Since tables in documents provide important information in compact form, table understanding has been an essential topic in document image processing. Researchers represented table structures in various formats for table understanding, such as simple grid structure, a graph with text/cell boxes as nodes, or a sequence of HTML tokens. However, these approaches have difficulties in handling regularities, e.g., global row and column information, and spanning cells simultaneously. In this paper, we propose a new table recognition method based on a grid shape graph and present grid localization and grid elements grouping networks. This approach is designed to exploit the grid structure and deal with spanning cells. To convert grid structure into cell structure, we only have to test adjacent pairs of grid elements, enabling efficient inference. In addition, we have discovered that predicting row/column-based relationships between grid elements improve cell-based connectivity estimation performance. We demonstrate the effectiveness of the proposed method through experiments on three benchmark datasets.
ACKNOWLEDGMENT This research was supported by LG AI Research. This work was also supported by the BK21 FOUR program of the Education and Research Program for Future ICT Pioneers, Seoul National University in 2022.