JP2010237909A - Knowledge correction program, knowledge correcting device and knowledge correction method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve the specifying precision of a proper character string even when characters are separately recognized. <P>SOLUTION: A misreading candidate information storage means 1a stores misreading candidate information defining a misreading candidate character string including one or more separate character strings obtained by separating characters included in each of a plurality of correction character strings into a plurality of characters in association with each correction character string. A recognition result character string generation means 1b generates a recognition result character string 3 as the candidate of the character string included in image information 2. A character string comparison means 1c specifies the misreading candidate character string which is the most matched with the recognition result character string 3 from among the respective misreading candidate character strings by referring to the misreading candidate information stored in the misreading candidate information storage means 1a, and also specifies a corrected character string 4 corresponding to the misreading candidate character string. An output means 1d outputs the corrected character string 4 specified by the character string comparison means 1c. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a knowledge correction program for recognizing a character string included in image information, a knowledge correction apparatus, and a knowledge correction method.

  Conventionally, a character recognition device that identifies a character string included in an image acquired by imaging a form has been used. By using the character recognition device, it is possible to efficiently acquire the character code data of the character string entered in the form. For example, at a financial institution, information such as an account number, name, and address entered by a customer in a form is read by a character recognition device. The read information can be used as input data for financial transactions.

The character recognition device specifies a character string included in the image as follows, for example.
The character recognition device stores in advance a template for a target character to be read. When the character recognition device acquires an image to be recognized as a character string, the character recognition device extracts characters included in the image in character units. Then, the character recognition device acquires character features (for example, line inclination, shape, curvature, area, etc.) extracted from the image. The character recognition device collates this feature with the character feature of the template to identify the extracted character. The character recognition device sequentially performs such processing for each character included in the image to identify the character string entered by the writer.

  Here, in character recognition, if there is a misrecognition of a character, it is necessary for an operator to visually confirm it in order to obtain correct information, and information acquisition efficiency is reduced. For this reason, it is desired to reduce erroneous recognition of characters.

  On the other hand, for example, a character string that can be entered in a predetermined item (for example, an address) is registered in advance, and this character string (hereinafter referred to as a registered character string) and a character string candidate read from an image (hereinafter referred to as “character string”). A method (knowledge correction) for identifying a character string entered in a form by comparing with a recognition result character string (see, for example, Patent Document 1) is known. As a specific method, there is known a method of identifying a character string by collating a registered character string generated by combining a financial institution name and a branch name of the corresponding financial institution with a recognition result character string ( For example, see Patent Document 2).

JP 04-188383 A JP 09-09731 2 A

  However, even if the methods described in Patent Documents 1 and 2 are used, the character string specifying accuracy decreases as the content intended by the writer at the time of entry differs from the recognition result character string. For example, even though the writer has intentionally entered the character string “read”, “word-of-mouth” may be acquired as the recognition result character string because of the wide interval between the radicals. Conceivable. In this case, even if “read” exists in the registered character string, the number of characters is different (the recognition result character string is 4 characters, whereas the registered character string is 2 characters), and the character characteristics themselves Therefore, it becomes difficult to specify an appropriate character string.

  The present invention has been made in view of such a point, and even when characters are separated and recognized, a knowledge correction program, a knowledge correction device, and a knowledge correction program capable of improving the accuracy of specifying an appropriate character string An object is to provide a knowledge correction method.

  In order to solve the above problems, a knowledge correction program is provided. A computer that executes the knowledge correction program functions as a character string comparison unit and an output unit. The character string comparing means is defined as misread candidate information in which a misread candidate character string including one or more separated character strings obtained by separating a character included in each of the plurality of corrected character strings into a plurality of characters is associated with each corrected character string. The recognition result generated as a candidate for the character string included in the image information by the recognition result character string generation means among the misread candidate character strings, with reference to the misread candidate information stored in the misread candidate information storage means The misreading candidate character string that most closely matches the character string is specified, and the correction character string corresponding to the misreading candidate character string is specified. The output means outputs the corrected character string specified by the character string comparison means.

  In order to solve the above problem, a knowledge correction apparatus having the same function as a computer that executes the knowledge correction program is provided. Moreover, in order to solve the said subject, the knowledge correction method which performs the process similar to the computer which runs the said knowledge correction program is provided.

  According to the knowledge correction program, the knowledge correction apparatus, and the knowledge correction method, it is possible to improve the accuracy of specifying a proper character string even when characters are separated and read.

It is a figure which shows the outline | summary of this Embodiment. It is a figure which shows the hardware constitutions of a computer. It is a figure which shows the function structure of a computer. It is a figure which illustrates a form image. It is a figure which shows the example of the table which a misreading candidate information storage part memorize | stores. It is a figure which shows the example of a data structure of a prefecture table. It is a figure which shows the example of a data structure of a recognition result character string table. It is a figure which shows the 1st data structure example of a comparison character string table. It is a figure which shows the example of a data structure of an alternative recognition result character string table. It is a figure which shows the 2nd data structure example of a comparison character string table. It is a figure which shows the example of a data structure of an adjustment constant table. It is a figure which shows the example of a data structure of an accuracy definition table. It is a flowchart which shows the procedure of a character recognition process. It is a 1st flowchart which shows the procedure of a comparison character string production | generation process. It is a 2nd flowchart which shows the procedure of a comparison character string production | generation process. It is a flowchart which shows the procedure of the calculation process of alternative difference total. It is a figure which shows the 1st specific example of an evaluation result. It is a figure which shows the 2nd specific example of an evaluation result. It is a figure which shows the example of a display of a character recognition result confirmation window.

Hereinafter, the present embodiment will be described in detail with reference to the drawings.
FIG. 1 is a diagram showing an outline of the present embodiment. The computer 1 executes a knowledge correction program and recognizes a character string included in the image information 2. The image information 2 is generated, for example, by imaging a form with an imaging device. The computer 1 includes misreading candidate information storage means 1a, recognition result character string generation means 1b, character string comparison means 1c, and output means 1d.

  The misreading candidate information storage unit 1a defines a misreading candidate character string including one or more separated character strings obtained by separating a character included in each of a plurality of corrected character strings into a plurality of characters, in association with each corrected character string. Candidate information is stored.

The recognition result character string generation unit 1 b generates a recognition result character string 3 as a character string candidate included in the image information 2.
The character string comparison unit 1c refers to the misread candidate information stored in the misread candidate information storage unit 1a, and recognizes each misread registered in the recognition result character string 3 generated by the recognition result character string generating unit 1b and the misread candidate information. Compare and evaluate candidate strings. Then, the character string comparison unit 1c specifies the misreading candidate character string that most closely matches the recognition result character string 3, and specifies the correction character string 4 corresponding to the misreading candidate character string.

The output unit 1d outputs the corrected character string 4 specified by the character string comparison unit 1c.
According to the computer 1, the character string comparing unit 1c refers to the misreading candidate information stored in the misreading candidate information storage unit 1a, and the recognition result character string generating unit 1b generated among the misreading candidate character strings. The misreading candidate character string that most closely matches the result character string 3 is specified. The corrected character string 4 corresponding to the specified misreading candidate character string is specified by the character string comparison unit 1c based on the misreading candidate information. The corrected character string 4 specified by the character string comparison unit 1c is output by the output unit 1d.

  Thereby, even when characters are separated and read, it is possible to improve the accuracy of character string identification. Specifically, the misreading candidate information stored in the misreading candidate information storage unit 1a is associated with a character string “Kanagawa” that is the corrected character string 4, for example, “Kanano 11”, “Kana 1 Li”, or the like. The misreading candidate character string is defined. In this example, “No 11” and “1 li” correspond to the separation character string for “river”. At this time, it is assumed that the recognition result character string generation unit 1 b generates, for example, “Kanna 111” as the recognition result character string 3. In this case, the character string comparison unit 1c refers to the misreading candidate information stored in the misreading candidate information storage unit 1a, for example, as the misreading candidate character string that most closely matches the recognition result character string 3 (“Kanna 111”). Specify “Kanano 11”. Then, the recognition result character string generation unit 1 b refers to the misreading candidate information and identifies the character string “Kanagawa” associated with the misreading candidate character string “Kanano 11” as the corrected character string 4. As described above, the character string comparison unit 1c identifies the corrected character string with reference to the misreading candidate information stored in the misreading candidate information storage unit 1a, so that the separated character string is included in the recognition result character string. Even in this case, it is possible to improve the accuracy of specifying an appropriate correction character string. In the misreading candidate information storage unit 1a, a character string in which a character that is easily recognized as a separated character string among characters included in the corrected character string is replaced with a corresponding separated character string is registered in advance as a misreading candidate character string. The

  By the way, the computer 1 is useful when, for example, a window service such as a financial institution recognizes characters entered in a form by a customer and uses them as input data for financial transactions. Below, the case where the computer 1 is used for such an application is taken as an example, and it demonstrates more concretely.

  FIG. 2 is a diagram illustrating a hardware configuration of the computer. The computer 100 is installed at a window of a financial institution or the like, and is used for inputting form information by an operator. The computer 100 includes a CPU (Central Processing Unit) 101, a RAM (Random Access Memory) 102, an HDD 103, a graphic processing device 104, input interfaces 105 and 106, and a communication interface 107.

The CPU 101 controls the operation of the entire computer 100.
The RAM 102 temporarily stores at least part of an OS (Operating System) program and application software (hereinafter referred to as an application) program to be executed by the CPU 101. The RAM 102 stores various data necessary for processing by the CPU 101.

  The HDD 103 stores OS programs and application programs. The HDD 103 stores various data necessary for processing by the CPU 101. Instead of the HDD 103, for example, another nonvolatile storage device such as an SSD (Solid State Drive) can be used.

The graphic processing device 104 is connected to the monitor 11. The graphic processing device 104 displays an image on the screen of the monitor 11 in accordance with a command from the CPU 101.
The input interface 105 is connected to the keyboard 12 and the mouse 13. The input interface 105 transmits a signal sent from the keyboard 12 or the mouse 13 to the CPU 101.

  The input interface 106 is connected to the imaging device 14. The input interface 106 acquires a form image generated by the imaging device 14 imaging a form. The input interface 106 stores the form image acquired according to the processing of the CPU 101 in the RAM 102 or the HDD 103.

The communication interface 107 is connected to the network 20 and transmits / receives data to / from other information processing apparatuses.
FIG. 3 is a diagram illustrating a functional configuration of the computer. The computer 100 includes an image acquisition unit 110, a recognition result character string generation unit 120, a misreading candidate information storage unit 130, a control information storage unit 140, a comparison character string generation unit 150, an alternative dissimilarity calculation unit 160, and a corrected character string specification unit 170. And an output unit 180.

The image acquisition unit 110 acquires a form image generated by the imaging device 14 imaging a form. The image acquisition unit 110 outputs the acquired form image to the recognition result character string generation unit 120.
The recognition result character string generation unit 120 reads characters included in the form image acquired from the image acquisition unit 110 by item (prefecture name, first name, and the like), and generates a plurality of recognition result character strings based on the read characters. To do. The recognition result character string generation unit 120 generates a recognition result character string by the following procedure, for example.

  (1) The recognition result character string generation unit 120 reads a predetermined area considered to be one character in a character string of a predetermined item included in the form image, and acquires a characteristic for the corresponding character. Here, the feature is, for example, an evaluation of the position of a line included in the image, the degree of inclination of the line, or the like.

  (2) The recognition result character string generation unit 120 refers to a character template extracted from a form image with reference to a template of a character to be read (hereinafter referred to as a template character) stored in advance in a read character template storage unit (not shown). Compare the characteristics of each template character. As a result of the comparison, the recognition result character string generation unit 120 calculates a difference indicating the degree of discrepancy between the feature extracted from the form image and the feature of each template character. For example, a Bayesian method can be employed for calculating the degree of difference. The degree of difference is calculated as a value that is smaller as the value is smaller. That is, among the template characters, the lower the degree of difference, the higher the priority as a reading candidate.

  (3) The recognition result character string generation unit 120 generates a recognition result character string by sequentially combining characters with small differences calculated for each template character with respect to each character extracted from the form image. For example, a combination of those having the smallest difference for each character is the recognition result character string of the first candidate having the highest priority.

  In this way, the recognition result character string generation unit 120 generates a plurality of recognition result character strings from the characteristics of each character read from the form image. The recognition result character string generation unit 120 stores each generated recognition result character string in the control information storage unit 140.

  The misreading candidate information storage unit 130 stores misreading candidate information that defines a plurality of misreading candidate character strings in consideration of separated character strings that can be included in the recognition result character string. Here, the separated character string means that a character that should be recognized as one character has been recognized as a plurality of two or more characters because the interval between the radicals of the characters entered in the form is wide. Indicates a character string. For example, for the character “God”, a separated character string “Ne-Shen” can be considered. In addition, in the misreading candidate information, a correction character string that can be actually entered is registered in association with each misreading candidate character string in accordance with the item in which the corresponding recognition result character string is entered in the form. Here, the correction character string that can be actually entered is, for example, a character string that indicates the name of a prefecture that actually exists if the item is a name of a prefecture.

  The control information storage unit 140 stores the recognition result character string generated by the recognition result character string generation unit 120. Further, the control information storage unit 140 stores a comparison character string used for comparison with the recognition result character string generated by the comparison character string generation unit 150. In addition, the control information storage unit 140 stores control information used when comparing the plurality of misread candidate character strings extracted by the comparison character string generation unit 150 and the corresponding recognition result character strings. The control information includes various conditions used when numerically evaluating the comparison result.

  When the generation of the recognition result character string for each item by the recognition result character string generation unit 120 is completed, the comparison character string generation unit 150 selects one of the recognition result character strings for each item stored in the control information storage unit 140 (for example, , “Kanano 11” included in the item of prefecture. The comparison character string generation unit 150 refers to the misreading candidate information storage unit 130 and acquires misreading candidate information corresponding to the acquired recognition result character string. Further, the comparison character string generation unit 150 refers to the misreading candidate information stored in the misreading candidate information storage unit 130 and matches the number of characters included in the corresponding recognition result character string or the number of characters ± n (for example, n = 1) A plurality of misread candidate character strings that match within the range of Then, the comparison character string generation unit 150 generates a plurality of comparison character strings used for comparison with the recognition result character string based on the acquired misreading candidate character strings.

The comparison character string generation unit 150 generates a comparison character string corresponding to each of the following cases for each misreading candidate character string.
(First Case) The comparison character string generation unit 150 is a case where the number of valid characters of the recognition result character string is equal to the number of valid characters of the corresponding misreading candidate character string. In this case, the corresponding misreading candidate character string is set as a comparison character string.

  Here, the number of valid characters refers to the number of characters excluding the character or character string indicating the item (such as “city” or “prefecture”) among the character strings. Note that if the recognition result character string or the like does not include characters such as “city” or “prefecture”, the number of characters in the recognition result character string is the number of valid characters. The number of valid characters corresponds to “number of characters” in the claims.

  (Second case) The comparison character string generation unit 150 is a case where the number of valid characters of the recognition result character string is smaller than the number of valid characters of the corresponding misreading candidate character string. In this case, a character string in which characters are deleted from the misreading candidate character string by the amount corresponding to the character difference between the corresponding misreading candidate character string and the recognition result character string is generated as a comparison character string. For example, when the misreading candidate character string “Ne Shina 111” (the number of valid characters is 6) is extracted from the recognition result character string “Kanano 11” (the number of valid characters is 5), 6-5 =) The next character string excluding one character is generated. That is, “Senna 111” which is a character string from which the first character is removed, “Nenshina 11” which is a character string from which the last character is removed, and “Nen 111” which is a character string from which the middle character is removed. , “Ne-Shen 111” and “Ne-Sena 11” are generated as comparison character strings. In this way, when the number of valid characters in the misreading candidate character string is larger than the number of valid characters in the recognition result character string, the number of valid characters in the comparison character string is matched with the number of valid characters in the recognition result character string. This makes it possible to easily compare characters at corresponding positions in both character strings (for example, the second character from the beginning).

  (3rd case) The comparison character string production | generation part 150 is a case where the number of effective characters of a recognition result character string is larger than the number of effective characters of a corresponding misreading candidate character string. In this case, a character string obtained by deleting characters from the recognition result character string by the amount corresponding to the character difference between the corresponding misreading candidate character string and the recognition result character string is generated as an alternative recognition result character string. Then, the comparison character string generation unit 150 stores the generated alternative recognition result character strings in the control information storage unit 140. For example, when the misreading candidate character string “Kanna 11” (the number of effective characters is 4) is extracted from the recognition result character string “Kanano 11” (the number of effective characters is 5), -4 =) The next character string excluding one character is generated. That is, “Nano 11” which is a character string from which the first character is removed, “Kanano 1” which is a character string from which the last character is removed, and “Kamino 11” which is a character string from which the middle character is removed. , “Kanna 11” is generated as an alternative recognition result character string. In this case, the comparison character string generation unit 150 sets the corresponding misreading candidate character string as a comparison character string. In this way, when the number of valid characters in the recognition result character string is larger than the number of valid characters in the misreading candidate character string, the number of valid characters in the recognition result character string is matched with the number of valid characters in the comparison character string. As a result, as in the above (second case), it is possible to easily compare characters at corresponding positions in both character strings. The alternative recognition result character string need only be generated once for the corresponding character difference. For example, if an alternative recognition result character string corresponding to a difference of 1 in the number of effective characters is generated once for “Kanano 11”, the number of effective characters is 4 as a misreading candidate character string next by holding that information. When a character string is extracted, there is no need to generate a substitute recognition character string.

  (4th case) The comparison character string production | generation part 150 is a case where the number of effective characters of a recognition result character string is smaller than the number of effective characters of the correction | amendment character string corresponding to a corresponding misreading candidate character string. In this case, a character string excluding characters from the corrected character string by the amount corresponding to the character difference between the corresponding corrected character string and the recognition result character string is generated as a comparison character string. For example, when the effective character part of the recognition result character string is “Shingo” (the number of effective characters is 2), the correction character string corresponding to the misreading candidate character string is “Kanagawa” (the number of effective characters is 3). As a character string, the next character string obtained by removing (3-2 =) one character from “Kanagawa” is generated. In other words, “Nagawa”, the character string with the leading character removed, “Kana”, the character string with the trailing character removed, and “Kamikawa”, the character string with the middle character removed, are generated as comparison character strings. Is done. As an example of the integrated character, for the recognition target character “Kanagawa”, the “go” (integrated character in which “Nagawa” is integrated) included in the recognition result character string “Kango” exemplified above may be considered. . In this way, when the number of valid characters in the correction character string is larger than the number of valid characters in the recognition result character string, a character string in which the number of valid characters in the comparison character string matches the number of valid characters in the recognition result character string is generated. . For this reason, even when an integrated character is included, the matching probability between the recognition result character string and the comparison character string is improved. Thereby, even when the recognition result character string includes an integrated character in which two characters are integrated into one character, correction for the integrated character can be performed.

  The comparison character string generation unit 150 generates each comparison character string, the number of comparison characters of each comparison character string (subtracting the number of characters deleted from the number of valid characters), and a correction character string corresponding to the corresponding comparison character string Are stored in the control information storage unit 140.

  When the comparison character string generation unit 150 completes the generation of the comparison character string, the alternative dissimilarity calculation unit 160 acquires the number of comparison characters corresponding to each comparison character string stored in the control information storage unit 140. Then, the alternative difference calculation unit 160 compares the recognition result character string (or the alternative recognition result character string) with each comparison character string. As a result of the comparison, the alternative dissimilarity calculation unit 160 includes the difference calculated by the recognition result character string generation unit 120 for each character included in the recognition result character string, and the control information stored in the control information storage unit 140. Based on this, the alternative difference degree of each character of each comparison character string with respect to each character of the recognition result character string (or alternative recognition result character string) is determined. Here, the alternative dissimilarity is calculated based on the dissimilarity calculated by the recognition result character string generation unit 120 and the control information stored in the control information storage unit 140. It is a value indicating the degree of discrepancy between a character and each character of each comparison character string. That is, the alternative dissimilarity is a value indicating the degree of discrepancy between the two character strings in “character” units. The degree of difference is calculated as a value that is smaller as the value is smaller.

  Further, the substitution difference calculation unit 160 is a value (evaluation value) indicating the degree of discrepancy of each comparison character string with respect to the recognition result character string (or the substitution recognition result character string) based on the calculated substitution difference of each character. Is calculated. That is, the evaluation value is a value indicating the degree of discrepancy in units of “character string”. An evaluation value is calculated | required as a thing with a small degree of discrepancy as a value is small. Specifically, first, the alternative dissimilarity calculation unit 160 obtains the sum of the alternative dissimilarities of the characters included in the comparison character string. Then, the alternative dissimilarity calculation unit 160 uses the quotient obtained by dividing the sum of the alternative dissimilarities by the number of comparison characters corresponding to the corresponding comparison character string as the evaluation value of the comparison character string. The substitution difference calculation unit 160 outputs the calculated evaluation value of each comparison character string to the corrected character string specifying unit 170.

  Based on the evaluation value acquired from the alternative dissimilarity calculation unit 160, the corrected character string specifying unit 170 specifies the misreading candidate character string of the first candidate having the highest priority (that is, the lowest evaluation value). The corrected character string specifying unit 170 refers to the misreading candidate information stored in the misreading candidate information storage unit 130 and acquires a corrected character string corresponding to the specified misreading candidate character string of the first candidate. The corrected character string specifying unit 170 outputs the acquired corrected character string to the output unit 180.

The output unit 180 causes the monitor 11 to display information indicating the corrected character string acquired from the corrected character string specifying unit 170.
Note that the processing of the comparison character string generation unit 150, the alternative dissimilarity calculation unit 160, and the corrected character string specification unit 170 is sequentially executed for each item. For example, first, the correction character string is specified by the above-described units for the item indicating the prefecture name. Next, a correction character string is specified by the above-described units for the item indicating the city name. Further, next, the correction character string is specified by each of the above sections for the item indicating the district / address.

In this way, the computer 100 identifies the correction character string for all items and notifies the operator of the result.
FIG. 4 is a diagram illustrating a form image. The form image 200 is provided with areas 211, 212, 221, 222, 231, 232,. The form image 200 is generated when a customer enters a character string in a form installed in a financial institution and the imaging device 14 images the form. Then, the image acquisition unit 110 acquires the form image 200 from the imaging device 14. The image acquisition unit 110 outputs the acquired form image 200 to the recognition result character string generation unit 120.

The area 211 is an area in which a character string indicating a prefecture name is entered.
The area 212 is an area in which a character string indicating a city name is entered.
The area 221 is an area where a character string indicating a surname is entered.

The area 222 is an area in which a character string indicating a name is entered.
The area 231 is an area in which a character string indicating a bank name is entered.
The area 232 is an area in which a character string indicating a branch name is entered.

  The recognition result character string generation unit 120 determines which item of each recognition result character string depends on the read position of the regions 211, 212, 221, 222, 231, 232,. Can be identified. The recognition result character string generation unit 120 identifies the character string entered in the area 211 as corresponding to the item of the prefecture. In addition, the recognition result character string generation unit 120 identifies the character string entered in the area 212 as corresponding to the item of city. Hereinafter, similarly, the recognition result character string generation unit 120 specifies items such as first and last names, bank names, and branch names.

  FIG. 5 is a diagram illustrating an example of a table stored in the misreading candidate information storage unit. In the misreading candidate information storage unit 130, a prefecture table 131, a municipal table 132, a district table 133, a surname table 134,. The prefecture table 131, the municipal table 132, the district table 133, the surname table 134,... Correspond to the misreading candidate information.

The prefecture table 131 is defined by associating a corrected character string that can be entered as a prefecture name with a misreading candidate character string that includes a separated character string for the prefecture name.
The municipality table 132 is defined by associating a corrected character string that can be entered as a municipality name with a misreading candidate character string that includes a separation character string for the municipality name.

The district table 133 defines a correction character string that can be entered as a district name and a misreading candidate character string that includes a separation character string for the district name in association with each other.
The surname table 134 is defined by associating a corrected character string that can be entered as a surname with a misreading candidate character string that includes a separated character string for the surname.

In the prefecture table 131, the municipality table 132, the district table 133, the surname table 134,..., Misreading candidate character strings including separated character strings are registered in advance.
FIG. 6 is a diagram illustrating an example of the data structure of the prefecture table. The prefecture table 131 includes an item indicating a misreading candidate character string, an item indicating a correction character string, and an item indicating the number of valid characters. Information arranged in the horizontal direction of each item is associated with each other, and indicates information regarding one misreading candidate character string. The municipal data table 132, the district table 133, the surname table 134,... Have the same data structure.

  In the item indicating the misreading candidate character string, the misreading candidate character string for the corresponding correction character string is set. In the item indicating the correction character string, a character string indicating the name of a prefecture that actually exists is set. In the item indicating the number of valid characters, a value indicating the number of characters in a portion of the misreading candidate character string excluding the character or character string (such as “city” or “prefecture”) indicating the classification is set.

  In the prefecture table 131, for example, information that the misreading candidate character string is “Ne Shina 111 prefecture”, the correction character string is “Kanagawa prefecture”, and the number of valid characters is “6” is set. This indicates that “Ne Shina 111 prefecture” can be included in the recognition result character string as a misreading candidate including the separation character string in the corrected character string “Kanagawa prefecture”. This indicates that the number of characters excluding “” is “6”.

  FIG. 7 is a diagram illustrating an example of the data structure of the recognition result character string table. The recognition result character string tables 141a, 141b, 141c,... Are generated by the recognition result character string generation unit 120 for each item included in the form image 200 and stored in the control information storage unit 140. In the following description, only the recognition result character string table 141a will be described. However, the recognition result character string tables 141b, 141c,.

  The recognition result character string table 141a is associated with an item indicating a prefecture. The recognition result character string table 141a is provided with items indicating priority and items indicating recognition result character strings. Information arranged in the horizontal direction of each item is associated with each other to indicate information related to one recognition result character string.

  In the item indicating the priority order, information indicating the priority of the corresponding recognition result character string is set. In the item indicating the recognition result character string, a recognition result character string read from the form image 200 is set.

  In the recognition result character string table 141a, for example, information that the priority is “first candidate” and the recognition result character string is “Kanano 11” is set. In addition, a dissimilarity indicating a difference between each character included in the recognition result character string and the template character is set. For example, the degree of difference “135” is set for the character “God”. Also, the difference degree “157” is set for the character “NA”. Thus, the dissimilarity is set for all characters included in the recognition result character string.

  Note that the recognition result character string generation unit 120 sets a high priority for the recognition result character string in which all the characters included in each recognition result character string are combined with those having the smallest difference.

  FIG. 8 is a diagram illustrating a first data structure example of the comparison character string table. The comparison character string table 142 a is generated by the comparison character string generation unit 150 and stored in the control information storage unit 140. The comparison character string table 142a exemplifies what is generated when the recognition result character string “Kanano 11” (number of valid characters 5) is acquired.

  The comparison character string table 142a includes an item indicating a comparison character string, an item indicating a correction character string, and an item indicating the number of comparison characters. Information arranged in the horizontal direction of each item is associated with each other to indicate information of one comparison character string.

  In the item indicating the comparison character string, a comparison character string for comparison with the recognition result character string is set. In the item indicating the correction character string, a correction character string corresponding to the corresponding comparison character string is set. In the item indicating the number of comparison characters, the number of comparison characters corresponding to the corresponding comparison character string is set.

  In the comparison character string table 142a, for example, information that the comparison character string is “(N) Shinna 111 prefecture”, the correction character string is “Kanagawa prefecture”, and the number of comparison characters is “6” is set. In this case, the comparison character string is generated corresponding to the case described in FIG. 3 (second case), and the first character of the misreading candidate character string is deleted. The character “ne” in parentheses is a deleted character and is not used as a comparison target. This indicates that “6-1 = 5” is obtained by subtracting one character excluded from the number of comparison characters from the number of valid characters.

  The other comparison character strings “Kan 111 prefecture” and “Kaori 111 prefecture” illustrated in the comparison character string table 142a correspond to (first case) and the misreading candidate character strings are directly adopted as the comparison character strings. Is. In this case, the value of the number of valid characters is set as the number of comparison characters as it is.

  FIG. 9 is a diagram illustrating a data structure example of the alternative recognition result character string table. The substitution recognition result table 142b is generated by the comparison character string generation unit 150 and stored in the control information storage unit 140. The alternative recognition result character string table 142b illustrates an example generated when “Kanano 11” is acquired as the recognition result character string.

  The substitution recognition result character string table 142b includes an item indicating the substitution recognition result character string and an item indicating the number of target characters. Information arranged in the horizontal direction of each item is associated with each other to indicate information of one alternative recognition result character string.

  In the item indicating the alternative recognition result character string, an alternative recognition result character string generated for the recognition result character string is set. In the item indicating the number of target characters, the number of characters of the corresponding alternative recognition result character string is set.

  In the alternative recognition result character string table 142b, for example, information that the alternative recognition result character string is “Nano 11” and the number of target characters is “4” is set. This corresponds to the case described in FIG. 3 (third case), and “Nano 11” excluding the first character “God” of the recognition result character string “Kanano 11” is generated as an alternative recognition result character string. It has been done. The number of characters in the alternative recognition result character string is “4”. Note that the recognition result character string “Kanano 11” includes two consecutive character strings “1”. Therefore, the comparison character string generation unit 150 generates two alternative recognition result character strings “Kanano 1”. Only one of them is given priority in the alternative recognition result character string table 142b. sign up.

  FIG. 10 is a diagram illustrating a second data structure example of the comparison character string table. The comparison character string table 142 c is generated by the comparison character string generation unit 150 and stored in the control information storage unit 140. The comparison character string table 142c exemplifies what is generated when “Shingo” is acquired as the recognition result character string.

Since the configuration of the comparison character string table 142c is the same as that of the comparison character string table 142a, the description thereof is omitted.
In the comparison character string table 142c, for example, information that the comparison character string is “Nagawa”, the correction character string is “Kanagawa”, and the number of comparison characters is “2” is set. In this example, the comparison character string is generated corresponding to the case described in FIG. 3 (fourth case), and the first character of the correction character string is excluded. At this time, a value obtained by subtracting the number of characters excluded from the number of characters in the correction character string is the number of comparison characters. Also, the comparison character string “Kanagawa” is generated in the same manner.

The other comparison character string “Ko 1 prefecture” exemplified in the comparison character string table 142c corresponds to (first case), and the misreading candidate character string is directly adopted as the comparison character string.
FIG. 11 is a diagram illustrating a data structure example of the adjustment constant table. The adjustment constant table 143 is stored in the control information storage unit 140 in advance. The adjustment constant table 143 includes an item indicating an item name and an item indicating an adjustment constant. Information arranged in the horizontal direction of each item is associated with each other to indicate information regarding one adjustment constant.

In the item indicating the item name, information indicating the item name of the adjustment constant is set. In the item indicating the adjustment constant, information indicating the value of the corresponding adjustment constant is set.
In the adjustment constant table 143, for example, the following item names and adjustment constants are set.

  (A) “Character mismatch difference degree (V)” is a substitution of a corresponding character when the corresponding character included in the misreading candidate character string is not included in the recognition result character string (or alternative recognition result character string). A difference value V is shown. As the value of V, for example, V = 1000 is set.

  (B) “One-character separation difference (W)” is the difference between the number of valid characters in the misread candidate character string and the number of valid characters in the recognition result character string, and is 1 for each character included in the misread candidate character string. A value W to be added to the total of the alternative dissimilarities is shown. For example, W = 1100 is set as the value of W. The comparison character string generation unit 150 may extract a misreading candidate character string that has a difference in the number of valid characters from the recognition result character string of two or more characters. In this case, a value reflecting that there is a difference of two characters with respect to the alternative dissimilarity is added to the alternative dissimilarity. In this case, it may be possible to add, for example, 10000 as a value larger than W to the total of alternative dissimilarities of each character.

  (C) “Character unit adjustment difference threshold (X)” and “Character adjustment difference difference threshold (Y)” are used when the corresponding character included in the misread candidate character string is included in the recognition result character string. Threshold values X and Y for determining whether or not to give a superior evaluation to the substitution difference degree of the corresponding character are shown. For example, by evaluating the value of the alternative dissimilarity of the corresponding character to be small, it can be handled as having a high accuracy of the corresponding character. The “character unit adjustment difference threshold (X)” is used when the difference calculated by the recognition result character string generation unit 120 with respect to the corresponding character of the recognition result character string (or the alternative recognition result character string) is X or less. , It can be given a favorable evaluation. As the value of X, for example, X = 200 is set. The “character unit adjustment difference difference threshold (Y)” is a corresponding position in the recognition result character string of the second candidate generated by the recognition result character string generation unit 120 for the corresponding character of the recognition result character string. It shows that a superior evaluation can be given when the difference in degree of difference from the letter is greater than or equal to Y. In this example, when both of the above conditions are satisfied, the substitution difference calculation unit 160 sets a value obtained by multiplying the difference calculated by the recognition result character string generation unit 120 by 1/10 to the substitution difference of the corresponding character. It shall be a degree. In this way, when the recognition result character string generation unit 120 generates the first candidate recognition result character string, the character having a large difference in difference from the character at the same character position in the other recognition result character strings is dominant. Can be handled.

  FIG. 12 is a diagram illustrating a data structure example of the accuracy definition table. The accuracy definition table 144 is stored in the control information storage unit 140 in advance. The accuracy definition table 144 includes an item indicating an evaluation value, an item indicating an evaluation value difference, and an item indicating the accuracy. Information arranged in the horizontal direction of each item is associated with each other to indicate information on one accuracy.

  Information indicating the range of the evaluation value is set in the item indicating the evaluation value. The item indicating the evaluation value difference is set with information indicating a range of evaluation value difference between the first candidate misreading candidate character string and the second candidate misreading candidate character string. In the item indicating the accuracy, information indicating the accuracy of the misreading candidate character string of the first candidate when the evaluation value range and the evaluation value difference are set is set.

  In the accuracy definition table 144, for example, information that the evaluation value is “600 or less”, the evaluation value difference is “50 or more”, and the accuracy is “high” is set. This indicates that when the evaluation value is “600 or less” and the evaluation value difference is “50 or more”, the accuracy of the misreading candidate character string of the corresponding first candidate is set to “high”.

The alternative difference calculation unit 160 can determine the accuracy of the misread candidate character string of the first candidate based on the accuracy definition table 144.
In the adjustment constant table 143 and the accuracy definition table 144, optimum parameters according to the use environment are set in advance.

Next, processing of the computer 100 having the above configuration will be described.
FIG. 13 is a flowchart illustrating a procedure of character recognition processing. In the following, the process illustrated in FIG. 13 will be described in order of step number.

  [Step S <b> 11] The image acquisition unit 110 acquires a form image 200 generated by the imaging device 14 capturing an image. The image acquisition unit 110 outputs the acquired form image 200 to the recognition result character string generation unit 120.

  [Step S12] The recognition result character string generation unit 120 reads the character strings written in the areas 211, 212, 221, 222, 231, 232,... Included in the form image 200 acquired from the image acquisition unit 110. , Recognition result character string tables 141a, 141b, 141c,... Are generated and stored in the control information storage unit 140.

  [Step S13] The comparison character string generation unit 150 identifies an item to be processed next. For example, the order of processing for each item is predetermined. Specifically, for each item included in the form image 200, if it is an address, it may be processed in the order of prefecture name, city name, district name, address, building name, and the like. In the case of a name, it may be possible to process in the order of last name and first name. Furthermore, in the case of a financial institution name, it is possible to process in the order of the financial institution name and the branch name.

  [Step S14] The comparison character string generation unit 150 refers to the recognition result character string table 141a stored in the control information storage unit 140, and recognizes the recognition result character string of the item to be processed (for example, “Kanano 11”). ) To get. Then, the comparison character string generation unit 150 acquires the number of valid characters of the extracted recognition result character string (for example, “5” for “Kanano 11”). For example, the comparison character string generation unit 150 can specify the item of each character string based on the position where the corresponding character string is recognized in the form image 200. For example, when the prefecture name and city name are recognized consecutively, such as “Sagamihara City, Kanagawa Prefecture”, the characters “prefecture” or “city” It is conceivable to identify the character “”, specify the name of the prefecture or city, and obtain the corresponding character string. In the case of a full name, if the first and last names are described consecutively, it may be possible to distinguish the first name and the last name by identifying a delimiter such as a space between the first name and the last name.

  [Step S15] The comparison character string generation unit 150 refers to the misreading candidate information (such as the prefecture table 131) stored in the misreading candidate information storage unit 130, and performs processing based on the number of valid characters acquired in Step S14. The misreading candidate character string corresponding to the item is extracted. Specifically, the comparison character string generation unit 150 extracts from the misreading candidate information the misreading candidate character string that has the number of valid characters in the range of the acquired number of valid characters ± n. As the value of n, an optimum value is determined in consideration of the accuracy and processing speed of subsequent processing. Here, n = 1 is selected as such a value of n. In this case, for example, for the recognition result character string with the number of valid characters “5”, the comparison character string generation unit 150 has the number of valid characters registered in the misreading candidate information “6”, “5”, “4”. Get misreading candidate character string group. Note that n may be an integer of 2 or more when the number of comparison targets is increased to further improve the accuracy of collation.

[Step S16] The comparison character string generation unit 150 acquires a correction character string group corresponding to the acquired misreading candidate character string group.
[Step S17] The comparison character string generation unit 150 generates a comparison character string table based on the acquired recognition result character string, misreading candidate character string group, and correction character string group. The comparison character string generation unit 150 stores the generated comparison character string table in the control information storage unit 140.

  [Step S18] The substitution difference calculation unit 160 calculates a substitution difference for each character of the recognition result character string for each character of each comparison character string included in the comparison character string table stored in the control information storage unit 140. To do. Then, the calculated alternative dissimilarity total T is obtained.

  [Step S19] The alternative difference calculation unit 160 calculates a quotient obtained by dividing the alternative difference total T for each comparison character string by the number of comparison characters corresponding to the corresponding comparison character string as an evaluation value of each comparison character string.

  [Step S20] The alternative dissimilarity calculation unit 160 determines the priority of each comparison character string based on the calculated evaluation value. The priorities are ranked as being superior as the evaluation value is smaller. The alternative difference calculation unit 160 outputs information on the evaluation value calculated for each comparison character string to the corrected character string specifying unit 170.

  [Step S21] The corrected character string specifying unit 170 acquires the first candidate comparison character string having the highest priority among the evaluation values of the comparison character strings calculated by the alternative dissimilarity calculation unit 160, and corresponds to this. Specify the correction character string. The corrected character string specifying unit 170 refers to the accuracy definition table 144 stored in the control information storage unit 140 and determines the accuracy of the specified corrected character string. The corrected character string specifying unit 170 outputs the specified correction character string and the corresponding accuracy to the output unit 180.

  [Step S22] The comparison character string generation unit 150 determines whether or not an item to be processed next exists among items to be processed. If it exists, the process proceeds to step S13. If not, the process moves to step S23. For example, the comparison character string generation unit 150 completes the process up to the last item in the predetermined order (for example, the last item in the series of prefecture name, city name, district name,...). The above determination can be made by detecting whether or not.

  [Step S23] The output unit 180 generates a screen for displaying a list of the corrected character strings and the accuracy acquired from the corrected character string specifying unit 170 on the monitor 11, and causes the monitor 11 to display the generated screen.

In this way, the computer 100 performs knowledge correction considering that the recognition result character string read from the form image 200 includes the separated character string.
FIG. 14 is a first flowchart illustrating the procedure of the comparison character string generation process. In the following, the process illustrated in FIG. 14 will be described in order of step number. The following process shows the process in step S17 in FIG. 13 in detail.

[Step S31] The comparison character string generation unit 150 extracts one misreading candidate character string for which no comparison character string has been generated from the misreading candidate character string group acquired in step S15 of FIG.
[Step S32] The comparison character string generation unit 150 compares the recognition result character string with the extracted misreading candidate character string.

  [Step S33] The comparison character string generation unit 150 determines whether the number of valid characters in the recognition result character string matches the number of valid characters in the misreading candidate character string. If they match, the process proceeds to step S34. If not, the process moves to step S35.

  [Step S34] The comparison character string generation unit 150 registers the extracted misread candidate character string as a comparison character string in the comparison character string table 142a stored in the control information storage unit 140. At this time, the comparison character string generation unit 150 registers the corresponding correction character string and the number of comparison characters of the comparison character string in the comparison character string table 142a for the corresponding comparison character string. Then, the process proceeds to step S44.

  [Step S35] The comparison character string generation unit 150 determines whether or not the number of valid characters in the misreading candidate character string is larger than the number of valid characters in the recognition result character string. If larger, the process proceeds to step S36. If not, the process moves to step S39.

  [Step S36] The comparison character string generation unit 150 acquires the difference between the number of valid characters in the recognition result character string and the number of valid characters in the misread candidate character string. Here, 1 is acquired as the difference in the number of valid characters.

[Step S <b> 37] The comparison character string generation unit 150 generates a character string excluding the difference characters acquired from the misread candidate character strings.
[Step S38] The comparison character string generation unit 150 registers the character string generated in step S37 as a comparison character string in the comparison character string table 142a. At this time, the comparison character string generation unit 150 registers the corresponding correction character string and the number of comparison characters of the comparison character string in the comparison character string table 142a for the corresponding comparison character string.

[Step S39] The comparison character string generation unit 150 acquires the difference in the number of valid characters between the recognition result character string and the misreading candidate character string.
[Step S40] The comparison character string generation unit 150 determines whether or not an alternative recognition result character string of the corresponding difference in the number of valid characters has not been generated. If not yet generated, the process proceeds to step S41. If it has already been generated, the process proceeds to step S43. For example, when the number of valid characters in the recognition result character string is 5 and the number of valid characters in the misreading candidate character string is 4, the comparison character string generation unit 150 stores the alternative recognition result character string table stored in the control information storage unit 140. With reference to 142b, the above determination can be made based on whether or not there is an alternative recognition result character string of the target character string “4”.

[Step S41] The comparison character string generation unit 150 generates a character string excluding the difference characters acquired from the recognition result character string.
[Step S42] The comparison character string generation unit 150 registers the generated character string as an alternative recognition result character string in the alternative recognition result character string table 142b stored in the control information storage unit 140.

  [Step S43] The comparison character string generation unit 150 registers the extracted misreading candidate character string as a comparison character string in the comparison character string table 142a stored in the control information storage unit 140. At this time, the comparison character string generation unit 150 registers the corresponding correction character string and the number of comparison characters of the comparison character string in the comparison character string table 142a for the corresponding comparison character string.

  [Step S44] The comparison character string generation unit 150 determines whether or not comparison character strings have been generated for all the misreading candidate character strings included in the misreading candidate character string group. If it has been generated, the process proceeds to step S45. If not generated, the process proceeds to step S31.

  FIG. 15 is a second flowchart illustrating the procedure of the comparison character string generation process. In the following, the process illustrated in FIG. 15 will be described in order of step number. The following processing is executed continuously from the processing in step S44 in FIG.

[Step S45] The comparison character string generation unit 150 extracts one unextracted correction character string from the correction character string group acquired in step S16 of FIG.
[Step S46] The comparison character string generation unit 150 compares the recognition result character string with the extracted corrected character string.

  [Step S47] The comparison character string generation unit 150 determines whether or not the number of valid characters in the correction character string is smaller than the number of valid characters in the recognition result character string. If it is smaller (including the case where the number of valid characters is the same), the process proceeds to step S51. If not, that is, if the number of valid characters in the corrected character string is greater than the number of valid characters in the recognition result character string, the process proceeds to step S48.

  [Step S48] The comparison character string generation unit 150 acquires the difference between the number of valid characters in the recognition result character string and the number of valid characters in the correction character string. For example, if the recognition result character string is “Shingo”, the number of effective characters is “2”, and if the correction character number is “Kanagawa”, the number of effective characters is “3”. The difference is “3-2 = 1”.

[Step S49] The comparison character string generation unit 150 generates a character string excluding the difference characters acquired from the corrected character string.
[Step S50] The comparison character string generation unit 150 registers the character string generated in step S49 in the comparison character string table 142a as a comparison character string. At this time, the comparison character string generation unit 150 registers the corresponding correction character string and the number of comparison characters of the comparison character string in the comparison character string table 142a for the corresponding comparison character string.

  [Step S51] The comparison character string generation unit 150 determines whether or not all correction character strings included in the correction character string group have been extracted. If all have been extracted, the process is complete. If not all have been extracted, the process proceeds to step S45.

  In this way, the computer 100 generates a comparison character string according to the difference in the number of valid characters between the recognition result character string and the misreading candidate character string and the difference in the number of valid characters between the recognition result character string and the correction character string. Then, the comparison character string and the recognition result character string are compared with each other at the same character position (for example, the second character from the top).

  When the comparison character string generation unit 150 attempts to register the comparison character string in the comparison character string table 142a in steps S34 and S38 of FIG. 14, the same comparison character string as the corresponding comparison character string has already been registered. Cases are also conceivable. In this case, only the already acquired character string is used as the comparison character string, and the comparison character string to be newly registered is discarded. For example, if “Ne Shina 11” has already been registered as a comparison character string, “Ne Shina 11” to be newly registered is discarded. Thereby, it is possible to prevent duplicate comparison processing from being performed for the same comparison character string.

  FIG. 16 is a flowchart showing the procedure of the calculation process of the total alternative difference. In the following, the process illustrated in FIG. 16 will be described in order of step number. The following process shows the process in step S18 in FIG. 13 in detail.

  [Step S61] The alternative dissimilarity calculation unit 160 extracts one comparison character string for which the alternative dissimilarity has not been calculated from the comparison character strings included in the comparison character string table 142a stored in the control information storage unit 140. .

[Step S62] The substitution difference calculation unit 160 acquires the number of comparison characters of the extracted comparison character string.
[Step S63] The substitution difference calculation unit 160 determines whether the number of valid characters in the recognition result character string is equal to the number of comparison characters. If equal, the process proceeds to step S64. If not equal, the process proceeds to step S65.

  [Step S64] The alternative dissimilarity calculation unit 160 extracts a recognition result character string as a comparison source character string to be compared with the comparison character string. Then, the process proceeds to step S66. That is, the subsequent processing is executed on the recognition result character string.

  [Step S65] The substitution difference calculation unit 160 substitutes the comparison character number equal to the comparison character number from the substitution recognition result character string table 142b stored in the control information storage unit 140 as a comparison source character string to be compared with the comparison character string. One recognition result character string is extracted.

[Step S66] The substitution difference calculation unit 160 sets the counter i = 0.
[Step S67] The substitution difference calculation unit 160 extracts the i-th character from the beginning of the characters included in the recognition result character string and the comparison character string. Then, it is determined whether or not both extracted characters match. If they match, the process proceeds to step S68. If not, the process moves to step S69.

  [Step S68] The alternative difference calculation unit 160 acquires the difference calculated by the recognition result character string generation unit 120 for the corresponding character as an alternative difference. The alternative dissimilarity calculation unit 160 can acquire the dissimilarity with reference to the recognition result character string table of the item to be processed stored in the control information storage unit 140.

  [Step S69] The substitution difference calculation unit 160 refers to the adjustment constant table 143 stored in the control information storage unit 140, and sets the substitution difference degree for the corresponding character as the character mismatch difference degree (V). In the example of FIG. 11, V = 1000.

  [Step S <b> 70] The alternative dissimilarity calculation unit 160 determines whether or not the acquired alternative dissimilarity is equal to or less than the character unit adjustment dissimilarity threshold (X) set in the adjustment constant table 143. If it is X or less, the process proceeds to step S71. If it is greater than X, the process proceeds to step S73. In the example of FIG. 11, X = 200.

  [Step S71] The substitution difference calculation unit 160 refers to the recognition result character string table of the item to be processed, and compares the corresponding character with the i-th character from the beginning of the second candidate recognition result character string. The difference degree difference is calculated. For example, in the recognition result character string table 141a, when comparing the second character from the head, the character “NA” included in the first candidate character string is compared with the “match” included in the second candidate character string. A difference degree difference “241−157 = 84” is calculated. Then, the alternative difference calculation unit 160 determines whether or not the calculated difference difference is equal to or greater than the character unit adjustment difference difference threshold (Y). If it is greater than or equal to Y, the process proceeds to step S72. If smaller than Y, the process proceeds to step S73. In the example of FIG. 11, Y = 100.

[Step S <b> 72] The substitution difference calculation unit 160 newly acquires a value obtained by multiplying the substitution difference obtained for the character by 1/10 as the substitution difference.
[Step S73] The alternative difference calculation unit 160 determines whether there is a character for which the alternative difference has not been calculated for the comparison character string extracted in step S61. If there is an uncalculated character, the process proceeds to step S74. If there is no uncalculated character, the process proceeds to step S75. For example, the alternative dissimilarity calculation unit 160 can make this determination by determining whether the value of the counter i is equal to the value of “(number of comparison characters) −1” of the corresponding comparison character string.

[Step S74] The substitution difference calculation unit 160 increments the counter i. Then, the process proceeds to step S67.
[Step S75] The alternative dissimilarity calculation unit 160 calculates the sum K of alternative dissimilarities acquired for each character included in the corresponding comparison character string.

  [Step S76] The substitution difference calculation unit 160 acquires valid characters of the recognition result character string. Also, the alternative difference calculation unit 160 refers to the comparison character string table 142a and acquires the number of valid characters of the corresponding comparison character string. Here, of the comparison character strings included in the comparison character string table 142a, for example, the number of effective characters of “Kan 111 prefecture” is “5”. In addition, the number of effective characters of “(N) Shina 111 prefecture” is “6”. The substitution difference calculation unit 160 determines whether or not there is a difference between the acquired numbers of valid characters. If there is a difference, the process proceeds to step S77. If there is no difference, the alternative dissimilarity calculation unit 160 sets the alternative dissimilarity total T = K, and the process proceeds to step S78.

  [Step S77] The substitution difference calculation unit 160 adds the one character separation difference (W) to the sum K of substitution differences. In the example of FIG. 11, W = 1100. The alternative dissimilarity calculation unit 160 sets this value as the alternative dissimilarity total T (= K + W). In this example, when there is a difference in the number of valid characters between the recognition result character string and the comparison character string, the difference is 1. Therefore, here, a value W reflecting that there is a difference of one character is added to T. However, when a comparison character string having a difference in the number of valid characters of 2 or more is extracted, it is conceivable to add a value corresponding to the difference.

  [Step S78] The substitution difference calculation unit 160 obtains the corresponding substitution difference total T for the substitution recognition result character string, and uses the number of the corresponding target characters registered in the substitution recognition result string table 142b. It is determined whether or not all alternative recognition result character strings have been calculated. If the corresponding alternative dissimilarity total T is calculated for the recognition result character string, or if all the alternative recognition result character strings have been calculated, the process proceeds to step S79. If there is an alternative recognition result character string that has been calculated and the target alternative recognition result character string is not yet calculated, the process proceeds to step S65.

  [Step S79] The alternative dissimilarity calculation unit 160 determines whether or not the alternative dissimilarity total T has been calculated for all the comparison character strings included in the comparison character string table 142a. If it has been calculated, the process is completed. If not calculated, the process proceeds to step S61.

  In this way, the computer 100 calculates the alternative dissimilarity total T for each comparison character string. At this time, the computer 100 compares the character at the same character position in the recognition result character string and the comparison character string, determines whether or not these characters match, and is stored in the control information storage unit 140. With reference to the adjustment constant table 143, the alternative dissimilarity is determined. Furthermore, in addition to the sum of the alternative dissimilarities for each character, an adjustment value corresponding to the difference between the number of effective characters in the recognition result character string and the effective character number corresponding to the comparison character string is added to calculate the total alternative dissimilarity. .

  If the alternative dissimilarity determined in step S69 is a value that is clearly larger than the character unit adjustment dissimilarity (X), the step S73 may be executed following the step S69.

  FIG. 17 is a diagram illustrating a first specific example of an evaluation result. The evaluation result table 310 is generated by the alternative difference calculation unit 160 and is output to the corrected character string specifying unit 170. The evaluation result table 310 exemplifies a case where the recognition result character string is “Kanano 11”.

  The evaluation result table 310 is provided with an item indicating the ranking, an item indicating the comparison source character string, an item indicating the comparison character string, an item indicating the total alternative dissimilarity, an item indicating the evaluation value, and an item indicating the correction character string. ing. Information regarding the evaluation result of one comparison character string is shown by associating pieces of information arranged in the horizontal direction of each item with each other.

  In the item indicating the rank, a rank indicating the priority obtained as a result of the comparison is set. The lower the corresponding evaluation value, the higher the priority. In the item indicating the comparison source character string, a recognition result character string or an alternative recognition result character string is set as a character string to be compared. In the item indicating the comparison character string, a comparison character string for comparison with the comparison source character string is set. In the item indicating the total alternative difference, a value of the total alternative difference is set. An evaluation value is set in the item indicating the evaluation value. In the item indicating the correction character string, a correction character string corresponding to the corresponding comparison character string is set.

  In the evaluation result table 310, for example, “Kan 111 prefecture” is shown as a comparison character string of rank “1”. The comparison source character string at this time indicates “Kanano 11”, which is the same character string as the recognition result character string. The alternative dissimilarity calculation unit 160 obtains an evaluation value of “Kan 111 prefecture” (however, “prefecture” is not a comparison target) for the comparison source character string “Kanano 11” as follows.

  (A1) The substitution difference calculation unit 160 compares the first characters. The character “God” in the comparison source character string matches the character “God” in the comparison character string. Therefore, the degree of difference “135” calculated by the recognition result character string generation unit 120 with respect to “God” of the comparison source character string is set as the alternative degree of difference with respect to “God” of the comparison character string. Note that the character match or mismatch can be determined, for example, based on whether the character codes match.

  (A2) The substitution difference calculation unit 160 compares the second characters. The character “NA” in the comparison source character string matches the character “NA” in the comparison character string. Therefore, the degree of difference “157” calculated by the recognition result character string generation unit 120 for “N” of the comparison source character string is set as the alternative degree of difference for “N” of the comparison character string.

  (A3) The substitution difference calculation unit 160 compares the third characters. The character “NO” in the comparison source character string does not match the character “1” in the comparison character string. Therefore, the alternative dissimilarity with respect to the third “1” in the comparison character string is assumed to be the character disagreement dissimilarity V = “1000”.

  (A4) The substitution difference calculation unit 160 compares the fourth characters. The fourth character “1” in the comparison source character string matches the fourth character “1” in the comparison character string. For this reason, the difference “102” calculated by the recognition result character string generation unit 120 for the fourth “1” of the comparison source character string is set as the alternative difference for the fourth “1” of the comparison character string. Here, referring to the recognition result character string table 141 a, the difference between the fourth characters of the second candidate “Shin 1 and 1” is different from the first candidate “Kanna 11”. 215−102 = 113 ″, which satisfies the character unit adjustment difference difference threshold (Y) = 100 or more. The alternative dissimilarity “102” satisfies the character unit adjustment dissimilarity threshold (X) = 200 or less. For this reason, the alternative dissimilarity calculation unit 160 newly acquires a value obtained by reducing the alternative dissimilarity 102 to 1/10 as an alternative dissimilarity. In other words, the alternative dissimilarity with respect to the fourth “1” is “(102/10) = 10” (decimal point truncation).

  (A5) The substitution difference calculation unit 160 compares the fifth characters. The fifth character “1” in the comparison source character string matches the fifth character “1” in the comparison character string. For this reason, the difference “108” calculated by the recognition result character string generation unit 120 for the fifth “1” of the comparison source character string is set as the alternative difference for the fifth “1” of the comparison character string.

(A6) The substitution difference calculation unit 160 adds up the substitution differences of the characters obtained in (a1) to (a5) to obtain a substitution difference total T = “1410”.
(A7) The alternative dissimilarity calculation unit 160 sets “282”, which is a quotient obtained by dividing the alternative dissimilarity total T by the number of comparison characters “5” corresponding to the comparison character string “Kan 111 prefecture” as an evaluation value.

  In the evaluation result table 310, for example, “Kan 11 prefecture” is shown as a comparison character string of rank “2”. The alternative dissimilarity calculation unit 160 obtains an evaluation value of “Kan 11 prefecture” (however, “prefecture” is not a comparison target) for the comparison source character string “Kan 11” as follows.

  (B1) The substitution difference calculation unit 160 compares the first characters. The character “God” in the comparison source character string matches the character “God” in the comparison character string. Therefore, the degree of difference “135” calculated by the recognition result character string generation unit 120 with respect to “God” of the comparison source character string is set as the alternative degree of difference with respect to “God” of the comparison character string.

  (B2) The alternative difference calculation unit 160 compares the second characters. The character “NA” in the comparison source character string matches the character “NA” in the comparison character string. Therefore, the degree of difference “157” calculated by the recognition result character string generation unit 120 for “N” of the comparison source character string is set as the alternative degree of difference for “N” of the comparison character string.

  (B3) The substitution difference calculation unit 160 compares the third characters. The third character “1” in the comparison source character string matches the third character “1” in the comparison character string. For this reason, the degree of difference “102” calculated by the recognition result character string generation unit 120 for the third “1” of the comparison source character string is set as the alternative difference degree for the third “1” of the comparison character string. Here, referring to the recognition result character string table 141 a, the difference between the fourth characters of the second candidate “Shin 1 and 1” is different from the first candidate “Kanna 11”. 215−102 = 113 ″, which satisfies the character unit adjustment difference difference threshold (Y) = 100 or more. The fourth character of the first candidate corresponds to the third “1” of the comparison source character string “Kanna 11”. The alternative dissimilarity “102” satisfies the character unit adjustment dissimilarity threshold (X) = 200 or less. For this reason, the alternative dissimilarity calculation unit 160 newly acquires a value obtained by reducing the alternative dissimilarity 102 to 1/10 as an alternative dissimilarity. In other words, the alternative dissimilarity with respect to the third “1” is “(102/10) = 10” (decimal point truncation).

  (B4) The substitution difference calculating unit 160 substitutes by summing the one-character separation difference W = 1100 reflecting the difference in the number of effective characters in the sum of substitution differences of the characters obtained in (b1) to (b3). The total difference T = “1402” is obtained.

(B5) The alternative dissimilarity calculation unit 160 sets “350”, which is a quotient obtained by dividing the alternative dissimilarity total T by the number of comparison characters “4” corresponding to the comparison character string “Kan 11 prefecture” as an evaluation value.
In the evaluation result table 310, for example, “(N) Shinna 111 prefecture” is shown as a comparison character string of rank “10”. The comparison source character string at this time indicates “Kanano 11”, which is the same character string as the comparison source character string. The alternative dissimilarity calculation unit 160 obtains an evaluation value of “(N) Shinna 111 prefecture” (however, “prefecture” is not subject to comparison) for the comparison source character string “Kanano 11” as follows. “(Ne) Shina 111 prefecture” is generated corresponding to the misreading candidate character string “Ne Shina 111 prefecture”, and “(Ne)” is not subject to comparison. It is shown that.

  (C1) The substitution difference calculation unit 160 compares the first characters. The characters “Sen” in the comparison source character string do not match the characters “God” in the comparison character string. For this reason, the alternative dissimilarity with respect to “comment” of the comparison source character string is assumed to be character mismatch dissimilarity V = “1000”.

  (C2) The substitution difference calculation unit 160 compares the second characters. The character “NA” in the comparison source character string matches the character “NA” in the comparison character string. Therefore, the degree of difference “157” calculated by the recognition result character string generation unit 120 for “N” of the comparison source character string is set as the alternative degree of difference for “N” of the comparison character string.

  (C3) The substitution difference calculation unit 160 compares the third characters. The character “NO” in the comparison source character string does not match the character “1” in the comparison character string. Therefore, the alternative dissimilarity with respect to the third “1” in the comparison character string is assumed to be the character disagreement dissimilarity V = “1000”.

  (C4) The substitution difference calculation unit 160 compares the fourth characters. The fourth character “1” in the comparison source character string matches the fourth character “1” in the comparison character string. For this reason, the difference “102” calculated by the recognition result character string generation unit 120 for the fourth “1” of the comparison source character string is set as the alternative difference for the fourth “1” of the comparison character string. Here, referring to the recognition result character string table 141 a, the difference between the fourth characters of the second candidate “Shin 1 and 1” is different from the first candidate “Kanna 11”. 215−102 = 113 ″, which satisfies the character unit adjustment difference difference threshold (Y) = 100 or more. The alternative dissimilarity “102” satisfies the character unit adjustment dissimilarity threshold (X) = 200 or less. For this reason, the alternative dissimilarity calculation unit 160 newly acquires a value obtained by reducing the alternative dissimilarity 102 to 1/10 as an alternative dissimilarity. That is, the alternative dissimilarity with respect to the fourth “1” is (102/10) = 10 (decimal point truncation).

  (C5) The substitution difference calculation unit 160 compares the fifth characters. The fifth character “1” in the comparison source character string matches the fifth character “1” in the comparison character string. For this reason, the difference “108” calculated by the recognition result character string generation unit 120 for the fifth “1” of the comparison source character string is set as the alternative difference for the fifth “1” of the comparison character string.

  (C6) The substitution difference calculation unit 160 replaces the sum of the substitution differences of each character obtained in (c1) to (c5) with a one-character separation difference W = 1100 that further reflects the difference in the number of effective characters. The total difference T = “3375” is obtained.

(C7) The alternative dissimilarity calculation unit 160 uses “562”, which is a quotient obtained by dividing the alternative dissimilarity total T by the number of comparison characters “5” corresponding to the comparison character string “(N) Shinna 11 prefecture” as an evaluation value. To do.
The alternative difference calculation unit 160 calculates the evaluation value of each comparison character string as described above, and sets the comparison character string having a smaller evaluation value in the evaluation result table 310 as a higher priority.

The evaluation result table 310 can identify the comparison character string having the highest priority for the recognition result character string including the separated character string.
FIG. 18 is a diagram illustrating a second specific example of the evaluation result. The evaluation result table 320 is generated by the alternative difference calculation unit 160 and is output to the corrected character string specifying unit 170. The evaluation result table 320 exemplifies a case where the recognition result character string is “Shingo”.

The configuration of the evaluation result table 320 is the same as that of the evaluation result table 310 shown in FIG.
In the evaluation result table 320, for example, “Kanagawa” is shown as a comparison character string of rank “1”. The alternative dissimilarity calculation unit 160 obtains an evaluation value of “Kan prefecture” (where “prefecture” is not a comparison target) for the recognition result character string “Shingo” as follows. Note that “Kan Prefecture” is generated by the comparison character string generation unit 150 because the number of effective characters “2” of the recognition result character string “Kango” is smaller than the number of effective characters “3” of the correction character string “Kanagawa”. It has been done.

  (D1) The substitution difference calculation unit 160 compares the first characters. The character “God” in the recognition result character string matches the character “God” in the comparison character string. For this reason, the degree of difference “135” calculated by the recognition result character string generation unit 120 for “god” of the recognition result character string is set as an alternative degree of difference for “god” of the comparison character string.

  (D2) The substitution difference calculation unit 160 compares the second characters. The character “go” in the recognition result character string does not match the character “na” in the comparison character string. For this reason, the alternative dissimilarity with respect to “N” of the recognition result character string is assumed to be character mismatch dissimilarity V = “1000”.

(D3) The substitution difference calculation unit 160 adds up the substitution differences of the characters obtained in (d1) and (d2) to obtain a substitution difference total T = “1135”.
(D4) The alternative dissimilarity calculation unit 160 sets “567”, which is a quotient obtained by dividing the alternative dissimilarity total T by the number of comparison characters “2” corresponding to the comparison character string “Kanagawa” as an evaluation value.

  Evaluation values are similarly calculated for the comparison character strings “Nagawa Prefecture” and “Kaori 1 Prefecture”. Here, since there is no character that matches the recognition result character string in both comparison character strings, the evaluation value is calculated as “1000”. For this reason, both comparison character strings have the same priority.

FIG. 19 is a diagram illustrating a display example of a character recognition result confirmation window. The character recognition result confirmation window 400 is generated by the output unit 180 and displayed on the monitor 11.
In the character recognition result confirmation window 400, a candidate display area 410, a confirmation button display area 420, and another candidate selection button display area 430 are provided.

  In the candidate display area 410, the corrected character string with the highest accuracy is displayed as a result of knowledge correction. At this time, information indicating the accuracy of knowledge correction regarding the corresponding correction character string is also displayed. Information indicating the accuracy is generated by the corrected character string specifying unit 170.

  For example, when the correction character string specifying unit 170 acquires the evaluation result table 310 from the alternative dissimilarity calculation unit 160, the correction character string specifying unit 170 acquires the correction character string “Kanagawa” corresponding to the comparison character string having the highest priority. Then, the corrected character string specifying unit 170 refers to the accuracy definition table 144 stored in the control information storage unit 140 and specifies the accuracy for the acquired corrected character string “Kanagawa Prefecture”. Specifically, the evaluation value for the first candidate “Kan 111 prefecture” is “282” and “350” or less, and the evaluation value for the second candidate “Kan 11 prefecture” is “562”. Since the difference is “350−282 = 68” and “50” or more, the accuracy is specified as “high”.

  In this way, the corrected character string specifying unit 170 specifies the accuracy of the corrected character string displayed in each item and notifies the output unit 180 of the accuracy. The output unit 180 includes accuracy information in the candidate display area 410 together with the corrected character string of each item acquired from the corrected character string specifying unit 170.

The confirmation button display area 420 is a button for confirming each item candidate displayed in the candidate display area 410 as input data.
The other candidate selection button display area 430 is a button for instructing the output unit 180 to display a character string different from the character string displayed in the candidate display area 410. Upon receiving this instruction, the output unit 180 acquires another correction character string from the correction character string specifying unit 170 and displays it on the monitor 11.

  The operator can check the character recognition result by browsing the character recognition result confirmation window 400 and performing an operation input to each button using the keyboard 12 or the mouse 13, and can re-correct if necessary.

  For example, the recognition result character string generation unit 120 generates “Kanano 11” as the recognition result character string, and the recognition result character string is not yet stored in the prefecture table 131 stored in the misreading candidate information storage unit 130. It may be a registration. In this case, it is conceivable that a misreading candidate character string “Kanano 11” is newly registered in the prefecture table 131 in association with the correction character string “Kanagawa” at the timing when the confirmation button is pressed. In this way, in the subsequent knowledge correction, “Kanano 11” is added to the comparison target, and the correction accuracy can be improved.

  As described above, according to the computer 100, knowledge correction is performed using misreading candidate information that takes into consideration a separated character string in which a single character is recognized by being separated into a plurality of characters. In the misreading candidate information, a misreading candidate character string including a separated character string is registered in association with the correction character string. The computer 100 can improve the identification accuracy of the corrected character string by comparing the misread candidate character string with the recognition result character string read from the form image 200.

  At this time, the computer 100 generates a comparison character string used for comparison with the recognition result character string according to the difference between the number of effective characters in the recognition result character string and the number of effective characters in the misreading candidate character string. Thereby, the comparison of the recognition result character string with respect to the comparison character string can be easily performed by comparing characters at the same character position (for example, the second position from the top).

  In addition, the degree of discrepancy (substitution difference) between the recognition result character string and the comparison character string is evaluated based on the match or mismatch for each character, the difference in the number of valid characters, the superiority to other candidates when generating the recognition result character string, and the like. . Thus, by evaluating the alternative dissimilarity in detail, a comparative character string with higher accuracy can be specified with high accuracy.

  In addition, when the number of valid characters in the correction character string corresponding to the misreading candidate character string is larger than the number of valid characters in the recognition result character string, the computer 100 excludes characters corresponding to the difference in the number of characters from the correction character string. The generated character string is generated as a comparison character string. Thereby, the correction | amendment with respect to the integrated character by which the several character was integrated into the single character can be performed.

  The functions that the computer 100 should have are realized by executing a knowledge correction program on the computer 100. The program describing the processing contents can be recorded on a recording medium readable by the computer 100. Examples of the recording medium readable by the computer 100 include a magnetic recording device, an optical disc, a magneto-optical recording medium, and a semiconductor memory.

  When the program is distributed, for example, a portable recording medium such as an optical disk on which the program is recorded is sold. It is also possible to store the program in a storage device of a server computer and transfer the program from the server computer to another computer via a network.

  The computer 100 stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. Then, the computer 100 reads the program from its own storage device and executes processing according to the program. The computer 100 can also read a program directly from a portable recording medium and execute processing according to the program. In addition, each time the program is transferred from the server computer, the computer 100 can sequentially execute processing according to the received program.

  As described above, the knowledge correction program, the knowledge correction apparatus, and the knowledge correction method of the present invention have been described based on the illustrated embodiments. However, the present invention is not limited thereto, and the configuration of each unit is an arbitrary configuration having the same function. Can be replaced. Moreover, other arbitrary structures and processes may be added. Further, the present invention may be a combination of any two or more configurations (features) of the above-described embodiments.

DESCRIPTION OF SYMBOLS 1 Computer 1a Misreading candidate information storage means 1b Recognition result character string production | generation means 1c Character string comparison means 1d Output means 2 Image information 3 Recognition result character string 4 Correction character string

Claims (10)

Computer
A misreading candidate that stores misreading candidate information in which a misreading candidate character string including one or more separated character strings obtained by separating a character included in each of a plurality of corrected character strings is associated with each correction character string. With reference to the misreading candidate information stored in the information storage means, among the misreading candidate character strings, the recognition result character string generated as a character string candidate included in the image information by the recognition result character string generating means. A character string comparison unit that identifies the misread candidate character string that most closely matches and identifies the correction character string corresponding to the misread candidate character string;
Output means for outputting the corrected character string specified by the character string comparison means;
Knowledge correction program characterized by functioning as   The character string comparison unit is configured to recognize the recognition candidate character string from among the misreading candidate character strings, which is the number of characters included in a character number range increased or decreased by a predetermined value with respect to the number of characters of the misreading candidate character string. The knowledge correction program according to claim 1, wherein the misreading candidate character string that most closely matches the result character string is specified.   The character string comparison unit compares the characters in the recognition result character string and the characters included in the misreading candidate character strings by comparing a match or mismatch between characters at the same character position, and Calculating a predetermined evaluation value indicating a degree of difference between the result character string and each of the misread candidate character strings, and identifying the misread candidate character string most closely matching the recognition result character string based on the evaluation value The knowledge correction program according to claim 2.   The character string comparison means calculates the evaluation value so that the degree of difference increases, and when there is a difference between the number of characters of the recognition result character string and the number of characters of the misread candidate character string, 4. The knowledge correction program according to claim 3, wherein a value corresponding to the difference in the number of characters is added to the evaluation value.   If the number of characters in the misreading candidate character string is larger than the number of characters in the recognition result character string, the character string comparison unit obtains the difference in the number of characters, and the same number as the difference in the number of characters from the misreading candidate character string. A plurality of first comparison character strings excluding the character is generated in association with the correction character string corresponding to the misreading candidate character string, and the evaluation value of the recognition result character string and each of the first comparison character strings The knowledge correction program according to claim 3, wherein the knowledge correction program is calculated.   If the number of characters in the misread candidate character string is smaller than the number of characters in the recognition result character string, the character string comparison unit obtains the difference in the number of characters, and the same number as the difference in the number of characters from the recognition result character string. A plurality of comparison source character strings excluding the character is generated in association with the recognition result character string, and the evaluation values of the comparison source character strings and the misreading candidate character strings are calculated. Item 6. The knowledge correction program according to any one of Items 3 to 5.   When the number of characters of the correction character string corresponding to the misread candidate character string is larger than the number of characters of the recognition result character string, the character string comparison unit obtains the difference in the number of characters and calculates the difference from the correction character string. A plurality of second comparison character strings excluding the same number of characters as the difference in the number of characters are generated in association with the correction character strings, and the evaluation values of the recognition result character strings and the respective second comparison character strings are calculated. The knowledge correction program according to any one of claims 3 to 6, characterized by:   When the output means obtains the corrected character string, the evaluation value calculated by the character string comparison means for the correction character string is referred to the accuracy definition information storage means for storing the accuracy information according to the evaluation value. The knowledge correction program according to any one of claims 3 to 7, wherein the accuracy information according to the information is acquired and the accuracy information is displayed together with the correction character string on a display device. A knowledge correction device for recognizing a character string included in image information,
A misreading candidate storing misreading candidate information in which a misreading candidate character string including one or more separated character strings obtained by separating a character included in each of a plurality of corrected character strings is associated with each correction character string. With reference to the misreading candidate information stored in the information storage means, among the misreading candidate character strings, the recognition result character string generated as a character string candidate included in the image information by the recognition result character string generating means. A character string comparison unit that identifies the misread candidate character string that most closely matches, and that identifies the correction character string corresponding to the misread candidate character string;
Output means for outputting the corrected character string specified by the character string comparison means;
A knowledge correction apparatus comprising: A knowledge correction method of a knowledge correction device for recognizing a character string included in image information,
A misreading candidate in which a character string comparison unit defines a misreading candidate character string including one or more separated character strings obtained by separating a character included in each of a plurality of corrected character strings into a plurality of characters, in association with each of the corrected character strings. With reference to the misreading candidate information stored in the misreading candidate information storing means for storing information, among the misreading candidate character strings, the recognition result character string generating means is generated as a character string candidate included in the image information. Identify the misreading candidate character string that most closely matches the recognition result character string, identify the correction character string corresponding to the misreading candidate character string,
An output unit that outputs the corrected character string specified by the character string comparison unit;
A knowledge correction method characterized by that.

Images