Using CoordConv for Tabular Data Detection and Structure Recognition
Apoorva Ambulgekar, Naman Lad, Krunal Doshi, Pranit Bari · 2022 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS) · 2022
This paper explores the usage of CoordConv, the novel upgrade to general convolutional layers in the problem of Tabular Data Detection and Cell-Based Structure Recognition. CoordConv has been shown to provide considerably better results in the domain of Object Detection than its counterpart. The authors integrate it within the established Anchor optimization approach which leverages guided anchors to accomplish the task of recognizing rows and columns present in tabular data. In contrast to the majority of techniques implemented for Table Structure Recognition, the authors attempt to recognize the cells present in the tabular images instead of the rows and columns. They evaluate this method on the coveted ICDAR-19 dataset (International Conference on Document Analysis and Recognition - 2019) which comprises of scanned document images containing tabular regions and achieve results surpassing those of many popular techniques. They also apply this approach for the task of Table Detection and achieve results comparable to other established techniques.