A New Method “ProjectionP” for Table Structure Recognition

This paper introduces computer vision techniques to address the challenge of analyzing tabular data in digital documents. It details the research and experiments conducted by the authors aimed at automating the invoice registration process within the accounting system of a particular private company. Based on the results obtained, the authors assert that in certain contexts, pattern recognition processing techniques can continue to be effectively utilized for this problem, providing commercial benefits while circumventing some of the limitations inherent in newer, high-performance deep learning approaches. The article shows an overview of the latest advancements in the field and presents the authors’ proposed table structure recognition (TSR) algorithm, which is based on the typical profile-projection approach. It also features comparison of the developed TSR algorithm with selected methods described in the literature, undertaking a verification of its efficacy using a dataset of reals documents derived from the accounting department of a private company.