LIST RECOGNIZING METHOD AND LIST RECOGNIZING SYSTEM

• US 20150095022A1
• Filed: 12/04/2013
• Published: 04/02/2015
• Est. Priority Date: 09/29/2013
• Status: Abandoned Application

• Current Assignee: Peking University Founder Group Company Limited (Beijing Peking University Asset Management Company Limited), Founder Apabi Technology Limited
• Original Assignee: Peking University Founder Group Company Limited (Beijing Peking University Asset Management Company Limited), Founder Apabi Technology Limited
• Inventors: XU, Canhui, TANG, Zhi, XU, Jianbo, TAO, Xin
• Application Number:
  US14/096,431
• Publication Number:
  US 20150095022A1
• Time in Patent Office: Days
• US Class Current: 704/10
• CPC Class Codes: G06V 30/413 Classification of content, e.g. text, photographs or tables
• CPC Class Codes: G06V 30/416 Extracting the logical structure, e.g. chapters, sections or page numbers; Identifying elements of the document, e.g. authors

Abstract

A list recognizing method and system, which comprises: parsing and analyzing metadata information within an original fixed-layout document, and extracting basic elements within a page; segmenting the basic elements, extracting segmented text lines within the page to obtain fragments; building an undirected graph with respect to the fragments; detecting indent features of a bullet according to features of the basic elements; training a learning model according to the indent features, local features of the fragments and neighborhood relation features among the fragments, obtaining model parameters, and establishing a list recognizing model; and invoking the list recognizing model to perform list recognizing on the required document, so as to get recognition result. This machine learning method may recognize not only a list, but also the contextual relationship between the first line and its subsequent lines of a list, and realize analyzing and understanding a layout of the list of the fixed-layout document ultimately. The accuracy of list recognizing on a fixed-layout document can be improved even if the bullets of the first line of the list are various.

Claims

1. A list recognizing method, comprising:

parsing and analyzing metadata information within an original fixed-layout document, and extracting basic elements within a page;

segmenting the basic elements, extracting segmented text lines within the page to obtain fragments;

building an undirected graph with respect to the fragments;

detecting indent features of a bullet according to features of the basic elements;

training a learning model according to the indent features, local features of the fragments and neighborhood relation features among the fragments, obtaining model parameters, and establishing a list recognizing model; and

invoking the list recognizing model to perform list recognizing on the required document, so as to get recognition results.

2. The method according to claim 1, wherein

the training a learning model according to the indent, local features of the fragments and the neighborhood relation features among the fragments, obtaining model parameters, and establishing a list recognizing model comprises:

extracting the local features of each of the fragments in the undirected graph, classifying the local features and converting scores of classifications into a pseudo-probability function used as a unary feature function of a Conditional Random Fields (CRF) model; and

extracting the neighborhood relation features among the fragments as a binary feature function according to neighborhood relations of the undirected graph,

wherein the learning model is a CRF model.

3. The method according to the claim 1, wherein

the segmenting the basic elements, extracting segmented text lines within the page to obtain fragments comprises:

segmenting continuous texts in the text lines into one fragment.

4. The method according to the claim 1, wherein

the extracting segmented text lines within the page comprises:

using clustering method for extracting segmented text lines.

5. The method according to the claim 1, wherein

the building an undirected graph with respect to the fragments comprises:

building the undirected graph by using the neighborhood relations among the fragments.

6. The method according to the claim 1, wherein

the building an undirected graph with respect to the fragments comprises:

using a Minimal Spanning Tree (MST) method or a triangulation method to build the undirected graph.

7. The method according to the claim 1, wherein

the detecting indent features of a bullet according to features of the basic elements comprises:

detecting indent level of the bullet, relative indents and whether the indents of the bullets are identical to those of other bullets.

8. The method according to the claim 1, wherein

the local features of the fragments comprise a length-width ratio, a normalized area, an indent level, and image texture features of the fragments.

9. The method according to the claim 2, wherein

the extracting the local features of each of the fragments in the undirected graph, classifying the local features and converting scores of classifications into a pseudo-probability function comprises:

classifying by a Support Vector Machine (SVM) classifier, selecting Radial Basis Function (RBF), and converting the scores of the classifications into the pseudo-probability function.

10. The method according to the claim 1, wherein

the indent features comprise an indent level of the bullet, relative indents and whether the indents of the bullets are identical to those of other bullets.

11. A list recognizing system, comprising:

an extracting unit, configured to parse and analyze metadata information within an original fixed-layout document, and extract basic elements within a page;

a segmenting unit, configured to segment the basic elements, extract segmented text lines within the page to obtain fragments;

a building unit, configured to build an undirected graph with respect to the fragments;

a detecting unit, configured to detect indent features of a bullet according to features of the basic elements;

a modeling unit, configured to train a learning model according to the indent features, local features of the fragments and neighborhood relation features among the fragments, obtain model parameters, and establish a list recognizing model; and

an invoking unit, configured to invoke the list recognizing model to perform list recognizing on the required document, so as to get recognition results.

12. The system according to the claim 11, wherein the modeling unit comprises:

a first feature extraction subunit, configured to extract the local features of each of the fragments in the undirected graph, classify the local features and convert scores of classifications into a pseudo-probability function used as a unary feature function of a CRF model;

a second feature extraction subunit, configured to extract the neighborhood relation features among the fragments as a binary feature function according to neighborhood relations of the undirected graph,

wherein the learning model is a CRF model.

13. The system according to the claim 11, wherein

the segmenting unit is configured to segment continuous texts in the text lines into one fragment.

14. The system according to the claim 11, wherein

the extracting unit is configured to extract the segmented text lines by using a clustering method.

15. The system according to the claim 11, wherein

the building unit is configured to build the undirected graph by using the neighborhood relations among the fragments.

16. The system according to the claim 11, wherein

the building unit is configured to use a MST method or a triangulation method to build the undirected graph.

17. The system according to the claim 11, wherein

the detecting unit is configured to detect indent level of the bullet, relative indents and whether the indents of the bullets are identical to those of other bullets.

18. The system according to the claim 11, wherein

the local features of the fragments comprise a length-width ratio, a normalized area, an indent level, and image texture features of the fragments.

19. The system according to the claim 12, wherein

the first feature extraction subunit is configured to classify by a SVM classifier, select RBF, and convert the scores of the classifications into the pseudo-probability function.

20. The system according to the claim 11, wherein

the indent features comprise an indent level of the bullet, relative indents and whether the indents of the bullets are identical to those of other bullets.