JP2006343797A - Character recognition device, character recognition method, and computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To realize higher character recognition accuracy in a structure equipped with a user dictionary. <P>SOLUTION: Document image is input, and character recognition is performed by collating selected one of a plurality of used dictionaries with an updating dictionary GU. The updating dictionary GU has a data structure which uses groups of each item F1, F2, F3, F4 of "character code", "correspondence dictionary identification information", "number of times of updating", and "feature quantity" as one record. Such one record expresses one character. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a technique for recognizing a character image.

  The character recognition device includes a recognition dictionary that links a character image and a character code, and performs character recognition by collating the character image with the recognition dictionary. 2. Description of the Related Art Conventionally, some character recognition devices have a configuration including a plurality of types of recognition dictionaries depending on fonts (Patent Document 1 below).

  The plurality of recognition dictionaries are prepared in advance, that is, recognition dictionaries built in the system (hereinafter referred to as “system dictionaries”). The character recognition device recognizes that the operator can manage in addition to the system dictionaries. Some have a dictionary (hereinafter referred to as a “user dictionary”) (see Patent Document 2 below). By providing the user dictionary, recognition accuracy is improved.

JP-A-11-85908 JP-A-6-266902

  However, in the technology provided with the above user dictionary, adding a character image to the user dictionary may make it impossible to recognize similar characters that have been recognized so far, and the accuracy of character recognition is necessarily sufficiently increased. There was a problem that could not. This is because when a character image and a correct character code are added to the user dictionary due to misrecognition, the added character image is similar to a character pattern registered in a dictionary different from the dictionary that performed the misrecognition. This is because the characters that could be recognized by this different dictionary until now become unrecognizable.

  The problem to be solved by the present invention is to realize higher character recognition accuracy in a configuration including a user dictionary.

  As means for solving at least a part of the problems described above, the following configuration is adopted.

The character recognition device of the present invention is
As a dictionary for character recognition, a first dictionary whose registration contents are fixed in advance and a second dictionary that can update the registration contents are provided. A document image is input, and the character image represented in the document image is converted into the character image. A character recognition device that performs character recognition by collating with a first dictionary and a second dictionary,
A first dictionary storage means for storing a plurality of types of dictionaries that store character image feature quantities and character codes in association with each other as the first dictionary;
Second dictionary updating means for updating the second dictionary,
The second dictionary updating means includes
A character image cutting means for cutting out a character image of one character from the document image;
A character code specifying means for specifying a character code for the character image cut out by the character image cut-out means;
Dictionary selection means for selecting one dictionary from a plurality of types of dictionaries stored in the first dictionary storage means;
Feature quantity acquisition means for obtaining a feature quantity of the character image cut out by the character image cutout means;
The feature quantity corresponding to the character code designated by the character code designation means is read from the dictionary selected by the dictionary selection means, and the read feature quantity is the feature quantity obtained by the feature quantity acquisition means. A feature amount correcting means for correcting by,
A set of a character code designated by the character code designation unit, a feature amount after correction by the feature amount correction unit, and identification information for identifying the dictionary selected by the dictionary selection unit is set for one character. And a data adding means for adding the data to the second dictionary as data.

  According to the character recognition device having the above configuration, the character image is recognized by collating with the first dictionary and the second dictionary. The first dictionary further includes a plurality of dictionaries, The dictionary is updated. The update of the second dictionary is performed in response to the extraction of the character image, the specification of the character code for the extracted character image, and the selection of one dictionary from a plurality of types of dictionaries as the first dictionary. . The feature amount of the cut character image is acquired, the feature amount corresponding to the specified character code is read from the selected dictionary, and the read feature amount is corrected with the acquired feature amount. To do.

  For this reason, since the feature amount is corrected for each dictionary, when a character image and a correct character code are added to the second dictionary due to erroneous recognition, the added character image performs erroneous recognition. It is not used for character recognition using a dictionary different from the dictionary. Therefore, by adding a character image to the second dictionary, similar characters that have been recognized so far will not become unrecognizable, so that the accuracy of character recognition can be sufficiently increased.

  In the character recognition device having the above configuration, the data for one character including the character code specified by the character code specifying means and the identification information for identifying the dictionary selected by the dictionary selecting means is the second data. A registration determination unit that determines whether or not the dictionary has already been registered, and when the registration determination unit determines that it has not been registered, the registered data is used as a feature amount corrected by the feature amount correction unit. It is good also as a structure provided with the data correction means corrected by it.

  According to this configuration, when the registration is already made in the second dictionary using the character code and the dictionary identification information as a key, the registered information can be corrected.

  In the character recognition device, the character code designating unit is configured to designate a character code representing a character input in response to an operator operation command, and the dictionary selecting unit is configured to respond to the operator operation command. It is good also as a structure which selects the dictionary designated according to it.

  According to this configuration, it is possible to easily register the character code intended by the worker and the identification information of the dictionary in the second dictionary.

  In the character recognition device, a character image of one character is sequentially cut out from the document image, a feature amount of the cut-out character image is calculated, and selected from a plurality of dictionaries stored in the first dictionary storage means A character recognition unit that performs character recognition by selecting a character code that is close to the feature amount obtained by the calculation from the used dictionary and the second dictionary, and the dictionary selection unit includes the use dictionary It is good also as a structure made into the selected dictionary.

  According to this configuration, since the dictionary used in character recognition can be immediately specified, the configuration is simple.

  The feature amount correction unit may include an averaging unit that calculates an average value of the feature amount read from the dictionary and the feature amount obtained by the feature amount acquisition unit. According to this configuration, the feature amount can be easily corrected.

The character recognition method of the present invention includes:
As a dictionary for character recognition, a first dictionary whose registration contents are fixed in advance and a second dictionary whose registration contents can be updated are prepared, a document image is input, and a character image represented in the document image A character recognition method for recognizing characters by comparing them with the first dictionary and the second dictionary,
(A) a step of previously storing a plurality of types of dictionaries for storing the character image feature quantity and the character code in association with each other as the first dictionary;
(B) a step of updating the second dictionary;
The step (b) is:
(B-1) a step of cutting out one character image from the document image;
(B-2) a step of designating a character code for the character image cut out in the step (b-1);
(B-3) a step of selecting one dictionary from a plurality of types of dictionaries as the first dictionary;
(B-4) a step of obtaining a feature amount of the character image cut out in the step (b-1);
(B-5) A feature amount corresponding to the character code specified in the step (b-2) is read from the dictionary selected in the step (b-3), and the read feature amount is A process of correcting with the feature amount obtained by the process (b-4);
(B-6) The character code designated by the step (b-2), the corrected feature amount by the step (b-5), and the dictionary selected by the step (b-3) are identified. And a step of adding a set of the identification information to the second dictionary as data for one character.

The computer program of the present invention is:
For performing character recognition by inputting a document image and collating the character image represented in the document image with a first dictionary whose registration contents are fixed in advance and a second dictionary whose registration contents can be updated. A computer program,
(A) a function of preliminarily storing in the storage device a plurality of types of dictionaries that store character image feature quantities and character codes in association with each other as the first dictionary;
(B) causing the computer to realize the function of updating the second dictionary;
The function (b) is
(B-1) a function of cutting out one character image from the document image;
(B-2) a function of designating a character code for the character image cut out by the function (b-1);
(B-3) a function of selecting one dictionary from a plurality of types of dictionaries as the first dictionary;
(B-4) a function of obtaining a feature amount of the character image cut out by the function (b-1);
(B-5) A feature amount corresponding to the character code designated by the function (b-2) is read from the dictionary selected by the function (b-3), and the read feature amount is A function of correcting with the feature amount obtained by the function (b-4);
(B-6) The character code specified by the function (b-2), the feature value after correction by the function (b-5), and the dictionary selected by the function (b-3) are identified. And a function of adding a set of the identification information to the second dictionary as data for one character.

  According to the character recognition method and the computer program of the present invention, as in the character recognition device of the present invention, the second dictionary capable of updating the registered contents with respect to the first dictionary with the fixed registered contents is provided. As a result, it is possible to achieve higher character recognition accuracy.

  The recording medium of the present invention is characterized by a computer-readable recording medium that records the computer program of the present invention. This recording medium has the same operations and effects as the computer program of the present invention.

  The present invention includes other aspects as follows. The 1st aspect is an aspect as a program supply apparatus which supplies the computer program of this invention via a communication path. In this first aspect, the above-described method and apparatus are realized by placing a computer program on a server or the like on a computer network, downloading a necessary program to a computer via a communication path, and executing the program. Can do.

The best mode for carrying out the present invention will be described based on examples. This embodiment will be described in the following order.
1. Device configuration:
2. Computer processing:
3. Action / effect 4. Other embodiments:

1. Device configuration:
FIG. 1 is an explanatory diagram showing a schematic configuration of a computer system to which an embodiment of the present invention is applied. The computer system of this embodiment mainly includes a personal computer 10 constituting the character recognition device of the present invention, and includes a display 20, a keyboard 22, and a mouse 24 as peripheral devices. Furthermore, a scanner 26 and a printer 28 are connected to the personal computer 10.

  The personal computer 10 includes a memory 13, a hard disk drive 14, an input control unit 15, a display control unit 16, and the like connected to each other by a bus 12 around a CPU 11 as a central processing unit. The memory 13 stores various data and becomes a work area of the CPU 11.

  The hard disk drive 14 stores a computer program PR as software of the character recognition device and a dictionary for character recognition. The character recognition dictionary includes a system dictionary prepared in advance and a user dictionary that can be managed by the operator. The system dictionary has registered contents fixed in advance, and corresponds to the “first dictionary” of the present invention. The user dictionary can update the registered contents, and corresponds to the “second dictionary” of the present invention. As the system dictionary, n (n is a plurality) dictionaries such as a basic recognition dictionary G1, a collapsed character recognition dictionary G2, a faint character recognition dictionary G3, a Mincho recognition dictionary G4, and a Gothic recognition dictionary G5 are prepared. . That is, as a system dictionary, a basic recognition dictionary G1, a collapsed character recognition dictionary G2 in which the degree of character crushing / fading is changed, a blurred character recognition dictionary G3, etc., and a Mincho recognition dictionary in which the font of characters is changed. G4, Gothic recognition dictionary G5, etc. are prepared. The user dictionary is an update dictionary GU. The hard disk drive 14 corresponds to the “first dictionary storage means” of the present invention.

  The dictionaries G1 to Gn store character image feature quantities and character codes in association with each character. The “feature amount” is for expressing the shape of the character image in numerical form. In this embodiment, a known mesh feature is used as the feature amount. In the mesh feature, a character region is divided into a plurality of small regions, and the number of black pixels included in the small region is used as a feature amount.

  The update dictionary GU is a dictionary for correcting the feature amount of characters registered in the dictionaries G1 to Gn serving as system dictionaries. Characters are registered (added) and corrected by the user. The configuration of the update dictionary GU will be described later. The hard disk drive 14 stores a word dictionary GW and a registered image information file PF, which will be described later.

  The scanner 26 optically scans a manuscript such as a photograph or document, measures the intensity of the reflected light or transmitted light, and converts the light into analog-digital data to capture an image (image) into a computer. Device. Here, it is used to obtain document image data DD using a document as a document.

  The input control unit 15 is a control unit that captures input operations from the keyboard 22 and the mouse 24 and captures document image data DD from the scanner 26. The display control unit 16 is a control unit that controls signal output to the display 20.

  The computer program PR is originally stored in a CD-ROM (not shown) as a recording medium, and is installed in the hard disk drive 14 from the CD-ROM by starting a predetermined installation program. is there. Similarly, dictionaries G1 to Gn as system dictionaries are installed in the hard disk drive 14 from the CD-ROM. Various constituent requirements of the character recognition device of the present invention are realized by the CPU 11 executing the computer program PR.

  In FIG. 1, various constituent requirements are indicated by functional blocks implemented in the CPU 11. The CPU 11 inputs the document image data DD by executing the computer program PR, and performs character recognition by collating the character image represented in the document image data DD with the dictionaries G1 to Gn and the update dictionary GU. The update dictionary GU is updated. The update dictionary GU is updated using the second dictionary update unit 18. The second dictionary update unit 18 is performed using a character image cutout unit 18a, a character code designation unit 18b, a dictionary selection unit 18c, a feature amount acquisition unit 18d, a feature amount correction unit 18e, and a data addition unit 18f. The function of each part 18a-18f will be described in detail later.

  The computer program PR is provided as a configuration stored in another portable recording medium (portable recording medium) such as a flexible disk, a magneto-optical disk, and an IC card in place of the CD-ROM. It can be. The computer program PR can also be provided via a network from a specific server connected to an external network. The network may be the Internet or a computer program obtained by downloading from a specific homepage. Alternatively, it may be a computer program supplied in the form of an email attachment.

2. Computer processing:
The control process according to the computer program PR will be described in detail below. FIG. 2 is a flowchart showing a main routine of this control process. This main routine is started when an instruction to start the computer program PR is given. As shown in the figure, when the process is started, the CPU 11 first performs a process of displaying a work window WD on the CRT display 12 (step S100).

  FIG. 3 is an explanatory diagram showing the window WD. As illustrated, the window WD is provided with a menu bar BR1 and a tool bar BR2. The toolbar BR2 is provided with “scan”, “recognition”, and “save” buttons BT1, BT2, and BT3 for executing commands in the order according to the work procedure. Below the toolbar BR2, a document image display field FD1 and a recognition result display field FD2 are provided. Each field FD1, FD2 is displayed in the figure, but is blank at the time of step S100.

  Returning to FIG. 2, the CPU 11 then executes a scan process for scanning a document that is a character recognition target in response to an operation in which the “scan” button BT <b> 1 on the toolbar BR <b> 2 is clicked by the mouse 24 ( Step S110). As a result, the document image data DD obtained by optically reading the document image is stored in the memory 13 and displayed in the document image display field FD1 of the window WD. Subsequently, the CPU 11 executes a character recognition process for recognizing the character image represented in the document image data DD obtained in step S110 (step S120).

  FIG. 4 is a flowchart showing details of the character recognition process executed in step S120. As shown in the figure, when the process is started, the CPU 11 performs a process of cutting out one character image from the document image data DD (step S200). Specifically, the CPU 11 checks whether the character sequence of the document is horizontal or vertical, cuts out a row in the horizontal direction and a column in the vertical direction, and converts the row into one character from the cut-out sequence. One corresponding character image is cut out.

  Next, in response to an operation in which the “recognition” button BT2 on the tool bar BR2 is clicked by the mouse 24, the CPU 11 uses a dictionary used as a system dictionary for the character image clipped in step S200. A process of selecting one of Gn is performed (step S210). Specifically, by looking at the thickness of the characters in the character image cut out in step S200, it is possible to determine whether the collapsed character recognition dictionary G2 and the blurred character recognition dictionary G3 are suitable for the use dictionary. Further, the character code determined to have a high degree of similarity is once compared (matched) with the basic recognition dictionary G1, and the feature amount of the character image corresponding to the character code is determined as the Mincho character recognition dictionary G4. Then, it is read from each font-specific recognition dictionary such as the Gothic recognition dictionary G5, and it is determined whether the read feature value is similar to the character image feature value cut out in step S200. Then, the dictionary including the character image having the highest similarity is selected as the use dictionary. The selection of the use dictionary is not limited to such a method, and can be made by another method depending on the type of the dictionary G1 to Gn. In short, any method may be used as long as it can select dictionaries G1 to Gn most suitable for recognition of the character image cut out in step S200.

  Next, the CPU 11 collates the character image cut out in step S200 with the dictionary (any one of G1 to Gn) selected as the use dictionary in step S210 and the update dictionary GU, thereby matching the character image. Character recognition is performed (step S220). More specifically, the feature amount of the extracted character image is calculated, a character code close to the calculated feature amount is selected from the use dictionary and the update dictionary GU, and the selected character code is recognized. Output as a result. “From the use dictionary and the update dictionary GU” is to perform matching on both dictionaries and select a character code having a similar feature amount from both dictionaries. As a recognition result, not only a character code having the closest feature amount (highest similarity) but also a plurality of candidates are obtained with priority (higher priority is higher priority).

  Thereafter, the CPU 11 stores the character recognition result obtained in step S220 in the memory 13 (step S230). Next, the CPU 11 determines whether or not the character image cut out in step S200 is the last character image included in the document image data DD (step S240). If it is determined that it is not the last one, the process returns to step S200 to cut out the next character image and perform character recognition. On the other hand, if it is determined in step S230 that the character image is the last character image, the CPU 11 proceeds to step S250, assuming that all character images included in the document image data DD have been recognized. In step S250, the CPU 11 performs a process of displaying the recognition result, that is, the character code string of the first candidate stored in step S230, in the recognition result display field FD2 of the window WD. After execution of step S250, the CPU 11 advances the process to “return” and exits the routine of this character recognition process.

  In this way, after exiting the character recognition process of step S120 in FIG. 2, the CPU 11 advances the process to step S130. In step S130, the CPU 11 performs a correction process for correcting the character recognition result obtained in step S120. There are three types of correction processing executed here: automatic learning correction, semi-automatic learning correction, and manual correction.

  FIG. 5 is an explanatory diagram for explaining the above three types of learning correction using an example of correction results. The first row in the figure shows the arrangement of character images obtained as a result of the scan processing in step S110, and the third to sixth rows show the result of character recognition processing in step S120. From the first candidate to the fourth candidate obtained as shown in FIG. The second level in the figure shows the result of the correction process in step S130.

  In the automatic learning correction, the character recognition result is automatically corrected using the word dictionary GW without the help of an operator. That is, since the first candidate as the recognition result obtained by the character recognition processing in step S120 is not necessarily correct, the character recognition result is obtained from the recognition result obtained as a word stored in the word dictionary GW. Are searched from the second candidate and below and designated as the correct answer to correct the result. The designation of the correct character code constitutes the character code designation unit 18b (see FIG. 1) at the time of automatic learning correction. In the illustrated example, the second character is due to automatic learning correction. In the character recognition process, “book” is recognized as the second candidate, but since the word “dictionary” is registered in advance in the word dictionary GW, correction by automatic learning correction using the word dictionary GW in step S120. By performing the process, "book" is adopted as the recognition result.

  The semi-automatic learning correction selects one of a plurality of character codes obtained as the second or lower candidate as a character recognition result in accordance with the operator's operation command and corrects the selected character code. Is. In the illustrated example, the seventh character is due to semi-automatic learning correction. In the character recognition process, “ni” is recognized as the first candidate and “ha” is recognized as the second candidate, but even if it is “ni” as the sentence, there is no problem, so correction using the word dictionary GW is performed. It is not performed, and the result of the character recognition process remains as it is. Therefore, with the help of the operator, semi-automatic learning correction for correcting “to” to “to” is performed.

  FIG. 6 is an explanatory diagram showing an example of a recognition result display field FD2 at the time of semi-automatic learning correction. As shown in the drawing, the recognition result including the seventh error character is displayed in the recognition result display field FD2 provided in the window WD. The operator designates a correct character from a plurality of candidate characters according to the following procedure. First, the operator operates the mouse 24 to move the cursor to the error character (the character “”). As a result, a plurality of candidate characters are displayed at the bottom of the recognition result display field FD2 such that the higher the similarity is, the more left the character is. Next, the operator performs an operation of clicking a correct character from the displayed candidate characters. In the figure, the candidate character TC “ha” is clicked, and the error character (“character” in “) is corrected to the clicked candidate character TC (“ character ”in“). The designation of clicking the correct character code constitutes the character code designation unit 18b (see FIG. 1) at the time of semi-automatic learning correction.

  In manual correction, when it is desired to correct a character other than the candidate character obtained by the character recognition process, a character input is directly received from the operator and the input character is corrected. In the example shown in FIG. 5, the 18th character is due to manual correction. In the character recognition process, the characters “call” are not displayed as candidates. In this case, since the word “dictionary” cannot be constructed from the candidate characters, it cannot be corrected and an incorrect result is output. Therefore, manual correction for correcting “thing” to “call” is performed with the help of the operator.

  The operator performs processing for correcting an incorrect recognition result in the following procedure. First, the operator operates the mouse 24 to move the cursor to an error character in the character string displayed in the recognition result display field FD2 of the window WD. Next, a “manual correction” button BT4 (see FIG. 3) on the tool bar BR2 of the window WD is clicked with the mouse 24 to display a dialog box DB for manual correction.

  FIG. 7 is an explanatory view showing a dialog box DB for manual correction. The dialog box DB includes “character image” and “list” display fields FD11 and FD12, and “character” and “dictionary” input fields FD13 and FD14. In the “character image” display field FD11, the document image data DD displayed in the document image display field FD1 of the window WD is displayed as it is. At this time, the error character specified earlier from the recognition result display field FD is surrounded by a red frame RF. The operator operates the mouse 24 to input identification information (ID) for identifying the dictionary from the “dictionary” input field FD14, and operates the keyboard 22 to input the “character” input field FD13. Key in characters. Thereafter, the “Add” button BT11 is clicked. As a result, the character code representing the character input from the “character” input field FD13 is designated, and the error character is corrected to the designated character code (character code of “book”). . Note that the configuration for designating the correct character code and the dictionary to be selected from the input fields FD13 and FD14 of “character” and “dictionary” in accordance with the operator's operation command is the character code designating portion 18b and dictionary at the time of manual correction. It constitutes a part of the selector 18c (see FIG. 1).

  In the correction process of step S130, a process of performing automatic learning correction is always executed (of course, there is no change of the character when the character recognition result is correct), and further, semi-automatic when receiving the above-described operation command from the operator. Learning correction or manual correction is performed. Each time a character is changed by any one of automatic learning correction, semi-automatic learning correction, and manual correction, a process of updating the update dictionary GU is performed. Next, the update processing of the update dictionary GU will be described in detail.

  FIG. 8 is an explanatory diagram showing the data structures of the update dictionary GU and the registered image information file PF. As shown in the figure, the update dictionary GU has a data structure in which a set of items F1, F2, F3, and F4 of “character code”, “corresponding dictionary identification information”, “update count”, and “feature” is one record. It has. One character is represented by one record. According to the update dictionary GU having such a configuration, a dictionary (any one of G1 to Gn) specified by the identification information stored in the item F2 of “corresponding dictionary identification information” is used as a correction target, and It can be seen that a correction is made for the character that matches the character code stored in the item “Character code” F1, and the feature amount as the correction amount is stored in the item “F4” item F4. You can see that

  On the other hand, the registered image information file PF has a data structure in which a set of items E1, E2, and E3 of “character code”, “update dictionary corresponding record number”, and “character image file name” is one record. Yes. The item “E3” of “character image file name” can be m (m is a positive number). The data stored in the item “E1 corresponding record number of the update dictionary” is for designating one record of the update dictionary GU. The data stored in the “character code” item F1 of one record of the update dictionary GU specified by the item E1 matches the data stored in the “character code” item E1. The registered image information file PF is for registering the character image used for the update every time the update dictionary GU is updated.

  In the update process of the update dictionary GU, the registered image information file PF is not directly used. The registered image information file PF is used when the update dictionary GU is damaged and re-created, or when displaying the character images registered in the update dictionary GU as a list.

  FIG. 9 is a flowchart showing an update process of the update dictionary GU. This update processing of the update dictionary GU is executed when a character is changed by any one of automatic learning correction, semi-automatic learning correction, and manual correction. As shown in the figure, when the process is started, the CPU 11 first performs a process of taking in the character code CC changed by any of the above-described corrections (step S300). Next, the CPU 11 performs processing for taking in the feature amount CV1 of the character image before the change (step S310). In each of the automatic learning correction, semi-automatic learning correction, and manual correction processes, the feature amount of the character image of one character cut out from the document image data DD is calculated. In step S310, the feature amount is captured. . The configuration for capturing the feature amount corresponds to the feature amount acquisition unit 18d (see FIG. 1).

  Next, in the case of automatic or semi-automatic learning correction, the CPU 11 stores identification information for identifying the used dictionary selected in step S210 as corresponding dictionary identification information Y, and in the case of manual correction, “dictionary”. The dictionary identification information input by the operator from the input field FD14 is stored as corresponding dictionary identification information Y (step S320). In the automatic and semi-automatic learning correction, the configuration for receiving the identification information of the used dictionary selected in step S210 of the character recognition process corresponds to the dictionary selection unit 18c (see FIG. 1), and in the manual correction, The “dictionary” input field FD14 provided in the dialog box DB described above, and the configuration for capturing the dictionary identification information input by the operator from the input field FD14 correspond to the dictionary selection unit 18c (see FIG. 1). .

  Thereafter, the CPU 11 stores in the updated dictionary GU, the corresponding character identification information that the item “F1” of “character code” matches the character code CC captured in step S300, and the item “F2” of “corresponding dictionary identification information” is read in step S320. It is determined whether or not a record matching Y is registered (step S330). Here, if a negative determination is made, that is, it is determined that it is not registered, the process proceeds to step S340.

  In step S340, the CPU 11 reads the feature amount CV0 corresponding to the character code CC captured in step S300 from the dictionary specified by the corresponding dictionary identification information Y read in step S320 (step S340).

  Thereafter, a process of adding a new record to the update dictionary GU is performed (step S350). Specifically, the CPU 11 transfers the character code CC fetched in step S300 to the item F1 of “character code”, and the correspondence dictionary identification information Y read in step S320 to the item F2 of “corresponding dictionary identification information”. The value 1 is transferred to the “update count” item F3, and the feature value CV0 read from the correspondence dictionary in step S340 and the feature value CV1 captured in step S310 are stored in the “feature value” item F4. A new record created by transferring the average value, that is, the calculation result of (CV0 + CV1) / 2 is added to the update dictionary GU. After execution of step S350, the process returns to “Return” to temporarily end the update process of the update dictionary GU.

  On the other hand, if it is determined affirmative in step S330, that is, if it is registered, the record determined to match, that is, the item “F1” of “character code” matches the character code CC captured in step S300, In addition, a process for correcting the record in which the item F2 “corresponding dictionary identification information” matches the corresponding dictionary identification information Y read in step S320 is performed (step S360). The details of the correction are as follows. The item F3 of “update count” of the record determined to match is incremented by a value 1, and the value stored in the item F4 of “feature” is acquired in step S310. It is averaged using the quantity CV1. In this averaging, the value stored in the item F4 “feature” is multiplied by the value of the item F3 “number of updates” (the value after the increment), and the multiplication result is captured in step S310. This is performed by adding the amount CV1 and dividing the addition result by a value obtained by adding the value 1 to the value of the item F3 of “update count”. The “character code” item F1 and the “corresponding dictionary identification information” item F2 are not changed. After executing step S360, the process returns to “Return” to temporarily end the update process of the update dictionary GU.

  Details of steps S350 and S360 will be further described by way of examples. When a new record is added to the update dictionary GU, that is, when the dictionary is updated for the first time, assuming that the feature value CV1 captured in step S310 is V1, the item F3 of “number of updates” has a value of “ The item F4 of “feature” is (CV0 + V1) / 2.

  At the time of the second update, assuming that the feature value CV1 captured in step S310 is V2, the “number of updates” item F3 is the value 2, and the “feature” item F4 is (CV0 + V1 + V2) / 3. At the time of the third update, assuming that the feature value CV1 captured in step S310 is V3, the “update count” item F3 is the value 3, and the “feature” item F4 is (CV0 + V1 + V2 + V3) / 4. That is, at the k-th update (k is a positive number), if the feature value CV1 captured in step S310 is Vk, the “number of updates” item F3 is k, and the “feature value” item F4 is ( CV0 + V1 + V2 + ... + Vk) / (k + 1).

  The update processing of the update dictionary GU described above is performed every time a character is changed in any one of automatic learning correction, semi-automatic learning correction, and manual correction in step S130 as described above. After exiting the correction process in step S130, the CPU 11 advances the process to step S140. In step S140, the CPU 11 executes a process of saving the character string of the recognition result in a storage device such as the hard disk drive 14 in response to an operation in which the “save” button BT3 on the toolbar BR2 is clicked by the mouse 24. To do. This storage is performed upon receiving a file name input from the keyboard 22. Thereafter, the CPU 11 closes the window WD (step S150) and ends this control process.

  The process of step S130 executed by the CPU 11 corresponds to the second dictionary update unit 18 shown in FIG. The character image cutout unit 18a, the character code designation unit 18b, and the dictionary selection unit 18c included in the second dictionary update unit 18 are realized in each of automatic learning correction, semi-automatic learning correction, and manual correction, and feature amount acquisition is performed. The unit 18d is realized by the process of step S310, and the feature amount correcting unit 18e and the data adding unit 18f are realized by the process of step S350.

  The updated dictionary GU that is added / modified as described above is used in step S220 (FIG. 4) in the character recognition process. In this step S220, the identification information of the used dictionary and the identification information stored in the item F2 of the “corresponding dictionary identification information” of the update dictionary GU are compared, and only the record having the matching item F2 is updated. And only the selected record content is used for matching.

3. Action / Effect According to the present embodiment configured as described above, the feature amount is corrected for each dictionary by the updated dictionary GU. For this reason, when a character image and a correct character code are added to the updated dictionary GU due to misrecognition, the added character image is used for character recognition using a dictionary that is different from the erroneously recognized dictionary. It will not be done. Therefore, by adding a character image to the update dictionary GU, similar characters that have been recognized so far do not become unrecognizable, so that the accuracy of character recognition can be sufficiently increased.

  In this embodiment, the correction of the character recognition result can be performed according to various recognition result situations by automatic learning correction, semi-automatic learning correction, and manual correction. It can be updated and has excellent operability. At the time of manual correction, the update dictionary GU can be updated manually, that is, the character code and the dictionary can be manually specified, and the update intended by the operator can be easily performed.

4). Other embodiments:
The present invention is not limited to the above-described examples and embodiments, and can be implemented in various modes without departing from the gist of the present invention. For example, the following modifications are possible.

(1) In the above-described embodiment, the feature amount correcting unit 18e is configured by obtaining a simple average between the feature amount CV0 read from the corresponding dictionary and the feature amount CV1 of the character image that has been erroneously recognized. Alternatively, the feature amount read from the dictionary may be corrected with the feature amount of the character image that has been erroneously recognized by another method such as weighted averaging. In the case of weighted average, weighting that places importance on the feature amount CV0 read from the corresponding dictionary can be performed without significantly affecting the recognition accuracy before learning.

(2) Further, as one of the other methods described above, a comparison between CV0 and CV1 is performed for each small region representing the feature amount, so that a portion (CV0 It is also possible to extract a portion having a large difference from CV1 and correct only the extracted portion.

(3) In the above-described embodiment, when the same character code and the same corresponding dictionary identification information are corrected a large number of times in the updated dictionary GU, a large amount of data is used for averaging. There is a risk that there will be no dictionary data. For this reason, if the number of corrections for the same character code and the same corresponding dictionary identification information is limited, and the number of corrections exceeds the limit, the same character code and the same corresponding dictionary identification information are used. However, the record may be changed and registered. According to this configuration, the recognition accuracy can be further enhanced.

It is explanatory drawing which shows schematic structure of the computer system to which one Example of this invention is applied. It is a flowchart which shows the main routine of a control process. 12 is an explanatory diagram showing a window WD displayed on the display 20. FIG. It is a flowchart showing the detail of the character recognition process performed by step S120 of FIG. It is explanatory drawing for demonstrating automatic learning correction | amendment, semiautomatic learning correction | amendment, and manual correction using an example of a correction result. It is explanatory drawing which shows an example of the field FD2 for a recognition result display at the time of semiautomatic learning correction | amendment. It is explanatory drawing which shows dialog box DB for manual correction | amendment. It is explanatory drawing which shows the data structure of the update dictionary GU and the registration image information file PF. It is a flowchart which shows the update process of the update dictionary GU.

Explanation of symbols

10 ... Personal computer 11 ... CPU
12 ... Bus 13 ... Memory 14 ... Hard disk drive 15 ... Input control unit 16 ... Display control unit 18 ... Second dictionary update unit 18a ... Character image cutout unit 18b. ..Character code designation part 18c ... Dictionary selection part 18d ... Feature quantity acquisition part 18e ... Feature quantity correction part 18f ... Data addition part 20 ... Display 22 ... Keyboard 24 ... Mouse 26 ... Scanner 28 ... Printer WD ... Window BR1 ... Menu bar BR2 ... Toolbar BT1, BT2, BT3, BT4 ... Button DD ... Document image data FD1 ... Document Image display field FD2 ... Recognition result display field G1 ... Basic recognition dictionary G2 ... Shattered character recognition dictionary G3 ... Blurred character recognition dictionary G4 ... Mincho body recognition dictionary G5 .. .Gothic recognition dictionary GU ... update dictionary GW ... word dictionary PF ... registration Image information file PR ... Computer program DB ... Dialog box FD11 ... "Character image" display field FD12 ... "List" display field FD13 ... "Character" input field FD14 ... Dictionary input field

Claims (10)

As a dictionary for character recognition, a first dictionary whose registration contents are fixed in advance and a second dictionary that can update the registration contents are provided. A document image is input, and the character image represented in the document image is converted into the character image. A character recognition device that performs character recognition by collating with a first dictionary and a second dictionary,
A first dictionary storage means for storing a plurality of types of dictionaries that store character image feature quantities and character codes in association with each other as the first dictionary;
Second dictionary updating means for updating the second dictionary,
The second dictionary updating means includes
A character image cutting means for cutting out a character image of one character from the document image;
A character code specifying means for specifying a character code for the character image cut out by the character image cut-out means;
Dictionary selection means for selecting one dictionary from a plurality of types of dictionaries stored in the first dictionary storage means;
Feature quantity acquisition means for obtaining a feature quantity of the character image cut out by the character image cutout means;
The feature quantity corresponding to the character code designated by the character code designation means is read from the dictionary selected by the dictionary selection means, and the read feature quantity is the feature quantity obtained by the feature quantity acquisition means. A feature amount correcting means for correcting by,
A set of a character code designated by the character code designation unit, a feature amount after correction by the feature amount correction unit, and identification information for identifying the dictionary selected by the dictionary selection unit is set for one character. A character recognizing device comprising: data adding means for adding data to the second dictionary as data. The character recognition device according to claim 1,
Whether or not data for one character including the character code specified by the character code specifying means and identification information for identifying the dictionary selected by the dictionary selecting means has been registered in the second dictionary. Registration determination means for determining whether or not
A character recognition apparatus comprising: a data correction unit that corrects the registered data with the feature amount corrected by the feature amount correction unit when the registration determination unit determines that the registration is not completed. The character recognition device according to claim 1 or 2,
The character code specifying means is configured to specify a character code representing a character input in response to an operator's operation command,
The said dictionary selection means is a character recognition apparatus which is the structure which selects the dictionary designated according to the operator's operation command. The character recognition device according to claim 1 or 2,
A character image of one character is sequentially cut out from the document image, a feature amount of the cut out character image is calculated, and a use dictionary selected from a plurality of dictionaries stored in the first dictionary storage means, Character recognition means for recognizing the character by selecting a character code close to the feature value obtained by the calculation from the second dictionary;
The said dictionary selection means is a character recognition apparatus which is the structure which uses the said use dictionary as the selected dictionary. The character recognition device according to any one of claims 1 to 4,
The feature amount correcting means includes
A character recognition device comprising: averaging means for obtaining an average value of a feature value read from the dictionary and a feature value obtained by the feature value acquisition means. As a dictionary for character recognition, a first dictionary whose registration contents are fixed in advance and a second dictionary whose registration contents can be updated are prepared, a document image is input, and a character image represented in the document image A character recognition method for recognizing characters by comparing them with the first dictionary and the second dictionary,
(A) a step of previously storing a plurality of types of dictionaries for storing the character image feature quantity and the character code in association with each other as the first dictionary;
(B) a step of updating the second dictionary;
The step (b) is:
(B-1) a step of cutting out one character image from the document image;
(B-2) a step of designating a character code for the character image cut out in the step (b-1);
(B-3) a step of selecting one dictionary from a plurality of types of dictionaries as the first dictionary;
(B-4) a step of obtaining a feature amount of the character image cut out in the step (b-1);
(B-5) A feature amount corresponding to the character code specified in the step (b-2) is read from the dictionary selected in the step (b-3), and the read feature amount is A process of correcting with the feature amount obtained by the process (b-4);
(B-6) The character code designated by the step (b-2), the corrected feature amount by the step (b-5), and the dictionary selected by the step (b-3) are identified. And a step of adding a set of identification information to the second dictionary as data for one character. For performing character recognition by inputting a document image and collating the character image represented in the document image with a first dictionary whose registration contents are fixed in advance and a second dictionary whose registration contents can be updated. A computer program,
(A) a function of preliminarily storing in the storage device a plurality of types of dictionaries that store character image feature quantities and character codes in association with each other as the first dictionary;
(B) causing the computer to realize the function of updating the second dictionary;
The function (b) is
(B-1) a function of cutting out one character image from the document image;
(B-2) a function of designating a character code for the character image cut out by the function (b-1);
(B-3) a function of selecting one dictionary from a plurality of types of dictionaries as the first dictionary;
(B-4) a function of obtaining a feature amount of the character image cut out by the function (b-1);
(B-5) The feature amount corresponding to the character code designated by the function (b-2) is read from the dictionary selected by the function (b-3), and the read feature amount is A function of correcting with the feature amount obtained by the function (b-4);
(B-6) The character code designated by the function (b-2), the feature value after the correction by the function (b-5), and the dictionary selected by the function (b-3) are identified. And a function of adding a set of identification information to the second dictionary as data for one character. The computer program according to claim 7, further comprising:
(B-7) Data for one character comprising the character code designated by the function (b-2) and identification information for identifying the dictionary selected by the function (b-3) A function for determining whether or not it is already registered in the dictionary 2;
(B-8) a function of correcting the registered data with the feature amount corrected by the function (b-5) when it is determined that the function (b-7) has not been registered. Computer program to be realized. A computer program according to claim 7 or 8,
The function (b-5) is
A computer program comprising a function for obtaining an average value of a feature value read from the dictionary selected by the function (b-3) and a feature value obtained by the function (b-4).   10. A computer-readable recording medium on which the computer program according to claim 7 is recorded.

Images