JP3428512B2 - Character recognition device and computer-readable recording medium - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device having an automatic learning function, and more particularly to a character recognition device capable of automatically learning a personal character style that is difficult to recognize with a general-purpose dictionary.

[0002]

2. Description of the Related Art An optical character recognition device (OCR device) that determines the category having the shortest distance from the recognition target character by collating the characteristics of the recognition target character with standard features stored in advance in a dictionary for each category. ) Or an online character recognition device, it is important to improve the performance of the dictionary in order to improve the recognition accuracy and the recognition speed.

Methods for improving the performance of the dictionary include, for example, Japanese Patent Publication No. 6-32087 and Japanese Patent Laid-Open No. 10-21340.
As can be seen in Japanese Patent Publication, a method for changing general standard features for each character category in a general dictionary by learning,
For example, Japanese Examined Patent Publication No. 60-30993, Japanese Examined Patent Publication No. 2-53
As disclosed in Japanese Patent No. 833, it is roughly classified into a method of adding a new standard feature to a dictionary by learning without changing the general standard feature for each character category.
However, in the former method, since general-purpose standard features are lost as learning progresses, recognition accuracy for a specific individual increases, but recognition accuracy for other individuals decreases. Therefore, in the case of improving the recognition accuracy for a specific individual while preventing the recognition accuracy for a third person from being lowered, the latter method is generally adopted. The present invention relates to improvements in this latter scheme.

A configuration example of a character recognition device (hereinafter referred to as a first conventional technique) described in Japanese Patent Publication No. 53833/1990 is shown in FIG. 16, and the characters described in Japanese Patent Publication No. 60-30993 are disclosed. FIG. 17 shows a configuration example of a recognition device (hereinafter, referred to as a second related art).

A conventional character recognition apparatus having an automatic learning function shown in FIG. 16 has a data input unit 10, a feature extraction unit 20, a general dictionary 30, a personal dictionary 35, a collation unit 45, and a result correction unit. 50, a display unit 51, a correction information input unit 52, and a new feature registration unit 65.

The conventional character recognition device having an automatic learning function having such a structure operates as follows. That is, in FIG. 16, in the case of the OCR device, the characters written on the form or the like are input as a character image image by the data input means 10 such as a scanner, and the feature extraction means 20 is used.
Extracts the recognition feature from the character image. The general-purpose dictionary 30 stores general-purpose standard features for each character category of recognition features in advance. The personal dictionary 35 stores personal standard features for each character category of recognition features (initial state is empty). The collating unit 45 uses the general dictionary 3 to identify the recognition feature of the input character obtained by the feature extracting unit 20.
0 and the standard features stored in the personal dictionary 35 are sequentially compared to obtain candidate character categories of standard features having a small distance value. The result correction means 50 displays the first candidate on the display means 51, and if the result is erroneously recognized, the user specifies the correct category by the correction information input means 52 such as a keyboard. The new feature registration means 65, when erroneously recognized, the correct category c obtained by the result correction means 50 and the input character recognition feature fi obtained by the feature extraction means 20.
Using (i represents the element of the feature), the template {learning feature number n (c), standard feature Si (c)} of the category c already stored in the personal dictionary 35 is updated by the following equation (1). (If it does not exist, register it newly).

[0007] Si (c) ← {n (c) ・ Si (c) + fi} / {n (c) +1} n (c) ← n (c) +1 … (1)

The conventional character recognition apparatus having an automatic learning function shown in FIG. 17 has a data input means 10, a feature extraction means 20, a dictionary 36, a collation means 46, a result correction means 50, and a display. Means 51 and correction information input means 5
2, a new feature registration means 66, and a temporary registration buffer 67.

The conventional character recognition device having an automatic learning function having such a structure operates as follows. That is, in FIG. 17, in the case of the online character recognition device,
The handwritten character is input as a stroke coordinate point sequence by the data input means 10 such as a tablet, and the feature extraction means 20 extracts a recognition feature from the stroke coordinate point sequence. The dictionary 36 stores standard features of recognition features for each character category (only general-purpose standard features in the initial state). The collation unit 46 sequentially collates the recognition feature of the input character obtained by the feature extraction unit 20 with each standard feature stored in the dictionary 36, and obtains a candidate character category of the standard feature having a small distance value. The new feature registration means 66, when erroneously recognized, uses the correct category c obtained by the result correction means 50 and the recognition feature fi of the input character obtained by the feature extraction means 20 to register the recognition feature in the temporary registration buffer 67. . Here, if the standard feature of the temporary template already stored in the temporary registration buffer 67 is close, the standard feature is updated, and if it is not close, the standard feature is newly registered. Further, when the number of learning features (the number of registrations) of the temporary template reaches a certain number, the temporary template is registered in the dictionary 36.

[0010]

As described above, according to the method of adding new standard features to the dictionary by learning without changing the general standard features for each character category, the learning progresses. However, since general-purpose standard features are not lost, it is possible to prevent deterioration of recognition accuracy for third parties, while recognition accuracy for specific individuals can be improved because new standard features by learning are added to the dictionary. it can. However, there is a problem that the dictionary capacity becomes huge.
That is, since there is no limit to the addition of new standard features in both the first and second conventional techniques, the number of categories stored in the personal dictionary 35 increases in the first conventional technique, and the second standard feature increases. In the related art, the number of categories stored in the personal dictionary portion of the dictionary 36 increases, and the capacity of the entire dictionary becomes huge.

Another problem of the prior art is that if the dictionary capacity is limited to prevent the dictionary capacity from becoming huge, new standard features cannot be added and the learning function stops.

The present invention is intended to solve such a conventional problem, and an object thereof is to learn a new standard feature by dictionary without changing the general standard feature for each character category. In order to prevent the dictionary capacity from becoming huge without stopping the learning function in the character recognition device of the method to be added to.

[0013]

[0014]

A first character recognition apparatus of the present invention has a data input means (10 in FIG. 1) for acquiring input character data, and a recognition feature from character data obtained by the data input means. Feature extraction means (2 in FIG. 1) for extracting
0), a general-purpose dictionary (30 in FIG. 1) in which general-purpose standard features for each character category of recognition features are stored in advance, and a predetermined number of personal standard features (maximum number of stored templates) are sequentially learned by learning. A personal dictionary that will be remembered (31 in Fig. 1)
And collating means for sequentially collating the recognition features obtained by the feature extracting means with the standard features stored in the general-purpose dictionary and the personal dictionary, and outputting the character category and template number of the standard features having a small distance value. (4 in FIG. 1
0), the recognition result is displayed by the display unit, the correct category is specified by the correction information input unit (52 in FIG. 1), and the recognition result is corrected and confirmed by the result correction unit (50 in FIG. 1). In the case of recognition, when registering the correct answer category obtained by the result correcting means and the recognition feature obtained by the feature extracting means in the personal dictionary, if the same template as the correct answer category already exists, between the standard features If the distance value is compared with a predetermined threshold value, the standard feature is updated by performing a weighted sum if it is smaller, a new template is registered otherwise, and when the number of templates exceeds the maximum number of stored templates, Using the new feature registration means (60 in FIG. 1) for deleting the oldest template and the correct category and recognition candidate obtained by the result correction means, the standard feature of the personal dictionary is used. When is effectively used for recognition, the order of the template containing the standard feature in the personal dictionary is moved to the newer direction. Conversely, when it adversely affects the recognition, the template individual containing the standard feature is A feature order changing unit (70 in FIG. 1) for moving the order in the dictionary to the older one is provided.

Here, the characteristic order changing means may change the physical order of the template in the personal dictionary, or may change the order number indicating the order of the template in the personal dictionary. . Further, by making the predetermined threshold value used by the new feature registration means for comparison dependent on the number of learning features, integration of features is controlled according to the number of features used for creating standard features of the personal dictionary, and a small number of features are controlled. A stable template may be created from the data.

Further, the second character recognition device of the present invention stores the number of times of effective use of the template in the personal dictionary, and makes it difficult to delete the template having a large number of times of effective use. More specifically, when the personal dictionary (32 in FIG. 10) that stores the number of times of effective use in addition to the standard feature and the number of times of effective use of the template is larger than a predetermined threshold value when the oldest template is deleted. Is a new feature registration means (61 in FIG. 10) for moving the order of the template in the personal dictionary to the newer one and subtracting a predetermined number of effective use times, and when the standard feature of the personal dictionary is effectively used. In addition, when the order of the template containing the standard feature is moved to the new one, and the number of times of effective use is added at the same time, and when the recognition is adversely affected, the personal dictionary of the template containing the standard feature is added. Feature order changing means (FIG. 1) for moving the order in the order of
0) and 71).

Further, the third character recognition apparatus of the present invention uses the standard features of the general-purpose dictionary when registering the recognition features in the personal dictionary. More specifically, when registering the recognition feature in the personal dictionary, the standard dictionary of the same category in the general-purpose dictionary is
It has a new feature registration means (62 in FIG. 13) for performing weighted sum registration at a ratio according to the number of learning features.

Further, the fourth character recognition device of the present invention adds a penalty to the template of the personal dictionary at the time of matching. More specifically, when obtaining the distance value between the recognition feature and each standard feature of the dictionary, a collating unit (41 in FIG. 14) that adds a penalty according to the number of learning features for the standard feature of the personal dictionary is used. Have.

In the present invention, a general dictionary and a personal dictionary are used.
Although different types of dictionaries are used, the dictionaries do not necessarily have to be physically separated, and it is possible to use one dictionary having both a general-purpose dictionary part and a personal dictionary part.

[0020]

BEST MODE FOR CARRYING OUT THE INVENTION Next, embodiments of the present invention will be described in detail with reference to the drawings.

Referring to FIG. 1, the first embodiment of the present invention comprises a data input means 10, a feature extraction means 20,
General-purpose dictionary 30, personal dictionary 31, collation means 40, result correction means 50, display means 51, correction information input means 52, new feature registration means 60, feature order change means 7
It is composed of 0 and 0.

The general-purpose dictionary 30 pre-stores general-purpose standard features for each character category of recognition features. The standard feature is created by averaging the recognition features extracted from each data for each character category using a character database filled in by a large number of writers. The stored template information includes character category codes, template numbers, and standard features.

Although nothing is stored in the personal dictionary 31 in the initial state, personal standard features are sequentially stored up to a predetermined number (the maximum number of stored templates) by learning.
The stored template information includes character category codes, template numbers, learning feature numbers, and standard features. Here, the personal dictionary 31 is a dictionary for a specific user. When the user uses the character recognition device,
A personal dictionary dedicated to oneself is set as a personal dictionary 31 in a character recognition device for use. Since the general-purpose dictionary 30 is common to all users, each user only needs to have a personal dictionary individually.

The data input means 10 uses a data input device such as a scanner or a tablet to obtain character data for one character. The character data is stored as a character image image when input by a scanner and as a coordinate point sequence of strokes when input by a tablet.

The feature extraction means 20 is the data input means 10.
The recognition feature is extracted from the obtained character data. Various feature extraction methods for character recognition have been proposed in the past, and can be easily realized by conventional techniques. For example, a character image image is divided into meshes, and a grayscale feature characterized by the frequency of the number of black pixels in each region, or a character image image is divided into direction components, and the frequency of direction components in each region for each direction. Famous for directional features, etc. For example, as a document describing such a method, "Improvement of Directional Pattern Matching Method and Application to Handwritten Kanji Recognition" (1
June 990, IEICE Technical Report, PRU9
0-20). Further, even when the character data is a stroke coordinate point sequence, it can be handled in the same manner by converting it into a character image image.

The collating means 40 sequentially collates the recognition characteristics of the input data obtained by the characteristic extracting means 20 with the standard characteristics stored in the general dictionary 30 and the personal dictionary 31,
A candidate for a character category of a standard feature having a small distance value is obtained. As the distance measure, for example, Euclidean distance or city block distance can be used. Here, each recognition candidate includes the template number of the matched standard feature in addition to the code of the character category.

The result correcting means 50 displays the first candidate on the display means 51, and when the first candidate is erroneously recognized, the user specifies the correct category by the correction information input means 52 such as a keyboard.

When the incorrect recognition is performed, the new feature registration means 60 recognizes the correct category c obtained by the result correction means 50 and the recognition feature fi of the input data obtained by the feature extraction means 20.
Is registered in the personal dictionary 31. Here, if the template of category c already exists in the personal dictionary 31 and the distance value between the standard feature of the template and the recognition feature of the input data is smaller than a predetermined threshold value, the standard feature is updated by weighted summation. Otherwise, a new template is registered.
The number of templates that can be registered (the maximum number of stored templates) is set in advance in the personal dictionary 31, and if the number of stored templates exceeds the maximum number, the older template is deleted. When the standard feature is updated by the weighted sum, the standard feature and the distance value of the other template of the category c are calculated, and when the standard feature is smaller than a predetermined threshold value, the standard feature is further integrated by the weighted sum.

The feature order changing means 70 is the result correcting means 5
If the standard feature of the personal dictionary 31 is effectively used for recognition using the correct category c obtained from 0 and the recognition candidate,
When the order of the template containing the standard feature is moved to a new one in the personal dictionary 31 and adversely affects the recognition, the order of the template containing the standard feature in the personal dictionary 31 is changed to the old order. To move.

Next, the overall operation of this embodiment will be described in detail with reference to the flow charts of FIGS.

First, the character data obtained from the data input means 10 is supplied to the feature extracting means 20, and the feature extracting means 20 extracts the recognition feature from the character data (step S1 in FIG. 2). Next, the collation unit 40 sequentially collates the recognition feature fi of the input data obtained by the feature extraction unit 20 with each standard feature Si (b) stored in the general dictionary 30 and the personal dictionary 31, and the distance value D (B) Small standard feature character category candidates are obtained (step S2).

Further, the result correction means 50 is the display means 5
1 to display the first recognition candidate (step S
3) In the case of incorrect recognition, the user specifies the correct answer category by the correction information input means 52 such as a keyboard and confirms the recognition result (step S4).

Next, the feature order changing means 70 uses the correct category c and the recognition candidate obtained from the result correcting means 50. When the standard feature of the personal dictionary 31 is used for recognition, the personal feature dictionary of the standard feature is used. The order within 31 is moved (step S5). Step S5, referring to FIG.
The details will be described. First, regarding the first recognition candidate,
Whether the candidate was obtained from the general-purpose dictionary 30 from the feature number (template number) of the matched standard feature, or the personal dictionary 31.
Is determined from step S50 (step S50 in FIG. 3).
0). When it is obtained from the personal dictionary 31, the category of the first candidate is compared with the correct answer category to determine whether or not the correct answer is obtained (step S501). If the answer is correct, the order of the template including the standard feature in the personal dictionary 31 is moved to the newer direction by Pf (step S502). On the contrary, if the template is mistakenly recognized, the order of the template in the personal dictionary 31 is changed. Is moved to the old side by Pb (step S50
3) (If you move out of range,
Move to the newest and oldest position respectively). Where P
f and Pb are integers of 1 or more, which are predetermined values.

Referring again to FIG. 2, finally, the new feature registration means 60 checks whether or not it has been erroneously recognized (step S6 in FIG. 2), and if erroneously recognized, the correct answer category obtained from the result correction means 50. Using c and the recognition feature fi of the input character obtained from the feature extraction means 20, the recognition feature is registered in the personal dictionary 31 (step S7). Step S7
Will be described in detail with reference to FIGS. 4 and 5.

First, in the personal dictionary, a template of category c that has already been stored is selected (step S700 in FIG. 4). The number K of templates in category c is 0
It is checked whether or not (step S701), and if K = 0 (if it does not exist), the recognition feature fi is registered as a new template (proceeding to step S710).

If a template of category c already exists, the template is sorted in descending order of learning feature number (step S702). The template number sequence at this time is b (k) (k = 1, ..., K). The standard features Si (b) of the template are sorted in descending order of the number of learning features.
The distance value D (b) between (k)) and the recognition feature fi of the input character is calculated and compared with the threshold value T (n (b (k)) + 1) (steps S703 and S704).

Here, the threshold value T (n) is set depending on the learning feature number n when the features are weighted and summed.
This threshold value T (n) is set in advance by a preliminary experiment or the like. Generally, when the learning feature number n is small, it becomes unstable unless only very similar features are averaged.
(N) is set smaller as n is smaller and larger as n is larger. The distance value D (b) is the threshold value T (n (b
(K)) + 1), the standard features Si (b (k))
And the recognition feature fi are weighted and summed according to the following equation (2) to update the standard feature (step S707), and each template newer than b (k) is shifted one by one to the older side to obtain b (k). Move to the latest position (step S70)
8).

[0038] Si (b (k)) ← {n (b (k)) · Si (b (k)) + fi} / {n (b (k)) + 1} n (b (k)) ← n (b (k)) + 1 … (2)

The distance value D (b) is the threshold value T (n (b
If it is greater than (k)) + 1), the next (k + 1) template is examined (steps S705, S706 to S704). If there is no template to be weighted (k> K), the recognition feature fi is registered as a new template (proceeding to step S710).

If the stored template number m is equal to the maximum stored template number M (step S71)
0), the oldest (first) template is deleted, and each template is moved one by one to the older one (step S71).
1) The recognition feature fi is registered as a new template at the latest (m-th) position (step S713). m <M
In the case of, m is incremented by 1 (step S712) and registered at the m-th position (step S713).

Next, in step S707, the recognition feature fi
Is registered as a weighted sum, the template of the category c is integrated. Specifically, assuming that the weighted template number is b0 (step S750 in FIG. 5), b0
If there is a template of category c excluding (steps S751 and S752) (K is decremented by 1), the template is sorted in descending order of learning feature number (step S753). The template number column at this time is b
(K) (k = 1, ..., K). The distance value D (b) between the standard feature Si (b (k)) and the standard feature Si (b0) of the template is calculated in descending order of the learning feature number, and the threshold value T
It is compared with (n (b (k)) + n (b0)) (step S7).
54, S755).

The distance value D (b) is the threshold value T (n (b
(K)) + n (b0)), the standard features Si (b
(K)) and the recognition feature Si (b0) are weighted. At this time, the template in the new position is updated. That is, b
(K) is compared with the position of b0 (step S760), and b
If 0 is at a new position, the template of b0 is updated by the following equation (3) (step S761), and b is further updated.
Delete the template of (k) (m ← m-1), b
Each template newer than (k) is moved to the older one (step S762).

[0043] Si (b0) ← {n (b (k)) ・ Si (b (k)) + n (b (k)) ・ Si (b0)} / {n ( b (k)) + n (b0)} n (b0) ← n (b (k)) + n (b0) … (3)

If b (k) is located at a newer position in step S760, the template of b (k) is updated by the following equation (4) (step S763), and the template of b0 is deleted (m ←). m-1), move each template newer than b0 one by one to the old one (step S
764), and finally b (k) is newly set to b0 (step S765).

[0045] Si (b (k)) ← {n (b (k)) ・ Si (b (k)) + n (b (k)) ・ Si (b0)} / { n (b (k)) + n (b0) n (b (k)) ← n (b (k)) + n (b0) … (4)

If updated, the process returns to step S751 to re-integrate. In step S755, the distance value D (b) is the threshold value T (n (b (k)) + n (b0)).
If it is larger, the next (k + 1) template is checked (steps S756, S757 to S755). If there is no template for weighted sum, (k
> K), the process ends (step S757).

Next, the effect of this embodiment will be described.

In the present embodiment, since new standard features are added to the personal dictionary by learning without changing the general standard features of the general dictionary, it is possible to prevent deterioration of recognition accuracy for a third party. At the same time, the recognition accuracy for a specific individual can be improved, and since the maximum number of stored templates is set in the personal dictionary, it is possible to prevent the dictionary capacity from becoming huge.

Further, in the present embodiment, the learning effect can be obtained in a short period of time because the learning is all performed when the recognition error is made. Further, since the template effective for recognition is left and the template having an adverse effect is deleted one after another, the adverse effect on the recognition can be minimized. In addition, when the features are weighted and registered for registration, the threshold value is changed according to the number of learning features to create a stable template, so that appropriate learning without adverse effects can be realized.

Next, the operation of the present embodiment will be described using a concrete example.

It is assumed that 100 or less templates are stored in the general-purpose dictionary 30 and 5 templates as shown in FIG. 6 are stored in the personal dictionary 31 (maximum stored template number M = 5, standard feature). Is not shown) (In the case of kanji recognition, the number of templates in the general dictionary is actually 4
About 000, and the number of templates in the personal dictionary is about 200). Template numbers less than 100 are template numbers in the general dictionary 30, and numbers 100 or more are in the personal dictionary 31.
Represents the template number of. Moreover, each parameter is set as follows. Pf = 2, Pb = 2, T (n)
= 100-n.

(Example 1) When the character "A" is input from the data input unit 10, the feature extraction unit 20 extracts the recognition feature (step S1), and the matching unit 40 recognizes the recognition feature and the general dictionary 30 and the individual. It is assumed that a distance candidate with respect to each standard feature stored in the dictionary 31 is calculated, and a candidate having a small distance value is selected to obtain a recognition candidate as shown in FIG. 7A (step S2). In this case, since the recognition result is the correct answer, the result correcting means 50 determines without correction (steps S3 and S4). Feature order changing means 7
0 indicates that the template of the personal dictionary is used because the template number of the first candidate is 100 or more (step S500), and the correct answer category “A” determined by the result correcting means 50 is the first category. Since the candidate categories “A” are the same, it is determined that the first candidate is the correct answer (step S501). Since it is a correct answer, the template with the template number “150” (stored second in FIG. 6) in the personal dictionary corresponding to the first candidate template number “150” is ordered as shown in FIG. The second one (Pf) is moved to the newer one and the fourth one is stored. At this time, "day" and "a" (template number "153") stored in the third and fourth positions in FIG. 6 are stored in the second and third positions, respectively (step S502). The new feature registration means 60 detects that the recognition result is correct, and ends (step S6).

(Example 2) It is assumed that when the character "O" is input from the data input means 10, the matching means 40 obtains a recognition candidate as shown in FIG. 7A (step S2). In this case, the recognition result is an erroneous recognition, so the result correction means 5
At 0, the recognition result is corrected by selecting the second candidate "o" (steps S3 and S4). The characteristic order changing unit 70 detects that the template of the personal dictionary is used because the template number of the first candidate is 100 or more (step S500), and further the result correcting unit 5
Since the correct answer category “O” determined in 0 and the category “A” of the first candidate are different, it is determined that the first candidate is erroneous recognition (step S501). Because it is a misrecognition,
Since the template number of the first candidate is “150”, the template with the template number “150” in the personal dictionary (stored second in FIG. 6) is shown in FIG.
As shown in (b), move to the older one. At this time, P
Since b = 2, two are originally moved, but only the first one can be moved, so the first one is stored. At this time, 1
The “i” in FIG. 6 stored second is shifted to the second and stored (step S503).

Next, the new feature registration means 60 detects that the recognition result is erroneous recognition (step S6), and registers the recognition feature obtained by the feature extraction means 20 in the personal dictionary 31 (step S7). At this time, the template of "O" in the personal dictionary is searched (step S700),
It is detected that the "O" template does not exist (step S701), and the process proceeds to step S710 to register as a new template. It is detected that the current number of stored templates is equal to the maximum number of stored templates M = 5 (step S710), and the oldest (first) template (in FIG. 8B, template number 150 “A”) is deleted. The second to fifth templates are shown in FIG. 8 (c).
As described above, they are stored by being shifted from the first to the fourth (step S711), and the recognition feature is registered as the category "O" at the latest position (fifth) (step S71).
3).

(Example 3) It is assumed that when the character "A" is input from the data input unit 10, the collation unit 40 obtains a recognition candidate as shown in FIG. 7B (step S2). In this case, the recognition result is an erroneous recognition, so the result correction means 5
At 0, the recognition result is corrected by selecting the second candidate "A" (steps S3 and S4). The feature order changing unit 70 detects that the template of the general-purpose dictionary is used because the template number of the first candidate is less than 100 (step S500), and does nothing. The new feature registration means 60 detects that the recognition result is erroneous recognition (step S6), and registers the recognition feature obtained by the feature extraction means 20 in the personal dictionary 31 (step S6).
7). At this time, the template of "A" in the personal dictionary shown in FIG. 6 is searched (step S700), and two templates of "A" (template numbers "150" and "1") are searched.
53 ”) is detected (step S701),
The two template numbers are sorted in descending order of learning feature number (step S702). n (150) = 2, n
Since (153) = 1, b (1) = 150, b
(2) = 153.

Here, the recognition feature fi and the standard feature Si (b
The distance value D (b) of (k)) is sequentially calculated, and the threshold value T
When compared with (n) (steps S703 to S706),
For example, when D (150) = 400, D (150)> T (n (150)
When +1) = T (3) = 300 and D (153) = 350, D
Since (153)> T (n (153) +1) = T (2) = 200 and there is no template to be weighted and summed, the process proceeds to step S710 to register as a new template. Since the current number of stored templates is equal to the maximum number of stored templates M = 5 (step S710), the first template (“I” in FIG. 6) is deleted, and the second to fifth templates are changed to those in FIG. 9A. Each one like
Shift from the 4th to 4th and store (step S71
1) The fifth recognition feature is registered as a category "A" (template number "155") (step S71).
3).

If, in step S704, D (15
When 0) = 250, D (150) <T (n (150) +1), and the template of “150” is updated by the weighted sum (step S707), and the third to fifth templates in FIG. As shown in FIG. 9B, the second to the fourth are shifted and stored, and the template of "150" is moved to the fifth (step S708). Next, FIG.
In the state of (b), the template of "A" in the personal dictionary is searched for except the template of "150" (step S751), and one template of "A" (template number "153") exists. It is detected (step S752) and b (1) = 153 (step S75).
3). Here, the standard feature Si (150) and the standard feature Si (b
(K)) distance value D (b) is calculated, and for example, D (153) = 3
When 50, D (153) <T (n (153) +1) = T (4) =
It becomes 400, and the template of “150” is at the latest position rather than “153” (step S760),
The template of “150” is updated by the weighted sum (step S761), the third template of “153” is deleted, and the fourth and fifth templates are changed to those of FIG.
As shown in (c), they are shifted to the third and fourth positions and stored (step S762).

Next, a second embodiment of the present invention will be described in detail with reference to the drawings.

Referring to FIG. 10, in the second embodiment of the present invention, the data input means 10 and the feature extraction means 20 are provided.
, General dictionary 30, personal dictionary 32, and collating means 40
The result correction unit 50, the display unit 51, the correction information input unit 52, the new feature registration unit 61, and the feature order change unit 71.

Data input means 10 and feature extraction means 20
, General dictionary 30, collating means 40, result correcting means 5
0, the display means 51, and the correction information input means 52 are shown in FIG.
It is the same as that of the character recognition device of the first embodiment shown in FIG.

The point different from the character recognition apparatus of the first embodiment is that the personal dictionary 32 also stores the number of times of effective use.

In addition to the operation of the character recognition device of the first embodiment, the new feature registration means 61 has a number of effective use times of the template equal to or greater than a predetermined threshold value (R) when deleting the oldest template. In the case of, the template is moved to the newer one (Pf) and the number of effective uses is subtracted by a predetermined number (Q).

The characteristic order changing means 71 is the result correcting means 5
When the standard feature of the personal dictionary 32 is effectively used by using the correct category c obtained from 0 and the recognition candidate, the order of the template including the standard feature in the personal dictionary 32 is moved to the new one at the same time. When the effective use count is incremented by 1, and when the recognition is adversely affected, the order of the template including the standard feature in the personal dictionary 32 is moved to the older one, and at the same time the effective use count is decremented by 1.

Next, the effect of this embodiment will be described.

In this embodiment, since the number of times of effective use is stored, it is difficult to delete a template that is frequently used, and the learning effect can be maintained.

Next, the operation of this embodiment will be described using a concrete example.

In the personal dictionary 32, 5 as shown in FIG.
Assume that individual templates are stored (maximum stored template number M = 5), and each parameter is set as follows. R = 10, Q = 5.

It is assumed that when the character "A" is input from the data input unit 10, the collation unit 40 obtains a recognition candidate as shown in FIG. 7B (step S2). In this case, the recognition result is an erroneous recognition, so the result correcting means 50 uses the second
The recognition result is corrected by selecting the candidate "A" (steps S3 and S4). The feature order changing means 71
Since the template number of the first candidate is less than 100, it is detected that the template of the general-purpose dictionary 30 is used (step S500), and nothing is executed. The new feature registration means 61 detects that the recognition result is erroneous recognition (step S6), and registers the recognition feature obtained by the feature extraction means 20 in the personal dictionary 32 (step S7). At this time, the template of "A" in the personal dictionary 32 is searched (step S700), and it is detected that there are two templates of "A" (template numbers "150" and "153") (step S701). The two template numbers are sorted in descending order of learning feature number (step S702). Since n (150) = 2 and n (153) = 1, b (1) = 150 and b (2) = 153. Here, the distance value D (b) of the recognition feature fi and the standard feature Si (b (k)) is sequentially calculated and compared with the threshold value T (n) (steps S703 to S706), for example, D (150).
= 400, D (150)> T (n (150) +1) = T
When (3) = 300 and D (153) = 350, D (153)> T
Since (n (153) +1) = T (2) = 200 and there is no template to be weighted and summed, the process proceeds to step S710 to register as a new template. Since the current number of stored templates is equal to the maximum number of stored templates M = 5 (step S710), the oldest template is deleted (step S711). As shown in FIG.
Move "i" by 2 (Pf) to the newer one, and subtract 5 (Q) of the number of effective uses. Here, the second and third templates are shifted to the first and second, respectively, and stored. Since the number of times of effective use of the first "A" is less than or equal to the threshold value R, "A" is deleted, and the second to fifth templates are respectively shifted from the first to the fourth as shown in FIG. 12 (b). It is stored and the recognition feature is registered in the fifth category as a category "A" (step S713).

Next, a third embodiment of the present invention will be described in detail with reference to the drawings.

Referring to FIG. 13, in the third embodiment of the present invention, the data input means 10 and the feature extraction means 20 are provided.
, General-purpose dictionary 30, personal dictionary 31, and collating means 40
The result correction means 50, the display means 51, the correction information input means 52, the new feature registration means 62, and the feature order change means 70.

Data input means 10 and feature extraction means 20
, General-purpose dictionary 30, personal dictionary 31, and collating means 40
The result correcting means 50, the display means 51, the correction information inputting means 52, and the feature order changing means 70 are the same as those in the character recognition device of the first embodiment shown in FIG.

In the character recognition device of the first embodiment, the new feature registration means 62 weights the recognition feature of the category c with the standard dictionary of the category c of the general dictionary 30 when registering the recognition feature of the category c in the personal dictionary 31. There is a point to register. At that time, the weight of the weighted sum is controlled according to the number of learning features.

In step S713 of FIG. 4, when the recognition feature fi is newly registered, the synthetic feature Si is generated and registered by the following equation (5) using the standard dictionary gi of the category c of the general-purpose dictionary 30.

Si ← α (1) · fi + {1-α (1)} · gi (5) where α (n) is a parameter (0 <α (n) ≦ 1 depending on the learning feature number n. ). Generally, the number of learning features n
When is small, in order to increase the contribution of the general-purpose dictionary 30,
(N) is reduced, and α (n) increases as n increases.

In step S707 of FIG. 4, when the standard feature Si (b (k)) is updated by weighted summing the recognition features fi, it can be obtained by the following equation (6). Si (b (k)) ← α (n + 1) · f′i + {1-α (n + 1)} · gi where f′i = {n (b (k)) · f ″ i + fi} / {n (B (k)) + 1} f ″ i = {Si (b (k))-{1-α (n) · gi}} / α (n) n (b (k)) ← n (b ( k)) + 1 (6)

Next, the effect of this embodiment will be described.

In the present embodiment, since the recognition feature is configured to be weighted and registered with the standard dictionary of the general-purpose dictionary, a stable dictionary is created and recognition can be performed with high accuracy.

Next, a fourth embodiment of the present invention will be described in detail with reference to the drawings.

Referring to FIG. 14, in the fourth embodiment of the present invention, the data input means 10 and the feature extraction means 20 are used.
, General-purpose dictionary 30, personal dictionary 31, and collating means 41
The result correction means 50, the display means 51, the correction information input means 52, the new feature registration means 60, and the feature order change means 70.

Data input means 10 and feature extraction means 20
, General dictionary 30, personal dictionary 31, and result correcting means 5
0, the display unit 51, the correction information input unit 52, the new feature registration unit 60, and the feature order change unit 70 are the same as those in the character recognition device of the first embodiment shown in FIG. .

The collating means 41 sequentially collates the recognition characteristics of the input character obtained from the characteristic extracting means 20 with the standard characteristics stored in the general dictionary 30 and the personal dictionary 31, and determines the standard characteristics having the small distance value D. Find candidates for character categories. Here, when collating with the standard feature Si (b) of the personal dictionary 31, the distance value D (b) is calculated according to the learning feature number n (b).
A penalty is added according to the following equation (7). D (b) ← D (b) + E (n (b)) ... (7)

E (n) is a value that reflects the stability of the standard feature and is preset. Generally, the number of learning features n
When is small, the standard feature is unstable, so E (n) is set to be large, and as n becomes large, E (n) is set to be small.

Next, the effect of this embodiment will be described.

In the present embodiment, in the matching process using the personal dictionary, a penalty is added according to the number of learning features, so that the adverse effects of unstable standard features can be suppressed and highly accurate recognition can be performed. .

Next, a fifth embodiment of the present invention will be described in detail with reference to the drawings.

Referring to FIG. 15, in the fifth embodiment of the present invention, the same data input means 10, display means 51 and correction information input means 52 as those of the first to fourth embodiments are provided.
Computer 80 connected to the computer. The computer 80 includes a processing device 81 such as an MPU and a memory 82 connected thereto. The recording medium 83 is a CD
A machine-readable recording medium such as a ROM, a magnetic disk, and a semiconductor memory, in which a character recognition control program is recorded. The character recognition control program recorded in the recording medium 83 is read by the processing device 81 when the computer 80 is started up, and the operation of the processing device 81 is controlled to control the character recognition control program in the first embodiment. Is a feature extracting unit 20, a collating unit 40, a result correcting unit 50, a new feature registering unit 60, and a feature order changing unit 7.
0 is generated on the processing device 81, and the general dictionary 30 and the personal dictionary 31 are loaded on the memory 82. Also,
In the above-described second embodiment, the feature extraction means 2
0, collation means 40, result correction means 50, new feature registration means 61 and feature order change means 71 are generated on the processing device 81, and the general dictionary 30 and the personal dictionary 32 are loaded on the memory 82. In addition, in the above-described third embodiment, the feature extracting unit 20, the collating unit 40, the result correcting unit 50, the new feature registering unit 62, and the feature order changing unit 70 are generated on the processing device 81, and The general dictionary 30 and the personal dictionary 31 are loaded on the memory 82. In addition, in the above-described fourth embodiment, the feature extraction unit 20, the collation unit 41, the result correction unit 50, the new feature registration unit 60, and the feature order change unit 70 are generated on the processing device 81, and General-purpose dictionary 3 on the memory 82
0 and personal dictionary 31 are loaded.

The embodiments of the present invention have been described above.
In the new feature registration means, the new registration of the recognition feature has been described only in the case of erroneous recognition, but it may be learned in the case of a correct answer. For example, even if the answer is correct,
If the distance value of the candidate and the distance value of the second candidate are close to each other, there is a possibility of erroneous recognition next time, which is effective for learning. Also,
Also in the case of moving the template in the feature order changing means, only the case where the personal dictionary is used for the first candidate has been described, but the learning may be performed when the template is used for the second candidate and thereafter. Further, when the templates are integrated in the new feature registration means, the integration is performed for every two templates in order, but the integration may be considered for three or more templates at the same time. Further, although the number of templates that can be registered in the personal dictionary is limited to the maximum number of stored templates, it may be limited according to the dictionary capacity if the capacity of each template is not constant. Further, when changing the order of the templates in the personal dictionary in the feature order changing means, the physical order in the personal dictionary was changed, but the template is given a sequence number indicating the order in the personal dictionary, The order of each template in the personal dictionary may be managed by changing the order number without changing the physical order.

[0088]

As described above, according to the present invention, the following various effects can be obtained.

It is possible to prevent the dictionary capacity from becoming huge without stopping the learning function. The reason is that the maximum number of templates that can be registered in the personal dictionary is limited, the usage status of the currently registered templates is managed, and when new templates cannot be registered, personal templates that are not being effectively used. This is for deleting and then registering.

Appropriate learning can be realized with the learning effect appearing in a short period of time and less adverse effects. The reason is that it is configured to learn all when misrecognized. Further, it is configured such that the template effective for recognition is left and the template having an adverse effect is sequentially deleted. Also, when the features are weighted and registered for registration, the threshold is made variable according to the number of learning features, and a stable template is created. This is also because it is configured to store the number of times of effective use, and it is difficult to delete a template that is frequently used. This is also because the recognition feature is configured to be weighted with the standard dictionary of the general-purpose dictionary and registered. Also,
This is because in the matching process using the personal dictionary, a penalty is added according to the number of learning features, and the adverse effects of unstable standard features are suppressed.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration example of a first exemplary embodiment of the present invention.

FIG. 2 is a flowchart showing the operation of the first embodiment.

FIG. 3 is a flow chart showing an operation of changing a feature order according to the first embodiment.

FIG. 4 is a flowchart showing an operation of registering a new feature according to the first embodiment.

FIG. 5 is a flowchart showing an operation of template integration according to the first embodiment.

FIG. 6 is a diagram showing a specific example of a personal dictionary according to the first embodiment.

FIG. 7 is a diagram showing a specific example of a recognition candidate according to the first embodiment.

FIG. 8 is a diagram showing a specific example of learning of a personal dictionary according to the first embodiment.

FIG. 9 is a diagram showing a specific example of learning of the personal dictionary according to the first embodiment.

FIG. 10 is a block diagram showing a configuration example of a second exemplary embodiment of the present invention.

FIG. 11 is a diagram showing a specific example of a personal dictionary according to the second embodiment.

FIG. 12 is a diagram showing a specific example of learning of the personal dictionary according to the second embodiment.

FIG. 13 is a block diagram showing a configuration example of a third exemplary embodiment of the present invention.

FIG. 14 is a block diagram showing a configuration example of a fourth exemplary embodiment of the present invention.

FIG. 15 is a block diagram showing a configuration example of a fifth embodiment of the present invention.

FIG. 16 is a block diagram showing a configuration of an embodiment of a first conventional technique.

FIG. 17 is a block diagram showing a configuration of an embodiment of a second conventional technique.

[Explanation of symbols]

10 ... Data input means 20 ... Feature extracting means 30 ... General dictionary 31, 32 ... Personal dictionary 40, 41 ... Collation means 50 ... Result correction means 51 ... Display means 52 ... Correction information input means 60, 61, 62 ... New feature registration means 70, 71 ... Feature order changing means

─────────────────────────────────────────────────── ─── Continuation of the front page (56) Reference JP-A-8-315076 (JP, A) JP-A-11-143992 (JP, A) JP-A-9-35007 (JP, A) JP-A-9- 50491 (JP, A) JP 7-49951 (JP, A) JP 11-110487 (JP, A) JP 7-49928 (JP, A) JP 59-186075 (JP, A) JP-A-59-106085 (JP, A) JP-A-59-794 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06K 9/00-9/82

Claims (8)

(57) [Claims] 1. A data input unit for acquiring input character data, a feature extracting unit for extracting a recognition feature from the character data obtained by the data input unit, and a general-purpose standard feature for each character category of the recognition feature in advance. The stored general-purpose dictionary, the personal dictionary in which personal standard features are sequentially stored up to a predetermined number (the maximum number of stored templates) by learning, and the recognition features obtained by the feature extraction means are stored in the general-purpose dictionary and the Matching means that sequentially matches each standard feature stored in the personal dictionary and outputs the character category and template number of the standard feature with a small distance value, and the recognition result is displayed by the display means, and the correct answer category is displayed by the correction information input means. And the result correction means for correcting and confirming the recognition result in response to the designation of When registering the correct answer category and the recognition feature obtained by the feature extraction means in the personal dictionary, if the same template as the correct answer category already exists, the distance value between each standard feature is compared with a predetermined threshold value. If the number is smaller, the standard feature is updated by weighted summing, otherwise a new template is registered, and if the number of templates exceeds the maximum number of stored templates, new feature registration means for deleting the oldest template is added. If the standard feature of the personal dictionary is effectively used for recognition using the correct category and the recognition candidate obtained by the result correcting means, the order of the template including the standard feature in the personal dictionary is moved to the newer one. If, on the contrary, the recognition is adversely affected, the feature order change that moves the order of the template containing the standard feature in the personal dictionary to the older one And a character recognition device. Wherein said characteristic sequence changing means, the character recognition apparatus according to claim 1, wherein altering the physical order within personal dictionary template. Wherein the characteristic sequence changing means, the character recognition apparatus according to claim 1, wherein changing the sequence number that indicates the order in the personal dictionary template. 4. The character recognition device according to claim 2 , wherein the predetermined threshold value used by the new feature registration means for comparison depends on the number of learned features. 5. The personal dictionary stores, in addition to standard features, the number of times of effective use, and the new feature registration means sets the number of times of effective use of the template to a predetermined threshold value when deleting the oldest template. If it is larger, the order in the personal dictionary of the template is moved to the newer side and the number of times of effective use is subtracted by a predetermined number, and the feature order changing means, when the standard features of the personal dictionary are effectively used, When the order in the personal dictionary of the template containing the standard feature is moved to the new one and the number of times of effective use is added at the same time, which adversely affects the recognition, in the personal dictionary of the template containing the standard feature, The character recognition device according to claim 4, wherein the number of effective uses is subtracted at the same time that the order of is moved to the older one. 6. The new feature registering means, when registering the recognition feature in the personal dictionary, adds the standard features of the same category in the general-purpose dictionary with a weighted sum at a ratio according to the number of learning features to register. The character recognition device according to claim 4 or 5 . 7. The matching means adds a penalty according to the number of learning features for the standard feature of the personal dictionary when obtaining the distance value between the recognition feature and each standard feature of the dictionary. The character recognition device according to claim 4, 5 or 6 . 8. A recording medium having recorded thereon a program for causing a computer connected to a data input means, a display means and a correction information input means to function as a character recognition device, wherein the computer is provided with characters obtained by the data input means. Feature extraction means for extracting recognition features from data, a general-purpose dictionary in which general-purpose standard features for each character category of recognition features are stored in advance, and learning standard features up to a predetermined number (maximum number of stored templates) A personal dictionary that is stored, recognition features obtained by the feature extraction means are sequentially collated with standard features stored in the general-purpose dictionary and the personal dictionary, and character categories and template numbers of standard features with small distance values are stored. Collating means for outputting the recognition result by the display means, and correct by the correction information input means. Result correction means for correcting and fixing the recognition result in response to designation of a category, and when registering the correct category obtained by the result correction means and the recognition feature obtained by the feature extraction means in the case of erroneous recognition in the personal dictionary If the same template as the correct category already exists, the distance value between each standard feature is compared with a predetermined threshold depending on the number of learning features, and if it is smaller, the standard feature is updated by weighted summing. , Other than that, a new template is registered, and when the number of templates exceeds the maximum number of stored templates, new feature registration means for deleting the oldest template is used, and the correct category and recognition candidate obtained by the result correction means are used. , If the standard features of the personal dictionary are effectively used for recognition, the order of the template containing the standard features in the personal dictionary is updated. If it adversely affects the recognition, the template containing the standard feature is moved to the older one in the personal dictionary. recoding media.