Method and apparatus for automated pattern recognition

• US 5,784,485 A
• Filed: 08/25/1994
• Issued: 07/21/1998
• Est. Priority Date: 08/25/1994
• Status: Expired due to Fees

• Current Assignee: Chein-I Chang, Youguang Xu
• Original Assignee: Chein-I Chang, Youguang Xu
• Inventors: Chang, Chein-I, Xu, Youguang
• Primary Examiner(s): Boudreau, Leo
• Assistant Examiner(s): Tran, Phuoc
• Application Number:
  US08/296,039
• Time in Patent Office: 1,426 Days
• Field of Search: 382/155, 382/156, 382/159, 382/160, 382/161, 382/209, 382/220, 382/224, 382/225, 395/22, 395/23
• US Class Current: 382/156
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/18057 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]

Abstract

The present invention relates to a new and useful Automated Pattern Recognition Device comprising a neural-network system, implemented on a general purpose computer, and capable of recognizing not only printed characters but also handwritten characters and other patterns in n-dimensions. The system incorporates novel feature extraction which generates an additional dimension from an n-dimensional input pattern, for example, a three-dimensional feature pattern from a two dimensional input pattern, resulting in shift-invariance, scale-invariance, and invariance to slight rotation.

Claims

1. A neural network system for pattern recognition, comprising:

storage means for storing a plurality of reference vectors, each of which is associated with a known pattern;

input means for inputting a target pattern having n dimensions which is to be identified;

processing means for receiving said target pattern from said input means and processing said target pattern so as to generate a characteristic vector for said target pattern;

comparator means for comparing said characteristic vector with one or more of said reference vectors until a match has been found or until all of said reference vectors have been compared to said characteristic vector;

output means for outputting the result of a match between said characteristic vector and one or more of said reference vectors, or outputting the result that no match has been found;

digitizing means for converting the target pattern to a bitmap of cells; and

masking means for generating an additional dimension to the n-dimensional target pattern so as to create an n+1 dimensional hyperdimensioned target pattern by applying successive masking windows to said bitmap of the target pattern, assigning weights to each of the cells of the bitmap, said weights dependent on the degree of match between the cell being weighted and the successive masking windows.

2. A system as in claim 1 wherein the masking means generates a hyperdimensioned characteristic vector from the characteristic vector and wherein the comparator means compares said hyperdimensioned characteristic vector with hyperdimensioned reference vectors.

3. A system as in claim 1 wherein each reference vector is an array comprising at least 2 sections, one section comprising a pattern name and one section comprising memory units representing end directions of a character'"'"'s strokes.

4. A system as in claim 3 wherein one section of the reference vector array comprises memory units storing local peak altitudes of the character represented by the reference vector.

5. A system as in claim 4 wherein one section of the reference vector array comprises memory units storing the altitudes of neighboring cells of the cell with the highest peak.

6. A system as in claim 5 wherein one section of the reference vector array comprises memory units storing features specific to particular characters.

7. A system as in claim 1 wherein n=2.

8. A system for pattern recognition, implemented on a general purpose computer having a central processing unit (CPU), feature library input means, feature library storage means, comparator means, masking means, target pattern input means and output means, comprising:

target pattern input means suitable for acquiring an n-dimensional input target pattern in a form suitable for processing by said CPU;

masking means which receive said n-dimensional input target pattern and generate an n+1-dimensional hyperdimensioned vector to said comparator means;

feature library input means for acquiring a library of feature vectors of n+1 dimensions associated with n dimensional reference characters;

feature library storage means for storing said library of feature vectors;

comparator means for receiving said hyperdimensioned input target pattern vector, successively comparing said hyperdimensioned vector with feature vectors from said feature library storage means, and outputting the result of said comparing to output means.

9. A system as in claim 8 wherein:

said target pattern input means comprises a mouse;

said masking means and comparator means comprise said computer'"'"'s CPU;

said feature library storage means comprises a disk;

said feature library input means comprises a disk drive; and

said output means comprises a video display screen.

11. A system as in claim 8 wherein said masking means generates said hyperdimensioned vector by applying successive masks to said input target pattern so as to generate an (n+1)st dimension.

12. A system as in claim 11 wherein said (n+1)st dimension corresponds to a degree of fit between the input target pattern and the masks.

13. A system as in claim 12 wherein said successive masks comprise a 2×

2 cell mask and a 3×

3 cell mask.

10. A process for pattern recognition comprising the steps of:

providing a digitized image of an n-dimensional input target pattern to be recognized;

providing a library of hyperdimensioned reference vectors having n+1 dimensions, each corresponding to a known character;

hyperdimensioning said digitized image of said n-dimensional input target pattern so as to create an n+1 dimensional vector;

comparing said n+1 dimensional vector to each of said reference vectors until a match occurs, or until all of said reference vectors have been compared without a match occurring;

identifying the input target pattern as the character corresponding to the matching reference vector if a match occurs or declaring the input target pattern as not identifiable if no match occurs.