JP4116688B2 - Dictionary learning method and character recognition device - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a character recognition device, and more particularly to a dictionary learning method and a character recognition device that create a feature amount from image data of an input character and obtain a recognized character code by comparison with a feature amount of a recognition dictionary.
[0002]
[Prior art]
A character recognition device is a device that obtains a character code by reading a character image and converting it into an electrical signal to obtain an image image, cutting it out character by character, and performing character recognition using a recognition dictionary.
In the above character recognition process, first, the extracted feature amount of the extracted character is compared with the extracted feature amount and the feature amount of each template in the recognition dictionary, and the feature amount having the highest similarity is obtained. The candidate character is regarded as a correct character and its character code is acquired.
However, even if the character to be read is the same character, the shape, the size, the thickness of the line is different, or the line is blurred.
In particular, when recognizing handwritten characters, in addition to the above factors, factors such as personal writing habits are added to lower the recognition rate.
[0003]
  As a solution to the above-described problem, a method of updating (correcting / adding) a standard pattern registered in the recognition dictionary at the time of correction input of an unread character or a misrecognized character, so-called 'learning the recognition dictionary' There is a method to improve the recognition rate. As a conventional technology related to the recognition rate learning method,(1)Whether the character is to be learned one by one at the time of correction input as in the technique disclosed in “Character Recognition Device” of Japanese Patent Laid-Open No. 2-186484 or “Learning Method of Recognition Dictionary” of Japanese Patent Laid-Open No. 9-185682 (2) Learning simultaneously with correction as in the technique disclosed in “Character recognition device and dictionary learning method and dictionary creation method thereof” of Japanese Patent Laid-Open No. 7-129724 The method of performing,(3)As in the technique disclosed in “Handwritten Character Recognition Dictionary Creation Method and Device” of Japanese Patent Laid-Open No. 7-271919, “Optical Character Reading Device” of Japanese Patent Laid-Open No. 8-287191, etc. There is a method of checking validity by recognition simulation.
[0004]
[Problems to be solved by the invention]
  However, the above(1)In this technique, there is a problem that an operator's work load increases and a highly accurate 'learning of recognition dictionary' cannot be performed unless the operator has specialized knowledge for determining the necessity of character learning.(2)This technology does not consider the validity of learning characters, so there is a problem that the efficiency of learning and the stability of the recognition dictionary after learning are lacking.(3)This technique has the disadvantage that it takes time to simulate the validity of the recognition dictionary after learning, and it is necessary to add storage means for storing data necessary for the recognition simulation.
[0005]
  The present invention has been made in view of the above-described problems and disadvantages, and is a dictionary learning that can efficiently learn a recognition dictionary and can give high validity and stability to the recognized dictionary after learning.MethodAn object is to provide a character recognition device.
[0006]
[Means for Solving the Problems]
  In order to solve the above problem, the dictionary learning method according to the first aspect of the present invention is a dictionary learning method for a character recognition apparatus that performs learning processing of a recognition dictionary. Corrected from the entered characterfollowingExclusion conditionsAny of (1) to (5)Exclude the characters that matchStudyA feature amount extracting step for extracting a feature amount of the custom character c, a recognition distance calculating step for calculating a minimum recognition distance and a maximum recognition distance for each character code from the feature amount of the learned character c and the feature amount of the template of the recognition dictionary, and recognition When the normalized distance of the minimum recognition distance A of the character code different from the learning character c in the minimum recognition distance of each character code calculated by the distance calculation step is smaller than the threshold ω, the learning character c is excluded and calculated. A filter step of obtaining the learning character C by excluding the learning character c when the normalized distance of the maximum recognition distance B of the same character code as the learning character c is larger than the threshold ω among the maximum recognition distances of the character codes; It is characterized by comprising.
  Exclusion conditions: (1) character inserted by correction input, (2) correction target character deleted by correction input, (3) correction target character when the correction target character is in contact with the characters before and after it, (4) The correction target character when the size of the correction target character is smaller than the threshold, (5) The correction target character when the correction target character is drowned
[0007]
  Further, the invention according to claim 2 is the dictionary learning method according to claim 1, wherein the learning character C obtained by the filter step is further converted into a learning pair.ElephantA selection step of selecting a learning character D that has not yet been subjected to a learning process and has a minimum minimum recognition distance.Features of learning character D and features of learning target templateComposition ratio R of learning character DChangeWhile lettingAs a vector quantity based on the following formula GA feature amount combining step of combining and bringing the feature amount of the classified template close to the feature amount of the learning character D.
  Formula G: Fdic = (1-R) Fdic + RFunk
Fdic Is the feature after synthesis, Fdic Is the feature quantity of the learning target template in the recognition dictionary, Funk Is the character feature, R Is the composite ratio (0 < R <1)
[0008]
  Further, the invention according to claim 3 is the dictionary learning method according to claim 2, further comprising setting the minimum recognition distance of the learning character D at the end of learning in the dictionary learning step.Whether or not the learning character D is correctly recognizedAn update / add step for updating or adding the recognition dictionary data based on the determination result, and a learning character recognition process using a template in which the recognition dictionary data is updated or added by the update / add step. And a filter data updating step for changing filter data used in the learning character filter means based on the recognized distance data.
[0009]
  According to a fourth aspect of the present invention, there is provided a character recognition device that performs a learning process of a recognition dictionary, from a character that has been corrected and input to a recognition result.followingExclusion conditionsOne of (1) to (5)Exclude the characters that matchStudyFeature amount extraction means for extracting the feature amount of the custom character c, and recognition distance calculation means for calculating the minimum recognition distance and the maximum recognition distance for each character code from the feature amount of the learned character c and the feature amount of the template of the recognition dictionary. If the normalized distance of the minimum recognition distance A of the character code different from the learning character c in the minimum recognition distance of each character code calculated by the recognition distance calculation means is less than the threshold ω, the learning character c is excluded, A filter that obtains a learning character C by excluding the learning character c when the normalized distance of the maximum recognition distance B of the same character code as the learning character c is larger than the threshold ω among the calculated maximum recognition distances of the character codes And learning character C obtained by the filter meansElephantSelecting means for selecting learning characters D that have not yet been subjected to learning processing and have the smallest minimum recognition distance;Features of learning character D and features of learning target templateComposition ratio R of learning character DChangeWhile lettingAs a vector quantity based on the following formula GA feature amount synthesizing unit that synthesizes and approximates the feature amount of the classified template to the feature amount of the learning character D; and a minimum recognition distance of the learning character D at the end of the synthesis process by the feature amount synthesizing unit.Whether or not the learning character D is correctly recognizedUpdate / addition means for updating or adding recognition dictionary data based on the determination result, and learning character recognition processing using a template in which the recognition dictionary data is updated or added by the update / addition means Filter data updating means for changing filter data used in the learning character filter means based on the obtained recognition distance data.
Exclusion conditions: (1) character inserted by correction input, (2) correction target character deleted by correction input, (3) correction target character when the correction target character is in contact with the characters before and after it, (4) The correction target character when the size of the correction target character is smaller than the threshold, (5) The correction target character when the correction target character is drowned
  Formula G: Fdic = (1-R) Fdic + RFunk
Fdic Is the feature after synthesis, Fdic Is the feature quantity of the learning target template in the recognition dictionary, Funk Is the character feature, R Is the composite ratio (0 < R <1)
[0010]
  According to a fifth aspect of the present invention, there is provided a computer-readable recording medium on which a program according to the present invention is recorded.followingExclusion conditionsOne of (1) to (5)Exclude the characters that matchStudyFeature amount extraction means for extracting the feature amount of the custom character c, recognition distance calculation means for calculating the minimum recognition distance and the maximum recognition distance for each character code from the feature amount of the learning character c and the feature amount of the template of the recognition dictionary, recognition When the normalized distance of the minimum recognition distance A of the character code different from the learning character c in the minimum recognition distance of each character code calculated by the distance calculation means is smaller than the threshold ω, the learning character c is excluded and calculated. Filter means for obtaining the learning character C by excluding the learning character c when the normalized distance of the maximum recognition distance B of the same character code as the learning character c in the maximum recognition distance of each character code is larger than the threshold ω, Learning character C obtained by the filter meansElephantSelecting means for selecting learning characters D that have not yet been subjected to learning processing and have the smallest minimum recognition distance;Features of learning character D and features of learning target templateComposition ratio R of learning character DChangeWhile lettingAs a vector quantity based on the following formula GA feature amount synthesizing unit that synthesizes and approximates the feature amount of the template classified as the feature amount of the learning character D, and a minimum recognition distance of the learning character D at the end of the synthesizing process by the feature amount synthesizing unitWhether or not the learning character D is correctly recognizedUpdate / addition means for updating or adding recognition dictionary data based on the determination result, and learning character recognition processing using a template in which the recognition dictionary data is updated or added by the update / addition means. The computer-readable recording medium which recorded the program which functions as a filter data update means to change the filter data used when obtaining the learning character C based on the recognized distance data.
Exclusion conditions: (1) character inserted by correction input, (2) correction target character deleted by correction input, (3) correction target character when the correction target character is in contact with the characters before and after it, (4) The correction target character when the size of the correction target character is smaller than the threshold, (5) The correction target character when the correction target character is drowned
  Formula G: Fdic = (1-R) Fdic + RFunk
Fdic Is the feature after synthesis, Fdic Is the feature quantity of the learning target template in the recognition dictionary, Funk Is the character feature, R Is the composite ratio (0 < R <1)
[0015]
DETAILED DESCRIPTION OF THE INVENTION
<Configuration example of character recognition device>
FIG. 1 is a block diagram showing a configuration example of a character recognition device according to the present invention, (a) is a block diagram showing a configuration example of a character recognition device, and (b) is a configuration of a dictionary learning unit which is a main part of the present invention. It is a block diagram which shows an example.
In FIG. 1A, a character recognition device 100 includes an optical reading device, a handwritten character reading device, etc., reads a character (or a line, a dot), converts it into an electric signal, and further digitizes the image data. Character input unit 1 to be obtained, character cutout unit 2 that decomposes the image data from the character input unit 1 into character images in units of one character, and feature extraction to extract features of the extracted character image And a character recognition unit 4 that compares a feature amount of the input character with a feature amount of each template in the recognition dictionary 9 and outputs a recognition result (for example, a character code).
[0016]
Furthermore, the character recognition device 100 uses the character recognition unit when learning the recognition dictionary 9.4Recognition result storage unit 5 that temporarily stores the recognition result output from the computer, and a recognition result correction unit that displays the recognition result as characters on the monitor screen, and an operator corrects and inputs unread characters or misrecognized characters by keyboard operation. 6, a learning character extraction unit 7,Dictionary learningThe unit 8 and the recognition dictionary 9 are included. A learning character extraction unit 7;Dictionary learningThe unit 8 and the recognition dictionary 9 will be described in detail below.
[0017]
Moreover, the character recognition apparatus 100 has a control unit (not shown). The control unit has a microprocessor configuration having a CPU, a RAM, a ROM, and the like, and controls the operation of the entire character recognition device.
Note that the character input unit 1 may have a dedicated control unit as a reading device. In addition, the character extraction unit 2 to the dictionary learning unit 8 can be configured by a hardware circuit, but one of the modules of the character extraction unit 2 to the dictionary learning unit 8 is a hardware circuit, The module may be configured by a program. When a part of the character extraction unit 2 to the dictionary learning unit 8 is configured by a program, each module is recorded on a recording medium such as a ROM and is controlled by the CPU of the control unit under the control of the control program. Each processing is realized.
[0018]
In this embodiment, the character extraction unit 2 to the character recognition unit 4 and a part of the recognition result correction unit 6, the learning character extraction unit, and the dictionary learning unit 8 are configured by a program. The recognition result storage unit 5 is an internal memory such as a RAM, and the hardware part of the recognition result correction unit 6 is a device including a display device having a monitor screen and a keyboard.
[0019]
<Learning character extraction unit>
The learning character extraction unit 7 automatically excludes characters that satisfy certain exclusion conditions from characters to be corrected by the recognition result correction unit 6 (hereinafter, correction target characters), and learning that requires learning of other characters Outputs data stored as characters.
In the embodiment, the exclusion conditions in the learning character extraction unit 7 are the following five conditions, and characters other than the characters corresponding to any of these five conditions are the learning characters.
[0020]
  Exclusion conditions;
(I)Characters inserted by correction input,(B)Characters to be corrected deleted by correction input,(C)The correction target character when the correction target character is in contact with the preceding and following characters, (d) the correction target character when the size of the correction target character is smaller than the threshold, and (e) the correction target character In case the above correction target character(I), (B), (C)This exclusion condition is provided in order to exclude these characters as learning characters because they may have failed to be cut out.(B), (b)Can be determined based on information obtained from the recognition result correction unit 6,(C)The exclusion condition can be determined based on information obtained from the character cutout unit 2. Also, above(D), (e)Since the exclusion condition of can be expected that the shape of the character is inferior, it is provided to exclude these characters as learning characters as inappropriate,(D)The exclusion condition can be determined based on information obtained from the character cutout unit 2, and the exclusion condition (e) can be determined from the number of black dots (dots) of the correction target character.
[0021]
<Recognition dictionary>
The recognition dictionary is recorded on a removable recording medium such as a hard disk, floppy disk, or optical disk, and any one of these recording medium read / write devices, that is, a magnetic disk device, a floppy disk device, an optical disk device, etc. It is a component of the recognition device.
[0022]
  FIG. 2 is an explanatory diagram showing the structure of the recognition dictionary 9, and the recognition dictionary 9 has a template group 91 and a filter data group 92. The template has a character code, a character feature amount corresponding to the character code, and a template number. The filter data includes an average value of the maximum recognition distance for each character code, its mean square value (variance value), and the recognition distance. It consists of the minimum value, its root mean square value (variance value), and the number of characters used to create these values. The filter data is obtained by performing the same recognition process as the character recognition unit 4 on the character data used to create the recognition dictionary 9 using the recognition dictionary. The recognition distance here is the feature amount of the character to be recognized.,This is the distance calculated using the feature values of the recognition dictionary having the same character code as a vector.
[0023]
In the dictionary learning method of the present invention, the learned recognition dictionary is stored according to the type of the input character source separately from the initial recognition dictionary (standard dictionary). With this method, if the input character has been learned before, the recognition dictionary at that time can be used for recognition, and if it is an unlearned character, the standard dictionary is used, so a stable recognition condition can always be obtained. it can.
As the type of input character source, in the case of an input character obtained by reading a handwritten form or an input character obtained by a handwriting input device, etc. (because it has individuality such as a character font), the writer itself Can be used as the source, and in the case of type, the book name or font can be used as the source.
[0024]
<Dictionary learningDepartment>

[Configuration] In FIG.Dictionary learningThe unit 8 includes a feature extraction unit 11, a recognition distance calculation unit 12, a learning character filter unit 13, a learning character selection unit 14, a learning character synthesis unit 15, a template addition unit 16, and a filter data change unit 17. The feature extraction unit 11 to the filter data change unit 17 can be configured by a hardware circuit, but are configured by a program in this embodiment. A certain module among the modules of the feature extraction unit 11 to the filter data change unit 17 may be configured by a hardware circuit, and the other modules may be configured by a program. When the feature extraction unit 11 to the filter data changing unit 17 are configured by a program, each module is recorded on a recording medium such as a ROM, and is executed and controlled by the CPU of the control unit under the control of the control program. Realize dictionary learning.
[0025]
[Dictionary learning operation]
  Figure 3Dictionary learning10 is a flowchart showing a dictionary learning operation of unit 8. In the feature extraction unit 11, among the plurality of learning characters output from the learning character extraction unit 7 in FIG.Learning charactersData ofSpeciallyIt is checked whether or not the collected amount is included (S1). If the feature amount is included, the process proceeds to S3.LearningCharacterFrom the dataA feature quantity is extracted (created) (S2).
[0026]
  Next, the recognition distance calculation unit 12 performs the characteristics of the learning character.amountThen, recognition distance data including the minimum recognition distance and the template number at that time, and the maximum recognition distance and the template number at that time is created for each character code from the template feature quantity in the recognition dictionary 9 (S3). FIG. 4 shows the structure of the recognition distance data.
[0027]
The learning character filter unit 13 automatically excludes learning characters unsuitable for learning without requiring operator judgment.
Specifically, among the normalized distances of the minimum recognition distances of the respective character codes derived by the below-described expression 2 (recognition distance normalization expression), the normalized distances and threshold values of the character codes different from the learning characters are set. If there is a character smaller than the threshold, the learning character is excluded. That is, such learning characters exist in the normal distribution of the minimum recognition distance of different character codes, and adding the learning characters to the recognition dictionary may affect the recognition of the different character codes. Exclude the learned characters as occurring. Similarly, the normalization distance of the maximum recognition distance of the character code of the learning character is obtained from Expression 2 and compared with a threshold value. If the normalization distance is larger than the threshold value, the learning character is excluded. That is, since such a learning character does not exist within the normal distribution of the maximum recognition distance of the character code, the learning character is excluded as having low reliability (S4). If there is no learning character as a result of the filtering process based on the comparison between the normalized distance and the threshold value, the process ends (S5).
[0028]
The comparison processing with the character code different from the learning character in the learning character filter unit 13 described above is performed to check whether or not the learning character is statistically sufficiently separated from the dictionary template of the different character code. That is, the minimum recognition distance of a character code in a learning character is the distance between the learning character and the feature vector closest to the learning character in the recognition dictionary template of the character code, and the distance is short. This indicates that the learned character is close to the character code (see FIG. 6).
If the learning character is too close to another character code, and the learning process is performed for that character, the recognition performance of other character codes that are too close will be adversely affected. Therefore, the learning character is excluded.
[0029]
Further, the comparison of the character codes of the learning characters is performed to check whether or not the learning characters are included in the statistical distribution of the character codes. In other words, since the maximum recognition distance in a character code approximates the radius of the character code's feature space, the value of a character is far from the average value. This means that it is far from the feature space of the code (see FIG. 7). Therefore, a character that is far from other character features of the same character code is excluded from learning because the character feature has low reliability and is not suitable for learning.
[0030]
Based on the above two exclusion conditions, characters that are inappropriate for learning, that is, characters that affect the recognition of other character codes and characters that are far from the average characteristics of the same character code are automatically determined without operator judgment. Can be excluded.
In addition, the reason for the appearance of characters that are close to the characteristics of other character codes or characters that are far from the average characteristics of the same character code is that the image of the character is inferior due to the curling, crushing, tilting, and deformation of the characters. In addition, the operator may input a character with an incorrect character code when correcting the recognition result.
[0031]
The learning character selection unit 14 automatically excludes effective learning characters without requiring the operator's judgment.
Specifically, the learning characters that have passed the filtering process in the learning character filter unit 13 are classified according to the recognition dictionary template (learning target template) to be learned, and the recognition processing is not performed yet and is correctly recognized. The learning character with the shortest distance is selected (S6).
[0032]
Here, the learning target template refers to a template when the learning character recognition distance data has the same character code as the learning character and the minimum recognition distance, that is, a template closest to the learning character and the same character code. In addition, the minimum recognition distance at that time is set as a positive recognition distance.
If the minimum recognition distance in the same character code as the character to be recognized is the positive recognition distance, and the smallest recognition distance of the character code different from the character to be recognized is the smallest recognition distance, the correct recognition distance> the erroneous recognition distance. In this case, the recognition result is erroneous recognition.
In addition, the recognition result is rejected when the erroneous recognition distance ≧ the correct recognition distance ≧ the erroneous recognition distance−α (α: threshold for determination of rejection). Further, when the erroneous recognition distance-α> the correct recognition distance, the recognition result is the correct recognition.
[0033]
  The learning character is divided into the learning target templates. The learning character with the same character code input from the same source and the same correct recognition distance dictionary template (learning target template)recognitiondictionaryofThis is because, when viewed from the feature vector of the template, it has the same direction as shown in FIG.
[0034]
Therefore, by learning from the learning characters that are close to the learning target template, the amount of change in the learning target template is such that the learning character is recognized by gradually bringing the characteristic amounts of the learning target template closer to those feature amounts. Can be kept to a minimum. Usually, since the recognition dictionary has a balanced template in an initial state (in a state where learning is not performed), learning can be performed without greatly losing the balance by using such a learning method. Further, since it is not necessary for the operator to make a judgment, it is not necessary to require the operator to have specialized knowledge related to character characteristics or learning.
[0035]
The learning character synthesis unit 15 combines the learning character feature quantity selected by the learning character selection unit 14 and the feature quantity of the learning target template by using a feature quantity synthesis method (Equation 3: see FIG. 5) described later. It is synthesized as a vector amount while gradually changing the ratio from the initial value, and the feature amount of the learning target template is brought closer to the learning character. Then, when any of the following three composition end determination conditions is satisfied during repeated composition, the feature amount composition processing is terminated (S7).
[0036]
The template adding unit 16 checks whether or not the learning character is correctly recognized at the end of the repetition of the feature amount synthesizing process. If it is correctly recognized, the process proceeds to S9 (S8). If the character is not recognized correctly (described later), the feature amount of the learning character is added as a new template of the recognition dictionary (S9). In other words, the learning method using only feature amount synthesis does not increase the number of templates, so it does not affect the recognition speed after learning.However, if the dictionary changes significantly from the original dictionary or becomes too close to other character codes, the recognition rate May be reduced. In addition, in the learning method of only adding a template, if the template is not too close to other character codes, the recognition rate does not decrease. However, the increase in the template causes the recognition speed to decrease. Therefore, in the present invention, priority is given to a learning method based on feature value synthesis (S7).), LaterWhen (B) (Equation 5) or (C) (Equation 6) is satisfied, the learning method is inappropriate, and the learning method is switched to a template addition method. As a result, it is possible to select an optimal learning method for the learning character without requiring the operator's judgment.
[0037]
The filter data changing unit 17 corrects the filter data of the recognition dictionary 9 when the learned characters are correctly recognized by S7, S8 (learning character synthesizing unit 15) or S9 (template adding unit 16). .
Specifically, learning character recognition processing is performed in the recognition dictionary after learning, and the average value of the minimum value / maximum value of the recognition distance used in the learning character filter unit 13 at this time and the mean square value thereof, and By adding to each of the number of characters, the average value of the minimum value / maximum value of the recognition distance, the mean square value thereof, and the number of characters are corrected to values suitable for the correction target character (S10).
[0038]
After completion of all learning learning processes (S7 to S9) selected by the learning character selection unit 14, the recognition distance calculation unit 12 creates recognition distance data again using the recognition dictionary after learning, and then learns. Only for the learning characters that are not recognized and are still misrecognized, all of the stored learning characters are recognized correctly, or until the characters that pass the learning character filter unit 13 no longer exist. The learning process of S10 (recognition distance calculation unit 12 to file data change unit 17) is repeated (S11).
[0039]
[Normalization of recognition distance, etc.]
The normalized distance of the minimum / maximum recognition distance to be compared with the threshold value by the learning character filter unit 13 can be obtained as follows.
The dispersion value var (dc0) of the recognition distance dc0 in the character code co is the average value ave (dc0) of the recognition distance stored in the recognition dictionary 9 and its mean square value ave (dc0).²) From the following formula 1.
var (dco) = ave (dco²)-(Ave (dco))²      (Formula 1)
Using the variance value of the minimum recognition distance of each character code obtained by Expression 1, the minimum recognition distance in each character code of the recognition distance data of the learning character obtained by the recognition distance calculation unit 12 is normalized according to Expression 2 below. .
d'co = (dco-ave (dco)) / (var (dco))^1/2    (Formula 2)
Here, dco is a minimum recognition distance in the character code co of the learning character, d′ co is a normalized distance of the minimum recognition distance in the character code co of the learning character, and ave (dco) is a character code stored in the recognition dictionary 9. The average value of the minimum recognition distances at co, var (dco) is the variance of the minimum recognition distances of the character code co calculated by the above equation 1.
[0040]
Further, using the variance value of the maximum recognition distance of each character code obtained by Expression 1 and the above Expression 2, the maximum recognition distance in each character code of the recognition distance data of the learning character obtained by the recognition distance calculation unit 12 is normalized. can do.
In this case, the maximum recognition distance in the character code co of the learning character is dco, the normalization distance of the maximum recognition distance in the character code co of the learning character is d'co, and the maximum recognition distance in the character code co stored in the recognition dictionary 9 Equation 2 may be applied with ave (dco) as the average value of and var (dco) as the variance value of the maximum recognition distance of the character code co calculated by Equation 1.
[0041]
Thus, the range of observed values in a certain normalized distribution can be obtained as the range of observed values normalized by Equation 2 using the average value and variance value of the sample. For example, if the reliability is 95%, −1.96 to 1.96 is given as a normalized observation value range (confidence interval) from the normalized distribution table. Thereby, it can be determined whether or not the recognition distance exists within the normal distribution of the recognition distance of the character code.
[0042]
[Characteristic composition method and composition completion judgment condition]
(1) Feature synthesis method
The feature amount of the learning character and the feature amount of the learning target template are synthesized by the following equation (Equation 3) as a vector amount while gradually changing the learning character composition ratio from the initial value. FIG. 5 shows the transition state of the feature amount that is transitioned by learning by synthesis.
Fdic '= (1-R) * Fdic + R * Funk (Formula 3)
Here, Funk is a feature amount of a learning character, Fdic is a feature amount of a learning target template in the recognition dictionary, Fdic ′ is a feature amount after the above-described template is synthesized, and R is a synthesis ratio (0 <R <1).
[0043]
(2) Composition Completion Judgment Condition As described above, the process in the learning character composition unit 15 is terminated when any of the following three conditions is satisfied during the repetitive composition according to the above expression 3.
(A)When the learning character is recognized correctly and the difference between the correct recognition distance and the erroneous recognition distance is larger than the threshold, that is, when the following expression (Expression 4) is satisfied: | Fdicerr-Funk |-| Fdic-Funk |> 1 (Formula 4)
However, Fdicerr is the feature vector of the misrecognition distance template, and 1 (> 0) is a threshold value.
(B)When the movement value before and after the synthesis of the learning target template is larger than the threshold value, that is, when the following expression (Expression 5) is satisfied: | Fdic−Fdic |> 2 (Expression 5)
However, α2 (> 0) is a threshold value.
(C)When the distance between the feature vector of the learning target template after learning and the feature vector of the character template different from the character code of the learning character in the recognition dictionary is smaller than the threshold, that is, when the following equation (Equation 6) holds: −Fdic | <3 (Formula 6)
However, α3 (> 0) is a threshold value.
[0044]
(A)There are cases where the composite end determination condition is satisfied, that is, the case where the correct recognition distance and the incorrect recognition distance are larger than the threshold value α1. This is because there is a certain amount of variation in the amount, so that the variation is absorbed by learning one learning character so that there is a margin in the correct recognition distance.
[0045]
(B)As described above, since the recognition dictionary usually has a balanced template in the initial state, as described above, if the existing template in the recognition dictionary moves greatly, the balance is greatly lost, and the character code is recognized. This is to prevent the rate from decreasing.
[0046]
Also,(C)The combination end determination condition is a condition provided to prevent the learning target template from being too close to another character code template and adversely affecting the recognition of the character code.
[0047]
[Minimum recognition distance]
  FIG. 6 is an explanatory diagram of the minimum recognition distance. In FIG. 6, the feature vector of the learning character X is A (triangle mark),FIG.Character code41The distribution range of the feature vector of the right circle (broken line) a,FIG.Character code42The distribution range of the feature vector is set to the large circle (dashed line) d on the left side, and the character code41The character vector of the character is b (cross), the feature vector of the dictionary template is c (circle), and the dictionary template closest to the learning character is B (double circle).42E (cross mark), the feature value vector of the dictionary template is f (circle mark), and the dictionary template closest to the learning character is D (double circle mark). In this case, the learning character and its character code41Is the length of the line segment indicated by C, and the learning character XWhenThe character code42Is the length of the line segment indicated by E, and since C <E, the minimum recognition distance of the learning character X is the length of the line segment C.
[0048]
[Maximum recognition distance]
  FIG. 7 is an explanatory diagram of the maximum recognition distance. In FIG. 7, the learning character is A (triangle mark),4Character code41The feature vector distribution range is a circle (dashed line) a, and the character code41B is the feature vector of the dictionary, c is the feature vector of the dictionary template, c is the maximum recognition space (≈feature space radius), and B is the dictionary template most distant from the learning character A (double Circle)), character code in learning character A41Is the length of the line segment BA.
[0049]
[Direction of learning characters]
  FIG. 7 is an explanatory diagram of the directionality of learning characters. In FIG.FIG.Character code41If the distribution range of the feature vector of the right side is a circle (dashed line) a, the feature vector of the character code 41 is b (cross), and the feature vector of the dictionary template is c (circle), a certain template (black circle) The learning character group A having a minimum recognition distance as a template exists in a certain angle range centering on the template, and the learning character group B having another template (double circle) as a minimum recognition distance template is It exists in a certain angle range centering on the template (that is, it has directionality).
[0050]
【The invention's effect】
  As explained above,According to the dictionary learning method of the invention of claim 1, the character recognition method of claim 4, and the invention of claim 5, characters corresponding to a predetermined exclusion condition are detected from the corrected input characters of the recognition result. Since the minimum recognition distance and the maximum recognition distance are calculated after extracting the feature value from the learning characters obtained by exclusion, it is possible to prevent the recognition distance from being calculated using the correction target character inappropriate as the learning character as the learning character. And the calculation speed of the recognition distance is fast because no unnecessary calculation is performed. In addition, since the learning characters that meet the predetermined conditions can be excluded based on the calculated minimum recognition distance and maximum recognition distance, it is possible to exclude learning characters that have too long recognition distance and adversely affect the recognition performance of other character codes. Characters that are too far apart from other character features of the same character code can be excluded from learning.
[0051]
According to the dictionary learning method of the invention of claim 2, the character recognition method of claim 4, and the invention of claim 5, further learning is performed from the learning characters that are close to the learning target template. Therefore, the amount of change in the learning target template can be minimized by recognizing learning characters by gradually bringing the characteristic amounts of the learning target template closer to those feature amounts.
[0052]
According to the dictionary learning method of the invention of claim 3, the character recognition method of claim 4, and the invention of claim 5, the dictionary further changes greatly from the original dictionary or other characters Even if the code gets too close, it is possible to suppress a decrease in the recognition rate.
In addition to the initial recognition dictionary (standard dictionary), the learned recognition dictionary is stored according to the type of the input character source, and if the input character has been learned before, the recognition dictionary at that time is used for recognition. If the character is unlearned, a standard dictionary is used, so that a stable recognition condition can always be obtained.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration example of a character recognition device of the present invention.
FIG. 2 is a structural diagram showing the structure of a recognition dictionary.
FIG. 3 is a flowchart showing a dictionary learning operation of a learning dictionary unit.
FIG. 4 is an explanatory diagram showing a structure of recognition distance data.
FIG. 5 is an explanatory diagram illustrating a transition state of a feature amount that is transitioned by learning by synthesis.
FIG. 6 is an explanatory diagram of a minimum recognition distance.
FIG. 7 is an explanatory diagram of a maximum recognition distance.
FIG. 8 is an explanatory diagram of the directionality of learning characters.
[Explanation of symbols]
7 Learning character extraction unit (learning character extraction means)
9 recognition dictionary
13 Learning character filter section (first filter means, second filter means)
14 Learning character selection part (learning character selection means)
15 Learning character synthesis unit (feature amount synthesis means)
16 Template addition part (dictionary learning method judging means)
91 templates
92 Filter data (recognition dictionary data)
100 character recognition device

Claims (5)

A dictionary learning method of a character recognition device that performs learning processing of a recognition dictionary,
By the computer of the character recognition device,
Each character from the feature quantity of the feature with recognition dictionary template learning character c excluding the character that corresponds to one of the recognition result for fixed character entered from the following exclusion conditions (1) to (5) A recognition distance calculating step for calculating a minimum recognition distance and a maximum recognition distance for each code;
If the normalized distance of the minimum recognition distance A of the character code different from the learning character c in the minimum recognition distance of each character code calculated in the recognition distance calculation step is smaller than the threshold ω, the learning character c is excluded. When the normalized distance of the maximum recognition distance B of the same character code as the learning character c is greater than the threshold ω among the calculated maximum recognition distances of the character codes, the learning character C is excluded and the learning character C is excluded. Get the filter step,
A dictionary learning method characterized by comprising:
Exclusion conditions: (1) character inserted by correction input, (2) correction target character deleted by correction input, (3) correction target character when the correction target character is in contact with the characters before and after it, (4) The correction target character when the size of the correction target character is smaller than the threshold, (5) The correction target character when the correction target character is drowned Furthermore, selection step of the learning letter C obtained by filtering step classifies the learning pair Zote Plates another, minimum recognition distance and without to the still learning process from its selects the smallest learned character D,
The feature amount of the learning character D and the feature amount of the learning target template are combined as a vector amount based on the following formula G while changing the composition ratio R of the learning character D, and A feature value synthesis step to bring the feature values closer,
The dictionary learning method according to claim 1, further comprising:
Formula G: Fdic = (1-R) Fdic + RFunk
Fdic is the feature value after composition, Fdic is the feature value of the learning target template in the recognition dictionary, Funk is the character feature value, and R is the composition ratio (0 < R <1) Furthermore, at the end of learning in the dictionary learning step, it is determined whether or not the learning character D is correctly recognized using the minimum recognition distance of the learning character D , and the recognition dictionary data is updated or added based on the determination result. Update / add steps to perform,
Filter data update for changing filter data used in the learning character filter means based on recognition distance data obtained by performing recognition processing of learning characters using a template in which recognition dictionary data is updated or added in the updating / adding step. Step,
The dictionary learning method according to claim 2, further comprising: In a character recognition device that performs learning processing of a recognition dictionary,
Each character from the feature quantity of the feature with recognition dictionary template learning character c excluding the character that corresponds to one of the recognition result for fixed character entered from the following exclusion conditions (1) to (5) A recognition distance calculation means for calculating a minimum recognition distance and a maximum recognition distance for each code;
The learning character c is excluded when the normalized distance of the minimum recognition distance A of the character code different from the learning character c among the minimum recognition distances of the character codes calculated by the recognition distance calculation means is smaller than the threshold ω. When the normalized distance of the maximum recognition distance B of the same character code as the learning character c is greater than the threshold ω among the calculated maximum recognition distances of the character codes, the learning character C is excluded and the learning character C is excluded. Filter means to obtain
Selection means for said learning letter C obtained by the filter means to classify the learning pair Zote Plates another, minimum recognition distance and without to the still learning process from its selects the smallest learned character D,
The feature amount of the learning character D and the feature amount of the learning target template are combined as a vector amount based on the following formula G while changing the composition ratio R of the learning character D, and A feature amount combining means for bringing the feature amounts closer;
At the end of the synthesizing process by the feature amount synthesizing means, it is determined whether or not the learning character D is correctly recognized using the minimum recognition distance of the learning character D , and the recognition dictionary data is updated or added based on the determination result. Update / add means to perform;
Filter data update for changing filter data used in the learning character filter means based on recognition distance data obtained by performing recognition processing of learning characters using a template in which the recognition dictionary data is updated or added by the updating / adding means Means,
A character recognition device comprising:
Exclusion conditions: (1) character inserted by correction input, (2) correction target character deleted by correction input, (3) correction target character when the correction target character is in contact with the characters before and after it, (4) The correction target character when the size of the correction target character is smaller than the threshold, (5) The correction target character when the correction target character is drowned
Formula G: Fdic = (1-R) Fdic + RFunk
Fdic is the feature value after composition, Fdic is the feature value of the learning target template in the recognition dictionary, Funk is the character feature value, and R is the composition ratio (0 < R <1) In a character recognition device that performs learning processing of a recognition dictionary,
Computer
Each character from the feature quantity of the feature with recognition dictionary template learning character c excluding the character that corresponds to one of the recognition result for fixed character entered from the following exclusion conditions (1) to (5) Recognition distance calculation means for calculating the minimum recognition distance and the maximum recognition distance for each code,
The learning character c is excluded when the normalized distance of the minimum recognition distance A of the character code different from the learning character c among the minimum recognition distances of the character codes calculated by the recognition distance calculation means is smaller than the threshold ω. When the normalized distance of the maximum recognition distance B of the same character code as the learning character c is greater than the threshold ω among the calculated maximum recognition distances of the character codes, the learning character C is excluded and the learning character C is excluded. Obtaining filter means,
Selection means for said learning letter C obtained by the filter means to classify the learning pair Zote Plates another, minimum recognition distance and without to the still learning process from its selects the smallest learned character D,
The feature amount of the learning character D and the feature amount of the learning target template are combined as a vector amount based on the following formula G while changing the composition ratio R of the learning character D, and A feature amount combining means for bringing the feature amounts closer;
At the end of the synthesizing process by the feature amount synthesizing means, it is determined whether or not the learning character D is correctly recognized using the minimum recognition distance of the learning character D , and the recognition dictionary data is updated or added based on the determination result. Updating / adding means,
A filter for changing filter data used when obtaining the learning character C based on recognition distance data obtained by performing recognition processing of the learning character using a template in which the recognition dictionary data is updated or added by the updating / adding means. Data update means,
A computer-readable recording medium storing a program that functions as a computer.
Exclusion conditions: (1) character inserted by correction input, (2) correction target character deleted by correction input, (3) correction target character when the correction target character is in contact with the characters before and after it, (4) The correction target character when the size of the correction target character is smaller than the threshold, (5) The correction target character when the correction target character is drowned
Formula G: Fdic = (1-R) Fdic + RFunk
Fdic is the feature value after composition, Fdic is the feature value of the learning target template in the recognition dictionary, Funk is the character feature value, and R is the composition ratio (0 < R <1)

Images