JPH03185582A - Dictionary correction system for character recognizing device - Google Patents

Abstract

PURPOSE:To improve the accuracy of character recognition by recording the correcting state or the like of a dictionary finding an answer in each time correcting a reject or misreading, analyzing accumulated recording results and correcting the dictionary. CONSTITUTION:When a reject (disabled reading) or misreading is generated as the result of character recognition, a display part 4 displays a character image for correcting the reject or misreading. The display is observed to correct characters, and when the correction is necessary, the correcting state or the like of the dictionary finding an answer is recorded in an error recording part 6 at the time of correcting the reject or misreading. The recording is repeatedly executed and recording data are stored. The recording data stored in the error recording part 6 are analyzed by an error analysis part 6, and as to the data satisfying a prescribed condition as the result of the analysis, the dictionary 3 is corrected by a dictionary correction part 7. Thus, the accuracy of character recognition can be improved.

Description

[Detailed description of the invention] 〔table of contents〕 overview Industrial applications Conventional technology Problems that the invention aims to solve Means to solve the problem (Figure 1) Effect (Figure 1) Example (Figures 2 to 8) Effect of the invention 〔overview〕 Concerning dictionary correction methods for character recognition devices.

A dictionary that has not been verified and is not initially valid.

From judgment of operational status to effective dictionary.

The purpose is to be able to increase character recognition accuracy.

An input section that reads characters from the original and inputs image data, and from the input image data.

Equipped with a character recognition unit that recognizes characters using a dictionary.

In a dictionary correction method of a character recognition device that corrects a dictionary based on data of character recognition results, an error recording section records correction status of reading results, etc., and an error analysis section analyzes recording results of the error recording section.

We have a dictionary correction section that corrects the corresponding dictionary based on the analysis results of the error analysis section, and every time a rejection or misreading is corrected, the dictionary that gives the answer (2) records the correction status, etc., and stores it. The dictionary is configured to analyze the recorded results and correct the dictionary.

[Industrial application field]

The present invention relates to a dictionary correction method for a character recognition device.

More specifically, it is used in two-character recognition devices, etc.
Especially when illegible or misread characters.

Relating to a dictionary modification method for a character recognition device that improves the recognition accuracy of 52 characters by making it possible to modify the output contents (character recognition results) given to the dictionary without changing the basic contents of the character recognition dictionary .

[Conventional technology]

Conventional two-character recognition devices are equipped with a dictionary for two-character recognition, and this dictionary is used to recognize characters.

In order to improve the accuracy of two-character recognition using the character recognition device described above, it is well known to create a dictionary of non-rejectable characters (characters that cannot be read) and add them to the dictionary in the current device (simple dictionary addition). It was getting worse.

[Problem to be solved by the invention]

The above-mentioned conventional devices had the following drawbacks.

(1) When adding to a dictionary (simple addition), that is, by learning, incompleteness appears in the dictionary due to insufficient learning or data loss.

(2) Also, the correlation with the dictionary depends on the user's learning, which often reduces recognition accuracy.

It was only used as an addition to a dictionary for reading special characters by specific users.

(3) Since the recognition accuracy of a character recognition device is generally due to insufficient learning for dictionary creation, reading accuracy can be improved by increasing learning and increasing the number of dictionaries. However, character databases for learning and verification do not necessarily include all character data during actual operation.

Therefore, the character rate that appears relatively frequently during operation - Even in shape, the frequency of occurrence on the arc base is low.

Due to concerns about misreading due to lack of learning, dictionaries for two-character recognition have not been created. This means that the two-character recognition accuracy depends on the character database of the device provider.

The present invention solves these conventional drawbacks and makes it possible to change an unverified dictionary that is not initially effective into an effective dictionary based on judgment of the operating situation, thereby increasing the accuracy of single character recognition. The purpose is to make it possible.

[Means to solve the problem]

Figure 1 is a diagram of the principle of the present invention, and in Figure 2, 1 is an input section, 2
is a character recognition section, 3 is a dictionary, and 4 is a display section.

5 is an error analysis section, 6 is an error recording section, and 7 is a dictionary correction section.

In order to achieve the above object, the present invention includes an input section 1 that reads characters on a document and inputs image data, and a text section 2 and 1 that recognizes characters from the input image data.
In a dictionary correction method for a character recognition device that includes a dictionary 3 for character recognition and two display sections 4 and corrects the dictionary based on the results of two-character recognition.

An error recording section 6 that records correction status of reading results, an error analysis section 5 that analyzes the recording result of the error recording section 6, and a corresponding dictionary is corrected based on the analysis result of the error analysis section 5. A dictionary correction section 7 is provided.

Each time a rejection or misreading is corrected, the correction situation for the dictionary that gave the answer is recorded, and the dictionary is corrected by analyzing the accumulated recording results of the error recording section 6. It is.

[Effect]

Since the present invention is structured as described above, it has the following first-order effect.

The input unit 1 reads characters on a document and inputs image data. The input image data is taken into the 9-character recognition unit 2 and character recognition is performed using the dictionary 3.
/ If a rejection (unreadable) or misreading occurs as a result of this character recognition, a character image for correcting the rejection or misreading is displayed on the second display section 4.

Characters are corrected by looking at this display, and if correction is necessary, the correction status of the dictionary that gave the answer when correcting rejects or misreadings is recorded in the error recording section 6. This recording is performed repeatedly and the recorded data is accumulated.

Thereafter, the recorded data accumulated in the error recording section 6 is analyzed by the error analysis section 5. In this analysis result, the dictionary correction unit 7
Modify dictionary 3 by

In this way, equipment providers may use dictionaries that have not been properly trained or verified.
Providing many dictionaries in a format that was initially rejected, 5-
- Operational status of the server: Records and accumulates the status of corrections to the reading results in the machine.

After that, the accumulated data was analyzed, and a part of the dictionary was revised based on the analysis results (=), and it was written so that it could be changed.

Therefore, it is possible to change an insufficiently verified dictionary that is initially ineffective (rejected) into an effective dictionary based on the judgment of the operational status, thereby improving the single character recognition rate.

〔Example〕

Two embodiments of the present invention will be described below based on the drawings. Figures 2 to 8 are diagrams showing one embodiment of the present invention, and Figure 2 is a block diagram of a character recognition device. Figure 3 is 9
Figure 4 shows the operation flowchart of the character recognition device, Figure 4 shows the data format of the reading results, Figure 5 shows an example of an error table, and Figure 6 shows character recognition in the initial state. FIG. 7 is a diagram showing an example of a dictionary, FIG. 7 is a diagram showing a modified character recognition dictionary technique, and FIG. 8 is a diagram showing a flowchart of dictionary modification processing.

In the figure, the same reference numerals as in FIG. 1 indicate the same parts. Further, 8 is a work stage 1, 9 is a CPU, 10 is a reading result storage section, 11 is a key manual section, 12 is a control section, 13 is an image memory, and 14 is a feature extraction section.

Reference numeral 15 indicates a collation unit, and 16 indicates a dictionary.0 The human power unit 1 is composed of an image input means such as an image scanner, and the dictionary 16 is a dictionary for character recognition provided in the 9 character recognition w2, and is configured by RAM. do.

The work stage 8 is equipped with a CPU that performs overall control, etc.
9. A reading result storage unit 10 that stores reading result data.
Display section 4. Key human resources department 11. Error recording section 6. Error analysis section 5. A dictionary 3 etc. will be provided.

The dictionary 3 stores dictionary data in an auxiliary memory such as a magnetic disk drive.The operation of the character recognition apparatus having such a configuration will be explained with reference to the flowchart shown in FIG. Note that the numbers in parentheses indicate each processing number shown in 9.

When starting the device (see figure 3), when the power is turned on (
100), Send the dictionary data and character recognition program from the work stage w 78 to the character recognition unit 2 (10
1).

In this case, the dictionary data read from dictionary 3.

It is sent to the dictionary 16 and stored. With this operation, preparation for character recognition is completed (102).

Next, the nine character recognition and error table creation processing (see Figure 3) is carried out as follows.

When the operator issues a reading instruction (103), the input unit 1 reads the original one sheet at a time (104) and inputs image data to the character recognition unit 2. The character recognition unit 2 stores the input image data in the image memory 13. Under the control of the control unit 12, this data is subjected to character feature extraction in the feature extraction unit 14, character recognition by comparing it with the dictionary 16 in the collation unit 15 (105), and recognition to the upper work stage code 8. Send the results.

If the recognition result is unreadable or misread, the CP
The U9 requests the corresponding character image data from the control unit 12, and displays the received character images for correction of objects and misreadings on the display unit 4 (106).

The operator looks at this display and inputs the corrected characters from the key manual section 11. This will correct 9 characters (1
07). At this time, it is determined whether misreading cannot be tolerated or whether the character shape is rejected and the character is desired to be read, and the corresponding dictionary correction key is used to notify the device of the intention.

When this dictionary correction key is pressed (108), the CPU 9
detects the error and instructs the error recording unit 6 to create an error table.

The error recording unit 6 records data such as the correction status for the character recognition dictionary that gave the answer (creates an error table) when rejecting or misreading is corrected (109). Also, at this time, unless the dictionary correction key is pressed, an error table is not created.

The above-mentioned character correction and error table creation are continuously executed until the end key is pressed (110), and when the end key is pressed, the above processing ends (111). After that 2
After waiting for the next reading instruction, the same character recognition process as above is performed.

Letter i! When the internal processing is completed and data is accumulated in the error recording section 6, the results are analyzed and the dictionary is corrected (third step
(See Figure C).

In this process, first, when an instruction for analysis and dictionary correction is given from the key human resource section 11 (112), the error analysis section 5 analyzes the accumulated data in the error recording section 6, and based on the result, the C
The PU 9 corrects the corresponding part of the dictionary 3 (113) and ends the process (114).

The data of Dictionary 3 corrected as above is as follows.

When the power is turned on next time (see Figure 3), the data is stored in the dictionary 16 of the single character recognition unit, so that character recognition processing is executed using the corrected dictionary.

Next, a specific example of dictionary correction will be explained based on FIGS. 4 to 8.

The data format of the reading result sent from the character recognition unit 2 to the work stage 8 is as shown in FIG. 4, for example. In this example, the character code of the answer is rejected for dictionary hidden 1230, and this dictionary has "A" as candidate 1. Furthermore, the character code of the answer is rBJ in the dictionary t4a5, and this dictionary has rBJ in candidate 1.

By the way, the character recognition dictionary in the initial state is as shown in Figure 6.2 For example, dictionary Na 1230 has a flag of "0" and is an invalid dictionary. Even if there is, it will not be used as a dictionary, so the answer will be as above.

Further, the dictionary N15 is a valid dictionary with the flag set to "1", and since candidate 1 is rBJ and there is no data after candidate 2 (there is one candidate character), the answer is given as rBJ.

The error table (data table of character correction status etc. recorded in the error recording section 6) is as shown in FIG.

To create this error table, when you want to read a reject correctly, operator operation (operation of the dictionary correction key) gives a correction result (correct answer) corresponding to the dictionary base of the corrected character, adds +1 to the count value, and also adds 1 to the count value. This is done by adding a correction result (correct answer) corresponding to the dictionary basis of the corrected character and adding +1 to the count value by operator operation when the user wants to reject the corrected character.

As a result, in the dictionary 2411230, answer 1 is "A", count 1 is "20", and there is no data after answer 2, so the number of times the operator corrected the answer as "A" is 20 times.

Also, in the dictionary t4a5, answer 1 is rBJ and count 1 is "0".
”, answer 2 is “8” and count 2 is “3”, so the answer is rBJ is not corrected, and the answer is corrected to “8” is 3
times.

Since this dictionary N15 is a valid dictionary (the flag is l),
Answer "B" if the two letters are numbers.

``8'' may be correct instead of rBJ 0 If there is such a misreading, change answer 2 to ``8'' and count.

The process of analyzing the accumulated data of the error table recorded as described above and correcting the dictionary will be explained based on the flowchart of FIG. 8. In addition.

The numbers in parentheses indicate the respective processing numbers shown in the two figures.

When the beginning of the error table is specified (121) by starting the process (120), the first error table is analyzed (122).

Next, the dictionary analyzes whether there is a misreading state (123).

As a result, if it is not a misreading state, the count value of the error table is searched at the acceptance threshold, and if the condition of count 1>threshold is satisfied and there is no answer after answer 2, (124),
The dictionary is corrected so that the 7 lag is set to "1" and becomes an effective dictionary (125).If the above conditions are not met, the dictionary is not corrected.

Also, as a result of the above analysis, if it is not a misreading state (123)
, the count values of the error table are searched with the misreading threshold E, and the counts 2 to n
It is analyzed whether any of the following conditions is satisfied: count value> misreading threshold E (128).

As a result, it is necessary to judge whether there is a dictionary that satisfies the two conditions.12
9) If there is, modify the dictionary (add it to candidates).130
), otherwise no corrections will be made.

After that, it is necessary to judge whether the error table has ended or not (126).
, if not completed, analyze the secondary error table and repeat the above process. When there are no more error tables, the dictionary correction process ends (131).

For example, in the above example, the top of the error table is dictionary N11.
1230, answer 1 is "a" and count 1 is r20J
, and there is no answer after answer 2. If the threshold value is set to 20, this dictionary needs to be modified, and the flag is set to "1" to make it an effective dictionary, as shown in FIG.

Also, the second error table is dictionary Na5.

Answer 1 is "B", Count 1 "OJ", Answer, t2 Cal'
-8J.

Count 2 is "3". In this case, if the threshold E<3
If so, the dictionary needs to be revised.

As a result, a correction is made to add answer 2, ``8'', as candidate 2 (see Figure 7).

Although the embodiments have been described above, there are two aspects of the present invention.

It is not limited to the above example, but can also be implemented in the following manner. (1) The character recognition device can recognize not only two characters but also symbols and the like.

(2) The present invention can be similarly applied not only to character recognition devices but also to speech and graphic recognition devices.

〔Effect of the invention〕

As explained above, the two inventions have the following effects.

(1) An insufficiently verified dictionary that is initially ineffective (rejected) can be changed into an effective dictionary based on the judgment of the operational status, and the level of character recognition can be improved (2) Operational status You can change the accuracy of the dictionary according to the
The device provider's dictionary learning limit can be significantly reduced.

[Brief explanation of drawings]

FIG. 1 is a diagram showing the principle of a dictionary correction method for a character recognition device according to the present invention. 2 to 8 are diagrams showing two embodiments of the present invention. FIG. 2 is a block diagram of a two-character recognition device. FIG. 3 is an operation flowchart of the two-character recognition device. FIG. 4 is a diagram showing the data format of the reading results. FIG. 5 is a diagram showing an example of an error table. FIG. 6 is a diagram showing an example of a character recognition dictionary in an initial state. FIG. 7 is a diagram showing an example of a corrected character recognition dictionary. FIG. 8 is a diagram showing a 70-chart of dictionary correction processing. 1... Input section, 2... Character recognition section. 3...Dictionary, 4...Display section. 5... Error analysis section, 6... Error recording section. 7...Dictionary correction section.

Claims (1)

[Claims] An input unit (1) that reads characters from a document and inputs image data, and a character recognition unit (2) that recognizes characters from the input image data using a dictionary (3). , in a dictionary correction method of a character recognition device that corrects the dictionary (3) based on data of character recognition results, an error recording section (6) that records correction status of reading results, etc.; and the error recording section (6). An error analysis section (5) that analyzes the recorded results of the error analysis section (5), and a dictionary correction section (7) that corrects the corresponding dictionary based on the analysis result of the error analysis section (5) are provided. , record the correction status for the dictionary that gave the answer, analyze the accumulated recorded results,
A dictionary correction method for a character recognition device characterized by correcting a dictionary.