Image analysis – Pattern recognition – Feature extraction
Reexamination Certificate
2000-01-22
2003-12-30
Boudreau, Leo (Department: 2621)
Image analysis
Pattern recognition
Feature extraction
Reexamination Certificate
active
06671404
ABSTRACT:
FIELD OF THE INVENTION
The invention relates to a pattern recognition apparatus and a pattern recognition method and in particular to a pattern recognition method that combines different recognition modules to improve the recognition accuracy.
BACKGROUND OF THE INVENTION
The demand for information processing involving pattern recognition is currently large and is rapidly increasing. Pattern recognition is used in such applications as image processing, text processing, and sound processing performed by computers. Consequently, improvements in pattern recognition technology are strongly desirable.
Pattern recognition is the process by which a physical phenomenon, such as an image, hand written or printed character or a sound, is converted to an electronic signal representing the phenomenon, a determination is made of which one of a number of possible categories in a category set the phenomenon belongs to, and a code indicating the determined category is generated. For example, in character recognition, an unknown printed character, such as the letter “A,” may be scanned by an electronic scanner. The scanner generates an electronic pattern signal that represents a pattern composed of an array of a few thousand bytes that represent the unknown character. The pattern signal is then analyzed to determine to which of the categories in the category set corresponding to the letters A-Z the unknown character belongs. A code identifying this category as the category of the unknown character is then generated. For example, the ASCII code
33
representing the letter A may be generated.
Pattern recognition processing is preferably performed using features of the pattern extracted from the pattern signal instead of using the raw pattern signal. Processing the features extracted from the pattern signal is preferable because these features can often be processed faster, more accurately and cheaper than the raw pattern signal. If pattern signals containing extremely large quantities of information are processed, features are sometimes extracted, and the features must be processed. One objective of pattern recognition is to compress the information by representing the patterns using the features extracted from the pattern signal. Of course, the features must be extracted from the pattern signal in a way that does not impair the ability of the pattern recognition processing to recognize the pattern.
The feature f of a pattern p is usually defined by a set {x(p;m); m=1,2,3, . . . , M} of a finite number M of feature components x(p;m). The feature f tangibly and quantitatively represents the characteristic qualities of the pattern. Consequently, since the feature f is represented by an M-dimensional vector whose m-th component is the feature component x(p;m), the vector representation of the feature f is the feature vector X(p)=(x(p;1), x(p;2), . . . , x(p;m))
t
. The argument p indicates the feature vector X(p) is the feature vector of the pattern p. The superscript t denotes vector transposition.
Even though the feature components are qualitative, they can be quantified and used.
If the pattern p undergoes various deformations, the value of the feature component x(p;m) changes. Consequently, the feature vector X(p) changes. However, as long as the deformed pattern belongs to its original category, the pattern recognition process must recognize it as belonging to that category.
A particular pattern that specified as being representative of the patterns belonging to a particular category or as being representative of a feature of the category is called the reference pattern of the category. The feature vector of the specified pattern is called the reference vector of the category. As an alternative to using a particular pattern as the reference pattern for the category, a hypothetical pattern obtained by averaging the patterns belonging to the category can be used as the reference pattern, and the feature vector of such hypothetical pattern can be used as the reference vector of the category.
In pattern recognition, an unknown pattern p is received and a pattern recognition process is performed. The pattern may determine whether the unknown pattern is similar to a known pattern q, or to determine which category the unknown pattern belongs to. Pattern recognition is essential in recognizing diagrams, characters, symbols, images, and sounds. General information about pattern recognition and the problems of pattern recognition can be found in Reference 1, Kazuo Nakada, Ed.:
Pattern Recognition and its Application
, Corona Co. (1978) (in Japanese) and Reference 2, Hidemitsu Ogawa, Ed.:
New Developments in Pattern Recognition and Understanding—The Challenges—Denshi Jyoho Tsushin Gakkai
(1992) (in Japanese).
Examples of pattern recognition in which the unknown patterns are character patterns will described below on the understanding that the principles set forth in the description can easily be applied to other forms of pattern recognition, such as image recognition and sound recognition. Character patterns are patterns representing letters, numbers, Kanji characters and the like. A pattern representing a character will from now on be referred to as a character pattern. Examples of possible feature components of character patterns include:
the length-to-width ratio of the character,
the number of horizontal lines,
the number of loops,
whether each square of a grid overlaid on the character is black or white,
the number of crossing points with a straight line in a specific direction, and
the transform coefficients of a Fourier transform of the character pattern.
A set of feature components such as that listed above is used to construct the feature vector so that the resulting feature vector can optimally represent the characters in the character set. The dynamic range of each feature component is selected to improve the accuracy of the pattern recognition to be described later. The feature component may be normalized using the standard deviation when this is needed.
Pattern recognition generates a category name for each character pattern subject to pattern recognition. The category code represents the reading, meaning, or code of the character pattern. For example, the category name of the category to which the character “A” belongs may be “category A.” As noted above, a specific character pattern belonging to the category is selected as the reference pattern for the category. Alternatively, a hypothetical pattern obtained by averaging a number of character patterns belonging to a category may be used as the reference pattern. The feature vector of the reference pattern is adopted as the reference vector of the category.
At the heart of pattern recognition is a recognition processor that has the objective of determining that all unknown character patterns that represent the character “A,” belong to category A, irrespective of whether the character pattern is deformed, and, further, that such character patterns do not belong to categories other than category A.
The processing performed by a character recognition apparatus after character pattern observation and reading is usually divided into a series of process modules that perform character pattern preprocessing, feature extraction, and recognition. Each process module can primarily be implemented using a computer and is realized by the computer performing a specific set of operations. All of the process modules, including observation of the character pattern, affect the result generated by the recognition module.
The strategy for increasing the accuracy of character recognition is to maximize the recognition ratio and to reduce the misrecognition ratio to zero. The recognition ratio is the fraction of character patterns that should belong to each category that are correctly recognized as belonging to that category. The misrecognition ratio is the fraction of characters patterns that do not belong to each category that are misrecognized as belonging to that category. In particular, many applications strongly demand that misre
Kawatani Takahiko
Shimizu Hiroyuki
Akhavannik Hussein
Boudreau Leo
LandOfFree
Method and apparatus for recognizing patterns does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method and apparatus for recognizing patterns, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method and apparatus for recognizing patterns will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3182385