Image analysis – Pattern recognition – Feature extraction
Reexamination Certificate
1997-01-21
2002-04-09
Couso, Jose L. (Department: 2621)
Image analysis
Pattern recognition
Feature extraction
C382S186000, C382S226000, C382S253000
Reexamination Certificate
active
06370269
ABSTRACT:
FIELD OF THE INVENTION
The present invention relates to the field of optical character recognition (OCR) of cursive, normal handwriting by individuals. More particularly, it relates to the OCR of text that is written or printed in any of a plurality of languages where letters of the alphabet, even though small in number, may assume different shapes dependent on their position within a word, and which may connect to an adjacent character at their left, right, both, or not at all. It further relates to translation from one language, as represented by cursive script, to another. The method of the invention does not attempt to segment words into characters before recognition; rather it follows the writing strokes or traces from beginning to end; and only then attempts recognition of characters in a word (as in some English script) or in a sub-word or word (as in Arabic and cursive representations of many languages). An important feature of the invention is that it recognizes that sub-words may exist in a plurality of languages, and that an existing text may contain several languages; for example, it recognizes the common phenomenon that a quotation may be in a language different from the main language of the text.
BACKGROUND OF THE INVENTION
Examples of prior art directed to character segmentation are the following U.S. patents:
U.S. Pat. No. 4,024,500 granted May 17, 1977, and titled “Segmentation Mechanism for Cursive Script Character Recognition Systems”.
U.S. Pat. No. 4,654,873 granted Mar. 31, 1987, and titled “System and Method for Segmentation and Recognition of Patterns”.
U.S. Pat. No. 5,001,765 granted Mar. 19, 1991, and titled “Fast Spatial Segmenter for Handwritten Characters”.
U.S. Pat. No. 5,101,439 granted Mar. 31, 1992, and titled “Segmentation Process for Machine Reading of Handwritten Information”.
U.S. Pat. No. 5,111,514 granted May 5, 1992, and titled “Apparatus for Converting Handwritten Characters onto Finely Shaped Characters of Common Size and Pitch, Aligned in an Inferred Direction”.
U.S. Pat. No. 5,151,950 granted Sep. 29, 1992, and titled “Method for Recognizing Handwritten Characters Using Shape and Context Analysis”.
In U.S. Pat. No. 4,773,098 granted Sep. 20, 1988, and titled “Method of Optical Character Recognition”, individual characters are recognized by means of assigning directional vector values in contour determination of a character.
In U.S. Pat. No. 4,959,870 granted Sep. 25, 1990, and titled “Character Recognition Apparatus Having Means for Compressing Feature Data”, feature vectors having components which are histogram values are extracted and compressed then matched with stored compressed feature vectors of standard characters.
U.S. Pat. No. 4,979,226 granted Dec. 18, 1990, and titled “Code Sequence Matching Method and Apparatus”, teaches code sequence extraction from an input pattern and comparison with a reference code sequence for character recognition.
U.S. Pat. No. 3,609,685 granted Sep. 28. 1971, and titled “Character Recognition by Linear Traverse”, teaches character recognition in which the shape of the character is thinned to be represented by a single set of lines and converted to a combination of numbered direction vectors, and the set of direction vectors is reduced to eliminate redundant consecutive identical elements.
U.S. Pat. No. 5,050,219 granted Sep. 17, 1991, and titled “Method of Handwriting Recognition” is abstracted as follows:
“A method of recognition of handwriting consisting in applying predetermined criterions(sic) of a tracing of handwriting or to elements of this tracing so that several characterizing features of this tracing or of these elements be determined, comparing characterizing features thus determined to characterizing features representative of known elements of writing and identifying one element of the tracing with one known element of writing when the comparison of their characterizing features gives a predetermined result, wherein the improvement consists in the setting up of a sequence of predetermined operating steps in accordance with predetermined characterizing features by applying criterions to the tracing elements.”
The above United States patents are incorporated herein by reference, where permitted. None of the known prior art, however, teaches how to deal with units of interconnected text tracings wherein vectors remain unused after all characters have been recognized, nor how to deal with the appearance of multiple languages within a single document or on a single page.
SUMMARY OF THE INVENTION
It has been found that a more efficient character recognition is achieved using encoded units of interconnected text tracings as a sequence of directions in a plane where the units are recognized as sub-words, where all vectors in the text tracings are used to create the character or language fragment being recognized, and where the vector sequences are tested against one or a plurality of sets of language-specific rules.
It has further been found that the amount of pre-processing, before recognition but after acquisition of the text image and noise reduction and filtering, is reduced if the input text is not segmented into constituent characters before it is presented to the recognition engine. Thus, the natural segmentation inherent in the text image (due to spacing between words and sub-words) is adhered to and exploited.
In the present disclosure and claims, “sub-words” mean the intra-connected portions of words that are bounded by a break in the cursive text, i.e. where successive characters are not bound by a ligature. Sub-words can be as long as the entire word or as short as one character, or even a portion of a character if, for example, the character includes a secondary feature.
The present invention provides an improvement to the known methods of optical character recognition in which the characters can comprise a plurality of languages, comprising an intermediate step wherein an acquired text image consisting of a sequence of planar directional vectors is analyzed by the recognition engine in chunks of intra-connected sub-words, the cursive text is parsed and a character marker is entered upon the recognition of each successive sub-word, and if unused vectors remain following the recognition of connected sub-units of text, then the text is reparsed by moving the character marker forward or backward one vector at a time until each vector in the sequence contributes to recognition of the characters of the text, as described in copending Canadian patent application S. N. 2139094. The recognition engine further uses a first set of language-specific rules, and if after exhausting the entries in the first set of language-specific rules a particular sub-word is not recognized, it compares that sub-word with a second set of language-specific rules until the sub-word is recognized.
The present invention further provides an apparatus for recognition of cursive text in one or more of a plurality of languages from a scanned image, including means for recognizing a sequence of directional vectors as characters only if all of the vectors have contributed to the recognition, means for reparsing the sequence of directional vectors until all of the vectors do contribute to recognition, at least two language-specific dictionaries, and means for comparing the sequence of direction vectors with the language-specific dictionaries. Code to control a computer for carrying out the steps of the method can be programmed onto a suitable medium, for example a magnetic storage diskette or a programmable read-only memory.
The present invention further provides a computer-usable medium containing program code executable by the computer to perform a method for recognition of cursive text in one or more of a plurality of languages from a scanned image, including reparsing a sequence of directional vectors by moving a character marker one vector at a time until each vector in the sequence contributes to recognition of the characters of the text from at least one set of language-specific rules. Examples of media suitab
Al-Karmi Abdel Naser
Singh Meenu
Singh Shamsher S.
Soor Baldev Singh
Couso Jose L.
Desire Gregory
F. Chau & Associates LLP
International Business Machines - Corporation
Singh Meenu
LandOfFree
Optical character recognition of handwritten or cursive text... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Optical character recognition of handwritten or cursive text..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Optical character recognition of handwritten or cursive text... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2839535