JP2016095662A - Information processing device and information processing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an information processing device and an information processing method for recognizing a character string in a prescribed area of a document to be processed, inferring a correct character string from a character string supposed to be erroneously recognized, and facilitating the correction of an erroneously recognized character string.SOLUTION: A character string to be processed for registration in a sorter 162 is scanned to acquire a first image, and a character string to be recognized in a necessary area is recognized. Image information corresponding to an erroneously recognized candidate character string whose degree of correctness of the recognized recognition character string is low is displayed, and a correct character string is entered, and the recognition character string is registered in the sorter in association with the corresponding correct character string. Thereafter, the document to be processed is scanned, a character string to be recognized in a prescribed area is recognized, and, referring to the sorter, a correct character string is inferred by referring to the recognition character string corresponding to the correct character for the character string to be recognized that is displayed in the document to be processed, and the inferred character string is outputted.SELECTED DRAWING: Figure 2

Description

  The present invention relates to an information processing apparatus and an information processing method capable of efficiently correcting a character string recognition result recognized by character recognition processing from acquired image information. For example, the present invention relates to an information processing apparatus registered in advance with a recognized character string. The present invention relates to an information processing apparatus and an information processing method capable of correcting a recognized character string by comparing with a character string.

  Although digital information technology is now widespread, paper manuscripts are still widely used as information transmission media. In the case of converting information written on a paper document into digital data with high accuracy, characters to be converted are manually input from a keyboard or the like. In particular, when performing a large amount of manual input, a double entry process is performed in which different input persons input the same character from the same manuscript, and if the input results are the same (if they match), the input is correct. The data of the entry result was adopted, and when the input result was different, the input character was confirmed as one of the inputs was incorrect.

  As a method for replacing entry by manual input, a method of recognizing character data in a character string portion of a paper document by an OCR (optical character reading) apparatus and converting it into data is widely adopted. High recognition accuracy can be obtained by improving the resolution of the scanner and development of image analysis technology, but there is a problem of recognition accuracy due to the inevitable misrecognition, and a large amount of character recognition by the OCR device. Processing cannot be an alternative to manual entry.

  For example, medical fee bills are being digitized, but they may still be provided on paper. Among them, it is necessary to convert the name of a printed disease and the like into data, and high accuracy is required for the conversion into data. For this reason, manual entry is performed twice, and if the entry result matches, the data is adopted, and if it does not match, the entry is adopted again and the matched data is adopted. It has become enormous.

  In order to improve this point, the character data misrecognized by the OCR device is referred to the word dictionary of correct data, checked using grammatical rules, and valid candidate character strings are generated as Japanese sentences. A method (Patent Document 1) has been proposed in which an evaluation value is calculated for a character string from the word length, the appearance frequency of the word, etc., and the candidate character string having the maximum evaluation value is used as a misread correction result.

  Prepare a similar word dictionary that collects characters that are easily misrecognized among the characters read by the OCR device. If it is determined to be misrecognized, refer to the similar dictionary to find the same misread character and replace it with the correct character. There has also been proposed a method of replacing a mark that can be easily recognized when there is a possibility of erroneous recognition when a misread character is not found or when a character that is likely to be erroneously recognized even after replacement is included (Patent Document 2).

Japanese Patent Laid-Open No. 07-028956 JP 2011-150436 A

However, in order to replace one of the manually entered entries with the OCR process by the character recognition process by the conventional OCR apparatus, a correct answer rate exceeding 90% is absolutely necessary. Although high recognition accuracy can be obtained by improving the reading resolution of the OCR device and the development of image analysis technology, the accuracy rate is often below 90% in the past.
For this reason, there has been a demand for a method of providing a correct answer rate exceeding 90% by using a character recognition process by an OCR device but estimating a more probable character from misrecognized characters.

  That is, in the OCR process, a range where one character is drawn is determined from the image, and the one character is recognized. For this reason, there are various causes of errors, such as an error in judging one character range or an error in recognizing one character. In order to increase the correct answer rate, it is necessary to provide a method of analogizing correct characters against such various misrecognitions.

  In Patent Document 1, the character data erroneously recognized by the OCR device is referred to the word dictionary of correct data. In the OCR misrecognition, “right knee joint sprain” is often misrecognized by a plurality of characters such as “right forehead W”, and a candidate character string is searched even by simply referring to the word dictionary of correct data. It is difficult. In addition, the judgment as to whether it is appropriate as a Japanese sentence does not have any meaning in the name of a disease that is just a word.

  In Patent Document 2, there is “specify a character that is easily misrecognized”, for example, “mouth” (kanji for kanji) and “ro” (katakana). The OCR misrecognizes “right knee joint sprain” as “right forehead W”. That is, the extraction of the character range fails and two characters are recognized as one character (“joint” is “鸚”), or the noise behind the character is recognized as a character (“挫” is “挫 W "). For this reason, it is difficult to increase the accuracy rate only by substitution in units of characters.

  The present invention has been made for the purpose of providing an information processing apparatus and an information processing method capable of solving the above-mentioned problems and improving the accuracy rate of a document character string read by an OCR apparatus. As one means to achieve, for example, the following configuration is provided.

  In other words, the image acquisition unit includes a display unit that displays image information, acquires a processing target document as image information, and recognizes a character by extracting a necessary character string from the image information acquired by the image acquisition unit. Correct character string acquisition means for displaying a necessary character string in the image information acquired by the image acquisition means on the display means, confirming the display, and acquiring an input character string as a correct character string, and the recognition means A classifier that pre-registers the recognized character string recognized in step 2 and the corresponding correct character string acquired by the correct character string acquisition means, and a character newly recognized by referring to the registered character string pair of the classifier An analogizing means for analogizing the correct character string of the sequence is provided.

  And, for example, a determination unit that further includes a registration master in which correct character strings are registered in advance, and determines whether or not the character string recognized by the recognition unit is a misrecognized character string based on whether or not the character string is registered in the registration master And a misrecognized character string extracting unit that extracts a recognized character string determined by the determining unit as a misrecognized character string, and a misrecognition among character strings newly recognized with reference to a registered character string pair of the classifier. An analogizing means for analogizing a correct character string from the determined character string is provided.

  Also, for example, a reading master that further includes a registration master in which correct character strings are registered in advance, determines whether or not a character string similar to the recognized character string is registered in the registration master, and reads out a similar character string. When the similar character string is not extracted by the reading means, registration of the recognized character string recognized by the recognition means to the classifier is stopped.

  Further, for example, a registration master in which a correct character string is registered in advance, a determination unit that determines whether a character string similar to the recognition character string is registered in the registration master, and the recognition in the registration master A registration canceling unit for canceling registration of the recognized character string recognized by the recognition unit in the classifier when a character string similar to the character string is not registered, and a character string newly recognized by the determination unit When a character string similar to is registered in the registration master, a reading means for reading out a similar character string, a recognition character string registered in association with the classifier, and a correct character string are referred to. A correct character string for a character string recognized as a character is analogized.

  For example, the recognition means performs character recognition using a plurality of types of character recognition methods, and the classifier includes a recognized character string recognized by the character and a corresponding correct character string acquired by the correct character string acquisition means. It is characterized by registering in advance in association.

  Further, for example, it comprises analogy character string display means for displaying the analogy character string estimated by the analogy means on the display means, and the recognition character string can be corrected with reference to the analogy character string displayed by the display means. It is characterized by.

  Also, for example, the analogy character string display means displays a plurality of analogy character strings estimated by the analogy means in order of high probability of correct answer, and displays the analogy character string until the sum of correct probability reaches a certain value. It is characterized by.

  Further, for example, it comprises a separating means for separating the recognized character string recognized by the recognizing means for each character, and the classifier includes a character group separated from the recognized character string separated by the separating means as the recognized character string. It is characterized in that it is registered in association with a correct character string.

  According to the present invention, it is possible to provide an information processing apparatus and an information processing method that can improve the accuracy rate of an original character string read by an OCR device or the like, for example, and improve the accuracy rate to 90% or more.

1 is a block diagram illustrating a basic configuration of an information processing system according to an embodiment of the present invention. FIG. It is a functional diagram for demonstrating the outline | summary of the character string recognition process of the embodiment of one invention concerning this invention. It is a figure which shows the example of the judo reduction treatment medical treatment expense supply application processed in this embodiment. It is a figure which shows the example of the injury name printed on the judo reduction medical treatment payment application used in the example of this embodiment.

It is a figure which shows the example of the character string obtained by reading the injury name printed on the judo reduction medical treatment payment application used in this embodiment as image information, and recognizing the character. It is a figure which shows the example which decomposed | disassembled the character string which recognized the character in this Embodiment into a character one by one. It is a figure which shows the example of the correct character string identified by confirming the application form reading image information read in the example of this embodiment.

It is a figure which shows the example which arranged and displayed the correct character string corresponding to the character string which decomposed | disassembled the character string recognized in this Embodiment one by one. It is a flowchart which shows the schematic process of the 2nd Embodiment based on this invention. It is a figure which shows the example of the character string determined with the misread character in the example of 2nd Embodiment. It is an example of the character string decomposed | disassembled for every character of the character string determined to be misread of 2nd Embodiment.

It is a figure which shows the example which compares the character string and character string which decomposed | disassembled the character string determined to be misread used by the example of 2nd invention one character at a time and a pair. It is a flowchart which shows the outline of the misread correction method which concerns on the embodiment of the 3rd invention concerning this invention.

It is a figure which shows the example which displayed the example of the character string which recognized the character of 2nd Embodiment, and the example of the similar character string of a recognized character string so that comparison was possible. It is a figure which shows the example of the correct answer character string entered in the example of 3rd Embodiment. It is a figure which shows the state which extracted the string which decomposed | disassembled the character string determined to be the misread character string in the character string recognized in the example of 3rd Embodiment one by one, and the character string similar to the misread character string. It is a figure which shows the example which displayed as a pair the character string which decomposed | disassembled the misread character string character by character in the example of 3rd Embodiment, and the character string determined to be a correct character string. It is a flowchart for demonstrating operation | movement of the misread correction method of the 4th Embodiment based on this invention. It is a figure which shows the example of the entry screen which concerns on the 4th example of embodiment which concerns on this invention. It is a figure which shows the example of the entry screen in each embodiment of this invention. It is a figure for demonstrating the example of an effect of the misreading correction method in each embodiment of this invention.

DESCRIPTION OF SYMBOLS 100 Central processing unit 110 Display apparatus 120 Input apparatus 130 Image reading apparatus 150 Communication apparatus 160 Misreading correction process part 162 Classification | category 163 Master 210 Recognition result database 220 Correct character string database 230 Application form database 240 Recognition area extraction part 500 Customer terminal

Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.
According to an embodiment of the present invention related to the present invention, high accuracy is required when data on the names of wounds and the like printed in a medical fee bill or the like is converted into data. Therefore, in the past, manual entry was performed twice, and if the entry result matched, the data was adopted, and if it did not match, the entry was made again and the matched data was adopted. The effort was enormous.

  As an alternative to manual entry, there is a method of recognizing character data by OCR (optical character reading) and converting it into data, but conventionally, the accuracy rate is often less than 90%. However, according to the OCR misread correction method according to an embodiment of the present invention, it is possible to estimate correct data with a correct answer rate of 90% or more for characters misread by OCR, and manually enter entries. One of them can be replaced with OCR, and the data conversion effort can be reduced.

[First Embodiment]
FIG. 1 is a block diagram showing a basic configuration of an information processing system according to an embodiment of the present invention. In FIG. 1, reference numeral 100 denotes a central processing unit that performs overall control of each component of the present system. For example, the central processing unit 100 registers read image data of paper original paper (for example, a form) read by controlling the image reading device 130 in the application form data file 230 and controls the recognition region extracting unit 240 to recognize the recognition region. A specific process, a character recognition process described in the specified recognition area, and the like are performed, and a process of temporarily registering the recognized character string in the recognition result database 210 is also performed.

  Reference numeral 110 denotes a display device that displays various types of information to be described later. An input device 120 includes a keyboard and the like, and can check the recognition result and the read image information displayed on the display device 110 and correct the recognition result as necessary. Reference numeral 130 denotes an image reading apparatus capable of reading a paper document such as various forms and recognizing characters in a specific area described as needed, and reads an image of insurance application form written on the form or paper at high speed. In addition to conversion into information, character information in a specific area can be recognized and converted into character data.

  A communication apparatus 150 can communicate with other apparatuses and systems via a communication medium. For example, various types of data communication can be performed with the clinic terminal device 500 of the clinic. Although only one clinic terminal is shown in FIG. 1, there is no limit on the number of treatment hospital terminals that can be communicated, and in fact, communication with several hundred or more medical terminals is possible. is there. Note that communication from the clinic is not limited to digital data communication, and includes, for example, sending documents by facsimile communication.

  Reference numeral 210 denotes a recognition result database for temporarily registering recognition results read and recognized by the image reading apparatus 130, and is registered so as to be readable in association with the application form data file 230.

  An error reading correction processing unit 160, which will be described in detail later, includes a classification kit 162, a master 163, and the like, and corrects the recognition result registered in the correct character string database 220 as necessary using the classifier 162, for example. Correct by referring to the character string. Reference numeral 220 denotes a correct character string database in which a character string such as a name of a disease to be described in the application form is registered in advance as a correct character string.

An application form database 230 registers image information of the application form and character-coded data of the description.
A schematic function of the present embodiment having the above configuration will be described with reference to FIG. In the present embodiment, the correct character string 145, which is the correct character string of the character string to be recognized in advance, and the recognized character string recognized by the recognition process read from the paper manuscript before performing the actual character string recognition process Are registered in the classifier 162 in association with each other.

  In the misread correction method (S101) of FIG. 2, first, the left classifier generation process S102 is performed. First, the first scan process S111 for reading the paper document to be processed and generating the first image information 141 is executed. Specifically, a paper original sheet on which a character string to be recognized is displayed is set in the image reading device 130. Then, an instruction from the input device 120 is given to scan the paper original paper set on the image reading device 130 under the control of the central processing unit 100 to acquire image information of the paper original paper and register it in the application form data file 230.

Subsequently, the process proceeds to a first character string extraction process in step S112, where a character string area to be recognized in advance in the read document (application form) is extracted, and a first extracted character string 142 is extracted. . In the first character string recognition process in the next step S113, character recognition processing is performed on the first extracted character string 142 extracted in step S112, and a first recognized character string 143 is obtained as a recognition result.
In the first character string decomposition process in the next step S114, the character string decomposition process S114 is performed on the first recognized character string 143 recognized in step 113, and the first recognized character group 144 is obtained.

Further, since the paper document image information read in the first scan process (S111) is displayed on the display device 110, for example, the first entry process in step S115 is performed, and the area corresponding to the extracted character string 142 in step S112 is displayed. The correct character string 145 displayed in the area is instructed and inputted from the input device 120, for example, while confirming the read image information. Since this process only needs to be performed once, the entry process is not performed when all the character strings previously written in the application form have been entered.
The correct character string 145 represented in the same area input in this way and the recognized character group 144 are associated with each other and registered in the classifier 162.

  When the correct character string 145 and the first recognized character group 144 are registered in the pair classifier 162 for all the character strings to be processed, preparation for performing a specific recognition process is completed. The correct analogy process in step S103 is executed.

  First, in step S131, a paper original sheet to be actually processed is set in the image reading device 130 and image reading processing is performed, and the read image information is subjected to second scanning processing in which it is registered in the application form data file 230. Image information 151 is obtained. Subsequently, in step S132, a second character string extraction process for extracting an image of a recognition target character string (second extracted character string 152) is performed by cutting out an image of an area in the paper original sheet that needs to be recognized. In step S133, character recognition processing is performed on information including the character string 152 extracted to generate a second recognized character string 153. In the second decomposing process in the next step S134, the character string decomposing process S134 is performed on the second recognized character string 153 recognized in step 133 to obtain the second recognized character group 154.

  Next, in step S135, correct analogy processing is performed, and an analogy character string 155 that is analogized with the correct character string by referring to the correct character string 145 and the first recognized character group 144 registered in the separator 162 registered earlier is analogized. To do. Alternatively, a plurality of analogized character strings are output together with their correct answer probabilities (165).

  Details of the embodiment described above will be described below. In the following description, an example will be described in which a judo reduction treatment medical treatment payment application is processed as a paper manuscript sheet to be processed. FIG. 3 shows an example of an application form for judo reduction treatment medical treatment paid in the present embodiment. FIG. 3 is a diagram showing an example of a judo reduction treatment medical treatment payment application processed in the present embodiment.

The character string to be extracted is an injury name indicated by 402 in FIG. 3, and the display character string in the injury name display area is recognized and converted into digital data. For example, in the judo reduction treatment medical expenses supply application 401 shown in FIG. 3, the injury name description column 402 is printed by a printer.
FIG. 4 shows an injured name printing example. The names of wounds are printed as shown by 501 to 503 in FIG. The character string may be a handwritten character string instead of printing. However, in the case of handwriting, since the correct rate of character recognition is lowered by the writer, a printed character string is desirable.

  In actual processing, when a plurality of judo reduction treatment and medical expenses payment application forms to be processed are read from the image reading device 130 and scanned, the classifier generation process 102 has a plurality of judo reduction treatment and medical expenses payment application forms 401. Are acquired as image information in the first scan S111, and the first images 141 corresponding to the number of read images are obtained.

  Next, in the first character extraction process of step S112 for the plurality of read image information (first image 141), for example, an injury name column 402 that is an area of the character string to be extracted is extracted, and the first extracted character is extracted. A column 142 is obtained. Next, a character recognition process is performed by the character recognition process (OCR process) of step S113, and the 1st recognition character string 143 is obtained. The injury name 402 of the judo reduction treatment medical expenses supply application 401 is printed as shown in FIG. 4, for example, and the first extracted character string 142 is recognized as shown in FIG. 5, for example.

  In this state, in the first recognized character string decomposing process S114, the first recognized character string 143 is decomposed into characters, and a first recognized character group 144 is created. For example, the recognized character string examples 601, 602, and 603 in FIG. 5 are decomposed into individual characters as in the first recognized character group examples 701, 702, and 703 shown in FIG. 6.

  In the first entry process, the central processing unit 100 registers the image information read in the first scan process on the display device 110 in the application form data file 230 and also corresponds to the first recognized character string 143 of the first image 141. The operator confirms this display and prompts the user to input the correct character string 145 in the first entry process S115.

For example, by highlighting a character string portion to be input, an object to be input by the operator can be visually confirmed. The operator who has confirmed the character string sequentially inputs the correct character string 145 while visually confirming the read image that is highlighted.
For example, the correct character string examples 801, 802, and 803 in FIG. In addition, the accuracy of data may be improved by performing this entry twice and matching the result.

  The correct character string 145 thus obtained and the first recognized character group 144 are registered in the classifier 162 in association with each other. In the classifier generation process S116 of the present embodiment, a pair 146 of the first recognized character group 144 and the correct character string 145 is created. For example, a specific example is a pair 901, 902, 903 shown in FIG. In the classifier generation process S116, a large number of pairs 146 of first recognized character groups and correct character strings are collected, and after the necessary number is collected, the classifier 162 is generated by, for example, a machine learning algorithm.

  There are various types of machine learning algorithms such as naive Bayes and support vector machines, which can also be adopted in this embodiment. The details of the machine learning algorithm will be described using Naive Bayes as an example. Naive Bayes determines, as a probability, which class has a higher probability when a certain attribute is given between two classes. In this example, the attribute is a recognized character group, and the class is a correct character string. One classifier is created between two classes. In the present embodiment, there are about 400 correct character strings as injury names, and the number of generated classifiers is the number of combinations, for example, about 400 × 399/2 = 79,800 are generated.

  As an example of the classifier 162, a case where a classifier that compares a pair of the right knee joint sprain 902 and the left knee joint sprain 903 is generated will be described as an example. From all the pairs of the correct character string 145 and the first recognized character group 144, a pair in which the right knee joint sprain 902 or the left knee joint sprain 903 enters the correct character string is extracted, and the probability is calculated from the number of appearances of the first recognized character group. calculate.

  For example, if the total number of pairs is 100 and the number of appearances of the third character heel of the right forehead 鸚 挫 W906 is 1, three characters in the classifier of the right knee joint sprain 902 or the left knee joint sprain 903 When an eyelid appears as an attribute to the eye, the probability of being a right knee joint sprain 902 is 1/100, and the probability of being a left knee joint sprain 903 is zero.

  Thus, the appearance probability for each attribute is calculated, and the result of multiplication is the total probability. In this case, if the attribute has an appearance probability of zero, the total becomes zero. In order to avoid such a phenomenon, 1 is added to the numerator and denominator of the appearance probability calculation, and the log function is multiplied to convert the multiplication into the addition.

  When a new recognized character group is given, there are 79,800 separators, for example, and a class with a high probability that a new recognized character group appears, that is, an analogy character string that seems to be correct is selected, and the most frequently selected one is selected. The analogy character string is assumed to be correct.

  There is a method called Bag of Words as a method other than decomposing the recognition character string into characters. Bag of Words first creates a set that uniquely collects all OCR recognition characters. This is a method of creating a table by making a table that is True if the characters of the recognized character group are included in the set and False otherwise. When Bag of Words is used instead of for each character in the classifier generation processing of the present embodiment, Bag of Words is also used in the correct analogy processing described later.

  When the classifier is set in the machine learning algorithm described above and a recognition character group is given, the correct character string is inferred and its probability is returned. In order to improve the accuracy of the classifier, it is necessary to collect a large number of pairs of the first recognized character group and the correct character string. Since the judo rehabilitation medical treatment payment application 401 is tabulated on a monthly basis, the judo rehabilitation medical treatment payment application 401 is read at least for one month, and the judo reduction medical treatment payment application 401 for one month is read. A classifier 162 is generated from the image.

  Furthermore, in the misreading correction for the next month, the classifier generation process S102 is performed using the image, the recognized character string, and the correct character string for the current month. Since the image, the recognized character string, and the correct character string for this month are already present, the first scan S111, the first entry S115, the first character string extraction S112, and the first character string recognition S113 are not necessary. It is not necessary to perform the operation once, and it can be operated continuously and smoothly with high accuracy.

  The machine learning algorithm includes a support vector machine (SVM) and a neuron, but is not limited to these as long as it is a machine learning algorithm that can generate a classifier.

Next, the details of the correct analogy processing unit 103 will be described. A plurality of new judo reduction treatment medical application payment application forms to be processed are set in the image reading device 130, and an image is read in the second scan processing S131 to obtain a plurality of second images 151.
Since the second character string extraction process S132 for extracting the second extracted character string 152 from the second image 151 is the same as the first character string extraction process in step S112, detailed description thereof is omitted.
The extracted second extracted character string 152 is recognized in the second character string recognition process S133 similar to the first recognized character string recognition process in step S113, and the second recognized character 153 is similar to the first character string decomposition process. A second character string decomposition process 134 divides the character string into a second recognized character group 154.

In the correct analogy process in step S135, the classifier 162 is set to the previous machine learning, and the second recognized character group 154 is given to the machine learning, thereby obtaining the analogy character string 155.
When the second recognized character group 154 is right, forehead, 鸚, 挫, W, 1102, when the right, forehead, 鸚, 挫, W, 1102 are input as attributes to the 79,800 classifiers, From the classifier of the right knee joint sprain 902 or the left knee joint sprain 903, the probability of the right knee joint sprain 902 returns high, and the right knee joint sprain 902 is selected as an analogy character string candidate. Similarly, when all the classifiers are tried, the right knee joint sprain 902 is selected as the analogy character string candidate most frequently and is selected as the analogy character string 155.

Alternatively, the correct analogy 135 may instruct machine learning for one second recognized character group 154 to obtain a plurality of analogy strings and their probabilities 165.
In this case, a specific example of a learning effect and a specific example of a method of using a plurality of analogized character strings specifically estimated for the recognized character group 154 and their probabilities 165 are shown below.
The printed character string may be “Right Knee Sprain” and the character recognition result may be misrecognized as “Notice Sprain”. The printed character string may be “Left Knee Sprain” May be misrecognized as a “joint sprain”. Of course, as the probability of the classifier 162, “right knee joint sprain” is high, but the probability of “left knee joint sprain” is also a certain probability.
When the analogy result is used for the entry, “right knee joint sprain” is displayed as the first candidate. However, when “left knee joint sprain” is also displayed as the second candidate, the efficiency of the entry increases.

[Second Embodiment]
In the above description, for example, a standard paper document or the like is scanned in advance, the classifier 162 is generated from all recognized character strings, and then the paper document to be processed is scanned and the classifier 162 previously generated is used. An example of analogizing the correct character string has been described.

  However, the present invention is not limited to the above example. A master 163 in which all injury names are registered in advance is generated, and whether or not the same injury name as the recognized character string is registered is registered. For example, the control may be performed so that it is determined as a misread. It is also conceivable that the classifier 162 is generated from a misread character group and a correct character string pair and only the misread character string is used for analogy. In this case, high accuracy can be expected.

A second embodiment according to the present invention configured as described above will be described below.
FIG. 9 is a flowchart showing a schematic process of the second embodiment according to the present invention. In FIG. 9, the same step numbers are assigned to the same processes as those shown in the flowchart of FIG. 2 described above, and detailed description thereof is omitted.

  First scan processing step S111, first character string extraction processing step S112, first character string recognition step S113, first character string decomposition step S114, first entry processing step S115, second scan processing step S131, second character string The extraction processing step S132, the second character string recognition step S133, and the second character string decomposition step S134 are the same as the processing of the first embodiment shown in FIG.

  All injured names are registered in advance in the master 163 used in the second embodiment, and there are about 400 types. When, for example, the first recognized character string 143 shown in FIG. 5 extracted in the first character string recognition process in step S112 is extracted, the process proceeds to the first master comparison process shown in step S213, and the first recognized character string 143 is sequentially mastered. Reference is made to 163 to compare whether or not the character string is registered in the master 163.

  Then, a character string that is not registered in the first recognized character string is selected as the first misread character string 243. For example, when the character string of FIG. 5 is recognized, “right forehead W” 1002 of the misread character string example of FIG. 10 is not included in the injury name registered in the master 163 and is determined as a misread character string. To do.

  If the character string registered in the master 163 is included in the first extracted character string 142, it is determined that the entry in the first entry processing step S116 is not necessary and is a correct character string. Good.

  Subsequent to the first master comparison process S213, in a first misread character string decomposition process S114, the first misread character string 243 is decomposed into characters to create a first misread character group 144. For example, the “right forehead W” 1001 of the misread character group example shown in FIG. 10 is decomposed into characters one by one as in the first misread character group example 1102 shown in FIG.

  Next, in the classifier generation processing step S216, a first misread character group and correct character string pair 246, which is a pair of the first misread character group 144 and the correct character string 145, is created. For example, in the second embodiment, the pair 246 of the first misread character group and the correct character string is a pair character string indicated by 1202 shown in FIG.

  In the second embodiment, the classifier generation 216 collects a number of first misread character group / correct character string pairs 246 and then creates the classifier 222 by a machine learning algorithm.

Next, the correct analog inference unit processing step S203 shown on the right side of FIG. 9 will be described. A new plurality of judo reduction treatment medical application payment applications are read in the same process as in the second scan processing step S131 of FIG. 2, and a predetermined number of second images 151 are obtained.
Next, the second extracted character string 152 is obtained by the second character string extraction process in step S132 similar to FIG. 2 for the plurality of second images 151 obtained.
Next, a second misread character string 153 is obtained by the second character string recognition process in step S133 similar to FIG. 2 for the plurality of second extracted character strings 152 obtained.

  In the subsequent second master comparison process in step S233, the master 163 is referred to for the second misread character string 153, and when the second recognized character string 153 not registered in the master 163 is detected, the detected character string Is taken out as a second misread character string 253.

Next, in the second misread character string disassembling process in step S134, the second misread character string 253 is decomposed into characters one by one to create a second misread character group 154.
Next, in the correct answer analogy process in step S235, the classifier 162 is set to the previous machine learning algorithm, and the second misread character group 154 is given to the machine learning to obtain the analogy character string 255.

  As in the embodiment of the first invention, in the correct analogy process S235, one second misread character group 253 is instructed to machine learning, and a plurality of analogy character strings and their probabilities 265 are obtained. You may do it.

  According to the embodiment of the second invention, a recognized character string with a high accuracy rate can be obtained, and an injury name is registered in the master 163, so it is determined whether or not the recognized character string is correct. In this case, the target can be easily narrowed down.

[Third Embodiment]
Next, a third embodiment according to the present invention will be described. An outline of the character string recognition misreading correction method according to the third embodiment will be described with reference to FIG. In FIG. 13, the same steps as those shown in FIG. 2 and the processing shown in FIG.
The third embodiment shown in FIG. 13 is an improvement of the misreading correction method shown in FIG. 2, and includes a first scan process S111, a first character string extraction process S112, a first character string recognition process S113, The one entry process S115, the first character string decomposition S114, the second scan process S131, the second character string extraction process S132, the second character string recognition process S133, and the second character string decomposition S134 are the same as the processes shown in FIG. .
Also in the third embodiment, all injured names are registered in the master 163 as in the second embodiment.

  In FIG. 13, in the first similar character string search process in step S <b> 313, a plurality of similar character strings 343 corresponding to the first recognized character string 144 are extracted from the master 163. For example, a case where the first recognition character string 144 is a soda bruise 1301 shown in FIG. 14 will be described as an example.

  In this case, in the first similar character string search process S313, the back contusion 1311, the buttock contusion 1312, the buttocks contusion 1313, and the lumbar contusion 1314 shown in FIG. 16 are searched and extracted as a plurality of similar character strings 343 similar to each other. To do. For example, the Levenshtein edit distance is used as the character similarity.

  If the correct character string 145 obtained from the first entry process S115 is the back contusion 1401 shown in FIG. 15, in the first interruption determination process S317, the back contusion 1401 that is the correct character string is a plurality of similar character strings 343. Since it corresponds to the back contusion 1311 in the inside, the process proceeds to the classifier generation process S318.

  On the other hand, if the correct character string 145 obtained from the first entry process S115 is the thigh sprain 1411 shown in FIG. 15, a plurality of thigh sprains 1411 are not included in the first interruption determination process S319. Since there is no matching character string in the similar character string 343, the process is interrupted without proceeding to the classifier generation process S318.

  In the case where the first recognized character string 144 is a soda bruise 1301 and the correct character string 145 is a thigh sprain 1411, the similarity is low. Generating a classifier from such data only increases noise and increases the amount of calculation, but does not improve accuracy, so the misclassification correction processing classifier generation process is interrupted.

  Alternatively, in the first character string extraction process S112, in the case where the position detection of the injury name in the judo reduction treatment medical treatment payment application 401 has failed, the injured date on the right is read and the first extracted character is read. In the case where the column 142 is “23 · 9 · 5” 1302 in FIG. 14, the similar character string cannot be searched in the first similar character string search process S313, and a similar character string may be returned. Since it cannot be performed, in the first interruption determination process S319, the process does not proceed to the classifier generation process S318, and the registration process to the classifier is interrupted.

  When the first extracted character string 142 is “23 · 9 · 5” 1302, the similarity is low regardless of the correct character string 145. Generating a classifier from such data only increases noise and increases the amount of calculation, but does not improve accuracy. In this case, the classifier generation process is interrupted.

  Next, for example, if the first extracted character string 142 is a soda bruise 1301 shown in FIG. 14, the first character string disassembling process S144 performs the soda, part, flaw, scratch,... ,, 1501 to generate a first recognized character group 144.

  Next, the classifier generation processing S318 creates a first recognized character group and correct character string pair 146 that is a pair of correct character strings 145 associated with the first recognized character group 144. For example, as shown in FIG. A large number of these data are collected and a classifier 162 is created by a machine learning algorithm.

  A classifier for machine learning typified by a support vector machine (SVM) is composed of a large number of classifiers that classify one correct character string or another correct character string. The data of Cao, part, contusion, wound, ..., back contusion 1601 is also used to generate a classifier called back contusion vs thigh contusion.

  However, in consideration of a back contusion 1311, a buttocks contusion 1312, a buttocks contusion 1313, and a lumbar contusion 1314, which are a plurality of similar character strings 343 of the soda contusion 1301, a soda, part, contusion, wound,. By using the data of the back contusion 1601 only for the generation of the classifier 162 of the back contusion vs the buttock contusion, the back contusion vs the buttock contusion, the back contusion vs the lumbar contusion, the amount of calculation for generating the classifier 162 is reduced and the accuracy is improved. It can also be raised.

Next, the correct analogy process S303 shown on the right side of FIG. 13 will be described. A plurality of new judo reduction treatment medical expenses application forms are scanned in the second scan process S131 to obtain a plurality of second images 151.
Next, in the second character string extraction process S132, the second extracted character string 152 is obtained for the obtained plurality of second images 151.
Next, the second recognized character string 153 is obtained in the second character string recognition process S133 for the plurality of second extracted character strings 152 obtained.

In the second similar character string search S333, a plurality of similar character strings 353 are extracted from the master 163 for the second recognized character string 153.
When a character string similar to the second recognized character string 153 is not registered in the master 163 and cannot be searched, the correct answer analogy process is interrupted in the determination in the second interruption determination process S339.

In the second character string decomposition process S134, the second recognized character string 153 is decomposed to generate a second recognized character group 154.
Next, in the correct analogy processing S235, when the classifier 162 is set to the previous machine learning algorithm and the second recognized character group 154 is given to the machine learning algorithm, the analogy character string 355 can be generated.

Alternatively, in the correct analogy processing S235, it may be configured such that one second recognized character group 154 can be instructed to machine learning to obtain a plurality of analogy character strings and their probabilities 365.
In the classifier generation process S318, when a similar character string (similar character string) is not searched, the process is interrupted, and the data of the recognized character string is not used for generation of the classifier 162. For this reason, in the correct analogy process S338 as well, the correct analogy process is interrupted when a similar character string (similar character string) is not searched. However, since there is a possibility of analogy, the correct analogy process may be continued without interruption.

  In addition, by instructing machine learning to make an analogy by referring to only the classifier 162 that targets a plurality of similar character strings 353, the machine learning can narrow down the range of analogy, and the calculation time required for the analogy Can be significantly reduced.

  For example, when the plurality of similar character strings searched in the second similar character string search processing S333 are a back contusion 1311, a buttocks contusion 1312, a buttocks contusion 1313, and a lumbar contusion 1314, a back contusion vs. By calculating the probabilities only with the classifiers of buttock contusion, back contusion vs buttock contusion, back contusion vs lumbar contusion, buttock contusion vs genital contusion, buttock contusion vs lumbar contusion, gluteal contusion vs lumbar contusion It can be greatly reduced.

  Furthermore, the accuracy can be increased by adding the similarity to the probability calculation. For example, if the second similar character search S333 is performed with the similarity increased, the thigh contusion becomes a similar character as the similar character string of the soda contusion 1301. The back contusion is one letter different from the Sobe contusion, and the thigh contusion is two letters different. By reflecting this difference in similarity in the probability calculation, the accuracy can be improved.

  Furthermore, a process of replacing an apparently misread character with a correct character in the second recognized character string before the second similar character string search S333 is also effective in improving accuracy. For example, if the second recognized character string is a treated leg bruise W, the second similar character string search S333 is performed after the treated lower leg part is replaced with the right lower leg part, whereby the second similar character string The probability that the correct character string is included in the column 353 is increased, and the accuracy is improved.

[Fourth Embodiment]
Next, a fourth embodiment of the present invention will be described with reference to the flowchart of FIG. In the first to third exemplary embodiments described above, an example in which one type of classifier generation processing and one type of correct analogy processing are prepared has been described. However, the present invention is not limited to the above examples, and a plurality of types may be provided. Alternatively, each example may be repeated twice or more.

  In the fourth embodiment, two different types of image processing are performed, and image reading processing and character recognition processing are performed, respectively. These processes are referred to as character recognition processing A and character recognition processing B, respectively. In the fourth embodiment, two classifier generators and two correct analogies are also prepared.

  As a classifier generation processing unit S1702. A first character string extraction / recognition process S1711 and a classifier generation process S1712 are executed. In the correct analogy process S1703, a second character string extraction / recognition process S1713 and a correct analogy process S1714 are executed. Then, an analogy character string 1715 or a plurality of analogy character strings and their probabilities 1716 are generated.

  On the other hand, as the classifier generation processing unit S1704. A third character string extraction / recognition process S1721 and a classifier generation process S1722 are executed. In the correct analogy process S1705, a fourth character string extraction / recognition process S1723 and a correct analogy process S1724 are executed. Then, an analogy character string 1725 or a plurality of analogy character strings and their probabilities 1726 are generated.

  In the fourth embodiment, for example, one of the two classifier generation processes is controlled to perform the classifier generation process of the first embodiment, and the classifier generation process 1704 is the second embodiment. You may make it employ | adopt the process of.

  For example, different character string extraction and recognition processes have different character recognition capabilities, and the registered contents of the classifier 1621 and the classifier 1622 generated from the processes are different. Therefore, the correct analogy process S1714 and the correct analogy process S1724 select the analogy character string 1715 and the analogy character string 1725 from the misread character group based on different grounds.

  In the fourth embodiment, the analogy character strings 1715 and 1725 generated by different recognition methods are taken in by the comparison process S1731 in step S1731, and the analogy character string 1715 and the analogy character string 1725 are compared and matched. For example, the analogy character string 1732 is output, and if it does not match, it outputs that it does not match. Even if they do not match, the probability that either is correct is increased and improved.

Alternatively, even if they do not match, the processing of the third embodiment is applied, and the correct analogy 1714 and the correct analogy 1724 instruct machine learning, and a plurality of analogy strings, their probabilities 1716, and a plurality of analogy strings When the probability 1726 is obtained, the possibility of a correct answer increases by selecting an analogy character string with a high probability.
In the present embodiment, an example in which two types of character recognition processes are used has been described. However, three or more types of character recognition methods can be applied.

[Fifth Embodiment]
A fifth embodiment according to the present invention will be described with reference to FIG. In the fifth embodiment, differences from the above-described embodiments will be mainly described.
The fifth embodiment is characterized in that the analogy character string 115 obtained in each of the above embodiments is used after being obtained. For example, when the operator needs to make an entry with reference to the injury name of the second image 151, the central processing unit 100 displays the entry screen 1801 shown in FIG.

  The entry screen 1801 has an area 1802 for displaying the second image 151. The injury name entered with the injury name displayed therein is highlighted 1803, and an enlarged 1804 of the injury name is displayed on the upper right. Control to do. An injured name entry area 1806 is provided on the right side.

  The analogy character string 155 is displayed as an entry candidate character 1807 below the entry area 1806. Therefore, if the injured name expansion 1804 and the entry candidate character 1807 are equal, the operator can input the entry candidate character 1807 into the injured name entry area 1806 only by inputting the enter key, and the next entry area Automatically move the cursor. This greatly reduces the time required for entry.

[Sixth embodiment]
A sixth embodiment of the present invention will be described with reference to FIG. In the sixth embodiment, differences from the above-described embodiments will be mainly described.

  The sixth embodiment is characterized by the utilization after obtaining a plurality of analogy character strings and their probabilities 165 obtained in the above embodiments. For example, when the operator needs to make an entry with reference to the injury name of the second image 151, the central processing unit 100 displays an entry screen 1801 shown in FIG.

  In the sixth embodiment, for example, there is a feature in the effective utilization processing after obtaining the analogy character string 165 and its probability in the first embodiment. For example, when the operator looks at the correct character string with reference to the injury name of the second image 151 and makes an entry, the entry screen 1801 is displayed. The entry screen 1801 has an area 1802 for displaying the second image 151. The injury name entered with the injury name displayed therein is highlighted 1803, and an enlarged 1804 of the injury name is displayed on the upper right. An entry area 1806 for the injured name is provided.

  Specifically, for example, analogy character strings are displayed in descending order of probability as follows, so that the correct character string can be easily identified. Suppose that we got multiple analogy strings and their probabilities. For example, the first analogy string is “Right Knee Sprain” and the probability is 50%. The second analogy string is “Right Elbow Sprain” and the probability is 30%. The third item in the column is “Right shoulder sprain” and its probability is 20%. When the summed probability value is 80%, the display is stopped.

  The probability of “right knee sprain” is the highest at 50%, and the probability of “right elbow sprain” is the second highest, at 30%. When the probability is summed, it becomes 80%, and the display so far is obtained. That is, a right knee joint sprain 1901 and a right elbow sprain 1902 are displayed.

The operator selects the right knee joint sprain 1901 and right elbow joint sprain 1902 that is equal to the injured name enlargement 1104 and inputs an enter key, so that the selected analogy string is the injured name entry. Input is made in area 1806, and the cursor moves to the next entry area. This shortens the time required for entry and increases the probability that a correct analogy character string will appear.
According to the sixth embodiment, the correct character string can be reliably identified by an extremely simple operation.

  The misread correction method according to the present invention has been described above, but the misread correction method according to the present invention is not limited to the above-described embodiment, and various modifications can be made within the scope of the present invention. Needless to say.

  An example of industrial use of the present invention will be described with reference to FIG. For example, high accuracy is required when data such as a name of a sickness or the like printed in a medical fee bill or the like is converted into data. Therefore, in the conventional process 2001 before applying the misreading correction method of the present invention, the first entry process 2012 and the second entry process 2013 are manually performed by different operators, and the two entry results are compared to extract data that does not match. A different data retrieval 2014 is performed, and a third entry is performed for making an entry regarding the mismatched data.

  After applying the misreading correction method according to the present invention, the first entry 2012 is abolished, and the data output by the misreading correction method of each embodiment according to the present invention is used instead of different extracted data. As a result, the number of entries can be greatly reduced. Such data that requires high accuracy has many prescription drug names, etc., operation of performing entry twice is widely performed, and the effective range of the present invention is also wide.

Claims (10)

A display means for displaying image information;
Image acquisition means for acquiring a processing target document as image information;
Recognizing means for recognizing characters by extracting necessary character strings in the image information obtained by the image obtaining means;
Correct character string acquisition means for displaying the necessary character string in the image information acquired by the image acquisition means on the display means and acquiring the input character string as a correct character string by confirming the display;
A classifier that pre-registers the recognized character string recognized by the recognizing unit in association with the corresponding correct character string acquired by the correct character string acquiring unit;
A misreading correction method comprising analogical reasoning means for analogizing a correct character string for a newly recognized character string with reference to a recognized character string and a correct character string registered in association with the classifier. In addition, it has a registration master in which correct character strings are registered in advance.
Determination means for determining whether or not the character string recognized by the recognition means means is a misrecognized character string based on whether or not the character string is registered in the registration master;
A misrecognized character string extracting unit for extracting a recognized character string determined by the determining unit as a misrecognized character string;
A misreading correction method comprising analogical reasoning means for analogizing a correct character string from a character string determined to be misrecognized among character strings newly recognized by referring to a registered character string pair of the classifier. In addition, it has a registration master in which correct character strings are registered in advance.
Reading means for determining whether or not a character string similar to the recognized character string is registered in the registration master and reading out a similar character string;
2. The misread correction method according to claim 1, wherein when a similar character string is not registered by the reading unit, registration of the recognized character string recognized by the recognition unit to the classifier is stopped. In addition, it has a registration master in which correct character strings are registered in advance.
Determining means for determining whether or not a character string similar to the recognized character string is registered in the registration master;
A registration canceling unit for canceling registration of the recognized character string recognized by the recognition unit in the classifier when a character string similar to the recognized character string is not registered in the registration master;
Reading means for reading out a similar character string when a character string similar to a character string newly recognized by the determination means is registered in the registration master;
The method of correcting misreading according to claim 1, wherein the correct character string for the newly recognized character string is inferred by referring to the recognized character string and the correct character string registered in association with the classifier.   5. The method of correcting misreading according to claim 1, wherein the analogizing means analogizes a plurality of analogical character strings with probability. The recognition means performs character recognition using a plurality of types of character recognition methods,
6. The classification device according to any one of claims 1 to 5, wherein a recognition character string recognized as a character and a corresponding correct character string acquired by the correct character string acquisition unit are registered in advance in association with each other. The misread correction method described in 1. An analogy string display means for displaying the analogy string estimated by the analogy means on the display means;
The misread correction method according to claim 1, wherein the recognition character string can be corrected with reference to the analogy character string displayed by the display unit.   The analogy string display means displays a plurality of analogy strings estimated by the analogy means in order of high probability of correct answer, and displays the analogy string until the sum of correct probabilities reaches a certain value. The misread correction method according to claim 7.   The analogy string display means displays a plurality of analogy strings estimated by the analogy means in order of high probability of correct answer, and displays the analogy string until the sum of correct probabilities reaches a certain value. The misread correction method according to claim 7. Separating means for separating the recognized character string recognized by the recognition means for each character;
The said classifier registers the character group which the recognition character string isolate | separated by the isolation | separation means as said recognition character string in association with the said correct character string, It registers in any one of Claim 1 thru | or 9 characterized by the above-mentioned. Misreading correction method of description.