JP2894305B2 - Recognition device candidate correction method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a data recognition device such as a character recognition device for recognizing characters written on a form.
More specifically, when an operator corrects a recognition result to an incorrect one because of an error in the recognition result, the correction information is used in a subsequent recognition process to improve a recognition accuracy of the recognition device. About.

[0002]

2. Description of the Related Art It is extremely difficult for a character recognition device, especially a handwritten character recognition device, to achieve a recognition rate of 100%. For this reason, a semi-automatic method of displaying a recognition result, visually checking by an operator, and replacing misread characters with correct characters is generally adopted. Various techniques have conventionally been proposed to assist such an operator in correcting the recognition result.

FIG. 14 is a block diagram showing an example of the configuration of a conventional character recognition device. The functional block diagram of the configuration of the character recognition device described in Japanese Patent Application Laid-Open No. 61-74081 (hereinafter referred to as Reference 1) is shown in FIG. It is a thing.

In FIG. 14, a character written on a form or the like is input as a character image image by a data input means 10 such as a scanner and is given to a recognition means 22.
The recognition dictionary 21 stores standard features of character patterns for each character category, and the recognition unit 22 extracts recognition features from the character image image, and extracts the recognition features and the standard features for each character category in the recognition dictionary 21. , For example, the recognition candidate category up to the fifth candidate is obtained, and the candidate adding means 50
Output to

[0005] The candidate adding means 50 performs a candidate adding process described later as necessary, and then outputs all recognition candidate categories to the result correcting means 40. The result correcting means 40 outputs the first candidate recognition candidate. The category is displayed on the display means 41 as a recognition result.

The operator looks at the recognition result displayed on the display means 41 and checks whether the characters on the form or the like have been correctly recognized. In the case of misreading, the fact that the misreading is performed is input from the correction information input means 42 such as a keyboard. The result correction unit 40 displays the second and subsequent recognition candidate categories output from the candidate addition unit 50 on the display unit 41 in response to this.

Accordingly, for example, regardless of whether the character written on the form is "mi", the recognition candidates determined by the recognizing means 22 are "width", "ke", "sa" in descending order. ,
If “H” or “Y”, and no new candidate has been added by the candidate adding unit 50, the display unit 41 displays the “width” of the first candidate as a recognition result, indicating that the operator misread the message. And the remaining candidates "ke", "sa",
“H” and “ya” are displayed on the display means 41.

If the displayed candidate contains a correct answer, the operator selects the correct answer to cause the result correcting means 40 to change the recognition result to the selected candidate. If the correct answer is not among the candidates as described above, the correct answer "mi" is input from the correction information input means 42 by, for example, kana-kanji conversion or the like, and the correction is instructed. When the correct answer is input, the result correcting means 40 performs a process of changing the recognition result to the correct answer, and performs a misreading category (a category of “width” which was the first candidate) and a correct answer category (input by the operator). "Mi" category)
Is sent to the table updating means 52. Table updating means 5
2 additionally registers the set of the misread category and the correct answer category in the category table 51 if it is not registered.

[0009] When such registration in the category table 51 is performed, the same typeface “mi” is again used.
For example, as described above, the recognition unit 22 outputs recognition candidates of “width”, “ke”, “sa”, “H”, and “ya” in order from the top and outputs the candidates to the candidate addition unit 50. Then, “mi” is added to a new candidate in the candidate adding process by the candidate adding unit 50. That is, the candidate adding unit 50 searches the category table 51 for the misread category “width” that is the same category as the category “width” of the first candidate sent from the recognizing unit 22. Is added as a second candidate to the candidates. Therefore, in the above case, the new recognition candidate is
.., And the first recognition result displayed on the display means 41 is “width”, which is the same as the previous one, but is not displayed last time when the operator instructs misreading. "Mi" is displayed on the display means 41 as a recognition candidate, and correction to the correct answer can be made only by selecting the displayed "mi".

A technique similar to the prior art shown in FIG. 14 is also described in Japanese Patent Application Laid-Open No. 1-168686 (hereinafter referred to as Document 2).

On the other hand, JP-A-59-27375 (hereinafter referred to as Reference 3), JP-A-63-208180 (hereinafter referred to as Reference 4) and JP-A-5-73709 (hereinafter referred to as Reference 3). 5), when a recognition result of a certain character is corrected in a series of character strings displayed as a recognition result, there is a portion where the same correction should be performed in subsequent recognition results. A technique is described that automatically detects whether or not to perform the same operation and corrects a plurality of erroneously read portions of the same type collectively by a single correction operation by an operator.

[0012]

As described above, in the conventional character recognition apparatus, various techniques have been proposed to assist the operator in correcting the recognition result, but have the following problems. .

[0013] In the techniques described in Documents 1 and 2,
Since the correct category registered in the category table 51 is only added as a recognition candidate of a fixed rank (the second rank in the above example), the first displayed recognition result always indicates that the recognition means is the first rank candidate. Limited to That is, the recognition result correction information by the operator has no effect on the first candidate. Therefore, even if the correct answer content rate in which the correct answer is included in the candidates is improved, the recognition rate at which the first candidate becomes the correct answer cannot be improved. For this reason, it is difficult to reduce the number of corrections themselves by the operator.

The techniques disclosed in References 3 to 5 are:
Although the efficiency of the correction work for a series of recognition results can be improved, the recognition result correction information input by the operator in the correction work is used only for correcting the current recognition result. Recognition of the recognition target character string brings the same result again. That is, the recognition rate cannot be improved by utilizing the recognition result correction information of the operator.

The present invention has solved such a conventional problem. The purpose of the present invention is to utilize the operator's recognition result correction information in the subsequent recognition processing, and to determine only the correct answer content ratio in which the correct answer is included in the candidates. Instead, it is to improve the recognition rate at which the first candidate is the correct answer.

[0016]

According to the present invention, a feature for recognition is extracted from recognition target data such as a character image and compared with standard features for each category registered in a recognition dictionary in advance. A plurality of upper recognition candidate categories having a small distance value, which is a difference between a feature and a standard feature, a recognition unit for obtaining the distance value, and a correction table for storing a set of a misread category, a correct category, and correction information. And when a set including an erroneous reading category that matches the first-ranked recognition candidate category in the group of recognition candidate categories determined by the recognition unit is stored in the correction table, the set matches the correct category in the set. Correction means for reducing the distance value of the recognition candidate category to be performed by a value corresponding to the correction information in the set; A result correction means for outputting the recognition candidate category of the rank as a recognition result, and correcting the recognition result in accordance with the correction information input from the correction information input means; Table updating means for adding a set with the correction information to the correction table.

Further, the candidate correcting means, when the recognition candidate category matching the correct answer category does not exist in the recognition candidate category group obtained by the recognition means, sets a predetermined distance value matching the correct answer category. The configuration is such that the distance value is corrected after adding the given recognition candidate category to the recognition candidate category group.

The correction information in the set added by the table updating means to the correction table includes, for example, (1) a distance difference which is a difference between a distance value of the misread category and a distance value of the correct category. (2) A distance difference, which is a difference between the distance value of the misread category and the distance value of the correct category, and a recognition feature extracted from the recognition target data. Is used.

When using the correction information as in the above (1), the candidate correction means corrects the distance value by, for example, the following equation.

D ′ = D−w · s where D ′ is a distance value after correction, D is a distance value before correction,
w is a weight coefficient, and s is a distance difference as correction information.

When the correction information as described in (2) is used, the candidate correction means corrects the distance value by the following equation, for example.

D ′ = D−w · s · a / d where D ′ is a distance value after correction, D is a distance value before correction,
w is a weighting factor, s is the distance difference in the correction information, d is the distance value between the recognition feature in the correction information and the recognition feature of the data to be recognized this time, and a is a constant for normalizing the distance value d. This is a certain reference distance value. In addition, the candidate correcting unit is configured to generate a new first candidate category when the first recognition candidate category in the group of recognition candidate categories obtained by the recognition unit is changed by candidate correction using the correction table. Is used to correct a candidate again.

In the present invention configured as described above, when the operator corrects a correct category at the time of erroneous reading, the correct category, the erroneous reading category (the category output as the first candidate by the apparatus), the correction information, Are registered in the correction table. Thereafter, when the same category as the misread category is obtained as the first recognition candidate category, the distance value of the recognition candidate category of the same category as the correct category corresponding to the misread category is reduced. Therefore, depending on the magnitude of the value to be reduced, there is a case where it jumps to the first candidate. This is a major difference from the techniques described in the above-mentioned Documents 1 and 2. The recognition result displayed first is not limited to the one in which the recognition means is always set as the first candidate, but the category previously input as the correct answer category. May be output as a recognition result. By such an operation, the recognition rate at which the first-rank candidate becomes a correct answer can be improved.

[0024]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram of an embodiment of a character recognition apparatus to which the present invention is applied. As shown in FIG. 1, the character recognition device according to the present embodiment includes a data input unit 1 and a recognition unit 2.
, A recognition dictionary 3, a candidate correction unit 4, a correction table 5, a result correction unit 6, a display unit 7, a correction information input unit 8, and a table update unit 9.

FIG. 2 is an operation flowchart of the embodiment of FIG. Hereinafter, the function and operation of each unit of the present embodiment will be described with reference to FIGS.

The data input means 1 is constituted by, for example, a scanner or the like. For example, the data input means 1 scans handwritten characters written on a form (not shown) to generate a character image image for each character and outputs it to the recognizing means 2.

The recognition dictionary 3 stores in advance the recognition characteristics (standard characteristics) of the standard character pattern for each character category. The recognition means 2 extracts a recognition feature from the character image image input from the data input means 1 and
In comparison with the standard features for each character category stored in advance, a plurality of recognition candidate categories (e.g., five to the fifth place) having smaller distance values and their respective distance values are obtained (S1). ). The obtained recognition candidate category group and each distance value are output to the candidate correcting means 4. Here, the distance value is a difference between the recognition feature and the standard feature, and the value decreases as the similarity increases. Various methods have been proposed for performing character recognition by obtaining a distance value between a feature for recognition of a recognition target character and a standard feature in a recognition dictionary, and the present invention may use any of the methods. it can. For example,
A document describing such a technique is described in "Improvement of Direction Pattern Matching Method and Application to Handwritten Kanji Character Recognition" (199).
June 2000, IEICE Technical Report, PRU90-
20).

The correction table 5 is a table for storing a set 50 of a misread category 51, a correct answer category 52, and a distance difference 53, and has a capacity to store, for example, 200 sets. Here, in the present embodiment, a distance difference 53 which is a difference between the distance value relating to the misread category 51 and the distance value relating to the correct category 52 is used as the correction information. The initial state of the correction table 5 is an empty state, that is, a state in which no set is stored. The correction table 5 can be exchanged for each user or each type of recognition target (type of form, etc.).

The candidate correction means 4 refers to the correction table 5 and newly performs a new recognition on the recognition candidate category group (the first or fifth highest recognition candidate category) inputted from the recognition means 2. This is a means for adding a candidate category or subtracting a distance value of a predetermined recognition candidate category and then sorting by distance values to generate a new recognition candidate category group, and executes the processing of steps S2 to S10 in FIG. I do.

First, the candidate correction unit 4 searches the misread categories 51 in the set 50 in the correction table 5 one by one in order from, for example, the set closest to the head of the table, and the misread category is input from the recognition unit 2. It is checked whether it is the same as the category of the first candidate in the recognized candidate category group (S2, S3, S4). Each time a misread category of the same category as the first candidate is found (YE in S4)
S), the following processing is executed.

First, it is determined whether or not a recognition candidate category of the same category as the correct answer category 52 in the set 50 including the found misread category is included in the recognition candidate category group input from the recognition means 2. Search and examine the recognition candidate categories one by one from the group one by one (S5, S5
6, S7).

Then, if a recognition candidate category of the same category as the correct answer category 52 exists (YE in S7)
S), the distance value of the recognition candidate category is corrected by the following equation (S9).

D ′ = D−w · s where D ′ is a distance value after correction, D is a distance value before correction,
w is a weight coefficient, and s is a distance difference 53 that is correction information in the set 50 including the correct answer category 52. Here, the weight coefficient w may be fixed to, for example, 1, or may be set to a value that is smaller than 1 when the set 50 in the correction table 5 is old.

On the other hand, when there is no recognition candidate category of the same category as the correct answer category 52 (YE in S6)
S), a recognition candidate category of the same category as the correct answer category 52 is added to the recognition candidate category group (S8). At this time, as the distance value of the added recognition candidate category, for example, a predetermined constant α (for example, 200) is added to the distance value of the lowest recognition candidate category in the recognition candidate category group input from the recognition means 2. Value. And
Proceeding to step S9, the distance value of the added recognition candidate category is corrected by the above equation.

After the above processing, the candidate correcting means 4
The recognition candidate categories in the recognition candidate category group are sorted in ascending order by their distance values (S10). At this time, if the distance value has been corrected, the candidate order may be different from the time when the recognition unit 2 outputs the candidate order. Out of the newly ordered recognition candidate category groups, the upper plurality (for example, five to the fifth) of the recognition candidate categories and their distance values are output to the result correction means 6.

The result correction means 6 displays the first recognition candidate category in the recognition candidate category group input from the candidate correction means 4 as a recognition result on the display means 7 constituted by, for example, a CRT (S11). . In addition, the second and lower recognition candidate categories are displayed in another display area of the display means 7 in preparation for correction in the case of misreading. The display of the second or lower recognition candidate category may be displayed when the operator instructs that the reading is erroneous.

The operator checks the recognition result displayed on the display means 7 and, if erroneous reading is performed, inputs correction information from the correction information input means 8 such as a keyboard to make correction. At this time, if a correct answer category exists in the second or lower recognition candidate categories displayed on the display means 7, information for selecting the correct answer category is given from the correction information input means 8 (YES in S12). Then, the result correcting means 6 corrects the recognition result to the selected correct answer category (S13). If the correct category does not exist in the second and lower recognition candidate categories, the operator inputs the correct category through the correction information input means 8 by, for example, kana-kanji conversion or the like, and causes the correction to be performed. At this time, the result correcting means 6 corrects the recognition result with the input correction category (S13).

When the result of the recognition is corrected, the result correction means 6 updates the category before correction (misread category) and its distance value and the corrected category (correct answer category) and its distance value in the table updating means 9. To update the correction table (S14). When the correct category does not exist in the recognition candidate category group, for example, a value obtained by adding a predetermined constant α (for example, 200) to the distance value of the lowest recognition candidate category in the recognition candidate category group is used as the correct category. The distance value.

The table updating means 9 first obtains a distance difference by subtracting the distance value of the correct answer category from the distance value of the misread category notified from the result correcting means 6. And
A set 50 including the distance difference, the correct answer category, and the misread category is additionally registered in the correction table 5.

At this time, the table updating means 9 registers, for example, new sets 50 in order from the top of the correction table 5, and when the sets 50 are registered up to the end of the table and becomes full (ie, (In the above example, when 200 sets are registered), the table is returned to the top of the table again and the oldest registered set 50 is overwritten and additionally registered. In this way, the number of sets to be registered is limited to, for example, about 200, and after the 200 sets are registered, the area of the table 5 is cyclically used, thereby preventing an increase in the memory required for the correction table 5.

In the present embodiment, the table updating means 9 determines whether a group including the same misreading category and correct answer category as the currently registered set of misreading and correct answer categories has already been registered in the correction table 5. The current distance difference is added to the registered distance difference.

Next, this embodiment will be described in more detail with reference to specific examples.

In the initial state where the contents of the correction table 5 are empty as shown in FIG. 3 (a), certain characters handwritten on a form or the like are recognized, and
It is assumed that a recognition candidate category group and a distance value as shown in FIG. That is, the first candidate is “width”, the second candidate is “yu”, the third candidate is “ke”, the fourth candidate is “ya”, and the fifth candidate is “sa”. Distance values of 1300, 1500, 1600, 1700, 1800
Assume that At this time, in the candidate correction unit 4, since no group is registered in the correction table 5, FIG.
Since the search is immediately determined to be completed in step S3, the distance value is not corrected. Therefore, the new recognition candidate category group and the distance value sorted by the distance value in step S10 become the same as those in FIG. 3B as shown in FIG. 3C. For this reason, the "width" of the first candidate is displayed on the display means 7 as a recognition result, and the second and lower candidates are displayed (S11). Here, since the original input character is "Y", the operator determines that the reading is erroneous, and selects the second candidate "Y" displayed on the display means 7, and the result correction means 6 changes the recognition result to "Y". Is corrected from "width" to "yu" (S
13). In addition, the misread category “width” and its distance value 13
00, the correct answer category “Y” and its distance value 1500 are notified to the table updating means 9. In the table updating means 9, 200 obtained by subtracting the distance value 1300 of the misread category “width” from the distance value 1500 of the correct answer category “yu” is set as a distance difference 53, and the misread category “width” and the correct answer category “yu” are added thereto. The set 50 is registered in the correction table 5 (S14). As a result, the contents of the correction table 5 are shown in FIG.
(D).

Next, the contents of the correction table 5 are shown in FIG.
In the state shown in FIG. 3A (the same as FIG. 3D),
Recognition of another certain character is performed.
It is assumed that a recognition candidate category group and a distance value as shown in FIG. That is, the first candidate is “width”, the second candidate is “ke”, the third candidate is “sa”, the fourth candidate is “H”, and the fifth candidate is “yu”. Distance values of 1100, 1150, 1200, 1400, 1400
Assume that At this time, the candidate correction means 4
Since there is one set in the correction table 5 containing the misread category “width” of the same category as the rank candidate “width”, the recognition candidate “yu” is determined according to the correct answer category “yu” and the distance difference 200 of that set. The distance value is changed from 1400 to 1200.
Therefore, the new recognition candidate category group and the distance values sorted by the distance values in step S10 are as shown in FIG. Next, the display means 7 displays the "width" of the first candidate as a recognition result, and also displays the second and lower candidates (S11). Here, since the original input character is "mi", the operator determines that the reading is erroneous and inputs "mi" from the correction information input means 8, and the result correction means 6 changes the recognition result from "width" to "mi". (S13). Further, the table updating unit 9 is notified of the misread category “width” and its distance value 1100 and the correct answer category “mi” and its distance value. Here, since the correct category “mi” does not exist in the recognition candidate category group, the distance value 1400 of the lowest candidate “yu” is set.
Is added to 200 (= α), and 1600 is set as the distance value of the correct answer category “mi”. The table updating means 9 subtracts the distance value 1100 of the misread category "width" from the distance value 1600 of the correct answer category "mi" to obtain 500 as the distance difference 53, and adds the misread category "width" and the correct answer category "mi". The set 50 is registered in the correction table 5 (S1).
4). As a result, the contents of the correction table 5 are as shown in FIG.
become that way.

Next, the contents of the correction table 5 are shown in FIG.
In the state shown in FIG. 4A (the same as FIG. 4D),
Another certain character (here, “Yu”, which is also the previous “Yu”)
(It is assumed that the handwriting is more hand-drawn compared to the above.), And the recognition unit 2 outputs a recognition candidate category group and a distance value as shown in FIG. That is, the first candidate is “width”, the second candidate is “yu”, the third candidate is “ke”, the fourth candidate is “sa”, and the fifth candidate is “ya”. Distance value of 120
0, 1350, 1500, 1800, and 1900. At this time, the candidate correction unit 4 corrects the distance value because two sets including the misread category “width” of the same category as the first candidate “width” exist in the correction table 5. First, one set of correct answer category “Y” and distance difference 2
00, the distance value of the recognition candidate “Y” is changed from 1350 to 1150. Next, a recognition candidate "mi" is added according to the remaining one set of the correct answer category "mi" and the distance difference 500, and the distance value is set to 1600 (= 1900 +
200-500). Therefore, the new recognition candidate category group and the distance values sorted by the distance values in step S10 are as shown in FIG. Next, the display means 7 displays the first candidate “Y” as a recognition result.

As described above, when the misreading is performed under the condition that the distance difference is as small as 200, and the group of the same category “width, yu” is again generated under the condition that the distance difference is as small as 150, The category "Yu" corrected as the correct answer comes to the top. In other words, it indicates that any character that is more similar to the character that was recognized when the set was registered may be the first candidate.

Next, the contents of the correction table 5 are shown in FIG.
In the state shown in (a) (the same as in FIGS. 4 (d) and 5 (a)), another certain character (here, "width" is assumed) is recognized, Suppose that a group of recognition candidate categories and distance values as shown in FIG. That is, the first candidate is “width”, the second candidate is “ke”, the third candidate is “sa”, the fourth candidate is “yu”, and the fifth candidate is “ya”, respectively. Distance value of 120
0, 1350, 1400, 1650, 1800. At this time, the candidate correction unit 4 corrects the distance value because two sets including the misread category “width” of the same category as the first candidate “width” exist in the correction table 5. First, one set of correct answer category “Y” and distance difference 2
In accordance with 00, the distance value of the recognition candidate “Y” is changed from 1650 to 1450. Next, a recognition candidate "mi" is added according to the remaining one set of the correct answer category "mi" and the distance value 500, and the distance value is set to 1500 (= 1800 +
200-500). Accordingly, the new recognition candidate category group and the distance values sorted by the distance values in step S10 are as shown in FIG. Next, the display means 7 displays the “width” of the first candidate as a recognition result.

As described above, when the erroneous reading is performed under the condition that the distance difference is as small as 200, and the pair “width, yu” of the same category occurs again under the condition that the distance difference is as large as 450, the recognition means 2 The recognition result is maintained.

Next, the contents of the correction table 5 are shown in FIG.
In the state shown in FIG. 4A (the same as FIG. 4D),
Another certain character (here, "mi", which is also the previous "mi"
(It is assumed that the handwriting is slightly more "compared to handwriting"), and a recognition candidate category group and a distance value as shown in FIG. That is, the first candidate is "width", the second candidate is "ke", the third candidate is "sa", the fourth candidate is "H", and the fifth candidate is "ya", respectively. Distance value of 120
0, 1250, 1300, 1400, 1450. At this time, the candidate correction unit 4 corrects the distance value because two sets including the misread category “width” of the same category as the first candidate “width” exist in the correction table 5. First, one set of correct answer category “Y” and distance difference 2
00, the recognition candidate “Y” is added, and the distance value is set to 1450 (= 1450 + 200−200). Next, the remaining one set of the correct answer category “mi” and the distance value 50
According to 0, the recognition candidate “mi” is added, and the distance value is set to 1150 (= 1450 + 200−500). Therefore, the new recognition candidate category group and the distance values sorted by the distance values in step S10 are as shown in FIG. 7C. Next, the display unit 7 displays the first candidate “mi” as a recognition result.

As described above, when the misreading is performed under the condition that the distance difference is as large as 500 and the pair of the same category “width, mi” appears again, even if the distance difference is large, the lower candidate becomes higher. It is possible.

Next, the contents of the correction table 5 are shown in FIG.
In the state shown in FIG. 4A (the same as FIG. 4D),
Recognition of another certain character (here, “Y”, which is also assumed to be handwritten that does not seem more “Y” than the “Y” at the time of the correction) is performed, and the recognition unit 2 From FIG. 8
It is assumed that a recognition candidate category group and a distance value as shown in FIG. That is, the first candidate is “width”, the second candidate is “yu”, the third candidate is “ke”, the fourth candidate is “ya”, and the fifth candidate is “sa”. Distance values of 1300, 1550, 1600, 1700, 1800
Assume that At this time, the candidate correction means 4
Since two sets including the misread category “width” of the same category as the rank candidate “width” exist in the correction table 5, the distance value is corrected. First, the distance value of the recognition candidate “Y” is set to 15 according to one set of the correct answer category “Y” and the distance difference 200.
Change from 50 to 1350. Next, a recognition candidate "mi" is added according to the remaining one set of the correct category "mi" and the distance value 500, and the distance value is set to 1500 (= 1800).
+ 200-500). Therefore, the new recognition candidate category group and the distance value sorted by the distance value in step S10 are as shown in FIG. 8C. Next, display means 7
Displays the "width" of the first candidate as a recognition result. Here, since the original character is “Y”, the operator determines that the reading is erroneous, and selects the second candidate “Y” displayed on the display means 7. Is corrected to "Y" (S13). In addition, the misread category “width” and its distance value 1300 and the correct answer category “yu”
And the distance value 1350 are notified to the table updating means 9. In the table updating means 9, the correct answer category "Y"
Distance value 13 of misread category "width" from distance value 1350 of
A set 50 in which 50 obtained by subtracting 00 is added to the distance difference 53 and the misread category “width” and the correct answer category “yu” is added to the correction table 5 (S14). At this time, since the registered table {width, yu, 200} having the same misread category "width" and the same correct category "yu" exists in the correction table 5, the distance difference 200 is added to the current distance difference 50. Is added. As a result, the contents of the correction table 5 are as shown in FIG.
become that way. Note that a registered group having the same misread category and correct category may not be deleted, but the current group may be registered separately. That is, in the above example, the pair {width, yu, 200} is left while leaving the pair {width, yu, 200}.
50 ° may be additionally registered.

As described above, if the misreading is performed under the condition that the distance difference is as small as 200, and the same category pair “width, yu” appears with a larger distance difference, the category “corrected as a correct answer previously” “Yu” cannot necessarily be the first candidate. However, since the distance difference in the correction table 5 is changed by performing the correction again, the distance difference that can be absorbed can be gradually increased. Therefore, as shown in the following example, in the case of reading “Y” that is more or less “Y” compared to the previous “Y”, “Y” is output again as a recognition result.

That is, the contents of the correction table 5 are shown in FIG.
In the state shown in FIG. 8A (the same as FIG. 8D),
Recognition of “Y” that is more or less “Y” compared to “Y” assumed in FIG. 8 is performed, and a recognition candidate category group and a distance value as shown in FIG. Suppose. That is, the first candidate is “width”, the second candidate is “yu”, the third candidate is “ke”, the fourth candidate is “ya”, and the fifth candidate is “sa”. Is 130
It is assumed that they are 0, 1500, 1600, 1700, and 1800. At this time, the candidate correction unit 4 corrects the distance value because two sets including the misread category “width” of the same category as the first candidate “width” exist in the correction table 5. First, one set of correct answer category “Y” and distance value 2
According to 50, the distance value of the recognition candidate “Y” is changed from 1500 to 1250. Next, a recognition candidate "mi" is added according to the remaining one set of the correct answer category "mi" and the distance value 500, and the distance value is set to 1500 (= 1800 +
200-500). Accordingly, the new recognition candidate category group and the distance value sorted by the distance value in step S10 are as shown in FIG. 9C, and the first candidate "Y"
Is displayed as a recognition result.

FIG. 10 is another operation flowchart of the embodiment of FIG. This operation flowchart differs from that of FIG. 2 in that the first-rank candidate category is held in step S15, and after the processing of steps S2 to S10 in FIG. 2 (step S16), the first-rank candidate category is changed to step S15. It is checked whether it is different from the candidate category held in (step S17), and if different (YE in S17)
The point of S) is that the processes of steps S2 to S10 are performed again (step S18). The candidate correcting means 4 of FIG.
0, the processes of steps S15 to S18 are executed.

Next, a more specific example will be described.

When the contents of the correction table 5 are in the state shown in FIG. 11A (the same as FIG. 4A), another certain character (here, "width" is assumed) is used. It is assumed that the recognition has been performed, and a recognition candidate category group and a distance value as shown in FIG. That is, the first candidate is “width”, the second candidate is “yu”, the third candidate is “sa”, the fourth candidate is “H”, and the fifth candidate is “ya”. Distance values of 1100, 1150, 120
It is assumed that the values are 0, 1400, and 1400. First, the candidate correcting means 4 stores the first candidate "width" (step S15). Next, a set including the misread category “width” of the same category as the first candidate “width” is added to the correction table 5 by one.
Therefore, the distance value of the recognition candidate “Y” is set to 1150 according to the correct answer category “Y” of the group and the distance difference 200.
To 950. Therefore, the new recognition candidate category group and the distance value are as shown in FIG. 11C (step S16). Next, it is checked whether the new first-place candidate “Y” is different from the stored candidate (in this case, “width”) (step S17). In this case, because the difference is different, a set including the misread category of the same category as the first candidate “Y” is searched in the correction table 5, but is not corrected because it does not exist (step S18). Next, the display means 7 displays the first candidate "Y" as a recognition result, and also displays the second and lower candidates (S11). Here, the original input character is "width"
Therefore, when the operator determines that the reading is erroneous and inputs "width" from the correction information input means 8, the result correcting means 6 corrects the recognition result from "y" to "width" (S13). The table updating means 9 registers the misread category “Y”, the correct answer category “width”, and the distance difference 150 (= 1100−950) in the correction table 5 (S14). As a result, the contents of the correction table 5 are as shown in FIG.

Next, the contents of the correction table 5 are shown in FIG.
In the state shown in FIG. 11A (same as FIG. 11D), another certain character (here, “width” is assumed)
Is recognized, and a recognition candidate category group and a distance value as shown in FIG. That is, the first candidate is “width”, the second candidate is “yu”,
The third candidate is “ke”, the fourth candidate is “sa”, the fifth candidate is “ya”, and the respective distance values are 1200 and 135.
It is assumed that they are 0, 1500, 1800, and 1900. First, the candidate correcting means 4 stores the first candidate "width" (step S15). Next, since there is one set including the misread category “width” of the same category as the first candidate “width” in the correction table 5, the distance value of the recognition candidate “yu” is changed from 1350 to 1150 according to the distance difference 200. Change to Accordingly, the new recognition candidate category group and the distance value are as shown in FIG. 12C (step S16). Next, it is checked whether the new first-place candidate “Y” is different from the stored candidate (in this case, “width”) (step S17).
In this case, since a difference is found, a set including an erroneous reading category of the same category as the first candidate “Y” is searched in the correction table 5, and 1
Since there are two, the recognition candidate “width” according to the distance difference 150
Is changed from 1200 to 1050. Therefore,
The new recognition candidate category group and the distance value are shown in FIG.
(Step S18).

As described above, the set “yu” and “width” that are easily misread.
If both are recognized many times, it is easy for misreading to be the opposite character in one correction process. However, when the correction process is performed twice, the result of the first recognition process is emphasized and the recognition rate is reversed. Can be prevented.

FIG. 13 is a block diagram of another embodiment of the character recognition apparatus to which the present invention is applied. The character recognition device of this embodiment differs from that of the embodiment of FIG. 1 in that a distance difference 53 and a recognition feature 54 are used as correction information. Along with this difference, the recognition means 2 ', the candidate correction means 4', the correction table 5 ', and the table updating means 9' are modified as follows with respect to the corresponding means shown in FIG.

The correction table 5 'contains the misread category 51
, A correct category 52, a distance difference 53, and a recognition feature 54.

The recognizing means 2 ′ extracts a feature for recognition from the character image image input from the data input means 1, compares it with standard features for each character category stored in advance in the recognition dictionary 3, Are smaller (for example, 5th
(5) to the recognition candidate categories and their respective distance values are output to the candidate correcting means 4 ', and the recognition features extracted from the character image are output to the table updating means 9'.

The candidate correcting means 4 'replaces the above equation with
The distance value is corrected by the following equation.

D ′ = D−w · s · a / d where D ′ is the distance value after correction, D is the distance value before correction,
w is a weighting factor, s is the distance difference in the correction information, d is the distance value between the recognition feature in the correction information and the recognition feature extracted from the current character image, and a is the distance value d. This is a reference distance value that is a constant. The calculation of the distance value d between the features for recognition may be the same as the calculation of the distance between the standard feature of each category of the recognition dictionary 3 and the features for recognition obtained from the input character image image in the recognition means 2 ', for example, the Euclidean distance. Or city block distance.

The table updating means 9 'is provided with the result correcting means 6
Is notified of the misread category and its distance value and the correct answer category and its distance value, the distance difference 53 is obtained in the same manner as the table updating means 9 in FIG.
In addition, the information for recognition input from the recognizing means 2 'is added to create correction information, and the misreading category and the correct answer category are further added to the correction information. 'Is generated.

The recognition feature registered in the correction table 5 'may be a compressed feature obtained by compressing the recognition feature. By doing so, the required capacity of the correction table 5 'can be reduced.

Other functions and operations of the character recognition device of this embodiment are the same as those of the character recognition device of FIG. In the present embodiment, when the recognition feature 54 of the correction table 5 'and the recognition feature obtained from the input character are close to each other and the distance value d is small, the distance value D greatly decreases. It is easy to come to. Conversely, when the recognition feature 54 of the correction table 5 ′ is far from the recognition feature obtained from the input character, and as a result the distance value d is large, the distance value D does not decrease so much. Indicates a characteristic that does not increase very much.

The embodiments of the present invention have been described above.
The present invention is not limited to the above embodiments, and various other additions and changes are possible. For example, the input data obtained by the data input unit is a character image image input by a scanner or the like, but may be online character data input by a tablet or voice data input by a microphone. The misread category stored in the correction table may be a template number of the recognition dictionary. If the template number is used, the template can be distinguished when a plurality of templates of the same category are stored in the recognition dictionary.

[0069]

As described above, according to the present invention, the recognition result correction information of the operator is used for correcting the distance value in the subsequent recognition processing, and the category which has been misread once is output to the output of the recognition means. When appearing as the first candidate, the distance value of the recognition candidate category that matches the correct category registered in the correction table is appropriately reduced, leaving the possibility of jumping to the first candidate. The recognition rate at which the candidate becomes a correct answer can be improved as compared with the conventional one.

When the first candidate is corrected by the correction processing, the correction processing is performed again, so that the original recognition result is obtained for the character set in which the misread category and the correct category are mutually registered. Since the possibility of lowering the recognition rate is reduced with emphasis, the recognition rate can be improved as compared with the conventional case.

[Brief description of the drawings]

FIG. 1 is a block diagram of an embodiment of a character recognition device to which the present invention is applied.

FIG. 2 is a flowchart showing an operation procedure of the character recognition device of the embodiment of FIG.

FIG. 3 is an explanatory diagram of the operation of the embodiment of FIG. 1 and shows a situation in which a first set including correction information is registered in a correction table in an initial state due to correction of misreading; FIG.

FIG. 4 is an operation explanatory diagram of the embodiment of FIG. 1, showing a situation where a second set including correction information is registered in the correction table due to correction of misreading again; is there.

FIG. 5 is an explanatory diagram of the operation of the embodiment of FIG. 1, showing an example in which a correct category can be derived by a set including correction information registered in a correction table.

FIG. 6 is an explanatory diagram of the operation of the embodiment of FIG. 1, showing an example in which a correct answer category can be derived without being affected by a set including correction information registered in a correction table.

FIG. 7 is an operation explanatory diagram of the embodiment of FIG. 1, showing another example in which a correct answer category can be derived by a set including correction information registered in a correction table.

FIG. 8 is an explanatory diagram of the operation of the embodiment of FIG. 1, in which the correct category cannot be derived even by the set including the correction information registered in the correction table, FIG. 8 is a diagram showing a situation in which a set including important correction information is registered in a correction table.

9 is an explanatory diagram of the operation of the embodiment of FIG. 1, showing an example in which a correct category can be derived by another correction.

FIG. 10 is another flowchart showing an operation procedure of the character recognition device of the embodiment of FIG. 1;

11 is an explanatory diagram of the operation of the embodiment of FIG. 1 according to the flow of FIG. 10, and shows a situation in which a set in which the misread category and the correct category are reversed is registered in a correction table.

12 is an operation explanatory diagram of the embodiment of FIG. 1 according to the flow of FIG. 10, and is a diagram showing an example in which a correct answer category can be derived by two correction processes.

FIG. 13 is a block diagram of another embodiment of the character recognition device to which the present invention is applied.

FIG. 14 is a block diagram illustrating a configuration example of a conventional character recognition device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Data input means 2, 2 'Recognition means 3 Recognition dictionary 4, 4' Candidate correction means 5, 5 'Correction table 50, 50' Set 51 Misread category 52 Correct answer category 53 Distance difference 54 Recognition feature 6 Result correction means 7 Display means 8 Correction information input means 9 'Table updating means

Claims (7)

(57) [Claims] 1. A feature for recognition is extracted from data to be recognized,
Recognition for obtaining a plurality of upper-ranked recognition candidate categories having a smaller distance value, which is a difference between the recognition feature and the standard feature, compared with the standard feature for each category registered in advance in the recognition dictionary, and the distance values thereof Means, a correction table storing a set of a misread category, a correct answer category, and correction information; and a set including a misread category that matches the first-ranked recognition candidate category in the recognition candidate category group obtained by the recognition means. Is stored in the correction table, the candidate correction means for reducing the distance value of the recognition candidate category that matches the correct answer category in the set by a value corresponding to the correction information in the set; Output the first-ranked recognition candidate category in the recognition-candidate category group after processing by the means as a recognition result, and follow the correction information input from the correction information input means. And a table updating means for adding a set of a misread category, a correct answer category, and correction information to the correction table at the time of correction by the result correction means. Method for correcting candidate recognition devices. 2. The method according to claim 1, wherein when the recognition candidate category matching the correct answer category does not exist in the recognition candidate category group obtained by the recognition means, the predetermined distance value matching the correct answer category is set. 2. The apparatus according to claim 1, further comprising a step of correcting the distance value after adding the given recognition candidate category to the recognition candidate category group.
The candidate correction method for the described recognition device. 3. The method according to claim 2, wherein the table updating means uses a distance difference between the distance value of the misread category and the distance value of the correct category as the correction information.
The candidate correction method for the described recognition device. 4. When the distance value before correction is D, the weighting factor is w, and the distance difference as correction information is s, the candidate correction means calculates the corrected distance value D ′ by D ′ = D The candidate correction method for a recognition device according to claim 3, wherein the candidate correction method is obtained by −w · s. 5. The table updating means uses, as the correction information, a distance difference between a distance value of a misread category and a distance value of a correct category, and a recognition feature extracted from recognition target data. 3. The candidate correction method for a recognition device according to claim 2, wherein: 6. The candidate correcting means recognizes a distance value before correction as D, a weighting coefficient as w, a distance difference in correction information as s, a recognition feature in the correction information and a current recognition target data. Assuming that the distance value to the feature for use is d and the reference distance value which is a constant for normalizing the distance value d is a, the corrected distance value D ′ is represented by D ′ = D−w · s · a / d 6. The method according to claim 5, wherein the candidate is corrected. 7. The method according to claim 1, wherein the first correction candidate category in the group of recognition candidate categories determined by the recognition means is changed by a candidate correction using the correction table. 7. The method according to claim 1, further comprising the step of manually correcting the candidate using the rank candidate category.