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
Link
ICDAR 2019 competition datasetCascadeTabNet | 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}
}