DeTable: Table data extraction model based on deep
Yuxuan Fan, Huobin Tan, Yu Liu, Jingxuan Zhang · 2021
The rapid development of the information age leads to the mass production and frequent transmission of data, which is difficult to deal with by human alone. With the rise and development of artificial intelligence, the use of data is becoming more efficient. Table, as a special data form, has attracted wide attention gradually. However, extracting data from table subimages presents a number of challenges, including accurately detecting table regions in the image, and then detecting and extracting information from detected table rows and columns. While some progress has been made in table detection, extracting table contents remains a challenge because it involves more fine-grained table structure identification. In this paper, DeTable: table data extraction model based on deep learning is proposed. The model uses the interdependence between the twin tasks of table detection and table structure recognition to divide the text region and the box-line region of the table. Then, rows based on semantic rules are extracted from the identified table regions. The proposed models and extraction methods were evaluated on publicly available ICDAR 2019 and Marmot table datasets and the most advanced results were obtained.