Flagro / ExcelTableCNN

This repository is for ExcelTableCNN project - open source automatic table detection on Excel sheets with computer vision

ExcelTableCNN

This repository is for ExcelTableCNN project - open source automatic table detection on Excel sheets with computer vision.

The project's goal is to implement OpenSource library for easy and convenient detection of multiple Table Header/Data coordinates by the use of algorithms combined with Deep Learning/Computer Vision approaches.

The model design

The first version of the ExcelTableCNN would try to recreate the TableSense model. So the model structure would look like this:

1. DataLoader with minibatch=1
2. FCN backbone (this is where the number of channels would be normalized to be 3)
3. RPN - generationg of RoI
4. For each RoI RoIAlign would be implemented and for each RoIAlign segment would be done
   • Table classification - change of table presence in the RoI
   • BBR + TableSense smart RoIAlign for each bounding box prediction
   • Table segmentation - binary mask for table present in each cell of RoI
5. For each RoI bounding boxes would be ranked and filtered with NMS

Performance Metrics

TODO

Training Data

For this project the data that's going to be used is based on VEnron2 dataset with markup from TableSense project.