JP4718699B2 - Character recognition device, character recognition method, program, and computer-readable recording medium - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
  The present invention relates to a character recognition device,Character recognitionMethod,program,andComputer readableThe present invention relates to a recording medium, and more particularly to post-processing of recognition results after character recognition processing results.
[0002]
[Prior art]
In recent years, optical character reading (OCR) software for identifying a character image in document image data taken into a computer from a scanner or the like and outputting it as a character code has been very widely used. This is because information converted into electronic data as a character code is easier to reuse and can be stored and exchanged more efficiently than paper-based information or image data.
However, in character recognition, it is extremely difficult to completely eliminate misrecognition by recognizing one character unit from a character image. For this reason, as a post-processing, a character string obtained as a result of character recognition is often subjected to linguistic restrictions to select a candidate or automatically corrected by word matching, morphological analysis, or the like. .
This method can eliminate many errors, but depending on the application method, a new error may be created for error correction.
In addition, since the correction is not 100% correction, a recognition result having a high certainty factor is adopted as compared with a recognition result obtained by another character segmentation method (for example, see Japanese Patent Laid-Open No. 9-274645). , A certainty factor is given to each character, the user is notified, and the correction is left to the user.
For example, as a post-processing method for character recognition, the technique of Japanese Patent Laid-Open No. Hei 5-40853 creates candidate words for each word of the character recognition result and calculates a certainty factor. This corrects the recognition result of words with low confidence using grammatical relations.
As the calculation of the certainty factor, for example, as in the technique of JP-A-9-134410, language processing prior to the certainty factor calculation, the notation length of the word, the part of speech, the appearance frequency, the strength of the connection with the preceding word, etc. It is calculated by combining parameters.
[0003]
[Problems to be solved by the invention]
However, even if language processing is performed using word dictionary and grammatical knowledge as post-processing of character recognition, it is expected to improve the results in grammatical errors, unrecognizable areas, grammatical special areas, etc. Can not.
Eventually, the user is presented with a character recognition result and settled on correcting it.
As a technique for helping to correct the recognition result presented to the user, for example, in Japanese Patent Laid-Open No. 5-46803, the recognition result with low probability of the recognition result is shaded, and can be further distinguished according to the character type. By displaying the identifiers at the same time, it is easy for the user to correct the recognition result for similar characters such as “ri” in katakana and “ri” in hiragana.
In order to easily correct erroneous recognition results, Japanese Patent Application Laid-Open No. 6-176189 displays a total of three candidate characters, one character before and after the character judged to be suspicious. In Japanese Patent Laid-Open No. 7-113951 and Japanese Patent No. 2915417, the recognition result and the original image are displayed simultaneously.
However, since the above conventional technique is a display method mainly relying on visual information, a visually impaired person or the like cannot determine which character is an uncertain recognition result.
In order to solve the above-described problems, the present invention corrects a recognition result using visual information, auditory information, tactile information, or the like to improve the efficiency of post-processing, a method thereof, and a recording The purpose is to provide a medium.
In particular, it is an object of the present invention to provide a character recognition device, a method thereof, and a recording medium that improve the work efficiency of post-processing after character recognition even for visually impaired persons.
[0004]
[Means for Solving the Problems]
  In order to achieve the above object, the present inventionAn image input unit for inputting a document image, a character recognition unit for character recognition for a character area of the image input by the image input unit, and a character recognition result for each character obtained from the character recognition unit.ConfidenceA certainty factor calculation unit for calculatingA word extraction unit that extracts a word from characters recognized by the character recognition unit, and a character that cannot obtain a predetermined certainty factor from the recognition result for each character calculated by the certainty factor calculation unit, before and after the character When there is an adjacent single character that is not two or more words, the character that cannot obtain the predetermined certainty and the single character adjacent to the character are combined and extracted as one character string. There is provided a character recognition device including a certainty factor extraction unit and a result output unit that outputs the character string extracted by the low certainty factor extraction unit in distinction from other words with respect to the recognition result by the character recognition unit.
  Also,As aboveIn a character recognition device,The word extraction unit searches for a word including a character recognized by the character recognition unit using a word dictionary held in a predetermined storage unit, and obtains the searched word and the part of speech information of the searched word. The evaluation value of the searched word is calculated based on the part-of-speech information, and the word having the highest evaluation value may be extracted from the searched words.
  Furthermore, as aboveIn the character recognition device, the result output unit includes:When a word including a character for which the predetermined certainty factor cannot be obtained is corrected, the word including the character for which the predetermined certainty factor cannot be obtained is replaced with a corrected word, and the next predetermined certainty factor is obtained. You may want to move on to correcting words that contain no letters.
[0005]
  Also,A character recognition method executed by a character recognition device, comprising: an image input step for inputting a document image; a character recognition step for recognizing a character region of an image input in the image input step; and the character recognition A certainty factor calculating step for calculating the certainty factor of the character recognition result for each character obtained from the steps, a word extracting step for extracting a word from the characters recognized by the character recognition step, and the respective confidence factor calculating steps. If there is an adjacent single character that is not a word of two or more characters before and after the character for which the predetermined certainty cannot be obtained from the recognition result for the character, the predetermined certainty cannot be obtained. A low certainty factor extraction step for extracting a character and a single character adjacent to the character as a single character string, and a recognition result obtained by the character recognition step are extracted by the low certainty factor extraction step. And to provide a character recognition method comprising the result output step of outputting the character string to distinguish from other words.
Further, in the character recognition method as described above, the word extraction step searches for a word including the character recognized by the character recognition step using a word dictionary held in a predetermined storage unit, and the search is performed. Extracting a word and part of speech information of the searched word, calculating an evaluation value of the searched word based on the part of speech information, and extracting a word having the highest evaluation value from the searched word It should be included.
Also provided is a program for causing a computer to execute each step of the character recognition method as described above.
Furthermore, a computer-readable recording medium recording the program as described above is also provided.
[0006]
DETAILED DESCRIPTION OF THE INVENTION
The configuration and operation of the embodiment of the present invention will be described below in detail with reference to the drawings.
<Example>
(1) Configuration of the embodiment
FIG. 1 is a block diagram showing a configuration of a character recognition apparatus according to an embodiment of the present invention.
A character recognition device 100 according to an embodiment of the present invention includes a control unit 10, an image input unit 20, a character recognition unit 30, a certainty factor calculation unit 40, a result output unit 50, an original image storage unit 25, a recognition dictionary 35, and a recognition result storage. It consists of a unit 45 and a language dictionary 55.
Further, the result output unit 50 includes a low confidence extraction unit 60.
The control unit 10 reads the image from the scanner or file, recognizes characters from the image information, and controls the entire process from outputting the recognition result as the final result.
The image input unit 20 reads an image from a scanner or a file and stores the image data in the original image storage unit 25.
The character recognition unit 30 discriminates a character image area from the image information stored in the original image storage unit 25, cuts out a line from the character area, cuts out a character from the cut out line, and pairs of rectangles surrounding the character part. Extract the angular coordinate value, normalize the size of the character part, remove noise (dirt, etc.), calculate the feature value, perform pattern matching with the recognition dictionary 35 that holds the feature value and the standard pattern, One or a plurality of recognition candidate characters per character, their ranks, and distance values between the corresponding standard patterns are stored in the recognition result storage unit 45.
The recognition dictionary 35 holds information such as a character code for each character and a feature amount of a standard pattern of the character.
[0007]
Next, the character recognizing unit 30 generates word candidates starting from each character position, searches the word dictionary of the language dictionary 55, extracts the matched word and its part of speech information, and selects candidates from the top of the processing target area. The word path is generated by connecting the words, and at the same time, the cost of the word path is calculated using the part-of-speech connection cost table of the language dictionary 55. The word paths are selected in descending order of cost so that the number of generated word paths is below a certain number, the recognition candidate characters are corrected based on the word path with the lowest cost, and the recognition result storage unit 45 is updated. .
The language dictionary 55 includes a word dictionary that holds information such as word notation, reading, part of speech, and a part-of-speech connection cost table that holds weights indicating whether or not the parts of speech of connected words can be connected. For example, in the case of Japanese, “Kanji + Kana”, “Kanji + English”, etc. can be registered in the word notation, in addition to words of only Kanji.
The certainty factor calculation unit 40 assigns numerical values (weights) that meet the following conditions to the parameters for calculating the certainty factor, and calculates the certainty factor as a linear combination or an average value of these parameters (for example, JP-A-9-134410).
(A) The certainty level is higher as the character ranks higher in the recognition result for the same character in the image.
(B) The higher the degree of similarity in character recognition of the character, the higher the certainty level.
(C) The certainty factor increases as the notation length of the word determined to belong to the character is longer.
(D) The greater the likelihood of connection between the character or category of the character and that of the preceding and succeeding characters, the higher the certainty.
(E) The greater the possibility of connection between the word and the preceding and following words, the greater the certainty.
This rule is stored in advance in, for example, a rule table or a part-of-speech connection cost table as statistical properties or heuristic rules of many documents.
The result output unit 50 outputs the recognition result stored in the recognition result storage unit 45 in correspondence with the original image data corresponding to the recognition result for post-processing.
The user corrects the recognition result if there is an error by referring to the output result, and outputs the final result to an output device such as a display, a printer or a file, or transmits it to another computer via a network. To do.
[0008]
In this post-processing, the conventional method is intended to correct only misrecognized characters, and therefore, when inputting correct characters, one “single character” is input.
However, in the case of kanji, there are many single kanji characters with the same sound, so it is difficult to select the correct answer character. However, when considered in terms of words, the number of homophones is less than that of single kanji characters and misrecognized. In many cases, it is more efficient not to correct only the characters, but to correct them in units of words, including correct characters before and after.
On the other hand, the fact that there is a word including a character with low confidence in the recognition result calculated by the certainty calculation unit 40 often fails to match the language dictionary 55.
In such a case, the low confidence extraction unit 60 first searches for a character with a low recognition result in the word path, and is registered as a word of two or more characters before and after the detected recognized character. Searching for non-adjacent single characters in the word path, concatenating them as one character string, and extracting the character string as one word (hereinafter, this character string is also referred to as a word in the present invention).
For example, referring to FIG. 2, as a result of character recognition of image data “peripheral image display processing”, a character string “peripheral image display processing” is obtained, and word paths are “peripheral”, “image”, “wheat ”,“ Indication ”, and“ Process ”are considered. In addition, the certainty for each character is "90", "89", "80", "75", "70", "80", "50", "80", "99", "92" Suppose that it is calculated.
Since “wheat” with a low certainty factor of “50” is detected, when searching forward and backward from “wheat”, there is the word “image” in the forward direction and the single letter “show” in the backward direction. Since there is a word “processing”, “barley” is extracted as a word with low probability.
[0009]
Next, when performing the post-processing of the recognized result, the result output unit 50 distinguishes a word having a low certainty extracted from the low certainty extraction unit 60 as described above from a word having a high certainty. indicate.
For example, as a display method, it is possible to distinguish visually by changing a color, changing a font size, underlining or winking.
In this post-processing, when one word is corrected, the recognition result storage unit 45 is updated with the corrected word, and then the next word with a low certainty level is transferred.
Thus, the user can correct erroneous recognition efficiently in units of words without taking his hands off the keyboard.
By configuring the embodiment in this way, it is possible to display words with low confidence level in units of words and correct them in units of words without correcting them in units of characters as in the past. This makes it easier to correct words and improves the efficiency of editing.
[0010]
(2) Processing procedure
FIG. 3 is a flowchart showing the processing procedure of this embodiment.
An image from a scanner or a file is read, and the image data is stored in the original image storage unit 25 (step S100). Thus, the image input unit 20 is configured.
A character image region is determined from the image information stored in the original image storage unit 25, and a line is cut out from the character region (step S110).
A character is cut out from the cut out line, a diagonal coordinate value of a rectangle surrounding the character portion is extracted, the size of the character portion is normalized, noise (dirt, etc.) is removed, and a feature amount is calculated (step S120). ).
Pattern matching is performed from the feature dictionary and the recognition dictionary 35 that holds the standard pattern, and one or more recognition candidate characters per character, their ranks, and distance values between the corresponding standard patterns are recognized. (Step S130).
Based on the recognition candidates registered in the recognition result storage unit 45, word candidates starting from each character position are generated, the word dictionary of the language dictionary 55 is searched, the matched word and its part of speech information are extracted, and the processing target Candidate words are connected from the top of the region to generate a word path, and at the same time, the cost of the word path is calculated using the part-of-speech connection cost table of the language dictionary 55. The word paths are selected in descending order of cost so that the number of generated word paths is below a certain number, the recognition candidate characters are corrected based on the word path with the lowest cost, and the recognition result storage unit 45 is updated. (Step S140).
The character recognition unit 30 is configured from step S110 to step 140.
For the parameters for calculating the certainty factor, for example, the weights held in the rule table, the part-of-speech connection cost table, etc. are extracted, and the certainty factor is calculated as a linear combination or an average value (step S150).
Thereby, the certainty factor calculation unit 40 is configured.
In the word path generated in step S140, a character with a low recognition result is searched for, and adjacent single characters not registered as two or more words before and after the detected recognized character are entered in the word path. They are searched and connected as one character string, and the character string is extracted as one word (step S160).
Thereby, the low certainty degree extraction unit 60 is configured.
In order to perform post-processing of the recognition result, the recognition result stored in the recognition result storage unit 45 is associated with the corresponding original image data, and the word with low confidence and the word with high confidence extracted in step S160. By changing the color, changing the font size, underlining, or winking, you can visually distinguish them.
By referring to the output result, the user corrects the recognition result if there is an error (step S170).
In this post-processing, when one word is corrected, the recognition result storage unit 45 is updated with the corrected word, and then moved to the next word with a low certainty factor.
After the post-processing of all the recognition results is completed, the final result stored in the recognition result storage unit 45 is output to an output device such as a display, a printer or a file, or transmitted to another computer via a network. (Step S180).
The result output unit 50 is configured by steps S160 to S180.
[0011]
<Modification 1>
On the other hand, visually impaired people often obtain information by reading the recognition result. However, in the method of the above embodiment, it is possible to make a visual judgment depending on the certainty, but it cannot be used for a visually impaired person.
Therefore, in the first modification, in the post-processing of the result output unit 50, instead of extracting a word with a low certainty level and visually displaying this word, the sound quality of the voice information to be read is changed according to the certainty level. Or by changing the volume of the sound. When reading a word with low confidence, it does not make sense as a word, so a plurality of candidates are read out for each character, not as a word.
For example, when outputting sound in stereo format, when the confidence level is high, the right and left signals are output at the same time. The destination may be changed.
In addition, before and after reading out the recognition result, a signal sound or a message for distinguishing may be inserted.
Also, visually impaired people can not see the image, so it may not be possible to correct it to a completely correct character, but it can be guessed and corrected from the flow of front and back, etc. Will.
Therefore, after correcting a low-confidence word in post-processing, instead of directly reading the next low-confidence word, what the visually impaired person does is Make sure you know what to fix.
[0012]
Or in the case of a hearing-impaired person, you may make it read in order instead of jumping to the word of the next low certainty degree.
For healthy people, it is more efficient to jump to the next low confidence word without reading this way, so specify whether to jump between healthy and visually impaired people. You may make it use properly.
With the configuration as in the first modification, a visually impaired person reads a word with a low certainty factor in units of words, so that the user can easily correct it to a correct word, and the efficiency of editing work is improved.
Also, when outputting the recognition result in this post-processing, if visual and auditory are output together, the visually impaired and their caregiver can work together, so the editing efficiency is higher. Will improve.
Also, when visually checked, similar characters may cause omissions, so if you output both visual and auditory, you can read similar characters aloud. Thus, even a healthy person can easily find an error, and the efficiency of editing work is further improved.
[0013]
<Modification 2>
In the case of a visually impaired person, the recognition result is often obtained by Braille.
However, in the method of the above embodiment, it is possible to make a visual judgment depending on the certainty, but it cannot be used for a visually impaired person.
Therefore, in the second modification, in the post-processing of the result output unit 50, instead of extracting a word with a low certainty factor and visually displaying the word, a method that can be tactilely determined by the braille output device For example, the height of the unevenness of the Braille is changed, or a mark or the like for distinguishing before and after a character or word with low confidence is inserted.
Also, when correcting low confidence words in post-processing, the jump to this low confidence word is a braille cursor line (a braille pin appears) around or around the line containing the low confidence word. As shown.
By configuring as in the second modification, a visually impaired person outputs a word with low confidence in word units as braille, so that the user can easily correct the word to a correct word, and the efficiency of editing work is improved.
In addition, when outputting the recognition result in this post-processing, if tactile and auditory are output together, it can be confirmed tactilely when a visually handicapped person hears it audibly. The efficiency of editing work is improved.
In the present embodiment, a word path is created from the candidates recognized by the character recognition unit 30 by using the language dictionary 55, and a character string (word) including a character with a low certainty factor is selected from the word path. Even if it is configured to search for an unknown word that is not registered as a word in the language dictionary 55 for the recognition result in the post-processing of the recognition result by the result output unit 50, the same effect can be obtained. Can do.
[0014]
<Example by computer>
Furthermore, the present invention is not limited only to the above-described embodiment. For example, the character recognition device 100 shown in FIG. 1 can also be realized by a computer device 200 having a hardware configuration as shown in FIG.
That is, the computer device 200 includes a keyboard, a mouse, a touch panel, a scanner, a Braille input device, and the like. The input device 1 used for inputting information, various output information, information input from the input device 1, and the like. Is displayed, or is output to a printer, a Braille output device, or the like, a CPU (Central Processing Unit) 3 that operates various programs, and the program itself is held. A memory 4 that stores information temporarily created when executed, an original image storage unit 25, a recognition dictionary 35, a recognition result storage unit 45, a language dictionary 55, and a program or program of the character recognition device of the present invention A storage device 5 that holds temporary information and the like at the time of execution and a recording medium that stores programs, data, etc. Are connected to each other by a bus 8 and a medium connection device 6 used for connecting to the network 9 and a network connection device 7 which is an interface for connecting to the network 9. Yes.
The network 9 is a transmission path for connecting the computer apparatus 200 and another computer apparatus 200, and is generally realized by a cable, and TCP / IP is used as a communication protocol. However, the transmission path is not limited to cables, and may be any of wireless, wired, and broadcast waves as long as the communication protocol between them is the same. For example, LAN (Local Area Network), WAN (Wide Area Network) Internet, analog telephone network, digital telephone network (ISDN: Integral Service Digital Network), PHS (Personal Handyphone System), mobile phone network, satellite communication network, and the like can be used.
[0015]
In such a configuration of the computer apparatus 200, each function constituting the character recognition apparatus shown in FIG. 1 is programmed and written in a recording medium such as a CD-ROM in advance, and this CD-ROM is stored in the CD of each site. The above-mentioned program is mounted on a computer device 200 equipped with a medium driving device 6 such as a ROM drive, stored in the memory 4 or the storage device 5 of each computer device 200, and executed as described above. Functions similar to those in the embodiment can be realized.
As the recording medium, a semiconductor medium (eg, ROM, IC memory card, etc.), an optical medium (eg, DVD, MO, MD, CD-R, etc.), a magnetic medium (eg, magnetic tape, flexible disk, etc.), etc. Either may be sufficient.
Further, not only the functions of the above-described embodiments are realized by executing a program loaded into the memory 4 of the computer device 200, but an operating system or the like can execute a part of actual processing or The case where all the functions are performed and the functions of the above-described embodiments are realized by the processing is also included.
When the program for realizing the above-described embodiment is a semiconductor recording medium such as a ROM, the program is loaded directly into the memory 4 and executed instead of the medium driving device 6.
[0016]
<Operation in Network Environment of the Present Invention>
FIG. 5 shows a configuration of an embodiment in which the present invention is operated by connecting to a wired or wireless communication network.
For example, a server 210 that holds a character recognition program and a terminal 220 that is used by a plurality of users are connected via the network 9.
In this case, the server 210 and the user terminal 220 are constituted by the general-purpose computer apparatus 200 shown in FIG.
The user logs in to the server 210 from the terminal 220 or inputs image data for character recognition, and requests the character recognition program of the server 210 to execute character recognition.
The character recognition program of the server 210 returns the character recognition result for the character area of the transmitted image data to the requesting terminal 220.
The user's terminal 220 outputs the recognition result or the original image data while comparing them, or performs post-processing.
This has the advantage that the latest character recognition program can always be used.
When the server 210 and the terminal 220 are connected via a wired or wireless communication network as shown in FIG. 5, a character recognition program for realizing the functions of the present invention is stored in a storage device such as a magnetic disk of the server 210. The terminal 220 can be distributed in the form of download or the like.
Furthermore, a character recognition program that implements the functions of the present invention may be provided by distribution through a medium or broadcast wave.
[0017]
【The invention's effect】
As described above, according to the present invention, it is possible to confirm the certainty of the recognition result regardless of the presence or absence of obstacles such as vision and hearing.
In addition, since a word that is not certain is extracted in units of words, it is easy to correct the word, and the work efficiency of post-processing is improved.
[Brief description of the drawings]
FIG. 1 is a block diagram illustrating a functional configuration of an embodiment.
FIG. 2 is a diagram illustrating a character recognition result of a conventional example and the present invention.
FIG. 3 is a flowchart illustrating a processing procedure according to the embodiment.
FIG. 4 is a diagram showing a computer device for operating the character recognition device of the present invention.
FIG. 5 is a diagram for explaining an operation example in a network environment according to the present invention.
[Explanation of symbols]
1 input device, 2 output device, 3 CPU, 4 memory, 5 storage device, 6 medium drive device, 7 network connection device, 8 bus, 9 network, 10 control unit, 20 image input unit, 25 original image storage unit, 30 Character recognition unit, 35 recognition dictionary, 40 confidence calculation unit, 45 recognition result storage unit, 50 result output unit, 55 language dictionary, 60 low confidence extraction unit, 100 character recognition device, 200 computer device, 210 server, 220 terminal

Claims (7)

An image input unit for inputting a document image, and a character recognizing character recognition unit for the character region of the image input by the image input unit,
A confidence factor calculation unit that calculates a confidence of the character recognition result for each character obtained from the character recognition unit,
A word extraction unit for extracting words from the characters recognized by the character recognition unit;
When there is an adjacent single character that is not two or more words before and after the character for a character for which a predetermined certainty factor cannot be obtained from the recognition result for each character calculated by the certainty factor calculation unit A low certainty factor extraction unit that extracts a character that cannot obtain the predetermined certainty factor and a single character adjacent to the character as a single character string;
A character recognition apparatus, comprising: a result output unit that outputs the character string extracted by the low certainty factor extraction unit as distinguished from other words with respect to a recognition result by the character recognition unit. The word extraction unit searches for a word including a character recognized by the character recognition unit using a word dictionary held in a predetermined storage unit, and stores the searched word and the part of speech information of the searched word. 2. The character recognition apparatus according to claim 1, wherein the character recognition device extracts the word having the highest evaluation value from the searched word by taking out the calculated word and calculating the evaluation value of the searched word based on the part of speech information. . The character recognition device according to claim 1 or 2,
The result output unit, when the word containing the predetermined confidence is not obtained character is modified, replaced by the word after correction words with the predetermined confidence is not obtained character, following the predetermined A character recognition apparatus characterized in that it moves to correction of a word including a character for which the certainty cannot be obtained . A character recognition method executed by a character recognition device, comprising: an image input step for inputting a document image; a character recognition step for character recognition with respect to a character region of an image input in the image input step; and the character recognition A certainty factor calculating step for calculating the certainty factor of the character recognition result for each character obtained from the step, a word extracting step for extracting a word from the characters recognized by the character recognition step, and the certainty factor calculating step If there is an adjacent single character that is not a word of two or more characters before and after the character for which the predetermined certainty factor cannot be obtained from the recognition result for the character, the predetermined certainty factor cannot be obtained. About the recognition result by the low reliability extraction process which extracts a character and the single character adjacent to the character connected as one character string, and the recognition result by the character recognition process, it was extracted by the low reliability extraction process Character recognition method characterized by comprising the serial strings from a result outputting step of outputting separately from other words. The word extraction step searches for a word including the character recognized by the character recognition step using a word dictionary held in a predetermined storage unit, and obtains the searched word and the part of speech information of the searched word. 5. The method according to claim 4, further comprising: extracting, calculating an evaluation value of the searched word based on the part of speech information, and extracting a word having the highest evaluation value from the searched words. Character recognition method. The program for making a computer perform each process of the character recognition method of Claim 4 or 5. A computer-readable recording medium on which the program according to claim 6 is recorded.

Images