table_structure_recognition

Table detection and table structure recognition using Yolov5

Dataset

You can download PubTables-1M from Microsoft Research Open Data, and uncompress PubTables-1M-Image_Page_Detection_PASCAL_VOC.tar.gz and PubTables-1M-Image_Table_Structure_PASCAL_VOC.tar.gz to the directory PubTables-1M/.

Then, you can run the 2 notebooks (voc2coco_detection.ipynb and voc2coco_structure.ipynb) to convert VOC format to COCO format.

Train Model

You can clone latest Yolov5 code from https://github.com/ultralytics/yolov5 to the directory yolov5/, and run the 2 scripts to train table detection model (train_PubTables-1M_detection.sh) and table structure recognition model (train_PubTables-1M_structure.sh). You may need to change the variable path of the yaml files according to your environment in the directory yolov5/data/.

I have trained each model using yolov5s for 10 epochs, and you can use the models in the directory yolov5/runs/ for fast try or finetune from the checkpoints.

Use Model

You can run the notebook table_structure_recognition.ipynb to convert a table image to an excel file. Please pay attention to the ocr function, you should use all-in-one-ai, or PaddleOCR, or any OCR service to get the ocr result.