JPH11184976A - Dictionary learning system and character recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a dictionary learning system and a character recognition device that efficiently perform learning of a recognition dictionary and, at the same time, can give high validity and stability to the recognition dictionary after learning. SOLUTION: A learning character extraction part 7 automatically deletes a character pertinent to deletion condition from a character of a correction object. A learning character filter part 13 of a learning dictionary part 8 automatically deletes a learning character inappropriate to learning by taking recognition distance as reference; moreover, a learning character selection part 14 classifies the learning character into a template (a learning object template) of the recognition dictionary of the learning object, selects the learning character which is not learnt processed by taking the recognition distance as reference, a learning character synthesis part 15 gradually changes the synthesis ratio of the feature amount of the learning character and the feature amount of the learning object template from their initial values and synthesizes them into vector amount and brings the feature amount of the learning object template closer to the learning character. At the time of synthesis processing completion at a template addition part 16, necessity/unnecsessity of addition of the template is decided by taking the minimum recognition distance of the learning character as reference, a filter data change part 17 updates/adds recognition dictionary data. A recognition dictionary 9 retains a learnt dictionary in addition to a standard dictionary in accordance with kinds of an inputted character source.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition apparatus and, more particularly, to a dictionary learning method and a character learning method for generating a characteristic amount from image data of an input character and obtaining a character code recognized by comparing the characteristic amount with a characteristic amount of a recognition dictionary. It relates to a recognition device.

[0002]

2. Description of the Related Art A character recognizing device is a device for reading a character image and converting it into an electric signal to obtain an image image, extracting the image one character at a time, performing character recognition using a recognition dictionary, and obtaining a character code. is there. In the above character recognition process, first,
The feature amount of the extracted character is extracted, and the extracted feature amount is compared with the feature amount of each template in the recognition dictionary.
The candidate character having the feature amount having the highest similarity is determined as the correct character, and the character code is obtained. However, the characters to be read cannot always be accurately recognized because the same characters have different shapes, sizes, and different line thicknesses, or have blurred lines. In particular, when recognizing handwritten characters, in addition to the above factors, factors such as personal writing habits are added, and the recognition rate is reduced.

As a solution to the above-mentioned problem, a method of updating (correcting / adding) a standard pattern registered in a recognition dictionary when correcting and inputting an unread character or an erroneously recognized character, that is, a so-called 'recognition dictionary. There is a method to improve the recognition rate by 'learning'. As a prior art relating to a learning method of a recognition rate, Japanese Patent Application Laid-Open No. 2-186484 discloses a "character recognition apparatus".
A method in which an operator determines whether or not to learn each character at the time of correction input as in a technique disclosed in “Learning method of recognition dictionary” in Japanese Patent Application Laid-Open No. Hei 9-185682 or the like, and gives an instruction. Japanese Patent Laid-Open No. 7-129724, "Character Recognition Apparatus and Dictionary Learning Method and Dictionary Creation Method", a method of performing learning simultaneously with correction. A method of examining the validity of a recognition dictionary after learning by recognition simulation, such as the technique disclosed in "Dictionary Creation Method and Apparatus" and "Optical Character Reading Device" in Japanese Patent Application Laid-Open No. 8-287191.

[0004]

However, according to the above-mentioned technique, the accuracy of the operation is high unless the operator has the technical load for determining the necessity of the character learning and the point that the workload of the operator increases. There is a problem that the 'learning of the recognition dictionary' cannot be performed, and the technology does not consider the validity of the learning character, so there is a problem that the learning efficiency and the stability of the recognition dictionary after learning are lacking. However, the technique of (1) has a disadvantage that it takes time to simulate the validity of the recognition dictionary after learning, and that it is necessary to add a storage unit for storing data necessary for the recognition simulation.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems and disadvantages, and has a dictionary learning method capable of efficiently learning a recognition dictionary and providing the learned dictionary with high validity and stability after learning. It is intended to provide a character recognition device.

[0006]

In order to solve the above-mentioned problems, a dictionary learning method according to the present invention provides a character recognition device for performing a learning process of a recognition dictionary. Automatically removing based on the information.

[0007] In the dictionary learning method,
A step of classifying learning characters for each template of a recognition dictionary to be learned, a step of selecting learning characters that have not yet undergone learning processing based on the recognition distance, a feature amount and classification of the selected learning characters And a step of bringing the classified template feature amount closer to the learning character by performing synthesis while gradually changing the synthesis ratio with the obtained template feature amount.

[0008] Further, in the dictionary learning method,
At the end of the synthesizing process, the method includes a step of determining whether or not to add a template based on the minimum recognition distance of the learning character and updating or adding the recognition dictionary data.

Each of the dictionary learning methods has a step of automatically excluding a character corresponding to a predetermined exclusion condition from characters to be corrected to obtain a learning character.

Further, the character recognition device of the present invention is a character recognition device for performing a learning process of a recognition dictionary, wherein the normalized distance between each character code different from the learning character and the learning character is smaller than the first threshold value. First filter means for excluding the learning character as a learning character unsuitable for learning of the recognition dictionary; and a learning character when the normalized distance between the same character code as the learning character and the learning character is larger than a second threshold value. And second filter means for excluding as a learning character unsuitable for learning a dictionary.

Further, the character recognition apparatus further classifies the learning characters according to templates of a recognition dictionary to be learned, and selects learning characters that have not yet undergone learning processing from the classification based on the recognition distance. Selecting means, and characteristic amount synthesizing means for bringing the characteristic amount of the classified template closer to the learning character by changing the combination ratio between the characteristic amount of the selected learning character and the classified template characteristic amount. It is characterized by having.

Further, in the above character recognition device,
It is provided with a dictionary learning method determining means for determining whether or not to add a template based on the minimum recognition distance of the learning character at the end of the combining process, and updating or adding the recognition dictionary data based on the determination result. I do.

Further, each of the above character recognition devices has a learning character extracting means for obtaining a learning character by automatically excluding a character corresponding to a predetermined exclusion condition from characters to be corrected.

Further, in the above-mentioned character recognition device, when a character to be corrected satisfies any one of the following exclusion conditions, the learning character extracting means automatically excludes the character to be corrected. The character inserted by the correction input, the correction target character deleted by the correction input, the correction target character when the correction target character is in contact with the character before and after it, and the correction target character when the size of the correction target character is smaller than the threshold value. The correction target character, or the correction target character when the correction target character is blurred.

[0015]

DESCRIPTION OF THE PREFERRED EMBODIMENTS <Configuration Example of Character Recognition Apparatus> FIG. 1 is a block diagram illustrating a configuration example of a character recognition apparatus according to the present invention. FIG. 1A is a block diagram illustrating a configuration example of a character recognition apparatus.
FIG. 2B is a block diagram illustrating a configuration example of a dictionary learning unit that is a main part of the present invention. In FIG. 1A, the character recognition device 100
Is composed of an optical reader, a handwritten character reader, etc., and reads a character (or a line or a dot), converts it into an electric signal, and further digitizes it to obtain image data; Character extracting unit 2 for decomposing image data from a character into character images in units of one character, a characteristic extracting unit 3 for extracting characteristics of the extracted character image to obtain a characteristic amount, and a characteristic amount of an input character Is compared with the feature amount of each template in the recognition dictionary 9 and a recognition result (for example,
And a character recognition unit 4 for outputting a character code.

Further, the character recognition device 100 includes a recognition dictionary 9
A recognition result storage unit 5 for temporarily storing a recognition result output from the character recognition unit 8 when learning is performed, and displaying the recognition result as a character on a monitor screen so that the operator can recognize an unread character or a misrecognized character. It has a recognition result correcting unit 6 for correcting and inputting by a keyboard operation, a learning character extracting unit 7, a learning dictionary unit 8, and a recognition dictionary 9. The learning character extracting unit 7, the learning dictionary unit 8, and the recognition dictionary 9 will be described in detail below.

The character recognition device 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 separate dedicated control unit as a reading device. Further, the character extracting unit 2 to the dictionary learning unit 8 can be configured by a hardware circuit, but a certain module among the modules of the character extracting unit 2 to the dictionary learning unit 8 is configured by a hardware circuit. The module may be configured by a program. When a part of the character extracting 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 executed and controlled by the CPU of the control unit under the control of the control program. To realize each processing.

In this embodiment, the character extracting section 2 to the character recognizing section 4 and a part of the recognition result correcting section 6, the learning character extracting section and the dictionary learning section 8 are constituted by a program. And
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.

<Learning Character Extraction Unit> The learning character extraction unit 7
Data in which characters that satisfy certain exclusion conditions among characters to be corrected by the recognition result correction unit 6 (hereinafter, correction target characters) are automatically excluded, and other characters are stored as learning characters requiring learning. Is output. In the embodiment, the exclusion conditions in the learning character extraction unit 7 are set as the following five conditions.
Characters other than those that meet any of the conditions are set as learning characters.

Exclusion condition: Character inserted by correction input Character to be corrected deleted by correction input Character to be corrected when the character to be corrected touches characters before and after it The size of the character to be corrected is larger than the threshold. The character to be corrected when the character is small The character to be corrected when the character to be corrected is blurred The exclusion conditions described above are inappropriate to use these characters as learning characters because the extraction may have failed. The exclusion condition can be determined based on information obtained from the recognition result correction unit 6, and the exclusion condition can be determined based on information obtained from the character extraction unit 2. . In addition, the above exclusion condition can be expected that the shape of the character is poor,
These characters are provided in order to exclude them as inappropriate learning characters. The exclusion condition can be determined based on information obtained from the character cutout unit 2, and the exclusion condition can be determined based on the correction target character. It can be determined from the number of black points (dots).

<Recognition Dictionary> The recognition dictionary is recorded on a removable recording medium such as a hard disk, a floppy disk, or an optical disk, and a read / write device for these recording media, that is, a magnetic disk device, a floppy disk device, and an optical disk. Any of the devices and the like is a component of the character recognition device.

FIG. 2 is an explanatory diagram showing the structure of the recognition dictionary 9. The recognition dictionary 9 has a template group 91 and a filter data group 92. The template includes a character code, a feature amount of the character corresponding to the character code, and a template number. The filter data includes an average value of a maximum value of a recognition distance for each character code, a root-mean-square value (variance value), and a The minimum value includes the root mean square value (variance value), and the number of characters used in creating these values. The filter data is obtained by performing the same recognition processing as that of the character recognition unit 4 on the character data used to create the recognition dictionary 9 using the recognition dictionary. Here, the recognition distance refers to a distance calculated as a vector of a feature amount of a recognition dictionary having the same character code as the feature amount of a character to be recognized.

In the dictionary learning method according to the present invention, the learned recognition dictionary is stored separately from the initial recognition dictionary (standard dictionary) according to the type of input character source. With this method, if the input characters have been learned before, the recognition dictionary at that time can be used for recognition.If the characters have not been learned, the standard dictionary is used, so that stable recognition conditions can always be obtained. it can. In addition, 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 or the like (because it has a personality such as a character habit), the writer itself Can be the source,
In the case of print, the title or font can be used as the source.

<Learning Dictionary Unit> [Configuration] In FIG. 1B, the learning dictionary unit 8 includes a feature extracting unit 1.
1, 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 extracting unit 11 to the filter data changing unit 17 can be configured by a hardware circuit, but in the present embodiment, is configured by a program. Some of the modules of the feature extracting unit 11 to the filter data changing unit 17 may be configured by a hardware circuit, and the other modules may be configured by a program. When the feature extracting 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. A dictionary learning process is realized.

[Dictionary Learning Operation] FIG. 3 is a flowchart showing a dictionary learning operation of the learning dictionary section 8. Feature extraction unit 1
In step S1, it is checked whether or not the data includes a feature amount of a character among a plurality of learning characters output from the learning character extraction unit 7 in FIG. 1A (S1). If so, the process proceeds to S3, and for the learning data that does not include the feature, the feature of the character is extracted (created) (S2).

Next, the recognition distance calculation unit 12 determines the minimum recognition distance and the template number at that time for each character code, and the maximum recognition distance and the template number at that time for each character code based on the features of the learning characters and the feature amounts of the templates in the recognition dictionary 9. The following recognition distance data is created (S3). FIG. 4 shows the structure of the recognition distance data.

The learning character filter unit 13 automatically excludes learning characters unsuitable for learning without requiring the operator to make a determination. Specifically, among the normalized distances of the minimum recognition distances of the respective character codes derived by Expression 2 (recognition distance normalization expressions) described below, the normalized distances of the respective character codes different from the learning characters and the threshold are set. If the learning character is smaller than the threshold value, the learning character is excluded. That is, such a learning character exists in the normal distribution of the minimum recognition distance of different character codes, and adding the learning character to the recognition dictionary may affect the recognition of the different character codes. Exclude the learning character as occurring. Similarly, the normalized distance of the maximum recognition distance of the character code of the learning character is obtained from Expression 2 and compared with a threshold. If the normalized distance is larger than the threshold, the learning character is excluded. That is, since such a learning character does not exist in the normal distribution of the maximum recognition distance of the character code, the learning character is excluded as having low reliability (S4). If the result of the filtering process based on the comparison between the normalized distance and the threshold value indicates that there is no learning character, the process ends (S5).

The above-described comparison processing with the learning character filter unit 13 using the character code different from the learning character is performed to check whether 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 feature vector of the character dictionary closest to the learning character in the template of the character code and the learning character. This indicates that the character is close to the character code (see FIG. 6). If the learning process is performed for a character that is too close to another character code, the recognition performance of the other character code that is too close will be adversely affected. The learning character is excluded because it is recognized as a code).

The comparison of the character codes of the learning characters is performed to check whether or not the statistical characters of the character codes include the learning characters. That is, since the maximum recognition distance in a certain character code approximates the radius of the feature space of the character code, the fact that the value of a certain character is far from the average value indicates that the character is not the target character. This means that it is far from the feature space of the code (see FIG. 7). Therefore, a character that is far from the characteristics of other characters of the same character code has low reliability of the characteristics of the character and is not suitable for learning, and the learning character is excluded from learning.

Based on the above two exclusion conditions, characters that are unsuitable for learning, that is, characters that affect the recognition of other character codes and characters that are far from the average feature of the same character code, can be determined without the operator's judgment. Can be automatically excluded. In addition, characters that are close to the characteristics of other character codes or characters that are significantly different from the average characteristic of the same character code may be caused by character blurring, crushing, inclination or deformation, and the image of the character is poor. Or the operator inputs a character with an incorrect character code when correcting the recognition result.

The learning character selecting section 14 automatically excludes valid learning characters without requiring the operator to make a decision. Specifically, the learning characters that have passed the filtering process in the learning character filter unit 13 are classified according to the template of the recognition dictionary to be learned (learning template), and the learning process is not yet performed and the correct recognition is performed. A learning character having the shortest distance is selected (S6).

Here, the learning target template refers to a template having the same character code as the learning character and a minimum recognition distance in the recognition distance data of the learning character, that is, a template closest to the same character code as the learning character. Further, the minimum recognition distance at that time is defined as a positive recognition distance.
If the minimum recognition distance in the same character code as the character to be recognized is the correct recognition distance, and the smallest one of the minimum recognition distances of the character codes different from the character to be recognized is the incorrect recognition distance, the correct recognition distance> the incorrect recognition distance In this case, the recognition result is erroneous recognition. Further, when the erroneous recognition distance ≧ the correct recognition distance ≧ the erroneous recognition distance−α (α: threshold for reject determination), the recognition result is rejected. Further, when the erroneous recognition distance−α> the correct recognition distance, the recognition result is the correct recognition.

The learning characters are divided into the learning target templates. The learning characters having the same character code input from the same source and having the same dictionary template of the correct recognition distance (learning target template) have the feature amount vector as the dictionary template. This is because they have the same directionality as shown in FIG.

Therefore, by learning from the learning characters that are close to the learning target template, the feature amount of the learning target template is gradually brought closer to those feature amounts, and the learning character is recognized. Can be minimized. Normally, the recognition dictionary has a balanced template in an initial state (a state in which learning is not performed). Therefore, by using such a learning method, learning can be performed without significantly breaking the balance. Further, since it is not necessary for the operator to make a judgment, it is not necessary to request the operator for character characteristics and specialized knowledge relating to learning.

The learning character synthesizing unit 15 learns the characteristic amount of the learning character selected by the learning character selecting unit 14 and the characteristic amount of the learning target template by using a characteristic amount synthesizing method described later (formula 3: see FIG. 5). The character synthesis ratio is gradually changed from the initial value and is synthesized as a vector amount to bring the feature amount of the learning target template closer to the learning character. Then, when any of the following three synthesis termination determination conditions is satisfied during the repetitive synthesis, the feature amount synthesis processing ends (S7).

At the end of the repetition of the feature amount synthesizing process, the template adding unit 16 checks whether or not the learning character is correctly recognized. If the learning character is correctly recognized, the process proceeds to S9. If the recognition is not performed correctly (described later), the feature amount of the learning character is added as a new template of the recognition dictionary (S8). In other words, the learning method using only the feature amount synthesis does not affect the recognition speed after learning because the number of templates does not increase. May decrease. In the learning method using only the template addition, the recognition rate does not decrease unless the template is too close to other character codes. However, the increase in the template causes the recognition speed to decrease. Therefore, in the present invention, the learning method based on the feature amount synthesis is prioritized (S7), and when the later-described synthesis end determination condition (Equation 5) or (Equation 6) is satisfied, the learning method is determined to be inappropriate. Switch to learning method by adding template. As a result, it is possible to select an optimal learning method for the learning character without requiring an operator's judgment.

In the filter data changing section 17, the S
7, when the learning character is correctly recognized by S8 (learning character synthesizing unit 15) or S9 (template adding unit 16), the filter data of the recognition dictionary 9 is corrected. More specifically, the learning character recognition process is performed using the recognition dictionary after learning, and the recognition distance data at this time is used as an average value of the minimum / maximum recognition distances used in the learning character filter unit 13 and its root mean square value. By adding to each of the number of characters, the average value of the minimum value / maximum value of the recognition distance, its squared average value, and the number of characters are corrected to values suitable for the correction target character (S10).

After the learning process (S7 to S9) for all the learning selected by the learning character selecting unit 14, the recognition distance calculating unit 12 creates recognition distance data again using the learning dictionary after learning. Thereafter, the learning process is not performed, and only the learning characters that are still erroneously recognized are correctly recognized until all stored learning characters are recognized correctly, or until there are no characters that pass through the learning character filter unit 13. S3 to S
10 (recognition distance calculation unit 12 to file data change unit 1)
The learning process of 7) is repeated (S11).

[Normalization of recognition distance, etc.] The normalized distance of the minimum / maximum recognition distance to be compared with the threshold value in the learning character filter unit 13 can be obtained as follows. Variance value var recognition distance dc0 in the character code co (dc0) obtains the average value ave of the recognition distance stored in the recognition dictionary 9 (dc0) and its mean square value ave than (dc0 ²⁾ by equation 1 below Can be var (dco) = ave (dco ² ) − (ave (dco)) ² (Equation 1) Learning obtained by the recognition distance calculation unit 12 using the variance value of the minimum recognition distance of each character code obtained by Expression 1 The minimum recognition distance of each character code of the character recognition distance data is normalized according to the following equation (2). d'co = (dco-ave (dco)) / (var (dco)) ^1/2 (Equation 2) where dco is the minimum recognition distance in the character code co of the learning character, and d'co is the character of the learning character. The normalized distance of the minimum recognition distance in the code co, ave (dco) is the average value of the minimum recognition distance in the character code co stored in the recognition dictionary 9, and var (dco) is the character code co calculated by the above equation 1. Is the minimum recognition distance variance.

Also, using the variance of the maximum recognition distance of each character code obtained by Expression 1 and Expression 2 above, the maximum recognition distance of each character code of the recognition distance data of the learning character obtained by the recognition distance calculation unit 12 Can be normalized. In this case, the maximum recognition distance in the character code co of the learning character is dco, the normalized 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 by setting the average value of ave (dco) and var (dco) to the variance of the maximum recognition distance of the character code co calculated by equation 1.

As described above, the range of the observed value in a certain normalized distribution can be obtained as the range of the observed value normalized by the equation 2 using the average value and the variance value of the sample.
For example, assuming that the reliability is 95%, from the normalized distribution table,
1.96 to 1.96 are given as the range (confidence interval) of the normalized observations. This makes it possible to determine whether or not the recognition distance exists within the normal distribution of the recognition distance of a certain character code.

[Feature Amount Synthesis Method and Synthesis Completion Judgment Condition] (1) Feature Amount Synthesis Method The feature amount of the learning character and the feature amount of the learning target template are synthesized by gradually changing the synthesis ratio of the learning character from the initial value. While performing the following equation (Equation 3) as a vector amount.
Note that FIG. 5 shows a transition state of a feature amount that transitions by learning by synthesis. Fdic ′ = (1−R) × Fdic + R × Funk (Equation 3) Here, Funk is the feature amount of the learning character, Fdic is the feature amount of the learning target template in the recognition dictionary, and Fdic ′ is the feature of the template after synthesis. The amount and R are the synthesis ratio (0 <R
<1).

(2) Conditions for Determining Combination Termination As described above, when any of the following three conditions is satisfied during repetitive composition according to the above equation 3, the processing in the learning character composition unit 15 is terminated. When the learning character is correctly recognized, and the difference between the correct recognition distance and the erroneous recognition distance is larger than the threshold, that is,
Is satisfied, | Fdic ^err −Funk | − | Fdic′−Funk |> α1 (Equation 4) where Fdic ^err is a feature vector of the template of the misrecognition distance, and α1 (> 0) is a threshold value. When the movement value before and after the synthesis of the learning target template is larger than the threshold, that is, when the following equation (Equation 5) is satisfied: | Fdic'−Fdic |> α2 (Equation 5) where α2 (> 0) is a threshold is there. If the distance between the feature vector of the learning target template after learning and the feature vector of the template of a character different from the character code of the learning character in the recognition dictionary is smaller than the threshold, that is, if the following equation (Equation 6) holds, | Fdicerr- Fdic '| <α3 (Equation 6) where α3 (> 0) is a threshold value.

A case where the condition for judging the completion of synthesis is satisfied,
That is, in some cases, the magnitudes of the correct recognition distance and the erroneous recognition distance are larger than the threshold value α1, but the reason is that even with characters of the same character code input from the same source, there is a certain degree of variation in the feature amount. This is because the learning is performed so as to absorb the variation by learning the two learning characters so that the correct recognition distance has a margin.

As described above, since the recognition dictionary is normally balanced in the initial state of the template as described above, if the existing template in the recognition dictionary moves greatly, the balance is largely lost. This is to prevent the possibility that the recognition rate of the character code may be reduced.

The synthesis termination determination condition is a condition provided to prevent the template to be learned from being too close to the template of another character code, thereby adversely affecting the recognition of the character code.

[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), the distribution range of the characteristic amount vector of the character code in FIG. 5 is a right circle (broken line) a, and the distribution range of the characteristic amount vector of the character code in FIG. The feature vector of the code is b (cross), the feature vector of the dictionary template is c (circle), the dictionary template closest to the learning character is B (double circle),
It is assumed that the feature vector of the character code is e (cross), the feature vector of the dictionary template is f (circle), and the dictionary template closest to the learning character is D (double circle). In this case, the minimum recognition distance between the learning character and its character code is the length of the line segment indicated by C, the minimum recognition distance of the learning character X and its character code is the length of the line segment indicated by E, and C <E Therefore, the minimum recognition distance of the learning character X is the length of the line segment C.

[Maximum Recognition Distance] FIG. 7 is an explanatory diagram of the maximum recognition distance. In FIG. 7, the learning character is A (triangle mark), and the distribution range of the feature vector of the character code in FIG. 5 is a circle (dashed line).
a, and the character vector of the character code is b (cross),
Assuming that the feature vector of the dictionary template is c (circle), the maximum recognition space (≒ feature space radius) is d, and the dictionary template that is most separated from the learning character A is B (double circle), the learning character A The maximum recognition distance of the character code is the length of the line segment BA.

[Direction of Learning Character] FIG. 7 is an explanatory diagram of the direction of the learning character. In FIG. 7, the distribution range of the characteristic amount vector of the character code in FIG. 5 is a circle on the right (broken line) a, the characteristic amount vector of the character code is b (cross), and the characteristic amount vector of the dictionary template is c (circle). In this case, the learning character group A having a certain template (black circle) as a template of the minimum recognition distance exists within a certain angle range around the template, and another template (double circle) is a template of the minimum recognition distance. The learning character group B exists in a certain angle range around the template (that is, has a directionality).

[0050]

As described above, according to the dictionary learning method and the character recognition device of the present invention, learning characters unsuitable for learning can be automatically excluded based on the recognition distance.
Further, since the learning characters can be classified according to the template of the recognition dictionary to be learned and selected based on the recognition distance, there is no need for an operator's judgment, and there is no need to require the operator to provide character characteristics or specialized knowledge related to learning. .

Further, since the necessity of adding a template is determined based on the minimum recognition distance of a learning character to update or add the recognition dictionary data, an optimum learning method for the learning character is selected without an operator's judgment. it can.

Further, since a learning character is obtained by automatically excluding a character corresponding to a predetermined exclusion condition from characters to be corrected, a character to be corrected that is inappropriate as a learning character, for example, a character adjacent to the correction target character is contacted. Thus, it is possible to prevent a character that has failed to be cut out from being a learning character, thereby preventing unnecessary learning processing.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration example of a character recognition device of the present invention.

FIG. 2 is a structural diagram showing a structure of a recognition dictionary.

FIG. 3 is a flowchart illustrating a dictionary learning operation of a learning dictionary unit.

FIG. 4 is an explanatory diagram showing the structure of recognition distance data.

FIG. 5 is an explanatory diagram illustrating a transition state of a feature amount that transitions 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 a learning character.

[Explanation of symbols]

7 learning character extraction unit (learning character extraction unit) 9 recognition dictionary 13 learning character filter unit (first filter unit, second filter unit)
14 learning character selecting unit (learning character selecting unit) 15 learning character synthesizing unit (feature amount synthesizing unit) 16 template adding unit (dictionary learning method determining unit) 91 template 92 filter data (recognition dictionary data) 100 character recognition apparatus

Claims (9)

[Claims] 1. A dictionary learning method for a character recognition device for performing a learning process of a recognition dictionary, comprising a step of automatically excluding a learning character unsuitable for learning of the recognition dictionary based on a recognition distance. 2. The dictionary learning method according to claim 1, wherein
Further, a step of classifying learning characters for each template of a recognition dictionary to be learned, and a step of selecting learning characters that have not yet undergone learning processing based on the recognition distance,
Combining the characteristic amount of the selected learning character and the characteristic amount of the classified template while gradually changing the combination ratio to bring the characteristic amount of the classified template closer to the learning character. Dictionary learning method. 3. The dictionary learning method according to claim 2,
The dictionary learning method further includes a step of determining whether or not to add a template based on the minimum recognition distance of the learning character and updating or adding the recognition dictionary data at the end of the combining process. 4. The dictionary learning method according to claim 1, further comprising a step of automatically excluding a character corresponding to a predetermined exclusion condition from characters to be corrected to obtain a learning character. A dictionary learning method characterized by the following. 5. A character recognition device for performing a learning process of a recognition dictionary, wherein, when a normalized distance between each character code different from the learning character and the learning character is smaller than a first threshold value, the learning character is trained in the recognition dictionary. A first filter means for excluding as a learning character unsuitable for learning a learning character which is inappropriate for learning a dictionary when a normalized distance between the same character code as the learning character and the learning character is larger than a second threshold value A character recognition device comprising: a second filter unit that excludes characters. 6. The character recognition device according to claim 5, wherein
Further, a learning character selecting means for classifying learning characters according to a template of a recognition dictionary to be learned and selecting learning characters which have not yet undergone learning processing based on the recognition distance, and a feature of the selected learning characters. A character amount synthesizing means for synthesizing while changing a synthesis ratio of the amount and the feature amount of the classified template to change the feature amount of the classified template closer to the learning character. . 7. The character recognition device according to claim 6, wherein
Further, at the end of the combining process, the method further comprises a dictionary learning method determining means for determining whether or not to add a template based on the minimum recognition distance of the learning character, and updating or adding the recognition dictionary data based on the determination result. Character recognition device. 8. A learning character extracting apparatus according to claim 5, wherein a learning character is obtained by automatically excluding a character corresponding to a predetermined exclusion condition from a correction target character. A character recognition device comprising means. 9. The character recognition device according to claim 8, wherein
A character recognition device, wherein a learning character extracting unit automatically excludes a correction target character when a correction target character satisfies one of the following exclusion conditions. The character inserted by the correction input, the correction target character deleted by the correction input, the correction target character when the correction target character is in contact with the character before and after it, and the correction target character when the size of the correction target character is smaller than the threshold value. The correction target character, or the correction target character when the correction target character is blurred.