ROTATION-FREE RECOGNITION OF HANDWRITTEN CHARACTERS

• US 20130251249A1
• Filed: 03/23/2012
• Published: 09/26/2013
• Est. Priority Date: 03/23/2012
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Huo, Qiang, Du, Jun
• Granted Patent: US 8,977,042 B2
• Time in Patent Office: Days
• US Class Current: 382/161
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06F 18/24133 Distances to prototypes
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/1478 of characters or characters lines
• CPC Class Codes: G06V 30/166 Normalisation of pattern dimensions
• CPC Class Codes: G06V 30/19173 Classification techniques
• CPC Class Codes: G06V 30/2445 Alphabet recognition, e.g. Latin, Kanji or Katakana
• CPC Class Codes: G06V 30/287 of Kanji, Hiragana or Katakana characters
• CPC Class Codes: G06V 30/333 Preprocessing; Feature extraction

Abstract

A character recognition system receives an unknown character and recognizes the character based on a pre-trained recognition model. Prior to recognizing the character, the character recognition system may pre-process the character to rotate the character to a normalized orientation. By rotating the character to a normalized orientation in both training and recognition stages, the character recognition system releases the pre-trained recognition model from considering character prototypes in different orientations and thereby speeds up recognition of the unknown character. In one example, the character recognition system rotates the character to the normalized orientation by aligning a line between a sum of coordinates of starting points and a sum of coordinates of ending points of each stroke of the character with a normalized direction.

Claims

1. A system comprising:

one or more processors;

memory, communicatively coupled to the one or more processors, storing executable instructions that, when executed by the one or more processors, configure the one or more processors to perform acts comprising;

receiving a plurality of handwritten textual characters;

rotating each handwritten textual character to a predetermined orientation based on information of a starting point and an ending point of each stroke of each handwritten textual character;

extracting features from the plurality of rotated, handwritten textual characters;

training a character recognition model based on optimizing a sample-separation-margin (SSM) based minimum classification error (MCE) objective function using a resilient propagation (Rprop) algorithm;

compressing parameters of the character recognition model using a split vector quantization (SVQ) technique; and

constructing a two-level fast-match tree for recognition of an unknown textual character at run time.

2. One or more computer-readable media storing executable instructions that, when executed by one or more processors, configure the one or more processors to perform acts comprising:

receiving a plurality of textual characters;

rotating each textual character to a normalized orientation based on information of a starting point and an ending point of each of a plurality of strokes in the textual character;

extracting features of each rotated textual character; and

training a recognition model based on the extracted features of each rotated textual character.

3. The one or more computer-readable media as recited in claim 2, the acts further comprising compressing parameters of the trained recognition model using a split vector quantization technique.

4. The one or more computer-readable media as recited in claim 2, the acts further constructing a two-level fast-match tree for use at run time to aid in recognition of an unknown textual character using the trained recognition model.

5. The one or more computer-readable media as recited in claim 2, wherein the rotating each textual character comprises:

computing a first coordinate as a first weighted sum of coordinates of the starting points of the plurality of strokes of the respective textual character;

computing a second coordinate as a second weighted sum of coordinates of the ending points of the plurality of strokes of the respective textual character; and

aligning a line between the first coordinate and the second coordinate in the normalized direction.

6. The one or more computer-readable media as recited in claim 5, wherein the first weighted sum or the second weighted sum comprises:

an average sum of the coordinates of the starting points or the ending points, or

a sum of the coordinates of the starting points or the ending points with coefficients weighted with lengths of strokes associated with the coordinates of the starting points or the ending points.

7. The one or more computer-readable media as recited in claim 2, wherein the training comprises training the recognition model based on a sample-separation-margin (SSM) based minimum classification error (MCE) objective function using a resilient propagation (Rprop) algorithm.

8. The one or more computer-readable media as recited in claim 2, wherein the plurality of strokes comprises all strokes of the respective textual character, a predetermined number of longest strokes of the respective textual character, a predetermined percentage of longest strokes of the respective textual character, one or more strokes of the respective textual character having lengths greater than or equal to a predetermined length threshold, and/or one or more strokes of the respective textual character having lengths greater than or equal to an average length of the plurality of strokes.

9. A method comprising:

under control of one or more processors configured with executable instructions;

receiving a textual character comprising multiple strokes; and

rotating the textual character to a normalized direction based on at least two points of a stroke of the multiple strokes.

10. The method as recited in claim 9, further comprising recognizing the rotated textual character based on a recognition model.

11. The method as recited in claim 9, wherein the at least two points comprise a starting point and an ending point of the stroke of the multiple strokes.

12. The method as recited in claim 9, wherein the rotating comprises:

computing a first coordinate as a first weighted sum of coordinates of starting points of a subset of the multiple strokes; and

computing a second coordinate as a second weighted sum of coordinates of ending points of the subset of the multiple strokes, wherein the rotating comprises rotating the textual character to cause a line between the first coordinate and the second coordinate to align with the normalized direction.

13. The method as recited in claim 12, wherein the first weighted sum or the second weighted sum comprises:

an average sum of the coordinates of the starting points or the ending points, or

a sum of the coordinates of the starting points or the ending points with coefficients weighted with lengths of the subset of the multiple strokes associated with the coordinates of the starting points or the ending points.

14. The method as recited in claim 12, wherein the subset of the multiple strokes comprises all strokes of the multiple strokes, a predetermined number of longest strokes of the multiple strokes, a predetermined percentage of longest strokes of the multiple strokes, a stroke of the multiple strokes having lengths greater than or equal to a predetermined length threshold, and/or a stroke of the multiple strokes having lengths greater than or equal to an average length of the multiple strokes.

15. The method as recited in claim 9, wherein the rotating is independent of an order of the multiple strokes of the textual character written by a user.

16. The method as recited in claim 9, wherein the receiving comprises receiving coordinates of starting points and ending points of the multiple strokes of the textual character.

17. The method as recited in claim 9, further comprising determining starting points and ending points of the multiple strokes without knowing or tracking directions of writing of the multiple strokes.

18. The method as recited in claim 9, further comprising determining starting points and ending points of the multiple strokes based on heuristic assumptions of conventions of writing a stroke.

19. The method as recited in claim 17, where the conventions of writing the stroke comprise writing from left to right and writing from top to bottom.

20. The method as recited in claim 9, further comprising:

computing a direction of each of the multiple strokes from first ones of the at least two points of the multiple strokes to second ones of the at least two points of the multiple strokes; and

computing a vector sum direction as a vector sum of the directions of the multiple strokes, wherein the rotating comprises rotating the textual character to cause the vector sum direction to align with the normalized direction.