JPH06266906A - Character recognition system - Google Patents

Abstract

PURPOSE:To decrease recognition errors and to improve character recognition rate by detecting the recognition errors by correcting characters while using knowledge processing, and reconstructing the recognition dictionary of a character recognition part. CONSTITUTION:A character recognition part 1 recognizes a document image 10 and outputs pieces of candidate characters per character. A candidate clause is calculated from a candidate character set 11 by using a word dictionary 6 and a grammar dictionary 7. A clause evaluated value arithmetic part 4 calculates the vocabular and grammatical rightness of the clause, the clause is selected by a clause selection part 5 with the evaluated value of the clause as a reference, and a corrected character string 14 is outputted. The candidate character set 11 is compared with the corrected character string 14, and an additional learning character 15 is decided by a candidate character comparison part 9. Based on the additional learning character 15, a recognition dictionary 16 of the character recognition part 1 is reconstructed.

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 for reading characters.

[0002]

2. Description of the Related Art In recent years, with the development of databases, there has been an increasing demand for a character recognition device having a high speed and a high recognition rate.

As a conventional character recognition device, for example, the character recognition device shown in FIG. 9 as disclosed in Japanese Patent Laid-Open No. 2-214990 has been proposed. The character correction unit 8 receives N candidate characters per character from the character recognition unit 1 as input. The automatic correction unit 61 compares the candidate character with the correction rule table 63 and corrects the character according to the correction rule. The correction output result of the automatic correction unit 61 is displayed to the operator, and the operator corrects the character that is erroneously recognized. Based on the information of this correction operation, the manual correction control unit 62 creates a correction rule, registers this rule in the correction rule table 63, and applies the correction rule to subsequent recognition results to automatically correct the recognition error. . As a result, based on the correction made by the operator, it is possible to perform character recognition that matches the font of the document.

[0004]

However, in the above-mentioned character recognition device, since the correction rule table is created based on the correction made by the operator, the correction rule is automatically created without human intervention. I can't.

The present invention solves such a conventional problem, and automatically reconstructs the recognition dictionary of the character recognizing unit based on the character string corrected by the knowledge processing. The purpose is to automatically recognize the font of the document and increase the character recognition rate.

[0006]

In order to achieve the above object, the present invention compares a character corrected in a character correction unit with a candidate character in a character recognition unit to detect an error-prone character in the character recognition unit. Extract. The extracted characters are sent to the character recognition unit, and by reconstructing the recognition dictionary in the character recognition unit based on these characters, the processing of the character recognition unit is adapted to the characters of the document to be recognized, and the recognition error is eliminated. To lose.

[0007]

With the above-described structure, the present invention extracts a character that is likely to be erroneous in the character recognizing unit from the information in which the character correcting unit has corrected the character, and reconstructs the recognition dictionary of the character recognizing unit based on this character. This reduces recognition errors in the character recognition unit and improves the character recognition rate.

[0008]

EXAMPLES Examples of the first invention of the present invention will be described below. FIG. 1 shows the overall configuration of the character recognition device of this embodiment. The character recognition unit 1 performs character recognition from the document image 10 using the recognition dictionary 16 and outputs a candidate character set having n candidate characters from the first candidate character to the nth candidate character for each character.

The word search unit 2 searches the word dictionary 6 to select a candidate word set 12 which is a combination of candidate characters matching a word existing in the word dictionary 6 from the combinations of the candidate character set 11. The phrase searching unit 3 refers to the grammar dictionary 7 and selects a candidate phrase set 13 that is a combination of words that can be a phrase from the candidate word set 12. The bunsetsu evaluation value calculation unit 4 calculates an evaluation value based on the lexical and grammatical correctness of the bunsetsu searched by the bunsetsu searching unit 3 based on the length and frequency of the words in the bunsetsu. The phrase selecting unit 5 selects the phrase having the largest evaluation value among the phrase candidates and outputs the corrected character string 14.

The candidate character comparison unit 9 compares the modified character string 14 with the candidate character set 11 and extracts a character whose modified character string is different from the first candidate character of the candidate character set.
5 is sent to the character recognition unit 1.

Character recognition is performed as follows in the character recognition device having the above configuration. First, the document image 1 to be recognized
The character recognition unit 1 character-recognizes 0 using the recognition dictionary 16 and outputs a candidate character set having n candidate characters from the first candidate character to the n-th candidate character per character.

Further, the word search unit 2 searches the word dictionary 6 to select a candidate word set 12 which is a combination of candidate characters matching a word existing in the word dictionary 6 from the combinations of the candidate character sets 11. Pick out. Further, the phrase searching unit 3 refers to the grammar dictionary 7 and sets the candidate word set 12
A candidate phrase set 13, which is a combination of words that can be a phrase, is selected. The lexical evaluation value is calculated based on the lexical and grammatical correctness of the bunsetsu searched by the bunsetsu searching unit 3 based on the length and frequency of the words in the bunsetsu. Based on the phrase evaluation value for the candidate phrase for which the phrase evaluation value is obtained, the phrase selection unit 5 selects the correct combination of phrases and the modified character string 14 is selected.
Is output.

The candidate character comparison unit 9 compares the corrected character string 14 with the candidate character set 11. The characters of the corrected character string at the same character position are compared with the first candidate character of the candidate character set, and if these two characters are different, the additional learning character 15 is output.

The character recognition unit 1 receives the additional learning character 15 and adds a dictionary of the additional learning character 15 to the recognition dictionary 16 so that the additional learning character can be recognized from the character image of the additional learning character and the character of the corrected character string. To do.

As a result, the character that cannot be recognized by the first character recognition in the character recognition unit 1 can be recognized by adding the dictionary of additional characters to the recognition dictionary.

The recognition dictionary 16 in the character recognition unit 1
The learning character 15 is added to the dictionary by the character recognition unit 1.
May be configured by a neural network and the weight of the network may be changed by additional learning to perform the processing.

In the present embodiment, the candidate character comparison unit 9
The corrected character string was compared with the first candidate character, but the corrected character was compared with the mth candidate character (1 ≦ m ≦ i <n), and the corrected character was included in the i candidate characters. If not, the corrected character may be output as the additional learning character 15.

Next, a second embodiment of the present invention will be described. FIG. 2 shows the overall configuration of the character recognition device of this embodiment.

The character recognition unit 1, the word search unit 2, the phrase search unit 3, the phrase evaluation calculation unit 4, the phrase selection unit 5, and the candidate character comparison unit 9 are the same as those in the first embodiment of the invention.

The same character extraction unit 21 extracts the same character from the characters output from the candidate character comparison unit, and when the character is included in different words, the character is additionally learned character 15 Output as.

Character recognition is performed as follows in the character recognition device having the above configuration. First, the document image 1 to be recognized
The character recognition unit 1 character-recognizes 0 using the recognition dictionary 16 and outputs a candidate character set having n candidate characters from the first candidate character to the n-th candidate character per character.

Further, the word search unit 2 searches the word dictionary 6 to find a candidate word set 12 which is a combination of candidate characters matching a word existing in the word dictionary 6 from among the combinations of the candidate character sets 11. Pick out. Further, the phrase searching unit 3 refers to the grammar dictionary 7 and sets the candidate word set 12
A candidate phrase set 13, which is a combination of words that can be a phrase, is selected. The lexical evaluation value is calculated based on the lexical and grammatical correctness of the bunsetsu searched by the bunsetsu searching unit 3 based on the length and frequency of the words in the bunsetsu. Based on the phrase evaluation value for the candidate phrase for which the phrase evaluation value is obtained, the phrase selection unit 5 selects the correct combination of phrases and the modified character string 14 is selected.
Is output.

The candidate character comparison unit 9 compares the corrected character string 14 with the candidate character set 11. The characters of the modified character string at the same character position are compared with the first candidate character of the candidate character set, and the characters differing from these are output.

The same character extraction unit 21 extracts the same character from the characters output by the candidate character comparison unit 9, and when the character is included in a different word, the character is additionally learned character 15 Output as. For example, the character correction unit 8
When the corrected character string is output as shown in FIG. 3, the candidate character comparison unit 9 outputs the sentence “sentence”, the recognition “recognition”, the grammar “sentence”, and the correction “correct”. With respect to these characters, the same character extraction unit outputs the “sentence” as the additional learning character 15 because the “sentence” of the sentence and the “sentence” of the grammar are the same character and are included in different words. To do.

The character recognition unit 1 receives the additional learning character 15 and adds a dictionary of the additional learning character 15 to the recognition dictionary 16 so that the additional learning character can be recognized from the character image of the additional learning character and the character of the corrected character string. To do.

As a result, the character which cannot be recognized by the first character recognition in the character recognition unit 1 can be recognized by adding the dictionary of additional characters to the recognition dictionary.

The recognition dictionary 16 in the character recognition unit 1
The learning character 15 is added to the dictionary by the character recognition unit 1.
May be configured by a neural network and the weight of the network may be changed by additional learning to perform the processing.

Further, in this embodiment, the candidate character comparison unit 9
The corrected character string was compared with the first candidate character, but the corrected character was compared with the mth candidate character (1 ≦ m ≦ i <n), and the corrected character was included in the i candidate characters. If not, the corrected character may be output as the additional learning character 15.

Next, a third embodiment of the present invention will be described. FIG. 4 shows the overall configuration of the character recognition device of this embodiment.

FIG. 5 shows the configuration of the character recognition unit 1 shown in FIG.
The character recognition unit 1 recognizes a candidate character using a neural network. The similarity calculation unit 36 calculates the similarity between each character from the character image and the weighting coefficient 37 and outputs a candidate character. The weighting factor updating unit 38 can also update the weighting factor based on the error between the candidate character and the additional learning character.

The word search unit 2, the phrase search unit 3, the phrase evaluation value calculation unit 4, and the phrase selection unit 5 are the same as those in the first embodiment of the present invention.

The keyword extracting unit 31 in FIG. 4 extracts the keywords of the document to be recognized from the corrected character string 14 output from the phrase selecting unit 5, and creates the keyword set 35. The keyword is extracted, for example, from the difference between the frequency of words in a document and the frequency of words in a general document. The keyword partial match search unit 32 performs a partial match search between the obtained keyword set 35 and the candidate character set 11. For example, if “recognition” is extracted as the keyword, “recognition *” or “* knowledge” in the correction character string 14 is extracted as a partially matching character. The candidate word addition unit 33 adds the partially matched keywords to the candidate word set. In the above-described example, the partially matching “recognition *” and “* knowledge” are added to the candidate word set 12 as “recognition”. As a result, the character not output from the character recognition unit 1 can be used for character correction.

The non-candidate character detection unit 34 detects the non-candidate character added by the candidate word addition unit 33 in the corrected character string 14 output by the phrase selection unit 5, and sets that character as the additional learning character 15. Output.

Character recognition is performed as follows in the character recognition device having the above configuration. First, the document image 1 to be recognized
In the character recognition unit 1, 0 is referred to the weighting factor 37 shown in FIG. 5, and in the similarity calculation unit 36, a candidate character set having n candidate characters from the first candidate character to the nth candidate character per character. Is output.

Further, the word search unit 2 searches the word dictionary 6 to find a candidate word set 12 which is a combination of candidate characters matching a word existing in the word dictionary 6 from the combinations of the candidate character sets 11. Pick out. Further, the phrase searching unit 3 refers to the grammar dictionary 7 and sets the candidate word set 12
A candidate phrase set 13, which is a combination of words that can be a phrase, is selected. The lexical evaluation value is calculated based on the lexical and grammatical correctness of the bunsetsu searched by the bunsetsu searching unit 3 based on the length and frequency of the words in the bunsetsu. Based on the phrase evaluation value for the candidate phrase for which the phrase evaluation value is obtained, the phrase selection unit 5 selects the correct combination of phrases and the modified character string 14 is selected.
Is output.

Next, the keyword extracting unit 31 extracts the keyword set 35 from the corrected character string 14.

The keyword partial match search section 32 performs a partial match search between the keyword set 35 and the candidate character set 12. Next, the candidate word addition unit 33 adds the words output by the keyword partial match search unit 32 to the candidate word set 13.

Again, the phrase search unit 3 and the phrase evaluation value calculation unit 4
Then, the candidate phrase is searched from the added candidate word to obtain the phrase evaluation value.

Further, the phrase selecting section 5 selects a phrase having a large evaluation value among the phrase candidates and outputs the corrected character string 14.

In the non-candidate character detection unit 34, the corrected character string 14
A character added by the candidate word addition unit 33 is detected in the table, and the character is output as the additional learning character 15.

The weighting factor updating unit 38 updates the weighting factor 37 based on the error between the additional learning character and the candidate character to perform additional learning.

As a result, a character that cannot be recognized by the first character recognition in the character recognition unit 1 can be recognized by adding a dictionary of additional characters to the recognition dictionary.

The character recognition unit 1 may use a character recognition method that does not use a neural network. As this method, for example, an average value vector of each character may be provided as a recognition dictionary, and the character may be recognized by comparing it with an image. When the recognition dictionary is used, the recognition dictionary is reconstructed based on the added characters to perform additional learning.

Next, a fourth embodiment of the present invention will be described. FIG. 6 shows the overall configuration of the character recognition device of this embodiment.

The character recognition unit 1, the word search unit 2, the phrase search unit 3, the phrase evaluation value calculation unit 4, and the phrase selection unit 5 are the first
This is the same as the embodiment of the invention.

The keyword extracting unit 31, the keyword partial match searching unit 32, and the non-candidate character detecting unit 34 are the same as those in the fourth embodiment of the present invention.

The word erroneous correction degree calculation unit 41 uses the corrected character string 1
The probability that the word corrected from 4 is an erroneous correction, that is, the word erroneous correction degree is calculated. The reject character determination unit 42 determines the reject character based on the word error correction degree output by the word error correction degree operation unit 41. The candidate word addition unit 43 adds a character that is a reject character among the keywords searched by the keyword partial match search unit 32 to the candidate word set.

Character recognition is performed as follows in the character recognition device having the above configuration. First, the document image 1 to be recognized
The character recognition unit 1 character-recognizes 0 using the recognition dictionary 16 and outputs a candidate character set having n candidate characters from the first candidate character to the n-th candidate character per character.

Further, the word search unit 2 searches the word dictionary 6 to select a candidate word set 12 which is a combination of candidate characters that match a word existing in the word dictionary 6 from the combinations of the candidate character sets 11. Pick out. Further, the phrase searching unit 3 refers to the grammar dictionary 7 and sets the candidate word set 12
A candidate phrase set 13, which is a combination of words that can be a phrase, is selected. The lexical evaluation value is calculated based on the lexical and grammatical correctness of the bunsetsu searched by the bunsetsu searching unit 3 based on the length and frequency of the words in the bunsetsu. Based on the phrase evaluation value for the candidate phrase for which the phrase evaluation value is obtained, the phrase selection unit 5 selects the correct combination of phrases and the modified character string 14 is selected.
Is output.

Next, the keyword extracting unit 31 extracts the keyword set 35 from the corrected character string 14.

The keyword partial match search unit 32 performs a partial match search between the keyword set 35 and the candidate character set 12.

Next, in the word error correction degree calculation unit 41, the length of the corrected word, the evaluation value of the character included in the word in the character recognition unit 1, the corrected character and the character recognition unit 1 of the first candidate character are detected. The word erroneous correction degree is calculated from the difference in the evaluation value of, the type of characters forming the word, and the statistical probability that the corrected word is correct. The reject character determination unit 42 determines the reject character from the corrected word and the word error correction degree of the words before and after the corrected word.

Next, the candidate word addition unit 43 compares the words output by the keyword partial match search unit 32 with the characters output by the reject character determination unit 42, and the words that match both are selected as a candidate word set. Add to 13.

Again, the phrase retrieval unit 3 and the phrase evaluation value calculation unit 4
Then, the candidate phrase is searched from the added candidate word to obtain the phrase evaluation value.

Further, the phrase selecting section 5 selects a phrase having a large evaluation value among the phrase candidates and outputs the corrected character string 14.

In the non-candidate character detection unit 34, the corrected character string 14
A character added by the candidate word addition unit 43 is detected in the table, and the character is output as the additional learning character 15.

The character recognition unit 1 receives the additional learning character 15 and adds a dictionary of the additional learning character 15 to the recognition dictionary 16 so that the additional learning character can be recognized from the character image of the additional learning character and the character of the corrected character string. To do.

As a result, the character which cannot be recognized by the first character recognition in the character recognition unit 1 can be recognized by adding the dictionary of additional characters to the recognition dictionary.

The recognition dictionary 16 in the character recognition unit 1
The learning character 15 is added to the dictionary by the character recognition unit 1.
May be configured by a neural network and the weight of the network may be changed by additional learning to perform the processing.

Next, a fifth embodiment of the present invention will be described. FIG. 7 shows the overall configuration of the character recognition device of this embodiment.

The character recognition unit 1, the word search unit 2, the phrase search unit 3, the phrase evaluation value calculation unit 4, and the phrase selection unit 5 are the first
This is the same as the embodiment of the invention.

The character type determining section 51 refers to the character type dictionary 52 from the corrected character string 14 to determine the character type of each character and outputs it.

Character recognition is performed in the character recognition device having the above-described configuration as follows. First, the document image 1 to be recognized
The character recognition unit 1 character-recognizes 0 using the recognition dictionary 16 and outputs a candidate character set having n candidate characters from the first candidate character to the n-th candidate character per character.

Further, the word search unit 2 searches the word dictionary 6 to select a candidate word set 12 which is a combination of candidate characters matching a word existing in the word dictionary 6 from among the combinations of the candidate character sets 11. Pick out. Further, the phrase searching unit 3 refers to the grammar dictionary 7 and sets the candidate word set 12
A candidate phrase set 13, which is a combination of words that can be a phrase, is selected. The lexical evaluation value is calculated based on the lexical and grammatical correctness of the bunsetsu searched by the bunsetsu searching unit 3 based on the length and frequency of the words in the bunsetsu. Based on the phrase evaluation value for the candidate phrase for which the phrase evaluation value is obtained, the phrase selection unit 5 selects the correct combination of phrases and the modified character string 14 is selected.
Is output.

The character type determining unit 51 determines the character type from the corrected character string 14 according to the character type dictionary 52. In the character type dictionary 52, for example, the characters sandwiched between the pre-numerical prefix (e.g., about, etc.) and the number particle (book, individual, etc.) are numbers, and the characters sandwiched between katakana and katakana are katakana. There is a registered rule that there is a high probability that there is a high probability that the letters sandwiched between letters and letters are letters, and the character type is determined according to this rule. FIG. 8 is an example in which the character type is determined from the corrected character string. In FIG. 8, the character string “Nura-ne Network” in the modified character string is a katakana “9G.
The character string "5" was determined to be a number. Other than that, the character type was not determined. The character whose character type has been determined is sent to the character recognition unit 1, and the character recognition unit 1 performs character recognition again by limiting the character type according to the output of the character type determination unit.

The word search unit 2 adds the characters output by the character recognition unit 1 to the candidate word set 12 and searches again for candidate words. Furthermore, the phrase search unit 3 and the phrase evaluation value calculation unit 4 search for candidate phrases from the added candidate words,
Find the phrase evaluation value.

Further, the phrase selecting section 5 selects a phrase having a large evaluation value among the phrase candidates and outputs the corrected character string 14.

As described above, when the character types are limited, the character recognition unit 1 improves the recognition rate because the character types are limited. As a result, the recognition rate also improves as a result of character correction.

In this embodiment, the character type determining unit 51 determines the character type and limits the character type. However, the character type determining unit 51 may limit the number of characters to a certain number. For example, if it is determined from the corrected character string 14 that the character string represents a city name, it is sufficient to limit the characters used in the city name.

[0070]

As is apparent from the above embodiments, by using the character recognition device having the configuration of the present invention, the additional learning character is determined by using the information of the correction result of the character correction unit,
It is possible to automatically configure a recognition dictionary that matches the font of the document to be recognized. For this reason, the recognition rate is improved in order to perform character recognition suitable for the document to be recognized, and its practical effect is great. Further, the recognition rate of the character recognition unit can be improved by limiting the character type of the recognized character in the character recognition unit from the information of the correction result of the character correction unit.

[Brief description of drawings]

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

FIG. 2 is a configuration diagram of another embodiment of the character recognition device of the present invention.

FIG. 3 is an output diagram of the same character extraction unit of the present invention.

FIG. 4 is a configuration diagram of another embodiment of the character recognition device of the present invention.

FIG. 5 is a configuration diagram of a character recognition unit of the present invention.

FIG. 6 is a configuration diagram of another embodiment of the character recognition device of the present invention.

FIG. 7 is a configuration diagram of another embodiment of the character recognition device of the present invention.

FIG. 8 is an output diagram of the character type determination unit of the present invention.

FIG. 9 is a configuration diagram of a conventional character recognition device.

[Explanation of symbols]

 1 character recognition unit 2 word search unit 3 phrase search unit 4 phrase evaluation value calculation unit 5 phrase selection unit 6 word dictionary 7 grammar dictionary 8 character correction unit 9 candidate character comparison unit 10 character image 11 candidate character set 12 candidate word set 13 candidate Phrase set 14 Modified character string 15 Additional learning character 16 Recognition dictionary 21 Same character extraction unit 31 Keyword extraction unit 32 Keyword partial match search unit 33 Candidate word addition unit 34 Candidate non-character detection unit 35 Keyword set 36 Similarity calculation unit 37 Weight coefficient 38 Weighting coefficient update unit 41 Word error correction degree calculation unit 42 Rejected character determination unit 43 Candidate word addition unit 51 Character type determination unit 52 Character type dictionary 61 Automatic correction unit 62 Manual correction control unit 63 Correction rule table

Front page continued (72) Inventor Hidetsugu Maekawa 1006 Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. (72) Inventor, Satoshi Ito 1006 Kadoma, Kadoma City Osaka Prefecture (72) Invention Person Yoshihiro Kojima, 1006, Kadoma, Kadoma, Osaka Prefecture, Matsushita Electric Industrial Co., Ltd. (72) Inventor, Koji Yamamoto, 1006, Kadoma, Kadoma, Osaka Prefecture Matsushita Electric Industrial Co., Ltd.

Claims (6)

[Claims] 1. A character recognition unit that recognizes a document image and outputs N candidate characters per character, a word search unit that obtains a candidate word set from a candidate character set using a word dictionary, and a candidate word set. A phrase search unit that obtains candidate phrases from a grammar dictionary, a phrase evaluation value calculation unit that calculates the lexical and grammatical correctness of a phrase, and a phrase that selects and corrects a phrase based on the phrase evaluation value. A character recognition device comprising: a phrase selection unit that outputs a candidate character set and a candidate character comparison unit that compares a candidate character set with a modified character string and outputs an additional learned character. 2. A character recognition unit that recognizes a document image and outputs N candidate characters per character, a word search unit that obtains a candidate word set from a candidate character set using a word dictionary, and a candidate word set. A phrase search unit that obtains candidate phrases from a grammar dictionary, a phrase evaluation value calculation unit that calculates the lexical and grammatical correctness of a phrase, and a phrase that selects and corrects a phrase based on the phrase evaluation value. And a candidate character comparison unit that compares the candidate character set and the modified character string, and a same character extraction unit that extracts the same character in a different word and outputs an additional learning character. Character recognizer. 3. A character recognition unit that recognizes a document image and outputs N candidate characters per character, a word search unit that obtains a candidate word set from a candidate character set using a word dictionary, and a candidate word set. A phrase search unit that obtains candidate phrases from a grammar dictionary, a phrase evaluation value calculation unit that calculates the lexical and grammatical correctness of a phrase, and a phrase that selects and corrects a phrase based on the phrase evaluation value. A phrase selection unit that outputs, a keyword extraction unit that extracts a keyword of a document to be recognized from a corrected character string, a keyword partial match search unit that performs a partial match search with the keyword in the candidate character string set, A candidate word addition unit that adds a partially matched keyword to the candidate word set, and a non-candidate character detection unit that detects the characters added by the candidate word addition unit from the corrected character string and outputs additional learning characters. Character recognition apparatus characterized by comprising. 4. A character recognition unit that recognizes a document image and outputs N candidate characters per character, a word search unit that obtains a candidate word set from a candidate character set using a word dictionary, and a candidate word set. A phrase search unit that obtains candidate phrases from a grammar dictionary, a phrase evaluation value calculation unit that calculates the lexical and grammatical correctness of a phrase, and a phrase that selects and corrects a phrase based on the phrase evaluation value. A phrase selection unit that outputs, a keyword extraction unit that extracts a keyword of a document to be recognized from a corrected character string, a keyword partial match search unit that performs a partial match search with the keyword in the candidate character string set, The word error correction degree calculation unit that determines the error correction degree of the words in the corrected character string, the reject character determination unit that determines the reject character from the word error correction degree, and the keyword that partially matches the reject character are compared. And a candidate word addition unit that adds matching characters to the candidate word set, and a non-candidate character detection unit that detects the characters added by the candidate word addition unit from the corrected character string and outputs additional learned characters. Character recognition device. 5. A character recognition unit updates a weighting factor from a similarity calculation unit that calculates a character similarity from a character image and a weighting factor and outputs a candidate character, and an error between an additional learning character and a candidate character. The character recognition device according to claim 4, further comprising: 6. A character recognition unit that recognizes a document image and outputs N candidate characters per character, a word search unit that obtains a candidate word set from a candidate character set using a word dictionary, and a candidate word set. A phrase search unit that obtains candidate phrases from a grammar dictionary, a phrase evaluation value calculation unit that calculates the lexical and grammatical correctness of a phrase, and a phrase that selects and corrects a phrase based on the phrase evaluation value. A character recognition device characterized by comprising a phrase selection unit for outputting a character type and a character type determination unit for determining a character type from a corrected character string using a character type dictionary.