A Structure-Focused Deep Learning Approach for Table Recognition from Document Images

In this paper, we present a nuanced exploration of deep-learning techniques (DL) for extracting structural information from document images generated from the digitization of business processes. The driving example presented is the extraction of columns and rows of tables using a simple stacked CNN architecture and a combination of ensemble techniques. In addition, the component models of the ensemble are diversified by training on datasets created by applying a "semantics-preserving" transformation on the base dataset. This "semantics-preserving" transformation also aims to alleviate hard recognition in certain noisy images commonly encountered in practice. Our experiments demonstrate how DL techniques can be applied and innovatively combined to measurably improve the accuracy of structure extraction.