JPH11120293A - Character recognition/correction system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To precisely detect a character string of low quality which is entered into various slips at irregular character intervals or by an irregular entering method. SOLUTION: A specific character or specific character string is extracted from an input character string 101 by performing a 1st matching process between the input character string 101 and a specific character standard pattern dictionary 107. Then a candidate word group which belongs to a specific category and has the possibility that the candidate is positioned in areas in the input character strings 101 preceding or following to each specific character or specific character string extracted from the input character string 101 is extracted from a specific character dictionary 110 and a knowledge dictionary 111 linked to it. Then a 2nd matching process using a standard pattern dictionary 113 is performed for respective areas in the input character string 101 by candidate words belonging to the candidate word group according to information regarding the respective candidate words to recognize the characters constituting the input character string 101.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character string which is usually written on various types of slips, which are written at irregular character intervals or irregular writing methods, and contact and separation occur between adjacent characters. The present invention relates to a technique for recognizing a low-quality character string that may occur.

[0002]

2. Description of the Related Art OCR for reading image data and converting it into character code data
(Optical character reading device) has been used for various tasks as its application field has expanded. Different forms are used for each job, and the character strings and the persons who fill in the forms are various.

[0003] In a conventional OCR form, a character frame in which a character entry frame is printed one character at a time is used. In particular, when a kanji is entered, a large character frame is used.
This is to make it easier for the OCR device to detect the characters to be entered one by one, and to encourage the writer to avoid the characters to be entered at the time of entry from touching adjacent characters.

[0004] In the case of such a form, for example, even if the address or name is entered only in two or three places, the number of characters to be entered becomes tens of characters, and as a result, a large-sized form is required, which is costly. In addition, the writer has to complicate the need to fill in the boxes one character at a time.

[0005] As the application field of OCR becomes widespread, Kanji character strings can be entered in small forms such as ordinary forms without being bound by character frames, and characters which can be recognized with practical recognition accuracy and which cannot be read are used. There is a need for a character recognition / correction technique that can efficiently correct the correction.

In the conventional typical character recognition method, the entered characters are referred to by referring to a file called a definition body in which the coordinates of the character frame in which the character string to be recognized is entered on the form are stored. Each character is detected and cut out. Then, a recognition process is executed on each of the cut-out characters, so that a candidate character group as a recognition result is output.

Recognition processing of the cut-out character is executed, for example, as follows. First, characters written by a large number of unspecified writers in accordance with a predetermined format are collected, and a characteristic amount depending on the recognition method is extracted from these characters, and the characters are extracted by a statistical method or the like (for example, a clustering method). A standard pattern is created. Then, a standard pattern dictionary is created from the standard patterns for each target character type.

The standard pattern is created as an average pattern obtained by averaging the collected character patterns, for example. More specifically, this average pattern is expressed by an average feature amount obtained by calculating the average of the feature amounts corresponding to the collected characters.

In the process of recognizing handwritten characters, a large character deformation is caused by a writer, so that a plurality of standard patterns are created for each character type. Usually, one standard pattern is called a template, and a dictionary created from a plurality of standard patterns for each character type is called a multiple template dictionary.

The character recognition processing is executed using the above-described standard pattern dictionary or plural template dictionary. Specifically, a feature amount is extracted from one character cut out from the input form, and between this feature amount and the feature amount of each template (standard pattern) constituting the standard pattern dictionary (or a plurality of template dictionaries), Similarity or distance (Euclidean distance, Mahalanobis distance, etc.) is calculated. Then, each character type category to which each template up to a predetermined rank (for example, the eighth rank) belongs in the order of larger similarity or smaller distance is output as a candidate character group.

If the recognized character is a character representing an address or a name, generally, knowledge processing using the address word and the name word is performed on the candidate character group. More specifically, first, a candidate character string group is output by combining candidate character groups for each entry position over the entire entry position.

Next, for each candidate character string constituting this candidate character string group, it is compared whether or not each word character string in the address dictionary or name dictionary to be subjected to knowledge processing exists in the candidate character string. Is done.

A score is given to the candidate character string according to the comparison result and, for example, the order of each candidate character constituting the candidate character string. After this process is performed on all candidate character strings, the candidate character string with the highest score is output as a knowledge processing result.

As a prior art relating to such knowledge processing, for example, Japanese Patent Laid-Open Publication No. Sho 61-107
No. 486 is known. Here, when a kanji character string such as an address written at a free pitch and a name is recognized as in a normal slip, it is generally common for adjacent characters to contact each other, Conversely, some kanji characters are written separately, such as hen and tsukuri.

For this reason, in the conventional character recognition method in which the input characters are detected, cut out, and recognized one by one, it is difficult to determine which range is the range of one character, and a recognition that can withstand practical use. It is difficult to achieve accuracy.

Further, if each character cannot be recognized correctly,
Sometimes it is not possible to judge how many characters are entered,
In conventional knowledge processing on the assumption that the number of characters constituting a word has been determined, there is a limit to improving recognition accuracy.

In addition, in particular, in the recognition processing of an address, a place name, or the like, if a word at a higher level (for example, Tokyo, Osaka, etc.) cannot be recognized by the knowledge processing, a word at a lower level is not processed at that stage. In general, it is not possible.
It is necessary to correct all character strings sequentially from the first character.

As a first prior art for recognizing the above-described free-pitch character string, there is known a technique disclosed in Japanese Patent Publication No. 8-23875 "Word reading system". I have. In this first prior art,
The candidate character string as a recognition result and the word dictionary are collated by DP matching or the like, a word having many matching characters is selected, and a non-matching portion is cut out again, and further recognition is performed on the cut out character string. Will be

As a second prior art for recognizing a character string having a free pitch, Japanese Patent Laid-Open Publication No. Sho 63 is an example.
No. 136291 is known. In the second conventional technique, recognition processing is executed using a standard pattern dictionary having partial patterns indicating partial portions of a character as a standard pattern as a standard pattern. Is generated, and a matching process between the generated word dictionary and the word dictionary is executed.

As a third prior art for recognizing a character string of a free pitch, Japanese Patent Laid-Open Publication No. Hei 8-
The thing disclosed in 171614 "character string reading device" is known. According to the third conventional technique, when a correct character is not included in a candidate character string and skipping occurs, or when a plurality of read candidates occur due to the presence of a character candidate that conflicts with the correct character, an expected character is determined. The existence of the column is verified. Several realization methods are disclosed as verification means in this case.

However, a recognition process for a low-quality character string that we frequently enter, that is, a contact between adjacent characters frequently occurs, the character width changes greatly for each character, and the character string is often crushed or blurred. In consideration of the above, the first to third conventional techniques have the following problems.

First, in the first prior art, it is uncertain which character of a candidate character string is to be preferentially treated, and all characters in the candidate character string are treated equally. In some cases, there is a problem that only an inappropriate word may be selected as a candidate.

Next, in the second prior art, there is a problem in processing for an area where adjacent characters are in contact with each other. Furthermore, the third
In the prior art, several methods are described as a method of realizing the verification means, but each method uses a combination of character candidates, and their verification performance is greatly affected by the result of extracting the first character. It has a problem of dependence.

An object of the present invention is to accurately recognize a low-quality character string by focusing on a specific character.

[0025]

According to the present invention, there is provided a character recognition / correction method for recognizing a character constituting an input character string entered in an entry field having a predetermined category, a character recognition device having a function equivalent thereto, or A computer-readable recording medium is assumed.

In the present invention, first, the input character string and the first
By performing the first matching process with the recognition dictionary (specific character standard pattern dictionary 107), a specific character or a specific character string is extracted from the input character string. More specifically, a standard pattern corresponding to a specific character or a specific character string is stored in the first recognition dictionary, and a first pattern between the pattern of the input character string and each standard pattern in the first recognition dictionary is stored. Is executed, a specific character or a specific character string is extracted from the input character string. The above-mentioned specific character or specific character string is, for example, one that frequently appears in a predetermined category or one that has high recognition accuracy.

Next, candidates which belong to a predetermined category (for example, an address character string) and may be located in each specific character extracted from the input character string or an area in the input character string before and after the specific character string A word group is extracted from the word dictionary for each category (specific character dictionary 110, knowledge dictionary 111).

Then, for each candidate word belonging to the extracted candidate word group, a second recognition is performed on each region in the input character string where the candidate word is located based on information on the candidate word. By executing the second matching process using the dictionary (the standard pattern dictionary 113), the characters constituting the input character string are recognized. More specifically, a standard pattern corresponding to a character or a character string related to a candidate word belonging to the candidate word group is stored in the second recognition dictionary, and for each candidate word belonging to the candidate word group, Based on the information about the word, a second matching process is performed between the pattern of each candidate word and each standard pattern in the second recognition dictionary for each region in the input character string where each candidate word is located. By executing, the characters constituting the input character string are recognized. In this case, as information on each candidate word, for example, information on the number of characters of each candidate word is used. Further, the second recognition dictionary may be configured to include the first recognition dictionary.

According to the configuration of the invention described above, a specific character or a specific character string in an input character string is first recognized first, and candidate words before and after the specific character are assumed based on the recognition result. The characters that make up the input character string that are entered at irregular intervals and in a manner that would normally be entered into various forms (forms) by re-recognizing the characters that make up the input character string using Can be recognized with high accuracy.

In the configuration of the invention described above, the recognition result of the characters constituting the input character string is displayed in parallel with the input character string, and a desired area on the displayed input character string is designated by the user and the desired area is displayed. The character or character string corresponding to the area is corrected, and based on the information on the correct character or correct character string given by the correction, the extraction processing of the candidate word group and the second matching processing are executed again, and the input character string Can be configured to be recognized again. In this case, in response to designation of a desired area on the displayed input character string, a plurality of candidate recognition results in the desired area can be displayed.

With such a character correction technique, it is possible to automatically correct other unrecognizable parts only by correcting a specific character or character string. Further, in the configuration of the invention described above, it is possible to configure so that a word having fluctuation in notation with respect to each candidate word is output as a new candidate word belonging to the candidate word group.

With such a notation fluctuation control technique, it is possible to flexibly cope with various writing methods.

[0033]

Embodiments of the present invention will be described below in detail with reference to the drawings. Configuration and Schematic Operation Figure 1 of the embodiment of the present invention is a configuration diagram of an embodiment of the present invention.

First, the character extracting unit 103 uses the entry field definition 104 defining information on the entry field position of the form, from the beginning of the input character string 101 written in the form read from the image memory 102. Cut out one character at a time.

Next, the feature extracting unit 105 extracts a feature amount from the extracted character. Subsequently, the matching unit 106 performs a matching process between the feature amount of the cut-out character and the feature amount of each specific character standard pattern in the specific character standard pattern dictionary 107, and determines a predetermined amount in descending order of matching degree. The character type category of each specific character to which each specific character standard pattern belongs up to the rank is set as a candidate specific character for the cut-out character.
08.

Character extraction unit 103, feature extraction unit 10
5 and the above-described series of specific character recognition processing by the matching unit 106, the character cutout unit 103
Is executed for each character cut out in order from the beginning. As a result, the input character string 101 is stored in the candidate character string buffer 108.
The candidate identification characters corresponding to each character are held in the arrangement order corresponding to the arrangement order of the characters extracted from.

The candidate word search unit 109 extracts all adjacent two sets of specific characters (specific character sets) from the candidate specific character strings obtained in the candidate character string buffer 108, and A search is performed to determine whether or not the character set is registered in the specific character dictionary 110.

When one specific character set is registered in the specific character dictionary 110, the candidate word search unit 109 constructs the specific character set from a record in the knowledge dictionary 111 linked to the registered record. A word group sandwiched between two specific characters is searched, and the searched word group is held in the candidate word buffer 112 as a candidate word group.

The candidate word search unit 109 extracts a candidate word group corresponding to each of the specific character sets extracted from the candidate character string buffer 108 and stores the group in the candidate word buffer 112.

After all, in the candidate word buffer 112, one or more candidate word groups are obtained for one specific character set.
Finally, a set of candidate word groups for a plurality of specific character sets is obtained.

Each candidate word belonging to the candidate word group obtained in the candidate word buffer 112 for one specific character set is
They are sequentially read out, and the following series of processing is executed for each of them.

First, the character cutout unit 103 uses the information of the candidate word output from the candidate word buffer 112 in the input character string 101 read from the image memory 102 to form a specific character set to which the candidate word belongs. The character string in the character string area sandwiched between two specific characters is cut out again.

The feature extracting unit 105 extracts a feature amount from the character string cut out again. Further, the matching unit 106
Executes a matching process between the feature amount of the character string cut out again and the feature amount of each standard pattern in the standard pattern dictionary 113 which is the second dictionary, and each of the feature amounts up to a predetermined order in descending order of matching degree. The category of the character string to which the standard pattern belongs is held in the candidate character string buffer 108 as a candidate recognition result group for the candidate word.

Character extraction unit 103, feature extraction unit 10
5 and the series of re-recognition processing by the matching unit 106 are executed for each candidate word belonging to the candidate word group obtained in the candidate word buffer 112 for the one specific character set, and a predetermined The candidate recognition result group up to the rank is obtained in the candidate character string buffer 108.

The matching unit 106 determines the most appropriate and reliable reliability among all the candidate recognition result groups up to a predetermined order obtained in the candidate character string buffer 108 for each candidate word belonging to the one specific character set. Of the candidate recognition result having the highest matching degree,
The recognition result is output to the knowledge processing unit 114 as a recognition result of a portion sandwiched between two specific characters constituting the specific character set of the set.

Character extraction unit 103, feature extraction unit 10
5, and the above series of re-recognition processing for each candidate word belonging to the candidate word group of one specific character set by the matching unit 106 is executed for each specific character set registered in the candidate word buffer 112. . As a result, a recognition result corresponding to each character region sandwiched between two specific characters constituting each specific character set is output to the knowledge processing unit 114.

The knowledge processing unit 114 performs the knowledge processing using the entry field definition 104 and the knowledge dictionary 111 on the recognition result corresponding to each character area sandwiched between two specific characters constituting each specific character set. , And determines the final recognition result of the entire character area composed of the above character areas, and outputs the result to the recognition result buffer 115.

In the above-described series of recognition processing, reject (unrecognizable) information is added to a portion of a character or a character string that does not satisfy the recognition conditions to the end. In this case, the recognition result obtained in the recognition result buffer 115 is displayed on the display unit 117 via the interface unit 116. The user operates the input unit 11 including a mouse, a keyboard, and the like to display the recognition result on the display unit 117.
8, the unrecognizable character / character string can be corrected.

The user simply specifies a specific correct character in the unrecognizable character / character string from the input unit 118, and information on the correct character is output from the interface unit 116 to the correct character buffer 119 and the area coordinate buffer 120. You.

The candidate word search unit 109 uses the information on the correct character obtained in the correct character buffer 119 as the information on the specific character, as the specific character dictionary 110 and the knowledge dictionary 1 described above.
By executing the candidate word search process using No. 11, unrecognizable characters can be correctly re-recognized.
The character cutout unit 103 stores the cutout position of the correct character designated by the user in the area coordinate buffer 120.
By extracting from, correct character segmentation can be performed.

As described above, in this embodiment, the input character string 101 such as an address, a name, and a product name entered in each entry field in a form is a character or a specific character having a high appearance frequency in each field. By focusing on characters / strings,
Using the word information held by the knowledge dictionary 111 and, in the case of a character string such as an address having a hierarchical structure, connection information for each character region, a candidate word of a character region sandwiched between the specific characters can be selected. it can. Further, in the present embodiment, the extraction of the character region sandwiched between the specific characters from the input character string 101 and the re-recognition processing on the character region are performed using the information of the candidate word, so that the adjacent character A character string written in a style that often causes contact and separation between
Recognition can be performed with high recognition accuracy. Detailed Operation of the Embodiment of the Present Invention FIGS. 2 to 4 are operation flowcharts showing overall control realized by the embodiment of the present invention having the configuration shown in FIG. < Specific Character Recognition Processing > First, the character cutout unit 103
Uses the entry field definition 104 defining information on the entry field position of the form, and
Characters are cut out one by one from the beginning in the input character string 101 written on the form, read out as binary image data from 02 (step 201 in FIG. 2).

FIG. 5 is a diagram showing an example of the data format of the entry field definition 104 used by the character cutout unit 103. For example, when fields 1 and 2 are arranged on a form and character strings entered in these two fields are recognized, the entry field definition 104 is determined as follows.

First, the top of the form is defined as the coordinate origin, the x-axis is defined in the horizontal direction, and the y-axis is defined in the vertical direction. For each of the fields 1 and 2, the coordinates of the upper left position of the field (field origin) 5A, and field size data comprising a field width in the x-axis direction and a field height in the y-axis direction are defined as shown in FIG. The unit of length is millimeters or inches.

Next, for each of the fields 1 and 2, a field type indicating what type of character string is written in each field is defined. These pieces of information are stored in a storage device (not shown) as the entry field definition 104 in the form of a table shown in FIG.

The character cutout unit 103 determines the character area for each field on the image data read from the image memory 102 by using the above-described entry field definition 104, and then determines the image data in the character area. , The character cutout control shown by the operation flowchart shown in FIG. 6 is executed.

Here, as shown in FIG. 8A, the field origin coordinates of the target area extracted from the entry field definition 104 are (x ₀ , y ₀ ), and the field width in the x-axis direction is dx, y. The field height in the axial direction is dy.

First, the character cutout unit 103 accumulates the number of black pixels for each scanning line in the x-axis direction, thereby forming a horizontal histogram indicating the appearance frequency of black pixels in the x-axis direction for each y-coordinate position. , Calculated as shown in FIG. 8B (step 601 in FIG. 6).

Next, as shown in FIG. 8B, the character cutout unit 103 scans the horizontal histogram from above and below the horizontal histogram, and first finds the positions α and β that exceed the frequency value C. Then, the value α-β calculated from them is set as the character string height h in the target area (step 602).

Next, the character cutout unit 103 accumulates the number of black pixels for each scanning line in the y-axis direction,
A vertical histogram indicating the appearance frequency of black pixels in the y-axis direction at each x-coordinate position is calculated as shown in FIG. 8C (step 603 in FIG. 6).

Subsequently, as shown in FIG. 8C, the character cutout unit 103 scans the vertical histogram from the left, and changes the frequency value from the threshold d or less to the threshold d or more. Points x ₁ , x ₃ , x ₅ ,... (X _2n-1 : n
= 1, 2,...) As clipping candidate positions,
Further, points x ₂ , x ₄ , x ₆ ,... (X _2m : m = 1,
2,...) Are also calculated as clipping candidate positions (step 604).

Next, the character cutout unit 103 calculates an area [x _2m , x _2n-1 ] that satisfies the following conditional expression, and sets it as a character cutout result (step 605).

[0062]

## EQU1 ## ht ₁ ≤x _2m -x _2n-1 ≤h + t ₂ (m = 1,2,3,...), (N = 1,2,3,...)
Here, h is the character string height calculated in step 602 described above, and t ₁ and t ₂ are parameters determined by the learning sample of the input character string 101. In the example of FIG. 8 (c), the following three areas are calculated as a character cutout result. [X ₁ , x ₂ ] [x ₃ , x ₄ ] [x ₅ , x ₈ ] As a result of the processing in step 605, the character cutout unit 103 determines whether or not an area satisfying the following conditional expression remains (step 606).

[0063]

[Number 2] _{x 2l -x 2l-1> h} + t 2 (l = 1,2,3, ···) If the determination in step 606 is NO, the character segmentation unit 1
03 ends the control processing.

If the determination in step 606 is YES, the character cutout unit 103 determines that the frequency value of the vertical histogram calculated in step 603 is larger than the threshold value d in the area [x _{21 -1} , x ₂₁ ]. Less than or equal to a predetermined value, and
A value k that satisfies the following conditional expression is calculated.

[0065]

## _EQU3 ## h ≒ (x _2l −x _2l−1 ) / k As a result, each position obtained by dividing the area [x _2l−1 , x _2l ] into k is calculated as a character cutout position (step 60).
7). In the example of FIG. 8 (d), l = 1 , k = 2 , and the range [x _1, x _2] 2 divided position x 'is calculated as a character cut-out position.

After that, the character cutout unit 103 ends the control processing. The operation flowchart of FIG. 6 described above corresponds to the character cutout processing executed by the character cutout unit 103 for a field for which the number of characters has not been given in advance.

On the other hand, as in the case where re-recognition processing is executed based on information on candidate words read from the candidate word buffer 112, the character extracting unit 103 In some cases, character cutout processing may be performed on a field in which the number of characters in the field is given in advance.

In this case, the character extracting unit 103
The processing in step 701 in FIG. 7 is executed instead of the processing group in steps 605 to 607 in FIG. That is, the x coordinate of the left end of the area to be extracted is x _s , and the x coordinate of the right end is x
When the coordinates are given as x _t and the number of characters in the area is given as n, the character cutout unit 103 executes step 60 in FIG.
The position where the frequency value of the vertical histogram calculated in step 3 is equal to or less than a predetermined value and which has an adjacent interval close to the value _Xn satisfying the following conditional expression is calculated as a character cutout position.

[0069]

## EQU4 ## Specifically, (x _t −x _s ) / n = X _n Specifically, two adjacent character cutout positions are defined as x _i ,
x _{i + 1} (i = 1, 2,..., x _s ≦ x _i , x _{i + 1} ≦ x
when a _t), the character segmentation unit 103, character segmentation positions x _i satisfying the following condition _{(x i ≠ x s, x} t)
Is calculated.

[0070]

Equation 5] where _{X n -t 5 ≦ x i +} 1 -x i ≦ X n + t 6, t 5 and t ₆ are parameters determined by the learning sample of the input string 101.

After the character extraction processing by the character extraction unit 103 described above, the feature extraction unit 105 extracts a feature vector, which is a feature amount for recognition, from the extracted one character (step 202 in FIG. 2). ).

Specifically, the feature extracting unit 105 extracts a feature vector by, for example, the following series of processes. That is, first, the feature extraction unit 105 extracts a character contour pixel from the image data of the cut out character.

Next, the feature extracting unit 105 divides the cut-out region into a plurality of divided regions. Further, the feature extracting unit 105 extracts a directional component (for example, four directional components of a vertical direction, a horizontal direction, a leftward diagonal direction, and a rightward diagonal direction) for each contour pixel in the divided region for each divided region. By calculating the directional components of all the contour pixels in the divided region, a total value for each directional component in the divided region is calculated, and a partial feature vector having these as element values corresponding to each directional component is calculated.

Finally, the feature extracting unit 105 extracts a feature vector by integrating the elements of the partial feature vectors of all the divided areas. After the feature extraction unit 105 extracts the feature vector of the cut-out character as described above, the matching unit 106 compares the feature vector of the cut-out character with each specific character standard pattern in the specific character standard pattern dictionary 107. (Step 203 in FIG. 2), and the character type category of each specific character to which each specific character standard pattern up to a predetermined order belongs in the descending order of the matching degree is extracted from the extracted character. Are output to the candidate character string buffer 108 as a candidate specifying character group for (step 204 in FIG. 2).

More specifically, the matching unit 106
For example, a distance (Euclidean distance, Mahalanobis distance, etc.) is calculated between the feature vector of the extracted character and the feature vector of each specific character standard pattern in the specific character standard pattern dictionary 107. Then, the matching unit 106 determines the character type category of each specific character to which the specific character standard pattern belongs to the predetermined order (n-th) in ascending order of the distance as a candidate specific character group for the cut-out specific character. Output to the column buffer 108.

[0076] The distance 1 of the specific character standard pattern if greater than a predetermined threshold value T _1, at its cut-out character rejected (unrecognizable) information is added. Here, the specific character standard pattern dictionary 107 will be described with a specific example.

Now, consider a case where the input character string 101 is an address character string. In the present embodiment, initially, for example, in an address character string, the frequency of appearance indicating the break of the hierarchical structure is high, such as “city”, “road”, “fu”, “prefecture”, “city”, “ward”, “county”.
Only one or two specific characters such as "town", "village", "character", and "large" need be recognized. Further, in the address character string, specific characters such as “East”, “West”, “South”, and “North” also have a high appearance frequency.

For this reason, in this embodiment, in order to improve the recognition accuracy of these specific characters, a specific character standard pattern dictionary 107 having only a standard pattern of these specific characters and having a small dictionary capacity is used.

The specific character standard pattern dictionary 10
7 is prepared separately from the standard pattern dictionary 113, so that the recognition processing speed can be reduced and the recognition accuracy can be increased.

The specific character standard pattern dictionary 107 is configured as the same dictionary as the standard pattern dictionary 113. In order to improve the recognition accuracy of specific characters, many templates (standard patterns) are stored for each specific character. May be configured.

On the other hand, when the input character string 101 is a name character string, there is no character that serves as a delimiter like an address character string, but a character type with a high appearance frequency exists. For example, the characters used for surnames are the top 50 in appearance frequency.
Since the character types up to the 0th place cover about 82%,
The specific character standard pattern dictionary 107 can be configured with the upper N characters.

Alternatively, the configuration may be such that only the N character types are selectively used from the standard pattern dictionary 113 for recognition of a specific character. The specific character dictionary 110 is configured to correspond to the specific character type described above.

Instead of selecting a character type based on the frequency of appearance, characters that are easy to recognize may be statistically determined from a large number of actual data, and the determined character type may be selected. .

Character extraction section 103, feature extraction section 10
5 and the above-described series of specific character recognition processing by the matching unit 106, the character cutout unit 103
Is executed for each character cut out in order from the top (repetition of steps 205 → 202 in FIG. 2). As a result, the candidate character string buffer 108 holds candidate specific character groups for each character in the arrangement order corresponding to the arrangement order of the characters cut out from the input character string 101. < Search for candidate words in specific inter-character area and recognize in that area
Knowledge processing > candidate word search unit 109 extracts all sets of two adjacent specific characters (specific character sets) from the set of candidate specific character groups obtained in candidate character string buffer 108, Is a specific character set of the specific character dictionary 110
Search whether it is registered in. Then, the candidate word search unit 109 determines that one specific character set is the specific character dictionary 11
0, a group of words sandwiched between two specific characters constituting the specific character set is searched from a record in the knowledge dictionary 111 linked to the registered record, and the searched word group is selected as a candidate. It is stored in the candidate word buffer 112 as a word group (step 20 in FIG. 3).
6).

Now, consider a case where the input character string 101 is an address character string. It should be noted that the name character string, the product name character string, and the like other than the address character string do not have a hierarchical structure, and thus can be realized in the same manner as the case of the address character string except for the part related to the hierarchical structure.

The structure of the knowledge dictionary 111, which is an address dictionary, is as follows.
For example, as shown in FIG. 10, according to the hierarchical structure of the address, the words are divided into level 1: prefecture, level 2: city / county, level 3: town, and so on. Is stored.

On the other hand, as shown in FIG. 11, the specific character dictionary 110 stores, for each record corresponding to a specific character set including two specific characters “character 1” and “character 2”,
A data set consisting of pointer information and information on the number of data starting from the pointer for indicating a set of records on the knowledge dictionary 111 in which a word group sandwiched between two specific characters constituting the specific character set is stored is shown. Is stored. As this data set, a plurality of sets can be designated as shown in FIG. 11, and the record for each specific character set in the specific character dictionary 110 has the pointer information and the number of data as shown in FIG. Pointer number information N corresponding to the number of data sets of information is also stored.

In the example of FIG. 12, a record in the specific character dictionary 110 corresponding to a specific character set composed of two specific characters of a blank character and “ken” is a level in the knowledge dictionary 111 shown in FIG. N ₁ records starting with the word “Aomori” in one area, n ₂ records also starting with the word “Kanagawa” in the level 1 area (pointer information and data number information), and a pointer The number N = 2 is registered.

In the example of FIG. 13, the specific character dictionary 110
In the record corresponding to the specific character set consisting of two specific characters “To” and “ku”, n ₃ starting from the word “Chiyoda” in the level 2 area in the knowledge dictionary 111 shown in FIG. Records and the number of pointers N = 1 are registered.

The address usually ends in the form of "... chome ... address ... how". However, such specific characters "chome", "address", "number", "ho" and "go" are used. In many cases, numbers, not words, are written in the interposed area. In such a case, as shown in FIG. 14, a record in the specific character dictionary 110 corresponding to the specific character set made up of the specific characters described above includes data including pointer information and data number information as described above. "(Number) *
A symbol such as "n" is set. Candidate word search unit 1
09 indicates that when a record in which the above-described symbols are set is searched from the specific character dictionary 110, numbers are continuously written in the area between the specific characters as described above. Then, a detection result indicating the detection is written to the candidate word buffer 112.

Further, for example, the configuration of the specific character dictionary 110 and the knowledge dictionary 111 as shown in FIG. 15 is also possible. That is, in the example of FIG.
Records corresponding to a specific character set consisting of a blank character and two specific characters “kawa” include pointer information indicating the four-character word “Kanagawa” in the knowledge dictionary 111 and the number of data = 1, and the knowledge dictionary 111 The pointer information indicating the two-letter word "Kana" and the number of data = 1 are set.

A record in the specific character dictionary 110 corresponding to a specific character set composed of two specific characters “kawa” and “medium” indicates the two-character word “Sakiichi” in the knowledge dictionary 111. The pointer information and the number of data = 1 are set.

Further, a record corresponding to a specific character set consisting of two specific characters “medium” and “medium” in the special character dictionary 110 includes a five-character word “Kamioda” in the knowledge dictionary 111. Is set and pointer information indicating the number of data = 1 is set.

As described above, it is also possible to store information corresponding to specific characters and words appearing frequently in an address character string in the specific character dictionary 110 and the knowledge dictionary 111. Next, as shown in FIG.
When the word “Marunouchi” in the knowledge dictionary 111 is linked to a record corresponding to a specific character set including two specific characters of “ku” and a specific character indicating the end of the address in 0, As a fluctuation, there is a possibility that a character string "Marunouchi" instead of "Marunouchi" is entered. In such a case, it is useless to store the words corresponding to the fluctuations of all the notations in the knowledge dictionary 111.

Therefore, in the present embodiment, when searching for a word in the knowledge dictionary 111 linked from the specific character dictionary 110, a control operation for coping with the fluctuation of the notation shown in the operation flowchart of FIG. 9 is executed. .

First, the candidate word search unit 109 applies a specific character dictionary 110 and a knowledge dictionary 11 to one specific character set.
1 is searched according to the rules described so far, and the searched word group is written to the candidate word buffer 112 as a candidate word group corresponding to the specific character set currently being processed (step 901 in FIG. 9). This step 901 is a part of step 206 in FIG.

Next, as a part of step 206 in FIG. 3, the candidate word search unit 109 performs a search on each candidate word group obtained in the candidate word buffer 112 for one specific character set. Nine steps 902 to 910 are repeatedly executed.

That is, when there are hiragana in the characters constituting the detected word, the candidate word search unit 109 changes the hiragana to katakana, and converts the resulting word into the specific character set currently being processed. It is written into the candidate word buffer 112 as another corresponding candidate word (steps 902 → 903 in FIG. 9).

Next, if there are katakana in the characters constituting the detected word, the candidate word search unit 109 changes the katakana to hiragana, and replaces the resulting word with the specific character set currently being processed. Is written into the candidate word buffer 112 as another candidate word corresponding to (step 904 → 905 in FIG. 9).

Next, the candidate word search unit 109 changes the Kanji numerals to Arabic numerals when the characters constituting the detected word include Kanji numerals, and replaces the resulting word with the currently processed word. It is written into the candidate word buffer 112 as another candidate word corresponding to the specific character set (steps 906 → 907 in FIG. 9).

Next, the candidate word search unit 109 determines whether or not the characters constituting the detected word include Arabic numerals.
The Arabic numerals are changed to Chinese numerals, and the resulting word is written into the candidate word buffer 112 as another candidate word corresponding to the specific character set currently being processed (steps 908 → 909 in FIG. 9).

Finally, the candidate word search unit 109 replaces the characters constituting the detected word with optional characters (for example, “Mizonokuchi”).
Is abbreviated as "Mizoguchi"), the character string obtained by omitting the optional character is used as another candidate word corresponding to the specific character set currently being processed.
Write to candidate word buffer 112 (step 9 in FIG. 9)
08 → 909).

If there is a candidate word group for which the control process for the fluctuation of the notation has not been executed in the candidate word buffer 112 for one specific character set, the candidate word searching unit 109 shown in FIG. The series of processing shown in steps 902 to 910 is repeatedly executed (step 91 in FIG. 9).
1 → 902-910 → 911).

As described above, the control for the fluctuation of the notation is realized for the candidate word group obtained in the candidate word buffer 112 for one specific character set. As described above, a candidate word group is obtained in the candidate word buffer 112 for one specific character set selected from the candidate character string buffer 108.

Now, for example, the input character string 1 shown in FIG.
01 is entered, steps 201 to 2 in FIG.
In the recognition processing of the specific character 05, the area 1701 is recognized as the specific character “To” and the area 1702 is recognized as the specific character “ku”.

In response to the recognition result, the candidate word search unit 1
In step 206 of FIG. 3 described above, a record of a specific character set consisting of a blank character and a specific character “tsu” is detected in the specific character dictionary 110 in step 206 in FIG. , One word “Tokyo” sandwiched between two specific characters constituting the specific character set, and the searched word corresponds to a specific character set consisting of a blank character and a specific character “To” The candidate word group is held in the candidate word buffer 112. In this case, the number of candidate words for the specific character set is one, and as shown in FIG. 18, the number of characters of the candidate word “Tokyo” is two.

Also, the candidate word search unit 109 determines the specific characters “U” and “K” in the specific character dictionary 110 in step 206 in FIG. 3 which is executed for the second time after the determination in step 211 in FIG. From the entry in the knowledge dictionary 111 shown in FIG. 10 linked to the registered record, the 23 words “Chiyoda” sandwiched between two specific characters constituting the specific character set are detected. "", "Center", "port",..., And the searched word group is held in the candidate word buffer 112 as a candidate word group corresponding to the specific character set. In this case, the number of candidate word groups for the specific character set is 2
As shown in FIG. 19, the number of characters of each candidate word is one of three, two, or one.

Thus, the candidate character string buffer 10
After a candidate word group is obtained in the candidate word buffer 112 for one set of specific character sets selected from No. 8, the character cutout unit 103, the feature extraction unit 105, And matching unit 106
Performs a series of re-recognition processes in steps 207 to 211 of FIG. 3 to extract a group of candidate recognition results up to a predetermined order for each candidate word.

First, in the input character string 101 read from the image memory 102, the character cutout unit 103 uses the information on the candidate word output from the candidate word buffer 112 to form a specific character set to which the candidate word belongs. The character string in the character string area sandwiched between two specific characters is cut out again (step 207 in FIG. 3).

In this case, the number of characters of the candidate word is, for example, as shown in FIG.
In the case of two characters such as “Tokyo” shown in FIG. 8 or “center” shown in FIG.
In accordance with the operation flowchart shown in steps 601 to 604 in FIG. 6 and step 701 in FIG. Is determined.

If the number of characters of the candidate word is three, for example, “Chiyoda” shown in FIG. 19, the character cutout unit 103 sets the area for character cutout to three.
The character is divided (n = 3 in Equation 3 described above), and the cutout position of each character is determined.

Further, when the number of characters of the candidate word is one, for example, "port" shown in FIG. 19, the character extracting unit 103 assumes that only one character exists in the area to be extracted. (N in Equation 3 described above)
= 1).

Next, the feature extracting unit 105 extracts a feature vector for each character from the character string cut out again as described above (step 208 in FIG. 3). Furthermore,
The matching unit 106 determines, for each character, the feature vector of the character and the standard pattern dictionary 11 that is the second dictionary.
3 (step 209 in FIG. 3), and the respective character type categories to which the standard patterns up to a predetermined order belong in the descending order of the matching degree are candidates for the characters. The character group is output to the candidate character string buffer 108 (step 210 in FIG. 3).

More specifically, the matching unit 106
For example, a distance (Euclidean distance, Mahalanobis distance, etc.) is calculated between the character feature vector and the feature vector of each standard pattern in the standard pattern dictionary 113. Then, the matching unit 106 outputs, to the candidate character string buffer 108, each character type category to which each of the standard patterns up to the predetermined order (nth order) belongs in the order of smaller distance as a candidate character group for the above-described characters.

For each of the characters constituting the character string cut out again by the character cutout unit 103, a candidate character group up to a predetermined order is obtained in the candidate character string buffer 108 in ascending order of distance as described above. Step 207 is performed for other candidate words belonging to the candidate word group obtained in the candidate word buffer 112 for one specific character set.
A series of processes from to 210 are repeatedly executed.

For all the candidate words belonging to the candidate word group obtained in the candidate word buffer 112 for one particular character set, the candidate character group up to a predetermined order is obtained in the candidate character string buffer 108 for each character constituting each word. When it is
The matching unit 106 generates, for each of the candidate words, a candidate character string group by combining all of the candidate character groups up to a predetermined rank for each of the characters constituting each of the candidate words. The average distance is calculated by the equation (step 212 in FIG. 3).

[0117]

(D ₁ + D ₂ +... + D _m ) / m where m is the number of characters of the target candidate word, and D _i (1 ≦ 1)
i ≦ m) indicates the distance of the candidate character selected to form the target candidate character string at the i-th character in the target candidate word.

Then, the matching unit 106 determines, from the candidate character string group generated corresponding to all the candidate words for one specific character set, the average distance corresponding to each of the candidate character strings constituting the group. A predetermined number (P) of candidate character strings are selected in ascending order, and these are set as 2
The recognition result is output to the knowledge processing unit 114 as a recognition result of a character area sandwiched between two specific characters.

When the recognition result of the character area sandwiched between the two specific characters constituting one specific character set is obtained in this way, the process returns to step 206 from step 213 in FIG.

Then, steps 201 to 2 in FIG.
From the set of candidate specific character groups obtained in the candidate character string buffer 108 by the specific character recognition process of step 05, another specific character set consisting of any two adjacent specific characters is extracted again. The series of control processing of steps 206 to 212 in FIG. 3 is executed again for the specific character set, whereby the recognition result of the character area sandwiched between two specific characters constituting the specific character set is calculated. Is repeated for each specific character set (steps 213 → 206 to 212 → 213 in FIG. 3 are repeated).

The knowledge processing unit 114 performs the knowledge processing using the entry field definition 104 and the knowledge dictionary 111 on the recognition result corresponding to each character area sandwiched between two specific characters constituting each specific character set. The final recognition result of the entire character area composed of the above character areas is determined, and is output to the recognition result buffer 115 (step 21 in FIG. 4).
4).

Steps 201 to 4 in FIG.
The final recognition result for each entry field is determined by repeatedly executing the series of control processes of step 214 for each entry field position of the form (step 215 in FIG. 4 → repetition of step 201 in FIG. 2).

In the above-described series of recognition processing, reject (unrecognizable) information is added to a portion of a character or a character string that does not satisfy the recognition conditions to the end. In this case, the recognition result obtained in the recognition result buffer 115 is displayed on the display unit 117 via the interface unit 116. The user operates the input unit 11 including a mouse, a keyboard, and the like to display the recognition result on the display unit 117.
8, the unrecognizable character / character string can be corrected.

The user simply specifies a specific correct character in the unrecognizable character / character string from the input unit 118, and information on the correct character is output from the interface unit 116 to the correct character buffer 119 and the area coordinate buffer 120. You.

In the example of FIG. 21, an image 2101 of the target character string is displayed on the display unit 117 in parallel with the recognition result 2102. When the user designates a specific area 2103 on the image 2101 using the mouse or the like as the input unit 118, the corresponding recognition result character 2104 is highlighted or highlighted. In response to this display, the user
When the correct character "To" is input from the keyboard 18 or the like, information on the correct character "To" is output from the interface unit 116 to the correct character buffer 119 and the area coordinate buffer 120. Naturally, the user
When an area corresponding to, for example, “Tokyo” on 101 is designated and the corresponding recognition result “Bunch length” is corrected to “Tokyo”, information on the correct character “Tokyo” is transmitted from the interface unit 116 to the correct character buffer 119 and The data is output to the area coordinate buffer 120.

The candidate word search unit 109 uses the information on the correct character obtained in the correct character buffer 119 as the information on the specific character, as the specific character dictionary 110 and the knowledge dictionary 1 described above.
By executing the candidate word search process using No. 11, unrecognizable characters can be correctly re-recognized.
The character cutout unit 103 stores the cutout position of the correct character designated by the user in the area coordinate buffer 120.
By extracting from, correct character segmentation can be performed.

In the example shown in FIG. 22, the display 117 displays
An image of the target character string is displayed in parallel with the recognition result 2202. The user can select a specific area 220 on the image.
When the user designates “1” from the mouse or the like serving as the input unit 118, the corresponding recognition result character 2203 is highlighted or highlighted, and a recognition result candidate 2204 is displayed at the designated portion. In response to this display, when the user selects the correct character “To” from the keyboard or the like serving as the input unit 118, information on the correct character “To” is displayed in the interface unit 11.
6 to correct character buffer 119 and area coordinate buffer 1
20. In this case, the recognition result candidates 2204 displayed in the instruction portion are displayed in the order of appearance frequency of the displayed characters, or in the case of having a hierarchical structure such as an address character string, in the order determined by the hierarchical structure, or simply in the character order. It can be configured to be displayed in code order.

As shown in FIG. 23 following the example of FIG. 22, similar correction processing is further performed on the designated position 2301 and the corresponding recognition result position 2302 to correctly recognize the character string 2303 again. It is possible to do.

Here, regarding the re-recognition processing for each character area sandwiched between two specific characters constituting each specific character set, in steps 207 to 212 in FIG. The re-recognition process is individually performed for each candidate word, and the recognition result for the candidate word is finally output.

In this case, the character type searched from the standard pattern dictionary 113 by the matching unit 106 is limited to the character type of the category to which the candidate word belongs, so that efficient re-recognition processing is realized.

On the other hand, for the entire character region sandwiched between two specific characters, the feature vector extraction and matching unit 10
6 may be configured to execute the matching process. In this case, the standard pattern dictionary 113 includes
A feature vector of a standard pattern in which each of the words “Kawasaki”, “Yokohama”, “Yokosuka”,. Then, the matching process is performed with the feature vector of each standard pattern in the standard pattern dictionary 113.

In this case, the word group searched from the standard pattern dictionary 113 by the matching unit 106 is limited to the word group of the category to which the candidate word belongs, so that efficient re-recognition processing is realized.

More specifically, for example, in recognition of an address character string, the matching unit 106
3. The efficient re-recognition processing is realized by limiting the word group searched from above to the word group constituting the hierarchical level to which the candidate word belongs.

For example, as shown in FIG. 20, in the re-recognition processing of an area sandwiched between two specific characters “prefecture” and “city”, the standard pattern dictionary 113 stores “Kawasaki”, “Yokohama”
It can be limited to only words that represent the city such as "Yokosuka".

For example, in recognition of an address character string,
When a higher-level recognition result is obtained, the word group searched by the matching unit 106 from the standard pattern dictionary 113 is limited to a word group belonging to the higher-level recognition result and constituting a lower level to which the candidate word belongs. As a result, more efficient re-recognition processing is realized.

For example, if the recognition result of the address character string at level 1 is "Aomori", the standard pattern at level 2 is
Instead of all the word groups that can appear between two specific characters “prefecture” and “city”, it is possible to limit to a word group representing a city belonging to “Aomori prefecture”.

Contrary to the above, for example, in the recognition of an address character string, when a lower-level recognition result is obtained, the word group searched from the standard pattern dictionary 113 by the matching unit 106 is replaced with the lower-level recognition. By being limited to a word group forming a higher level to which the result belongs and to which the candidate word belongs, an unrecognizable state at a higher level can be relieved. Recording medium on which a program for realizing the present embodiment is recorded
The present invention is configured as a computer-readable recording medium that, when used by a computer, causes the computer to perform the same functions as the functions realized by the above-described embodiments of the present invention. You can also.

In this case, as shown in FIG. 24, according to the embodiment of the present invention via a portable recording medium 2402 such as a floppy disk, CD-ROM disk, optical disk, removable hard disk or the like, or a network line 2403. Programs for realizing various functions are stored in a memory (RAM) in the main body 2404 of the computer 2401.
Or a hard disk) 2405 and executed.

[0139]

According to the character recognition technique of the present invention, a specific character or a specific character string in an input character string is first recognized first, and candidate words before and after the specific character are assumed based on the recognition result. By re-recognizing the characters that make up the input character string using the information on the candidate words, irregular intervals, such as those normally entered in various forms (slips),
Characters constituting the input character string entered by the entry method can be recognized with high accuracy.

According to the character correcting technique of the present invention, it is possible to automatically correct other unrecognizable portions only by correcting a specific character or character string. According to the notation fluctuation control technique of the present invention, it is possible to flexibly cope with various writing methods.

[Brief description of the drawings]

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

FIG. 2 is an overall control operation flowchart (part 1) of the embodiment of the present invention.

FIG. 3 is an overall control operation flowchart (part 2) of the embodiment of the present invention.

FIG. 4 is an overall control operation flowchart (3) according to the embodiment of the present invention.

FIG. 5 is a diagram showing an example of a data format of an entry field definition.

FIG. 6 is a control operation flowchart (part 1) of a character cutout unit.

FIG. 7 is a control operation flowchart (part 2) of the character cutout unit.

FIG. 8 is an explanatory diagram of a control operation of a character cutout unit.

FIG. 9 is a flowchart of a control operation regarding the fluctuation of the notation.

FIG. 10 is a structural diagram of a knowledge dictionary (address).

FIG. 11 is a structural diagram of a specific character dictionary.

FIG. 12 is a diagram showing a structural example (part 1) of a specific character dictionary 110.

FIG. 13 is a diagram showing a structural example (part 2) of the specific character dictionary 110.

FIG. 14 is a diagram showing a structural example (part 3) of the specific character dictionary 110.

FIG. 15 is a diagram showing a structural example (part 4) of the specific character dictionary 110.

FIG. 16 is an explanatory diagram of the control operation of the fluctuation of the notation.

FIG. 17 is an explanatory diagram (part 1) of the operation of the candidate word search unit.

FIG. 18 is a diagram (part 2) illustrating the operation of the candidate word search unit.

FIG. 19 is a diagram (part 3) illustrating the operation of the candidate word search unit.

FIG. 20 is an explanatory diagram of a character string detection / recognition operation using a standard pattern dictionary.

FIG. 21 is an explanatory diagram (part 1) of the operation of the input unit and the display unit.

FIG. 22 is a diagram (part 2) illustrating the operation of the input unit and the display unit.

FIG. 23 is an explanatory diagram (part 3) of the operation of the input unit and the display unit.

FIG. 24 is an explanatory diagram of a recording medium on which a program for realizing the present embodiment is recorded.

[Explanation of symbols]

 Reference Signs List 101 Input character string 102 Image memory 103 Character cutout unit 104 Entry field definition 105 Feature extraction unit 106 Matching unit 107 Specific character standard pattern dictionary 108 Candidate character string buffer 109 Candidate word search unit 110 Specific character dictionary 111 Knowledge dictionary 112 Candidate word buffer 113 Standard pattern dictionary 114 Knowledge processing unit 115 Recognition result buffer 116 Interface unit 117 Display unit 118 Input unit 119 Correct character buffer 120 Area coordinate buffer

Claims (12)

[Claims] 1. A character recognition method for recognizing characters constituting an input character string entered in an entry field having a predetermined category, wherein a first matching is performed between the input character string and a first recognition dictionary. By executing the process, a specific character or a specific character string is extracted from the input character string, belongs to the predetermined category, and the specific character or the specific character string before and after each specific character or the specific character string extracted from the input character string is extracted. A candidate word group that may be located in an area in the input character string is extracted from the category-based word dictionary, and for each candidate word belonging to the extracted candidate word group, the candidate word is determined based on information about the candidate word. A character constituting the input character string is recognized by performing a second matching process on each area in the input character string where each candidate word is located using a second recognition dictionary. A character recognition method characterized by including a process. 2. The method according to claim 1, wherein a standard pattern corresponding to the specific character or the specific character string is stored in the first recognition dictionary, and the pattern of the input character string and the first pattern are stored in the first recognition dictionary. Extracting the specific character or the specific character string from the input character string by executing the first matching process with each standard pattern in the recognition dictionary of the character string. Recognition method. 3. The method according to claim 1, wherein a standard corresponding to a character or a character string related to a candidate word belonging to the candidate word group is stored in the second recognition dictionary. A pattern is stored, and for each candidate word belonging to the candidate word group, a pattern of the candidate word is determined for each region in the input character string where the candidate word is located based on information on the candidate word. Recognizing characters constituting the input character string by executing the second matching process between the input character string and the standard pattern in the second recognition dictionary. Method. 4. The method according to claim 1, further comprising: using information on the number of characters of each candidate word as the information on each of the candidate words. Character recognition method. 5. The method according to claim 1, wherein the second recognition dictionary including the first recognition dictionary is used as the first recognition dictionary. A character recognition method comprising: 6. The method according to claim 1, wherein a first matching process is performed between the input character string and the first recognition dictionary. Extracting a specific character or a specific character string that frequently appears in the predetermined category from the input character string. 7. The method according to claim 1, wherein a first matching process is performed between the input character string and the first recognition dictionary, thereby performing the first matching process. A character recognition method characterized by including a step of extracting a specific character or a specific character string having high recognition accuracy from an input character string. 8. A character correction method using the character recognition method according to claim 1, wherein recognition results of characters constituting the input character string are arranged in parallel with the input character string. Specifying a desired area on the displayed input character string, and correcting a character or a character string corresponding to the desired area, based on information on the correct character or the correct character string given by the correction Re-recognizing characters constituting the input character string by executing the candidate word group extraction process and the second matching process again. 9. The method according to claim 8, further comprising: displaying a plurality of candidate recognition results in the desired area in response to designation of a desired area on the displayed input character string. A character correction method, characterized in that: 10. The method according to claim 1, wherein a word having a notational fluctuation for each of the candidate words is output as a new candidate word belonging to the candidate word group. A character recognition method or a character correction method characterized by including the following. 11. A character recognition device for recognizing characters constituting an input character string entered in an entry field having a predetermined category, wherein a first matching is performed between the input character string and a first recognition dictionary. A specific character / specific character string extracting means for extracting a specific character or a specific character string from the input character string by executing a process; and each specific attribute belonging to the predetermined category and extracted from the input character string. Candidate word group extraction means for extracting a candidate word group that may be located in an area in the input character string before or after a character or a specific character string from a category-based word dictionary; For each candidate word, a second matching process is performed for each area in the input character string where the candidate word is located, using a second recognition dictionary, based on information about the candidate word. An input character string recognizing means for recognizing characters constituting the input character string by executing the input character string. 12. A recording medium storing a program read by the computer when used by a computer, wherein a first character recognition dictionary is provided between an input character string entered in an entry field having a predetermined category and a first recognition dictionary. A function of extracting a specific character or a specific character string from the input character string by executing the first matching process; and a specific character or a specific character belonging to the predetermined category and extracted from the input character string. A function of extracting a group of candidate words that may be located in an area in the input character string before and after a column from the word dictionary for each category; and for each candidate word belonging to the extracted candidate word group, A second matching process is performed on each area in the input character string where each candidate word is located based on information about the word using a second recognition dictionary. And a computer-readable recording medium storing a program for causing the computer to perform a function of recognizing characters constituting the input character string.