TesseractTrainer

by Balthazar Rouberol - Mor Consulting, rouberol.b@gmail.com

TesseractTrainer is a simple Python API, taking over the tedious process of manually training Tesseract3, as described in the wiki page.

The longest part of the training process is checking the box file, generated by tesseract using a reference tif image, as explained here. This file contains the coordinates of each character detected in the training tif. However, if Tesseract made some mistakes, you have to manually correct the boxfile, allowing Tesseract to "learn" from its mistakes.

TesseractTrainer allows you to skip this part, by automatically generating a tif (and the associated boxfile) using a text and a font that you specify, thus guaranteeing the total accuracy of the box file.

TesseractTrainer intends to provide both a python API and a bash command line tool.

Dependencies

• a Unix/Linux system
• Tesseract3.01 (Tesseract 3.02 compatibility is on the way)
• python 2.6 - 2.7, 3.2 - 3.3
• Pillow (Python Imaging Library fork) Following libraries should be already installed before Pillow installation
  • libjpeg62
  • libjpeg-dev
  • libfreetype6
  • libfreetype6-dev
  • zlib1g-dev otherwise, some freetype/zlib/jpeg operations will not be supported.
• ImageMagick

Installation

You can install TesseractTrainer in your virtualenv via

pip install TesseractTrainer

or

easy_install TesseractTrainer

if you really (really) have to.

Command line tool

usage: tesstrain [-h]
                  --tesseract-lang TESSERACT_LANG
                  --training-text TRAINING_TEXT
                  --font-path FONT_PATH
                  --font-name FONT_NAME
                  --font-properties FONT_PROPERTIES
                  [--experience_number EXPERIENCE_NUMBER]
                  [--font-size FONT_SIZE]
                  [--tessdata-path TESSDATA_PATH]
                  [--word_list WORD_LIST]
                  [--verbose]

tesstrain arguments

  -h, --help            show this help message and exit

**Required arguments:**
  --tesseract-lang TESSERACT_LANG, -l TESSERACT_LANG
                        Set the tesseract language traineddata to create.
  --training-text TRAINING_TEXT, -t TRAINING_TEXT
                        The path of the training text.
  --font-path FONT_PATH, -F FONT_PATH
                        The path of TrueType/OpenType file of the used training font.
  --font-name FONT_NAME, -n FONT_NAME
                        The name of the used training font. No spaces.
  --font-properties FONT_PROPERTIES, -f FONT_PROPERTIES
                        The path of a file containing font properties for a list of training fonts.

**Optional arguments**
  --experience_number EXPERIENCE_NUMBER, -e EXPERIENCE_NUMBER
                        The number of the training experience.
                        Default value: 0
  --font-size FONT_SIZE, -s FONT_SIZE
                        The font size of the training font, in px.
                        Default value: 25
  --tessdata-path TESSDATA_PATH, -p TESSDATA_PATH
                        The path of the tessdata/ directory on your filesystem.
                        Default value: /usr/local/share/tessdata
  --word_list WORD_LIST, -w WORD_LIST
                        The path of a file containing a list of frequent words.
                        Default value: None
  --verbose, -v         Use this argument if you want to display the training
                        output.

Examples

In this example, we would like to create a helveticanarrow dictionary:

• using an OpenType file located at ./font/Helvetica-Narrow.otf
• the font name is set to helveticanarrow
• with training text located at ./text
• the font_properties file is located at ./font_properties. It contains the following line: helveticanarrow 0 0 0 0 0
• the experience number is set to 0
• a tif font size of 25px
• the tessdata directory is located at /usr/local/share/tessdata
• no frequent word list

The command would thus be:

$ tesstrain --tesseract-lang helveticanarrow --training-text ./text --font-path font/Helvetica-Narrow.otf --font-name helveticanarrow  --font-properties ./font_properties --verbose

or using the short options names:

$ tesstrain -l helveticanarrow -t ./text -F ./font/Helvetica-Narrow.otf -n helveticanarrow -f ./font_properties -v

Python API

The tesseract_train.py file offers a very simple API, defined through the class TesseractTrainer. This class has only 4 public methods:

• __init__(self, text, exp_number, dictionary_name, font_name, font_size, font_path, font_properties, tessdata_path, word_list): returns a TesseractTrainer instance
• training(self): performs all training operations, thus creating a traineddata file.
• add_trained_data(self): copies the generated traineddata file to your tessdata directory
• clean(self): deletes all files generated during the training process (except for the traineddata one).

Example

from tesseract_trainer import TesseractTrainer

trainer = TesseractTrainer(dictionary_name='helveticanarrow',
                            text='./text',
                            font_name='helveticanarrow',
                            font_properties='./font_properties',
                            font_path='./font/Helvetica-Narrow.otf')
trainer.training()  # generate a multipage tif from args.training_text, train on it and generate a traineddata file
trainer.clean()  # remove all files generated in the training process (except the traineddata file)
trainer.add_trained_data()  # copy the traineddata file to the tessdata/ directory

Note that the same default values apply than when using the tesstrain command:

font_size = 25
exp_number = 0
tessdata_path = "/usr/local/share/tessdata"
word_list = None
verbose = True

You can override these constants when instanciating a TesseractTrainer object, to better suit your needs.

Remarks

• For now, only Tesseract 3.01 training can be automated. Adding Tesseract 3.02 support seems fairly simple, but I'm facing a tricky bug from tesseract. I'm hoping investigation with the tesseract dev team will resolve it (see here.
• UTF-8 encoding is supported.
• For now, training on "right to left" languages (ie: Arabic, for example) is not supported.
• If your tessdata directory is not writable without superuser rights, use the sudo command when executing your python script.
• Do not forget to describe your font properties in a file, following these instructions.

Installation troubleshoot

PIL installed without JPEG, PNG and freetype support in a virtualenv

If you install TesseractTrainer in a virtualenv, PIL will have to be installed along. A well known problem is that it might not be installed with JPEG, PNG and freetype support (see this thread).

In that case, follow these instructions, to have a working PIL installation in your virtualenv.

Attributions

TesseractTrainer was completed whilst working on StrongSteam for MorConsulting.