TableGraph: An Image Segmentation-Based Table Knowledge Interpretation Model for Civil and Construction Inspection Documentation

There are many manuals and codes to normalize each procedure in civil and construction engineering projects. Data tables in the codes offer various references and are playing a more and more valuable role in knowledge management. However, research has focused on regular table structure detection. For nonconventional tables— especially for nested tables—there is no efficient way to conduct automatic interpretation. In this paper, an automatic table knowledge interpretation model (TableGraph) is proposed to automatically extract table data from table images and then transform the table data into table cell graphs to facilitate table information querying. TableGraph considers that a table image is composed of three types of semantic pixel classes: Background, table border, and table cell contents. Because TableGraph only considers pixel semantic meaning rather than structural rules or form features, it can handle nonconventional and complex nested table situations. In addition, a cross-hit algorithm was designed to enable fast content queries on the generated table cell graphs. Validation of a real case of automatic interpretation of inspection manual table data is presented. The results show that the proposed TableGraph model can interpret the structure and contents of table images.