Tesseract Training Steps¶

SETUP / DOWNLOADS¶

Download Tesseract tesseract-ocr-setup-3.02.02.exe (https://github.com/tesseract-ocr/tesseract/wiki/Downloads and https://sourceforge.net/projects/tesseract-ocr-alt/files/)
Download Tesseract tessdata files tesseract-ocr-3.02.eng.tar.gz (https://github.com/tesseract-ocr/tesseract/wiki/Data-Files#data-files-for-version-302 or https://sourceforge.net/projects/tesseract-ocr-alt/files/)
Install tesseract and copy tessdata folder into application directory
Download and install / unzip qt-box-editor-1.08win32.zip into the tesseract folder (http://zdenop.github.io/qt-box-editor/ via https://github.com/zdenop/qt-box-editor/downloads)
Download and unzip DeSkew (http://galfar.vevb.net/deskew/)

PROCEDURE¶

Train Tesseract using scanned image(s)
Filename format: [language].[font name].exp[num]
input file (sample): eng2.FontName001.exp0.tiff
language: eng2
font: FontName001
Deskew image (auto rotate)

> deskew.exe -b FFFFFF -o eng2.FontName001.exp0.deskew.tiff eng2.FontName001.exp0.tiff
> move /Y eng2.FontName001.exp0.tiff eng2.FontName001.exp0.tiff.original
> move /Y eng2.FontName001.exp0.deskew.tiff eng2.FontName001.exp0.tiff

Create box file - contains exact locations of each found characters / shapes

> tesseract.exe eng2.FontName001.exp0.tiff eng2.FontName001.exp0 batch.nochop makebox

In batch, the following can be used: tesseract.exe %1 %~n1 batch.nochop makebox
this process will create box file: eng2.FontName001.exp0.box

Start qt-box-editor-1.08.exe and fix invalid mappings. Save fixed box file.

Run the following command to start training

> tesseract eng2.FontName001.exp0.tiff eng2.FontName001.exp0 nobatch box.train

Detect the Character set used in your box files (can pass *.box for all box files):¶

> unicharset_extractor eng2.FontName001.exp0.box

Create a font_properties file or add to the existing one

# font name <italic> <bold> <fixed> <serif> <fraktur>

> echo FontName001 0 0 0 0 0 > font_properties

you may add more lines for each of the types, Ex. for bold: FontName001 0 1 0 0 0 Execute shape clustering routine based on the font properties
```
> shapeclustering -F font_properties -U unicharset eng2.FontName001.exp0.tr
```

Run Training using MF and CN training executables

> mftraining -F font_properties -U unicharset -O eng2.unicharset eng2.FontName001.exp0.tr
> cntraining eng2.FontName001.exp0.tr

Prefix relevant files with language code

> move /Y inttemp eng2.inttemp
> move /Y normproto eng2.normproto
> move /Y pffmtable eng2.pffmtable
> move /Y shapetable eng2.shapetable

Combine data into tessdata format (dot at the ned is important)

> combine_tessdata eng2.

Copy newly generated trained data file into the tessdata folder

> copy eng2.traineddata .\tessdata

Now you can run tesseract to get text into trained-output.txt file

> tesseract eng2.FontName001.exp0.tiff trained-output.txt -l eng2