Architecture

1. Table detection: Using SOTA detectron2
2. Table Line: Using architecture Unet + rule base
3. OCR: Using SOTA easyocr

Train

Prepare dataset:

Data is private not public, you can learn on internet about tabular data, You can label data by labelme (wkentaro/labelme: Image Polygonal Annotation with Python (polygon, rectangle, circle, line, point and image-level flag annotation). (github.com))

Refers datasets:

• https://www.icst.pku.edu.cn/cpdp/sjzy/index.htm
• https://paperswithcode.com/dataset/icdar-2013
• https://doc-analysis.github.io/tablebank-page/

Training

Config params: file base_config.yaml

bash sh scripts/train.sh 

Demo Table Line

bash sh scripts/infer_table_line.sh 

Step 1: Table detection

Step 2: Table Line

Input:

Output:

Table OCR

Step 1: Table detection

Step 2: Table line

Step 3: Crop image according line

Step 4: OCR

Step 5: Save file csv/excel

sh scripts/infer_table_ocr.sh

Input: ./datasets/demo_examples/demo2.png

Output: ./results/demo.csv

Docker

docker run --name table_extraction nam157/table_extraction:v1.0.0

References

1. nanonets-blog
2. table-detection-structure-recognition
3. table-transformer
4. TableNet: Deep Learning Model for End-to-end Table Detection and Tabular Data Extraction from Scanned Document Images