JPH10320501A - Character recognition method and device therefor, and recording medium recording character recognition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a character recognition device capable of preventing erroneous recognition due to the writing characteristics of an individual even in the case that a person who writes characters can not be specified. SOLUTION: By judging whether or not a character pattern group recognized as the same category as the result of executing a character recognition processing can be defined as the same category even in an individual, by comparing the standard deviation of character distribution and a distance between the characters in the individual with a value averaged for plural persons and correcting the category of the character pattern of low recognition reliability in the case that it is the different category in the individual, the erroneous recognition of the character to be easily mistaken is prevented.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an optical character recognition device for optically capturing and recognizing characters written on paper or the like. In particular, the present invention relates to a character recognition process in an optical character recognition device that recognizes handwritten characters having greatly different character shapes depending on the writer, such as numbers and hiragana.

[0002]

2. Description of the Related Art Conventionally, when a handwritten character is read by a character recognition device of this type, a predetermined reference pattern (referred to as a dictionary or a template) is used to determine to which category a single pattern input to the device belongs. ) Was determined. When constructing a reference pattern of a certain category, for example, register an average pattern of character patterns of the same category described by a large number of people as a reference pattern, or form a plurality of representative patterns for the same category in a form called a multi-template. It was created by registering a pattern as a reference pattern. However, in this method, since the writing characteristics of the individual are not necessarily reflected in the reference pattern, there is a problem that a specific character is erroneously recognized as a different specific character at a high frequency depending on the writer. For example, as shown in FIG.
If "No" written by the person A and "9" written by the person B are very similar patterns, these character patterns are "no" for Mr. A and "9" for Mr. B. It is difficult to recognize correctly. That is, Mr. A is correctly recognized, but Mr. B is frequently misrecognized as "no", or Mr. B is correctly recognized, but Mr. A is recognized as "9". Very likely to be either. In order to solve this problem, for example, "Yukio Fukuda: Japanese Patent Application No. 6
In the case where a person who writes a character can be specified as described in “Pattern Recognition Method and Storage Medium Used therein”, a method of performing recognition by using a reference pattern specialized for each individual is known. However, in this method, when it is difficult to specify who has written the characters, such as a general form, it is not possible to use a reference pattern specialized for each individual. There is a problem that it is not possible. Even if it is difficult to determine who wrote the character, a method that can be applied is, for example, "
Yoshinobu Hotta et al .: "Character Recognition Method Based on Individual Writing Characteristics by Clustering" IEICE Technical Report, PRMU96-183
(1997) ", when characters of the same category are described in a certain number or more (for example, seven characters or more) in the same document, character patterns are grouped (clustered) for each document and personalized. There is known a method of statistically reflecting the writing characteristics of the character.

[0003]

However, it is difficult to specify who has written the characters, such as ordinary mail.
If the number of characters in the same category is small in the same document, not only can individual reference patterns not be used individually, but also cluster character patterns for each document to statistically write individual writing characteristics. Therefore, it is difficult to apply the conventional technology. According to the present invention, even when a person who describes a character cannot be specified, and even when the number of characters in the same category in the same document is small, a specific character generated due to the writing characteristics of an individual has a high frequency. It is an object of the present invention to provide a character recognition device capable of coping with the problem that a wrong specific character is erroneously recognized.

[0004]

According to the present invention, there is provided a character recognition method for recognizing a character string pattern by inputting a character string pattern, wherein the character recognition results in the recognition of the same category. For the pattern group, it is verified whether or not the same individual is included in the fluctuation range generated when the characters of the category are described, by using an average value of the fluctuation range obtained in advance for a plurality of persons, and the character When a part of the pattern group is compared with another part of the character pattern group and is determined that the characters are not in the same category described by the same individual, the category of the part or the other part is changed. It is characterized by doing. In order to solve the above-mentioned problems, the present invention provides a character recognition device for inputting and recognizing a character string pattern, comprising: an input unit for inputting an image of a character string; reading the image of the character string from the input unit; A character cutout unit for cutting out a character pattern, an initial recognition unit for reading out the character pattern cut out from the character cutout unit, performing character recognition, and storing the character pattern in association with a category, and the same individual described. A personal variation range storage unit that stores, for each category, a variation range of a character pattern that can be determined to be the same category, and reads the character pattern and the category associated with the character pattern from the initial recognition unit. Read the fluctuation range from the personal fluctuation range storage unit, and read the characters associated with the same category. In the same category determination unit that determines whether the turn group is within the fluctuation range, and in the same category determination unit, a part of the character pattern group is identical to another part of the character pattern group. A character re-recognition unit that changes the category of the part or the other part when it is determined that the character does not belong to the same category described by the individual. In order to solve the problem, the present invention provides a character recognition method for performing character recognition using a plurality of reference patterns per the same category, wherein a distance in a feature space between a plurality of reference patterns used in the same category is calculated, For a plurality of persons, a variation range that occurs when the same individual describes a character of the same category is determined, an average value thereof is calculated, the distance is compared with the average value, and a part of the plurality of reference patterns is the plurality of reference patterns. When it is determined that the characters are not in the same category described by the same individual as compared with the other parts of the reference pattern, the part and the other part are used when recognizing a character string written by the same individual. Is characterized in that it is not simultaneously selected as a reference pattern in a character string. The present invention, in order to solve the above problems,
In a character recognition device that performs recognition by inputting a character string pattern, an input unit that inputs a character string image, a character cutout unit that reads out the character string image from the input unit, and cuts out a character pattern one by one, An initial recognition unit that reads out a character pattern cut out from the character cutout unit, performs character recognition using a plurality of reference patterns per the same category, and stores the character pattern in association with a category. Calculate the distance in the feature space used for character recognition in the initial recognition unit between the reference patterns, and compare the distance with the variation range of the character pattern that can be determined to be in the same category when the same individual is described. In each category, a set of reference patterns that can be Reading and personal fluctuation range storage unit for storing each, the category associated to the character pattern and the character pattern from the initial recognition section,
The same category that reads the set of reference patterns from the individual variation range storage unit and determines whether a character pattern group associated with the same category can be associated with the same category in an individual according to the set of reference patterns. A determination unit, in the same category determination unit, when a part of the character pattern group is compared with another part of the character pattern group, and is determined not to be a character of the same category described by the same individual; , A character re-recognition unit that changes the category of the part or the other part.

Hereinafter, the principle of the present invention will be described. For example,
When the character pattern of the character "no" of Mr. A as shown in FIG. 2 is presented alone, it is a matter that even a human cannot determine whether the character pattern is "no" or "9". It is thought that more people answered "9" from their impressions. However, as shown in FIG. 3, when the character pattern “no” is presented as a part of a character string, and both “no” and “9” are described in the character string,
By determining whether the character pattern “9” is within the average variation range that can occur when the same individual describes the character pattern “no”, it is determined that the two are not at least in the same category can do. Further, on the premise that the character string is an address, it is possible to recognize the entire character string as "19 of 16/19", but this principle will be described later. Therefore, an average variation range that can occur when the same individual is described for each category (hereinafter, referred to as an individual average variation range) is obtained in advance. As a result of character recognition, if there is more than one character pattern associated with the same category, it is determined whether those character patterns belong to the same category in individuals based on the average variation range of the individual in that category. To judge. If it is determined that they are not the same, it is considered that one of the character patterns has been erroneously recognized, so that erroneous recognition can be reduced by correcting one of the categories. The following method can be considered as a method of obtaining the individual average fluctuation range. The first method is a method of using information indicating how wide a character pattern is when a person writes, that is, the character pattern is distributed in a feature space with a standard deviation. As shown in FIG. 4, when a certain individual writes a character pattern “9”, the distribution of the character pattern becomes smaller than the distribution of the entire category “9”. This is because the distribution of the entire category is a distribution in which the variation between individuals is taken into account, and the distribution of “9” of Mr. A, the distribution of B
This is because the distribution of many individuals such as the distribution of “9” of Mr. and the distribution of “9” of Mr. C corresponds to the distribution of the entire category of “9”. Therefore, "9"
When the standard deviation of the distribution is calculated for each individual with respect to the character pattern, and the average is calculated, the average individual standard deviation is smaller than the standard deviation of the category “9” as a whole. This means that character patterns that are simultaneously included in the distribution of the entire category but may not be simultaneously included in the distribution of individuals may exist. Hereinafter, an example will be described. No. 302, No. 305,
307, 308, 311 character patterns and "no"
It is assumed that the distribution of the entire category “9” and “9” has a relationship as shown in FIG. Here, 30
The second and third character patterns are both in the overall distribution of the category "9". FIG. 6 shows the average individual standard deviation of “9” in a circle centered on the 311 character pattern and the average individual standard deviation of “no” in a circle centered on the 302 character pattern. Show. In FIG. 6, the circles indicated by the two average individual standard deviations do not include the center of each other circle. This means that if the average individual standard deviation is the individual average variation range, assuming that the 311th character pattern is “9”, the 302nd character pattern is not within the “9” individual average variation range. Similarly, assuming that the character pattern of No. 302 is “No”, the character pattern of No. 311 is not within the personal average variation range of “No”. in this way,
302 based on the average individual standard deviation of each category
By evaluating the distance in the feature space between the character pattern of number 311 and the character pattern of number 311, can the character patterns 311 and 302, which can be in the same category as a whole, be accepted as the same category in an individual? Can be determined. A second method is to use information indicating how far a character pattern is distributed in a feature space from other character patterns when an individual writes. That is, the average value of the distribution center of an individual character pattern and the distance between distribution centers of other categories written by the same individual who is closest to the distribution center, that is, the average individual character distance is used. Now, for example, for the 311 and 302 character patterns in the overall distribution of the category “9”, the circle between the 311 character pattern and the circle is the average personal character distance of “9” by 302. FIG. 7 shows a case where the average personal character distance of “no” is represented by a circle centered on the number character pattern. In FIG. 7, the circles indicated by the two average individual standard deviations do not include the center of each other circle. This means that assuming that the average personal character distance is the personal average variation range, assuming that the character pattern of No. 311 is “9”, the character pattern of No. 302 is within the personal average variation range of “9”. Means that the character pattern of No. 302 is "no".
The character pattern of No. 311 means that it is not within the individual average fluctuation range of “no”. As described above, by evaluating the distance in the feature space between the character pattern of No. 302 and the character pattern of No. 311 on the basis of the average inter-character distance of each category, No. 311 which can be in the same category as a whole is obtained. And the 302 character pattern
It is possible to determine whether an individual is allowed as the same category. As described above, the method of determining whether an individual can be regarded as the same category by comparing the distance between the character patterns associated with the same category and the individual average variation range has been described. When character recognition is performed using a reference pattern, the following method is also conceivable. That is, by comparing the distance between a plurality of reference patterns belonging to the same category and the individual average variation range, a set of reference patterns that can be simultaneously adopted by an individual is determined. If the reference patterns that are not the same set are simultaneously adopted, it is determined that the character patterns respectively associated with the reference patterns that are not the same set do not belong to the same category in the individual. Next, a method of correcting the category when it is determined that the category is not the same in the individual will be described. For example, at the time of character recognition, a character pattern is associated with a category, and the likelihood of association is also stored, and the character pattern with the highest likelihood and the character pattern determined to be the same category in the individual as the character pattern , The category is not changed, and the category is corrected for the character pattern determined to be not the same category in the individual from the character pattern having the highest probability. As a correction method, for example, when a plurality of categories are ranked and associated with one character pattern at the time of character recognition, the second-ranked category is adopted as the corrected category. If the character string describes an address, the category can be determined by utilizing the knowledge of the address. For example, in FIG. 3, 302, 305, 307,
The category "9" was initially associated with the numbers 308 and 311.
Set of No. 02, 305, 308 and 307, 31
Assume that the first set is determined not to be in the same category. It is also assumed that “no” is associated with these character patterns as the second category. In this case, any set is associated with “9” or “no”
Depending on whether or not it is associated with
And the combination "910 of 191691" can be considered. Here, the knowledge that the character string is an address, that is, the knowledge that there is no address having a large crest of 191691, indicates that the latter combination is inappropriate, and the former is selected. As a result, 30
The categories of the character patterns of Nos. 2, 305 and 308 are not changed, and the categories of the character patterns of Nos. 307 and 311 are corrected to "no".

[0006]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an example of the embodiment of the present invention. In the present embodiment, an input unit 11 for inputting an image of a character string, a character cutout unit 12 that reads the image of the character string from the input unit 11 and cuts out a character pattern one by one, and a character cutout unit 12 The initial recognition unit 13 reads out the written character pattern, performs character recognition, and stores the character pattern in association with a category, and a variation range of the character pattern that can be determined to be the same category when the same individual is described. A personal variation range storage unit 14 for storing
Reading the character pattern and the category associated with the character pattern from the initial recognition unit 13;
The same category determination unit 15 reads the variation range from the individual variation range storage unit 14 and determines whether a character pattern group associated with the same category is within the variation range. , When a part of the character pattern group is compared with another part of the character pattern group, and it is determined that the character is not a character of the same category described by the same individual, the part or the other part And a character re-recognition unit 16 for changing the category. Next, the operation of the present embodiment will be described. The input unit 11 is a unit for inputting an image of a binarized character string. FIG. 3 shows an example of a character string image input to the input unit 11. The character cutout unit 12 is a means for cutting out and inputting a character pattern for each character from an image of a character string. The character cutout section 12 can be easily realized by using, for example, the technique described in Japanese Patent Application No. 07-008793. FIG. 8 shows an example of the extracted character pattern. The initial recognition unit 13 includes the character cutout unit 12
This is a means for performing character recognition processing on a character pattern input from, and storing, as character recognition result information, a category corresponding to the character pattern and reliability indicating the likelihood of association. For example, Tsukun: "Improvement of direction pattern matching method and application to handwritten kanji recognition"
IEICE Technical Report, PRU90-20 (1990) can be used. As the reliability, a distance value or a similarity in a feature space can be used.
1433 ". Here, there may be a plurality of reference patterns used for character recognition per category. When a plurality of reference patterns are used, a serial number is added to each reference pattern, and the serial number of the reference pattern is also output as character recognition result information. FIG. 9 shows a specific example of the character recognition result information stored in the initial recognition unit 13. In FIG. 9, 1
The ranking candidate and the second-ranking candidate represent the category having the highest possibility and the category having the second highest possibility as a result of the character recognition, and the reliability is indicated by percentage. When the reliability of the category of the first candidate is high, the category of the second candidate is not associated. For example, “1-2” of the reference pattern serial number indicates that it is the second reference pattern of the category “1”. The individual variation range storage unit 14 is means for storing, for example, the following information as the individual average variation range. As a first example, the average individual standard deviation is stored. The average individual standard deviation is obtained by averaging the standard deviations of the distribution of individual character patterns for a plurality of persons for each category. As a second example, the average personal character distance is stored. The average inter-individual character distance is the average of the distance between the distribution center of an individual's character pattern and the distribution center of another category written by the same individual that is the closest to each other for each category. Means what you do. Same category judgment unit 15
Reads out the character pattern and character recognition result information stored in the initial recognition unit 13, reads out the individual average variation range stored in the individual variation range storage unit 14, and reads a character pattern group associated with the same category. This is a means for determining whether or not the individual can be said to be in the same category based on the individual average fluctuation range. The operation of this means will be described for the case where the average individual standard deviation is stored in the individual fluctuation range storage unit 14. When a plurality of character patterns in a character string are associated with the same category in the initial recognition unit 13, a distance Dk (i, j) between these character patterns in the feature space is obtained. Here, the symbol k represents a category, and i and j are categories k
Represents the serial number of the character pattern associated with. If Dk (i, j) is not more than α1 times the average individual standard deviation of the category k stored in the individual variation range storage unit 14, it is determined that the character pattern i and the character pattern j are the same category,
If it exceeds α1, the category is determined to be different. Here, the value α1 may be a positive real number, and the frequency of the category correction in the character re-recognition unit 16 can be adjusted according to its magnitude. However, it is desirable to set the value in the range of 0.5 to 3.5. . As a specific example, in FIG. 3, characters 302, 305,
Consider a case where the character pattern of No. 308 and the character patterns of No. 307 and No. 311 which are originally “9” are all associated with the same category of “9”. FIG. 9 shows character recognition result information in this case. That is, five character patterns are associated with the category “9”, and these five character patterns (character pattern numbers: 302, 305, 307,
308, 311) = (No. In category: 1, 2, 3,
The serial numbers correspond to each other as in 4, 5). Next, a distance Dk (i, j) between the five character patterns in the feature space is obtained. For example, when calculating the distance Dk (i, j) for the 305th and 311th numbers, the code k represents a category, so that k = 9, and the 305th category number is 2 and the 311th category number is 5; D9 (2,5) will be calculated. In FIG.
Distance Dk (i, j) between these five character patterns
Here is an example in which the values are obtained in a table. This distance Dk (i, j)
And a value obtained by multiplying the average individual standard deviation for category “9” stored in the individual variation range storage unit by α1. For example, when the average individual standard deviation is 10 and α1 is 2.5, the distance Dk (i, j) is compared with 25. In the example shown in FIG. 10, there are four combinations in which the distance is 25 or less, and if the groups are grouped so as not to have any combination in which the distance is 25 or less, 3
Groups 02, 305, 308, 307,
It can be divided into the 311th group. This means that groups 302, 305 and 308
The 7th and 311th groups indicate that individuals cannot be said to be in the same category. Next, the operation of the present means will be described for the case where the average individual character distance is stored in the individual variation range storage unit 14. The process of obtaining the distance Dk (i, j) is the same as that in the case where the average individual standard deviation is stored.
If it is less than 1 time, character pattern i and character pattern j
Are judged to be the same category, and when they exceed β1 times, they are judged to be different categories. Here, the value β1 may be a positive real number, and the character re-recognition section 1
Although the frequency of the category correction in 6 can be adjusted, it is desirable to set the frequency in the range of 0.2 to 1.5. As described above, the operation of determining whether or not an individual can be considered to be the same category by comparing the distance between the character patterns associated with the same category with the individual average variation range has been described. When character recognition is performed using a plurality of reference patterns, the following embodiment is also conceivable. That is, by comparing the distance between a plurality of reference patterns belonging to the same category and the individual average variation range, a set of reference patterns that can be simultaneously adopted by an individual is determined. Then, when the reference patterns that are not the same set are simultaneously adopted, it is determined that the character patterns associated with the reference patterns that are not the same set do not belong to the same category in the individual. Specifically, when it is determined that a plurality of character patterns in a character string belong to the same category k, a distance Dk (iDi, iDj) in a feature space between reference patterns associated with each character pattern is obtained. . Here, the symbol k represents a category, iDi and i
Dj is the serial number of the reference pattern associated with the character pattern of the category number i, j. This Dk (i
If (Di, iDj) is equal to or smaller than the value obtained by multiplying the average individual standard deviation stored in the individual fluctuation range storage unit 14 by α2, for example, it is determined that the categories are the same category, and if it exceeds α2 times, the categories are different. Is determined. Where the value α
2 may be a positive real number, and the frequency of the category correction in the character re-recognition unit 16 can be adjusted according to the magnitude of the real number. However, it is desirable to set the number in the range of 0.5 to 3.5. When the average individual character distance is stored in the individual variation range storage unit 14, if the average individual character distance is equal to or less than β2 times the average individual character distance, the category is determined to be the same category. Judge as a category. Here, the value β2 may be a positive real number, and the frequency of the category correction in the character re-recognition unit 16 can be adjusted according to the magnitude. However, it is desirable to set the value in the range of 0.2 to 1.5. . The following embodiment is also conceivable. That is, the distance between the reference patterns is calculated in advance for each category, and the individual average variation range is compared with the distance. A set of reference patterns is determined and stored in the personal variation range storage unit 14 in the form of, for example, a table as shown in FIG. In FIG.
The same reception serial number means a serial number of a reference pattern that can be received simultaneously as the same category by an individual. For example, for a reference pattern with a reference pattern serial number of 9-2, the same reception serial number is 9-3, which means that the reference patterns of 9-2 and 9-3 may be simultaneously referred to by an individual. It represents that. The same-category determining unit 15 compares the serial number of the reference pattern of the character recognition result information read from the initial recognition unit 13 with a set of reference patterns that can be read as the same category in the individual read from the individual variation range storage unit 14. Thus, it is determined whether or not each character pattern can be said to be the same category in an individual. Also, for example, it is determined that the pattern A and the pattern B and the pattern B and the pattern C are in the same category.
Are determined to be different categories, patterns A, B, and C can be treated as the same category.
The character re-recognition unit 16 is a unit for re-categorizing a character pattern that is determined by the same category determination unit 15 as not belonging to the same category in an individual. This processing can be realized as follows. For example,
Among the character patterns determined not to be the same category in the individual, the character pattern with the highest reliability is selected. The corresponding category is not changed for such a character pattern and another character pattern determined to be the same category in the individual as the character pattern, and the corresponding category is changed for other character patterns. I do. As a change method, for example, when a plurality of categories are associated with one character pattern in a state of being ranked as the character recognition result information, the second category is adopted as the changed category. FIG. 9 shows the result of character recognition information in the case where one character pattern is associated with up to the second largest candidate. When a plurality of reference patterns are used for one category, the following embodiment can be considered. For example, in the character re-recognition unit 16, priorities are added in advance to the reference patterns for each category in the order of difficulty in making the reference patterns different from other categories and stored. Then, the reference pattern having the highest priority is selected from the reference patterns associated with the character pattern. The category is not changed for the reference pattern and the character pattern associated with the reference pattern that is considered to be simultaneously referred to by the individual, and the category associated with the other character patterns is changed. . As a change method, for example, when a plurality of categories are associated with one character pattern in a state of being ranked as the character recognition result information, the second category is adopted as the changed category. Further, when the character string describes an address, the following embodiment can be considered. For example, in the initial recognition unit 13, the numbers 302 and 305 shown in FIG.
Although the category “9” is associated with the character patterns No. 307, No. 308 and No. 311, the same category determination unit 15
Assume that it is determined that the sets of Nos. 305 and 308 and the sets of Nos. 307 and 311 do not belong to the same category. It is also assumed that "no" is associated with these character patterns as the second category. The character re-recognition unit 16 receives the information that the sets of Nos. 302, 305, and 308 and the sets of Nos. 307 and 311 are different categories, and changes the category of one set to “no”. As a result, "1 of 16 of 19 of 10
A combination “9” and a combination “910 of 191691” are created. On the other hand, the character re-recognition unit 16 stores the range of the street, address, and number as address knowledge in advance. Among the created combinations, it is determined whether or not the numbers corresponding to the streets, addresses, and numbers are within the range, and combinations that fall outside the range are rejected. In this example, the number 191691 in the latter combination is out of range for any of the streets, addresses, and numbers and is rejected.
Therefore, the former combination of “19 of 16 of 19” is selected, and the character patterns of Nos. 302, 305, and 308 are not changed in category, and are associated with the character patterns of Nos. 307 and 311. Change the category to “no”. In addition, as a method of using the address knowledge, a technique described in Shimomura: "Street address knowledge processing method in handwritten address reading", Information 51st, 4R-8 (1995) can be used.

Next, another embodiment of the present invention will be described in detail with reference to the drawings.

Referring to FIG. 12, an embodiment of the present invention comprises:
It has a recording medium 17 on which a character recognition program is recorded.
The recording medium 17 may be a CD-ROM, a magnetic disk, a semiconductor memory, or another recording medium, and includes a case where the recording medium is distributed via a network.

The character recognition program is read from the recording medium 17 to the data processing device 18 and controls the operation of the data processing device 18. Under the control of the character recognition program, the data processing device 18 extracts characters from the character string image input from the input unit 11, performs character recognition on the extracted character patterns, and performs character recognition. For the character pattern group recognized as the same, whether or not the same individual is included in the fluctuation range generated when the characters of the category are described, the average value of the fluctuation range obtained in advance for a plurality of persons. And verifying that the part of the character pattern group is not a character of the same category described by the same individual as compared with the other part of the character pattern group. Alternatively, the category of the other part is changed.
That is, the data processing device 18 controls the character recognition program, and performs the same processing as the character cut-out unit 12, the initial recognition unit 13, the individual variation range storage unit 14, the same category determination unit 15, and the character re-recognition unit 16 in FIG. Execute the process.

[0010]

As described above, according to the present invention,
When it is not possible to prepare a personalized reference pattern because the individual cannot be identified,
In addition, even if character patterns of the same category do not exist in the character string enough to statistically extract individual writing characteristics, erroneous recognition caused by individual writing characteristics can be reduced.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram showing a configuration of the present invention.

FIG. 2 is an explanatory diagram showing a difference in a character pattern for each individual.

FIG. 3 is a diagram showing a character string.

FIG. 4 is an explanatory diagram showing the relationship between the spread of the distribution of the character pattern “9” described by an individual and the spread of the distribution of the entire category “9”.

FIG. 5 shows the position of a character pattern in the feature space and “9”.
FIG. 9 is an explanatory diagram showing the distribution of the entire category of “and”.

FIG. 6 is an explanatory diagram showing a circle having a radius of a position of a character pattern in a feature space and an average individual standard deviation.

FIG. 7 is an explanatory diagram showing a circle having a radius of a position of a character pattern in a feature space and an average distance between individual characters.

FIG. 8 is an explanatory diagram showing a specific example of an output of a character cutout unit.

FIG. 9 is an explanatory diagram showing a specific example of character recognition result information.

FIG. 10 is an explanatory diagram showing a specific example of a distance value between character patterns associated with a category “9”.

FIG. 11 is an explanatory diagram showing a specific example of a table representing a set of reference patterns that may be simultaneously referred to by an individual.

FIG. 12 is an explanatory diagram showing an embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 11 Input part 12 Character cut-out part 13 Initial recognition part 14 Individual variation range storage part 15 Same category judgment part 16 Character re-recognition part 17 Storage medium 18 Data processing device

Claims (14)

[Claims] In a character recognition method for recognizing a character string pattern by inputting a character string pattern, a character of a character group belonging to the category is recognized by a character pattern group that is recognized as having the same category as a result of the character recognition. Whether it is included in the variation range that occurs when described, is verified by using the average value of the variation range obtained in advance for a plurality of people, a part of the character pattern group, the other of the character pattern group, A character recognition method characterized by changing the category of the part or the other part when it is determined that the character does not belong to the same category described by the same individual as compared with the part. 2. An input unit for inputting a character string image in a character recognition device for inputting and recognizing a character string pattern.
Reading the image of the character string from the input unit, a character cutout unit that cuts out a character pattern one character at a time, and reads a character pattern cut out from the character cutout unit;
An initial recognition unit that performs character recognition and stores a category for each character pattern in association with each other, and a variation range of a character pattern that can be determined to be the same category when the same individual describes, for each category. A personal variation range storage unit, the character pattern and the category associated with the character pattern are read from the initial recognition unit, the variation range is read from the personal variation range storage unit, and the characters associated with the same category are read. In the same category determination unit that determines whether the pattern group is within the fluctuation range, in the same category determination unit, a part of the character pattern group is identical to another part of the character pattern group. If it is determined that the characters do not belong to the same category described by the individual, the part or the other part Character recognition apparatus comprising: a character re-recognition unit to change the category. 3. In a recording medium storing a character recognition program operating on a computer, the same individual performs a character pattern group recognized as having the same category as a result of the character recognition. Verifying by using an average value of the fluctuation range obtained in advance for a plurality of persons to determine whether the character pattern group is included in the fluctuation range generated in the case described, and a part of the character pattern group, Compared to other parts,
When it is determined that the characters are not in the same category described by the same individual, the method includes a step of changing a category of the part or the other part,
A recording medium that stores a character recognition program that runs on a computer. 4. An individual standard deviation of a feature amount used for character recognition in the initial recognition unit is obtained in advance for a plurality of persons, and if the same individual is described in the individual fluctuation range storage unit, the same category is determined. The character recognition device according to claim 2, wherein an average value of the standard deviation within the individual is stored as a range in which the character pattern can be determined. 5. In a feature space used for character recognition in the initial recognition unit, a category distribution for each individual is obtained in advance for a plurality of persons, and a distance between distribution centers becomes minimum for each category in the category distribution for each individual. The distance to another category is calculated as the distance between distribution centers, and the individual variation range storage unit stores the distribution center as a variation range of a character pattern that can be determined to be in the same category when the same individual is described. 3. The character recognition device according to claim 2, wherein an average value of the distances is stored. 6. The same category determination section compares the variation range read from the individual variation range storage section with a distance between character patterns in a feature space used for character recognition in the initial recognition section. 6. The character recognition device according to claim 2, wherein the determination is performed. 7. The initial recognition section performs character recognition using a plurality of reference patterns per the same category, and the same category determination section reads the variation range read from the individual variation range storage section and the initial recognition section. 6. The character recognition apparatus according to claim 2, wherein the determination is made by comparing a distance between reference patterns associated with the character patterns in the feature space used for character recognition. 8. In a character recognition method for performing character recognition using a plurality of reference patterns per same category, a distance in a feature space between a plurality of reference patterns used in the same category is calculated, and a plurality of persons are identified in advance by the same individual. Calculates the average value of the variation range that occurs when characters in the same category are described, calculates the average value, compares the distance and the average value, and a part of the plurality of reference patterns is different from the plurality of reference patterns. When it is determined that the characters are not in the same category described by the same individual as compared to the part, when recognizing the character string described by the same individual, the part and the other part are included in the character string. A character recognition method, wherein the character is not simultaneously selected as a reference pattern. 9. A character recognition apparatus for performing recognition by inputting a character string pattern, comprising: an input unit for inputting an image of a character string;
Reading the image of the character string from the input unit, a character cutout unit that cuts out a character pattern one character at a time, and reads a character pattern cut out from the character cutout unit;
An initial recognition unit that performs character recognition using a plurality of reference patterns per the same category and stores a category in association with each character pattern; and a character recognition unit in the initial recognition unit between reference patterns in the same category in advance. Calculate the distance in the feature space used, and compare the distance with the variation range of the character pattern that can be judged to be the same category when the same individual is described. A personal variation range storage unit that stores a set of patterns for each category; and the character pattern and the category associated with the character pattern are read from the initial recognition unit, and the reference pattern of the reference pattern is read from the personal variation range storage unit. Read out the set and character pattern associated with the same category According to the set of reference patterns, the same category determination unit that determines whether the individual is also associated with the same category, in the same category determination unit, a part of the character pattern group, A character re-recognition unit that changes the category of the part or the other part when it is determined that the character is not in the same category described by the same individual as compared with the other part. Character recognition device. 10. An intra-individual standard deviation of a feature amount used for character recognition in the initial recognition unit is obtained for a plurality of persons in advance, and if the same individual is described in the individual fluctuation range storage unit, the same category is determined. 10. The character recognition device according to claim 9, wherein an average value of the standard deviation within the individual is used as a change range of the character pattern that can be determined. 11. In a feature space used for character recognition in the initial recognition unit, a category distribution for each individual is obtained in advance for a plurality of persons, and a distance to a category in which the distance between distribution centers is the smallest for each category is determined. Calculated as an inter-distance, and in the individual variation range storage unit, as a variation range of a character pattern that can be determined to be in the same category when the same individual is described,
The character recognition device according to claim 9, wherein an average value of the distance between the distribution centers is stored. 12. The reliability of the association is highest when the initial recognizing unit associates a category with the character pattern and outputs the reliability of the association. 12. The character recognition apparatus according to claim 2, wherein the category is not changed for a character pattern included, and the category is changed for a character pattern that is not included. 13. The method according to claim 13, wherein the initial recognition unit performs character recognition using a plurality of reference patterns per the same category, and when changing the category in the character re-recognition unit, the priority of the reference pattern is determined in advance. 12. The character recognition apparatus according to claim 2, wherein the category is not changed for a reference pattern including the highest reference pattern, and the category is changed for a non-reference pattern. . 14. The same category determination unit determines that a part of the character pattern group is not a character of the same category described by the same individual as compared with another part of the character pattern group. When changing the category in the character re-recognition unit, assuming that the input character string pattern represents an address, a restriction was placed on the correspondence of the character pattern to the category, and the other part was changed without changing the part In the case and the reverse case, the one that more satisfies the restriction is adopted.
The character recognition device according to 6, 9, 10 or 11.