Skip to main content

Typegroups classifier for OCR

Project description


Typegroups classifier for OCR


From PyPI

pip3 install ocrd_typegroup_classifier

From source

If needed, create a virtual environment for Python 3 (it was tested successfully with Python 3.7), activate it, and install ocrd.

virtualenv -p python3 ocrd-venv3
source ocrd-venv3/bin/activate
pip3 install ocrd

Enter in the folder containing the tool:

cd ocrd_typegroups_classifier/

Install the module and its dependencies

make install

Finally, run the test:

sh test/



The model densenet121.tgc is based on a DenseNet with 121 layers, and is trained to classify pages or regions, on the following 12 classes:

  • Antiqua

  • Bastarda

  • Fraktur

  • Gotico-Antiqua

  • Greek

  • Hebrew

  • Italic

  • Rotunda

  • Schwabacher

  • Textura

  • other_font

  • not_a_font

The confusion matrix obtained with a DenseNet-121 on the pages with a single font from the dataset (see "Training a classifier" below) is:

Antiqua Bastarda Fraktur Got.-Ant. Greek Hebrew Italic Rotunda Schwabacher Textura Other font Not a font Recall
Antiqua 1531 10 5 2 5 98.6%
Bastarda 286 6 10 1 94.4
Fraktur 1933 1 5 1 2 99.5%
Gotico-Antiqua 269 1 99.6
Greek 58 1 1 96.7%
Hebrew 1 326 99.7%
Italic 1 187 99.5%
Rotunda 9 1495 5 11 1 98.3%
Schwabacher 16 4 2 452 95.4%
Textura 2 371 1 99.2%
Other font 288 15 94.1%
Not a font 4 2 2 1 5 1 7 4 2331 98.9%
Precision 99.7% 94.7% 99.1% 95.4% 96.7% 99.4% 94.9% 99.1% 94.2% 96.4% 98.3% 99.0%

Colwise classifier

The model colwise_classifier.tgc is based on a pixel column classifier composed of an RNN and a CNN backbone, it is trained to classify text lines on the following 12 classes :

  • Antiqua

  • Bastarda

  • Fraktur

  • Textura

  • Schwabacher

  • Greek

  • Italic

  • Hebrew

  • Gotico-antiqua

  • Manuscript

  • Rotunda

  • no class/ignore

Updating PyTorch

If you update PyTorch, it is possible that the model cannot be loaded anymore. To solve this issue, proceed as follows.

  1. Downgrade to a version of PyTorch which can load the model,

  2. Run the following code:

import torch
from ocrd_typegroups_classifier.typegroups_classifier import TypegroupsClassifier
tgc = TypegroupsClassifier.load('ocrd_typegroups_classifier/models/densenet121.tgc'), '')
  1. Upgrade to the desired version of PyTorch

  2. Run the following code:

import torch
from import densenet121
from ocrd_typegroups_classifier.typegroups_classifier import TypegroupsClassifier
print('Creating the network')
net = densenet121(num_classes=12)
print('Creating the classifier')
tgc = TypegroupsClassifier(
  1. delete model.mdl

If PyTorch cannot load model.mdl, then you will have to train a new model from scratch.

Training a classifier

The data used for training the densenet classifier provided in this repository is freely available at the following address:

The script in tool/ can be used to extract a suitable amount of crops to train the network, with data balancing.

The script in tools/ continues the training of any existing densenet121.tgc in the models/ folder. If there is none present, then a new one is created and trained from scratch.

Note that you might have to adapt the paths in these scripts so that they correspond to where data is in your system.

Generating activation heatmaps

For investigation purpose, it is possible to produce heatmaps showing where and how much the densenet network gets activated for specific classes.

You need first to install an additional dependency which is not required by the OCR-D tool with:

pip install tqdm

Then, you can run

python3 --layer 9 --image_path sample2.jpg

You can specify which layer of the network you are interested in, between 0 and 11. Best results are to be expected with larger values. If no layer is specified, then the 11th is used by default.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

ocrd_typegroups_classifier-0.5.0.tar.gz (26.5 MB view hashes)

Uploaded Source

Built Distribution

ocrd_typegroups_classifier-0.5.0-py3-none-any.whl (26.5 MB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page