Method and system for data extraction from images of semi-structured documents

Abstract

The present invention is directed to a method of extracting data from fields in an image of a document. In one implementation, a text representation of the image of the document is obtained. A graph for storing features of the text fragments in the text representation of the image of the document and their links is constructed. A cascade classification for computing the features of the text fragments in the text representation of the image of the document and their link is run. Hypotheses about the belonging of text fragments to the fields in the image of the document are generated. Combinations of the hypotheses are generated. A combination of the hypotheses is selected. And data from the fields in the image of the document is extracted based on the selected combination of the hypotheses.

1 - 19 . (canceled) 
     
     
         20 . A method comprising:
 obtaining a text representation of an image of a document, wherein the text representation comprises one or more text fragments;   constructing a graph comprising one or more features values of the one or more text fragments and one or more links between the one or more text fragments;   generating one or more hypotheses about the one or more text fragments belonging to one or more fields in the image of the document using the one or more text fragments and the one or more links in the graph;   generating one or more combinations of the one or more hypotheses;   selecting a combination of the one or more combinations; and   extracting data from one or more fields in the image of the document based on the selected combination.   
     
     
         21 . The method of  claim 20 , wherein the text representation of the image of the document is a result of an optical character recognition (OCR). 
     
     
         22 . The method of  claim 20 , wherein the graph further comprises nodes and edges, and wherein constructing the graph further comprises:
 matching at least one of the nodes of the graph to at least one of a word, a word combination, or a word fragment of the text representation of the image of the document; and   connecting the nodes in a linear order by the edges.   
     
     
         23 . The method of  claim 20 , further comprising computing one or more features of the one or more text fragments and the one or more links using a cascade classification. 
     
     
         24 . The method of  claim 23 , wherein the generating the one or more hypotheses about the one or more text fragments belonging to the one or more fields in the image of the document is at least in part based on the computed one or more features of the one or more text fragments and the one or more links. 
     
     
         25 . The method of  claim 20 , wherein the selecting of the combination of the one or more combinations is based on a computed quality of the one or more combinations of the one or more hypotheses. 
     
     
         26 . The method of  claim 20 , wherein the selecting of the combination of the one or more combinations is based on comparing a first feature vector of a first combination of the one or more hypotheses with a second feature vector of a second combination of the one or more hypotheses. 
     
     
         27 . A system comprising:
 a memory; and   a processor communicably coupled to the memory, the processor to:
 obtain a text representation of an image of a document, wherein the text representation comprises one or more text fragments; 
 construct a graph comprising one or more features values of the one or more text fragments and one or more links between the one or more text fragments; 
 generate one or more hypotheses about the one or more text fragments belonging to one or more fields in the image of the document using the one or more text fragments and the one or more links in the graph; 
 generate one or more combinations of the one or more hypotheses; 
 select a combination of the one or more combinations; and 
 extract data from one or more fields in the image of the document based on the selected combination. 
   
     
     
         28 . The system of  claim 27 , wherein the text representation of the image of the document is a result of an optical character recognition (OCR). 
     
     
         29 . The system of  claim 27 , wherein the graph further comprises nodes and edges, and wherein, to construct the graph, the processor is further to:
 match at least one of the nodes of the graph to at least one of a word, a word combination, or a word fragment of the text representation of the image of the document; and   connect the nodes in a linear order by the edges.   
     
     
         30 . The system of  claim 27 , wherein the processor is further to compute one or more features of the one or more text fragments and the one or more links using a cascade classification. 
     
     
         31 . The system of  claim 30 , wherein the processor is further to generate the one or more hypotheses about the one or more text fragments belonging to the one or more fields in the image of the document based at least in part on the computed one or more features of the one or more text fragments and the one or more links. 
     
     
         32 . The system of  claim 27 , wherein the processor is further to select the combination of the one or more combinations based at least in part on a computed quality of the one or more combinations of the one or more hypotheses. 
     
     
         33 . The system of  claim 27 , wherein the processor is further to select the combination of the one or more combinations based at least in part on a comparison of a first feature vector of a first combination of the one or more hypotheses and a second feature vector of a second combination of the one or more hypotheses. 
     
     
         34 . A non-transitory computer-readable medium having instructions encoded thereon which, when executed by a processor, cause the processor to:
 obtain a text representation of an image of a document, wherein the text representation comprises one or more text fragments;   construct a graph comprising one or more features values of the one or more text fragments and one or more links between the one or more text fragments;   generate one or more hypotheses about the one or more text fragments belonging to one or more fields in the image of the document using the one or more text fragments and the one or more links in the graph;   generate one or more combinations of the one or more hypotheses;   select a combination of the one or more combinations; and   extract data from one or more fields in the image of the document based on the selected combination.   
     
     
         35 . The non-transitory computer-readable medium of  claim 34 , wherein the text representation of the image of the document is a result of an optical character recognition (OCR). 
     
     
         36 . The non-transitory computer-readable medium of  claim 34 , wherein the graph further comprises nodes and edges, and wherein, to construct the graph, the processor is further to:
 match at least one of the nodes of the graph to at least one of a word, a word combination, or a word fragment of the text representation of the image of the document; and   connect the nodes in a linear order by the edges.   
     
     
         37 . The non-transitory computer-readable medium of  claim 34 , wherein the processor is further to compute one or more features of the one or more text fragments and the one or more links using a cascade classification. 
     
     
         38 . The non-transitory computer-readable medium of  claim 37 , wherein the processor is further to generate the one or more hypotheses about the one or more text fragments belonging to the one or more fields in the image of the document based at least in part on the computed one or more features of the one or more text fragments and the one or more links. 
     
     
         39 . The non-transitory computer-readable medium of  claim 34 , wherein the processor is further to select the combination of the one or more combinations based at least in part on a computed quality of the one or more combinations of the one or more hypotheses.