JPH096923A - Character recognition device - Google Patents

Abstract

PURPOSE: To increase a processing speed for a compound word and lower a miscorrection rate. CONSTITUTION: A front stage 11 outputs plural recognition candidate characters and recognition precision to an output buffer 4. Then a two-word character extraction part 5 retrieves the candidate character of 1st order and wen the character is KANJI (Chinese character), the KANJI and the pointer of one character following it are stored in a pointer storage area. A word matching part 6 checks the continuity of the stored pointer to take out a two-character word consisting of two characters having successive KANJI at the head and when the recognition precision of the candidate character of 1st order of the taken-out two-character word is not in the top rank as to the two characters, a character similar to the candidate character of 1st order for each character is added to the candidate characters from a similar character dictionary 8 on condition that the candidate character of 1st order for the character. The continuity of a word dictionary 7 consisting of a set of two KANJI characters or one KANJI character and one Japanese syllabary added to a KANJI, and the pointer is checked to collate the combination of respective candidates for the taken-out two-character word, and the collation result is written in the output buffer 4.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device,
In particular, the present invention relates to a character recognition device and a word automatic correction method for automatically correcting a word after character recognition in character recognition processing.

[0002]

2. Description of the Related Art Conventionally, character recognition processing in a character recognition apparatus, that is, image data (image data) obtained by electrically converting manuscript data that has been optically read is character-recognized using a recognition dictionary and a recognition result is obtained. At the stage of processing to obtain the character code corresponding to
As disclosed in Japanese Patent Laid-Open No. 59-106083 (name “word reading device”), a word (a compound word composed of two or more characters) including an external character (meaning a character other than the recognition target character that is not included in the recognition dictionary) , Which is the same in the following), and a similar character table in which the external characters are similar in character shape are provided so that words including external characters can be recognized.

The technique disclosed in Japanese Patent Laid-Open No. 59-106083 mentioned above uses a similar character in the recognition process in addition to the recognition result at the time of word matching. The recognition target character and the external character not registered in the recognition dictionary are used. It is possible to increase the number of characters that can be recognized without increasing the capacity of the recognition dictionary by using the similar character table composed of the combination of and the word dictionary composed of external characters.

On the other hand, in the character recognition process, characters recognized one by one in the recognition process may potentially have erroneous recognition characters due to ambiguity of manuscript characters, erroneous characters, or a cause in recognition accuracy. . For such misrecognized characters, conventionally, the character recognition result (recognized character) and the image data are displayed in association with each other, and if the user visually inspects and there is a misrecognized character, the correct character is input from the input device. Although the work of inputting and correcting is performed, as an automatic correction method of erroneous recognition in the character recognition, as a post-stage processing after performing character recognition for each character in the recognition processing in the character recognition device. There is a correction process for a compound word of two or more characters. In a conventional character recognition device that performs the post-stage processing as described above, first, a compound word (hereinafter, referred to as a word composed of two or more characters) is used for a recognized character string at a transition of the character string or the like.
Then, the divided compound words are matched against the compound word dictionary, and the compound words obtained by word matching are automatically corrected for misrecognized characters while performing a connection check in the grammar dictionary. .

[0005]

However, in the character recognition device for performing the post-stage processing as described above, in order to automatically correct the recognized character, many characters are used by using the candidate characters of each character of the character string for word matching. It is necessary to match words in combination, and it is necessary to check whether the matched strings can be connected using a grammar dictionary and output the result. However, there is a problem that the processing time is long and the content of the dictionary must be increased in order to improve the accuracy, and the average processing time increases as the content of the dictionary increases.

Further, in recent years, the technical development of the character recognition device has remarkably improved the recognition accuracy due to the performance improvement of the character reading device. Therefore, in the conventional post-stage processing technique as described above, the recognition accuracy is improved with respect to the increase of the processing time. The degree of improvement is low,
For example, there is a problem in that, because long compound words are collated, erroneous corrections may be repeated during that time.

The present invention has been made in order to solve the above-mentioned conventional problems. In the character recognition process, a compound word consisting of two or more characters is collated using a two-character word dictionary to improve the processing speed. An object of the present invention is to provide a character recognition device capable of improving the error correction rate and reducing the error correction rate as compared with the conventional method.

[0008]

In order to achieve the above object, the character recognition device of the first invention converts image data into a character code and outputs the character code of a plurality of obtained recognition candidate characters. The first part of the recognition process, (1) the output storage part that stores the output of the first part, and (2) the character code of the first-ranked candidate character stored in the storage part is searched and if the character type is kanji A two-word character extraction unit that stores a character code of Kanji and a pointer of the character code of one character that follows,
(3) A two-character word dictionary containing a character code for two kanji characters or a character code for one kanji character and a character code for one kana kana, and (4) when the stored pointers are continuous, A character collating unit for collating the character code of the word dictionary with the character code of each candidate character referred to by the continuous pointer to obtain a correction character code.

A character recognition apparatus according to a second aspect of the present invention includes a pre-stage of a recognition process for converting image data into a character code and outputting the character codes and recognition accuracy of a plurality of obtained recognition candidate characters.
(1) An output storage unit for storing the output of the former stage unit, and (2)
A character type determination means for searching the character code of the first-ranked candidate character stored in the storage unit to determine the character type, and a character code of the Chinese character and a pointer for the subsequent one character code when the character type is Kanji. A two-word character extraction unit having a pointer storage unit for storing, and (3) a two-character word including a character code of two Kanji characters or a combination of a character code of one Kanji character and a character code of one kana. A dictionary, and (4) continuity determination means for checking the continuity of the stored pointers and extracting the character code of each candidate character and the recognition accuracy of the first-ranked candidate character for two characters beginning with a continuous Kanji character; Recognition accuracy rank determination means for determining the rank of the recognition accuracy of the first-ranked candidate characters for two characters that have been issued, and the word dictionary and the sentence of each candidate character read when the recognition accuracy is not the highest rank for both characters. A word dictionary matching unit that matches the code with a word, and a word matching unit that includes a word correcting unit that rewrites the corresponding character code stored in the output storage unit with the character code obtained as a result of the matching. Characterize.

A character recognition apparatus according to a third aspect of the present invention includes a front part of a recognition process for converting image data into a character code and outputting a character code and recognition accuracy of a plurality of obtained recognition candidate characters.
(1) An output storage unit for storing the output of the former stage unit, and (2)
A character type determination means for searching the character code of the first-ranked candidate character stored in the storage unit to determine the character type, and a character code of the Chinese character and a pointer for the subsequent one character code when the character type is Kanji. A two-word character extraction unit having a pointer storage unit for storing, and (3) a two-character word including a character code of two Kanji characters or a combination of a character code of one Kanji character and a character code of one kana. A dictionary, (4) a similar character dictionary composed of character codes of similar characters, and (5) a character code of each candidate character for two characters starting with a continuous kanji by checking the continuity of the stored pointers. And a continuity determination means for extracting the recognition accuracy of the first-ranked candidate character, a recognition accuracy rank determination means for judging the rank of the recognition accuracy of the read-out two-ranked first-ranked candidate characters, and the recognition accuracy of both characters is the highest. In the case of other than rank, 1
A candidate character adding means for adding a similar character obtained as a candidate character by referring to a similar character dictionary for position candidate characters,
A word dictionary collating means for collating the word dictionary and the read character code of each candidate character, and a word correcting means for rewriting the corresponding character code stored in the output storage unit with the character code obtained as a result of the collation. And a word matching unit including the word matching unit.

[0011]

With the above construction, the character recognition device of the first, second, and third inventions has a two-character word dictionary and stores two character strings.
The time required for collation is short because each character is divided into two kanji or one kanji and one kana is collated with a two-character word dictionary. Therefore, the character recognition processing speed is dramatically improved as compared with the prior art. Specifically, many words are registered in a conventional compound word dictionary, such as a two-letter word, a three-letter word, a four-letter word, and so on. 30,000 words, 3 letter words about 10,000 words,
Four-letter words are about 5,000 words, and so on. here,
For example, two-letter words are 30,000 words, three-letter words are 10,000 words, 4
The size of the word dictionary, which consists of 5,000 word words and 2,000 word words, requires a capacity of 2 bytes per character, (2 × 30000) + (3 × 10000) + (4 × 50
00) + 5 × 2000 = 120,000 characters and 120,000 × 2 = 240,000 bytes. On the other hand, since the word dictionary of the present invention is composed of only two-letter words, (2 × 30000) = 60,000 characters, (60,000 × 2) = 120,000 bytes, and the size of the word dictionary is conventionally. It will be half of the word dictionary of and the matching time will be half.

Further, since the character recognition device of the second aspect of the invention is configured to match two characters having low recognition accuracy with the word dictionary, the processing speed is faster than that of the character recognition device of the first aspect of the invention. Further, the character recognition device of the third invention is provided with a similar dictionary, and is configured to collate the similar dictionary for two first-ranked candidate characters having a low recognition accuracy rank and add the obtained result as the (n + 1) th candidate character. Since it has been done, the first
~ Even if the nth candidate character has no correct answer character, the correct answer character can be obtained.

[0013]

1 is a block diagram of a character recognition apparatus as an embodiment of the present invention, in which 1 is a reading unit, 2 is a recognition unit dictionary for a character recognition unit 3, 3 is a recognition unit, and 4 is an output buffer. ,
Reference numeral 5 is a two-word character extraction unit, 6 is a word matching unit, 7 is a word dictionary, and 8 is a similar character dictionary. The recognizing unit dictionary 2 and the recognizing unit 3 constitute a pre-stage unit 11 for performing character recognition in the recognizing process, and the two-word character extracting unit 5 to the similar character dictionary 8 are post-stage units for automatically correcting words in the recognizing process. Make up twelve.

In FIG. 1, the reading unit 1 optically reads the form data and converts it into an electric signal to obtain image data.
In the pre-stage part 11 of the recognition process, the character recognition part 3 converts the image data into a character code while referring to the recognition part dictionary 2 and ranks the recognition accuracy with a plurality of obtained recognition candidate characters (candidate character codes). The rank is output to the output buffer 4.

Next, in the latter stage part 12 of the recognition process, the two-word character extraction part 5 retrieves the first-ranked candidate character code from the output buffer 4 and judges the character type such as kanji, hiragana, katakana, numbers, symbols, etc. In the case of a Chinese character, the pointer of the Chinese character and the subsequent character (not necessarily Chinese character) is stored and stored in a pointer storage area (not shown).

The word collation unit 6 checks the continuity of the pointers stored in the pointer storage area to extract a two-character word consisting of two characters starting with a continuous Chinese character, and also the first-ranked candidate of the extracted two-character word. If the first character candidate character of each character is a Kanji character when the character recognition accuracy is not the highest rank for both characters, the similar character of the first character candidate character of each character is added to the candidate character from the similar character dictionary 8. A word dictionary 7 composed of two kanji characters or one kanji character and one kana character and a combination of each candidate character of a two-character word extracted by looking at the continuity of the pointer are word-matched, and the matching result is output in a buffer. Write to 4.

FIG. 2 is a flow chart showing an example of the operation of the latter part 12 of FIG. 1, and FIG. 3 shows a word dictionary 7 storing two-character words consisting of two kanji characters or one kanji character and one kana character. FIG. 4 is an explanatory diagram visually showing an example, and FIG. 4 is an explanatory diagram visually showing an example of the similar character dictionary 8.

In the heading column of FIG. 3, "number" means the number of characters added after "the first character", that is, a two-character word has a meaning consisting of the first kanji and one kanji following it.
When defined as a character, two characters consisting of the first kanji and one kana of kana, the "number" is a kanji or kanji which is a "subsequent character" that has a meaning when paired with the kanji of the "first character". The number of kanji registered in the word dictionary of kana.

For example, in the case of number 1 in FIG. 3, "first character"
"Letter" registered as a meaningful kanji for "Gen" is "Umi", "Seki", and "Person", and two-letter words "Genkai" and "Ken"","professional"
Is registered. In the case of number 2, the "starting character" is "present", while the "after character" is "kana", which is a syllabary, and the kanji "member", "bottom",
"..." is registered as a word.

In the heading column of FIG. 4, "number" indicates the number of similar characters, "first candidate character" represents the first candidate character of the recognition result, and "similar character" means "first candidate". "Letter" means a similar character. The “similar character” is a character other than the candidate character that has a high similarity (in character recognition) to the “first candidate character”, for example, a character obtained empirically (or as a tendency). They are registered in order of increasing tendency. For example, in the number 2 in FIG. 4, “first candidate character”
In addition to the characters output as candidate characters by the recognition unit 3 when is "Makoto", "knowledge" and "yaku" are added as 6th and 7th candidate characters as "similar characters".

FIG. 5 is an explanatory diagram of candidate characters, similar characters, output character strings and pointers in the recognition processing process when the input character string "Trend of modern Japanese character recognition device" is taken as an example. 6 is an explanatory view showing the relationship between the pointer of each character stored in the pointer storage area by the two-word character extraction unit 5 from the input character string of FIG. 5, the candidate character and the similar character (note that in FIG. 5 and FIG. 6). Only the similar characters of the character "sense" are shown as similar characters and the similar characters of other character strings are omitted.)

In FIG. 5, the number after the candidate character means the recognition accuracy. The recognition accuracy is classified into classes with high accuracy, and in the present embodiment, the classes are classified into three levels from “high accuracy” to “0, 1, 2”, and the number “3” means reject. Here, taking the character “present” in the input character string as an example, the recognition result by the character recognition unit 3 is 1
The candidate character for the rank is "present" and the precision is 1 and 2 is the candidate character for "rank"
The accuracy is 1st and 3rd. Candidate characters are "Peek" and the accuracy is "2", 4
The rank candidate character is "ring" and the precision is "2", and the rank 5 candidate character is "
The accuracy is "2" for "Ming".

Below, referring to the flow chart of FIG.
The operation of the latter stage section 12 of the recognition processing according to the present invention will be described by taking the input character string "Trend of modern Japanese character recognition device" as an example. Recognition unit 3 of the first stage 11 of the recognition process
The 1st to 5th candidate characters and the precision output from are written in the output buffer 4. The latter stage 12 starts searching the output buffer 4. In step S1, the two-word character extraction unit 5 searches only the first-ranked candidate character among the candidate characters stored in the output buffer 4, determines whether or not it is a kanji character, and determines whether it is a processing target (that is, a kanji character). In this case, the process proceeds to step S2, and in the case of a character other than Kanji, the process proceeds to step S4 (in the case of the character "gen" in FIG. 5, since it is a kanji character, the process proceeds to step S2. Character of
In the case of "no", the process moves to step S4 because it is not a kanji).

In step S2, a pointer for two characters (a pointer indicating the storage position of the character code and the recognition accuracy) is stored in the pointer storage area, and the process proceeds to step S3 (in the above example, "current" and the subsequent ones). Character "left"
The pointers "0" and "1" are stored in the pointer storage area and the process proceeds to step S3). 2 in step S3
The process skips to the character and moves to step S6 (in the above example, the pointer to be referred to next is the pointer "2" of "the first candidate character", and the process moves to step S5). In step S4, the character determined as the character type that is not recognized in step S1 is skipped by one character and the process proceeds to step S5 without storing the pointer (FIG. 5).
In the example of the character "NO" of "NO", the pointer to be referred to next is set as the pointer "3" of the first-ranked candidate character "day" and the process proceeds to step S5).

In step S5, it is determined whether or not the search for the character string is completed, and if it is completed, step S6
If not completed, the process returns to step S1 to search for the next candidate character for first rank.

The 2-word character extraction unit 5 performs the steps S1 to S1.
The process of S5 is repeated until the end of the first-ranked candidate character string, and a pair of pointers of two-character words "0, 1", "3, 4", "5" is set.
6 "," 7,8 "," 9, A "," B, C "," 11,
12 "is obtained (see FIG. 6). Next, in step S6, the word collation unit 6 searches the pointer stored in the pointer storage area from the beginning and checks the continuity with the next stored pointer to determine the character. The continuity of the columns is determined. If there is continuity, the process proceeds to step S7, and if there is no continuity, the process proceeds to step S12 (in the example of FIG. 6, pointer "0").
For the character string "current" indicated by the next pointer "1", the pointers "0" and "1" are continuous, so step S
Move to 7. On the other hand, in the case of the character string "in Japan" indicated by the pointer "1" and the pointer "3", the pointers are not continuous and the process proceeds to step S12).

In step S7, the accuracy of the two first-ranked candidate characters is checked, and the accuracy is 0 which is the highest rank accuracy for both characters.
If not, the process proceeds to step S8. If the precision of the two characters is 0, the process proceeds to step S12 (character string "
In the case of "present", the precisions of the two first-ranked candidate characters are both 1 and not 0, and thus the process proceeds to step S8. In step S8, the similar character dictionary 8 is referred to and the two 1
When the similar character of the rank candidate character is registered in the similar character dictionary 8, the similar character registered as the candidate character is added respectively (the first character candidate character of the Kanji character string “current” and “left” in FIG. 6). If similar characters are present in the similar character dictionary 8 for "present" and "left", they are added as 6 candidate characters, 7 candidate characters, ... Of the corresponding character and the process proceeds to step S9).

In step S9, word matching is performed on the word dictionary 8 for each of all candidate characters (first-ranked candidate characters to fifth-ranked candidate characters and similar characters as added candidate characters) and registered in the word dictionary 8. If it matches the one, step S
The process proceeds to step 10, and if they do not match, the process proceeds to step S12.

In step S10, the first candidate character (character code) of the output buffer 4 is determined by the word matching result (character code).
Writing for rewriting, then step S1
Move to 1. The word matching result to be written is selected with a high candidate rank. In step S11, the pointer for two characters is skipped, the pointer to be referred to next is obtained, and the process proceeds to step S13 (after referencing the word of the character string "current", in order to perform word matching of the next character string "Japan"). The pointer to be referred to is set to "3" and the process proceeds to step S13).

In step S12, the pointer is skipped by 1 and the pointer to be referred to next is obtained, and the process proceeds to step S13. In step S13, the word collation unit 6 continues from step S6 to step S12 until the search of all pointers is completed.
Is repeated. Here, in FIG. 6, the reference pointer is "
When it is "5", the processing target is "word" or "sentence", and the word matching results do not match in step S9, so the process proceeds to step S12. In step S12, the pointer is skipped by one character sentence and the next pointer "6" Since "gets", the next processing target is "sentence" or "U". If the next pointer is "8", the processing target is "acknowledgement" or "makoto", but in this case " Similar character dictionary in the candidate character of "Makoto" 8
There is "knowledge" added by, so it is correctly corrected as "recognition".

When the pointer to be referred to next is "A", the processing target is "apparatus" and "position". In this case, since the recognition accuracy is "0", it is determined in step S7. Then, the process proceeds to step S12. If the pointer to be referenced is "C", the next pointer is "1" in step S7.
It is 1 "and the pointers are not continuous, and therefore it is determined that the character strings are not continuous, and the process proceeds to step S12.
If the reference pointer is "11", the processing target is "
It becomes "work" and "direction", but "Trend" in steps S7 to S9
In step S10, "work" of the first-ranked candidate character string in the output buffer 4 is corrected (rewritten) to "trend". Since all the pointers are referenced here, the final result "about the current Japanese character recognition device" can be obtained.

The operations of steps S1 to S5 correspond to the two-word character extraction section in this embodiment, step S1 corresponds to the character type determination means, and step S2 corresponds to the pointer storage means. In addition, the above steps S6 to S13
This operation corresponds to the operation of the word matching unit in this embodiment. Step S6 is the continuity determining means, step S7 is the recognition accuracy rank determining means, step S8 is the candidate character adding means, and step S9 is the word dictionary matching. It corresponds to the means.

Here, in the conventional character recognition apparatus, comparing the case where the latter word correction processing is performed with the latter word correction processing according to the present invention, as described above in the prior art, first, the recognized character string is first recognized. Is divided into compound words at the transitions of character strings, then the divided compound words are matched with the compound word dictionary, and the compound words obtained by word matching are checked for connection with the grammar dictionary. Misrecognized characters are automatically corrected.

In the compound word dictionary used in this case, many words such as a two-letter word, a three-letter word, a four-letter word, ... Are registered, and only two-letter words are combined. Four-letter words, or three-letter words that consist of a two-letter word with a prefix or suffix (for example, "Japanese",
In some cases, the same two-letter word is actually registered in different forms, such as "encyclopedia" → "Japanese language encyclopedia", "mistake", "recognition" → "misrecognition". 2 registrations
The number of character words is about 30,000, the number of three-letter words is about 10,000, the number of four-letter words is about 5,000, and so on.

Here, for example, two-letter words are 30,000 words and three
The size of the word dictionary, which consists of 10,000 character words, 5,000 four-character words and 2,000 five-character words, is 2 per character.
Byte capacity required, (2 x 30000) + (3 x 10000) + (4 x 50
00) + 5 × 2000 = 120,000 characters and 120,000 × 2 = 240,000 bytes. On the other hand, since the word dictionary of the present invention is composed of only two-letter words, (2 × 30000) = 60,000 characters, (60,000 × 2) = 120,000 bytes, and the size of the word dictionary is conventionally. It is half of the word dictionary. When the recognition result is output up to the fifth-ranked candidate character, matching the candidate character with the two-character word dictionary in all combinations results in matching 25 times. Therefore, if the dictionary size is half, matching is performed. It takes half the time (that is, if the matching time of the word dictionary in the conventional method is 1, it becomes 0.5 in this method).

Furthermore, in the prior art, a grammar dictionary for checking the connection between compound words obtained by word matching is used. However, the present invention does not require a grammar dictionary, so that the processing speed is higher than the conventional one. It becomes 0.5 or less.

[0037]

As described above, according to the present invention, 2
Perform word matching by delimiting the character string on a character-by-character basis. At that time,
Since the recognition accuracy is determined and the word matching is performed using the word dictionary only for the two characters other than the case where the recognition accuracy of the two characters is the highest rank, the erroneously recognized characters can be corrected with high accuracy.

Further, since a character similar to the first-ranked candidate character is added as a candidate character for a character other than the case where the recognition accuracy of the two characters is the highest rank, it is possible to correct even if there is no correct character in the character recognition result. .

Furthermore, since the word dictionary of only two-letter words is used as the word dictionary, the dictionary size is 1/100 of that of the conventional one.
The processing speed can be further improved because it can be made as small as 2, and the connection check between compound words in the grammar dictionary is not required.

[Brief description of drawings]

FIG. 1 is a block diagram of a character recognition device as an embodiment of the present invention.

FIG. 2 is a flowchart showing an operation example of a latter part of the recognition process for automatically correcting a word.

FIG. 3 is an explanatory diagram visually showing an example of a word dictionary storing two-character words.

FIG. 4 is an explanatory diagram visually showing an example of a similar character dictionary.

FIG. 5 is an explanatory diagram of a candidate character, a similar character, an output character string, and a pointer in a recognition processing process when a certain input character string is used as an example.

FIG. 6 is an explanatory diagram showing a relationship between a pointer of each character stored in a pointer storage area by a two-word character extraction unit, a candidate character, and a similar character.

[Explanation of symbols]

 1 reading unit 2 recognition unit dictionary 3 recognition unit 4 output buffer 5 2 word character extraction unit 6 word matching unit 7 word dictionary 8 similar character dictionary

Claims (3)

[Claims] 1. A pre-stage part of a recognition process for converting image data into a character code and outputting a character code of a plurality of obtained recognition candidate characters; and (1) an output storage part for storing an output of the front-stage part. (2) A two-word character extraction unit that searches the character code of the first-ranked candidate character stored in the storage unit and stores the character code of the Kanji character and the pointer of the character code of the one character that follows when the character type is Kanji And (3) a two-character word dictionary containing a character code for two kanji characters or a character code for one kanji character and a character code for one kana kana,
(4) When the stored pointers are consecutive, a word collating unit that collates the character codes of the word dictionary with the character codes of the candidate characters referred to by the consecutive pointers to obtain a correcting character code. A character recognition device comprising: 2. A pre-stage portion of a recognition process for converting image data into a character code and outputting the obtained character code and recognition accuracy of a plurality of recognition candidate characters; and (1) an output storage for storing the output of the front-stage portion. Section, and (2) a character type determination means for determining the character type by searching the character code of the first-ranked candidate character stored in the storage section, and the character code of the Chinese character and the subsequent one character when the character type is Kanji. A two-word character extraction unit having a pointer storage means for storing a pointer of the character code of (3), a character code of two Kanji characters or a character code of one Kanji character and a character code of one Kana A two-character word dictionary including a set, and (4) the continuity of the stored pointers is checked to extract the character code of each candidate character and the recognition accuracy of the first-ranked candidate character for two characters starting with a continuous Chinese character. Continuity Means, a recognition accuracy rank determining means for determining the rank of the recognition accuracy of the read first-ranked candidate characters for two characters, and the word dictionary and the read-out when the recognition accuracy is not the highest rank for both characters. A word dictionary matching means for matching words with the character code of each candidate character,
A character recognition device comprising: a word matching unit including a word correction unit that rewrites a corresponding character code stored in the output storage unit with a character code obtained as a result of the matching. 3. A pre-stage of a recognition process for converting image data into a character code and outputting a character code and recognition accuracy of a plurality of obtained recognition candidate characters, and (1) an output storage for storing an output of the front-stage part. Section, and (2) a character type determination means for determining the character type by searching the character code of the first-ranked candidate character stored in the storage section, and the character code of the Chinese character and the subsequent one character when the character type is Kanji. A two-word character extraction unit having a pointer storage means for storing a pointer of the character code of (3), a character code of two Kanji characters or a character code of one Kanji character and a character code of one Kana A two-character word dictionary including a set, (4) a similar character dictionary composed of character codes of similar characters, and (5) two characters starting from a continuous Chinese character by checking the continuity of the stored pointers. Of each candidate character of Continuity determination means for extracting the character code and the recognition accuracy of the first-ranked candidate character, recognition accuracy rank determination means for judging the rank of the recognition accuracy of the read two characters of the first-ranked candidate character, and recognition accuracy of two characters When both are other than the highest rank, the candidate character adding means for referring to the similar character dictionary for the first candidate character and adding the obtained similar character as a candidate character, and the word dictionary and the characters of each read candidate character A word dictionary matching unit that matches the code with the word; and a word matching unit that includes a word correcting unit that rewrites the corresponding character code stored in the output storage unit with the character code obtained as a result of the matching. A character recognition device characterized by the above.