JPH06259595A - Device and method for processing character recognition - Google Patents

Abstract

PURPOSE:To improve recognition performance by enabling a same recognition result to be obtained for characters of a same form in the same document even when collation with a word dictionary, etc., is performed. CONSTITUTION:A recognition character in a character reading part 3 is corrected by collating with the word dictionary 7 in the unit of word, while, classification at every character of a same form is performed. Processing to obtain a same recognition result for the character of the same form is performed by reflecting the classification result of the character of a same form on character recognition by collation with the word dictionary. Thereby, it is possible to correct an erroneously recognized character by utilizing the result of the collation with another word for an unknown word.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition processing device having a function of optically reading a document image, recognizing the character thereof, and automatically correcting the result of character recognition using linguistic knowledge or the like. A recognition processing method.

[0002]

2. Description of the Related Art A character recognition processing device is used for automatically incorporating information described in a printed matter, a book, or the like into an information processing device. This apparatus optically reads a document image, converts it into an electric signal, and performs character recognition of characters included in the document image. In character recognition,
Usually, the characteristic amount of each read character is extracted, compared with the standard character characteristic amount, the character with the most similar characteristic amount is found, and the recognition result is obtained. However, even if the recognition process is performed for each character, it is not possible to completely prevent the recognition error, and in reality, the operator needs to evaluate the recognition result while reading the original text and correct it. . In order to reduce the burden on the operator, a technique has been introduced in which character recognition results are collated word by word with a word dictionary and post-processing is performed to improve the recognition performance (Japanese Patent Laid-Open No. 58-58).
-48181).

FIG. 2 shows such a conventional character recognition operation flowchart. First, in step S1, the feature amount extraction of each character is performed as described above. Next, in step S2, the feature amount is collated with reference to the standard character feature amount dictionary. As a result, a recognition result for each character is obtained. Next, in step S3, the character string is cut into word units. In English etc., there is a space between words, so it is easy to separate them. In Japanese,
Separation is performed by grammar analysis. Then, the word dictionary prepared in advance is collated (step S4). And
Find the word with the closest spelling. If the spelling is exactly the same, it may be left as it is, but if the spelling is not the same due to erroneous recognition of some characters, etc., if the word is recognized as a whole, in step S5, replace. In this way, after the erroneously recognized character is corrected, the process proceeds to step S6, and the character recognition result is output. Conventionally, character recognition processing has been performed by the above method.

[0004]

By the way, in the conventional method as described above, since recognition performance is improved by collating with the word dictionary, it is possible to correct unknown words that are not in the word dictionary. Can not. For example, "Ap
It is assumed that the word "lysia" is erroneously recognized as "Aly5: a". That is, in this case, "S"
Is mistakenly recognized as "5" and "i" as ":". However, if you can't find a word with a similar spelling,
You can't make this correction at all. Also, for example, "s
If "synaptic" is erroneously recognized as "syncptic", the above process causes "sy" to be detected from the word dictionary.
Two words, "noptic" and "synaptic" are given as candidate words. In this case, which one is selected depends on the comparison of the sum of the feature amounts. This result is generally determined by the font of characters and the habit of the recognition unit. Thus, even though the exact same word is used in the same document, it may be "synaptic" in some places.
It may be recognized as “Synaptic” in another place. That is, there is a problem that an erroneous correction occurs even when an appropriate candidate word exists in the word dictionary.

In particular, even if the recognition result for each character is the same with the characters having exactly the same shape, by combining with the recognition result of other characters by collating with the above-mentioned word dictionary, The unreasonable result that each is corrected to a different character may occur. The present invention has been made by paying attention to the above points, and in particular, for homomorphic characters, even if the collation with the word dictionary as described above is performed,
It is an object of the present invention to provide a character recognition processing device and a recognition processing method that improve the recognition performance by obtaining the same recognition result.

[0006]

According to a first aspect of the present invention, there is provided a character reading section for optically reading a document image and converting it into an electric signal for character recognition of a character included in the document image, and the recognition. A homomorphic character classification unit that classifies characters by homomorphic characters with similar characteristics, a recognition result correction unit that corrects the character recognition result on a word-by-word basis by matching the recognized characters with a word dictionary, and this recognition result correction A character recognition processing device, comprising: a correction control unit that controls the correction of each character by the unit so that the same recognition result can be obtained for the homomorphic character according to the classification result of the homomorphic character classification unit. .

According to the second aspect of the present invention, while character recognition of characters included in a document image is performed, classification is performed for each isomorphic character having a similar feature amount, and different recognition results are obtained for a plurality of characters of the same classification. At this time, the present invention relates to a character recognition processing method characterized by obtaining a unique recognition result by a majority vote.

According to a third aspect of the present invention, while character recognition of characters included in a document image is performed, classification is performed for each isomorphic character having a similar feature amount, and the recognition result is collated with a word dictionary on a word-by-word basis. When a specific character is determined to be erroneously recognized by replacing the character with a candidate word, the recognition result of the specific character and the isomorphic character is collectively corrected.

According to a fourth aspect of the present invention, while character recognition of characters included in a document image is performed, classification is performed for each isomorphic character having a similar feature amount, and the recognition result is collated with a word dictionary on a word-by-word basis. The present invention relates to a character recognition processing method characterized by selecting a candidate word that is recognized as the most valid when the recognition results of homomorphic characters are made the same when replacing the with.

[0010]

In this device, the recognized characters of the character reading unit are collated and corrected with the word dictionary in the unit of word as in the conventional case, and the same character is classified. Then, the classification result of the homomorphic characters is reflected in the character recognition result obtained by matching with the word dictionary, and the recognition processing is performed so that the same recognition result can be obtained for the homomorphic characters.
For homomorphic characters, the recognition results for the same sentence can be aggregated and unified if only a single recognition result is obtained by majority voting. When the recognition result of a specific character is corrected as a result of the word dictionary collation, the characters having the same shape as the character can be collectively corrected to be unified. Further, in the word matching, if the candidate word is selected in consideration of the same recognition result of the same characters, more accurate character recognition becomes possible. As a result, with respect to the unknown word, the misrecognized character can be correctly corrected by using the result of other word matching or the result of the homomorphic character classification.

[0011]

The present invention will be described in detail below with reference to the embodiments shown in the drawings. [First and Second Inventions] FIG. 1 is a block diagram showing an embodiment of a character recognition processing apparatus of the present invention. The apparatus shown in the figure includes a character reading unit 3 that reads a document image 2 in which a large number of characters 1 are written and recognizes each character. Also,
In addition to this, in order to classify the characters recognized by the character reading unit 3 for each isomorphic character having a similar feature, the isomorphic character classification unit 4 and the isomorphic character classification result storage unit 5 are provided in the present invention. Further, a recognition result correction unit 6, a word dictionary 7, and a candidate word storage unit 8 are provided to correct the character recognition result on a word-by-word basis by collating with a word dictionary similar to the conventional one.

Further, in the present invention, the correction control unit 9 is provided in order to adjust the correction of the recognition result according to the homomorphic character classification result. The correction control unit 9 receives the result stored in the homomorphic character classification result storage unit 5,
The correction operation by the recognition result correction unit 6 is controlled, or word-by-word re-correction is performed as necessary. The correction result thus obtained is stored in the recognition result storage unit 10
Is configured to be stored in.

Generally, the same character in the same document is printed in the same font and the same size, and the result read by the character reading unit 3 is recognized as the same shape. Therefore, the operation of collating and classifying such homomorphic characters can be performed relatively accurately. The device of the present invention is
By using such information, the same recognition result is obtained for the same characters in the same document.

The specific operation of the apparatus of the present invention will be described below in order. First, the character reading unit 3 compares a character pattern cut out from the document image 2 with a dictionary by pattern matching, similarly to that provided in a conventional apparatus, and performs a character recognition for each character. is there. FIG. 3 shows an explanatory diagram of such a character recognition operation.
For example, as shown in the figure, the character pattern is "Aly
When it is read as "sia", the feature quantities of these character patterns are extracted and the feature quantities are collated by referring to the standard character feature quantity dictionary, as already described with reference to FIG. Specifically, with the distance of the feature quantity as a parameter, the corresponding most appropriate candidate characters are listed. For example, in the example of FIG. 3, the character A has one candidate character, the character y has two candidate characters, and the other characters have three candidate characters. These are arranged in the order of the distance which is the difference between the feature amounts, and the one having the shortest distance is set as the first candidate. Therefore, the character reading unit 3
The first candidate of the candidate characters thus listed is output as a recognition result to the homomorphic character classification unit 4 and the recognition result correction unit 6 shown in FIG.

FIG. 4 shows an operation flowchart of the homomorphic character classification unit. The isomorphic character classification unit 4 operates according to the procedure shown in this figure. First, in step S1, the feature amount of each character is extracted. This feature extraction is
The homomorphic character classification unit 4 may perform the extraction independently, or the extraction result already performed by the character reading unit 3 may be used as it is. As a result, the processing speed can be improved and the efficiency can be improved.

Next, in step S2, each character in the same document is compared with the characteristic amount. That is, the character reading unit 3 extracts the feature amount for each character and performs character recognition by comparing each character with the dictionary, but the isomorphic character classification unit 4 determines the feature of each character in the same document. The quantities are compared with each other, and the same isomorphic character classification is given to each character having the same feature quantity (step S3). As described above, this comparison process has a much higher degree of matching than the comparison with the dictionary. Therefore, the classification can be performed extremely accurately. In step S4, it is determined whether or not all the characters have been classified, and the processes from step S1 to step S3 are repeated until all the characters are classified.

FIG. 5 is a diagram for explaining the recognition operation by isomorphic character classification. First, for each homomorphic character, for example, as shown in FIG. 5, a classification number based on, for example, a number is given. In this example, the character "s" is the classification number 1 and the character "y".
Is classified as 2, and the character “n” is classified as 3, and so on. In addition, in this example, it is assumed that four characters having the same classification number of 1 exist in the same document.
In the case of this example, it is assumed that there were three characters whose characters were recognized as “s” and were given a character code. The same applies to the following characters, and as for all the isomorphic characters, some are recognized as the same character and some are recognized as different characters such as the character "a".

Here, in the present invention, such a classification result of homomorphic characters is respected, and for all characters belonging to the same classification, the character code with the highest appearance is called the character code of the character of that classification. To obtain the only recognition result. For example, the classification number of homomorphic characters is 4
As for the character "a" which is, five homomorphic characters appear in the document. In this case, assuming that the number of cases where the character code is recognized as “a” is 3 and the number of cases where the character code is recognized as “o” and is 1 is 1, the character code of the character of the classification number 4 is “ a ”. Therefore, for example, the remaining one unrecognizable character is similarly recognized as "a". By adopting such a method, it is possible to obtain a more appropriate recognition result for each character.

[Third Invention] Next, the operation of the recognition result correction section 6 shown in FIG. 1 will be described. The recognition result correction unit 6
As in the conventional case, the recognition result output from the character reading unit 3 is accepted, the recognized characters are collated with the word dictionary in word units, and the correction work is performed. The specific operation is as already described in FIG. Here, for example, since homomorphic characters have the same characteristic amount, when the character reading unit 3 finds a character whose characteristic amount is closer than the correct recognized character, the recognition result may be all incorrect. Exists.

In view of this, in the present invention, when the recognition result correction unit 6 corrects a specific character by collating word by word and replacing the recognition result with a candidate word, the correction result is shown in FIG. It is reflected in the homomorphic character classification. That is, for example, the character “s” having the classification number 1 of the homomorphic character is recognized as the numeral “5” in the entire document, and the character is recognized as “5” by the majority decision by the process of the second invention. Even in such a case, when it is concluded that it is appropriate to recognize “s” by word matching, the recognition result is collectively corrected to “s”. As a result, the result of word matching or the like in the other part can be made more appropriate.

The correction control unit 9 reflects the above homomorphic character classification result in the word matching. Further, the modification control unit 9 further executes a process with high recognition performance by the following control. FIG. 6 shows an operation flowchart of the correction control unit. Further, FIG. 7 shows an operation explanatory view of the fourth invention. First, in FIG. 6, characters are read by the character reading unit 3 (step S1), word recognition is performed by the recognition result correction unit 6 in step S2, and in parallel, in the step S3, the isomorphic character classification unit 4 is executed. When the homomorphic character classification is performed, the character codes are associated based on the result (step S4). This association is the processing described with reference to FIG.

Further, using these results, the word matching operation in the recognition result correction section 6 is controlled in step S5. That is, the recognition result correction unit 6 determines whether or not there is a candidate word for each arbitrary word. Then, if there is a candidate word, the candidate word is selected using the homomorphic character classification in step S6. On the other hand, when there is no candidate word, in step S7,
Candidate characters are selected using homomorphic character classification.

The processing in step S6 of this figure is specifically the operation as shown in FIG. First, for example, for the character pattern “synaptic”, the result of the character recognition is “syncpti” as eight first candidate characters.
c ”. In this case, the candidate words are “synaptic” and “synapti” as shown in the figure.
c ”. In the conventional process comparison of feature amounts in word units, the former, that is, "synopic" is the first candidate word, and the latter is the second candidate word.

In the present invention, when each character code of this candidate word is focused on the homomorphic character classification, it is compared which one has a greater number of matches. At this time, the homomorphic character classification applies that already determined in FIG. Therefore, "a" is "c" as the first candidate character in this part.
However, the classification of the homomorphic characters corresponds to the classification number 4 shown in FIG. The character code is "a". Therefore, when comparing candidate words by focusing on such homomorphic character classification, the first candidate word “syno” is compared.
“Ptic” matches 7 characters, but the second candidate word “synaptic” matches 8 characters. As a result,
In the present invention, "synaptic" is used as the final recognition result. In this way, when the recognition results of homomorphic characters are the same, by selecting the most appropriate candidate word, a more reliable recognition result can be obtained.

Further, in FIG. 7, the character pattern "as
For “sociative”, the match between the actual character pattern and the first candidate character is extremely low. This is because only one candidate word can be obtained by the conventional method, so the recognition result of each character must be determined. Is possible.
Therefore, it is recognized that the seventh character "a" corresponds to the character code "a". This “a” has a homomorphic character classification of 4. Therefore, if the result is reflected in the matching of another word, for example, in the matching of "synaptic" described above, the candidate word can be selected similarly to the conclusion obtained previously. The character pattern "learnin" in FIG.
Also in "g", since the character "a" is recognized as corresponding to the character code "a", this isomorphic character classification is 4
It is possible to determine that it is more certain to recognize that a character is “a”.

On the other hand, the character pattern "Aply" shown in FIG.
For “sia”, no matching candidate word is found, and therefore word matching has failed. Conventionally, in such a case, the recognition result is output as it is. However, according to the invention, the character pattern "associative"
Depending on the processing result of the second or third character for "e",
It is determined that the character whose character classification is 1 is “s”.
Therefore, based on this result, it can be concluded that the fifth character of "Alysia" is "s". As for the sixth character, it can be concluded from the processing results of other character patterns that it is "i". As a result, even when word matching is impossible, more appropriate character recognition can be performed using the isomorphic classification result.

The present invention is not limited to the above embodiments. The above recognition processing device may be provided with hardware having the above function for each block, for example, or may be configured by a processing program or the like that performs the operations described above according to conditions. It doesn't matter. Further, it is possible to execute similar processing not only for English word matching but also for Japanese.

[0028]

According to the character recognition processing apparatus and the recognition processing method of the present invention described above, the characters included in the document image are recognized, and the recognized characters are classified into homomorphic characters having similar feature amounts. , When the recognized characters are collated with the word dictionary and corrected in word units, the classification result of the character classification is used, and the same recognition result can be obtained for the homomorphic characters, so it is more reliable than before. It becomes possible to perform character recognition processing. In addition, even if a candidate word cannot be obtained by word matching, it is possible to obtain the most appropriate recognition result by using another word matching result or the like.

Further, even if different recognition results are obtained for a plurality of characters of the same classification, the recognition error can be reduced by obtaining the unique recognition result by the majority vote. Furthermore, if a certain character is determined to be erroneously recognized by collating the word dictionary, the recognition result when the character collation is not possible can be improved by correcting the recognition results of the character and the isomorphic character collectively. Not only that, the reliability of other word matching can be improved. Furthermore, if the recognition results of homomorphic characters are the same, selecting the most appropriate candidate word will improve the reliability of the selected font of the candidate word as a whole and improve the post-processing performance of character recognition. To do.

[Brief description of drawings]

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

FIG. 2 is a flowchart of a conventional character recognition operation.

FIG. 3 is a diagram for explaining a character recognition operation.

FIG. 4 is an operation flowchart of a homomorphic character classification unit.

FIG. 5 is a diagram illustrating a recognition operation based on homomorphic character classification.

FIG. 6 is an operation flowchart of a correction control unit.

FIG. 7 is an operation explanatory diagram of the fourth invention.

[Explanation of symbols]

 3 Character reading unit 4 Homomorphic character classification unit 5 Homomorphic character classification result storage unit 6 Recognition result correction unit 7 Word dictionary 8 Candidate word storage unit 9 Correction control unit 10 Recognition result storage unit

Claims (4)

[Claims] 1. A character reading unit for optically reading a document image and converting it into an electric signal for character recognition of characters included in the document image; and the recognized characters classified by homomorphic characters having similar characteristics. A homomorphic character classification unit, a recognition result correction unit that matches the recognized characters with a word dictionary, and corrects the character recognition result on a word-by-word basis, and correction of each character by the recognition result correction unit is performed by the homomorphic character classification unit. A character recognition processing device, comprising: a correction control unit that controls the same character to obtain the same recognition result according to the classification result of the unit. 2. When the characters included in the document image are recognized, the same characters having similar feature quantities are classified, and different recognition results are obtained for a plurality of characters of the same classification. A character recognition processing method characterized by obtaining the only recognition result by. 3. Character recognition of characters included in a document image is performed, classification is performed for each homomorphic character having a similar feature amount, the recognition result is collated with a word dictionary for each word, and the recognition result is a candidate word. When the specific character is determined to be erroneously recognized, the character recognition processing method is to collectively correct the recognition results of the specific character and the isomorphic character. 4. Character recognition of characters included in a document image is performed, classification is performed for each isomorphic character having a similar feature amount, the recognition result is collated with a word dictionary for each word, and the recognition result is a candidate word. A character recognition processing method characterized by selecting a candidate word that is considered to be the most appropriate when the recognition results of the same-shaped characters are the same when replacing with.