Eynollah

Document Layout Analysis, Binarization and OCR with Deep Learning and Heuristics

Features

• Document layout analysis using pixelwise segmentation models with support for 10 segmentation classes:
  • background, page border, text region, text line, header, image, separator, marginalia, initial, table
• Textline segmentation to bounding boxes or polygons (contours) including for curved lines and vertical text
• Document image binarization with pixelwise segmentation or hybrid CNN-Transformer models
• Text recognition (OCR) with CNN-RNN or TrOCR models
• Detection of reading order (left-to-right or right-to-left) using heuristics or trainable models
• Output in PAGE-XML
• OCR-D interface

:warning: Development is focused on achieving the best quality of results for a wide variety of historical documents using a combination of multiple deep learning models and heuristics; therefore processing can be slow.

Installation

Python 3.8-3.11 with ONNX Runtime on Linux are currently supported.

For GPU support, NVidia drivers supporting CUDA 12 must be installed. The runtime dependencies will pull in ONNX, TensorRT and CUDA runtime libraries (including cuDNN) from PyPI.

You can either install from PyPI

pip install eynollah

or clone the repository, enter it and install (editable) with

git clone git@github.com:qurator-spk/eynollah.git
cd eynollah; pip install -e .

Alternatively, you can run make install or make install-dev for editable installation.

To also install the dependencies for the OCR engines:

pip install "eynollah[OCR]"
# or
make install EXTRAS=OCR

Note: Requirements for OCR are more involved, as they may need Tensorflow (with tf-keras) and/or Torch (with transformers). Those two frameworks may also have conflicting CUDA dependencies. An ONNX conversion for these models may be achieved soon. :construction:

Docker

Use

docker pull ghcr.io/qurator-spk/eynollah:latest

When using Eynollah with Docker, see docker.md.

Models

Pretrained models can be downloaded from Zenodo or Hugging Face.

For fast runtime inference, download the ONNX models distributed as models_inference_...zip.

For finetuning training, download the original (Tensorflow / Torch) models distributed as models_training...zip (and install the [training] extra).

For model documentation and model cards, see models.md.

Training

To train your own model with Eynollah, see train.md and use the tools in the train folder.

Usage

Eynollah supports five use cases:

1. layout analysis (segmentation),
2. binarization,
3. image enhancement,
4. text recognition (OCR), and
5. reading order detection.

Some example outputs can be found in examples.md.

The generic options shared by all subcommands are:

  -m <directory containing model files>
  -mv <model category> <model variant> <model path>
  -D <device specifier>
  -l <log level>

Layout Analysis

Detects layout elements, i.e. regions of various types and text lines, and determines their reading order using either heuristic methods or a pretrained model.

The command-line interface for layout analysis can be called like this:

eynollah [GENERIC_OPTIONS] layout \
  -i <single image file> | -di <directory containing image files> \
  -o <output directory> \
     [OPTIONS]

The following options can be used to further configure the processing:

The default is to only perform layout detection of main regions (background, text, images, separators and marginals).

The best output quality is achieved when RGB images are used as input rather than greyscale or binarized images.

Additional documentation can be found in usage.md.

Binarization

Performs document image binarization (thresholding) using pretrained pixelwise segmentation models.

The command-line interface for binarization can be called like this:

eynollah [GENERIC_OPTIONS] binarization \
  -i <single image file> | -di <directory containing image files> \
  -o <output directory> \
  [OPTIONS]

Image Enhancement

This enlarges and enhances images. Useful in case the scan quality is low.

eynollah [GENERIC_OPTIONS] enhancement \
  -i <single image file> | -di <directory containing image files> \
  -o <output directory> \
  [OPTIONS]

OCR

Performs text recognition using either a CNN-RNN model or a Transformer model. Needs a PAGE-XML input file.

The command-line interface for OCR can be called like this:

eynollah [GENERIC_OPTIONS] ocr \
  -i <single image file> | -di <directory containing image files> \
  -dx <directory of xmls> \
  -o <output directory> \

The following options can be used to further configure the ocr processing:

Reading Order Detection

Reading order can be detected either during layout analysis, or as a separate module, which requires a PAGE-XML input file.

The command-line interface for machine based reading order can be called like this:

eynollah [GENERIC_OPTIONS] machine-based-reading-order \
  -i <single image file> | -di <directory containing image files> \
  -xml <xml file name> | -dx <directory containing xml files> \
  -o <output directory> 

Use as OCR-D processor

See ocrd.md.

How to cite

@inproceedings{hip23rezanezhad,
  title     = {Document Layout Analysis with Deep Learning and Heuristics},
  author    = {Rezanezhad, Vahid and Baierer, Konstantin and Gerber, Mike and Labusch, Kai and Neudecker, Clemens},
  booktitle = {Proceedings of the 7th International Workshop on Historical Document Imaging and Processing {HIP} 2023,
               San José, CA, USA, August 25-26, 2023},
  publisher = {Association for Computing Machinery},
  address   = {New York, NY, USA},
  year      = {2023},
  pages     = {73--78},
  url       = {https://doi.org/10.1145/3604951.3605513}
}