eScriptorium Quick Guide

• https://escriptorium.readthedocs.io/en/latest/
• https://gitlab.inria.fr/scripta/escriptorium
• eScriptorium UNIGE instance: https://fondue.unige.ch/
• https://gitlab.unige.ch/fondue/escriptorium-fondue
• https://segmonto.github.io/
• https://github.com/mittagessen/kraken/tree/main
• https://github.com/mittagessen/kraken/blob/main/docs/user_guide/training_segmentation.rst
• https://github.com/PonteIneptique/YALTAi
• https://github.com/ultralytics/yolov5

• Model FoNDUE-GD for mss after the XVI century
• CATMuS Guidelines ⇒ Transcription principles (https://catmus-guidelines.github.io/)
• MUFI: The Medieval Unicode Font Initiative (https://mufi.info/q.php?p=mufi)

1. Create a new project * ATTENTION: Currently, it is not possible to delete or rename a project ⇒ choose a meaningful and sharable name
2. Create new document (= ms) * line offset: baseline
3. Add images to a document (wait until the upload is finished!)

• First test
  • Select first page [checkbox at the top of thumbnail]
  • Segment
    • model : default (blla.mlmodel)
    • Lines Baselines and Masks
    • Horizontal l2r
    • Check Override box
  • Check result [Edit mode : Segmentation panel]
• Manually correct the segmentation in eScriptorium [10-100 pages, the more the better]
  • Draw zones [Switch to region mode R]
  • Assign SegmOnto zone names to regions [Set the type on all selected lines/regions T]
    • DropCapitalZone for initials
    • MainZone
    • MarginTextZone
    • NumberingZone
    • QuireMarksZone
    • RunningTitleZone
    • TableZone
    • TitlePageZone
  • Re-segment the corrected images
  • Check and manually correct lines segmentation
    • Check the Toggle automatic reordering on line creation/deletion [little robot]
    • Cut lines if necessary [Cut through lines C]
    • Show order of lines [Toggle ordering display L]
    • Select all lines [ctrl+a]
    • Unlink selected lines from their regions [U]
    • Link selected lines to (the first detected) background region
    • Re-arrange lines [Text view, Toggle sorting mode]
  • Export corrected images as ALTO XML
• Train the model locally
  • Create folder and kraken environment

mkdir /Documents/kraken/
cd /Documents/kraken/
python3 -m venv kraken-env
source kraken-env/bin/activate
pip install kraken
kraken --version
pip install albumentations

• Folder structure:
• Train with kraken (not so good)

cd ../export_doc1631_averroes_logica_alto_20260325130540/
ls *.png | head -n 8 > train.txt
ls *.png | tail -n 2 > val.txt
realpath $(cat train.txt) > train_abs.txt
realpath $(cat val.txt) > val_abs.txt

ketos segtrain --load /Documents/kraken/blla.mlmodel -t export_doc1631_averroes_logica_alto_20260325130540/train_abs.txt -e export_doc1631_averroes_logica_alto_20260325130540/val_abs.txt -f alto --augment --resize union -N 50 -o averroes_seg

• Train with YALTAi and YOLO (recommended)

pip install YALTAi

cd yaltai_training/
yaltai convert alto-to-yolo Averroes*.xml my-dataset --shuffle 0.1 --segmonto region

yolo task=detect mode=train model=yolov8n.pt data=my-dataset/config.yml epochs=100 batch=4 imgsz=960 device=0

yaltai kraken --device cuda:0 -I "/Documents/kraken/averroes_complete/*.png" --suffix ".xml" segment --yolo runs/detect/train/weights/best.pt -i /Documents/kraken/blla.mlmodel

kraken alto -i Averroes_Logica_JuntSecunda_01-1_2.pdf_page_221.xml Averroes_Logica_JuntSecunda_01-1_2.pdf_page_221_alto.xml

python3 convert-json2ALTO.py 

for f in *.xml; do     sed -i 's|/home/padlina/Documents/kraken/averroes_complete/||g' "$f"; done

zip -j alto_output.zip *.xml