Automatic building of a large and complete dataset for image-based table structure recognition

Abstract

Table is one of the most common ways to represent structured data in documents. Existing researches on image-based table structure recognition often rely on limited datasets with the largest amount of 3,789 human-labeled tables as ICDAR 19 Track B dataset. A recent TableBank dataset for table structures contains 145K tables, however, the tables are labeled in an HTML tag sequence format, which impedes the development of image-based recognition methods. In this paper, we propose several processing methods that automatically convert an HTML tag sequence annotation into bounding box annotation for table cells in one table image. By ensembling these methods, we could convert 42,028 tables with high correctness, which is 11 times larger than the largest existing dataset (ICDAR 19). We then demonstrate that using these bounding box annotations, a straightforward representation of objects in images, we can achieve much higher F1-scores of table structure recognition at many high IoU thresholds using only off-the-shelf deep learning models: F1-score of 0.66 compared to the state-of-the-art of 0.44 for ICDAR19 dataset. A further experiment on using explicit bounding box annotation for image-based table structure recognition results in higher accuracy (70.6%) than implicit text sequence annotation (only 33.8%). The experiments show the effectiveness of our largest-to-date dataset to open up opportunities to generalize on real-world applications. Our dataset and experimental models are publicly available at shorturl.at/hwHY3