JPS594071B2 - character recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (1) Field of Application of the Invention The present invention relates to a character recognition device that is suitable for cases where there are many character categories, such as kanji, and is equipped with a recognition result verification function using context information.

(2) Prior Art Conventionally, most of the processing of various applications in government offices, for example, has been done manually.

These application forms are usually written in a mixture of kanji and kana, and if the application processing task were to be automated, the input unit would need a recognition device for Japanese characters, including kanji.
Currently, at the laboratory level, we have successfully conducted a principle experiment of a printed kanji recognition device with reading accuracy that is satisfactory for practical use.
For example, Journal of the Institute of Electronics and Communication Engineers, Volume 58-D, No. 2, 94
(see page), considering that most of the various application forms mentioned above are relatively high-quality typeprinted documents in Japanese type.
It can be said that the environmental conditions for using the printed kanji recognition device in the above application processing work are in place. However, if a printed kanji recognition device is to be put into practical use, a fairly high level of recognition accuracy is required due to the nature of the application process.

On the other hand, considering that there are an extremely large number of kanji character types, and that even if the print quality is relatively good, it is possible that an application form of relatively poor quality may be input, the reading accuracy is quite sufficient. I can't say that. However, it is conceivable that the rate of misrecognition can be significantly reduced by testing whether or not the recognition results are correct.

Conventionally, the above idea was carried out as follows. Character recognition devices that target numbers often handle monetary amounts, so for example, the amount of each item and the total amount are written on a form, and the recognition device can write the total amount of recognition results for each item, A method is used to detect errors by comparing the total recognition results. In addition, in character recognition devices that target English characters, N-Grade
A verification method using a method called m is being considered. However, the above conventional method cannot be directly applied to a character recognition device for Chinese characters.

The reason for this is that in the case of kanji, the number of character types is 2,000 to 4,000 compared to alphanumeric characters (50 characters at most), and the storage capacity of an N-Gram table, for example, becomes enormous, making it impossible to realize it as it is. It becomes possible. (3) Purpose of the Invention Accordingly, the purpose of the present invention is to provide a means for verifying whether the reading result is correct using the contextual information of the application form as a method suitable for cases where there are many character types. , the aim is to reduce the overall misrecognition rate.

(4) General explanation of the invention In order to achieve the above purpose, the present invention utilizes a format such as an application form (generally, context information).

For example, consider an application form with the following format. (Example) Type of application Registration application Purpose of registration All transfers Cause February 2, 1978 Name of purchase and sale right holder Name Kozan Abe Office 1-1 Kunitachi City, Tokyo Ownership 1/3rd Name Otsutsujibu Address 2-2 Tachikawa-shi, Tokyo Application date March 3, 1972 or later In the above example, the key items typed on the left side of the application form (type of application, purpose of registration, etc.) are Furthermore, the column to the right of the key item, that is, the character type of the fixed item, is also limited depending on what the key item is.

Therefore, taking advantage of this limitation, i.e., contextual information,
The validity of the reading results can be verified. Note that the key item here refers to a character string in a predetermined column on a form, and in the above example, it is a column of eight characters from the beginning of each line, as described above.

Further, the fixed item is a character string in a column that has a one-to-one correspondence with a key item, and in the above example, it is a character string in a column in the same line following the key item. An outline of the principle of the device of the present invention will be explained in the first section.
This will be explained using the flowchart shown in the figure.

First, at 201 and 202, characters on a form are photoelectrically converted, cut out within a certain frame, recognized line by line, and the recognition result is output for one line in the form of a character code. The recognition unit continues the above operation until all characters on the form are recognized. Everything up to this point is the same as the conventional character recognition device. Next, the recognition result verification unit, which is the core of the present invention, extracts the character sequence of the recognition result corresponding to the key item (the character string printed in the field of a predetermined length on the left side of the l line), and In step 203, it is recognized to which key item this character sequence corresponds from among the dictionary in which key items are stored. This is called key item word recognition. Note that character recognition may be incorrect, so the word recognition method described above needs to be devised. The method will be described later. Once you know the number of the key item, you can limit the character types that can appear in the fixed items following this key item.
Specify the character type with 4.

Therefore, in the character code string resulting from character recognition, the character code corresponding to the fixed item is checked to see if it is included in the above-mentioned permissible character types in step 205. At this time, if it is found that it is not included, the result of character recognition is incorrect.
Either the characters on the form were misspelled. Therefore, in this case, the above information is output along with the recognition result. For example, reverse the sign of the character code. As a result of the test, if the character type is included in the allowable character types, it is assumed to be read correctly and the character code is output as is. Continue the above operations until there are no more characters on the form.

In order to test the reading results based on the above idea,
It is necessary to have a means for recognizing key items as words all at once, and a means for specifying character types that can appear in fixed items following the recognized key items.

Before proceeding to the description of the embodiments of the present invention, the above two means will be explained. First, a method for recognizing a character sequence resulting from recognition as a word will be explained.

Generally, in order to recognize words, it is sufficient to prepare a word dictionary (a table consisting of character code strings that make up each word) and check which dictionary item the input character sequence matches.
However, in reality, not all input character sequences are read correctly, so there may be cases where a complete match cannot be achieved with any dictionary entry. Therefore, it is necessary to perform word recognition by determining the distance or equivalently the degree of similarity (to be defined later) between the input character sequence and each dictionary item, rather than determining whether or not there is a match with the dictionary item. For example, reading ``Application date'' may result in ``Application date'', but there is clearly no dictionary entry for ``Application date''.

If the degree of similarity between a character series and a dictionary entry is the degree of similarity between each character, then in the above example, the degree of similarity between "Monkey" and "A" is required.

However, such combinations of two characters are 4,000,000 combinations when the number of characters to be read is 2,000 characters, and it is impossible to store them. Therefore, in the device of the present invention, when it becomes necessary to measure the similarity between different characters (in the above example, “A” and “Mon”), the device calculates the similarity between the corresponding standard patterns in the recognition device. and use that value. The similarity between the same characters is always 1. Here, the degree of similarity is a numerical value that takes a value from O to 1, is defined between two character patterns, is easily calculated by a dedicated calculation circuit, and is well known, so its explanation will be omitted here.

The word recognition algorithm using the above method will be explained using the flowchart shown in FIG.

First, each dictionary entry has the number Nk of characters constituting a word and a character code string Wk-{Wi(k)1i-1,2,...,
Nk}. Let K be the total number of dictionary items. In the above, k is an item number (word number) and takes a value from l to K. In addition, the character sequence (character code string) of the character recognition result input to the word recognition unit is
,...,N}. The degree of similarity between the character series S and Wk is expressed as ρk. Figure 2 shows the structure of a dictionary necessary for word recognition.

The first word 501 (number D) of the dictionary holds the number K of key items, and then the address Al where the character code string of each item is stored.
, A2, . . . , a word 502 for storing AN follows. Next follows a word that stores the character code string of each key item. For example, address A1 503 holds the length (number of characters) N1 of the characters that make up the word of item number 1, and the following NO word 50
4 stores each character code. FIG. 3 illustrates a character code string that is a target of word recognition.

The character code string is stored in the working area of memory, and N
Consists of words. In FIG. 4, word recognition is performed as follows.

First, initialization is performed in steps 101 and 102. In 103,
Determine whether the word length matches the input character sequence length,
If there is no match, the similarity ρk remains 0 and the next word is examined.

When the word lengths match, the degree of similarity ρk is calculated in steps 105 to 112. Initialization is performed in 104, and in 105 it is checked whether the i-th character code Wi(k) of the k-th item in the dictionary matches the i-th character code Si of the input character series, and if they match, , 106 adds 1 to ρk, and if they do not match, it is checked in 107 whether it was impossible to determine.

When Si=0, it indicates that it is impossible to determine, in this case, execute step 106, and when Si\0, in step 108, use the standard pattern in the recognition device to determine the standard pattern of Wi(k) and the standard pattern of Si. , and add it to ρk.
It is determined in step 109 whether the value obtained by dividing ρk by the number of characters up to that point (1) exceeds the threshold value ε, and if it does not, item k is excluded from the candidates in step 113. If the number exceeds the number, the process advances to the next character, and when the processing in steps 105 to 111 is completed for all characters, in step 112, the degree of similarity between the character series is divided by the number of characters N to normalize it. When it is detected in 115 that all dictionary items have been processed, the maximum value ρ among all the similarities {ρKlk=1, 2, ..., K} found in 116 is the maximum value ρ2. , compare the absolute threshold δ and ρ1 (117), and then test whether there is a sufficient difference between ρl and ρ2 using the relative threshold γ. If it is not sufficient, output 120 indicating that it cannot be determined.

Next, a method for specifying character types that can appear in fixed items following a key item will be explained.

In the present invention, the flag table is as shown in FIG.
In addition, a bit number conversion table is prepared as shown in FIG. When the word recognition result for the key item is k*, first, the bit number conversion table is referred to to find the bit position number b(k*) indicating which bit in the flag table is to be used. Next, extract the contents of the flag table for any character, and b
It can be seen that when the value of the (k*)th bit is l, the same character is allowed as a character type following the same key item, and when it is 0, it is not allowed. Therefore, using this result,
The recognition results can be tested as described in the explanation of the principle.

(5) Examples Hereinafter, the present invention will be explained in detail with reference to Examples.

} FIG. 7 is a block diagram of one embodiment of the apparatus of the present invention.

An example will be described below with reference to the same figure. In the same figure, 1
is a conventional character recognition device, 3 is a character observation unit for observing unknown patterns, 4 is a character recognition processing device, and 5 is a standard pattern storage device. The above portions are well known and will not be described in detail here. The output 6 of the recognition processing device 4 is the result of recognizing characters on a form line by line, and is transferred in the form of a character code string. Here, when the character code is O, it means that the character was unrecognizable. The verification processing device 10, which is the central part of the present invention, is composed of a memory 11, a similarity calculation circuit 30, and a microprocessor 20.

The circuit 30 receives two character codes from the microprocessor 20, receives two standard patterns corresponding to the same character code from 5, calculates the similarity between the same standard patterns, and calculates the resulting similarity. Return to 20. The circuit 30 is the fourth
It is used when executing the process 108 in the figure. memory 1
1 is a section 12 for storing the flag table shown in FIG. 5;
A part 13 for storing the bit number conversion table shown in FIG. 6, a part 14 for storing the key item dictionary shown in FIG.
Furthermore, it consists of a work area 15. The microprocessor 20 performs word recognition (key item recognition) using the algorithm explained in FIG. Verify the character code that is the recognition result.

Next, the flow of processing as a character recognition device will be explained. The character pattern printed on the form is photoelectrically converted by 3, cut out within a certain frame, and transferred to 4.

In 4, the degree of similarity between the unknown pattern sent from 3 and each standard pattern in 5 is calculated, and the code of the character that gives the maximum degree of similarity is compiled for one line and output as a character code string on the output line 6. Output to.

However, here, in 4, it is tested whether the maximum similarity is greater than or equal to a predetermined threshold, and if the threshold is not reached, the output code is set to 0. The microprocessor 20 in the verification processing device 10 receives the character code string of the recognition result for each line through 6 and stores it in the memory 15.

First, a character code sequence corresponding to a key item is extracted from a character sequence for one line (blanks are also given one character code), and the process moves on to word recognition. An example of a character code string for l lines is shown in FIG. The l line consists of 25 characters, the first 8
Character 801 corresponds to the key item, and the latter 17 characters 802
corresponds to a fixed item. Character code 9999 means blank. Non-blunter character code in column 801 (
In FIG. 8, S1, S2, . . . , S6) are character code strings resulting from recognition of characters in key items. Word recognition is performed by a microprogram according to the algorithm shown in FIG.

However, in the same algorithm, processing 108 in FIG.
is performed by a similarity calculation circuit. In other words, 20 is 2
character code S1 and the i-th character code Wi(k) of the k-th dictionary entry (see FIG. 4) are transferred to 30, and a similarity calculation command is issued to 30. 30 receives the same command, reads two standard patterns corresponding to Si(5wi(k) from 5, and calculates the similarity ρ* between the same standard patterns.
Calculate and send back 20.

The above is the process 203 in FIG. When the microprogram completes word recognition, it moves on to verification processing. First, processing 204 in FIG. 1 is performed. First, when the key item number as a result of key item recognition is known, the bit number conversion table in the memory 13 is checked and the bit number b* of the flag table that specifies the character type of the fixed item following the same key item is determined. get. Subsequently, a verification process 205 of the fixed item recognition results is performed. 20 extracts one character code from the character code 802 corresponding to the fixed item from the recognition result character code string (Fig. 8) in the memory 15, and writes it to each character code in the flag table (see Fig. 5) in the memory 12. Check the b*th bit of the corresponding flag.

When the same bit is l, the character type is allowed, so nothing is done at that time, but when it is 0, the character type is not allowed, so the sign of the character code in 802 that gave the same result is reversed. let For example, if the recognition result character code with a fixed item is 500, and the character is not allowed as a result of the verification, the code is inverted to -500. Here, if the character code of the fixed item recognition result is a negative sign at the stage when it is sent from 4, the verification process for the same character code is not performed. Further, as for the character code corresponding to the key item, the character code string of the dictionary of the word recognition result is substituted for the character code string of the key item shown in FIG.

For example, if the character recognition result 801 is "Application date" but the word recognition result is the key item number corresponding to [-Application date], 20 will be replaced with "Application date". Since the corresponding character code string is retrieved from the dictionary stored in the memory 14 and 801 is rewritten, even if there is an error in the character recognition result, it is corrected correctly.The result of word recognition of the key item was undecidable. In this case, the subsequent character codes cannot be verified, so all the accompanying character codes are reversed to negative values.When the verification is completed and the character code string shown in Figure 8 is rewritten (if there is no error, the result is There is essentially no change.

), 20 outputs the same character code strings 801 and 802 to the output line 50
Output on top. The above process is executed for each line of the form.

(6) Summary As explained above, the device of the present invention verifies character recognition results using context information, and if the recognition result is determined to be incorrect, it reverses the sign of the character code and also changes the key item part. Since some characters that are misrecognized are correctly corrected and output, the misrecognition rate can be reduced.

The results of this character recognition device can be displayed, for example, as shown below, and the final judgment can be made manually.

In other words, if a positive character code is output, it is displayed normally, but if a negative character code is output, there is a high possibility of misrecognition, so you may change the display brightness or color, or change the displayed character. It can be displayed with special symbols added to the side, and can be corrected manually. The features of the device of the present invention are that it does not require major changes as it can be installed after the conventional character recognition device, and that the verification processing section can be easily removed and the recognition section can operate as a conventional recognition device as is. Therefore, this verification processing section can be handled as an option. Another feature of word recognition is that the measure of the closeness between any two characters, which is required when searching a dictionary from a character code sequence containing errors, is obtained from the similarity between standard patterns. It is.

Therefore, extensive storage for storing proximity measures is unnecessary. In addition, the present invention. In the embodiment described in the specification, the similarity calculation circuit 30 is provided in 10 in FIG. 5, but since 4 originally has the similarity calculation function, by slightly changing 4, 4, making it more efficient as a whole. In the present invention, the format is such that key items and fixed items are written for each line, but if the format changes, the microprogram of the verification processing section can be changed, so the present invention uses a line-by-line verification method. Not limited to.

Further, the format is stored in the storage device 11 in the form of a table,
It is also possible to avoid changing the microprogram. Further, although the character types of fixed items are limited by key items, it is also possible to further divide the fixed items and change the character types afterward while performing word recognition.

For example, in the case of place names, it is possible to recognize prefecture names and limit the character types of city, town, and village names. Furthermore, the first
It is also possible to create a large mark on each row and dynamically change the key item dictionary and flag table based on the word recognition results, and all of these are included in the present invention. As is clear from the above description, the verification processing section, which is the core of the present invention, does not depend on the specific method of the recognition processing section, which is the preceding stage, and can be similarly adopted.

Note that a method similar to the present invention is to limit the character types such as numbers, alphabets, or katakana for each column, but the principle of the present invention is to determine which character type the column is based on as a result of key item recognition. The character type is dynamically limited.

The present invention therefore differs substantially from such conventional methods.

[Brief explanation of drawings]

FIG. 1 is a flow chart for explaining the principle of the present invention.

Claims (1)

[Claims] 1. A character recognition device that inputs an unknown character pattern on a character field set on a recording medium and performs recognition by comparing the input unknown character pattern with a standard pattern prepared in advance, a storage means for storing a word dictionary; and a memory means for storing a word dictionary; and a memory means for storing a word dictionary; The present invention includes an identification means for identifying the context of the unknown character pattern string, and a testing means for testing the recognition result of the unknown character pattern string on each character field based on a subset of the character pattern specified by the identified context. A character recognition device featuring: 2 The above matching degree is determined by comparing the character pattern code string representing the recognition result with the character code string representing the word. If the character codes corresponding to each character pattern match, a similarity of 1 is given, and if they do not match, each character 2. The character recognition device according to claim 1, wherein the character recognition device is calculated by giving a degree of similarity between character patterns corresponding to codes. 3. The character recognition device according to claim 2, wherein in the above verification, if there is no match, a degree of similarity between standard character patterns corresponding to each character code is given. 4. The character recognition device according to claim 1, wherein the unknown character pattern string existing on the context identification area includes a key item specifying the type of the corresponding character field. 5. The character recognition device according to claim 1, wherein the unknown character pattern string existing on the context identification area includes a transition rule that specifies the type of the next character field following the corresponding character field. 6. The testing means is arranged in a matrix of M rows and N columns, with the number of categories of unknown character patterns being M (M: an integer of 2 or more) and the number of subsets of character patterns being N (N: an integer of 2 or more). n (n=1)
, 2, . . . , N)-th subset of character patterns is specified by the storage contents of storage elements corresponding to each row in the n-th column of the matrix. character recognition device. 7. The character recognition device according to claim 1, wherein the identification means includes means for changing the word dictionary according to the identification result of the context. 8. The character recognition device according to claim 6, wherein the identification means includes means for changing the stored content of the storage element based on the identification result of the context.