/TableMASTER-mmocr

2nd solution of ICDAR 2021 Competition on Scientific Literature Parsing, Task B.

Primary LanguagePythonApache License 2.0Apache-2.0

TableMASTER-mmocr

New

  • 2021/09: TableMASTER pretrained model release.
  • 2021/10: Lmdb training data release.
  • 2021/11: tablemaster_mmocr docker environment release.
  • 2021/12: End2end demo inference release, pretrained model release.

Contents
  1. About The Project
  2. Getting Started
  3. Usage
  4. Result
  5. Pretrained model
  6. Lmdb data
  7. License
  8. Acknowledgements

About The Project

This project presents our 2nd place solution for ICDAR 2021 Competition on Scientific Literature Parsing, Task B. We reimplement our solution by MMOCR，which is an open-source toolbox based on PyTorch. You can click here for more details about this competition. Our original implementation is based on FastOCR (one of our internal toolbox similar with MMOCR).

Method Description

In our solution, we divide the table content recognition task into four sub-tasks: table structure recognition, text line detection, text line recognition, and box assignment. Based on MASTER, we propose a novel table structure recognition architrcture, which we call TableMASTER. The difference between MASTER and TableMASTER will be shown below. You can click here for more details about this solution.

MASTER's architecture

Dependency

Getting Started

Prerequisites

  • Competition dataset PubTabNet, click here for downloading.
  • About PubTabNet, check their github and paper.
  • About the metric TEDS, see github

Installation

  1. Install mmdetection. click here for details.

    # We embed mmdetection-2.11.0 source code into this project.
# You can cd and install it (recommend).
cd ./mmdetection-2.11.0
pip install -v -e .

  2. Install mmocr. click here for details.

    # install mmocr
cd {Path to TableMASTER_mmocr}
pip install -v -e .

  3. Install mmcv-full-1.3.4. click here for details.

    pip install mmcv-full=={mmcv_version} -f https://download.openmmlab.com/mmcv/dist/{cu_version}/{torch_version}/index.html

# install mmcv-full-1.3.4 with torch version 1.8.0 cuda_version 10.2
pip install mmcv-full==1.3.4 -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.8.0/index.html

Docker environment

docker pull jiaquanye/tablemaster_mmocr:v1

Usage

Data preprocess

Run data_preprocess.py to get valid train data. Remember to change the 'raw_img_root' and ‘save_root’ property of PubtabnetParser to your path.

python ./table_recognition/data_preprocess.py

It will about 8 hours to finish parsing 500777 train files. After finishing the train set parsing, change the property of 'split' folder in PubtabnetParser to 'val' and get formatted val data.

Directory structure of parsed train data is :

.
├── StructureLabelAddEmptyBbox_train
│   ├── PMC1064074_007_00.txt
│   ├── PMC1064076_003_00.txt
│   ├── PMC1064076_004_00.txt
│   └── ...
├── recognition_train_img
│   ├── 0
│       ├── PMC1064100_007_00_0.png
│       ├── PMC1064100_007_00_10.png
│       ├── ...
│       └── PMC1064100_007_00_108.png
│   ├── 1
│   ├── ...
│   └── 15
├── recognition_train_txt
│   ├── 0.txt
│   ├── 1.txt
│   ├── ...
│   └── 15.txt
├── structure_alphabet.txt
└── textline_recognition_alphabet.txt

We also transfer the raw Pubtanet data to Lmdb files by the script lmdb_maker.py.

Click here to download the Pubtanet data Lmdb files (code:uxl1)

If you want to train your model via Lmdb files, please take a look at the TableMASTER lmdb config file and text-line MASTER lmdb config file

Train

  1. Train text line detection model with PSENet.

    sh ./table_recognition/expr/table_text_line_detection_dist_train.sh

    We don't offer PSENet train data here, you can create the text line annotations by open source label software. In our experiment, we only use 2,500 table images to train our model. It gets a perfect text line detection result on validation set.

  2. Train text-line recognition model with MASTER.

    sh ./table_recognition/expr/table_text_line_recognition_dist_train.sh

    We can get about 30,000,000 text line images from 500,777 training images and 550,000 text line images from 9115 validation images. But we only select 20,000 text line images from 550,000 dataset for evaluatiing after each trainig epoch, to pick up the best text line recognition model.

    Note that our MASTER OCR is directly trained on samples mixed with single-line texts and multiple-line texts.

  3. Train table structure recognition model, with TableMASTER.

    sh ./table_recognition/expr/table_recognition_dist_train.sh

Inference

To get final results, firstly, we need to forward the three up-mentioned models, respectively. Secondly, we merge the results by our matching algorithm, to generate the final HTML code.

  1. Models inference. We do this to speed up the inference.
python ./table_recognition/run_table_inference.py

run_table_inference.py wil call table_inference.py and use multiple gpu devices to do model inference. Before running this script, you should change the value of cfg in table_inference.py .

Directory structure of text line detection and text line recognition inference results are:

# If you use 8 gpu devices to inference, you will get 8 detection results pickle files, one end2end_result pickle files and 8 structure recognition results pickle files. 
.
├── end2end_caches
│   ├── end2end_results.pkl
│   ├── detection_results_0.pkl
│   ├── detection_results_1.pkl
│   ├── ...
│   └── detection_results_7.pkl
├── structure_master_caches
│   ├── structure_master_results_0.pkl
│   ├── structure_master_results_1.pkl
│   ├── ...
│   └── structure_master_results_7.pkl
  1. Merge results.
python ./table_recognition/match.py

After matching, congratulations, you will get final result pickle file.

Get TEDS score

  1. Installation.

    pip install -r ./table_recognition/PubTabNet-master/src/requirements.txt

  2. Get gtVal.json.

    python ./table_recognition/get_val_gt.py

  3. Calcutate TEDS score. Before run this script, modify pred file path and gt file path in mmocr_teds_acc_mp.py

    python ./table_recognition/PubTabNet-master/src/mmocr_teds_acc_mp.py

Result

Text line end2end recognition accuracy

Models Accuracy
PSENet + MASTER 0.9885

Structure recognition accuracy

Model architecture Accuracy
TableMASTER_maxlength_500 0.7808
TableMASTER_ConcatLayer_maxlength_500 0.7821
TableMASTER_ConcatLayer_maxlength_600 0.7799

TEDS score

Models TEDS
PSENet + MASTER + TableMASTER_maxlength_500 0.9658
PSENet + MASTER + TableMASTER_ConcatLayer_maxlength_500 0.9669
PSENet + MASTER + ensemble_TableMASTER 0.9676

In this paper, we reported 0.9684 TEDS score in validation set (9115 samples). The gap between 0.9676 and 0.9684 comes from that we ensemble three text line models in the competition, but here, we only use one model. Of course, hyperparameter tuning will also affect TEDS score.

Pretrained Model

The TableMASTER (TableMASTER_maxlength_500) pretrained model. In the validation set, the accuracy is 0.7767.

[Google Drive]

[BaiduYun Drive] code:irp6


The table textline detection model PSENet pretrained model.

[Google Drive]

[BaiduYun Drive] code:6b30

We also release the PSENet train data.

[Google Drive]

[BaiduYun Drive] code:rzu2


The table textline recognition model MASTER pretrained model. In the textline validation subset, the accuracy is 0.9798.

[Google Drive]

[BaiduYun Drive] code:kp25

PS: Due to some mistakes occur in making text-line recognition lmdb file, the accuracy of the MASTER pretrained model, is lower than the accuracy reported in this paper. In this version, we don't filter the text-line images of table head. In pubtabnet, content in table head is always surrounded by b /b, which means overstriking in visual. So, training with text-line images of table head will make model failed to recognition overstriking character. It will drop the word accuracy of text-line recognition model. Of course, it will drop the final TEDS score.

We will fix this bug when next version pretrained model releases. You can also train a text-line recognition model, after filtering the text-line images of table head.


The final TEDS score is 0.9618. You can get the log file in the following link.

[Google Drive]

[BaiduYun Drive] code:g3gx

Lmdb Data

We convert the raw Pubtabnet data to Lmdb files. You can download by this link:

[BaiduYun Drive] code: uxl1

License

This project is licensed under the MIT License. See LICENSE for more details.

Citations

@article{ye2021pingan,
  title={PingAn-VCGroup's Solution for ICDAR 2021 Competition on Scientific Literature Parsing Task B: Table Recognition to HTML},
  author={Ye, Jiaquan and Qi, Xianbiao and He, Yelin and Chen, Yihao and Gu, Dengyi and Gao, Peng and Xiao, Rong},
  journal={arXiv preprint arXiv:2105.01848},
  year={2021}
}
@article{He2021PingAnVCGroupsSF,
  title={PingAn-VCGroup's Solution for ICDAR 2021 Competition on Scientific Table Image Recognition to Latex},
  author={Yelin He and Xianbiao Qi and Jiaquan Ye and Peng Gao and Yihao Chen and Bingcong Li and Xin Tang and Rong Xiao},
  journal={ArXiv},
  year={2021},
  volume={abs/2105.01846}
}
@article{Lu2021MASTER,
  title={{MASTER}: Multi-Aspect Non-local Network for Scene Text Recognition},
  author={Ning Lu and Wenwen Yu and Xianbiao Qi and Yihao Chen and Ping Gong and Rong Xiao and Xiang Bai},
  journal={Pattern Recognition},
  year={2021}
}
@article{li2018shape,
  title={Shape robust text detection with progressive scale expansion network},
  author={Li, Xiang and Wang, Wenhai and Hou, Wenbo and Liu, Ruo-Ze and Lu, Tong and Yang, Jian},
  journal={arXiv preprint arXiv:1806.02559},
  year={2018}
}

Acknowledgements