Towards Visual Explanations for Document Table Detection using Coarse Localization Maps

Abstract

Computer-vision-based methods using deep neural networks offer considerable opportunities to extract tabular information from richly-structured documents. However, it is extremely challenging to build a unified framework for tabular information extraction, for example, due to a variety of document templates, as well as, diverse document table templates. Earlier, we proposed a transfer learning based table detection approach[1] and a supervised table detection framework initialized with pre-trained self-supervised image classification model weight for table detection [2] on domain specific document images. In this paper, we investigate different document table detection techniques with respect to explainability issues. These enable, e. g., diagnostics and method refinement, towards a complete tabular data extraction pipeline and tool. In particular, we present visual explanation approaches of earlier proposed table detection models on domain specific document images in order to enhance the explainability of the applied Convolutional Neural Network (CNN) based models. We discuss first experimental results for visual explanations of those models and outline several challenges in this context. © 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).