SRCT: A Structure Restoration of Complex Table Algorithm

Complex tables are widely used in fields such as financial reporting and industrial records. However, existing table structure recognition algorithms perform poorly on tables with complex structures. Existing methods for table structure extraction primarily fall into two categories: rule-based/template-based and deep learning-based methods. The former heavily relies on prior knowledge, while the latter requires extensive training and learning on a large number of complex tables, making it prone to overfitting. These methods are designed to address the challenges of recognizing everyday common tables but have not adequately addressed complex tables. To tackle this issue, we introduce a novel SRCT algorithm for table structure recognition. In the SRCT algorithm, we first conduct standardization processing and introduce algorithms such as PHT and DBSCAN to help establish logical coordinate sets, converting the originally unstructured tables into structured ones for better understanding and management. Then, through reconstruction processing, we build the structural correspondence between logical coordinate sets and physical coordinate sets, restoring the tables to their initial states and generating final physical coordinate sets based on reconstruction optimization formulas. We provide a clear and comprehensive explanation of the algorithm's principles, demonstrating its strong generalization ability in handling complex tables. To advance research on complex tables, we have collected and annotated a dataset called CTable. Preliminary experimental results show that our model performs excellently in complex table recognition tasks, achieving accuracies of 92.8%, 93.5%, and 90.8% on the ICDAR 2019, TableBank, and CTable datasets, respectively.