The research group led by Professor Wei Lu of the Key Laboratory of Intelligent Development and Service of Archives has published its latest research on figure and table detection in academic texts, entitled "An effective method for figures and tables detection in academic literature," in Information Processing and Management (IPM), an SSCI Q1 journal. IPM is a top journal in the field of information management. Postdoctoral researcher Dr. Fengchang Yu is the first author and corresponding author, supervised by Professor Wei Lu.

Figure 1: Screenshot of the Journal Paper

Figures and tables in academic literature are important data sources for various academic text data mining tasks, which require them to be complete when used as input. However, existing studies measure algorithm performance using the same IoU (Intersection over Union) or IoU-based metrics designed for natural scenes. In figure and table detection for academic texts, there is a gap between a high IoU and detecting the object as a whole. This study demonstrates the existence of this gap, attributes it mainly to detection errors in boundary regions, and proposes an effective detection method that cascades semantic segmentation and contour detection. The method achieves a page-level F1 of 0.983, surpassing state-of-the-art figure and table detection methods for academic texts. The results can significantly improve the quality of academic text data and reduce data cleaning costs for downstream applications.

Figure 2: Workflow of the Figure and Table Detection Method for Academic Literature

Original paper link: https://doi.org/10.1016/j.ipm.2023.103286