TableCell

Image-based table cell detection: a new dataset and an improved detection method

Introduction.

In this repo, we present TableCell, a new image-based dataset which contains 5262 samples with 170K high precision cell-wised annotations based on a novel semi-supervised method.

Furthermore, we come up with an efficient table projection method to facilitate capturing long-range global feature, which consists of row projection and column projection.

TableCell dataset.

Note: a detailed construction tutorial and code will be released to ease the burden of annotating a more adequate personal dataset after refactoring.

There are 5262 images with 170K annotations in total. TableCell dataset will be made available at https://github.com/weidafeng/TableCell upon our paper publication.

It contains four types tabular documents.

1. Electronic sample with complete lines
2. Scanner sample with partial lines
3. Electronic sample with partial lines
4. Scanned sample with complete lines

Some Annotation samples.

Table Projection Module.

Despite the performance of universal object detectors has been dramatically improved thanks to the advance of deep convolutional neural networks, there is a relative scarcity of deep learning approaches for table or even table cell detection. Even with strong learning ability, common object detection models don't suffice as the tabular data also require global information to identify table cell boundaries because both cells and cell separators contain large regions of white space.

Our proposed method named Table Projection Module designs architectures that can capture global information involving both row and column. Here is the architecture and simple visualization demo to understand.

Getting started.

We use the open source framework [MMDetection](https://github.com/open-mmlab/mmdetection) to train models on the TableCell. Please follow the official INSTALL.mdfor installation and dataset preparation, and see the official GETTING_STARTED.md for the basic usage of MMDetection.

To use our Table Projection Module for table cell detection, some channges should be made as follows.

1. backup the raw model

~$ cd YOUR_PATH_TO_MMDETECTION
mmdetection$ mv mmdet/models/backbones/resnet.py mmdet/models/backbones/resnet_raw.py  
mmdetection$ mv mmdet/ops/__init__.py mmdet/ops/__init__raw.py  

2. replace our modified model to the raw model

mmdetection$ cp src/resnet.py mmdet/models/backbones/resnet.py  
mmdetection$ cp src/__init__.py mmdet/ops/__init__.py  
mmdetection$ cp src/table_projection.py mmdet/ops/table_projection.py  

3. run You should specify your TableCell dataset path in src/faster_rcnn_r50_fpn_1x_table.py

mmdetection$ ./tools/dist_train.sh src/faster_rcnn_r50_fpn_1x_table.py 1 --validate  # train

4. test and save results

mmdetection$   python src/test_and_save_results.py -c config/xxx.py -m work_dirs/xxx.pth -i /path/to/test_images -o /path/to/save_results

5. pretrained model and results

We will share pretrained model with our Table Projection Module upon publication, please refer to the paper for more details.

Reference.

1. TableBank. https://github.com/doc-analysis/TableBank
2. MMDection. https://github.com/open-mmlab/mmdetection