This repository lets you train neural networks models for performing end-to-end full-page handwriting recognition using the Apache MXNet deep learning frameworks on the IAM Dataset.

These notebooks have been created by Jonathan Chung, as part of his internship as Applied Scientist @ Amazon AI, in collaboration with Thomas Delteil who built the original prototype.


git clone https://github.com/awslabs/handwritten-text-recognition-for-apache-mxnet --recursive

You need to install SCLITE for WER evaluation
You can follow the following bash script from this folder:

cd ..
git clone https://github.com/usnistgov/SCTK
export CXXFLAGS="-std=c++11" && make config
make all
make check
make install
make doc
cd -

You also need hsnwlib

pip install pybind11 numpy setuptools
cd ..
git clone https://github.com/nmslib/hnswlib
cd hnswlib/python_bindings
python setup.py install
cd ../..


The pipeline is composed of 3 steps:

The entire inference pipeline can be found in this notebook. See the pretrained models section for the pretrained models.

A recorded talk detailing the approach is available on youtube. [video]

The corresponding slides are available on slideshare. [slides]

Pretrained models:

You can get the models by running python get_models.py:

Sample results

The greedy, lexicon search, and beam search outputs present similar and reasonable predictions for the selected examples. In Figure 6, interesting examples are presented. The first line of Figure 6 show cases where the lexicon search algorithm provided fixes that corrected the words. In the top example, “tovely” (as it was written) was corrected “lovely” and “woved” was corrected to “waved”. In addition, the beam search output corrected “a” into “all”, however it missed a space between “lovely” and “things”. In the second example, “selt” was converted to “salt” with the lexicon search output. However, “selt” was erroneously converted to “self” in the beam search output. Therefore, in this example, beam search performed worse. In the third example, none of the three methods significantly provided comprehensible results. Finally, in the forth example, the lexicon search algorithm incorrectly converted “forhim” into “forum”, however the beam search algorithm correctly identified “for him”.


  • To use test_iam_dataset.ipynb, create credentials.json using credentials.json.example and editing the appropriate field. The username and password can be obtained from http://www.fki.inf.unibe.ch/DBs/iamDB/iLogin/index.php.

  • It is recommended to use an instance with 32GB+ RAM and 100GB disk size, a GPU is also recommended. A p3.2xlarge would be the recommended starter instance on AWS for this project


1) Handwritten area

Model architecture


2) Line Detection

Model architecture


3) Handwritten text recognition

Model architecture