TessPage

Toolset for Tesseract training with PageXML Ground-Truth

Install & Setup

1. Clone tesspage:

   $ git clone https://github.com/jahtz/tesspage
   

2. Install dependencies:

   $ sudo apt install -y tesseract-ocr libtesseract-dev libtool pkg-config make wget bash unzip bc
   $ cd tesspage
   $ pip install -r requirements.txt
   

3. Setup:

   $ python3 tesspage.py setup
   

Structure

(after setup)

tesspage
│ README.md                 
│ requirements.txt          required pip packages
│ LICENSE                   license
└─ tesspage                 tesspage files
│  └ ...
└─ tesstrain¹               tesstrain files
│  │ data                   trained model data 
│  └ ...
└─ data
│  │ eval                   default dir for evaluation
│  │ ground_truth           default dir for ground_truth output
│  │ ocr_input              default dir for tesseract image input
│  │ ocr_output             default dir for tesseract output
│  │ tessconfigs²           tesseract config files
│  │ tessdata_best³         training start model 
│  └ training_data          default dir for pagexml input
└─ tesspage.py              entry point

¹ tesstrain, ² tessconfigs, ³ tessdata_best

Usage

Generate Ground-Truth

Copy PageXML + Image Files to ./data/training_data (or custom folder)

Run script to generate single line image files and matching ground truth .txt files:

python3 tesspage.py generate [--training_data <input_folder>] [--ground_truth <output_folder>]

• --training_data: input folder containing pagexml and image files [default: ./data/training_data/]
• --ground_truth: output folder (line image and text files after exec) [default: ./data/ground_truth/]

Train Model

Run script to train custom Tesseract model from base model with single line image files and ground truth .txt files

python3 tesspage.py training [--model_name <name>] [--start_model <model>] [--data_dir <folder>] [--ground_truth <folder>] [--tessdata <folder>] [--max_iterations <number>] [ARGS ...]

• --model_name: name of trained model [default: foo]
• --start_model: select start model. Previously trained model or lang-code (e.g. "eng") from langdata [default: eng]
• --data_dir: tesstrain data dir [default: ./tesstrain/data/]
• --ground_truth: ground truth folder (line image and text files) [default: ./data/ground_truth/]
• --tessdata: training start model folder [default: ./data/tessdata_best/]
• --max_iterations: training iterations [default: 10000]
• ARGS: Full argument list here

Run Tesseract

Run Tesseract OCR with custom model

python3 tesspage.py tesseract --model_name <name> [--input <path>] [--output <path>] [--data_dir <folder>] [--config_dir <config_dir>] [--config <config>] [ARGS ...]

• --model_name: select model, either language or custom trained model
• --input: input directory or image file
• --output: output directory
• --data_dir: tesstrain data dir [default: ./tesstrain/data/]
• --config_dir: Output config directory. [default: ./data/tessconfigs/configs/]
• --config: Config file to be used (txt, pdf, hocr, tsv, pagexml, ...) [default: txt]
• ARGS: guide here

Evaluate Model

Run to evaluate trained models (CER/WER)

python3 tesspage.py eval [--eval_input <folder>]

• --eval_input: supports .txt, .hocr and .xml files [default: ./data/eval/]

Name Pattern:

• Reference files: <name>.gt.<extension>
• Prediction files: <name>.<extension>
• Supported extensions: .txt, .hocr, .xml
• Example: 0001.gt.xml / 0001.xml

Help

python3 tesspage.py -h

ZPD

Developed at Zentrum für Philologie und Digitalität at the Julius-Maximilians-Universität of Würzburg