Method and system for commercial document image classification

• US 8,831,361 B2
• Filed: 03/09/2012
• Issued: 09/09/2014
• Est. Priority Date: 03/09/2012
• Status: Active Grant

• Current Assignee: Ancora Software Incorporated
• Original Assignee: Ancora Software Incorporated
• Inventors: Pintsov, David A.
• Primary Examiner(s): MARIAM, DANIEL G
• Application Number:
  US13/417,037
• Publication Number:
  US 20130236111A1
• Time in Patent Office: 914 Days
• Field of Search: None
• US Class Current: 382/224
• CPC Class Codes: G06V 30/418 Document matching, e.g. of document images

Abstract

A completely automated method for classifying electronic images of commercial documents such as invoices, bills of lading, explanations of benefits, etc. is based on the layout of the documents. A document is classified as belonging to a class of similarly laid out images if its distance, determined by a set of pre-defined metrics, to any of the template layouts from the same class of template layouts does not exceed a certain user-defined threshold. Alternatively, if the sum of distances from a given image to several template layouts from the same class does not exceed a user-selected threshold, the image is deemed to belong to a class of images with a specific layout.

Claims

1. A method of automatic commercial document image classification using a computer for performing the steps of:

automatically obtaining the document layout using the salient features of the document images layout, said features consisting of text blocks, geometric line segments and the contents (sets of words) of said text blocks;

defining distances between such layouts;

automatically creating a plurality of classification template layouts directly from said distances;

Ordering said plurality of classification template layouts and generating a ranked list of candidate template layouts best matching a given input document layout in such a way that the classification template layout most similar to the document layout to be classified is at the top of the list (highest ranked layout), the next likely candidate template layout is in the second position and so forth;

classifying the input image as belonging to the class of the top template layout in said ranked list of template layouts.

2. The method of claim 1 comprising

generation of a set of text blocks for each said classification template layout and for any input image layout to be classified, each such text block consisting of two components, said components being the first the geometric figure (area) encompassing the text block, and second the collection of words contained within this figure;

using said text block components for computing a combined geometric/word distance between two text blocks, this distance having two parts, the geometric part as distance between two geometric figures containing the text blocks, and the word part reflecting the difference in the number of words contained within said geometric figures.

3. The method of claim 2 further

matching the two sets of text blocks generated for two document (page) images and calculating a distance between two sets of text blocks belonging to two document (page) images.

4. The method of claim 2 with further steps of

generating for each text block a set of text strings representing the content of said text block;

using a word distance to calculate a content distance between two text blocks.

5. The method of claim 4 additionally

computing a final distance between each classification template layout and an input image layout, said distance comprising both non-content and content distances;

building a list of template layouts ranked according to their final distances from the input image layout.

6. The method of claim 1 further

using a set of geometric horizontal line segments for each classification template layout and for any input image layout to be classified, each said horizontal line segment being determined by its vertical position (ordinate) and by its length;

calculating a distance between two horizontal line segments belonging to two layouts.

7. The method according to claim 1 further

using a set of geometric vertical line segments for each classification template layout and for any input image layout to be classified, each said vertical line segment being determined by its horizontal position (abscissa) and by its length;

calculating a distance between two vertical line segments belonging to two layouts.

8. The method of claim 1 further

calculating a combined (horizontal-vertical) line segment distance between each classification template layout and an input image layout.

9. The method of claim 8 additionally

matching two sets of horizontal line segments and two sets of vertical line segments belonging to each classification template layout and an input image layout and

calculating a line segment distance between said sets of horizontal and vertical line segments.

10. The method according to claim 1 further

calculating a combined text block-line segment distance between each classification template layout and an input image layout, said distance being a non-content distance.

11. The method according to claim 1 whereby the decision to accept the highest ranked template layout in the ranked list of template layouts as the classification outcome is made under at least two conditions, first, when said highest ranked template layout is found with sufficient confidence (sufficient confidence criterion), and second, when the sum of confidences of several highest ranked template layouts from the list exceeds a pre-defined threshold, and said highest ranked layouts correspond to the same document layout.

12. The method of claim 1 further applied to any sub-images of document images, said sub-images typically representing the upper half or one third of the entire document image and one eight of the lower half of the entire document image.