A Table Structure Recognition Method Based on Graph Neural Network

Abstract

The invention discloses a form structure recognition method based on a graph neural network. The method converts each page of a pdf document into an image, recognizes the position of the form for each input image, and intercepts the form area; for the form Region recognition text blob blocks; find adjacent blob sets for each blob, thereby establishing a blob graph structure; establish a dual graph structure for a blob graph, and transform the graph node connection prediction problem into a graph node classification problem; train graph node classification model; organize the relationship between blobs to obtain the cell structure of the table; the present invention applies the graph neural network to the table structure recognition, models the table structure recognition as graph node classification, and adds feedback adjustment network and condition random For airports, the classification results of graph nodes are corrected based on the rationality of the overall structure of the table, which improves the recognition accuracy.

Description

technical field [0001] The invention relates to image recognition technology, in particular to a table structure recognition method based on a graph neural network. Background technique [0002] In the application scenarios of big data and artificial intelligence, it is necessary to collect, process, and analyze a large amount of information, structure the data, and discover the laws in the data to guide production. Information exists in various and unstructured ways. A large amount of information exists in tables, and tables may exist in pdfs, web pages, and images. [0003] For tables in pdf, currently existing table parsing methods generally include parsing tables by reading xml information of pdf (such as xpdf tool), converting pdf to xml, html, word and other formats and then parsing (such as pdf-docx tool ), converting the pdf into an image and then performing structural recognition. The first two methods cannot be accurately analyzed due to the information loss of th...

AI Technical Summary

Problems solved by technology

[0003] For tables in pdf, currently existing table parsing methods generally include parsing tables by reading xml information of pdf (such as xpdf tool), converting pdf to xml, html, word and other formats and then parsing (such as pdf-docx tool ), converting the pdf into an image and then performing structural recognition. The first two methods cannot be accurately analyzed due to the information loss of the pdf file itself. The third method mainly relies on image recognition algorithms. The existing methods are not suitable for complex forms. can accurately identify

Images