Base of CycleGan is from https://github.com/aitorzip/PyTorch-CycleGAN with small changes.

Base of the reading discriminator is from https://github.com/georgeretsi/HTR-ctc and heavily changed. It was primarly used for the data loader and to have a functioning base to build on top of that. The model CRNN is adopted without changes from https://github.com/olivernina/nephi

All code additions and changes were done by me as part of my master thesis and the following publication. If there are any questions about the code I'm always happy to answer them at [email protected] .

The PDF to image transformation was done with the code you find here.

If you use our software, please cite our paper as:

@inproceedings{vogtlin2021generating,
    address = {Lausanne, Switzerland},
    author = {V{\"o}gtlin, Lars and Drazyk, Manuel and Pondenkandath, Vinaychandran and Alberti, Michele and Ingold, Rolf},
    title = {{Generating Synthetic Handwritten Historical Documents With OCR Constrained GANs}},
    booktitle = {2021 16th IAPR International Conference on Document Analysis and Recognition (ICDAR)},
    year = {2021},
    month = {sep},
}