Table structure recognition is a key component in document understanding. Many prior methods have addressed this problem with three sequential steps: table detection, table component extraction, and structure analysis based on pairwise relations. However, they have limitations in addressing complexly structured tables and/or practical scenarios (e.g., scanned documents). In this paper, we propose a novel graph-based table structure recognition framework. In order to handle complex tables, we formulate tables as planar graphs, whose faces are cell-regions. Then, we compute vertex (junction) confidence maps and line fields with the heatmap regression networks having a small number of parameters (about 1M) and reconstruct tables by solving a constrained optimization problem. We demonstrate the robustness of the proposed system through experiments on ICDAR 2019 dataset and on challenging table images. Experimental results show that the proposed method outperforms the conventional method for a range of scenarios and delivers good generalization performance.
This work was supported in part by the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2021-0-01062, Development of personal information processing technology for collection/utilization of high-quality and trusted training data for autonomous driving), and in part by LG AI Research.This work was supported in part by the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2021-0-01062, Development of personal information processing technology for collection/utilization of high-quality and trusted training data for autonomous driving), and in part by LG AI Research.
