JPH1185910A - Character recognition apparatus and method, and recording medium recording the method - Google Patents

Abstract

(57) [Summary] [PROBLEMS] To improve the recognition accuracy and speed of a character recognition device. SOLUTION: A character extracting unit 101 extracts a character image from a document image. The feature amount extraction unit 102 extracts a feature amount from a character image for each character. The category determination unit 104 obtains the similarity between the character feature amount of each character and the standard feature amount of the category in the category dictionary 103 to obtain the category of each character. The morpheme extraction unit 106
Morphological dictionary 1 with category or category string as search key
05 is retrieved to extract morphemes, a network is generated by connecting the morphemes, and grammatical processing is performed to obtain a combination of morphemes having a high priority. When one morpheme cannot be determined, the detailed recognition unit 107 determines the similarity of the characters constituting the morpheme by comparing it with the standard feature amount of each character in the category dictionary 103 to determine one morpheme.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device and method, and a recording medium on which the method is recorded.

[0002]

2. Description of the Related Art A character recognition device is a first character recognition device that uses a scanner or the like to read a document image by L / E (Light / Electric) conversion or the like, recognizes characters, and converts the characters into character codes. And a second character recognition device that receives input of coordinate data of handwritten characters online using a pen, a tablet, or the like, recognizes the characters, and converts the characters into character codes.

[0003] A first conventional character recognition device is described in, for example, Japanese Patent Application Laid-Open No. 62-219190. Hereinafter, the character recognition device will be described.

FIG. 34 is a configuration diagram of a conventional character recognition device. This character recognition device includes a character cutout unit 3401,
Feature amount extraction unit 3402, recognition dictionary 3403, similarity calculation unit 3404, morpheme dictionary 3405, and post-processing unit 3406
It is composed of

A character extracting unit 3401 projects a character image to be extracted on a vertical axis and a horizontal axis, and extracts an image corresponding to one character from a document image with a range in which a continuous image can be obtained as one character range. The correction is also performed using information on the character size and the pitch between characters.

[0006] The feature amount extraction unit 3402 extracts the feature amount of each cut-out character image. For example, as described in JP-A-63-34682,
The inclination or the like of a line segment in the character image is obtained, and vector data is extracted as a feature amount.

[0007] The recognition dictionary 3403 is a dictionary in which standard feature values and character codes are associated with each other in order to specify a character code from the feature values of a character image. Here, the standard feature value is
It refers to a standard character shape feature specified by a character code.

FIG. 35 shows a specific structure of the recognition dictionary 3403. The recognition dictionary 3403 registers a pair of a standard feature amount 3501 of a character and a corresponding character code 3502.

The similarity calculating section 3404 includes the feature extracting section 3
The feature amount of the character image extracted in 402 is recognized by the recognition dictionary 340.
3 is compared with the standard feature amount 3501, and a plurality of character codes 3502 are obtained as candidate characters in order from a character having a high degree of matching. A plurality of candidate characters are obtained for each character image.

The morphological dictionary 3405 registers character codes or character code strings representing morphemes and part-of-speech as attributes of morphemes in association with each other. Here, the morpheme refers to a minimum linguistic unit and includes not only independent words such as nouns and verbs, but also auxiliary words, prefixes, suffixes, and the like. FIG. 36 shows a morphological dictionary 3
405 shows the specific contents. Morphological dictionary 3405
Is registered as a set of the morpheme 3601 and the part of speech 3602 of the morpheme. Here, although a character code or a character code string is recorded in the morpheme 3601, a character or a character string that is a morpheme is shown so that the content can be easily understood.

The post-processing unit 3406 includes a similarity calculating unit 340
The morphemes in the morphological dictionary 3405 are compared with each candidate character or a character string obtained by combining the candidate characters obtained in step 4, and the morphemes that match each other are extracted. Further, grammatical processing is performed to obtain a sequence of morphemes having the highest priority. .

For example, a character string image 3 as shown in FIG.
In the case of recognizing 701, the character extracting unit 3401 extracts character images 3801 and 3802 shown in FIG.

[0013] The characteristic amount extracting section 3402 extracts the characteristic amount from each of the cut-out individual character images 3801 and 3802.

The similarity calculating section 3404 includes the feature quantity extracting section 3
The feature amount extracted in 402 is compared with the standard feature amount 3501 registered in the recognition dictionary 3403, and the similarity is calculated. Character code 3 corresponding to the descending order of similarity
FIG. 39 shows an example of extracting the character 502 up to the fourth candidate character of the similarity. That is, the character image 380
In 1, the sentence, height,
“Mata” and “branch” are extracted as candidate characters, and in the character image 3802, “good”, “author”, “bodhi”, “blue”
Is extracted as a candidate character.

Although the candidate characters shown in FIG. 39 are actually extracted as character codes, they are represented by characters so that the contents can be understood.

The post-processing unit 3406 has a morphological dictionary 3405
The words “bunsetsu”, “bunsetsu”, “bunbo”,..., Combining the morpheme 3601 and the candidate characters extracted by the similarity calculation unit 3404 or the first and second candidate characters And check if a match is registered. As a result, “sentence (noun)”, “blue (noun)”, “blue (adjective)”, and “good (adjective)” match. The post-processing unit 3406 rejects “blue (adjective)” and “good (adjective)” from the second candidate character by applying a grammatical rule that an adjective follows an inflected ending. Thus, the compound word of “sentence (noun)” and “blue (noun)” becomes the word with the highest priority, and “sentence blue” is obtained as a character string recognition result.

Next, the operation of this character recognition device will be described.
This will be described with reference to the flowchart of FIG.

The character cutout unit 3401 projects a character image on the vertical axis and the horizontal axis, sets a range in which a continuous image can be obtained as one character range, and cuts out an image corresponding to one character from the document image (S4001). The feature amount extraction unit 3402 extracts the feature amount of each cut-out character image (S
4002). Next, the similarity calculation unit 3404 refers to the recognition dictionary 3403, checks the extracted character feature amount against the standard feature amount, and obtains a character code as a candidate character in order from the character having the highest similarity. As a result, a plurality of candidate characters are obtained for each character image (S4003). Post-processing unit 3406
Matches the morpheme 3601 of the morpheme dictionary 3405 with each candidate character or a candidate character string combining each candidate character, extracts a matching morpheme, and further performs grammar processing to obtain a sequence of words having the highest priority. (S4004).
The obtained word is output as a character recognition result (S400
5).

Next, a second conventional character recognition apparatus will be described. In this device, a method of improving recognition performance using linguistic information in addition to the shape characteristics of the character represented by the coordinate data has been proposed. For example, there are techniques described in the IEICE Transactions Vol. J69-D No. 9, pp. 1292-1301, 1986 and Japanese Patent Application Laid-Open No. 6-124364.

FIG. 41 is a block diagram of this character recognition device. The character recognition device includes a coordinate input unit 4101, a character cutout unit 4102, a feature amount extraction unit 4103, a recognition dictionary 410
4, a matching unit 4105, a morphological dictionary 4106, a post-processing unit 4107, and an output unit 4108.

The coordinate input unit 4101 is, for example, a tablet. The tablet can detect whether the pen is in contact with the tablet (pen down) or not (pen up). If the pen is in contact, the coordinate data of the pen is displayed at regular intervals in the character cutout unit 4102. Is input to Here, the trajectory from pen down to pen up is called a stroke. A virtual stroke from pen-up to pen-down is called an off-stroke. The trajectory of the pen when the pen is down is displayed on the output unit 4108. When an input / output integrated tablet is used, the coordinate input unit 4101 and the output unit 4108 are integrated.

The character cutout unit 4102 includes a coordinate input unit 4
A character area is cut out from the coordinate data input from 101 using features such as stroke interval, pen-up time, character size, position, shape, aspect ratio, and character pitch. The method of extracting the character area is described in detail in, for example, Japanese Patent Application Laid-Open No. 6-124364, and the aspect ratio of the character, the character width, the character size, and the maximum space width in the character are used as the weight of the node. , A character pitch and a space width between characters are set as link weights, and a K-th shortest path is obtained to obtain a K-th character area.

The feature amount extraction unit 4103 extracts the feature amount of the extracted character region. Numerous methods have been proposed for this feature amount. For example, the feature points obtained by the following procedures (1) to (4) using a technique described in detail in Japanese Patent Application Laid-Open No. 8-249424 are disclosed. Is a feature amount.

(1) For each stroke, when the distance between the coordinate points is equal to or less than a fixed value Nlmin, the coordinate points are removed and the remaining coordinate point sequence is used as a feature point. (2) For consecutive feature points Pi-1, Pi and Pi + 1, Pi-
If the angle represented by the difference between the direction from 1 to Pi and the direction from Pi to Pi + 1 is equal to or smaller than a fixed value Ndist, Pi is removed and the remaining coordinate point sequence is set as a feature point. (3) Obtain a rectangle circumscribing the stroke sequence cut out as a character area, and normalize each feature point to a predetermined size. (4) Feature points are arranged in the order of stroke input for each stroke sequence cut out as a character area. At this time, if the distance between the feature points is equal to or greater than the fixed value Nfmax, a coordinate point corresponding to the feature point is selected, normalized, and added to the feature point. If the distance is equal to or smaller than Nfmin and the smaller angle between the feature points before and after the target feature point is smaller than the fixed value Nddst, the target feature point is removed. In the above method, since the feature point is added to the off-stroke by (4), the feature of “continuation character” or “break character” is also represented.

The recognition dictionary 4104 registers standard character amounts of characters and character codes in association with each other. FIG. 35 above
Has the same structure as that of the recognition dictionary 3403 shown in (1), but has standard feature values specific to the online character recognition device.

The matching unit 4105 calculates a D value between the extracted characteristic amount of the character area and the standard characteristic amount of the recognition dictionary 4104.
A P (Dynamic Programming) matching process is executed, and the similarity is calculated for each. Here, the similarity is a matching distance, and the shorter the matching distance, the higher the similarity.
The matching distance is a linear combination of the Euclidean distance between the feature points and the direction difference. As a result of the DP matching, a predetermined number of character codes are obtained as candidate characters in order from a character having a short matching distance (high similarity).

The post-processing unit 4107 checks each candidate character and a character string obtained by combining each candidate character with the morpheme 3601 of the morpheme dictionary 4106, and extracts a matching morpheme.
Further, grammatical processing is performed to obtain a row of morphemes having the highest priority. Various methods are conceivable as a method of obtaining the arrangement of the morphemes having the highest priority. Here, a network having the morpheme as a node is configured, and the route having the highest priority is obtained under the following conditions.

(1) It is checked whether or not two morphemes are connected by grammar processing. If not, the corresponding link on the network is rejected. (2) The number of clauses on the path from the beginning to the end is checked, and the path having the smallest number of clauses is selected. This is based on a generally known rule of thumb (the minimum number of phrases) that a sequence of morphemes with a small number of phrases has high reliability as a sentence. Note that the number of clauses is obtained as the number of independent words on the route. An independent word is a morpheme having a part of speech that can form a single phrase, and a noun, an adjective, and a verb are independent words. However, verbs (words to be used, which include adjectives and verbs) must have inflected endings. Adjuncts (particles and auxiliary verbs) and affixes (prefixes and suffixes) are not independent words. (3) Select a path having the smallest total recognition order of each character in all nodes on the path.

The output unit 4108 outputs a character code string of the obtained morpheme as a character recognition result so that the character code string can be visually recognized.

For example, a case in which the handwritten data shown in FIG. 42 is recognized by characters will be described. Writing data is input from a coordinate input unit 4101. The character cutout unit 4102
A set of strokes corresponding to one character is cut out from the character string 4201 in FIG. 42, and as shown in FIG.
01,4302.

The characteristic amount extracting section 4103 extracts the characteristic amount of each of the cut-out individual character regions 4301 and 4302.

The matching unit 4105 includes a recognition dictionary 410
4, the standard features (standard features 3501 shown in FIG. 35)
Are sequentially extracted, and the DP between the feature amount of each character area and
Matching is performed, and a character code (character code 3502 shown in FIG. 35) corresponding to the standard feature amount having a high degree of similarity is extracted. Here, the character codes up to the third place are taken out as candidate characters. As a result, in the character area 4301, “sentence”, “e”, “ぇ”
Is extracted. Similarly, for the character area 4302,
"Good", "pu" and "heat" are taken out.

The post-processing unit 4107 stores a character code of each candidate character or a character code string of a candidate character string obtained by combining the character codes and the morpheme 36 of the morphological dictionary 4106 (3405).
Compare with 01. Where the candidate characters "sentence",
In addition to “E”, “ぇ”, “Good”, “P”, and “Hot”, combining candidate characters for the character area 4301 and candidate characters for the character area 4302 yields “Bunzen”, “Bunpu”,
"Ben heat", "E good", "E phu", "E heat", "ぇ good",
Nine character strings of “ぇ PU” and “ぇ Hatsu” are obtained. on the other hand,
FIG. 36 shows an example of a morpheme registered in the morpheme dictionary 4106.
01 and the character area 43 (sentence (noun))
Although candidate characters “good (adjective)” and “hot (adjective)” for 02 are registered, no candidate character string combining candidate characters for the character region 4301 and candidate characters for the character region 4302 is registered. . Therefore, the post-processing unit 4107 obtains “sentence (noun)”, “good (adjective)”, and “hot (adjective)” as a search result of the morphological dictionary 4106.

The post-processing unit 4107 further performs grammar processing. Here, “good (adjective)” and “heat (adjective)” do not have grammatical endings that follow each, and are not grammatically established. Also, “sentence (noun)” is formed as a phrase alone, but there is no succeeding phrase. As described above, since the arrangement of the morphemes constituting the entire writing character string cannot be obtained, the post-processing fails.

Since the post-processing has failed, the output unit 4108 selects the character having the highest similarity by the matching unit 4105 and outputs "bunbun".

Next, the operation of the character recognition device will be described.
This will be described with reference to the flowchart of FIG. The coordinate input unit 4101 inputs the coordinate data on the tablet to the character cutout unit 4102 (S4401).

The character cutout unit 4102 includes a stroke interval, pen-up time, character size, position, shape,
An area of each character is cut out from the coordinate data using features such as an aspect ratio and a character pitch (S4402).

Next, the characteristic amount extracting unit 4103 extracts the characteristic amount of the coordinate data of each cut-out character area (S4403). The matching unit 4105 compares the feature amount obtained by the feature amount extraction unit 4103 with the recognition dictionary 4
Matching is performed with the standard feature amount of the reference character 104, and a plurality of character codes are obtained as candidate characters in order from the character corresponding to the standard feature amount with a high degree of matching (S4404).

Next, the post-processing unit 4107 converts each candidate character or a character string obtained by combining each candidate character with the morphological dictionary 410.
The morphemes are compared with each other, the morphemes that match each other are extracted, and further grammatical processing is performed to obtain a sequence of morphemes having the highest priority (S4405).

Next, the output unit 4108 outputs the morpheme on the selected route as a character recognition result (S440).
6).

[0041]

However, in the first and second character recognition devices, the similarity calculation unit 3404 or the matching unit 4105 uses the feature amount of the character area and the recognition dictionary 3403 to extract candidate characters. Since it is necessary to collate with all the standard feature values of 3404, it takes a long time to collate.

Although a predetermined number of character candidates having a high degree of similarity are extracted to reduce the amount of collation of the post-processing units 3406 and 4107 with the morphological dictionary 3405, correct recognition characters and If the second character "" in the above specific example is not included, the character string cannot be correctly recognized. Here, in order to correctly recognize a character string, it is necessary to increase the number of a predetermined number of character candidates so that the character candidates always include correct recognition characters. By doing so, the post-processing unit 340
It takes a long time for the comparison time of 6, 4107.
If the number of character candidates increases, the post-processing units 3406, 4406
107 may retrieve the recognition result of a plurality of character strings.

Further, in the second character recognition device, handwritten handwriting has a large individual difference, and a handwritten character deviating from a standard character shape may be input. For example, when the character elements “day” and “month” of the kanji “akira” are input at the same size and at the same interval as the characters before and after them, the matching unit 4105 correctly extracts “akira” as a candidate character. This is impossible in principle.

The present invention has been made in view of the above problems, and has as its object to provide a character recognition device having high recognition accuracy and high processing speed.

[0045]

According to the present invention, there is provided a character recognition device for recognizing a character from data representing a character string, wherein characters or character elements having similar shapes are classified as the same category. A category dictionary that stores a representative standard feature quantity representing a feature quantity extracted from a standard shape of a character or a character element classified in the same category and an identifier of the category in association with one another, A morpheme dictionary that stores a morpheme, which is a minimum linguistic unit composed of a character or a plurality of character elements, and a category identifier or a category identifier sequence to which the character or character string constituting the morpheme belongs in association with each other; A feature value extraction unit that extracts a feature value from a shape of each character representing data, and a feature value extracted by the feature value extraction unit and the category dictionary. Category determining means for comparing the stored representative standard feature quantity with an identifier of a category associated with the representative standard feature quantity having a high degree of similarity, and according to the category identifier determined by the category determining means. A candidate morpheme extracting unit that extracts a plurality of candidate morphemes from the morphological dictionary; and a morpheme determining unit that determines a morpheme to be prioritized based on a predetermined criterion from the plurality of candidate morphemes extracted by the candidate morpheme extracting unit. I have to do that.

[0046]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the character recognition device according to the present invention will be described below with reference to the drawings.

(Embodiment 1)

FIG. 1 is a block diagram of a first embodiment of a character recognition device according to the present invention. This character recognition device reads a document image using a scanner or the like and recognizes a character string, and includes a character cutout unit 101 and a feature amount extraction unit 10.
2, a category dictionary 103, and a category determination unit 104
, A morpheme dictionary 105, a morpheme extraction unit 106, and a detail recognition unit 107. Note that the character cutout unit 1
01 and the feature amount extraction unit 102 are the character cutout unit 3401 and the feature amount extraction unit 3 of the above-described conventional first character recognition device.
The description is omitted because it is the same as 402.

As shown in FIG. 2, the category dictionary 103 includes an upper layer 201 in which a category and its standard feature amount are paired.
And a lower layer 202, 203 in which character codes included in each category are associated with their standard feature amounts. Here, the category refers to a character or character element having a similar shape that is classified into one type. Also,
The standard feature amount of a category indicates a feature of a standard shape of a character or a character element included in the category.

In the upper layer 201, the number of categories 204, a set 205 of a category and its standard feature amount, and a pointer 206 to each category are registered. 20 categories
Reference numeral 4 denotes a number obtained by classifying all characters or character elements by cluster analysis of standard feature values representing features of their standard shapes. The cluster analysis is described in "Multivariate Statistical Analysis", by Yutaka Tanaka and Kazumasa Wakimoto, Gendai Mathematics, p.230-244. If the number of categories matches the number of character codes, it becomes the same as the above-described conventional recognition dictionary 3403. Further, the content of the morpheme registered corresponding to the category or the category string of the morpheme dictionary 105 described later changes according to the number of categories.

In the lower layer 202, the number of characters 207 included in each category 205 and their character codes 208 are registered, and are linked by each pointer 206 in the upper layer 201. The number of characters 207 included in each category 205 differs depending on the category. Lower layer 2 corresponding to this character code
In 03, a standard feature quantity representing a feature of a standard shape of a character corresponding to the character code is registered. The standard features are registered in the order of the character codes.

The category determining unit 104 includes the feature amount extracting unit 1
02 and the characteristic amount of the character extracted by the category dictionary 1
Standard feature amount 2 of each category of the upper layer 201 included in “03”
05 and a category to which the character image belongs is obtained. As the similarity, for example, the square distance between the extracted feature value and the standard feature value is used. In this case, the category with the smallest distance value is the category to be sought.

The morphological dictionary 105 is used for the category determining unit 10.
This is for obtaining a morpheme that is a minimum linguistic unit by using the category obtained in step 4 or a category string in which the categories are continuous as a search key. FIG. 3 shows an example of the morphological dictionary 105. The morpheme dictionary 105 is registered as a set of a category (including a category string) 301, one or more morphemes 302 corresponding to the category, and a part of speech 303 as an attribute thereof. Although morphemes are registered by character codes, corresponding character notations are described in this figure for easy understanding.

For example, the category “Ca” 304 is associated with the morpheme “noun / sentence, father”. in this case,
The morpheme includes two sentences, “sentence” and “father”. The category “CaCb” 305 is associated with the morpheme “noun / document”. In this case, the category string “CaCb” in which the category “Ca” and the category “Cb” are continuous is associated with one morpheme “document”. Further, the category “Cb” 306 is associated with the morpheme “noun / blue” and the morpheme “adjective / good, blue”. As described above, when the morpheme 302 corresponding to the category 301 includes the morphemes 302 of different parts of speech 303, the morpheme is registered for each part of speech 303. This is because, as will be described later, when morphological processing is performed by the morphological extraction unit 106, morphemes are extracted based on the part of speech.

[0055] The morpheme extraction unit 106 includes the category determination unit 1
The category obtained in step 04 is compared with the category 301 of the morphological dictionary 105 as a search key, and a matching morpheme is extracted. Furthermore, grammatical processing is performed on the extracted combination of morphemes, and morphemes that are not grammatically established are rejected.

When a plurality of morphemes remain, an arrangement of morphemes having the highest priority among combinations of morphemes is obtained. As a method of setting the priority, the minimum number of clauses method is used here. The minimum number of clauses method uses an empirical rule that, among combinations of morpheme strings, a combination having a small number of clauses has high reliability, and is a method of giving priority to a combination of morpheme strings having the smallest number of clauses.

The detail recognizing unit 107 includes a morpheme extracting unit 106
In the case where the arrangement of morphemes (character strings) cannot be determined as one set, the characteristic amount extracted by the characteristic amount extraction unit 102 for each character constituting the character string and the standard of characters in the lower layer 203 of the category dictionary 103 Calculate the similarity with the feature value. Using the similarity as an evaluation point, the evaluation points of the respective character strings are totaled, and the arrangement of the morphemes of the character string having the higher evaluation point is used as the recognition result.

Next, a specific example of character recognition of a character string image using the character recognition device will be described.

(Specific Example 1) The character cutout unit 101 uses the character string image 370 shown in FIG.
When 1 is input, the corresponding image is cut out to obtain character images 3801 and 3802 shown in FIG.

Next, the characteristic amount extracting unit 102 extracts the characteristic amount for each of the cut-out individual character images 3801 and 3802.

The category determining unit 104 collates the extracted character feature quantity with the standard feature quantity 205 of the category in the upper layer 201 of the category dictionary 103 to obtain a category having the most similar standard feature quantity. Here, the character image 380
1 for the category “Ca”, and the character image 38
For 02, a category “Cb” is obtained.

In the morpheme dictionary 105, categories and morphemes as shown in FIG. 3 are registered. The morpheme extraction unit 106 uses the categories “Ca”, “Cb”, and “CaCb” obtained by the category determination unit 104 as search keys, and sets a part-of-speech and morpheme “noun / sentence” in which the category 301 of the morpheme dictionary 105 matches. , Father "401," noun / document "402," noun / blue "403 and" adjective / good,
"Blue" 404 is extracted, and these morphemes are combined as a character string to obtain a combination diagram shown in FIG. In FIG. 4, since “noun / document” 402 is composed of two characters, a character string is completed by this combination. Combination pattern 50 obtained in FIG.
, 507 are shown.

Next, the morpheme extraction unit 106 performs grammatical processing to obtain a sequence of morphemes having a high priority. That is, "blue (adjective)" and "good (adjective)" are adjectives, respectively, and are not grammatically established because there is no inflectional ending following "blue (adjective)" and "good (adjective)". Therefore, “blue (adjective)” and “good (adjective)” 404 are rejected.

Further, when the minimum number of clauses method is applied, a one-character noun has two more words than a combination of morphemes composed of two phrases.
The morpheme in which the noun of the character is composed of one phrase has priority,
"Document (noun)" has the highest priority. In this case, there is only one morpheme with the highest priority,
The character string “document” is set as the recognition result.

As described above, in the character recognition device of the present embodiment, the feature amount extracted from the character image by the feature amount
The category is obtained by collating with the standard feature amount 205 of the upper layer 201 of the third category. At this time, the matching with the standard feature amount is not advanced until the character code is determined, but only the category corresponding to the group of similar character shapes is determined. Then, the morpheme extraction unit 106 refers to the morpheme dictionary 105 using the category as a search key, extracts a morpheme corresponding to the matching category, and further performs grammar processing to obtain a character string with the highest priority. Therefore, it is possible to eliminate a character recognition error caused by a failure to obtain a correct candidate character at the time of collation with the dictionary 3403 as in the related art, and to obtain a character code of “document” correctly for the character string image 3701 in FIG. .

(Embodiment 2) Next, a description will be given of a solution to the case where a combination of morphemes having high priority cannot be determined as one set as in Embodiment 1 above. Note that the category dictionary 103 and the morphological dictionary 105 are the same as those shown in FIGS. 2 and 3, respectively.

Here, the case where the character string image "Japan" is input will be described. The character cutout unit 101
Characters are cut out, the feature amount extraction unit 102 extracts feature amounts, and the category determination unit 104 determines “Cx” as the first character category and “Cy” as the second character category.

The category 301 and the morpheme 302 shown in FIG. 3 are registered in the morpheme dictionary 105, and a set of part of speech and morpheme “noun / day,
Eye, White, Mouth "and" Adjective / White "are registered, and part-of-speech and morpheme combination" Noun / Shiroki, Japan "is registered corresponding to category" CxCy ", corresponding to category" Cy " In addition, a set of part of speech and morpheme “noun / tree, book, water” is registered.

In this case, the morpheme extraction unit 106 collates the morpheme dictionary 105 to obtain a morpheme combination shown in FIG.
As a result, as shown in FIG.
, 708 are obtained. Next, the morpheme extraction unit 106
Checks the grammar and rejects the morpheme "white (adjective)" without inflection. Also, the morpheme "Shiraki (noun)" and the morpheme "Japan (noun)" are the morphemes of the character string with the highest priority by applying the phrase number minimum method.

In this case, the combination of morphemes cannot be determined as one set.
Refer to for more details on the parts that cannot be determined. In this case, the characters that cannot be determined are the first character image and the second character image.

First, the detail recognizing unit 107 calculates the feature amount extracted by the feature amount extracting unit 102 from the character image of the first character,
The similarity of each character belonging to the category of the upper layer 201 of FIG. 2 that has already been determined with the standard feature amount 203 is obtained, and candidate characters having a large similarity are extracted. Next, the detail recognition unit 107
Extracts a candidate character having a high degree of similarity for the second character image as in the case of the first character image.

At this time, since the character codes of the characters classified into the category of the upper layer 201 are registered in the lower layer 202, the standard characteristic amount 203 of each character is collated in accordance with this character code to obtain the standard characteristic amount of each character. Can be obtained.

FIG. 8 shows the candidate characters obtained by the detailed recognition unit 107 as described above. As shown in FIG. 8, as the first candidate character, “date” 801, “mouth” 802, “white” 803,
.. Are obtained, and “tree” 805, “book” 806,
"Water" 807, "e" 808, ... are obtained.

Next, the detail recognizing unit 107 selects a character string having the highest priority according to the order of similarity of the obtained candidate characters. For example, if a method of giving priority to a smaller addition value of the rank in the already obtained character string candidate characters is adopted, the addition value of “Shiroki” becomes “3” + “1” = “4”, and “Japan” Is “1” + “2” = “3”, so that the character string “Japan” is selected as the recognition result.

Even when the morpheme extraction unit 106 cannot determine one combination of morphemes as in the specific example 2 described above, the detailed recognition unit 107 can obtain “Japan” as a correct recognition result.

In the present embodiment, if the combination of morphemes can be determined by the morpheme extraction unit 106 into one set, it is not necessary to determine the character code of the candidate character using the category dictionary 103. The number of degrees calculation can be reduced. Therefore, the similarity calculation unit 3 of the conventional device
As compared with the case where the 404 always determines the character code of the candidate character, it is possible to configure a character recognition device that operates at an extremely high speed.

(Modification of Specific Example 2) As described in Specific Example 2, the morpheme extraction unit 106 determines that the morpheme “Shiraki (noun)” and the morpheme “Japan (noun)” To get. In this case, since the combination of morphemes cannot be determined as one set, the detailed recognition unit 107 outputs the four characters “white”, which constitute the morpheme “Shiroki” and the morpheme “Japan”,
The similarity between each of the character standard feature amounts of the category dictionary 103 of “tree”, “day”, and “book” and the feature amounts extracted by the feature amount extraction unit 102 is calculated.

As a result, a similarity is obtained as shown in FIG. The similarity between the first character image and “white” and “day” is “60” and “30”, respectively, and the similarity between the second character image and “tree” and “book” is "40",
It is "50".

Next, the detail recognizing unit 107 assigns priority to the morphemes “Shiroki” and “Japan” based on the similarity.
As a method of assigning priorities, for example, when priority is given to the smaller value of the similarity, the added value of the similarity of “Shiroki” is 1
00, since the added value of the similarity of “Japan” is 80,
"Japan" has a higher priority. Therefore, the character string “Japan” is finally selected as the recognition result.

As described above, in the present embodiment, even if the morpheme extraction unit 106 cannot determine one combination of morphemes, a correct recognition result “Japan” can be obtained.

At this time, for only the morphemes that could not be determined as a set, the detailed recognition unit 107 compares the standard feature quantity of each character constituting the morpheme with the feature quantity obtained by the feature quantity extraction unit. Therefore, a character recognition device that operates at a much higher speed than the conventional method in which the similarity calculation unit 3404 calculates the similarity of all the target characters can be configured.

Next, the operation of this embodiment will be described with reference to the flowchart of FIG. Character cutout unit 101
Cuts out an image corresponding to one character from the document image (S
1001), the characteristic amount extraction unit 102 extracts the characteristic amount of each cut-out character image (S100).
2). Next, the category determination unit 104 obtains a similarity between the extracted character feature amount and each of the standard feature amounts 205 of the category dictionary 103, and obtains a category to which the character belongs (S1).
003). The morpheme extraction unit 106 includes the category determination unit 10
Morphological dictionary 1 using the category extracted in step 4 as a search key
Then, morphemes are obtained by comparing the morphemes with each other, the morphemes thus obtained are combined, grammatical processing is performed on the morpheme combinations, and morphological combinations that are not grammatically established are rejected. Further, from the combinations of morphemes that have not been rejected, a sequence of morphemes having a high priority is obtained using the minimum number of phrases method (S10).
04). The detail recognizing unit 107 determines whether a combination of morphemes with high priority has been determined as one set (S100).
5) If affirmative, proceed to S1007. If not, the standard feature amount of the character constituting each morpheme is compared with the feature amount obtained by the feature amount extraction unit 102, and the morpheme with the highest priority is obtained (S1006). Finally, the morpheme extraction unit 106 outputs a character string of a set of morphemes having the highest priority as a recognition result (S1007).

In the above embodiment, the upper layer 2
Although a category dictionary composed of two layers, namely 01 and the lower layers 202 and 203, was used, a category dictionary composed of three or more layers and a category dictionary optimized by changing the depth of the layers depending on the category was used. You may.

In the above-described embodiment, the category determining unit 104 determines one category per one character, and collates the column of this category with the morpheme dictionary. A method may be adopted in which the 104 extracts a plurality of categories in descending order of priority and matches the character position and the matrix of category candidates with the morphological dictionary 105.

In the embodiment of the present invention, the character extracting unit 101 extracts characters in a single pattern. However, a plurality of patterns with different extracting positions are extracted, and the morpheme extracting unit 106 executes a combination pattern graph of categories. May be created and collated with the morphological dictionary 105.

(Embodiment 2) FIG. 11 is a block diagram of a character recognition apparatus according to Embodiment 2 of the present invention. This character recognition device recognizes characters of a handwritten character string input online, and includes a coordinate input unit 1101, a character cutout unit 1102, a feature amount extraction unit 1103, a category dictionary 1104, and a category determination unit 1105. , A morphological dictionary 1106, a morphological extracting unit 1107, a detailed recognizing unit 1108, and an output unit 1109. Note that the coordinate input unit 1101, the character cutout unit 1102, the feature amount extraction unit 1103, and the output unit 1109 have the same configurations as those of the above-described conventional second character recognition device, and thus description thereof will be omitted.

Although the category dictionary 1104 has the same structure (see FIG. 2) as the category dictionary 103 of the first embodiment, its contents are different. That is, in the present embodiment, input of coordinate data including time information is received instead of input of a character string image as in the first embodiment. For this reason, a character region is cut out according to the coordinate data, and the characteristic amount of the character is extracted, so that the standard characteristic amount to be compared with this characteristic amount is different. In addition, since the standard features are cluster-analyzed and classified into categories, the characters or character elements included in the categories are also different.

FIG. 12 is an explanatory diagram showing the contents of the category dictionary. The category dictionary 1104 registers a standard feature amount 1201 of characters or character elements classified into n categories in the upper layer 1202 for each category, and stores a character code 1204 included for each category in the lower layer 1203 and a lower layer 1
205 shows a case standard feature amount 1 corresponding to those character codes.
206 are registered. Here, the character element refers to, for example, "bias" on the left side of the kanji "saka" and "near" on the right side. This structure is the same as that of the category dictionary 103 shown in FIG.

For example, in the category Ca, the characters “heaven”,
Characters such as “gen”, “sentence”, and “branch” are classified. Similarly, the category Cb includes the characters "sound", "blue", "harm",
"Books" are classified.

The arrangement of the standard feature values 1206 registered in the lower layer 1205 corresponds to the value of the character code. Therefore, if the character code is known, the registration position of the standard feature amount of the character can be obtained. For example, when JIS X0208 is used as the character code, the character with the smallest value of the character code is the reading point ``, '' excluding the space, and the character code of the reading point is expressed as 2122 (hex), so the reading point The standard feature is registered at the head position of the array position. Further, since the period "." Is represented as 2123 (hex), the standard feature value of the period is registered after the period. When the standard feature amount is registered as described above, the standard feature amount can be extracted by referring to the registered position obtained by subtracting 2122 (hex) from the character code of the target character. Note that, for example, a character code group is divided into two groups, and two values to be subtracted from the character codes are set so that no useless space appears in the array of the standard feature amounts 1206. If the code is larger than the predetermined value, another set value is subtracted without subtracting 2122 (hex).

The category determining unit 1105 stores the character feature amounts extracted by the feature amount extracting unit 1103 and the category dictionary 11
Standard feature amount 1 of the category of the upper layer 1202 included in the category 04
DP matching is performed between the category and the category having the largest similarity (smallest matching distance). The definition of the matching distance is determined by the matching unit 410 of the conventional device.
It is assumed that the matching distance is the same as that described in 5.

As shown in FIG. 13, the morphological dictionary 1106 has the same category 130 as the morphological dictionary 105 described above.
1 and one or more morphemes 1302 corresponding to the category are registered for each part of speech 1303 as its grammatical attribute. Note that morphemes 1302 are grouped by part of speech 1303.

For example, corresponding to the category “Ca” 1304, the morpheme “noun / heaven, husband, sentence, hole, element, one” 130
5 are registered. Similarly, the category “Cb” 130
In correspondence with No. 6, the morphemes “noun / sound, blue” 1307, “sa noun / harm” 1308, and “adjective / good, blue” 1309 are registered for each part of speech. Further, a category string “CaCb” 1 in which the category “Ca” and the category “Cb” are continuous
A morpheme “noun / document” 1311 is registered corresponding to 310. Category “Cx” 1311, Category “C”
Corresponding morphemes are also registered for “xCy” 1312 and category “Cy” 1313.

The morpheme extraction unit 1107 collates each category obtained by the category determination unit 105 or a category string combined in the order of appearance of the category with a category of the morpheme dictionary 1106, and extracts a morpheme corresponding to a matching category. . Grammar processing is performed on the extracted combination of morphemes to obtain a sequence of morphemes with high priority. This specific processing will be described in detail in a specific example described later.

When the morpheme extracting unit 1107 cannot determine the combination of morphemes into one set, the detailed recognition unit 1108 examines the feature amount of the case characters constituting the morpheme in more detail, thereby obtaining the first embodiment. Detail recognition unit 1
Similarly to 08, the morpheme having the highest priority is obtained.

The output unit 109 outputs the obtained morpheme notation (character code sequence) as a character recognition result.

Next, a specific example of character recognition of a handwritten character string using the character recognition device will be described.

(Specific Example 1) The coordinate input unit 1101 uses the handwritten character string 420 shown in FIG.
1 is input, the coordinate data is input to the character cutout unit 1
Output to 102.

The character cutout unit 1102 is the coordinate input unit 1
An area corresponding to one character is cut out based on the coordinate data input from 101, and a character area 4301 shown in FIG.
4302 is obtained.

Next, the characteristic amount extracting unit 1103 extracts the characteristic amount for each of the extracted character regions 4301 and 4302.

The category determining section 1105 matches the extracted feature quantity with the standard feature quantity 1201 of each category in the upper layer 1202 of the category dictionary 1104 to determine a category having the highest similarity. As a result, the category “Ca” is assigned to the character area 4301 and the character area
The category “Cb” is determined for each of the two.

The morpheme extraction unit 1107 searches the morpheme dictionary 1106 using the categories “Ca” and “Cb” determined by the category determination unit 1105 and the category string “CaCb” obtained by combining the categories “Ca” and “Cb” as search keys. Retrieval and morpheme "Noun / heaven, husband, sentence, hole, element, person" 130
5, “Noun / document” 1310, “Noun / sound, blue” 130
6. "Sa noun / harm" 1307 and "adjective / good, blue"
1308 is extracted.

Next, the morpheme extraction unit 1107 generates a network combining the extracted morphemes. FIG.
Indicates a network of morphemes. Network 140
1 is the sentence head 1402 and the sentence end 1403 of the input character string
The morphemes 1305, 1310, 1306, 1307, and 1308 extracted by the morpheme extraction unit 1107 are connected as nodes to nodes in the order of their categories. Node N14 containing morpheme 1310
04 is connected to the beginning 1402 of the sentence by a link L1405, and is connected to the end 1403 of the sentence by a link L1406. Node N 1407 containing morpheme 1305
Is linked to the beginning 1402 of the sentence by a link L1408. Further, the node N14 in which the morpheme 1306 is stored
09 and a node N1410 containing a morpheme 1307 and a node N1411 containing a morpheme 1308 are connected to the node N1407 by links L1412, L1413, and L1.
414 respectively. Nodes N1409, N1410, and N1411 are linked to link L14.
15, L1416, and L1417, respectively, at the end of the sentence
403.

The morphological extraction unit 1107 determines whether or not the connection is grammatically established for the network 1401 based on the grammatical information stored therein, and rejects the link if the connection is not established.

First, the node N1411 is not grammatically established because the part of speech is an adjective and there is no inflection ending following it. Therefore, link L1417 is rejected.

Next, the phrase number minimum method is applied as an empirical rule of grammatical information. According to this, sentence head 1402 and sentence end 14
03 and “Link L1405-Node N1404-
The number of clauses in the path indicated by the link L1406 is “1”, and the number of clauses in all other paths is “2”. Therefore, the number of clauses in the path including the node N1404 becomes the minimum.
Node N1404 is selected. As a result, a morpheme 1310 is obtained as a morpheme with a high priority.

Since the morpheme 1310 has only one character code string “document”, the output unit 1109 outputs the notation “document” as a recognition result.

As described above, in the present embodiment, a category is obtained by grouping characters having similar shapes, and morphemes are extracted from the morpheme dictionary 1106 using the category sequence as a key. Therefore, it is possible to reduce a character recognition error caused by a failure to obtain a correct candidate character in matching with the category dictionary 1104. Therefore, it is possible to obtain a character code of “document” correctly for the character string 4201 in FIG. Becomes possible.

Further, since the category determining unit 1105 performs matching between the obtained characteristic amount of the character area and the standard characteristic amount 1201 of the category having a similar shape, the category determining unit 1105 performs matching with the standard characteristic amount prepared for each character. Compared with the conventional method, the number of times of matching can be greatly reduced, and the operation can be performed at extremely high speed.

(Specific Example 2) Next, the case where the handwritten character string 1501 shown in FIG. 15 is input will be described. The character cutout unit 1102 correctly cuts out the “day” and “book” regions, the feature amount extraction unit 1103 extracts the feature amount, and the category determination unit 1105 sets “C” as the category of the first character.
"Cy" is determined as the category of the "x" and the second character.

As shown in FIG. 13, the morphological dictionary 1106 includes categories “Cx” 1311, “CxCy” 131
2. A morpheme is registered corresponding to “Cy” 1313. Corresponding to the category “Cx” 1311, the part of speech of the morpheme and the morpheme “noun / day, month, white, eyes, mouth” 1314,
And “adjective / white” 1315 are registered, and the part of speech of the morpheme and the morpheme “noun / Shiroki, Japan” 1316 are registered corresponding to the category string “CxCy” 1312,
A morpheme part of speech and a morpheme “noun / tree, book” 1317 are registered corresponding to the category “Cy” 1313.

The category notation conversion unit 1107 searches the morpheme dictionary 1106 using the obtained categories “Cx” and “Cy” and the category string “CxCy” obtained by combining them as a search key, and finds the part of speech of the morpheme and the morpheme “noun”. /
Day, month, white, eyes, mouth "1314," adjective / white "131
5, "Noun / Shiroki, Japan" 1316, "Noun / Tree, book"
Extract 1317.

Next, the morpheme extraction unit 1107 generates a network combining the extracted morphemes. FIG.
Denotes a network of morphemes. Network 1601
Has a node N1602 containing a morpheme 1316
A node N1603 in which a morpheme 1314 is accommodated;
The node N1604 in which the morpheme 1315 is accommodated and the node N1605 in which the morpheme 1317 is accommodated are sentence head 160
6 and a sentence end 1607 via a link as shown in FIG.

The morphological extraction unit 1107 applies grammatical processing in the same manner as in the first embodiment, and then applies the minimum number of clauses method to obtain “sentence 1606-link-node N1602-link-
The path of “end of sentence 1607” is selected. As a result, the morpheme 1316 accommodated in the node N1602 is obtained as a morpheme with a high priority.

Since the detail recognition unit 1108 has two candidate morphemes “Shiroki” and “Japan” in the morpheme 1316,
One must determine the recognition result. Therefore, the upper layer 1202 determined by the category determination unit 1107
Tracing the pointers of the categories "Cx", "C
The character codes of the characters classified as "y" are obtained. The degree of similarity between the characteristic amount of the character region of the first character extracted by the characteristic amount extraction unit 1103 and each standard characteristic amount of the lower layer 1205 specified by the character code of the category “Cx” in the category dictionary 1104 is calculated. Similarly, the similarity between the feature amount of the character area of the second character and each standard feature amount of the lower layer 1205 of the category “Cy” is calculated. FIG. 17 shows the characters arranged in order from the character having the highest similarity.

The detail recognizing unit 1108 adds the values of the degrees of similarity between the two morphemes “Shiraki” and “Japan”. “Shiroki” is “5” + “1” = “6”,
“Japan” is “2” + “2” = “4”. "Japan" with a small addition value is selected as a morpheme, and the recognition result is a character string "Japan". The output unit 1109 outputs a recognition result.

As described above, in the present embodiment, even if the morpheme extraction unit 1107 cannot determine one combination of morphemes, “Japan” is regarded as a correct recognition result.
Can be obtained.

Even when the combination of morphemes cannot be determined as one set, the operation can be performed at an extremely high speed as compared with the conventional method of matching with the standard feature amount prepared for each character.

(Modification of Specific Example 2) In the specific example 2,
When two candidate morphemes are selected as high-priority morphemes in the morpheme extraction unit 1107, the detailed recognition unit 1
Reference numeral 108 denotes the feature amount of each character region extracted by the feature amount extraction unit 1103 and the category “C” of the category dictionary 1104.
Although the degree of similarity is obtained by matching with the standard feature amounts of all the characters classified as “x” and “Cy” and ranked, the following procedure is also possible.

The detail recognizing unit 1108 directly extracts the corresponding standard feature of the lower layer 1205 of the category dictionary 1104 from each character code constituting the candidate morpheme. This is because, as described above, the standard feature amount 1206 of each character is registered in the order of the character code.

Next, the similarity between the feature extracted from the first character area and the standard features of the characters “white” and “date”, and the feature extracted from the second character area The similarity between the characters “tree” and “book” and the standard feature amount is calculated. FIG. 18 shows the result. Candidate morpheme Shiraki
Are added, the value of “60” + “40” = “1”
00 ", and when the similarity value of the candidate morpheme" Japan "is added," 30 "+" 50 "=" 80 ", and the smaller value is determined as the morpheme of the recognition result. As a result, the recognition result can be correctly recognized as the character string “Japan”.

As described above, in the present embodiment, even if there are two candidate morphemes, the correct recognition result "Japan"
Can be obtained.

At this time, the detail recognizing unit 1108 determines the standard feature amount of each character, the feature amount obtained by the feature amount extracting unit 1103 and the standard feature amount of each character only for morphemes that could not be determined as a set. , So compared to the conventional method,
Processing can be performed at extremely high speed.

(Specific Example 3) Next, the case where the handwritten character string 1901 shown in FIG. 19 is recognized will be described. When the handwritten character string 1901 shown in FIG. 19 is input, the coordinate input unit 1101 outputs coordinate data to the character cutout unit 1102.

The character cutout unit 1102 cuts out an area corresponding to one character from the input coordinate data.
(A) As shown in (b), a plurality of clipping candidates are obtained.
In the cutout candidate 2001 shown in FIG.
Characters are cut out in four areas of 002, 2003, 2004, and 2005. In the clipping candidate 2006 in FIG. 20B, the character regions 2002, 2003, 2
Characters are cut out in four areas 007 and 2008. Here, the first and second character regions of the cutout candidates 2001 and 2006 indicate the same rectangular region.

Next, the feature amount extraction unit 1103 extracts the individual character regions 2002 to 2005, 2007, and 2
For 008, the feature amount is extracted. Next, the category determining unit 1105 determines whether the extracted feature amount
04 and the standard feature amount 1201 in the upper layer 1202, and a category having the highest similarity is determined for each character area. FIGS. 21 (a) and 21 (b) show the category determining unit 110.
5 shows the category determined. Character area 2002, 2
For 003, a category "Cp" is obtained, respectively.
The category "Cp" is obtained for the character area 2004, the category "Cq" is obtained for the character area 2005, and the category "Cr" is obtained for the character area 2008. However, since the similarity does not reach the threshold for the character area 2007, no category is obtained and the character area 2007 is rejected. Note that the character areas 2002 and 200
3 and 2004 indicate different character areas, but are determined to be in the same category “Cp” because of their similar shapes.

FIG. 22 shows an example of a morpheme registered in the morpheme dictionary 1106. Categories "Cp", "C
q ”,“ Cr ”, category columns“ CpCp ”,“ CpCp ”
The fact that a morpheme is registered for each part of speech corresponding to "Cp" is the same as described with reference to FIG.

In this morphological dictionary 1006,
The morphemes (minimum linguistic units registered by character codes) classified by the category “Cp” include the nouns “day, month, white,
Eye, mouth "and the adjective" white "are registered. Category column “CpCp” combining two of this category “Cp”
In the example, the noun “Monday, Sun” is registered as a morpheme.

On the other hand, in the category string “CpCpCp” in which three categories “Cp” are combined, the noun “tomorrow” and the adjective verb “clear” are registered as morphemes. The character shapes classified into the category “Cp” are stored in the category dictionary 1
As can be seen from the lower layer 1203 of FIG. 104 (FIG. 12), the format is all classified as a character code, but the category “Cp” includes “day” as the shape of a character element, ie, “partial”. And “month” as “nearby” are practically classified. Since two words “tomorrow” and “clear” are registered in the category column “CpCpCp”, it is understood that the shape of the character element has been classified into the category “Cp”. This content will be described in detail later in a modified example of this specific example.

The morpheme extracting unit 1107 extracts corresponding morphemes from the morphological dictionary 1106 using the category determined by the category determining unit 1105 and a category string obtained by combining the categories as a search key. A network is created by combining the extracted morphemes. FIG. 23 shows a network generated from the category determined by the extraction candidate 2001 shown in FIG. 21A, and FIG. 23 shows a network generated from the category determined by the extraction candidate 2006 shown in FIG. 24.

Since the category of character area 2007 has not been determined by category determining section 1105, network 2401 in FIG. 24 has a missing node N2402 corresponding to character area 2007.
It does not have a connected path from 03 to the end of sentence 2404. Therefore, the network 2401 is rejected by the morpheme extraction unit 1107.

The morphological extraction unit 1107 applies the same grammatical processing and the minimum number of clauses method as in the first embodiment to the network 2301 shown in FIG.
2-link L2303-node N2304-link L2
305-node N2306-link L2307-first sentence 2
308 "are obtained.

Although the morpheme “tomorrow” accommodated in the node N2304 can be uniquely determined, the morpheme of the node N2306 has two characters “ha” and “ni” and cannot be determined. Therefore, as described in the modification of the specific example 2, the detailed recognition unit 1108 calculates the similarity between the feature amount of the character area 2005 and the standard feature amounts of the character codes “ha” and “ni” of the category dictionary 1104. calculate. Character code "ha"
Is calculated to have a higher similarity than “Ni”, and the morpheme “Ha”
Select As a result, the character string “Tomorrow is” is the recognition result.

As described above, in the present embodiment, even if the extraction fails, a correct recognition result “Tomorrow is” can be obtained.

(Modification of Specific Example 3) Next, as a modification of Specific Example 3, a case where the handwritten character string 2501 shown in FIG. 25 is input will be described. Character cutout unit 110
2 are character areas 2601, 2602, as shown in FIG.
Cut out 2603. The feature amount extraction unit 1103 extracts feature amounts, and the category determination unit 1105 determines whether the category dictionary 11
04 and character areas 2601, 2602, 260
3 for each of the categories "Cs", "Ct",
"Cu" is determined. In the category “Cs” of the category dictionary 1104, characters “large”, “height”, “below”, and the like are classified. The category “Ct” includes “A”,
"A", "ya", "ya", etc. are in the category "Cu"
"Anti", "friend", etc. are classified respectively.

In the morphological dictionary 1106, as shown in FIG. 27, the category string "CsCtCu" and the category "C
The morpheme corresponding to "u" is registered. Category "Cs", Category column "CsCt", "CtCu"
Is not registered.

[0137] The morphological extraction unit 1107 outputs
06 to extract a morpheme “Noun / Osaka” corresponding to the category sequence “CsCtCu” and a morpheme “Noun / Friend” corresponding to the category “Cu”, and generate a network 2801 shown in FIG.

At the time of creating a category dictionary, if the partial or neighboring writing data is used as the data to be classified in addition to the writing data of the character, the character and the partial or neighboring can be classified at the same time. In this case, in addition to characters such as "A", "A", "Ya", "
The unbalanced part of “saka” is also classified into the category “Ct”. Further, in addition to characters such as "anti" and "friend", the part beside "saka" is also classified into the category "Cu". Using such a classification result, the category string “CtCu” is associated with “saka”, and the category string “CsC” is assigned to the morpheme “Osaka” including “saka”.
tCu ”can be associated. As described above, a morphological dictionary is obtained by associating not only characters but also a series of categories associated with character elements with morphemes.

The morphological extraction unit 1107 applies grammatical processing and the minimum number of clauses method to the generated network 2801,
"First sentence 2802-link L2803-node N2804
-The link "L2805-head 2806" is selected, and the morpheme "Osaka" is set as a morpheme with a high priority.

The output unit 1108 outputs "Osaka" as a recognition result. Thus, the character cutout unit 1102
Even if the extraction of the character area is not appropriate, the morphological dictionary 110 in which the correspondence between the category string and the morpheme is properly made.
By using 6, character strings can be correctly recognized.

(Specific Example 4) Next, the case where the handwritten character string 2901 shown in FIG. 29 is input will be described. As shown in FIG. 30, the character cutout unit 1102
001, 3002, and 3003 are cut out. The feature amount extraction unit 1103 extracts a feature amount from each character region. The category determination unit 1105 determines the upper layer 12 of the category dictionary 1104.
The similarity between each of the standard feature amounts 1201 and the extracted feature amount is calculated, and the category having the highest similarity is determined.
The category “C” is sequentially assigned from the character area 3001 of the first character.
1 ”,“ Cm ”, and“ Cn ”are determined.

FIG. 31 is a diagram showing the contents of the morphological dictionary 1106. "Noun / heaven," corresponding to the category "Cl"
The part-of-speech and morpheme of the morpheme of “husband, sentence, hole, element, and form” are registered, and “noun /
"Document" is registered, and "noun /
"Up" is registered. That is, in the morpheme corresponding to the category string of the morpheme dictionary 1105, a character code converted into a kanji when the hiragana is classified into each category constituting the category string is registered. In this specific example,
The hiragana “sho” corresponding to a part of the category column “CmCn” is converted into the kanji “sho”, and the category column “ClCmC
The morpheme “document” is registered corresponding to “n”.

The morpheme extracting unit 1106 determines the categories “Cl”, “Cm”,
A morpheme corresponding to a matching category in the morphological dictionary 1106 is extracted using “Cn” and a category string obtained by combining them as a search key.
01 is generated. "Sentence 3202-Link L3203-
Since there is no connected node on the route of the node N3204, it is rejected. Similarly, the route of “node N3205-link L3206-end of sentence 3207” is rejected because there is no link connected from the beginning of sentence 3202. Therefore,
"Sentence 3202-link L3208-node N3209
By applying grammar processing to the path of “−link L3210−end of sentence 3207”, the morpheme “document” accommodated in the node N3209 has a higher priority.

As a result, the output unit 1109 outputs “document” as a recognition result. As described above, in the present embodiment, the morpheme dictionary 1106 includes, as a morpheme, a character different from the character corresponding to the feature amount extracted from the coordinate data, and outputs this morpheme to form one of the handwritten character strings. If the part is written in Kana characters, it is converted to Kanji and output.

Next, the operation of this embodiment will be described with reference to FIG.
This will be described with reference to the flowchart of FIG. The coordinate input unit 1101 outputs the coordinate data of the handwritten character string drawn on the tablet to the character cutout unit (S330)
1). The character cutout unit 1102 is used to determine the stroke interval,
An area of each character is cut out from the stroke set using features such as pen-up time, character size, position, shape, aspect ratio, and character pitch (S3301).

Next, the characteristic amount extracting unit 1103 extracts the characteristic amount of the coordinate data of each cut-out character area (S3303). The category determining unit 1105 performs a DP matching process between the extracted character feature amount and each of the standard feature amounts 1201 in the upper layer 1202 of the category dictionary 1104 to obtain a category to which the character belongs for each character region (S3304). ).

The following steps S3305 to S3308 are the same as steps S1004 to S1007 of the first embodiment, and a description thereof will not be repeated.

In this embodiment, although the character cutout unit 1102 cuts out the character area based on the input coordinate data, when using the coordinate input unit 1101 in which a character input frame for each character is prepared in advance, The character cutout unit 1102 can be omitted.

Note that, in the embodiment of the present invention, 1
The category determination unit determines one category for each character,
A method of matching a column of this category with a morphological dictionary is used, but a category determining unit extracts a plurality of categories in descending order of priority, and matches the character position and a matrix of category candidates with a morphological dictionary. May be taken.

Note that the present invention is realized by a program, and is recorded on a recording medium such as a floppy disk and transferred, so that it can be easily implemented by another independent computer system.

[0151]

As described above, the present invention relates to a character recognition device for recognizing characters from data representing a character string, wherein characters or character elements having similar shapes are classified as the same category, and are classified into the same category. A category dictionary that stores a representative standard feature quantity representing a feature quantity extracted from a standard shape of a classified character or a character element and an identifier of the category in association with each other, and one or more characters or a plurality of characters A morpheme that is composed of elements and is the smallest linguistic unit;
A morphological dictionary for storing the identifiers of the categories to which the characters or character strings to which the morphemes belong or the identifier strings of the categories in association with each other, and a feature amount extraction for extracting a feature amount from the shape of each character representing the character string for each character Means, by comparing the feature quantity extracted by the feature quantity extraction means with the representative standard feature quantity stored in the category dictionary, and identifying a category identifier associated with the representative standard feature quantity having a high degree of similarity. A category deciding means for deciding, a candidate morpheme extracting means for extracting a plurality of candidate morphemes from the morpheme dictionary according to the category identifier decided by the category deciding means, and a plurality of candidate morphemes extracted by the candidate morpheme extracting means. And morpheme determining means for determining a priority morpheme based on a predetermined criterion. With such a configuration, each character constituting the character string is determined only as a category and does not recognize the character code.
It is possible to prevent character recognition errors caused by incorrectly obtaining candidate characters, and to increase the processing speed.

The morpheme of the morphological dictionary is composed of a character code, and the morpheme is stored for each attribute of the morpheme. The candidate morpheme extracting means stores the identifier of the category determined by the category determining means in the character Arrangement in the order of the characters of the data representing the columns, an identifier / identifier sequence acquisition unit for acquiring an identifier of a single category and an identifier sequence of a plurality of continuous categories, and an identifier of the category acquired by the identifier / identifier sequence acquisition unit; The morphological dictionary is searched using the identifier string of the category as a search key, and an extraction unit that extracts a morpheme associated with the identifier of the matching category and the identifier string of the category as a candidate morpheme, and the morpheme determination unit includes: The morpheme is determined based on the attribute of the morpheme. With such a configuration,
Since the priority morpheme is determined based on the attribute of the morpheme, the recognition accuracy can be improved.

The attribute of the morpheme stored in the morpheme dictionary is the part of speech of the morpheme, and the morpheme determining means converts the candidate morpheme extracted by the candidate morpheme extraction means from the beginning to the end of the sentence of the character string. An arrangement unit arranged on a network connected in the order of appearance of the identifier of the category or the identifier sequence of the category, and a rejection unit for rejecting candidate morphemes that do not conform to the grammar according to the part of speech of the candidate morpheme arranged in the arrangement unit And a selector for selecting a priority morpheme from candidate morphemes according to an empirical rule. With such a configuration, the priority morpheme is selected in accordance with the grammar check and the empirical rules by arranging the candidate morphemes on the network, thereby improving the accuracy of character recognition.

The category dictionary further stores a character code of a character classified into the category and a standard feature amount of the character corresponding to each category. When there are a plurality of morphemes to be compared, the standard feature quantity of each character constituting the morpheme stored in the category dictionary is compared with the feature quantity extracted by the feature quantity extraction means to obtain a similarity, The apparatus further includes a detailed recognition unit that determines one morpheme having a character with a high degree as its constituent character. With such a configuration, when there are a plurality of morphemes as the recognition result, the similarity at the character level constituting the morpheme is obtained, and one morpheme is determined as the recognition result. Only when characters cannot be determined, characters are simply recognized in accordance with the similarity in character units, and the processing speed can be increased.

Further, the character elements classified into the same category as the characters in the category dictionary are partial and adjacent characters constituting kanji, and the morphological dictionary combines the partial and adjacent characters in correspondence with the identifier string of the category. In this case, a morpheme including one character is stored. With such a configuration, even when the character area is cut out and the one character area is erroneously recognized, the character can be correctly recognized.

Further, the data representing the character string is a handwritten character string, and when the characteristic amount extracted by the characteristic amount extracting means is the characteristic amount of the hiragana character or the katakana character constituting the character string, It is assumed that the morpheme dictionary stores the identifiers of the categories corresponding to the hiragana characters or the morphemes of the category identifier column in Chinese characters. With such a configuration, even when a complex character is input in hiragana, a character string converted into kanji can be recognized.

Also, characters or character elements having similar shapes are classified as the same category, and a representative standard characteristic amount representing a characteristic amount extracted from a standard shape of the character or character element classified into the same category and the corresponding standard characteristic amount are described. A category dictionary that stores category identifiers in association with each other, a morpheme composed of one or more characters or a plurality of character elements, and a character code that is a minimum language unit, and a part of speech of the morpheme, A character recognition method for recognizing a character from data representing a character string, comprising: a category identifier to which the character string belongs or a morphological dictionary stored in association with the category identifier string. A feature amount extraction step of extracting a feature amount from a shape of each character; a feature amount extracted in the feature amount extraction step; A category determining step of determining a category identifier associated with a representative standard feature having a high similarity to the feature by comparing the representative standard feature stored in the document with the representative standard feature, and the category determining step. A candidate morpheme extraction step of extracting a plurality of candidate morphemes from the morphological dictionary according to the identifier of the category, and identifying the candidate morpheme extracted in the candidate morpheme extraction step from the beginning to the end of the sentence of the character string. Or, an arrangement step of arranging on a network connected in the order of appearance of the category identifier sequence, a rejection step of rejecting candidate morphemes that do not conform to the grammar according to the part of speech of the candidate morpheme arranged in the arrangement step, and a candidate morpheme according to an empirical rule. And a selecting step of selecting a priority morpheme from the above. With such a configuration, the accuracy of character recognition can be improved, and the processing speed can be increased.

Further, characters or character elements having similar shapes are classified as the same category, and a representative standard characteristic amount representing a characteristic amount extracted from a standard shape of the character or character element classified into the same category, and A category dictionary that stores category identifiers in association with each other, a morpheme composed of one or more characters or a plurality of character elements, and a character code that is a minimum language unit, and a part of speech of the morpheme, A computer-readable recording medium in which an identifier of a category to which a character string belongs or a morphological dictionary stored in association with the identifier sequence of the category is recorded, and a character recognition method for recognizing a character from data representing the character string is recorded. Extracting a feature value from a shape of each character representing the character string for each character; and extracting the feature value. A category that determines the identifier of the category associated with the representative standard feature having a high similarity to the feature by comparing the feature extracted in the step with the representative standard feature stored in the category dictionary. A determining step, a candidate morpheme extracting step of extracting a plurality of candidate morphemes from the morphological dictionary according to the category identifier determined in the category determining step, and extracting the candidate morpheme extracted in the candidate morpheme extracting step into the character string. An arrangement step of arranging on the network connected in the order of appearance of the identifier of the category or the identifier string of the category from the beginning of the sentence to the end of the sentence, and a candidate morpheme that does not conform to the grammar according to the part of speech of the candidate morpheme arranged in the arrangement step. A rejection step for rejection and a selection step for selecting a priority morpheme from candidate morphemes according to empirical rules. Tsu is the recording medium which records a program of character recognition method and a flop. With such a configuration, by using a computer having no character recognition function, the accuracy of character recognition can be improved and the processing speed can be increased.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a first embodiment of a character recognition device according to the present invention.

FIG. 2 is a diagram illustrating a structure of a category dictionary according to the embodiment.

FIG. 3 is a diagram illustrating an example of the contents of a morphological dictionary according to the embodiment.

FIG. 4 is a diagram illustrating an example of a network in which morphemes extracted by the morpheme extraction unit according to the embodiment are connected.

FIG. 5 is a diagram illustrating an example of a character string obtained by combining morphemes extracted by the morpheme extraction unit according to the embodiment.

FIG. 6 is a diagram illustrating an example of a network in which morphemes extracted by the morpheme extraction unit according to the embodiment are connected;

FIG. 7 is a diagram showing an example of a character string obtained by combining morphemes extracted by the morpheme extraction unit of the embodiment.

FIG. 8 is a diagram showing similarity rankings of characters constituting morphemes obtained by a detailed recognition unit according to the embodiment.

FIG. 9 is a diagram showing the similarity of characters constituting morphemes calculated by the detailed recognition unit of the embodiment.

FIG. 10 is a flowchart illustrating an operation of the embodiment.

FIG. 11 is a configuration diagram of a character recognition device according to a second embodiment of the present invention.

FIG. 12 is a diagram showing contents of a category dictionary according to the embodiment.

FIG. 13 is a diagram showing an example of the contents of a morphological dictionary according to the embodiment.

FIG. 14 is a diagram illustrating an example of a network generated by the morpheme extraction unit according to the embodiment.

FIG. 15 is a diagram illustrating an example of a handwritten character string input to the coordinate input unit according to the embodiment.

FIG. 16 is a diagram illustrating an example of a network generated by the morpheme extraction unit according to the embodiment.

FIG. 17 is a diagram showing similarity rankings of characters constituting morphemes obtained by the detailed recognition unit of the embodiment.

FIG. 18 is a diagram showing the similarity of characters constituting morphemes calculated by the detailed recognition unit of the embodiment.

FIG. 19 is a diagram illustrating an example of a handwritten character string input to the coordinate input unit according to the embodiment.

FIG. 20A is a diagram illustrating an example of a character region in which the handwritten character string illustrated in FIG. 15 is extracted by the character extraction unit according to the embodiment; FIG. 16B is a diagram illustrating an example of a character area in which the handwritten character string illustrated in FIG. 15 is extracted by the character extraction unit according to the embodiment.

FIG. 21A is a diagram showing the category of the character area shown in FIG. 20A determined by the category determining unit according to the embodiment. FIG. 20B is a diagram showing the category of the character area shown in FIG. 20B determined by the category determining unit of the embodiment.

FIG. 22 is a diagram showing an example of the contents of a morphological dictionary according to the embodiment.

FIG. 23 is a diagram illustrating an example of a network generated by the morpheme extraction unit according to the embodiment.

FIG. 24 is a diagram illustrating an example of a network generated by the morpheme extraction unit according to the embodiment.

FIG. 25 is a diagram illustrating an example of a handwritten character string input to the coordinate input unit according to the embodiment.

FIG. 26 is a diagram illustrating an example of a character area cut out by the character cutout unit according to the embodiment.

FIG. 27 is a diagram showing an example of the contents of the morphological dictionary according to the embodiment.

FIG. 28 is a diagram illustrating an example of a network generated by the morphological extraction unit of the embodiment.

FIG. 29 is a diagram illustrating an example of a handwritten character string input to the coordinate input unit according to the embodiment.

FIG. 30 is a diagram illustrating an example of a character area cut out by the character cutout unit according to the embodiment.

FIG. 31 is a diagram showing an example of the contents of the morphological dictionary according to the embodiment.

FIG. 32 is a diagram illustrating an example of a network generated by the morpheme extraction unit according to the embodiment.

FIG. 33 is a flowchart illustrating the operation of the above embodiment.

FIG. 34 is a configuration diagram of a first conventional character recognition device.

FIG. 35 is a diagram showing the structure of a conventional recognition dictionary.

FIG. 36 is a diagram showing an example of a conventional morphological dictionary.

FIG. 37 is a diagram illustrating a character string image input to the first character recognition device.

FIG. 38 is a diagram showing a character image cut out by a conventional character cutout unit.

FIG. 39 is a diagram showing the order of similarity of characters calculated by a conventional similarity calculator.

FIG. 40 is a flowchart illustrating an operation of the first character recognition device.

FIG. 41 is a configuration diagram of a first conventional character recognition device.

FIG. 42 is a diagram illustrating a handwritten character string input to the second character recognition device.

FIG. 43 is a diagram showing a character area cut out by a conventional character cutout unit.

FIG. 44 is a flowchart illustrating an operation of the second character recognition device.

[Explanation of symbols]

 101, 1102 Character extraction unit 102, 1103 Feature extraction unit 103, 1104 Category dictionary 104, 1105 Category determination unit 105, 1106 Morphological dictionary 106, 1107 Morphological extraction unit 107, 1108 Detailed recognition unit 1101 Coordinate input unit 1109 Output unit

Claims (8)

[Claims] 1. A character recognition apparatus for recognizing characters from data representing a character string, wherein characters or character elements having similar shapes are classified as the same category, and a standard character or character element classified into the same category is provided. A category dictionary that stores a representative standard feature quantity representing a feature quantity extracted from a simple shape and an identifier of the category in association with each other, and a morpheme that is composed of one or more characters or a plurality of character elements and is a minimum language unit And a morphological dictionary storing the identifiers of the categories or the identifier strings of the categories to which the characters or character strings constituting the morphemes are associated, and a feature of extracting a feature amount from the shape of each character representing the character strings for each character. A feature extraction unit that matches the feature amount extracted by the feature amount extraction unit with a representative standard feature amount stored in the category dictionary, Category deciding means for deciding an identifier of a category associated with a representative standard feature amount having a high value, and candidate morpheme extracting means for extracting a plurality of candidate morphemes from the morphological dictionary according to the category identifier decided by the category deciding means And a morpheme determining means for determining a priority morpheme based on a predetermined criterion from a plurality of candidate morphemes extracted by the candidate morpheme extraction means. 2. A morpheme of the morpheme dictionary is composed of a character code, and the morpheme is stored for each attribute of the morpheme. The candidate morpheme extracting means includes an identifier of the category determined by the category determining means, Arrangement in the order of the characters of the data representing the columns, an identifier / identifier sequence acquisition unit for acquiring an identifier of a single category and an identifier sequence of a plurality of continuous categories; and an identifier of the category acquired by the identifier / identifier sequence acquisition unit. The morpheme dictionary is searched using a category identifier string as a search key, and an extracting unit that extracts a morpheme associated with the identifier of the matching category and the identifier string of the category as a candidate morpheme, 2. The character recognition device according to claim 1, wherein a morpheme is determined based on the attribute of the morpheme. 3. The attribute of a morpheme stored in the morphological dictionary is a part of speech of the morpheme, and the morpheme determining means converts the candidate morpheme extracted by the candidate morpheme extracting means from the beginning to the end of the sentence of the character string. An arrangement unit arranged on the network connected in the order of appearance of the identifier of the category or the identifier sequence of the category, and a rejection unit for rejecting candidate morphemes that do not conform to the grammar according to the part of speech of the candidate morpheme arranged in the arrangement unit 3. The character recognition device according to claim 2, further comprising: a selection unit that selects a morpheme having priority from candidate morphemes according to an empirical rule. 4. The category dictionary further stores a character code of a character classified into the category and a standard feature amount of the character corresponding to each category, and stores the priority code selected by the selection unit. When there are a plurality of morphemes to be matched, the standard feature quantity of each character constituting the morpheme stored in the category dictionary is compared with the feature quantity extracted by the feature quantity extraction means to obtain a similarity, and the similarity is determined. 4. The character recognition device according to claim 3, further comprising a detailed recognition unit that determines one morpheme having a character with a high degree as a constituent character. 5. A character element classified into the same category as a character in the category dictionary is a partial and a partial constituting a kanji, and the morphological dictionary combines the partial and the partial according to an identifier string of a category. The character recognition device according to claim 1, wherein a morpheme including a character that becomes one character is stored. 6. When the data representing the character string is a handwritten character string, and the feature quantity extracted by the feature quantity extracting means is a feature quantity of a hiragana character or a katakana character constituting the character string, 6. The character recognition device according to claim 1, wherein the morpheme dictionary stores the identifiers of the categories corresponding to the hiragana characters or the morphemes in the category identifier column in kanji notation. 7. A character or character element having a similar shape is classified as the same category, and a representative standard characteristic amount representing a characteristic amount extracted from a standard shape of the character or character element classified into the same category, A category dictionary that stores category identifiers in association with each other, a morpheme composed of one or more characters or a plurality of character elements, and a character code that is a minimum language unit, and a part of speech of the morpheme, A character recognition method for recognizing a character from data representing a character string, comprising a category identifier to which the character string belongs or a morphological dictionary stored in association with the category identifier string, comprising: A feature value extraction step of extracting a feature value from a shape of each character; a feature value extracted in the feature value extraction step; A category determining step of comparing the stored representative standard feature quantity to determine a category identifier associated with the representative standard feature quantity having a high degree of similarity to the feature quantity; and a category determined in the category determining step. A candidate morpheme extraction step of extracting a plurality of candidate morphemes from the morphological dictionary according to the identifier of the morphological dictionary; and extracting the candidate morpheme extracted in the candidate morpheme extraction step from the beginning of the character string to the end of the sentence. An arrangement step of arranging the candidate morphemes on the network connected in the order of appearance of the identifier sequence, a rejection step of rejecting a candidate morpheme that does not conform to the grammar according to the part of speech of the candidate morpheme arranged in the arrangement step, And a selecting step of selecting a morpheme to be performed. 8. A character or character element having a similar shape is classified as the same category, and a representative standard characteristic amount representing a characteristic amount extracted from a standard shape of the character or character element classified into the same category, A category dictionary that stores category identifiers in association with each other, a morpheme composed of one or more characters or a plurality of character elements, and a character code that is a minimum language unit, and a part of speech of the morpheme, A computer-readable recording medium in which an identifier of a category to which a character string belongs or a morphological dictionary stored in association with the identifier sequence of the category is recorded, and a character recognition method for recognizing a character from data representing the character string is recorded. A feature amount extracting step of extracting a feature amount from a shape of each character representing the character string for each character; Category that matches the feature quantity extracted in the step with the representative standard feature quantity stored in the category dictionary to determine the identifier of the category associated with the representative standard feature quantity having a high similarity to the feature quantity A determining step; a candidate morpheme extracting step of extracting a plurality of candidate morphemes from the morphological dictionary according to the category identifier determined in the category determining step; and extracting the candidate morpheme extracted in the candidate morpheme extracting step into the character string. An arranging step of arranging on the network connected in the order of appearance of the identifier of the category or the identifier sequence of the category from the beginning of the sentence to the end of the sentence; A rejection step of rejecting, and a selection step of selecting a priority morpheme from candidate morphemes according to rules of thumb Recording medium for recording a program of a character recognition method and a flop.