Multi-level neural network based characters identification method and system

• US 20060123051A1
• Filed: 07/06/2005
• Published: 06/08/2006
• Est. Priority Date: 07/06/2004
• Status: Abandoned Application

• Current Assignee: Hi-Tech Solutions Incorporated
• Original Assignee: Hi-Tech Solutions Incorporated
• Inventors: Hofman, Yoram, Margolin, Alexandra
• Application Number:
  US11/175,982
• Publication Number:
  US 20060123051A1
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06V 20/625 License plates
• CPC Class Codes: G06V 20/63 Scene text, e.g. street names
• CPC Class Codes: G06V 30/1478 of characters or characters lines
• CPC Class Codes: G06V 30/244 using graphical properties, e.g. alphabet type or font
• CPC Class Codes: G06V 30/2504 Coarse or fine approaches, e.g. resolution of ambiguities or multiscale approaches

Abstract

A system and method, which enable precise and automatic identification of characters, perform and calibrate data verification to ensure data reliability. The system can process these identified characters, such as override adverse conditions, adjusting and correcting unclear characters and their images.

Claims

1. A multi-level method providing character recognition, comprising the steps of:

a) inputting original image to a buffer according to format file;

b) building matrix which includes information about said original image;

c) rotating said original image by an angle said angle is determined according to several parallel long straight lines found in said original image;

d) dividing said original image into rectangles;

e) activating coarse search and exact search on said rectangles;

f) locating candidate areas in said original image according to said coarse search and exact search;

g) separating said characters in said candidates area from background;

h) adjusting borders of said candidate areas;

i) stretching said candidate areas;

j) determining said characters color;

k) activating a projection process on said characters;

l) enhancing said candidates areas by activating high-pass filtering process;

m) sorting said candidates areas according to width and height ratio;

n) separating said candidate area into said characters;

o) detecting borders of said characters, junk space and anchor;

p) matching separation result to a structure format;

q) grading said matching according to a predefined formula;

r) analyzing said characters according to match area module;

s) activating a thinning process on said characters;

t) identifying said characters according to multi-level Neural Networks based on Fourier transform network letter features network and image scaling;

u) fitting said characters to said structure format;

v) evaluating the success of said character recognition process according to a mark table;

2. ) A method according to claim 1, wherein said format file includes data for initializing and processing said original image.

3. ) A method according to claim 1, wherein said matrix includes data concerning gradient and color of said original image.

4. ) A method according to claim 1, wherein said identifying further includes, Binarization, noise removal, border building and vector building.

5. ) A method according to claim 1, wherein said thinning process further includes closing small holes in said characters, diagonal filtering and skeleton calculation.

6. ) A method according to claim 1, wherein said identifying further includes identification by features and identification by scaling, using pre-trained neural networks.

7. ) A method according to claim 1, wherein said projection process include five projection types:

space, character, unknown, minus and junk.

8. ) A method for identifying characters comprising:

filters for enhancing said characters, structure formats for matching said characters to said structures format and mark table for grading said matching.

9. ) A method for executing a recognition in a ‘

real time’

fashion, using low computer resources requirements.