Nikyek/avestan-ocr-kraken-v1

Avestan OCR Training and Application – Kraken + eScriptorium

This folder contains the full training and application pipeline for Avestan OCR using Kraken and eScriptorium. It handles image segmentation, recognition model training, and output generation using ALTO XML. Outputs are later converted to CAB-compatible XML formats using tools from the xml_translator/ module.

---

Folder Structure

Applying_OCR/
├── Makefile                  # Defines all Kraken training and evaluation targets
├── models/                  # Stores trained segmentation and recognition models
│   ├── segmentation/
│   └── recognition/

---

Workflow Overview

Hugging Face model cards don’t render Mermaid.
See a rendered version on GitHub, or expand the source below.

View rendered diagram on GitHub →

---

Makefile Targets

The Makefile defines the Kraken training and evaluation pipeline. Example targets include:

train_seg:
    kraken train -i data/segmentation/*.xml -o models/segmentation/model.mlmodel

train_recog:
    kraken train -i data/recognition/*.xml -o models/recognition/model.mlmodel

eval:
    kraken eval -m models/recognition/model.mlmodel -i test/*.xml

Use make train_seg, make train_recog, or define your own targets for batch training/evaluation.

---

Input/Output Formats

Input:

• Line-segmented manuscript images (from eScriptorium or Kraken segmenter)
• ALTO XML files with gold-standard transcriptions

Output:

• .mlmodel files (segmentation + recognition)
• ALTO XML predictions from Kraken
• CAB-format XML via xml_translator/

---

Dependencies

This pipeline assumes you have:

• Kraken installed (pip install kraken)
• eScriptorium for GUI-assisted segmentation/transcription
• Pre-cleaned ALTO XML exported from eScriptorium