This repo is a clone of https://github.com/dhavalpotdar/Graph-Convolution-on-Structured-Documents which has be enriched with some scripts to handle RVL-CDIP Dataset.
The script label_data.py generate numpy matrix A, X and Y corresponding to adjacency, node features and labels respectively.
python label_data.py zenedo_dataset
This repo contains code to convert Structured Documents to Graphs and implement a Graph Convolution Neural Network (incomplete) for Node Classification, each node being an entity in the document.
Check out the article for an intuitive explanation on Towards Data Science: Using Graph Convolutional Neural Networks on Structured Documents for Information Extraction
The grapher.py file contains the code to convert a structured document to a graph.
An object map made using a Commercial OCR Tool is needed as the input which provides the bounding-box coordinates of each entity in the image along with it's recognized text. The script can then be used to generate an object_tree.png file and a
connections.csv file. The script joins each object to it's nearest object to the right and underneath thus generating a graph.
Here is what the generated graph looks like:
The implementation is still in progress and is being built using Tensorflow 1.8. The implementation details can be found in [1].
- Riba, Dutta et al - Table Detection in Invoice Documents by Graph Neural Networks - Link
- Adam W. Harley, Alex Ufkes, and Konstantinos G. Derpanis - Department of Computer Science, Ryerson University, Toronto, Ontario - Evaluation of Deep Convolutional Nets for Document Image Classification and Retrieval - Link
- Victor Garcia, Joan Bruna - Few-Shot Learning with Graph Neural Networks - Link