Semantic structure identification model for general contract documents using MASK_RCNN. Generates JSON file containing semantic structures of each component in a document.

Semantic labels are below:

• Title
• Subtitle
• Paragraph
• Footnotes
• Header
• Footer
• Page
• Signature

  git clone https://github.com/kasimebrahim/instance_segmentation
  cd instance_segmentation

  # Open conda_environment.yml and change the last line to your own conda environment path.
  conda env create -f conda_environment.yml

  pip install -r requirements.txt

You will need python>3.7 and optionally conda>4.7.10

To inspect and visualize your dataset use dataset_inspect.

To evaluate your model or visualize your output use model_eval.

To train your model

  # If you want to train from a pretrained model.
  python Segmentation.py train --datasets=datasets --log=log --model=models/mask_rcnn_pub_lay_seg_0100.h5
  # If you dont have.
  python Segmentation.py train --datasets=datasets --log=log
  # If you want to pickup from a stoped training.
  python Segmentation.py train --datasets=datasets --log=log --pickup=true

To segment your documents

  python Segmentation.py segment --model=models/mask_rcnn_doc_seg_0100.h5 --image=infer
  # --image is the directory where your documents to be segmented are stored.
  # Your documents should be stored in a directory in two ways.
  # One: Each documenat is in its own directory and every
  #      page of the document is in the directory.
  #      i:e ds/doc/p01.jpg
  # Two: All pages of all the documents are under one
  #      directory. And every image is named as document_page
  #      concatinated with page name/number.
  #      i:e ds/doc_p01.jpg

The out put of the segmented documents will be stored in a json file named "documents.json".

This repository heavily reuses code from the amazing tensorflow Mask RCNN implementation by @waleedka. Many thanks to all the contributors of that project. You are encouraged to checkout https://github.com/matterport/Mask_RCNN for documentation on many other aspects of this code.

Kasim Ebrahim [email protected]