JP3085107B2 - Character recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognizing device for reading and recognizing characters described in a form or the like.

[0002]

2. Description of the Related Art Hitherto, knowledge processing has been introduced to the results obtained by character recognition processing to improve recognition accuracy. This knowledge processing is a method of correcting the recognition result to the most likely character by comparing the recognition result for each character with a knowledge dictionary. Depending on the content of the recognition target, a word dictionary, a place name dictionary, a personal name dictionary, a part number dictionary, or the like is used as the knowledge dictionary.

FIG. 6 is a diagram showing a configuration of a conventional character recognition device. The operation of the device will be described below with reference to FIG. As shown in the figure, the character recognition unit 52 reads the form image 51 and outputs n candidate characters per character. String search section 53
Calculates a combination of characters constituting the character string included in the knowledge dictionary 54 from the candidate character string set using the knowledge dictionary 54, and calculates the number of matching characters with the knowledge dictionary 54 by the character string candidate evaluation value calculation unit 55. And a character string candidate evaluation value based on the similarity in the character recognition unit 52 and the like. The character string candidate selection unit 56 selects a character string candidate with the highest character string candidate evaluation value, and outputs this character string candidate as a recognition result. As described above, by recognizing the character string, the character that the character recognizing unit 52 has erroneously recognized can be corrected, and the recognition can be improved.

[0004]

However, in the above-described conventional configuration, when the image to be recognized is tilted or has noise, the degree of similarity to the candidate character output from the character recognition unit 52 greatly varies. Since the character string candidate evaluation value changes accordingly, an incorrect candidate may be selected in the knowledge processing.

[0005] In addition, in the entry of a form or the like, the entry person may mistakenly input a delimiter character (for example, "town" in an address).
In some cases, “characters” and “−” and “/” in product numbers) may be inserted or omitted. In this case, since the character position is shifted, the character string search unit cannot perform matching with the knowledge dictionary and cannot search for a correct character string candidate.

The present invention solves such a conventional problem. It is an object of the present invention to obtain an evaluation value based on probable characters and infer an error of a writer, thereby increasing a character recognition rate. I have.

[0007]

According to the present invention, in order to achieve the above object, a character string candidate evaluation value calculating unit calculates a similarity distribution of each candidate character in a character recognition unit and a character appearance frequency in a knowledge dictionary. A character string candidate evaluation value is obtained based on the evaluation value.
Further, it infers a typographical error of the writer and deletes and inserts erroneous characters to obtain a correct character string.

[0008]

According to the present invention, the character string candidate evaluation value is obtained from the tendency of the similarity of each candidate character even if noise is present in the image to be recognized. The contribution to the evaluation value is reduced, and erroneous recognition can be reduced. Further, by inferring a typographical error of the writer, a correct character string can be obtained even when there is a typographical error, and the character recognition rate is improved.

[0009]

An embodiment of the present invention will be described below.
FIG. 1 shows the overall configuration of the character recognition device of this embodiment.

A character recognizing unit 2 performs character recognition from the form image 1 and sets a candidate character set having n candidate characters from a first candidate character to an nth candidate character per character and the recognition similarity of each candidate character. Is output.

The character string search section 3 searches the knowledge dictionary 4 from combinations of the candidate character sets to select combinations that are character string candidates. The character string candidate evaluation value calculator 5 calculates the degree of coincidence of the character string candidate searched by the character string search unit 3 with the knowledge dictionary 4, the degree of similarity recognized by the character recognition unit 1, and the frequency of appearance of characters in the knowledge dictionary 4. The character string candidate evaluation value is calculated based on. The character string candidate selection unit 6 selects a character string candidate having the largest character string candidate evaluation value among the character string candidates.

The erroneous entry inference unit 7 infers from the character string candidate evaluation value of the character string candidate selected by the character string candidate selection unit 6 and a candidate character set whether there is any erroneous entry by the writer. The delimiter deletion / insertion unit 8 deletes and inserts an erroneously entered character when there is an erroneous entry by the writer.

FIG. 2 is a diagram showing the internal configuration of the character string candidate evaluation value calculation unit, which will be described. First, the similarity evaluation unit 12 performs similarity evaluation based on the tendency of similarity of each candidate character, and the character appearance frequency evaluation unit 13 performs character appearance frequency evaluation based on the appearance frequency information of the characters included in the knowledge dictionary. The character evaluation value calculation unit 14 obtains the character evaluation value of each candidate character based on the similarity evaluation and the character appearance frequency evaluation. The non-candidate character evaluation unit 15 evaluates non-candidate characters. String evaluation value derivation unit
Step 16 calculates a character string evaluation value of each character string candidate from the character evaluation value of each candidate character.

FIG. 3 is a diagram showing the internal configuration of the erroneous entry inference unit, which will be described below. First, the high character evaluation value character selection unit 22 selects a candidate character having a high character evaluation value,
The delimiter search unit 23 searches for a delimiter from among the candidate characters output from the high-character evaluation value character selection unit 22. The erroneous insertion inference unit 24 infers the erroneous insertion of the delimiter by the writer, and outputs the delimiter deletion instruction 28 when inferring that there is an erroneous insertion. The front part match search unit 25 searches the candidate characters output from the high character evaluation value character selection unit 22 for a character whose front part matches the character string included in the knowledge dictionary. The erroneous omission inference unit 26 infers an erroneous omission of the delimiter of the entry person and outputs a delimiter insertion instruction 29 when inferring that there is an erroneous omission.

The character recognition apparatus having the above-described structure performs character recognition as follows. First, the form image 1 is processed by the character recognizing unit 2, and a candidate character set having n candidate characters from the first candidate character to the n-th candidate character per character and the recognition similarity Aij (i is Character position, j is the j-th candidate). FIG. 4 shows the result of performing character recognition by the character recognition unit 2 and obtaining candidate characters from the first candidate character to the fifth candidate character for each character. The character string search unit 3 extracts character strings that partially match the combination of the candidate character sets from the character strings included in the knowledge dictionary 4 as character string candidates. For example, as a result of performing partial matching between the candidate character set of FIG. 4 and the knowledge dictionary 4 shown in FIG. 5, “FY-38N”, “JN-28”,
“JP-28M” and “LR-V08” are extracted as character string candidates.

From the character string candidates output from the character string search unit 3, the similarity evaluation unit 12 checks the similarity of candidate characters that match the character string candidates for each character position, and determines the similarity for the character position i from the similarity. The calculated similarity evaluation value Ri is output. When the character that matches the character string candidate is the j-th candidate, Ri is a function determined by Aij, for example, Ri = Aij. Also, normalized by the recognition similarity Ai1 of the first candidate,
Ri = Aij / Ai1. However, if there is no character that matches the character string candidate, Ri = 0.
The character appearance frequency evaluation unit 13 obtains a frequency probability Pi that each character of the character string candidate output from the character string search unit 3 appears at the same character position i of the character string included in the knowledge dictionary 4. For example, the frequency probability P1 of the character string candidate "FY-38N" is
In the character string included in the knowledge dictionary, "F"
Is the probability that appears.

The character evaluation value calculator 14 calculates a character evaluation value Bi at each character position of a character string candidate. Bi is Bi = Ri
× f (Pi), where f (Pi) is a monotonically decreasing function. For example, f (Pi) = 1 / Pi. By lowering the character evaluation value Bi for a character having a higher frequency probability Pi, the recognition character string determination authority for frequently occurring characters can be reduced. This is a frequently occurring character (eg, a punctuation character)
Because is included in many character strings, relying solely on that character to infer a correct character string is highly risky of misrecognition.

The non-candidate character evaluation unit 15 obtains a character evaluation value Bi when there is no character that matches the character string candidate. The character evaluation value Bi is Bi = −a × (Ai1−Ain), and a
Is a constant of 0 <a <1. For example, the character string candidate “FY
The character evaluation value B1 when "F" at character position 1 of "-38N" is not in the first to n-th candidate characters is B1 =
−a × (A11−A1n).

The character string evaluation value deriving section 16 obtains a character string candidate evaluation value C from the character evaluation value Bi of each character of the character string candidate. The character string candidate evaluation value C is obtained by C = ΣBi. The character string candidate selection unit 6 compares the character string candidate evaluation values of the respective character string candidates and selects the character string candidate having the largest character string candidate evaluation value. The high character evaluation value character selection unit 22 selects a candidate character having a character evaluation value of a certain value or more from a candidate character set. Alternatively, candidate characters within the m-th candidate (1 <m <
Select n).

The delimiter search unit 23 searches for a delimiter from the candidate characters selected by the high character evaluation value character selection unit 22. In the erroneous insertion inference unit 24, when the character string candidate evaluation value of the character string candidate output by the character string candidate selection unit 6 is equal to or less than a certain value, and the delimiter search unit 23 can search for the delimiter, Is output to the delimiter deletion insertion unit 8 as a delimiter deletion instruction. On the other hand, if the character string candidate evaluation value of the character string candidate output by the character string candidate selection unit 6 is equal to or greater than a certain value, or if the delimiter search unit 23 cannot search for a delimiter, the character string candidate selection unit 6 outputs the character string candidate output to the front part match detection unit 25.

In the front part matching search unit 25, the first part that does not match the front part of the character string of the knowledge dictionary 4 and the combination of the candidate characters selected by the high character evaluation value character selection unit 22 and does not match in the knowledge dictionary is used. A character string whose character is a delimiter is searched from the knowledge dictionary. The omitting omission inference unit 26 determines that the character string candidate evaluation value of the character string candidate output by the character string candidate selection unit 6 is equal to or less than a certain value, and that the front part matching search unit 25 can search for the front part matching character string. The character position of the first character that does not partially match is output to the delimiter deletion / insertion unit as a delimiter insertion instruction. On the other hand, if the character string candidate evaluation value of the character string candidate output by the character string candidate selection unit 6 is equal to or greater than a certain value, or if the front partial match search unit 25 cannot search for a front part match character string, The character string candidates output by the column candidate selection unit 6 are output as corrected character strings.

The delimiter deletion / insertion unit 8 edits a candidate character set based on a delimiter deletion instruction 28 or a delimiter insertion instruction 29. In the case of the delimiter deletion instruction 28, the candidate character at the character position of the delimiter is deleted from the candidate character set, and the candidate characters at the character positions subsequent to the character position are shifted one character position to the preceding character position. In the case of the delimiter insertion instruction 29, the candidate characters at the character positions subsequent to the delimiter character position are shifted one character at a time to the next character position with respect to the candidate character set, and the first candidate character position of the delimiter character is shifted. As a delimiter. Then, the candidate character set is output to the character string search unit.

[0023]

As described above, by using the character recognition apparatus of the present invention, a character string evaluation value is calculated from the tendency of recognition similarity and the frequency of appearance of characters in the knowledge dictionary. Even in the case where a correction character string is included, the correction character string can be determined more accurately. In addition, by the erroneous entry inference, a correct character string can be determined even when a writer inserts or omit a separator character by mistake. In order to perform character recognition in this way, the recognition rate is improved,
The effect is great.

[Brief description of the drawings]

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

FIG. 2 is a block diagram illustrating a configuration of a character string candidate evaluation value calculation unit according to the embodiment;

FIG. 3 is a block diagram illustrating a configuration of an erroneous entry inference unit according to the embodiment;

FIG. 4 is an output diagram of a character recognition unit according to the embodiment.

FIG. 5 is an output diagram of a character string search unit according to the embodiment;

FIG. 6 is a configuration diagram of a conventional character recognition device.

[Explanation of symbols]

 1 Form Image 2 Character Recognition Unit 3 Character String Search Unit 4 Knowledge Dictionary 5 Character String Candidate Evaluation Value Calculation Unit 6 Character String Candidate Selection Unit 7 Mistake Inference Unit 8 Delimiter Deletion and Insertion Unit 9 Corrected Character String 11 Character String Candidate 12 Similar Degree evaluation unit 13 Character appearance frequency evaluation unit 14 Character evaluation value calculation unit 15 Non-candidate character evaluation unit 16 Character string evaluation value derivation unit 17 Character string candidate evaluation value 21 Candidate character set 22 High character evaluation value character selection unit 23 Delimiter search Part 24 Incorrect insertion inference part 25 Partial match search part 26 Incorrect omission inference part 27 Corrected character string 28 Delimiter deletion instruction 29 Delimiter insertion instruction

────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Kazuhiro Kayashima 1006 Kadoma Kadoma, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. (56) References JP-A-4-120679 (JP, A) JP-A-4- 205457 (JP, A) JP-A-5-62022 (JP, A) JP-A-2-116988 (JP, A) JP-A-5-174195 (JP, A) JP-A-2-121078 (JP, A) JP-A-4-77980 (JP, A) JP-A-4-205457 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06K 9/72

Claims (6)

(57) [Claims] A character recognition unit for recognizing an input image character by character and outputting n candidate characters and recognition similarity for each character, and a partial match between a combination of a character string and a candidate character included in a knowledge dictionary. A character string search unit that searches for character strings to be output and outputs character string candidates, and n character strings at each character position of each character string candidate
Of the recognition similarity of the candidate character and the character string included in the knowledge dictionary
And string candidate evaluation value calculating unit for obtaining the character string candidate evaluation value appearance frequency of definitive to each character position based on the character string candidate selection unit for selecting a character string candidates with the highest string candidate evaluation value , a false description inference unit for inferring a registrant erroneous entry from the character string candidate and the candidate character set selected, the delimiter deletion insertion unit for deleting or insertion of false description character
Character recognition apparatus characterized by comprising. Character wherein character string candidate evaluation value calculation unit, and a similarity evaluation unit for recognizing similarities or et similarity evaluation of n candidate characters in each character position of the character string candidates, included in the knowledge dictionary A character appearance frequency evaluation unit that evaluates the appearance frequency of a character at each character position in the column; a character evaluation value calculation unit that obtains a character evaluation value at each character position of a character string candidate from similarity evaluation and character appearance frequency evaluation; character recognition system according to claim 1, characterized in that it comprises a string evaluation value deriving portion for obtaining the string evaluation value from the evaluation value. Statement 3. A string candidate evaluation value calculation unit, and a similarity evaluation unit for recognizing similarities or et similarity evaluation of n candidate characters in each character position of the character string candidates, included in the knowledge dictionary
A sentence that evaluates the frequency of occurrence of characters at each character position in the character string
A character appearance frequency evaluation unit, a non-candidate character evaluation unit that performs a non-candidate character evaluation when there is no matching character among the candidate characters from the similarity evaluation and the character appearance frequency evaluation, and a character character recognition system according to claim 1, characterized in that it comprises a string evaluation value introducing unit for obtaining the character string evaluation value from the evaluation value. 4. A high character evaluation value character selecting unit for selecting a candidate character having a character evaluation value equal to or greater than a predetermined value, and a candidate character selected by the high character evaluation value character selecting unit. and delimiter character search unit that searches for the delimiter character from the string climate
Complementary evaluation value is less than a certain value and separated by delimiter search part
When a character can be searched, the delimiter is
Erroneous insertion of characters and delimiter
Character recognition system according to claim 1, characterized in that it comprises a false insertion inference unit <br/> deletes the candidate character in character positions. 5. An erroneous entry inference unit comprising: a high character evaluation value character selection unit for selecting a candidate character having a character evaluation value equal to or greater than a certain value; and a character string of a knowledge dictionary and a high character evaluation value selection unit. Selected
Before performing a partial match search with the combination of candidate characters
Partial match search part, character string candidate evaluation value is below a certain value
And if a partial match string can be searched,
The first character position that was not entered was incorrectly omitted by the writer.
And the partial match does not occur for the candidate character set.
First character recognition system according to claim 1, characterized in that it comprises a false omitted inference unit <br/> and inserting the delimiter character positions. 6. A character string candidate evaluation value calculation unit calculates Ri as a character string
The recognition similarity of the candidate character at the character position i in the candidate and
Pi is a sentence in which each character of the character string candidate is included in the knowledge dictionary
F (Pi) is the probability of occurrence at the same character position i in the character string
Is a monotonically decreasing function, the product of Ri and f (Pi)
2. A character string evaluation value is obtained by:
Character recognition device according to the description.