Method and system for automatically segmenting and recognizing handwritten chinese characters

• US 6,665,436 B2
• Filed: 01/23/2003
• Issued: 12/16/2003
• Est. Priority Date: 01/13/1999
• Status: Expired due to Term

• Current Assignee: Lenovo PC International Limited (Lenovo Group Ltd.)
• Original Assignee: International Business Machines Corporation
• Inventors: Su, Hui, Wang, Qian Ying, Tang, Donald T.
• Primary Examiner(s): Tran, Phuoc
• Application Number:
  US10/350,244
• Publication Number:
  US 20030108239A1
• Time in Patent Office: 327 Days
• Field of Search: 382/178, 382/177, 382/185, 382/186, 382/187, 382/179, 382/202, 382/203, 382/228, 382/229, 382/230
• US Class Current: 382/177
• CPC Class Codes: G06V 30/196 using sequential comparisons of the image signals with a plurality of references
• CPC Class Codes: G06V 30/32 Digital ink

Abstract

This invention discloses a method for automatically segmenting and recognizing Chinese character strings continuously written by a user in a handwritten Chinese character processing system, comprising the steps of: creating a geometry model and a language mode; finding out all of potential segmentation schemes in the Chinese character strings continuously written by a user based on the associated timing information and said geometry model; recognizing the groups of strokes as defined by each of potential segmentation schemes and computing the probability characterizing the exactness of recognition results; correcting the probability characterizing the exactness of recognition results by said language model; and, selecting the recognition result and the corresponding segmentation scheme having the maximum probability value.

Claims

1. A method for automatically segmenting and recognizing character strings continuously written by a user in a handwritten character processing system, wherein said handwritten character processing system records character strings continuously written by a user in strokes and associated timing information thereof, said method comprising the steps of:

creating a geometry model which describes geometric characteristics of stroke sequences of handwritten character strings and a language model which describes dependency among characters or words;

determining potential segmentation schemes in the character strings continuously written by a user based on said associated timing information and said geometry model;

recognizing groups of strokes as defined by each of the potential segmentation schemes and computing a probability characterizing the exactness of the recognition result;

correcting the probability characterizing the exactness of the recognition result by said language model; and

selecting the recognition result having the maximum probability value and the corresponding segmentation scheme as the segmentation and recognition result of the character strings continuously written by a user.

2. A method for automatically segmenting and recognizing handwritten character strings according to claim 1, wherein said geometry model comprises a gap model G, which characterizes the probability of the later stroke of two adjacent strokes being a potential boundary based on the gap between said two adjacent strokes, said two adjacent strokes referred as a former one and a later one with respect to said associated timing information.

3. A method for automatically segmenting and recognizing handwritten character strings according to claim 2, wherein said model G is a monotone ascending function, wherein the argument of said function is the gap between said two adjacent strokes and the dependent variable of said function characterizes the probability of the later stroke being a potential boundary.

4. A method for automatically segmenting and recognizing handwritten character strings according to claim 1, wherein said geometry model comprises a distribution model D, which characterizes the probability of the strokes between the current stroke and the last confirmed potential character boundary constructing a character based on the distance between the current stroke and the last confirmed character boundary.

5. A method for automatically segmenting and recognizing handwritten character strings according to claim 4, wherein said model D is a normal-like distribution function, wherein the argument of said function is the distance between the current stroke and the last confirmed potential character boundary and the dependent variable of said function characterizes the probability of the strokes between the current stroke and the last confirmed potential character boundary constructing a character, when the argument varies within a continuous value range, the value of said normal-like function is larger, otherwise smaller.

6. A method for automatically segmenting and recognizing handwritten character strings according to claim 1, wherein the recognition results of the strokes between two adjacent character boundaries can be radical, single character or multi-characters word.

7. A method for automatically segmenting and recognizing handwritten character strings according to claim 1, wherein when the writing direction is horizontal and the recognition result is a character having separable components lined up from left to right, a character width model D'"'"' can be used to correct the probability characterizing exactness of the recognition results, model D'"'"' characterizes the probability of a component constructing a part of a character or a character based on the width of the component.

8. A method for automatically segmenting and recognizing handwritten character strings according to claim 1, wherein when the writing direction is vertical and the recognition result is a character having separable components stacked up from top to down, a character height model D'"'"' can be used to correct the probability characterizing exactness of the recognition result, model D'"'"' characterizes the probability of a component constructing a part of a character or a character based on the height of the component.

9. A method for automatically segmenting and recognizing handwritten character strings according to claim 1, wherein a tree is used for searching the potential segmentation schemes.

10. A method for automatically segmenting and recognizing handwritten character strings according to claim 9, wherein said tree is a binary tree, wherein the root represents the last confirmed character boundary, the left branch represents that the previous potential boundary is not accepted, the right branch represents that the previous potential boundary is accepted, and whenever a potential character boundary is generated, the tree creates a branch down.

11. A method for automatically segmenting and recognizing handwritten character strings according to claim 10, wherein said binary tree may be pruned in accordance with a gap model G, a distribution model D and the recognition results.

12. A system for automatically segmenting and recognizing handwritten character strings, comprising:

input means, for accepting character strings continuously written by a user, and recording the user input in strokes and the associated timing information;

model storage means, for storing a geometry model which describes geometric characteristics of stroke sequences in handwritten character strings and a language model which describes dependency among characters or words;

segmenting means, for determining potential segmentation schemes in the character strings continuously written by a user based on said associated timing information and said geometry model;

recognizing means, for recognizing groups of strokes as defined by each of the potential segmentation schemes and computing a probability characterizing the exactness of the recognition result; and

arbitrating means, for correcting the probability characterizing the exactness of the recognition result by said language model; and

selecting the recognition result and the corresponding segmentation scheme having the maximum probability value as the segmentation and recognition result of the character strings continuously written by a user.

13. Apparatus for automatically segmenting and recognizing character strings continuously written by a user in a handwritten character processing system, wherein said handwritten character processing system records character strings continuously written by a user in strokes and associated timing information thereof, said apparatus comprising:

at least one processor operative to;

(i) create a geometry model which describes geometric characteristics of stroke sequences of handwritten character strings and a language model which describes dependency among characters or words;

(ii) determine potential segmentation schemes in the character strings continuously written by a user based on said associated timing information and said geometry model;

(iii) recognize groups of strokes as defined by each of the potential segmentation schemes and computing a probability characterizing the exactness of the recognition result;

(iv) correct the probability characterizing the exactness of the recognition result by said language model; and

(v) select the recognition result having the maximum probability value and the corresponding segmentation scheme as the segmentation and recognition result of the character strings continuously written by a user.