Approaches for recognizing accents

Compared to texts with Roman letters or modern (or monotonal) Greek, classical (or polytonal) Greek uses a large number of accents that can be used in a wide range of combinations. Compared to the ordinary OCR process, this requires a special treatment both for attaching accents to characters and for recognizing the resulting combinations. From a general point of view, two different approaches are possible:

Wholistic approach:

Identify each character as a whole, including its accents. This approach requires that all possible character/accents combinations have been predefined and are present as samples in the training data.

Separatistic approach:

Identify core characters and accents separately and combine them subsequently. In this case, the training data contains the core characters and the individual accents.

This toolkit offers both possibilities. You must therefore make sure that your training data matches the chosen recognition approach.

Output code

The toolkit can generate the OCR result in two different codes:

• Unicode as specified in the Unicode standards Greek (Unicode range 0370-03FF) and Combining Diacritical Marks (Unicode range 0300-036F).
• LaTeX code with the Teubner style for representing polytonal Greek accents in combination with the Babel style option polutonikogreek.

The latter option is provided for generation of a human readable graphical representation as a Postscript or PDF file via LaTeX.

Limitations

As the segmentation of the individual characters is based on a connected component analysis, the toolkit cannot deal with touching characters, unless they have been trained as combinations. It is therefore in general only applicable to printed documents, rather than handwritten documents.

From a user's perspective, there are some points to beware in this toolkit:

• It does not include methods for preprocessing like skew correction or noise removal. For this purpose, the standard routines shipped with Gamera must be used beforehand, e.g. rotation_angle_projections for skew correction, or despeckle for noise removal.
• It does not provide prototypes of the Greek characters and accents. This means that characters must be trained on sample pages before using the toolkit.
• The standard page segmentation algorithm for textline separation is currently very basic.

Prerequisites

First you will need a working installation of the following software:

• Gamera 3.x, as available from the Gamera website. It is strongly recommended that you use a recent version, preferably from SVN.
• The OCR toolkit for Gamera, as available from the "Addons" section of the Gamera website.

If you want to generate the documentation, you will need two additional third-party Python libraries:

• docutils for handling reStructuredText documents.
• pygments for colorizing source code.

Building and Installing

To build and install this toolkit, go to the base directory of the toolkit distribution and run the setup.py script as follows:

# 1) compile
python setup.py build

# 2) install
sudo python setup.py install

Command 1) compiles the toolkit from the sources and command 2) installs it. As the latter requires root privilegue, you need to use sudo on Linux and MacOS X. On Windows, sudo is not necessary.

Note that the script greekocr4gamera.py is installed into /usr/bin or /usr/local/bin on Linux and newer versions of MacOS X, but into /System/Library/Frameworks/Python.framework/Versions/2.x/bin on older MacOS X versions. As the latter directory is not in the standard search path, you could either add it to your search path, or install the scripts additionally into /usr/bin on MacOS X with:

# install scripts into standard path (older MacOS X, optional)
sudo python setup.py install_scripts -d /usr/bin

If you want to regenerate the documentation, go to the doc directory and run the gendoc.py script. The output will be placed in the doc/html/ directory. The contents of this directory can be placed on a webserver for convenient viewing.

Installing without root privileges

The above installation with python setup.py install will install the toolkit system wide and thus requires root privileges. If you do not have root access (Linux) or are no sudoer (MacOS X), you can install the GreekOcr toolkit into your home directory. Note however that this also requires that Gamera is installed into your home directory. It is currently not possible to install Gamera globally and only toolkits locally.

Here are the steps to install both Gamera and the OCR toolkit into ~/python:

# install Gamera locally
mkdir ~/python
python setup.py install --prefix=~/python

# build and install the OCR toolkit locally
export CFLAGS=-I~/python/include/python2.3/gamera
python setup.py build
python setup.py install --prefix=~/python

Moreover you should set the following environment variables in your ~/.profile:

# search path for python modules
export PYTHONPATH=~/python/lib/python

# search path for executables (eg. greekocr4gamera.py)
export PATH=~/python/bin:$PATH

Uninstallation

The installation uses the Python distutils, which do not support uninstallation. Thus you need to remove the installed files manually:

• the installed Python library files of the toolkit
• the installed standalone scripts