---
tags: Table Detection
---
# TableNet: Deep Learning model for end-to-end Table detection and Tabular data extraction from Scanned Document Images
表格檢測 & 結構識別 e2e model.

## Contributions
- e2e deep multial-task architecture for both table detection and structure recognition.
- adding additional spatial semantic features.
- fine tuning it on an another new dataset will boost the performance of the model on the new dataset.
- manually annotated the Marmot dataset!!!!

## Model Structure
1. pre-train VGG-19 encoder and two decoder (table detection and structure recognition)
3. The images are preprocessed and modified using the Tesseract OCR
- spatial semantic features by highlighting the words.
4. output is mask.

## row segmentation
- Radon transform.
- The rows of the table which have maximum non-blank entries is marked as the starting point for a new row.
- All the columns are completely filled and there are no line demarcations, each line (level) can be seen as a unique row.

## dataset
[PAPER](https://arxiv.org/pdf/2001.01469.pdf)
[Marmot Dataset](https://www.icst.pku.edu.cn/cpdp/sjzy/index.htm)
[GITHUB](https://github.com/jainammm/TableNet)
[知乎](https://zhuanlan.zhihu.com/p/385673899)