Handwriting Recognition

---

This repository contains all code for the three tasks of the Handwriting Recognition, Master's course of the University of Groningen.

Team members

• Jeroen
• Milan
• Manuel
• Abhishek

---

Dead Sea Scrolls

NOTE: Unfortunately, the blob-line segmentation method used in this pipeline was only able to be implemented in Windows. However, the horizontal histogram projection method for line segmentation should work on all systems. Documentation on the method was not sufficient to find an appropriate adaptation to Linux and Mac. Our apologies for this inconvenience.

Matlab engine Setup

For the line segmentation, two approaches are used: a histogram method and a blob-line method. The histogram approach uses a peak detection algorithm by Roman Ptak et al.[1], that incorporates a variable threshold. The latter uses the implementation by Kurar Barakat et al.[2] in Matlab. For this to function correctly, the user must have the MATLAB Engine installed. This enables MATLAB code to be run directly from a Python script. The MATLAB engine can be installed as follows:

Verify installation

• Before you install, verify your Python and MATLAB configurations.
• Check that your system has a supported version of Python and MATLAB R2014b or later. Version R2021b was used in our case. This version supports Python 3.7, 3.8 and 3.9.
• Next, find the path to the MATLAB folder. Start MATLAB and type matlabroot in the command window. Copy the path returned by matlabroot.

Windows

Open a command prompt and type the following commands:

cd "MATLABROOT\extern\engines\python"
python3 setup.py install

Matlab Dependencies

• Image Processing Toolbox
• MATLAB Support for MinGW-w64 C/C++ Compiler
• MATLAB Compiler
• MATLAB Compiler SDK

Python Package dependencies

The python package dependencies can be found in dss_recognition/requirements.txt

Running the pipeline

To run the full pipeline for the line segmentation, character segmentation and character recognition, first ensure the dataset is using the correct naming scheme. This means all binarized versions of files contain 'binarized' in the filename. For example, a file can be named: P123-Fg001-R-C01-R01-binarized.jpg

For the pipeline to take the data, either deposite all image files in the image-data folder in the dss_recognition directory, or specify the location as a command line argument. By default, the resulting text files of the model are saved to a results directory in the dss_recognition directory. However, if the user prefers, a custom saving directory can also be used by specifying this in the command line as well. --line_segment_method is a flag to specify which line segmenter to use, it can either be the blob-line method ('blob') or horizontal histogram projection ('hhp').

Running the pipeline is done by running the following file from the dss_recognition directory:

python3 dss_recognizer.py --dir_images <path_to_dir> --dir_save_predictions <path_to_dir> --line_segment_method <method>

IAM Dataset

Code Inspiration

Some code inspired from https://github.com/GitYCC/crnn-pytorch and https://github.com/kris314/deep-text-recognition-benchmark

Setup

The python package dependencies can be found in iam_line_recognition/requirements.txt

To train the model

• To train the model run the following

python3 train.py

• To list all the commandline arguments, run the following

python3 train.py --help

Running the final trained model to generate predictions

• The script line_recognition/final_iam_line_recognizer.py can be run in the following way. Use --which_hw_model option to specify the model to be used. To run CRNN use --which_hw_model crnn and to run STN-CRNN use --which_hw_model stn_crnn
• The predictions of the model will be saved in individual .txt files i.e. one for each image, in a directory named results_crnn with CRNN model and results_stn_crnn with STN-CRNN model

python3 final_iam_line_recognizer.py --dir_images <full_path_to_dir> --which_hw_model <hw_model> --file_model <full_path_to_model_file>

References

[1] Ptak, R., Zygadlo, B., Unold, O. (2017). Projection-Based Text Line Segmentation with a Variable Threshold. International Journal of Applied Mathematics and Computer Science, 27, doi:10.1515/amcs-2017-0014. paper link

[2] Berat Kurar, B., Cohen, R., Droby, A., Rabaev, I. & El-Sana, J. (2020). Learning-Free Text Line Segmentation for Historical Handwritten Documents. Applied Sciences, 10, 8276; doi:10.3390/app10228276. paper link

[3] Dutta, Kartik & Krishnan, Praveen & Mathew, Minesh & Jawahar, C.V.. (2018). Improving CNN-RNN Hybrid Networks for Handwriting Recognition. 80-85, doi:10.1109/ICFHR-2018.2018.00023. paper link