Automatic Text Recognition (ATR) with post-ATR correction with LLM

Running the ATR Script

1. Set Up the Environment

Before running the script, it is recommended to install Kraken in a Python virtual environment. This helps isolate dependencies and prevents conflicts with other projects.

python3 -m venv env_name
source env_name/bin/activate
pip install kraken

2. Execute the Script

Once Kraken is installed, you can launch the script using the following command:

python3 atr_kraken.py <images_dir> <output_dir> <segmentation_model> <recognition_model> <output_format>

Argument Descriptions

• <images_dir>: Path to the folder containing the input images

• <output_dir>: Path to the folder where output files will be saved

• <segmentation_model>: Path to the segmentation model (Kraken .mlmodel)

• <recognition_model>: Path to the text recognition model (Kraken .mlmodel)

• <output_format>: Output format, choose one of the following:

  • -n: Plain text (.txt)
  • -a: ALTO XML format (.xml)
  • -x: PAGE XML format (.xml)

---

Selecting the Right Models

All the models have been collected from the Zenodo Kraken OCR/HTR models repository.

Segmentation

Model Name	Model Filename	Model Link
General segmentation model for print and handwriting	blla.mlmodel	https://doi.org/10.5281/zenodo.14602569
Kraken segmentation model for vital records	seg_vital_records.mlmodel	https://doi.org/10.5281/zenodo.11913537
Kraken segmentation model for two-column prints	seg_news_1.0.mlmodel	https://doi.org/10.5281/zenodo.10783346

Recognition

Model Name	Model Filename	Model Link
CATMuS Medieval	catmus-medieval-1.6.0.mlmodel	https://doi.org/10.5281/zenodo.15030337
CATMuS-Print [Large]	catmus-print-fondue-large.mlmodel	https://doi.org/10.5281/zenodo.10592716
Transcription model for Lucien Peraire's handwriting (French, 20th century)	peraire2_ft_MMCFR.mlmodel	https://doi.org/10.5281/zenodo.8193498
Printed Arabic-Script Base Model Trained on the OpenITI Corpus	all_arabic_scripts.mlmodel	https://doi.org/10.5281/zenodo.7050270
HTR-United - Manu Mc Fondue (Manuscripts of Modern and Contemporaneous French - Manu McFrench v4)	ManuMcFondue.mlmodel	https://doi.org/10.5281/zenodo.10886224
McCATMuS - Transcription model for handwritten, printed and typewritten documents from the 16th century to the 21st century	McCATMuS_nfd_nofix_V1.mlmodel	https://doi.org/10.5281/zenodo.13788177

---

Running the LLM Correction Script

1. Set Up the Environment

Before running the script, it is recommended to install OpenAI in a Python virtual environment. This helps isolate dependencies and prevents conflicts with other projects.

python3 -m venv env_name
source env_name/bin/activate
pip install openai

2. Execute the Script

Once OpenAI is installed, you can launch the script using the following command:

python3 correction_llm.py <input_folder> <output_folder>

Argument Descriptions

• <input_folder>: Path to the folder containing the input texts
• <output_folder>: Path to the folder where output files will be saved

WARNING: Don't forget to modify the script to put your own openai key!

---