Cursive character handwriting recognition system and method

• US 20080123940A1
• Filed: 11/29/2006
• Published: 05/29/2008
• Est. Priority Date: 11/29/2006
• Status: Active Grant

• Current Assignee: Mitre Corporation
• Original Assignee: Mitre Corporation
• Inventors: Van Guilder, Linda C., Phillips, Jon, Hines, Tom, Kundu, Amlan, Huyck, Ben
• Granted Patent: US 8,005,294 B2
• Time in Patent Office: Days
• US Class Current: 382/159
• CPC Class Codes: G06V 30/2268 using stroke segmentation

Abstract

A cursive character handwriting recognition system includes image processing means for processing an image of a handwritten word of one or more characters and classification means for determining an optimal string of one or more characters as composing the imaged word. The processing means segments the characters such that each character is made up of one or more segments and determines a sequence of the segments using an over-segmentation-relabeling algorithm. The system also includes feature extraction means for deriving a feature vector to represent feature information of one segment or a combination of several consecutive segments. The over-segmentation-relabeling algorithm places certain segments considered as diacritics or small segments so as to immediately precede or follow a segment of the associated main character body. Additionally, the system also includes classification means that processes each string of segments and outputs a number of optimal strings which could be matched against a given lexicon.

Claims

1. A method for recognizing unconstrained cursive handwritten words, comprising:

processing an image of a handwritten word of one or more characters, the processing step including segmenting each imaged word into a set of one or more segments and determining a sequence of segments using an over-segmentation-relabeling algorithm;

extracting feature information of one segment or a combination of several consecutive segments;

repeating said extracting step until feature information from all segments or combinations thereof have been extracted; and

classifying the imaged word as having a string of one or more characters using the extracted feature information.

2. The method of claim 1, wherein the feature information includes moment features, geometrical and positional information based features, pixel distribution features, topological features, stroke connectedness features, and reference line features.

3. The method of claim 2, wherein the imaged word is an Arabic word and the feature information further includes aspect ratio features, location and number of disconnected dots, and chain code features.

4. The method of claim 1, further comprising:

hypothesizing a set of one or more words from a given dictionary which are suggested by the string of one or more characters determined for the imaged word in the classifying step.

5. An unconstrained cursive character handwritten word recognition system, comprising:

image processing means for processing an image of a handwritten word of one or more characters, wherein the processing of the imaged word includes segmenting the imaged word into a finite number of segments and determining a sequence of the segments using an over-segmentation-relabeling algorithm, wherein each character includes one or more segments;

feature extraction means for deriving a feature vector to represent feature information of one segment or a combination of several consecutive segments; and

classification means for determining an optimal string of one or more characters as composing the imaged word.

6. The handwritten word recognition system of claim 5, wherein the processing means includes means for slant normalization and noise reduction.

7. The handwritten word recognition system of claim 5, wherein a feature vector is derived for each segment and each combination of two or more consecutive segments.

8. The handwritten word recognition system of claim 5, wherein the over-segmentation-relabeling algorithm segments each character into at most two to seven segments.

9. The method of claim 8, wherein each character of the imaged word is segmented into at most four segments.

10. The handwritten word recognition system of claim 9, wherein a feature vector is derived for each segment and each combination of two to four consecutive segments.

11. The handwritten word recognition system of claim 5, wherein the feature information of includes moment features, geometrical and positional information based features, pixel distribution features, and reference line features.

12. The handwritten word recognition system of claim 5, wherein the imaged word is an Arabic word.

13. The handwritten word recognition system of claim 12, wherein the feature information further includes aspect ratio features, location and number of disconnected dots, stroke connectedness features, and chain code features.

14. The handwritten word recognition system of claim 13, wherein the feature information represented by the feature vector consists of between 25 to 50 features.

15. The handwritten word recognition system of claim 14, wherein the feature information represented by the feature vector consists of 45 features.

16. The handwritten word recognition system of claim 5, wherein a modified Viterbi algorithm is used to determine the optimal string.

17. The handwritten word recognition system of claim 16, wherein the modified Viterbi algorithm is adapted to a variable duration hidden Markov model by incorporating state duration probabilities associated with each distinct character, wherein a state duration probability of a given distinct character represents a probability that the given character has a duration of a defined number of segments.

18. The handwritten word recognition system of claim 5, wherein the classification means uses a continuous-discrete hybrid probability modeling of features to determine if a given feature vector is indicative of a given distinct character.

19. The handwritten word recognition system of claim 5, further comprising:

a post-processing means for hypothesizing a set of one or more words from a given dictionary which are suggested by the optimal string.

20. The handwritten word recognition system of claim 5, wherein the post-processing means uses the Levenshtein edit distance function with empirically determined weights to hypothesize the set of one or more words.

21. A method for training an unconstrained cursive character handwritten word recognition system, comprising:

processing a corpus of handwritten word images, each imaged word having one or more characters, the processing step including segmenting each imaged word into a set of one or more segments and determining a sequence of the segments using an over-segmentation-relabeling algorithm;

extracting feature information of individual characters of the imaged words;

estimating symbol probability parameters associated with each distinct character so as to allow a statistical measure that given feature information is indicative of a distinct character; and

estimating state duration probabilities associated with each distinct character, wherein a state duration probability of a given distinct character represents a probability that a segmented image of the given character will have a duration of a defined number of segments.

22. The method of claim 21, wherein the estimating symbol probability parameters step includes calculating representative feature information for each distinct character based on the feature information extracted from character images of a like distinct character in the imaged words.

23. The method of claim 21, wherein the duration of any given character is from one to four segments, and wherein a state duration probability is estimated for each duration of one to four segments for each distinct character.

24. A method for determining a sequence of segments of a segmented image of a cursive written word processed in a word recognition system, comprising:

finding the number of segments, wherein the finding step includes locating the first segment and the last segment in the imaged word; and

determining the sequence of segments using an over-segmentation-relabeling algorithm, wherein the over-segmentation-relabeling algorithm includes;

characterizing segments as either situated segments or unsituated segments, wherein situated segments include the first and last segments and segments having an X-coordinate or Y-coordinate coverage that exceeds a threshold value, small segments that are cursively connected to segments on each side, and wherein unsituated segments are segments not characterized as situated segments; and

placing each unsituated segment having a situated segment above or below so as to either immediately precede or follow the situated segment in the sequence of segments.

25. The method of claim 24, wherein each unsituated segment having a situated segment below is placed so as to immediately follow the situated segment below, and wherein each unsituated segment having a situated segment above but not a situated segment below is placed so as to immediately precede the situated segment above.

26. The method of claim 24, wherein each unsituated segment having a situated segment below is placed so as to immediately precede the situated segment below, and wherein each unsituated segment having a situated segment above but not a situated segment below is placed so as to immediately follow the situated segment above.

27. The method of claim 24, further comprising:

rechecking the segment sequence to ensure continuity of segment placements and no multiple placement of a segment.

28. The method of claim 24, wherein the imaged word has one or more characters, wherein each character of the imaged word is segmented into at most between two to seven segments.

29. The method of claim 24, wherein the situated segments further include any special diacritic present in the imaged word, wherein a special diacritic includes a double consonant or a shadda.