Training Tesseract

Creating Training Text

To train tesseract, first create some training text. Make sure the text is not too long, because this will make training take forever, but make sure it includes around at least 10 of each character you want the language trained on.

Creating Training Images

Tesseract has an option to generate images from text doing training. To do this, run:

text2image --text=training_text.txt --outputbase=[lang_code].[fontname].exp0 --font='Font Name' --fonts_dir=/path/to/your/fonts

On Ubuntu, fonts are usually at /usr/share/fonts , but this path is platform specific. If you are training on multiple fonts, you will have to run this command multiple times. For the purposes of text2image, italics are considered a different font (you will have to run it once for Times, and once for Times Italic, for example)

For example, if you are training for Tuvan with Times New Roman:

text2image --text=training_text.txt --outputbase=tyv.TimesNewRoman.exp0 --font='Times New Roman' --fonts_dir=/usr/share/fonts

Training

Generating .tr files

The first step in training is generating tr files from the images you created. Do this by running:

$ tesseract [lang_code].[fontname].exp0.tif [lang_code].[fontname].exp0 box.train

You will have to run this command for each font. For Tuvan and Times New roman:

$ tesseract tyv.TimesNewRoman.exp0.tif tyv.TimesNewRoman.exp0 box.train

Character set

To get the charset, run:

$ unicharset_extractor [lang].*.exp0.box

This gets all the box files, so you can run that command verbatim for all fonts.

For Tuvan:

$ unicharset_extractor tyv.*.exp0.box

Fontproperties file

You must specify a font_properties file, with each line a font in the following format:

<fontname> <italic> <bold> <fixed> <serif> <fraktur> , where you fill in each property with a 1 or 0 depending on whether the property exists. For example for Times new roman italic, a serif font:

timesitalic 1 0 0 1 0

Clustering

Run the following 3 commands:

   $ shapeclustering -F font_properties -U unicharset [lang].*.exp0.tr  (only for indic languages)
   $ mftraining -F font_properties -U unicharset -O [lang].unicharset [lang].*.exp0.tr 
   $ cntraining [lang].*.exp0.tr 

Rename the files normproto, pffmtable, inttemp to be prefixed with <lang_code>.

For Tuvan:

   $ mftraining -F font_properties -U unicharset -O lang.unicharset tyv.*.exp0.tr 
   $ cntraining tyv.*.exp0.tr 

And prefix normproto, pffmtable, inttemp with tyv.

DAWG files

It is recommended that you have a list of word bigrams (line separated) and a wordlist (also line separated, but not necessarily complete). Run:

$ wordlist2dawg wordlist [lang].word-dawg lang.unicharset

$ wordlist2dawg bigram_list [lang].bigram-dawg lang.unicharset

For Tuvan:

$ wordlist2dawg wordlist tyv.word-dawg tyv.unicharset

$ wordlist2dawg bigram_list tyv.bigram-dawg tyv.unicharset

Final steps

Run $ combine_tessdata lang_code. to get the final .traineddata file.

For Tuvan:

   $ combine_tessdata tyv.

The output will be $ tyv.traineddata