JP6024118B2 - Character reader and character recognition method - Google Patents

Description

  The present invention relates to a character reader and a character recognition method.

  In a business for a transfer request or the like in a financial institution, the efficiency of the business is achieved by recognizing information written in a form designated by a customer in advance by a character reader.

  At this time, for the purpose of improving recognition accuracy (reading accuracy), Patent Document 1 stores customer transaction history for each form, searches the transaction history for each form, and corrects the character recognition. Corresponding to the item group correction unit to be performed, the two item correction unit for correcting the character recognition by searching the dictionary of the item by pairing the related items in the form, and corresponding to each item of the form Character recognition processing comprising: an item correction unit that searches for a dictionary to perform correction of the character recognition; and a related item correction unit that searches for a related dictionary based on the item attribute for each item and corrects the character recognition An apparatus is disclosed.

  Further, Patent Document 2 discloses that the characteristics of the arrangement of words and expressions that appear in the same type of text as the recognition target are digitized in advance, and the probability of candidate characters is ranked. Discloses a character recognition device that corrects missing correct characters.

  Furthermore, in Patent Document 3, the recognition result is compared with a pre-registered word. If these match, the recognition result is correct, and if they do not match, the recognition result is incorrect. A character reading device that performs character recognition with high accuracy by deleting the characters is disclosed.

Japanese Patent Laid-Open No. 5-258099 JP-A-9-138840 JP 2001-101340 A

  However, in each of the above-mentioned patent documents, the corresponding information is searched from a general dictionary prepared in advance, and the recognition result is corrected based on the searched information, so that high-precision recognition correction cannot be performed in a short time. There was a case that could not be.

  That is, when general character information is recognized, there is no restriction on the character range to be recognized. And the dictionary used for correction | amendment of the recognition character which is a recognition result is also a general dictionary without the restriction | limiting.

  On the other hand, when recognizing the characters entered in the transfer request form at a financial institution, the transfer requester (bank user) transfers money using a specific bank (such as a bank near the user or a bank with a transaction). I often request. For this reason, the range of the recognition character is limited to the range described by the bank user. Since the format of the transfer request form is prescribed, the characters entered are in a very limited range such as name and address. Therefore, when the recognition target is a form character, the dictionary used for correcting the recognition result need not be a general dictionary.

  However, in each of the above-mentioned patent documents, since a general correction dictionary is used, the search range when searching for the corresponding information is widened, the search load is increased, and the searched information is not valid. In some cases, the probability that the information is appropriate increases, and the recognized character cannot be corrected with high accuracy.

  Accordingly, a main object of the present invention is to provide a character reading apparatus and a character recognition method capable of correcting a recognition result at high speed and with high accuracy.

  In order to solve the above problem, a character recognition device for recognizing handwritten characters from image data of handwritten characters includes a character recognition dictionary used for character recognition of handwritten characters, and information about a customer who has entered the handwritten characters. A dictionary unit including a customer dictionary to store, image data is character-recognized using a character recognition dictionary, and the customer dictionary is searched using information of recognized characters included in the character recognition, A recognition unit that corrects the recognized character based on the retrieved customer information.

  A character recognition method for recognizing handwritten characters from image data of handwritten characters includes a character recognition procedure for recognizing image data using a character recognition dictionary, and basic customer information about a customer who has entered handwritten characters. A search procedure for searching for customer basic information using a recognition character from a customer dictionary composed of the customer basic information associated with the customer basic information, and a customer dictionary using the recognition character for the customer included in the recognition character as a search keyword And a correction procedure for acquiring the basic customer information and correcting the recognized character based on the acquired basic customer information.

  According to the present invention, the image data is character-recognized using the character recognition dictionary, and the customer dictionary is searched using the recognition result related to the customer, and the recognition result is corrected based on the searched customer information. Recognition result correction processing can be performed at high speed and with high accuracy.

It is a block diagram of the character recognition system concerning the embodiment of the present invention. It is a figure which shows the rough format of a transfer request form. It is the figure which illustrated the customer information table. It is a flowchart which shows the process sequence in a pre-processing part and a recognition part. It is a flowchart which shows the process sequence in a search word collation process.

  An embodiment of the present invention will be described. FIG. 1 is a block diagram of a character recognition system 2 using a character recognition device 6 according to the present invention. The character recognition system 2 includes a character recognition device 6, a server 4, a terminal 5, and the like, which are connected by a network L or the like. The scanner 3 is directly connected to the character recognition device 6 by a cable.

  The scanner 3 can be connected to the character recognition device 6 via the network L. Moreover, although the case where the terminal 5 and the character recognition device 6 are provided separately is shown, these may be integrated.

  The scanner 3 acquires image data of a form such as a transfer request form. The acquired image data is assigned an identification number and stored in the server 4 via the character recognition device 6.

  FIG. 2 is a diagram illustrating a transfer request form as an example of a form. The form 20 is provided with a plurality of fields (entry fields) for the user to fill in by hand. Typical fields include bank name field 20a, bank name field 20b, payee surname field 20c, payee name field 20d, account number field 20e, transfer amount field 20f, client telephone number field 20g, client The surname field 20h, the client name field 20j, the client name field 20k, and the client address field 20m can be exemplified.

  As shown in FIG. 1, the character recognition device 6 includes a dictionary unit 11 including a character recognition dictionary 11a and a customer dictionary 11b, and a recognition unit 12 including a preprocessing unit 15, a recognition unit 16, and a knowledge processing unit 17.

  The character recognition dictionary 11a in the dictionary unit 11 is a database composed of feature vectors (registered feature vectors) indicating character features. This registered feature vector is a numerical data string obtained by tracing the outline of a character image and extracting character features.

  The customer dictionary 11b is a customer information database composed of client information (customer basic information) and recipient information (customer incidental information) associated with the client information.

  FIG. 3 is a diagram in which such a customer information database is displayed as a customer information table 30. The customer information table 30 includes information such as customer information 31, transaction history 32, last name information 33, name information 34, financial institution name information 35, address information 36, and the like. The surname information 33, the first name information 34, and the address information 36 correspond to client information, and the financial institution name information 35 and the transaction history 32 correspond to client information. The surname information 33, the name information 34, the financial institution name information 35, and the address information 36 are associated with each item of the customer information 31, and the customer information 31 and the transaction information 32 are associated with each other.

  The preprocessing unit 15 in the recognition unit 12 performs a position correction process 15a, a character image cutout process 15b, a binarization process 15c, and a character image correction process 15d. The recognition unit 16 performs a feature extraction process 16a, a recognition process 16b, and a rejection determination process 16c. Then, the knowledge processing unit 17 performs a search word matching process 17a.

  FIG. 4 is a flowchart showing a processing procedure in the preprocessing unit 15 and the recognition unit 16.

  Step S1: The recognition unit 12 performs a position correction process 15a. This position correction process 15a detects information such as a print position shift / form cut error using the established information such as the reference mark and ruled line position in the image data, and corrects the position of the field to be recognized based on the detection result. It is a process to perform.

  Step S2: Next, a character image cutting process 15b is performed. The character image cutout process 15b is a process of cutting out image data corresponding to a field to be recognized from the image data position-corrected by the position correction process 15a and acquiring a character image. The image data is cut out in segment units in the length direction (horizontal direction) of the field (for example, in the case of Chinese characters, a unit corresponding to “hen” or “making”). Then, a character image is generated by concatenating images corresponding to a maximum of three adjacent segments. Note that the maximum number of segments to be combined is not limited to three segments.

  Step S3: Next, a binarization process 15c is performed on the character image. The binarization process 15c is a process of dynamically determining a binarization threshold value from the character image cut out by the character image cutout process 15b and converting the character image into a binary image. When the threshold value is dynamically determined, an average density of the extracted character image is calculated, and a density corresponding to, for example, 60% of the average density is set as a binarization threshold value.

  Step S4: Next, a character image correction process 15d is performed on the binary image for the purpose of correcting handwritten characters such as “characters”. The character image correction process 15d detects vertical and horizontal stroke inclinations of the binary image. Then, the character image is converted into a character image projected in the horizontal direction / vertical direction by performing geometric conversion in an amount corresponding to the inclination. Thereafter, the character image is nonlinearly enlarged and reduced and normalized so that the distribution of black pixels in the character image is constant. This normalized character image is output as a normalized image.

  Step S5: Next, a feature extraction process 16a for digitizing the character features included in the normalized image is performed. When the background color of the form is white and the character is black, the feature extraction processing 16a tracks a contour of a region composed of black pixels of the normalized image to extract a multidimensional feature vector (extracted feature vector) from the character image. ). As a result, the character image is digitized.

  Step S6: The recognition process 16b calculates a certainty parameter indicating the distance between the extracted feature vector obtained in the feature extraction process 16a and the registered feature vector stored in the character recognition dictionary 11a. Thus, the character image included in the normalized image is associated with the character code via the certainty factor parameter. The recognition process 16b is performed on a character basis. That is, when the normalized image includes three-character image data, the recognition process 16b is performed for each character. The character corresponding to the acquired character code is the recognition result.

  The certainty factor parameter indicates the similarity between the extracted feature vector and the registered feature vector. Accordingly, the recognition result of the certainty factor parameter having a large value has a similar surname to the characters registered in the character recognition dictionary 11a.

And the recognition process 16b selects a maximum of 10 recognition results in order of the recognition result with the largest value of a certainty factor parameter, and makes this a candidate character. For example, when the binary image includes three characters, a plurality of recognition results having different confidence parameter values are obtained for each character. Therefore, the recognition processing unit 16b sets 10 recognition results as candidate characters in descending order of the certainty parameter value. The recognition result corresponding to the confidence parameter with the largest value is described as a recognized character, and the character string is described as a recognized character string.
Step S7: The rejection determination process 16c determines whether or not the certainty factor parameter of the candidate character obtained in the recognition process 16b is greater than or equal to a preset threshold, leaves candidate characters that are greater than or equal to the threshold, and candidates that are less than or equal to the threshold Characters are a process to reject. Then, the remaining candidate characters are set as the Nth candidate character in descending order of the certainty factor parameter. For example, the candidate character having the largest certainty parameter is the first candidate character, and the candidate character having the third certainty parameter is the third candidate character.

  By this rejection determination process 16c, for example, the regular image includes three characters, the first character is “2” candidate characters, the second character is “5” candidate characters, and the third character is As in “0” candidate characters, different numbers of candidate characters remain for each character.

  Since there are a plurality of candidate characters, there are a plurality of candidate character strings formed by combining them. In the above example, the number of candidate characters for the third character is “0”, but this is missing because the third character is misrecognized (including typographical errors and omissions). This happens when

  The search word matching process 17a searches customer information and transaction history searched from the customer dictionary 11b using the candidate character string as a search keyword, and corrects the recognized character based on this information.

  FIG. 5 is a flowchart showing a processing procedure in the search word matching process 17a. One candidate character string corresponds to a character string in one field. The customer basic information related to the client, the customer incidental information related to the transfer financial institution and the account number, etc. are entered in different fields.

  Step S8: First, it is determined whether or not a number candidate character string exists among candidate characters related to the client. Candidate characters related to the client are, for example, the number information of the client telephone number field consisting of numbers, the character information of the client surname field and client name field consisting of kana, and the character information of the client surname field including kanji in FIG. This is character information corresponding to the client address field including kanji and numbers.

  In general, numeric information has higher recognition accuracy than character information. That is, the certainty parameter value is large. Therefore, a number candidate character string made up of such number information is set as a search keyword having a high priority. However, the number candidate character string does not always exist depending on the type of form. Therefore, it is determined whether or not a number candidate character string exists.

  Steps S9 to S11: If a number candidate character string exists, the number candidate character string is set as a search keyword. If no number candidate character string exists, a non-number candidate character string is set as a search keyword. Here, the non-numeric candidate character string refers to a candidate character string other than the numeric candidate character string. Then, the customer basic information is searched from the customer dictionary 11b using the search keyword.

  Steps S12 and S13: At this time, customer basic information corresponding to the search keyword may not be registered. Therefore, it is determined whether or not the customer basic information has been acquired. If it cannot be obtained, there are a plurality of candidate character strings related to the client, so it is determined whether or not a search has been performed using all candidate character strings as search keywords. If all candidate character strings have not been searched as search keywords, the process returns to step S8, and otherwise the process ends. In this case, the recognized character becomes the final recognition result from the recognition result having the largest certainty parameter value obtained by the recognition process.

  Step S14: It is determined whether or not the acquired customer basic information is appropriate. That is, there are cases where the search keyword is inappropriate or similar basic customer information exists, and the acquired basic customer information is not a transfer requester. Therefore, it is determined that the acquired customer basic information is appropriate only when the search keyword is completely included in the acquired customer basic information, and is determined to be inappropriate in other cases.

  For example, since the client's full name is “Patent Ichiro” and the candidate character includes a zip code which is a numerical candidate character string, the candidate character of this zip code was searched as a search keyword. It is assumed that the customer basic information obtained at this time is “Patent Jiro” instead of “Patent Ichiro”. In this case, since the acquired customer basic information of “Patent Jiro” cannot be used as the customer basic information of “Patent Ichiro”, it is determined to be inappropriate.

  If it is determined that the acquired customer basic information is inappropriate, the process proceeds to step S13, and the search keyword is changed and the search is performed again. On the other hand, if information that completely matches the search keyword is hit, the process proceeds to step S15.

  Step S15: Since the customer basic information related to the client is obtained by the above processing, a command for sorting the transfer history made by the client in the past is output to the dictionary unit 11. When the dictionary unit 11 receives the transfer history sorting command, the dictionary unit 11 sorts the transfer history by a preset method such as the frequency order of the transfer destinations or the order of the transfer dates.

  Steps S16 and S17: It is determined whether or not the recognition result needs to be corrected based on the acquired customer basic information. The candidate character regarding the transfer person includes a plurality of pieces of information such as the client's first and last name, address, and telephone number. Therefore, even if the first and last name of the client of the acquired customer basic information matches the telephone number, the address does not always match. In this case, correction relating to the address is required. Of course, if all the addresses are different, if correction is made based on the basic customer information, if the client actually moves, the correction is made with wrong information. Therefore, when the correction is made, the correction is made only when the degree of coincidence is a predetermined value or more (for example, 80% or more).

  For example, the client's address included in the acquired customer basic information is “5-7-1 Shiba, Minato-ku, Tokyo” (12 characters), but the candidate address is “5-chome Shiba, Minato-ku, Tokyo”. When “1-1” (11 characters) does not match, the matching degree is 91.7% (= 11/12 * 100), so the recognized character is corrected with the acquired address.

  Steps S18 and S19: As a result of the above processing, the recognition result regarding the client has been properly corrected, and the customer-accompanying information corresponding to the recipient information is searched using the corrected recognition character as a search keyword. And the suitability of the acquired customer accompanying information is determined. Appropriateness of the customer-accompanying information is performed in the same procedure as that for determining whether the customer basic information is appropriate.

  Steps S20 and S21: It is determined whether or not the recognized character is to be corrected based on the customer-accompanying information. If correction is necessary, the recognized character regarding the recipient is corrected.

  As described above, the image data is character-recognized using the character recognition dictionary, and the customer dictionary is searched using the recognition result regarding the customer, and the recognition result is corrected based on the searched customer information. Recognition result correction processing can be performed at high speed and with high accuracy.

2 character recognition system 3 scanner 4 server 5 terminal 6 character recognition device 11 dictionary unit 11a character recognition dictionary 11b customer dictionary 12 recognition unit 15 preprocessing unit 15a position correction process 15b character image cutout process 15c binarization process 15d character image correction Process 16 Recognition unit 16a Feature extraction process 16c Rejection determination process 16b Recognition process 17 Knowledge processing unit 17a Search word matching process 30 Customer information table

Claims (8)

A character recognition device for recognizing handwritten characters from image data of handwritten characters,
A dictionary unit comprising: a character recognition dictionary used for character recognition of the handwritten characters; and a customer dictionary for storing information about a customer who entered the handwritten characters;
The image data is character-recognized using the character recognition dictionary, and the customer dictionary is searched using information of recognized characters included in the character recognized, and the recognition is performed based on the searched customer information. A recognition unit for correcting characters,
The customer dictionary is composed of customer basic information and customer-accompanying information associated with the customer basic information ,
The recognition unit is
A preprocessing unit for obtaining a character image from the image data;
A recognition unit for obtaining a recognized character based on the acquired character image and the character recognition dictionary;
If the customer basic information is obtained by searching the customer dictionary using information about the customer included in the recognition character as a search keyword, and the acquired customer basic information is appropriate, the acquired customer basic information A knowledge processing unit that corrects the recognized character based on:
Character recognition apparatus according to claim Rukoto equipped with. The character recognition device according to claim 1,
The recognition unit includes a preprocessing unit that cuts a recognition target from the image data and acquires a character image;
A recognition unit that obtains a feature of a character included in the character image as an extracted feature vector, compares the extracted feature vector with a registered feature vector registered in the character recognition dictionary, and obtains a recognized character;
A knowledge processing unit that searches the customer dictionary using the information about the customer included in the recognized character as a search keyword to acquire the customer basic information, and corrects the recognized character based on the acquired customer basic information. A character recognition device. The character recognition device according to claim 2,
The knowledge processing unit corrects the recognized character based on the acquired customer basic information, and then corrects the recognized character based on the customer-accompanying information associated with the customer basic information. apparatus. The character recognition device according to claim 3,
When the recognition character regarding the customer included in the recognition character includes numeric information and character information, the character information is set as the search keyword having a higher search priority than the character information. apparatus. The character recognition device according to claim 3 or 4,
The knowledge processing unit instructs the dictionary unit to sort the customer-accompanying information associated with the customer basic information when the correction of the recognized character is started based on the acquired customer basic information. A character recognition device. A character recognition method for recognizing handwritten characters from image data of handwritten characters,
A character recognition procedure for recognizing the image data using a character recognition dictionary;
A search procedure for searching the customer basic information using the recognition character from a customer dictionary composed of customer basic information related to the customer who entered the handwritten characters and customer incidental information associated with the customer basic information;
A first correction procedure for acquiring the customer basic information by searching the customer dictionary using a recognition character related to a customer included in the recognition character as a search keyword, and correcting the recognition character based on the acquired customer basic information; ,
After correcting the recognized character based on the acquired customer basic information, see contains a second correction procedure that corrects the recognized character based on said customer associated information associated with the customer basic information,
The first correction procedure includes:
A preprocessing procedure for obtaining a character image from the image data;
A recognition procedure for obtaining a recognized character based on the acquired character image and the character recognition dictionary;
If the customer basic information is obtained by searching the customer dictionary using information about the customer included in the recognition character as a search keyword, and the acquired customer basic information is appropriate, the acquired customer basic information A knowledge processing procedure for correcting the recognized character based on:
Character recognition method, which comprises a. The character recognition method according to claim 6,
When the recognition character related to the customer included in the recognition character includes numerical information and character information, it includes a procedure for setting the numerical information as the search keyword having a higher search priority than the character information. Character recognition method. The character recognition method according to claim 7,
A character including a procedure for instructing the dictionary unit to sort the customer-accompanying information associated with the customer basic information when starting correction of the recognized character based on the acquired customer basic information. Recognition method.

Images