The unambiguous structure representation of tabular data for recognition

Recognizing the structure of tabular (image) data aims to convert the table on the image to a machine-readable representation. The structure of tabular data is complex and diverse, including cases such as cell spanning, nested cell spanning, cell vacancies, and slanted headers. Existing representations are plagued by the issue of representation ambiguity, complicating the structure recognition process and degrading performance. In this paper, we propose an unambiguous representation named Table Structure Graph (TSG), which can unambiguously represent tabular data with complex and diverse structures. This unambiguity enhances the accuracy and efficiency of structure recognition. Additionally, we introduce a dataset, Complex Table (CmpTab), which comprises tabular data with a wide range of complex structures aimed at advancing the task of complex structure recognition. In the proof, we first give the mathematical definition of the structure via the grid-geometry and then prove that the definition domain one-to-one maps to the graph domain, which means there is no ambiguity. In the experiments, we adapt TSG to existing structure recognition models to ensure a fair comparison among current representations, verifying the fact that TSG can enhance the recognition performance; and also demonstrate its 100 % transferring rate on the dataset, verifying its unambiguity. Moreover, after the model recognizes TSG, converting the TSG representation back to tabular data is a non-trivial step. To facilitate engineering implementation, we propose the tsg2ms conversion algorithm and provide corresponding API/GUI tools on GitHub: https://github.com/MaxKinny/tsg2ms.