Deep-learning and graph-based approach to table structure recognition

This is an official implementation of Graph-based-TSR on Python 3, TensorFlow.

Requirements

Use the following command:

    conda env create -f graphtsr.yml

You have to get the license of Gurobi optimizer.

Data

Document data used in paper are stored in data folder.

See explanation in data/config.txt.

Execution

Train

inside codes folder,

    python code_training/main.py --mode train --batch_size 6 --experiment_name model --data_dir ../data/ctdar19_B2_m/train/ --NUM_STACKS 2 --num_epochs EPOCHS --gpu GPU

See detailed arguments in codes/code_training/set_default_training_options.py.

Test

inside codes folder,

    python code_training/main.py --mode test --batch_size 1 --experiment_name bordered --test_data_dir ../data/ctdar19_B2_m/test/SCAN/img/ --test_name test --test_scan True

See detailed arguments in codes/code_training/set_default_training_options.py.

Table Structure Recognition results

ICDAR 2019 competition dataset Link

CascadeTabNet	TabStructNet	SPLERGE	Proposed

Scanned ICDAR 2019 competition dataset

CascadeTabNet	TabStructNet	SPLERGE	Proposed

Scanned hospital receipts

CascadeTabNet	TabStructNet	SPLERGE	Proposed

Scanned hand-drawn documents

CascadeTabNet	TabStructNet	SPLERGE	Proposed

Please use this to cite our work:

@article{lee2021deep,
  title={Deep-learning and graph-based approach to table structure recognition},
  author={Lee, Eunji and Park, Jaewoo and Koo, Hyung Il and Cho, Nam Ik},
  journal={Multimedia Tools and Applications},
  pages={1--22},
  year={2021},
  publisher={Springer}
}