KR100258923B1 - Hangeul and english name recognition and error correcting method - Google Patents

Abstract

PURPOSE: A method for recognizing Korean character and English names and correcting an erroneous recognition is provided to compare the results output from a Korean character and English recognizers with each other, and then automatically correct an erroneously recognized result, when a Korean character name and an English name of the same person are provided on a document. CONSTITUTION: An image that includes the Korean character and English names of the same person is inputted and processed(S21). An image corresponding to the Korean character name is extracted from the inputted image, and the extracted image is subjected to a Korean character recognizer after preprocessing(S22). The Korean character name is recognized(S23). An image corresponding to the English name is extracted, preprocessed, and then provided to an English recognizer(S24). The English name is recognized(S25). A recognition result of English description, which corresponds to a candidate character from a Korean character name recognizer, is extracted from a conversion table, and thereby the erroneous recognition is corrected(S27). The Korean character and English names corrected are output(S28,S29).

Description

Korean and English name recognition and misrecognition correction method

1 is a flowchart illustrating a conventional method of recognizing Korean and English names.

2 is a flowchart illustrating a method of recognizing a Korean name and an English name according to the present invention.

3 is a flowchart illustrating a method for correcting misrecognition of Korean and English names according to the present invention.

The present invention relates to a character recognition method, and more particularly, to a method for correcting recognition and misrecognition of Korean and English names of the same person input together.

As the use of computers has become commonplace, easy interface functions with humans are required. Representative means for performing such interface functions are input devices and output devices. Among them, the means by which humans directly express their intentions is an input device.

However, most existing computers use a keyboard as an input device, and thus there is a limit in processing a rapidly increasing amount of information. Therefore, there is a demand for an environment that can process information quickly by having a function similar to that of the five senses of a human, that is, directly recognizing a character or recognizing a voice.

Character Recognition is a field of pattern recognition. It uses preprocessing step to perform extraction, normalization and skeletalization of the text image from the image input through the image input device. This is done through a postprocessing step that corrects misrecognized characters.

As a post-processing step, a bottom-up method based on stochastic representation of knowledge, a top-down method based on structural expression of contextual knowledge, and a combination of bottom-up and top-down according to the method of expressing contextual knowledge. Hybrid methods exist. Representative algorithms of the bottom-up method are Viterbi Algorithm (also abbreviated as VA), and representative algorithms of the top-down method are dictionary look-up algorithm (DLA) and binary N-. The gram algorithm (BNA) is known, and the combined methods include the dictionary viterbi algorithm and the predictor-corrrector algorithm (PCA).

On the other hand, most automatic input methods using optical character recognition (OCR) technology generally recognize one or more languages as targets. However, since the number of characters to be recognized and the recognition method are different for each language, the existing automatic input method is implemented to use an independent recognizer for each language.

FIG. 1 is a flowchart illustrating a conventional method of recognizing a Korean name and an English name. In step S11, a document or a journal in which a Korean name and an English name of the same person are recorded is inputted through an image input device such as a scanner. If the input processing is performed as an image including the full name, the Korean name image is extracted from the input image using the Hangul recognizer in step S12, and the preprocessing is performed. After the recognition process, the Hangul name recognized in step S14 is output.

In the case of the English name, the English name image is extracted from the image input in step S15 to perform preprocessing, and the corresponding English name is recognized in the preprocessed English name image using the English recognizer in step S16. After that, the English name recognized in step S17 to be output.

However, in general, the character recognition performance for a language having a small number of characters is higher than the character recognition performance for a language having a large number of characters. Therefore, the character recognition performance for English characters may be higher than the character recognition performance for Korean characters. If the recognizer used in FIG. 1 has higher character recognition performance than the character recognition performance for the Korean characters as described above, the character recognition result for the Korean name is higher than the character recognition result for the English name. This can be high. However, the recognizer used in FIG. 1 may have a higher character recognition performance for Korean characters than a character recognition performance for English characters. In this case, the recognition rate of the English recognizer may be higher than that of the Korean recognizer.

In order to reduce the false recognition rate in each situation that can be generated by using the recognizers independently, a method of directly calibrating a person by referring to the output of the recognizers for each language has been used.

Accordingly, an object of the present invention is to correct a misrecognition method that automatically corrects a misrecognized result by cross-referencing a result outputted by a Korean recognizer and an English recognizer when a Korean name and an English name exist for the same person in a document or a document. To provide.

Another object of the present invention is to provide a statement recognition method using the above miscorrection correction method.

In order to achieve the above object, the misrecognition correction method according to the present invention, by inputting the image of the name described in the first letter and the second letter, the recognition result of the name described in the second letter of the name misrepresentation of the first letter A method for calibrating with reference to the method comprising: a first step of generating a representation of a second character from candidate characters of a name recognition result of the first character; Deleting a representation of the second character outside the recognition length of the name recognizer of the second character from the generated representation of the second character; Removing a character not included in the name recognition candidate set of the second character at the same position from the generated second character representation; A fourth step of aligning the notation of the second character that has passed through the third step according to the recognition reliability only when the performance of the name recognizer of the second character is higher than that of the name recognizer of the first character; A fifth step of converting the representation of the second character, which has passed through the fourth step, into a first character; A sixth step of removing the candidate character set of the first character out of the notation of the converted first character if not present; A seventh step of sorting according to the recognition reliability of the first character only when the performance of the second character is lower than that of the first name character recognizer; And selecting the candidate characters having the highest reliability among the first characters having passed through the seventh step and the candidate characters of the corresponding second characters as the first candidates of the first and second characters, respectively. .

Next, the present invention will be described in detail with reference to the accompanying drawings.

First, the present invention uses an image including both a Korean name and an English name for the same person as an input object. There must be at least one Korean recognizer and one English recognizer capable of outputting more than one recognition result and corresponding recognition reliability for each character, and the conversion table for the phoneme of the Hangul name and the corresponding English notation There must be a rule verifying that the result of an English statement is a legal English notation.

In addition, the correction by cross-reference between the recognition result of the Korean name and the recognition name of the English name is not based on the entire name, but on the English alphabet corresponding to one letter of the Korean name. Therefore, for statement post-processing, we need a table of which alphabets correspond to the phonemes used in Korean names. Here, the statement post-processing is a process of correcting misrecognized characters by cross-referencing the recognition result of the Korean name and the English name when both the Korean name and the English name field exist simultaneously in the document.

2 is a flowchart illustrating a method for recognizing a Korean name and an English name according to the present invention, and in step S21 to S11 of FIG. Input processing is performed on an image containing both Korean and English names for the same person.

In step S22, the image corresponding to the Hangul name is extracted from the input document image, and the extracted image is preprocessed through a process such as normalization and provided as an input image of a Hangul recognizer (not shown). Accordingly, the Hangul recognizer (not shown) performs the recognition process on the input image in the same manner as in the previous step S23, and at least one (10 in the embodiment of the present invention) 'candidate characters' and' Output the recognition reliability.

The recognition reliability is output in different forms according to the recognition method adopted by the Hangul recognizer. For example, when the Hangul recognizer adopts the recognition method by the minimum distance pattern classification among the statistical recognition methods, Euclid between the input character image and the candidate character The distance (Euclidean distance) is output as the reliability of the candidate character. At this time, the closer the Euclidean distance is, the higher the candidate character becomes. In addition, when the Hangul recognizer adopts the character recognition method based on the probability, the probability value that the input image belongs to the candidate character is output as the reliability of the candidate character. At this time, the closer the probability value is to 1, the more candidate characters are for reliability. Meanwhile, in step S24, an image corresponding to the English name is extracted from the input document image, and the extracted image is provided as an input image of the English recognizer after preprocessing such as normalization. Accordingly, the English recognizer outputs one or more candidate texts and corresponding recognition reliability for the input image in step S25. The output recognition reliability is output in a form determined according to the recognition method adopted by the English recognizer as described in step S23 described above.

In step S27, if the first candidate character among the candidate characters provided from the Korean name recognizer and the first candidate character among the candidate characters provided from the English name recognizer are not the same, the candidate characters of the Korean name recognizer The recognition result of the corresponding English notation is detected in the conversion table 50, and the misrecognition is corrected.

In step S28, the corrected Korean name is output in the same manner as before, and in step S29, the corrected English name is output in the same manner.

3 is a flowchart illustrating a method for correcting misrecognition of Korean and English names according to the present invention.

First, the process of inputting the 'Hangul full name candidate' recognized by the Hangul recognizer and the corresponding 'recognition reliability' in step S31 is corrected for misrecognition. 'Recognition Reliability' is processed so that it is input for misrecognition correction.

Then, it is verified whether the English name candidate character input in step S33 is a legitimate English notation of the Korean name. That is, the conversion table 50 is verified to verify whether the character string of the English name candidate character represents one character of a legitimate Korean name.

For example, when a table for verification as shown in Table 1 exists in the conversion table 50, the initial mapping function of the English name candidate character whose string is u {S, I, K} outputs 'ㅅ', Since the mapping function outputs 'ㅣ' and the final mapping function outputs 'ㄱ', this English name candidate character is considered to be a legitimate English notation, and the string is u {S, Y, I, G}. Since the initial mapping function of the letter outputs 'ㅅ', the neutral mapping function outputs 'ㅣ', and the final mapping function outputs 'ㄱ', this candidate name is also considered to be a legitimate English representation.

TABLE 1

However, in the case of English name candidate characters with the string u {S, Y, A, G}, the initial mapping function outputs 'ㅅ', the neutral mapping function outputs 'not applicable', and the final mapping function Since 'ㄱ' is output, it is determined to be an illegal English notation, and in the case of an English name candidate character whose string is u {S, Y, I, Q}, the initial mapping function outputs 'ㅅ', and the final mapping The function outputs 'ㅣ' and the final mapping function outputs 'Not applicable', so it is considered to be an illegal English notation.

In this case, if any of the mapping functions of the initial, neutral, and final character of the string is not applicable, the candidate character having the illegal English notation is determined.

Next, in step S34, it is compared whether the English name of the first candidate character is the same as the Korean name recognition result of the first candidate character. That is, one Korean character is generated by combining the results of the mapping functions of the initial, neutral, and final characters of the English name string corresponding to the first candidate character with reference to the conversion table 50. For example, the string of English name candidate character is u {S, I, K}, the output of initial mapping function is 'ㅅ', the output of neutral mapping function is 'ㅣ', and the output of final mapping function is If 'a', then the output of these mapping functions is combined to form an expression. One Korean character generated as described above is compared with one Korean character corresponding to the first candidate character. If the comparison result is the same, the misrecognition of the character is corrected to correct the misrecognition of the next character.

However, if not the same, perform the steps S35 to S42 to correct the misperception.

That is, in step S35, the English notation corresponding to all candidate characters applied as the Korean name recognition result is generated using the conversion table 50. For example, 'S' and 'SH' are the English letters corresponding to the initial consonant 'ㅅ', and 'I' and 'YI' are the English letters corresponding to the neutral 'ㅣ'. If the 'K' and 'G' exist as the English notation corresponding to the final word 'ㄱ', the generated English notation is 'SIK', 'SIG', 'SYIK', 'SYIG', 'SHYIK', 'SHYIG' 'Eight things will be. The conversion table 50 is provided with a Korean name notation and a corresponding English name notation as shown in Table 1 above so that the Korean name notation and the English name notation can be converted to each other for reference.

In step S36, the length of the generated English notation is not the same as the length of the English notation of the recognized image. In step S37, all the alphabets of the generated English notation are not present in the candidate character set of the English recognizer at the same position. In step S38, the generated English notation remaining in steps S36 and S37 is sorted based on the sum of the reliability of the candidate alphabets matched in step S37.

However, such alignment is optional. In other words, if the recognizer to be used has the recognition performance of the entire English name notation less than that of a single Hangul statement, the process proceeds directly to step S39. If the recognition performance of the recognizer to be used is reversed, the reliability is as described above. Sort the English notation accordingly.

In operation S39, the generated English notation is converted back to Korean using the conversion table 50. In step S40, if the converted character does not exist in the candidate character set of the Hangul name recognition result, it is removed. If present, in step S41, the candidate is sorted based on the Hangul reliability. However, in step S41 it is optionally made as in step S38 described above. That is, if the performance of the recognizer to be used is higher than the recognition performance of a single Korean name, the recognition performance of the entire English name notation is not performed. However, when the performance of the recognizer used is reversed, alignment is performed according to Hangul reliability in step S41.

In step S42, after selecting and storing the Hangul character candidate character and the corresponding English expression of the highest reliability among the sorted candidate characters as the first candidate character, the misrecognition correcting operation for the corresponding character ends. This process applies to all characters in the Hangul statement.

Table 2 is an example to explain the correction of Korean character recognition using 10 English recognition results.If the input Korean image is 'volume', the Korean name recognition result recognized correctly exists in the 4th candidate character but corresponds to this. In the English expression 'KWON', all alphabets are present in the first candidate character. Therefore, when all candidate characters in the Korean name recognition result are converted into English expressions corresponding to these by the misrecognition correction process, the expression corresponding to 'KWON' has the highest reliability, and if it is converted into Korean again, The first candidate character of the Hangul name is corrected from 'circle' to 'volume' because it exists in the candidate character set of the Korean name recognition result. In Table 2, the value placed next to each candidate character is the 'recognition reliability' of the character, and the lower the value, the higher the reliability.

TABLE 2

Table 3 is an example to explain the correction of the English character misidentification using 10 Korean recognition results. The Korean name recognition result correctly recognized for the input image 'Equation' exists in the first-class candidate character, and the corresponding English expression ' The recognition result of 'I' in SIG 'is present in the second candidate, but is corrected to the first candidate through the misrecognition correction process.

TABLE 3

As described above, the method of correcting the misperception by cross-referencing the recognition results of the Korean name and the English name can be established between any two characters. In particular, the above-described method can be applied between a Hangul statement with a superior recognition function and a Hanja statement with a relatively inferior function.

By using the misrecognition correction method of the present invention, the misrecognition can be easily corrected according to the performance of the recognizer to be used. That is, the recognition of the recognition can be easily corrected to be operated at the recognition rate of the superior side depending on whether the performance of the recognizer is superior to Korean or English.

Claims (6)

A method of correcting a misrecognition of a name written in the first letter by referring to a recognition result of the name written in the second letter by inputting an image of the name written in the first and second letters, the name of the first letter. A first step of generating a representation of the second character from candidate characters of the recognition result; Deleting a representation of the second character outside the recognition length of the name recognizer of the second character among the generated representations of the second character; Removing a character not included in the name recognition candidate set of the second character at the same position among the generated second character representations; A fourth step of aligning the notation of the second character that has passed through the third step according to the reliability of recognition only if the performance of the name recognizer of the second character is higher than that of the name recognizer of the first character; A fifth step of converting the notation of the second character that has passed through the fourth step back into the first character; A sixth step of removing the candidate character set of the first character from among the notation of the converted first character if not present; A seventh step of sorting according to the recognition reliability of the first character only when the performance of the name recognizer of the second character is lower than that of the name recognizer of the first character; And selecting candidate candidates having the highest reliability among the first characters having passed through the seventh step and candidate characters of the corresponding second characters as the first candidates of the first and second characters, respectively. Misconception correction method of the statement characterized by. The method of claim 1, wherein the first character is Korean and the second character is English. The method of claim 1, wherein the first character is Chinese and the second character is Korean. A name recognition method of correcting a misrecognition of a name of the first character by referring to a recognition result of the name of the second character by inputting the image of the name represented by the first and second characters, the image of the name A first step of extracting and preprocessing the name image of the first character from the first image; A second step of recognizing a name of the first character from a name image of a preprocessed first character; A third step of extracting and preprocessing the name of the second character from the image of the name; A fourth step of recognizing the name of the second character from the name image of the pre-processed second character; A fifth step of verifying a representation of a name of the second character; A sixth step of comparing a name of the first character recognized in the second step with a verification result of the second character transmitted through the fifth step, and correcting misrecognition of the first character if it is not the same; And a seventh step of outputting the corrected names of the first and second characters, respectively. The method of claim 4, wherein the first character is Korean and the second character is English. The method of claim 4, wherein the first character is a Chinese character, and the second character is a Korean character.