ocrd_eynollah

OCR-D wrapper for the Eynollah inference.

Work in progress.

Installation

For CPU use:

python -m pip install -e .[tests]

Usage

First, creating a workspace and adding image files to it:

mkdir myworkspace
cd myworkspace
ocrd workspace init
ocrd workspace add \
  -G {FILE_GRP} \
  -i {FILE_ID} \
  -m {MIMETYPE} \
  -g {PAGE_ID} \
  {PATH_TO_FILE} 

For example, OCR-D-IMG for FILE_GRP, FILE_001 for FILE_ID, image/tiff for MIMETYPE, PAGE_001 for PAGE_ID and /path/to/file.tif for PATH_TO_FILE.

Then download a specific trained Eynollah model or all available models, if needed. See ocrd-tool.json for the list of available models.

ocrd resmgr download ocrd-eynollah-inference eynollah-scale-bin-20260325-artbound-noheadings

or

ocrd resmgr download ocrd-eynollah-inference '*'

On Linux, the models will be downloaded to ~/.local/share/ocrd-resources/ocrd-eynollah-inference/model_name

Finally, run the Eynollah inference via an ocrd processor:

ocrd-eynollah-inference \
  -I {INPUT_FILE_GRP} \
  -O {OUTPUT_FILE_GRP} \
  -P model {MODEL_NAME}

For example:

ocrd-eynollah-inference \
  -I OCR-D-IMG \
  -O OCR-D-EYNOLLAH \
  -P model eynollah-scale-bin-20260325-artbound-noheadings

Results will be stored in the OUTPUT_FILE_GRP file group, including:

• PAGE XML files with the detected layout regions and their coordinates,
• Alternative image with the layout overlayed on the original image, and
• Alternative image with only the layout visualization