Enhancing TableNet´s Validation: Evaluating the Accuracy of the Existing Architecture with Novel Training Data

Abstract

This thesis introduces a innovative approach to automatically create annotated datasets in which different table layouts are systematically generated along with corresponding ground truth coordinates to enhance table detection in images. This thesis aims to implement the ex- isting deep learning architecture (TableNet), renowned for its table detection capabilities, and assess its performance on the established marmot dataset and our newly generated dataset through Bitwise XOR and Intersection over Union (IOU) metrics. This novel dataset facilitates enhanced evaluation of current state-of-the-art architectures and empowers the development of more effective models through comprehensive training. This evaluation seeks to assess the effectiveness of the proposed dataset generation method in improving the accuracy of table detection using TableNet.