Table structure recognition for scholarly papers

This project contains a pipeline that takes a folder of PDF files (academic papers) and outputs CSV files of tables.

Getting Started

Install the requirements found in requirements.txt using pip install -r requirements.txt

A dataset of generated tables will be published soon. This will include the ground truth .csv files, original .tex files, .png images of the tables, .png images of the table structure.

You can generate a dataset using /tablegenerator/tablegen.py. See the README file in the tablegenerator folder for more information on this process.

Running the pipeline requires a pretrained model. At least two pretrained models will be made available: pix2pixHD and SegNet. The pix2pixHD model is based on NVIDIA's https://github.com/NVIDIA/pix2pixHD/. The SegNet model is based on https://github.com/GeorgeSeif/Semantic-Segmentation-Suite. (Encoder-Decoder with skipconnections, InceptionV4)

Running the pipeline

You can run the pipeline using python ./pipeline/batch.py. Following options are available:

• --dataroot, folder of PDF files
• --model, options: 'pix2pixHD' and 'encoder-decoder-skip'
• --checkpoint_dir, required for pix2pixHD only
• --skip_generate_images, skips the extraction of tables using pdffigures2
• --skip_predict, skips the prediction phase
• --skip_find_cells, skips finding the cells based on the outlines
• --skip_extract_text, skips extracting the text from the cells using Fitz
• --skip_create_csv, skips the creation of a csv based on the found text and cells

Installing

pip install -r requirements.txt

Pretrained models

Pretrained models and a small annotated test set can be found in the following google drive: https://drive.google.com/drive/u/0/folders/1dgKISbhBNfR8XXnIxUD_sIhwYNurKHbb

Acknowledgments

• NVIDIA's pix2pixHD
• George Seif's Semantic-Segmentation-Suite
• Allen AI's Pdffigures2
• PyMuPDF's Fitz