JPH09311909A - Character recognition method and device therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To speed up a Japanese optical character reader(OCR). SOLUTION: The grouping of characters is executed for each feature amount and the number of matching object characters is reduced on the respective stages of matching so that recognition can be accelerated. At a group preparing part 103, the grouping results of respective features are stored in a group dictionary 104. At a representative feature amount dictionary preparing part 105, the feature amounts on behalf of groups are stored in a representative feature amount dictionary 106. When a character pattern 107 is inputted, the feature amount is extracted by a feature amount extracting part 108. At a matching processing part 109, the representative feature amount dictionary 106 corresponding to the candidate character of a candidate character storage table 110 is read out and stored in an evaluation value table 111 as the evaluation value of matching with all the characters belonging to the character group by performing matching processing and at a candidate character narrowing part 112, the candidate characters are narrowed down. By repeating this processing, the final candidate character is outputted.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition technique used in OCR or the like, and more particularly to a character recognition method and apparatus capable of improving recognition speed in Japanese OCR.

[0002]

2. Description of the Related Art Since there are various types of characters in Japanese, it is difficult for a Japanese OCR (optical character reader) to achieve both high speed character recognition and highly accurate character recognition. . If the character recognition method with a small processing amount is used for recognition, high speed can be realized, but the recognition rate is reduced. On the contrary, in the method of performing the complicated evaluation in order to improve the recognition accuracy, the processing amount becomes large and the recognition speed is lowered.

In order to realize high-speed character recognition while realizing a high recognition rate, conventionally, a plurality of features are extracted from a character pattern, and the features that are easy to process and the features having a small number of dimensions are selected in descending order. Hierarchical classification such as classification, middle classification, and detailed classification is used to speed up the process. Such a method is disclosed in, for example, Japanese Patent Laid-Open No. 7-
No. 306916, "Recognition method using neural network" and the like.

[0004]

According to the conventional hierarchical narrowing-down method, the features used for large classification have a small number of dimensions of the features and low classification ability. There is also a method of realizing high narrowing down by using many features with a small number of dimensions for large classification,
It takes a lot of processing time to extract the features, so it is difficult to improve the overall recognition speed. Generally, in the hierarchical narrowing method, since the features with a large number of dimensions are used in the later stages of middle classification and detailed classification, the processing time increases if the large classification is not sufficiently narrowed down. There is a problem of inviting. In particular, when recognizing a character having many character types such as kanji and kana characters and having many similar characters, the narrowing down ratio by the conventional large classification method cannot be said to be sufficient.

In view of the above-mentioned problems in the prior art, it is an object of the present invention to provide a character recognition method and device capable of realizing high-speed character recognition while realizing a high recognition rate.

[0006]

In order to achieve the above object, the present invention integrates (groups) characters having similar characteristics in advance at each stage of the above hierarchical classification,
By adding a method for reducing the number of matching target characters at each stage, the recognition speed is improved.

That is, the invention according to claim 1 is a character recognition method for recognizing a document image and converting it into code information, by extracting a plurality of features for character recognition from a standard character pattern of a character and extracting the extracted features. A step of storing the feature quantity in the feature quantity dictionary, and creating a plurality of character groups by grouping characters having similar features for each of the plurality of characteristics stored in the feature quantity dictionary. At the same time, a step of defining a representative character that is a representative of each character group, a step of extracting a plurality of features for character recognition from a character pattern of the recognition target extracted from the document image, and a character of the recognition target for each feature The matching process is performed by matching the feature amount obtained from the pattern with the representative feature amount that is the feature amount of the representative character of each character group created for that feature. By using the evaluation value acquisition step to obtain the evaluation value of, and the obtained evaluation value as the evaluation value of the matching between the character pattern to be recognized and all the characters belonging to each character group, candidate characters for narrowing down the candidate characters And a narrowing step.

According to a second aspect of the present invention, in a character recognition method for recognizing a document image and converting it into code information, a plurality of characteristics for character recognition are extracted from a standard character pattern of characters, and the extracted characteristics are extracted. Storing a quantity in a feature amount dictionary, and for a plurality of features stored in the feature amount dictionary, creating a plurality of character groups by grouping characters having similar features for each feature, A step of defining a representative character that is a representative of each character group, a step of creating a group dictionary storing information indicating which character group each character belongs to for each feature, and a feature related to each feature A step of creating a representative feature amount dictionary storing representative feature amounts which are the feature amounts of the representative characters of each character group, and the character pattern of the recognition target extracted from the document image. A step of extracting a plurality of features for character recognition from the character recognition process, and for each feature, perform a matching process between the feature amount obtained from the character pattern to be recognized and the representative feature amount of each character group related to that feature, An evaluation value acquisition step of obtaining an evaluation value, a step of storing the obtained evaluation value in the evaluation value table of each character as an evaluation value of matching between the character pattern to be recognized and all the characters belonging to each character group, And a candidate character narrowing step of narrowing down candidate characters by using the evaluation value stored in the evaluation value table for each character.

The invention according to claim 3 is the invention according to claim 1 or 2.
In the character recognition method described in [3], the feature quantity relating to a plurality of features for character recognition is set to the nth feature quantity (n = 1, 2, 3, ...
In the evaluation value acquisition step and the candidate character narrowing step, after performing matching and narrowing down on the nth feature amount, when performing matching and narrowing down on the next n + 1th feature amount, the narrowing down regarding the nth feature amount is performed. For the performed candidate characters, all character groups of the (n + 1) th feature amount including those candidate characters are obtained, and in the evaluation value acquisition step for the next (n + 1) th feature amount,
The obtained group is matched with the (n + 1) th feature amount.

[0010] The invention according to claim 4 is the invention according to claim 1 or 2.
In the character recognition method described in (1), the grouping of the characters is included within a predetermined threshold distance from all the characters stored in the feature dictionary, assuming that the character is a representative character. A temporary group is created from the characters that are written, the number of characters that belong to that temporary group is calculated, and the temporary group with the largest number of characters is determined as the regular group. In this case, the character belongs to a group having a shorter distance to the characteristic amount of the representative character.

According to a fifth aspect of the present invention, in a character recognition apparatus for recognizing a document image and converting it into code information, characteristic quantities relating to a plurality of character recognition features extracted from a standard character pattern of a character are stored. For a feature dictionary and a plurality of features stored in the feature dictionary, for each feature,
A group dictionary that stores information indicating which character group each character belongs to by grouping characters having similar characteristics, and a characteristic amount of a character that is a representative of the created character group for each characteristic. A representative feature quantity dictionary storing representative feature quantities, input means for inputting a character pattern to be recognized, means for extracting a plurality of features for character recognition from the character pattern to be recognized input by the input means, and Means for performing the matching process of the feature amount obtained from the character pattern of the recognition target and the representative feature amount of each character group for each feature, and obtaining the evaluation value of the matching, and the evaluation value as the character pattern of the recognition target. Means for storing in the evaluation value table of each character as an evaluation value for matching with all characters belonging to each character group, and the evaluation value table for each feature for each character A candidate character narrowing means for narrowing down candidate characters using the stored evaluation value; and means for outputting a recognition result which is a candidate character finally obtained by the candidate character narrowing means. To do.

[0012]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 shows a block diagram of a character recognition apparatus according to the present invention. The outline of the processing contents in this device is as follows.

In the character recognition device (101), a feature amount (referred to as a first feature amount, a second feature amount, ...) Associated with a plurality of features extracted from a standard character pattern is stored in advance for all characters. The feature amount dictionary (102) is created. The group creating unit (103) reads out the feature amount of each character from the feature amount dictionary (102), divides each feature into groups of similar characters (character groups) with respect to the feature, and represents each group. Find the character number of the character. The character number of the representative character is stored in the group dictionary (104) created for each feature as the group number of each character. In this figure, the character group having the first characteristic amount is stored in the secondary storage device 104-a, and the character group having the second characteristic amount is stored in the secondary storage device 104-b. The structures of the feature dictionary (102) and the group dictionary (104) will be described later with reference to FIGS. 2 and 3.

Next, in the representative feature quantity dictionary creating section (105), the representative character of each character group is extracted from the group dictionary (104), and the feature quantity of the extracted character is read out from the feature quantity dictionary (102). , Is stored in the representative feature amount dictionary (106) as the representative feature amount of the character group. In this figure, the representative feature amount of the first feature amount is the secondary storage device 106-
a, the representative feature amount of the second feature amount is the secondary storage device 106-b.
To be stored. The representative feature amount dictionary (106) will be described later with reference to FIG.

Creation of the above-mentioned dictionaries (102, 104, 106) is performed in advance as a pre-process of the character recognition process.

Next, when the character pattern (107) is input to the character recognition device (101), first, the feature amount extraction unit (1)
08) extracts the feature amount (first feature amount, second feature amount, ...) Used for recognition from the input character pattern (107). The extracted feature amount is stored in the matching processing unit (10
It is output to 9). In the matching processing unit (109),
As a preparation for the matching process, the feature quantity dictionary (102)
All the characters registered in (1) are initialized as candidate characters in the candidate character table (110) (the candidate character table (110) will be described in detail in FIG. 5).
Next, the matching processing unit (109) narrows down the candidate characters on the candidate character table (110) by performing the matching process in the order of the first feature amount, the second feature amount, .... The outline of the matching process regarding the nth feature amount is as follows.
First, for the candidate characters stored in the candidate character table (110), all the group numbers of the character groups of the nth feature amount including those candidate characters are obtained. Then, for those groups, the representative feature amount dictionary (10
A matching process of the representative feature amount stored in 6) and the feature amount of the input character pattern (107) is performed. The matching result is stored in the evaluation value table (111). The evaluation value table (111) has an area for storing evaluation values for all characters (the evaluation value table (111) will be described in detail with reference to FIG. 6), and the matching result with the representative feature amount is within the character group. Set as the evaluation value of all candidate characters of. Therefore, 1 with the representative feature
Since the matching result of a plurality of characters can be obtained by performing the matching process once, it is possible to reduce the matching processing time. Next, the candidate character narrowing unit (112)
Then, the candidate characters in the candidate character table (110) are narrowed down based on the evaluation value of each character stored in the evaluation value table (111). As described above, the matching regarding the nth feature amount and the narrowing down of the candidate characters are performed. This is sequentially performed for each feature amount, and when the final recognition result is obtained, the recognition result output unit (113) outputs the recognition result (114). The recognition result (114) is a set of candidate characters narrowed down to the level of large classification or middle classification.

FIG. 2 shows the contents of the feature quantity dictionary (102). As shown in the figure, in the feature quantity dictionary (102), the character number (201) in the dictionary and the feature quantity (202-a, 202-b, 202-c, ...) Having that character are stored. Is stored. Each feature amount has a multidimensional value, and each dimension value is stored as an element of the feature amount.

FIG. 3 shows the contents of the group dictionary (104). As shown in the figure, a character number (301) and a group number (302-a, 302-b, 302-c, ...) To which the character belongs are stored in the group dictionary (104) for each feature amount. To be done. As the group number, the character number of the representative character of the group (that is, the character number of the character having the representative feature amount) is used. For example, from the group number (302-a) of the first characteristic amount in FIG. 3, regarding the first characteristic amount, the character with the character number 1 belongs to the group whose representative is the character with the character number 3, and the character with the character number 2 is The representative belongs to the group whose character number is 2 (self), the character of character number 3 belongs to the group whose representative is character of character 3 (self), and the character of character number 4 is the character number It can be seen that it belongs to the group which is the character of 2. The same applies to other feature quantities.

FIG. 4 shows the contents of the representative feature quantity dictionary (106). This figure shows a representative feature quantity dictionary (106-
9A shows a group number (401) and a representative feature amount (402) of the group are stored. As described above, the representative characteristic amount (402) is the characteristic amount of the representative character (character whose character number is the group number (401)) of the group. For other characteristic amounts, the characteristic amounts are stored in the same format. The dictionary (1
02, 104, 106) are created as a pre-process for performing the character recognition process.

FIG. 5 shows the contents of the candidate character table (110). The candidate character table (110) is a dynamic table used when performing character recognition by inputting a character pattern (107). In the candidate character table (110), the number of candidate characters (501) and the character number of the candidate character (50
2) is stored. The candidate character table (110) is initialized (character numbers of all characters existing in the feature quantity dictionary (102) are set) when performing character recognition by inputting one character pattern. It is used to narrow down the candidate characters each time the feature amount is matched.

FIG. 6 shows the contents of the evaluation value table (111). The evaluation value table (111) is a table used when the character pattern (107) is input for character recognition. The evaluation value table (111) stores the character number (601) of each character and the evaluation value (602-a, 602-b, 602-c, ...) For each feature amount related to the character. In the field of character number (601),
Character numbers of all the characters existing in the feature dictionary (102) are set in advance.

FIG. 7 is a flow chart showing the flow of recognition processing in the apparatus of FIG. First, the character pattern (1
07) is read (process 701), and the feature amount of the input character pattern is extracted (process 702). Next, as an initial value of the candidate character to be recognized, the candidate character table (11
All characters are set as candidates in 0) (process 70).
3). The recognition process of the input character pattern (107) is
It is realized by substituting 1 as an initial value into the loop counter n of the feature number (process 704) and repeating matching and narrowing down candidate characters for each feature amount. This repetitive processing will be described below.

First, for the n-th feature amount, all group numbers (matching groups) of character groups including candidate characters currently set in the candidate character table (110) are obtained (process 705). Details of the process 705 will be described later with reference to FIG. Next, the representative feature amount of all the character groups obtained in the process 705 and the process 702.
A matching process is performed on the n-th feature amount with the feature amount of the input character pattern extracted in step S5 to obtain an evaluation value. The obtained evaluation value is stored in the evaluation value table (111) as the evaluation value of the character in the group (process 706). Since the evaluation value obtained by matching with the representative feature amount of the character group is stored in the evaluation value table (111) as the evaluation value of all the candidate characters belonging to the character group, a plurality of evaluation values are stored in the character group. If there are candidate characters,
This is the same as performing the matching process with all the candidate characters in the character group in one matching process using the representative feature amount, and the matching process can be reduced and the recognition speed can be improved.

When the matching process of step 706 is completed, the candidate characters in the candidate character table (110) are narrowed down based on the evaluation value obtained here (candidate characters whose evaluation value is equal to or more than a predetermined value remain. The candidate character in the table (110) is reset (process 707). Nth
When the recognition process for the feature amount is completed, the loop counter n for the feature number is incremented by 1 (process 70).
8). Then, it is determined whether or not the matching is over (process 7).
09). If the matching with all the feature amounts is not completed and the number of candidate characters is still a predetermined number or more, the process returns to the process 705 to perform the matching with the next feature amount and the narrowing down process. When the matching and narrowing down processing is completed for all the feature amounts, or when the number of candidate characters falls below a predetermined number and the recognition candidates are confirmed in the middle stage (processing 709), the matching and narrowing down processing is ended and the obtained recognition The candidates (candidate characters remaining in the candidate character table (110)) are output (process 710).

FIG. 8 is a flow chart showing the contents of the group extraction processing of the processing 705. In this process, the candidate character table (110)
All the group numbers including the candidate characters of (the group regarding the nth feature amount to be matched in the next process 706) are obtained. In the process of FIG. 7, for example, at the stage when the matching and narrowing-down process regarding the first feature amount is completed, the candidate character table (110) is set with candidate characters after the narrowing-down of the first feature amount.
Next, the matching and narrowing-down process regarding the second feature amount is performed. Since the first feature amount has already been narrowed down, it is necessary to perform matching with the representative feature amounts of all the character groups regarding the second feature amount. Not necessarily. Therefore, the candidate character table (11
All the group numbers of the character groups including the candidate character set to 0) are obtained. In the matching process of the second characteristic amount, it suffices to perform matching with the representative characteristic amount only for the group number thus obtained. FIG. 8 shows such a group extraction process.

First, in the process 801, as the initialization process, the group extraction management table (901) shown in FIG. 9 is initialized. The group extraction management table (901) includes a group number (902) and an extraction flag (903). A group with a group number (902) whose extraction flag (903) is 0 indicates a group that does not need to be matched, and a group with a group number (902) whose extraction flag (903) is 1 indicates a matching target.
In the initialization processing of the processing 801, the group numbers of all the groups of the nth feature quantity which are the processing targets are set in the field of the group number (902), and the corresponding extraction flags (903) are all cleared to 0. All the group numbers of the nth feature amount are obtained by referring to the representative feature amount dictionary (106).

Next, in processes 802 to 807, a process of setting the extraction flag (903) of the character group to be matched is performed to manage whether or not each character group has been extracted. Specifically, first, in process 802, one character remaining as a candidate character is read from the candidate character table (110). In processing 803, the group dictionary (10
4) is read and the group number to which the candidate character read in the process 802 belongs is obtained. In process 804, the group extraction management table (901) is referenced, and the extraction flag (903) corresponding to the group number obtained in process 803 is read. Then, in process 805, it is checked whether the group has already been extracted as a recognition target. If not yet extracted (ie, extraction flag (90
(3) is 0), in process 806, the extraction flag (903) of the group extraction management table (901) is set (that is, set to 1). In process 807, it is checked whether or not an unprocessed candidate character remains in the candidate character table (110), and if it remains, the process returns to process 802 to read the next character, and processes 803 to 806 are repeated. If not, the extraction flag (90
Only the group for which 3) is set is extracted as a group to be matched, and the process ends.

Next, the procedure for creating the group dictionary (104) will be described. 10 and 11 are diagrams showing the concept of a method of grouping feature amounts. In each case, the feature quantity of each character in the dictionary is mapped on a plane assuming that the feature quantity has two feature elements. As shown in FIG. 10, a dot (1001) indicated by a double circle.
Indicates the distribution of the feature quantity. A circle is drawn from the representative feature quantity (1002) at a distance of a fixed threshold value (1003), and the inside of this circle is set as a grouping range (1004). Characters included in the grouping range (1004) are extracted, and these are set as one group.

Specifically, for each feature amount, assuming that the feature amount of the character is the representative feature amount (1002) for all the characters in the feature amount dictionary (102), as shown in FIG. When a group is created with a predetermined threshold value (1003), the number of characters included in the group is calculated. Then, in order from the character in which more characters are included in the group, if the character is already included in another character group, the character is not represented, otherwise, The character is represented and the characteristic amount of the character is set as the representative characteristic amount (1002). By repeating this process, all the representative feature amount setting and grouping processes are performed. Details of the contents of this processing will be described later with reference to the flowchart of FIG.

If one character belongs to the range of a plurality of groups in the above process, it is included in the group of the closest representative feature amount as shown in FIG. That is, the characters (110
1) exists, the representative feature amount of the included group (representative feature amount 1: 1102-a, representative feature amount 2: 110).
2-b) distance (distance to representative feature amount 1: 110
3-a, distance to the representative feature amount 2: 1103-b), respectively. Then, the characters (1101) included in the plurality of groups are the groups of the representative feature amount having the closest obtained distance (in the figure, the representative feature amount 1 (1102-
Register as characters included in group a).

When the threshold value of the distance used for the grouping is given by the value B, whether the character is included in the group or not can be determined by the mathematical expression shown in the following Equation 1 if it is an n-dimensional feature. Xi (i = 1 ... n) in the equation 1 is the feature quantity of the character, and Gi (i = 1 ... n) is the representative feature quantity. The right side of Expression 1 is a value obtained by squaring the distance between Xi (i = 1 ... n) and Gi (i = 1 ... n).

[0033]

[Equation 1]

FIG. 12 is a flow chart showing the contents of the feature amount grouping process. In process 1201,
In order to determine the representative feature amount, the number of characters included in the group is calculated for all the characters in the feature amount dictionary (102) when the feature amount of each character is used as the representative feature amount. Next, in process 1202, the characters are extracted in descending order of the number of characters obtained in process 1201. If the extracted character is not registered in the group (all characters are not registered in the group in the initial stage) (process 1203), the characteristic of this character is set as the representative characteristic amount, and this representative characteristic is set. A group of characters whose distance from the amount is within a threshold is created (process 1204). Then, if the created group contains characters that already belong to any other group, the group representative feature amount that already belongs and the distance between this character and the representative feature of the group that is being registered now. The amount and the distance of this character are compared, and this character is registered only in the closer group (Process 1
205). The processes 1202 to 1205 are repeated until all the characters belong to any group (process 1206).

[0035]

According to the present invention, the matching result of a plurality of characters can be obtained by performing the matching with the representative feature value once, and the error of the recognition evaluation value obtained by this processing can be set to an appropriate threshold value. If the grouping is performed by using, the recognition rate is not so large as to affect the recognition rate, and thus the recognition speed can be improved without lowering the recognition accuracy.

[Brief description of drawings]

FIG. 1 is a block diagram of a character recognition device according to the present invention.

FIG. 2 is a diagram showing the contents of a feature dictionary.

FIG. 3 is a diagram showing the contents of a group dictionary

FIG. 4 is a diagram showing the contents of a representative feature quantity dictionary.

FIG. 5 is a diagram showing the contents of a candidate character table.

FIG. 6 is a diagram showing the contents of an evaluation value table.

FIG. 7 is a flowchart of character recognition processing.

FIG. 8 is a flowchart of group extraction processing.

FIG. 9 is a diagram showing the contents of a group extraction management table.

FIG. 10 is a diagram showing the contents of grouping processing.

FIG. 11 is a diagram showing processing of characters included in a plurality of groups.

FIG. 12 is a flowchart of grouping processing.

[Explanation of symbols]

101: character recognition device, 102: feature amount dictionary, 103:
Group creation unit, 104: Group dictionary, 105: Representative feature amount dictionary creation unit, 106: Representative feature amount dictionary, 107:
Character pattern, 108: feature amount extraction unit, 109: matching processing unit, 110: candidate character table, 111: evaluation value table, 112: candidate character narrowing unit, 113: recognition result output unit, 114: recognition result, 201: feature Quantity dictionary character number, 202: feature amount storage area, 301: group dictionary character number, 302: group number storage area, 4
01: Group number of representative feature amount dictionary, 402: Storage region of representative feature amount, 501: Number of candidate characters of candidate character table, 502: Candidate character number of candidate character table, 60
1: character number of evaluation value table, 602: evaluation value storage area of evaluation value table, 901: group extraction management table, 902: group number of group extraction management table, 903: extraction flag of group extraction management table,
1001: Mapping of character feature amount of dictionary, 100
2; mapping of representative feature amount, 1003: threshold of grouping, 1004: range of grouping, 1101: mapping of feature amount of characters included in a plurality of groups, 110
2: Representative feature amount, 1103: Distance to the representative feature amount.

Claims (5)

[Claims] 1. In a character recognition method for recognizing a document image and converting it into code information, a plurality of features for character recognition are extracted from a standard character pattern of characters, and the extracted feature quantities are stored in a feature quantity dictionary. The step of storing, and for the plurality of features stored in the feature amount dictionary, a plurality of character groups are created by grouping characters having similar features for each feature, and a representative of each character group is created. A step of determining a representative character, a step of extracting a plurality of features for character recognition from a recognition target character pattern extracted from a document image, and a feature amount obtained from the recognition target character pattern for each feature. An evaluation value acquisition step is performed to perform matching processing with the representative feature amount that is the feature amount of the representative character of each character group created for that feature, and obtain the matching evaluation value. And a candidate character narrowing step for narrowing down candidate characters by using the obtained evaluation value as an evaluation value for matching the character pattern to be recognized with all the characters belonging to each character group. Character recognition method that features. 2. A character recognition method for recognizing a document image and converting it into code information, extracting a plurality of features for character recognition from a standard character pattern of a character, and extracting the extracted feature quantity in a feature quantity dictionary. The step of storing, and for the plurality of features stored in the feature amount dictionary, a plurality of character groups are created by grouping characters having similar features for each feature, and a representative of each character group is created. , A step of creating a group dictionary that stores information indicating which character group each character belongs to, and a representative character of each character group related to that characteristic for each characteristic The step of creating a representative feature amount dictionary that stores the representative feature amount that is the feature amount of the A step of extracting a number of features, and for each feature, perform a matching process between the feature amount obtained from the recognition target character pattern and the representative feature amount of each character group related to that feature, and obtain a matching evaluation value. A value acquisition step, a step of storing the obtained evaluation value in the evaluation value table of each character as an evaluation value for matching the character pattern to be recognized with all the characters belonging to each character group, and for each character And a candidate character narrowing step of narrowing down candidate characters using the evaluation values stored in the evaluation value table. 3. When a feature amount relating to a plurality of features for character recognition is called an nth feature amount (n = 1, 2, 3, ...), the nth feature is obtained in the evaluation value acquisition step and the candidate character narrowing step. After performing the matching and the narrowing on the amount and the next matching and the narrowing on the n + 1th characteristic amount, the n + 1th candidate characters including the candidate characters narrowed down on the nth characteristic amount are included.
The character recognition method according to claim 1 or 2, wherein all character groups of feature quantities are obtained, and in the next evaluation value acquisition step relating to the (n + 1) th feature quantity, matching is performed regarding the n + 1th feature quantity with respect to the obtained group. 4. The grouping of characters is performed for all the characters stored in the feature dictionary by assuming that the character is a representative character and using characters included within a predetermined threshold distance from the character. When a temporary group is created, the number of characters belonging to the temporary group is calculated, the temporary group with the larger number of characters is determined as a regular group, and when there are characters included in multiple groups, The character recognition method according to claim 1 or 2, wherein the character belongs to a group having a shorter distance to the characteristic amount of the representative character. 5. A character recognition device for recognizing a document image and converting it into code information, and a feature quantity dictionary storing feature quantities relating to a plurality of features for character recognition extracted from a standard character pattern of characters, For a plurality of features stored in the feature amount dictionary, groups of characters with similar features are grouped for each feature, and a group dictionary that stores information indicating which character group each character belongs to, and each feature , A representative feature amount dictionary that stores representative feature amounts that are the feature amounts of the characters that are representative of the created character group, input means for the character pattern to be recognized, and the recognition target characters input by the input means. A means for extracting a plurality of features for character recognition from a pattern, and a matching process for a feature amount obtained from an input recognition target character pattern and a representative feature amount of each character group. Is performed for each feature, and the evaluation value of matching is stored in the evaluation value table of each character as an evaluation value of matching between the character pattern to be recognized and all characters belonging to each character group. Means, a candidate character narrowing means for narrowing down candidate characters using the evaluation values stored in the evaluation value table for each character, and a candidate character finally obtained by the candidate character narrowing means. A character recognition device comprising means for outputting a recognition result.