta/dnb-buchhandlerportrats

---

Overview

The detector identifies portrait regions on scanned pages containing historical engravings, illustrations, and printed portraits. The resulting bounding boxes can be used as a preprocessing step for:

• automatic portrait cropping
• image retrieval
• metadata enrichment
• large-scale analysis of historical portrait collections
• digital humanities workflows

The model was trained on manually annotated historical images from the Deutsche Nationalbibliothek (DNB). Github Repository: https://github.com/luanamoraescosta/Buchhandlerportrats

---

Model

The trained weights are available at:

runs/detect/portrait_detector-3/best.pt

---

Performance

Validation results show:

• Best F1 Score: 0.97
• Optimal confidence threshold: 0.34

The detector performs reliably across different engraving styles, page layouts, and portrait sizes.

---

Example Predictions

The model successfully detects portraits in historical printed material with high confidence.

Example predictions are available in:

runs/detect/portrait_detector-3/

including

• validation predictions
• precision-recall curves
• F1-confidence curve
• confusion matrix
• training statistics

---

Repository Structure

runs/
└── detect/
    └── portrait_detector-3/
        ├── best.pt
        ├── last.pt
        ├── results.png
        ├── PR_curve.png
        ├── F1_curve.png
        ├── confusion_matrix.png
        ├── val_batch*.jpg
        └── ...

---

Usage

Using Ultralytics YOLO:

from ultralytics import YOLO

model = YOLO("best.pt")

results = model.predict(
    source="image.jpg",
    conf=0.34
)

---

Project

This work was carried out within the HERMES-Forschungsstudienprogramm in collaboration with the Deutsche Nationalbibliothek (DNB).

The detector was developed to support automated processing of digitized historical collections in Digital Humanities research.

---

License

Please refer to the repository license for usage conditions.