计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (12): 62-76.DOI: 10.3778/j.issn.1002-8331.2209-0396

• 热点与综述 • 上一篇    下一篇

智能化表格识别技术综述

梁天恺,苏新铎,黄宇恒,徐天适,张华俊,曾碧   

  1. 1.广州广电运通金融电子股份有限公司 研究总院,广州 510000
    2.广东工业大学 计算机学院,广州 510000
  • 出版日期:2023-06-15 发布日期:2023-06-15

Survey on Intelligent Table Recognition

LIANG Tiankai, SU Xinduo, HUANG Yuheng, XU Tianshi, ZHANG Huajun, ZENG Bi   

  1. 1.Research Institute, GRG Banking Equipment Limited Company, Guangzhou 510000, China
    2.School of Computer, Guangdong University of Technology, Guangzhou 510000, China
  • Online:2023-06-15 Published:2023-06-15

摘要: 在大数据和互联网的历史背景下,信息技术的发展伴随着大量文档的产生。作为数据关系直观体现的表格常见于文档中,表格的归档也是文档处理的重要任务之一。如何在海量的文档中快速地对表格进行自动化识别成为妨碍文档处理迈向智能化的关键因素。作为人工智能研究领域重要分支之一的表格识别,能实现表格对象和结构的自动化检测与识别,被广泛应用在文档智能化处理等场景。因此总结与综述表格识别领域的概念、技术、应用与挑战显得尤为重要。阐述表格识别的概念,指出表格识别任务可被分为表格检测和表格结构识别两大子任务。针对表格检测研究方向主流的anchor-based和anchor-free算法进行介绍和分析,总结不同算法的优缺点。分别阐述基于语义分割、基于双向割并、融合神经网络以及端到端等四大类别的主流的表格结构识别算法的原理和优缺点。同时分析并讨论目前常见的有机融合表格检测和表格结构识别的非端到端与端到端的表格识别算法。最后总结并指出表格识别的应用、挑战与展望。

关键词: 表格识别, 表格检测, 表格结构识别, 人工智能, 大数据

Abstract: Under the background of big data and internet, a large number of documents are accompanied by the development of information technology. As the intuitive representation of data relationships, tables are commonly found in documents and table-filing is one of the important tasks during document processing. How to quickly and automatically  recognize a large number of tables from documents has become the key factor hindering the move to intelligent document processing. As the important branch of artificial intelligence, table recognition which can realize automatic detection and recognition of table objects and table structures is widely used in the scenes of intelligent document processing. Therefore, it is particularly important to summarize and survey the concept, technology, application and challenge of table recognition. Firstly, the concept of table recognition is expounded and pointed that the task of table recognition can be divided into two sub-tasks include table detection and table structure recognition. Secondly, the advantages and disadvantages of anchor-based and anchor-free mainstream table detection algorithms are introduced and analyzed. Thirdly, the principles, advantages and disadvantages of four major categories of mainstream table structure recognition algorithms include semantic segmentation based algorithm, bidirectional splitting and merging based algorithm, fusion neural network based algorithm, and end-to-end algorithm, are described respectively. Then, the currently common non-end-to-end and end-to-end table recognition algorithms which organically combines table detection and table structure recognition are analyzed and discussed. Finally, the application, challenge and prospect of table recognition are prospected.

Key words: table recognition, table detection, table structure recognition, artificial intelligence, big data