JP3157557B2 - 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 recognition device,
In particular, the present invention relates to a character recognition device having a function of additionally registering a character so that a character misidentified by a device user can be recognized.

[0002]

2. Description of the Related Art There is known a character recognition apparatus having a function of additionally registering a character pattern (feature amount) extracted from a misidentified character in a pattern dictionary. For example, the device described in Japanese Patent Application Laid-Open No. 57-187781 has a user dictionary as an auxiliary dictionary of a system pattern dictionary, and the device user can additionally register a character pattern of a misidentified character in the user dictionary. it can.

[0003] In addition, a table (confusion table) in which confusion characters similar to character patterns registered in the pattern dictionary are registered, and confusion characters of candidate characters obtained by pattern matching with the pattern dictionary are converted. There is known a character recognition device configured to take out a confusion character from a fusion table and provide the confusion character together with a candidate character to a post-processing unit to determine a final recognition result.

[0004]

Generally, pattern dictionaries are organized by various means such as clustering so as to achieve high recognition performance with as few records as possible. Showa 59-16368
No. 1), if additional registration of a character pattern in the pattern dictionary by the device user is allowed unconditionally, there is a concern that adverse effects on recognition of other characters may be caused.

[0005] In a character recognition device provided with a confusion table, there is a possibility that the misrecognized character can be recognized by additionally registering the misrecognized character as a confusion character of the candidate character in the confusion table. However, if additional registration in the confusion table is left to the discretion of the user, the confusion characters of the characters registered in the pattern dictionary will increase unnecessarily, causing adverse effects such as erroneous correction in post-processing. There is a concern that erroneous characters that would otherwise become recognizable unless additionally registered in the pattern dictionary may be meaninglessly added to the confusion table.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a character recognition apparatus having means for adding a confusion character to a candidate character by pattern matching, and means for post-processing using character cut-out information or linguistic knowledge. It is an object of the present invention to allow a user of a device to perform reasonable additional character registration so that a device can be recognized.

[0007]

According to the present invention, there is provided means for extracting a character from an input image, extracting a character pattern from the character extracted by the means, and matching a plurality of candidate characters by pattern matching with the pattern dictionary. A pattern recognition unit to be sought, a confusion table in which confusion characters having similar character patterns are registered, a unit for adding the confusion character to the candidate character, and a process for adding the candidate character and the confusion character added thereto. As a recognition result, in a character recognition device having means for performing post-processing for the correction, when registration of a misidentified character is instructed, the misidentified character is stored in either the pattern dictionary or the confusion table. Determining means for determining whether additional registration should be performed, and determining whether or not the confusion Table addition means for additionally registering the misidentified character as a confusion character of one or more candidate characters in the confusion table in the case of additional registration in the table, and determining by the determination means the addition to the pattern dictionary. In the case of registration, there is provided a dictionary adding means for additionally registering a character pattern extracted from the misrecognized character by the pattern recognition means in the pattern dictionary.

In the first embodiment of the present invention,
Means for determining the cut-out information obtained by the cut-out means for the misrecognized character, as information for determination (a match for the character line);
Position, height, width, area, etc.).

When the post-processing means can discriminate the erroneous character from the first candidate character based on the cut-out information , the judging means judges that the character should be additionally registered in the confusion table. Additional registration in the confusion table as a confusion character for the candidate character
I do.

In a second embodiment of the present invention,
The means uses the magnitude of the probability that the candidate character is erroneously recognized as information for determination .

The determining means determines that the first candidate character should be additionally registered in the confusion table when the probability of misrecognition is high, and the table adding means determines the misrecognized character as a confusion character of the first candidate character. Register additionally in the table .

The determination means determines that additional registration should be made in the confusion table when there is at least one candidate character having a high probability of being erroneously recognized. A misidentified character is additionally registered in the confusion table as a character confusion character .

In the third embodiment of the present invention, the judging means uses the cutout information of the erroneous character and the magnitude of the probability of erroneous recognition of the candidate character as information for the judgment .

The determining means determines that the first candidate character should be additionally registered in the confusion table when the probability of being erroneously recognized as the first candidate character is high and the first candidate character and the erroneously recognized character can be distinguished from cutout information. The table addition means additionally registers the erroneously recognized character in the confusion table as the confusion character of the first candidate character .

[0015] If the probability of misrecognition is high and there is at least one candidate character that can be identified as a misrecognized character based on the cut-out information , the determining means determines that additional registration should be made in the confusion table. Additionally registers the misidentified character in the confusion table as a confusion character of each candidate character that has a high probability of being misidentified and can be identified as an erroneous character by cutout information .

[0016]

According to the present invention , misidentified characters are additionally registered in the pattern dictionary or the confusion moon table by the judgment of the judging means. It will be possible to achieve the purpose of additional registration with reasonable control and avoiding adverse effects.

When a misidentified character is added to the confusion table as a confusion character of a candidate character obtained by pattern matching with respect to the misidentified character, the misidentified character is added to the candidate character as a confusion character, and the post-processing is performed. If the confusion character is likely to be sent and the confusion character is selected as the correct answer in the post-processing, it is effective to additionally register the misidentified character in the confusion table, and conversely, if the confusion character is not selected as the correct answer However, additional registration in the confusion table is meaningless and harmful, and misidentified characters should be additionally registered in the pattern dictionary.

In addition, a misidentified character added as a confusion character to a candidate character by pattern matching is
If it is selected as a correct answer by post-processing using the cutout information of the character (image) (relative position, height, width, area, etc. with respect to the line of the character), that is, a candidate character and a misidentified character according to the cutout information If it can be determined, additional registration of the erroneous character in the confusion table is effective. Otherwise, it is necessary to additionally register the erroneous character in the pattern dictionary.

According to the first embodiment of the present invention , it is possible to perform reasonable additional registration control based on such cut-out information. Further, the first candidate character by pattern matching is a character whose character pattern is most similar to the misrecognized character, and is likely to be a candidate character in subsequent recognition. According to the present embodiment , it is determined whether or not only such first candidate characters should be additionally registered in the confusion table, and an erroneous character is additionally registered as a confusion character of only the first candidate in the confusion table. It is efficient and can suppress an increase in the number of characters registered in the confusion table.

The misidentified character is additionally registered in the confusion table as a confusion character of the candidate character. If the misidentified character is added to the candidate character as a confusion character at the time of subsequent recognition, the confusion character is correctly corrected in post-processing. Can be selected. However, even for the same character, the candidate character may fluctuate due to a difference in the document state or slight deformation, and if the same candidate character is not obtained at the time of recognition after additional registration, the character cannot be recognized. The additional registration in the confusion table is meaningless and harmful, and should be additionally registered in the pattern dictionary.

According to the second embodiment of the present invention, the probability of erroneous recognition of a candidate character, that is, the possibility of becoming a candidate character by pattern matching with respect to the same character even in subsequent recognition, is determined by the magnitude of the probability. Since the additional registration destination is determined, such inappropriate determination and additional registration can be avoided. Further , according to the present embodiment , such a determination based on the magnitude of the probability of being misrecognized is performed only for the first candidate character, and the misrecognized character is additionally registered as a confusion character of only the first candidate character in the confusion table. The processing is efficient, and an increase in the number of registered characters in the confusion table can be suppressed. Further, according to the present embodiment , since the misrecognition character is additionally registered in the confusion table as a confusion character of each candidate character having a high probability of being misrecognized, when the order of the candidate characters is changed during subsequent recognition. Also, it is more likely that a correct character is added as a confusion character and a correct answer can be obtained by post-processing.

According to the third embodiment of the present invention , only when a candidate character that can be identified as an erroneous character is obtained from the cut-out information, the erroneous character is added to the confusion table as a confusion character of the candidate character. Registration can be performed, and rational additional registration control can be performed when only post-processing is performed on a character-by-character basis using cutout information. Further, even in the post-processing also performs device string units using the word knowledge such as, when you add a certain character as confusion character 1 is modified as long as that can not be corrected in characters occurs even be impossible, the According to the embodiment , it is possible to avoid additional registration of a confusion character that causes such inconvenience.

[0023]

EXAMPLES Hereinafter, some implementation of the invention with reference to the drawings
An embodiment (example) will be described.

Embodiment 1 FIG. 1 is a schematic block diagram of a character recognition device. The original is read as a binary image by the scanner 1 and input to the cutout unit 2. The cutout unit 2 cuts out a line image and individual characters (character images) from the input image. At this time, the relative position, height, width, area, and the like of each character with respect to the line are also obtained as cutout information.

In the pattern recognition unit 3, a cut-out unit
After normalizing the size of the character input from 2 in the normalizing section 4, the feature extracting section 5 extracts the characteristic amount (character pattern) of the character after the normalization. The pattern matching unit 6
Pattern matching is performed between the character pattern of the input character and the character pattern in the pattern dictionary 7, and a plurality of characters having similar character patterns are output as candidate characters in ascending order of distance. Here, since the distance is used as the similarity, the smaller the similarity, the more similar the character patterns are.

The confusion character adding unit 8 has therein a confusion table 9 in which confusion characters similar in character pattern to specific characters registered in the pattern dictionary 7 are registered in advance.
The confusion character of each input candidate character is extracted from the confusion table 9 and output to the post-processing unit 10 together with the candidate character.

FIG. 2 is a conceptual diagram showing the configuration up to the pattern recognition unit 3 and the confusion character adding unit 8. 30 is a character (gate) cut out from the input image, 31 is its character pattern (feature amount), 32 is a part of character patterns (feature amount) registered in the pattern dictionary 7, and 33 is obtained by pattern matching. Candidate characters (here, up to the third candidate), 34
Is a part of the registered contents of the confusion table 9, 35
Is a recognition result to which a confusion character is added.
In the example shown here, one confusion character (gate)
Is added as a fourth candidate.

In FIG. 1, the post-processing unit 10 treats both the candidate characters obtained by the pattern matching unit 3 and the confusion characters added by the confusion character adding unit 8 as candidate characters, and
The recognition result is corrected by selecting the correct character from the candidate characters by using the post-processing knowledge registered in. This post-processing includes processing on a character-by-character basis based on character cut-out information, and language processing on a character string basis using word knowledge, knowledge on character types, and the like. For post-processing based on the cutout information, the cutout information obtained by the cutout unit 2 is also input to the post-processing unit 10.

Reference numeral 12 denotes a console having an input device such as a display and a keyboard. The operation of displaying the recognition result after the post-processing, correcting the recognition error by the device user, and additionally registering a misrecognized (unrecognizable) character is performed on the console 1.
2 is performed.

An additional registration determination unit 20, a table additional registration unit 21, and a dictionary additional registration unit 22 are provided in association with the additional registration of misidentified characters specified by the apparatus user. The additional registration determination unit 20 determines whether the misrecognized character should be additionally registered in the pattern dictionary 7 or the confusion table 9, and according to the determination result, the table additional registration unit 21
And a section for controlling the dictionary addition registration unit 22.

FIG. 3 is a flowchart of the additional registration process. When additional registration of a misrecognized character is instructed, the additional registration determination unit 20
When the erroneous character is added to the candidate character and input to the post-processing unit 10, whether the candidate character and the first candidate can be determined by the cutout information, that is, the erroneous character is determined as the correct character in the post-processing by the cutout information It is determined whether selection can be made (processes 101, 103, 105, 107). For such a determination, the additional registration determination unit 20 captures cutout information of erroneously recognized characters from the cutout unit 2, captures first candidate characters from the pattern recognition unit 3, and performs post-processing based on the cutout information from the knowledge dictionary 11. Capture the knowledge for

In this embodiment, as the cut-out information for determination, the relative position of the character generally added to the cut-out character with respect to the line (process 101), the height of the character (process 103), and the character width ( Although the processing 105) and the character area (processing rate 107) are used, the order of the determination is arbitrary, and the determination may be made only for a part of the determination.

For example, it is assumed that the input character "," is erroneously recognized and the first candidate character is "'". The misidentified character “,” is a character located on the lower side of the line, whereas the first candidate character “′” is a character located on the upper side of the line. Therefore, the condition of the process 101 is satisfied. Therefore, the additional registration determination unit 20 determines that additional registration should be performed in the confusion table 9, and causes the table additional registration unit 21 to additionally register a misidentified character as a first candidate character confusion character in the confusion table 9 (process 10).
9).

The input character “x” is misidentified, and its first character is “x”.
It is assumed that the candidate character is “X”. Since the misidentified character “x” has a low character height and the first candidate character “X” has a high character height, these can be determined and corrected by post-processing based on character height information. Holds.
Therefore, the additional registration determination unit 20 causes the table additional registration unit 20 to additionally register the erroneously recognized character in the confusion table 9 as a confusion character of the first candidate character (process 109).

If "tsu" is erroneously recognized and the first candidate character is "tsu", "tsu" has a narrow character width and "tsu" has the full-width character width. Since the information can be determined and corrected based on the information, the condition of the process 105 is satisfied.
Therefore, this misidentified character is added to the confusion table 9 as a confusion character of the first candidate character.

If "@" is erroneously recognized and the first candidate character is "a", "@" has a small character area,
Since “A” has a character area of two-byte characters, these can be determined and corrected based on the character width information in post-processing, so that the condition of the process 107 is satisfied. Therefore, this misidentified character is additionally registered in the confusion table 9 as a confusion character of the first candidate character (process 10).
9).

On the other hand, processing 101 to 10 for erroneously recognized characters
If none of the conditions 7 is satisfied, the additional registration determination unit 2
0 causes the dictionary additional registration unit 22 to additionally register the character pattern (feature amount) of the erroneously recognized character extracted by the feature extraction unit 5 in the pattern dictionary 7 (process 111).

As described above, erroneous characters that can be corrected by post-processing based on cut-out information are added to the confusion table 9 and only other erroneous characters are added to the pattern dictionary 7. Can be suppressed, and the adverse effect of additional registration can be reduced.

Embodiment 2 A part of the structure relating to the additional registration of the character recognition device is changed as shown in FIG. 4, and the other structure is the same as that of the first embodiment. The additional registration determination unit 20A uses the pattern recognition unit 3 to determine the degree of similarity (distance) between the candidate character and the candidate character in order to use the magnitude of the probability that the candidate character is erroneously recognized with respect to the erroneous character as the determination information.
Information is input.

FIG. 5 is a flowchart of the additional registration process. When additional registration of a misrecognized character is instructed, additional registration determination unit 2
OA determines whether to add a misrecognized character to the pattern dictionary 22 or the confusion table 9 based on the magnitude of the misrecognition probability of the first candidate character (process 121). In the present embodiment, the similarities (distances) between the erroneously recognized character and the first and second candidate characters are D1 and D2,

## EQU1 ## When the condition of D1 / (D2-D1) <10 is satisfied, there is a high probability that the first candidate character is erroneously recognized, that is, the first candidate character becomes a candidate character when an erroneous character is input. It is determined that the probability is large, it is determined that additional registration should be made in the confusion table 9, and the misrecognized character is set as a first candidate confusion character by the table additional registration unit 21.
(Step 123). If the condition of (Equation 1) is not satisfied, the additional registration determination unit 20A sets the pattern dictionary 7
Is determined to be additionally registered, and the character pattern of the misrecognized character is additionally registered in the pattern dictionary 7 by the dictionary additional registration unit 22 (process 125). (Equation 1) is merely an example, and another determination formula may be used.

For example, "thick" is erroneously recognized, the first candidate character is "large", its similarity is 80, the second candidate character is "heaven",
It is assumed that the similarity is 120. Since the distance is used as the similarity of the character pattern, the value of the similarity is smaller as the character pattern is more similar. In the case of this misidentified character, since the condition of (Equation 1) is satisfied, the character is additionally registered in the confusion table 9. On the other hand, the “question” is misidentified, the first candidate character is “between”, its similarity is 150,
It is assumed that the candidate character is “Seki” and the similarity is 160. In this case, since the condition of (Equation 1) is not satisfied, the erroneously recognized character is additionally registered in the pattern dictionary 7.

Embodiment 3 The configuration of the character recognition device is the same as that of Embodiment 2.

FIG. 6 is a flowchart of the additional registration process. Processes 131 and 135 are the same processes as processes 121 and 125 in FIG. The processing 123 differs from the processing 123 in FIG. 5 in that, when there is a second or lower candidate character having a high probability of being erroneously recognized as an erroneous character, the erroneous character is additionally registered as the confusion character.

That is, when an erroneous character is additionally registered in the confusion table 9, not only is it additionally registered as a confusion character of the first candidate character, but also the similarity (distance) of the first candidate character to the erroneous character is D1. , When the similarity of each of the second and lower candidates is Di,

## EQU2 ## If there is an i-th candidate character that satisfies the condition of Di-D1 <30, the i-th candidate character is also considered to have a high misidentification probability, so that the misidentification character is additionally registered as the confusion character.

For example, "Musu" is misidentified, the first candidate character is "white" and the similarity is 100, the second candidate character is "day" and the similarity is 120, and the third candidate character is "eye" and similar. Degree 140
In the case of, the condition of (Equation 1) is satisfied, and additional registration in the confusion table 9 is performed. And
Since the condition of (Equation 2) is satisfied for the second candidate, the misidentified character is additionally registered as a confusion character of the first candidate character and the second candidate character.

Embodiment 4 The configuration of a character recognition device is the same as in Embodiments 2 and 3.
Further, the additional registration processing is expressed by the flowchart of FIG. 6 similarly to the third embodiment. However, in the process 131, when the similarity of the first candidate character is a fixed threshold value, for example, 100 or less,
It is determined that additional registration should be made in the confusion table 9. Then, in the process 133, an erroneously recognized character is additionally registered as a confusion character of each candidate character having a similarity lower than the threshold.

For example, "go" is erroneously recognized, the first candidate character is "contained", the similarity (distance) is 80, and the second candidate character is "sha".
And the similarity is 95, the third candidate character is “kura” and the similarity is 12
In the case of 0, it is considered that the probability of being erroneously recognized up to the second candidate character is large, and the erroneously recognized character is additionally registered as a confusion character of “contain” and “sha”. On the other hand, if the “sense” is misidentified and the similarity of the first candidate character “composition” is 115, the similarity exceeds the threshold, and the misidentified character is additionally registered in the pattern dictionary 7.

Fifth Embodiment The configuration of a portion related to additional registration of a character recognition device is changed as shown in FIG. 7, and the other configuration is the same as that of the first embodiment. The additional registration determination unit 20B uses the cutout information of the erroneous character and the magnitude of the probability that the candidate character is erroneously recognized as the determination information. , And knowledge for post-processing based on the cut-out information is given from the knowledge dictionary 11.

FIG. 8 is a flowchart of the additional registration process. The process 140 is the same as the process 121 of the second embodiment or the process 131 of the third or fourth embodiment. If the condition of this process 141 is not satisfied, it is determined that additional registration in the pattern dictionary 7 is to be performed, and the misrecognized character is additionally registered in the pattern dictionary 7 by the dictionary additional registration unit 22 (process 1).
47).

However, when the condition of the process 141 is satisfied, the additional registration judging unit 20B makes the judgment of the process 143. In this process 143, the first candidate character and (Equation 2)
For each of the lower candidate characters (candidate characters having a high misrecognition probability) satisfying the above condition or having a similarity value equal to or less than a fixed threshold, it is determined whether the post-processing based on the cutout information can determine and correct the misrecognition character. . The contents of this determination may be the same as the processing 101 to 107 in the first embodiment. If it is impossible to determine and correct any of the candidate characters, it is determined that the candidate character should be additionally registered in the pattern dictionary 7, but if there is at least one candidate character that can be determined and corrected, the confusion table 9 is entered. Judging that additional registration is required,
The table addition registration unit 21 additionally registers the erroneous recognition character in the confusion table 9 as a confusion character of each candidate character that has a high misrecognition probability and can be determined and corrected (process 145).

Embodiment 6 The configuration of the character recognition device is the same as that of Embodiment 5.

FIG. 9 is a flowchart of the additional registration process. The additional registration determination unit 20B first determines whether there is any candidate character that can be determined to be an erroneous character based on the cutout information (Process 15).
1). This determination method may be the same as in the first embodiment. If there is no candidate character that can be identified and corrected in the cutout information,
It is determined that additional registration should be made in the pattern dictionary 7, and the additional registration is performed by the dictionary additional registration unit 22 (process 157).

If there is a candidate character that can be determined and corrected based on the cut-out information, it is determined whether or not any of the candidate characters has a high false recognition probability (process 153). The determination of the magnitude of the misrecognition probability may be the same as the processing 141 of the fifth embodiment. If there is no candidate character having a high misrecognition probability, it is determined that the candidate character should be additionally registered in the pattern dictionary 7. Otherwise, the additional registration of the misrecognition character as a confusion character of each candidate character having a large misrecognition probability is made. Add table registration unit 2
1 (process 155).

Embodiment 7 In the processing 141 of the fifth embodiment, the magnitude of the misrecognition probability is determined only for the first candidate character, and in the processing 143, the possibility of determination / correction based on the cutout information is determined for only the first candidate. . Further, in processing 145, an erroneous character is additionally registered as a confusion character of only the first candidate. Except for this, it is the same as the fifth embodiment.

Eighth Embodiment In the processing 151 of the sixth embodiment, the possibility of discrimination / correction based on the cut-out information is determined only for the first candidate character, and only the magnitude of the misrecognition probability of the first candidate character is determined in the processing 153. In processing 155, an erroneously recognized character is additionally registered as a confusion character of only the first candidate. Other than this, it is the same as the sixth embodiment.

[0056]

According to the present invention , misidentified characters are additionally registered in the pattern dictionary or the confusion moon table by the judgment of the judging means. It allows you to control registration reasonably and achieve the purpose of additional registration while avoiding adverse effects.

Therefore, means for adding the confusion character to the candidate character by pattern matching,
In a character recognition device provided with post-processing means using character cut-out information, linguistic knowledge, and the like, the device user can make reasonable additional character registration for enabling recognition of misidentified characters.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a device configuration according to a first embodiment.

FIG. 2 is a conceptual diagram illustrating a configuration from pattern recognition to addition of a confusion character.

FIG. 3 is a flowchart for explaining additional registration processing of a misrecognized character in the first embodiment.

FIG. 4 is a block diagram showing a changed part of the device configuration of the second, third or fourth embodiment.

FIG. 5 is a flowchart for explaining additional registration processing of a misrecognized character in a second embodiment.

FIG. 6 is a flowchart for explaining additional registration processing of a misrecognized character in the third or fourth embodiment.

FIG. 7 is a book diagram showing a changed portion of the device configuration of the fifth, sixth, seventh or eighth embodiment.

FIG. 8 is a flowchart for explaining additional registration processing of a misrecognized character according to the fifth or seventh embodiment.

FIG. 9 is a flowchart for explaining additional registration processing of a misrecognized character in the sixth or eighth embodiment.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Scanner 2 Extraction part 3 Pattern recognition part 4 Normalization part 5 Feature extraction part 6 Pattern matching part 7 Pattern dictionary 8 Confusion character addition part 9 Confusion table 10 Post-processing part 11 Knowledge dictionary 12 Console 20, 20A, 20B Additional registration Judgment unit 21 Table addition registration unit 22 Dictionary addition registration unit

Claims (1)

(57) [Claims] 1. A means for extracting a character from an input image, a character recognizing means for extracting a character pattern from the character extracted by the means, and obtaining a plurality of candidate characters by pattern matching with the pattern dictionary, and a character pattern A means for adding a confusion character to the candidate character having a confusion table in which confusion characters similar to each other are registered, and as a recognition result of the candidate character and the confusion character added thereto,
In a character recognition device having means for performing post-processing for the correction, when registration of a misrecognized character is instructed, which of the pattern dictionary or the confusion table should be used to additionally register the misrecognized character A judging means for judging, and a table addition for additionally registering the misidentified character as a confusion character of one or more candidate characters in the confusion table when the judgment by the judging means is additional registration in the confusion table. Means,
Dictionary adding means for additionally registering, in the pattern dictionary, a character pattern extracted from the misrecognized character by the pattern recognizing means when the judgment of the judging means is additional registration in the pattern dictionary. Recognition device.