mptv

Tools necessary to perform a multi-fold pretrained voting approach utlizing OCRopus in order to significantly improve the achievable OCR error rate.
Please note that this code was only used to provide a proof of concept for the publications listed below including the output of several (unnecessary) intermediate results. In order to efficiently implement the concepts into an OCR workflow it is highly recommend to use Christoph Wick's (@chwick) Calamari (see [4]) which is under active development and natively supports cross fold training, confidence voting and pretraining.

Installing

Clone Repository

git clone https://github.com/chreul/mptv.git

Setup and Activate Virtual Enviroment

python -m pip install --user virtualenv

python -m virtualenv path/to/venv

source path/to/venv/bin/activate

Install Requirements and Run MPTV Setup

pip install -r requirements.txt

python setup.py install

Usage

Use Calamari instead...

mptv-setup_folds

Input: image list (*.png, *.bin.png, *.nrm.png, + corresponding GT (*.gt.txt))
Output: folder (--output, default: ./Data) with n (--folds, default = 5) sub folders which contain #lines/n lines

The folder id gets placed in front of the line name in order to prevent overwriting of lines with the same name.

mptv-run_multi_train

Input: folder containing the "Folds" sub folder.
Output: "Models" folder which containes the trained models for each fold.

--models: "LH,LH,FRK,ENG,ENG" means that the first two folds starts training from LH fold 3 from ENG, and fold 4 and 5 from FRK with LH, ENG,
and FRK being different mixed models from the pretraining folder. "LH,,,,LH" starts the training of fold 1 and 5 from LH and trains the rest from scratch.

mptv-find_best_model

Input: folder containing "Folds" and "Models" sub folders.
Output: the best model of each fold stored in "Models/BestModels".

mptv-recognize_lines

Input: folder(s) containing the to-be-predicted lines.
Output: folder "Voting" in each input folder which contains the results (*.txt and *.extLlocs) for each fold.

mptv-conf_vote

Input: "Voting" folder(s) resulting from the preceding step.
Output: "Voting/Voted" folders containing voted *.txts.

Corresponding Publications

Cross Fold Training and Confidence Voting

[1] Reul, Springmann, Wick, Puppe. (2017). Improving OCR Accuracy on Early Printed Books by Utilizing Cross Fold Training and Voting. Accepted for the 13th IAPR International Workshop on Document Analysis Systems, DAS2018. Preprint available at ArXiv e-prints.

@article{reul2017voting,
  title = {Improving {OCR} Accuracy on Early Printed Books by Utilizing Cross Fold Training and Voting},
  author = {Reul, Christian and Springmann, Uwe and Wick, Christoph and Puppe, Frank},
  journal = {ArXiv e-prints},
  url = {https://arxiv.org/abs/1711.09670},
  year = {2017}
}

PreTraining/Transfer Learning

[2] Reul, Wick, Springmann, Puppe. (2017). Transfer Learning for {OCR}opus Model Training on Early Printed Books. 027.7, Journal for Library Culture. Available online.

@article{reul2017transfer,
  author = {Reul, Christian and Wick, Christoph and Springmann, Uwe and Puppe, Frank},
  title = {Transfer Learning for {OCR}opus Model Training on Early
		  Printed Books},
  journal = {027.7 Zeitschrift für Bibliothekskultur / Journal for
		  Library Culture},
  volume = {5},
  number = {1},
  year = {2017},
  pages = {38-51},
  url = {http://dx.doi.org/10.12685/027.7-5-1-169}
}

Combination

[3] Reul, Springmann, Wick, Puppe. (2018). Improving OCR Accuracy on Early Printed Books by combining Pretraining, Voting, and Active Learning. Accepted for JLCL Volume 33 (2018), Issue 1: Special Issue on Automatic Text and Layout Recognition. Available online.

@article{reul2018combining,
  author = {Reul, Christian and Springmann, Uwe and Wick, Christoph and Puppe, Frank},
  title = {Improving OCR Accuracy on Early Printed Books by combining Pretraining, Voting, and Active Learning},
  journal = {ArXiv e-prints},
  url = {https://arxiv.org/abs/1802.10038},
  year = {2018}
}

Calamari

[4] Wick, Reul, Puppe. (2018). Calamari - A High-Performance Tensorflow-based Deep Learning Package for Optical Character Recognition. Submitted to Digital Humanities Quarterly. Preprint available at ArXiv e-prints.

@article{wick2018calamari,
  title = {Calamari - A High-Performance Tensorflow-based Deep Learning Package for Optical Character Recognition},
  author = {Wick, Christoph and Reul, Christian and Puppe, Frank},
  journal = {ArXiv e-prints},
  url = {https://arxiv.org/abs/1807.02004},
  year = {2018}
}