JPH07296102A - Data input method - Google Patents

Abstract

(57) [Summary] [Purpose] To improve the efficiency of correction work when a large amount of data is input using OCR. [Structure] Image data is stored on a memory (201), a recognition target character is extracted, correction is performed by character recognition and word matching (202), and a recognition result is stored in a recognition result storage disk (207). For the parts that were not corrected by word matching, check sheet (2
The image data, its appearance position, the recognition candidate character or the reject character is output to (08) (203). Next, when the candidate character and the image data do not match, the correction data (209) is manually input (204), and the recognition result (207) is corrected according to the correction data (205).

Description

Detailed Description of the Invention

[0001]

This invention recognizes large numbers of documents,
The present invention relates to a system for inputting data, and particularly to a highly accurate and efficient data input method.

[0002]

2. Description of the Related Art OC is generally used as a document input method.
There are a method using R and a method using key punch.

Many methods have been proposed for character recognition for character input (Japanese Patent Application No. 4-51305: character recognition method, etc.). In addition, as a method of improving the accuracy of the recognition result, a method of word check using a word dictionary (Japanese Patent Laid-Open No. 4-115
No. 384: Japanese OCR with word check function, and correction method based on the connection relation with the preceding and following characters (Japanese Patent Laid-Open No. 63-1932)
No. 87: How to read characters).

The higher the final required accuracy for the input data, the more the amount of work required for manual visual check due to erroneous recognition and erroneous input discovery. In order to reduce the amount of work, efficient erroneous recognition detection means and correction means are required.

As a conventional means for detecting and correcting erroneous recognition of OCR, a personal computer (PC), a workstation (WS) or the like is used to display the character image data and the recognition result on the display to make a distinction. If there are no characters, you can change the color or add a mark.
-24733: Character reader), displaying candidate characters,
There is a method of selecting and inputting (JP-A-57-20873: optical character reader).

As a method of inputting correction data for correcting the input code information, there is a method called "pinching". This is a method of identifying and correcting the position of the correction character by inputting several characters before and after the correction character.
This method requires many characters to be input in order to correct one character, but has the advantage that it is not necessary to input the position information of the correction character, and only the information of the code data to be corrected can be used for correction. is there.

[0007]

Of the above-mentioned conventional techniques,
The method of displaying indistinguishable characters on the display is good for correcting a small number of documents, but has a problem in work efficiency when correcting a large number of documents. There are problems such as a reduction in correction efficiency due to simultaneous checking and correction, and a reduction in the amount that can be checked at one time because the range that can be displayed on the display is limited. Although it is possible to share the work by outputting on paper, there is a problem that the number of terminals is limited when outputting on display. Further, since the operator's fatigue level increases by continuing to look at the display, check mistakes may increase.

When checking for errors in a large number of input documents, it is easier to work on paper, but the error detection rate is low by simply outputting the recognition result and visually checking the original. Not only that, the amount of work becomes enormous. Therefore, a method of outputting the recognition result that is easy to check is required.

Further, in the above-mentioned conventional "sandwiching" correction, it is necessary to input several characters before and after in order to correct one character. If the correction amount is small, there is no problem, but if the correction amount is large, the increase in the number of input characters becomes a problem. Therefore, if there are more than a few corrections,
There is a need for a method of outputting correction results by adding information for facilitating the input of correction data and efficiently inputting the correction data using the recognition results.

[0010]

[Means for Solving the Problems] The OCR recognition result is corrected by a logical check such as word matching, and the check data is not output for a portion that passes the logical check. Further, even for the portion which cannot be logically checked, the character recognition result having a high certainty and the character having a low certainty are separately output. Image data, characters of the recognition result, and position information for identifying the position of the character in the recognition result are output to a check sheet for checking and correcting the recognition result.

Further, based on the correction information such as the position information of the correction character and the correction character input from the keyboard or the like,
A means for correcting the recognition result by identifying the correction range in the recognition result of the correction target, replacing the characters in this range with the character code of the correction data, and correcting the recognition result.

[0012]

[Function] A document whose content is limited or whose wording is not changed has many parts that can be logically checked.
Further, the logically checked portion has sufficiently high accuracy. Therefore, by not outputting the logically checked portion to the check sheet, the target of manual check is limited, and the amount of check work is reduced.

It is necessary to change the display method of the recognition result such as outputting the image data and the character of the recognition result to the check sheet, and using the emphasized character and the reject character for the character with low certainty of recognition. Thus, it is possible to improve the checking efficiency and improve the detection rate of erroneous recognition, as compared with the case where the character code is simply output and compared with the original.

Further, when inputting the correction data after the check, the position information of the character to be corrected and the character code to be corrected, which are described on the check sheet, are input. As a result, the number of input characters for identifying the character position can be reduced, and efficient correction data input work can be performed.

By using the present invention, in the data input using the OCR, the efficiency of the correction work and the accuracy of the result after the check correction work can be improved.

[0016]

1 is a block diagram of a document data input system according to an embodiment of the present invention. The image data input from the scanner 101 is read out via the scanner controller 102, and is read out via the disk controller 103.
It is stored in the disk 104. The CPU 105 reads the program stored in the memory 106 and performs recognition processing. Image data from the disk 104 to the memory 106
The result of reading up is stored in the disk 104. Further, a check sheet for correction is output from the printer 108 via the printer controller 107.

To check and correct the recognition result, the check sheet is manually checked and the correction data is input to the keyboard 10.
It is done by inputting from 9. The correction data is stored in the memory 106 via the keyboard controller 110. Further, the recognition result stored in the disk 104 is read out on the memory 106, and the recognition result is corrected according to the input correction data. The correction result is stored in the disk again.

FIG. 2 is a schematic processing flow of an embodiment of the present invention.

Process 201: In this system, first, image data is read from the disc.

Process 202: The recognition target field is extracted from the image data, and the recognition target character string and paragraph are extracted from each field. The character pattern is recognized and the candidate character is output. Furthermore, a logical check such as word matching is added to the recognition candidate of each character to improve the recognition accuracy. The recognition result is stored in the recognition result storage disk (207).

Process 203: As for the result after the logical check is added, which is not passed by the logical check,
The image data, the recognition candidate, and the appearance position are output to a check sheet (208), and the correctness is manually checked. Further, depending on the evaluation value of recognition, the character for reject identification (such as "?") Is output as unrecognizable.

Here, the portion that has passed the logic check is
By not outputting the check sheet, the range of manual checks is reduced.

The check sheet output in the process 204: 203 is manually checked, and the correction data (209) is input from the keyboard or the like for the character that has been erroneously recognized.

Process 205: The input correction data (20
The recognition result (207) is corrected based on 9).

Process 206: The correction result is output to the correction result storage disk (210).

By the above processing, the image data is recognized and corrected.

FIG. 3 shows an example of the check sheet. Thirty
1 is the data name or file name of the recognition result. 302 is the number of pages of check sheets with the same data name. Reference numeral 303 is a paragraph number for identifying the paragraph position of the corrected portion. A correction character position 304 outputs the character position of the correction target in the paragraph. The position of the corrected character on the data file is represented based on the above data. Thirty
Reference numeral 5 is image data corresponding to characters that have failed the logical check. Image data including before and after undecidable characters is output when checking. 306 is a recognition result of the image data.

A portion for which the recognition result is determined to be correct by the logical check is distinguished by a method such as shading. Characters with low reliability such as low quality characters such as faint or crushed characters, or characters that cannot be distinguished because there are similar characters do not output the recognition result but output reject characters such as “?”. By doing so, you can prevent check mistakes.
The reliability is determined by comparing the evaluation values of the first and second recognition candidates, and if the difference is less than or equal to a threshold value, the reliability is determined to be low.

FIG. 4 shows an example of correction data that is input. The correction data is the correction paragraph number, the correction character start position,
Enter the correction character end position and the correction character code. If there is only one correction character, the end position of the correction character can be omitted.
If the correction paragraph number is the same as the correction data immediately before, this can also be omitted. One piece of correction data ends when a line feed code is input at the end.

401 is when one character is wrong,
Enter the correction paragraph number, character position, and correction character code.
As a result, the character at the target position is corrected.

Reference numeral 402 denotes a case in which noise is recognized as a character and is deleted by inputting only the corrected paragraph number and character position.

Reference numeral 403 denotes a case where one character originally has been recognized as two characters, and by inputting the correction paragraph number, the correction character start position, the correction character end position, and the correction character code, two characters are converted into one character. To fix.

Reference numeral 404 is a case where two characters are recognized as one character, and one character is corrected to two characters by inputting the corrected paragraph number, the corrected character position, and the corrected character code for two characters.

Reference numeral 405 denotes a case where a character that has not been extracted is added and is distinguished by inputting "a" instead of the correction end character position.

FIG. 5 shows reject numbers, fields,
It is a rejection number correspondence table that associates paragraphs. In the process 203, this table is created together with the output of the reject list.

Reference numeral 501 is a reject number. 502 is an image number. 503 is a field number in the image. Reference numeral 504 is a paragraph number in the field.

During the correction process, this table is used to associate the reject number with the paragraph to be corrected, whereby the correction position can be identified only by the reject number, and the number of characters input for correction can be reduced. It

The contents of the recognition process of the process 202 will be described below.

First, character recognition processing is performed according to the flowchart of FIG.

Process 601: From an image, a ruled line serving as a reference,
A mark or the like is extracted, and a recognition field corresponding to a predetermined format is set based on the position information.

Process 602: A character component is extracted for each set field. The character component is extracted based on the black pixel connected component in the field, but if there is contact between characters, they are separated according to the standard size of the character.

Process 603: A character string is generated by integrating the character components for each field in the character arrangement direction.

Process 604: Paragraphs are collectively extracted from information such as indentation of each line.

Process 605: Extract the character component of each paragraph,
Character recognition is performed and a recognition result (606) is output.

As the format analysis method and the character extraction method, "extraction of individual characters in horizontal Japanese document", IEICE Transactions, '85 / 11 Vol.J68-D No.11
The method of pp.1899-1909 can be applied. The method described in Japanese Patent Application No. 4-51305 can be applied to character recognition.

As the character recognition candidates, as shown in FIG. 7, a plurality of character candidates are output together with the evaluation value.

Reference numeral 701 is the number of characters, 702 is the number of output candidate characters, 703 is a character number, 704 is a candidate character, and 705 is a portion for storing an evaluation value.

As shown in FIG. 8, a plurality of recognition candidates are obtained for the character pattern, and the recognition result is corrected from the candidate characters by logical checking such as word matching. At this time, as a method of correcting the recognition result by word matching, the method disclosed in Japanese Patent Laid-Open No. 4-115384 can be used.

FIG. 9 shows a storage format of the recognition result 207 (same as the correction result of 210). The recognition result corrected by word matching is stored as a character string for each paragraph.

The image number 901 stores the number of the image that is uniquely determined. The field number 902 stores the number of the recognition field. The paragraph number in the field is stored in the paragraph number 903. 90
In the recognition result of No. 4, the character code data of the paragraph identified by 901 to 903 is stored.

By the above processing, highly accurate character recognition and efficient correction of the recognition result are realized.

[0052]

By using the logic check for the character recognition result, high recognition accuracy can be obtained. Therefore, the check sheet for checking and correcting the recognition result is not output for the portion that has passed the logic check. Further, even for the part that does not pass the logical check, the output form to the check sheet is changed according to the evaluation value of the character recognition, so that the target of the manual check is limited and the check efficiency is improved.

Further, the correction character position information output on the check sheet is used to reduce the number of input characters for designating the position of the correction character, thereby realizing efficient correction data input.

According to the present invention, in data input using OCR, highly accurate recognition, efficiency of check / correction work, and high accuracy of results after check / correction work can be realized.

[Brief description of drawings]

FIG. 1 is a block diagram of a system configuration of an embodiment of the present invention.

FIG. 2 is a schematic flowchart of processing contents according to an embodiment of the present invention.

FIG. 3 is a diagram showing an output example of a check sheet.

FIG. 4 is a diagram showing an example of inputting correction data for correcting a recognition result.

FIG. 5 is a diagram showing a correspondence table of reject numbers, images, fields, and paragraphs.

FIG. 6 is a flowchart of character recognition processing.

FIG. 7 is a diagram showing an output result of character recognition candidates.

FIG. 8 is a diagram showing an example of a result of word matching processing.

FIG. 9 is a diagram showing output contents of a character recognition result.

[Explanation of Codes] 101 ... Scanner, 102 ... Scanner Controller, 1
03 ... disk controller, 104 ... disk, 10
5 ... CPU, 106 ... Memory, 107 ... Printer controller, 108 ... Printer, 109 ... Keyboard, 11
0 ... Keyboard controller, 207 ... Recognition result storage disk, 208 ... Check sheet, 209 ... Correction data, 210 ... Correction result storage disk.

Claims (1)

[Claims] 1. A data input method for inputting characters in a document to a computer system, comprising: character extracting means for extracting each character pattern from a document image; character recognizing means for recognizing the extracted character pattern; Character recognition candidate output means for outputting a candidate character for recognition, recognition candidate correction means for correcting the output recognition candidate by logical check, and its reliability from the results obtained by the recognition candidate correction means and the character recognition means. A recognition result discriminating means for discriminating, and a recognition result check according to the discrimination result obtained by the recognition result discriminating means,
A check / correction information output means for outputting information for making correction, a means for inputting correction data based on the check / correction information, and a recognition result obtained by the character recognition means according to the input correction data. A data input method characterized by having means.