KR20040051437A - Performance improving method of grapheme based hangul recognizer - Google Patents

Abstract

PURPOSE: A method for improving the performance of a character recognizer based on a grapheme is provided to prevent the wrong recognition caused by the wrong recognition of a vowel. CONSTITUTION: A grapheme combination pattern for a character image included in a document image is classified(502). It is judged that a final consonant is included in the character image based on the classified grapheme combination pattern(503). The grapheme included in the character image is classified based on a judging result. While the first consonant, a middle consonant, and/or the final consonant are classified based on the judging result, the recognition of the middle consonant is executed based on a recognition reliability of the first and/or the final consonant. The recognition results of the first, the middle, and the final consonant are combined(517).

Description

How to improve the phoneme-based character recognizer {PERFORMANCE IMPROVING METHOD OF GRAPHEME BASED HANGUL RECOGNIZER}

The present invention relates to a method for improving a phoneme-based character recognizer. In particular, in recognizing a Hangul character image existing in a document image acquired through an image input device such as a scanner or a camera, it is possible to prevent the recognition of a character due to a misrecognition of a vowel. It's about how to make that happen.

Typically, Hangul has a large number of objects to recognize 11,172 because 19 letters, 21 neutrals and 27 finalities are composed of two-dimensional combinations of letters.

In addition, if only the characters with high frequency of use are recognized as objects, there are more than 1,000 to 2,000 types, so it is very difficult to implement a recognition system with high recognition rate. This is mainstream.

1 is a view illustrating a conventional method of recognizing a Korean character, and after receiving a character image to be recognized through an individual character image extraction step from a document image acquired by an image input device such as a scanner or a camera (step) 1), the combination type of the phonemes constituting the input text image is classified (step 2). Subsequently, partial images are separated for each phoneme by using the combination location information of the phonemes obtained by performing the phoneme combination type classification process (step 3), and feature value vectors are extracted from the separated phonemes, respectively. Independently recognize the phoneme of (step 4), and obtain the final recognition result by combining the recognized phonemes (step 5).

As described above, the conventional character recognition method independently recognizes each phoneme, and combines the phoneme recognition results. Therefore, if one of the phonemes constituting the letter is misrecognized, a character recognition error inevitably occurs. Here, Hangul phoneme has a consonant consonant, a neutral vowel, a final consonant, among which the most common misrecognition is a neutral vowel.

When the neutral vowels frequently misrecognized are described in more detail, the recognition rate of the vowels lower than the consonants can be explained by two reasons.

In other words, vowels are less redundant than information. Consonants can be deduced relatively easily from the original consonant, no matter how much loss or deformation of the vowel occurs (to some extent the information is lost).

However, vowels are distinguished by one small stroke in Korean alphabet. For example, ,, ㅓ, and ㅕ are the same as the vertical strokes, but are distinguished by the number of short horizontal strokes, so that a slight deformation in the phoneme greatly increases the possibility of misrecognition. This is especially true for unusual fonts, low resolution and low quality characters.

Next, the degree of deformation of the vowel image at the normalization of the phoneme image is very large compared to the degree of deformation of the consonant image. In general, the phoneme is normalized linearly or nonlinearly before recognition, and the feature vector of a certain dimension is extracted and recognized. In this process, the vowel deformation is very large compared to the consonant, so that the original vowel is transformed into a completely different vowel. As a result, the recognition rate of the vowels is relatively lower than the recognition rate of the consonants.

As such, the character recognition techniques include "Recognition of Printed Korean Characters Using the Improved Phoneme Recognition Method," published in the Journal of the Korean Information Science Society in August 1996, and "Hangul Character Recognition Method" registered as 0285765 on January 5, 2001. And "Method for Recognizing Korean Characters of Low-Quality Characters Including Joong-young" registered on October 20, 1999, 0239354.

As described above, when the disclosed prior art is described in detail, "high recognition rate printed Hangul recognition through an improved phoneme recognition method" is one of six types of Hangul after extracting a feature value vector for an input character image. After classifying as, separate the phonemes using the type information. In order to recognize the separated phoneme, the feature vector of the phoneme is extracted, and the phoneme is recognized through a phoneme recognition neural network having the input. Recognition of each phoneme is made independently of each other, and final character recognition is achieved by combining the recognized phonemes.

Since the separation of phonemes uses the classified Hangul type information, it is difficult to separate only the correct phoneme area. Therefore, the phoneme area is expanded to include the phoneme area sufficiently even if the miso is included. In order to recognize the separated phonemes, we extract the mesh feature of the phonemes.

The mesh feature is defined as the number of black pixels existing in the partial region by dividing the recognition target region into partial regions. The phoneme is recognized by inputting the mesh feature for the separated phoneme into the corresponding phoneme-aware neural network among 17 phoneme-aware neural network recognizers. Recognizes all phonemes in a letter and combines them to complete the final letter.

Next, the "Hangul character recognition method" determines the character recognition method to recognize each input character in the unit of phoneme or character unit, and if it is determined as the recognition method of phoneme separation, Recognize them separately and combine the recognized phonemes to output the characters, or if it is determined by the character-based recognition method, the input characters are recognized by character units and the recognized characters are output.

Application of the character unit method and character unit recognition method extracts horizontal vowels, vertical vowels and fragmented strokes, recognizes vowels using horizontal vowels and vertical vowels, and then identifies the vowel position, size and number of vowels If it compares with each other, it is judged that the vowels are recognized correctly. If it does not match, the vowel recognition is difficult.

Horizontal collection, vertical collection and fragment information which are information for determining the recognition method are composed of the following process. The horizontal vowel extraction is performed by extracting the horizontal vowel existence region, extracting the horizontal line in the region, and determining the existence of the up-down projection image. Similar to the horizontal vowel extraction method, the vertical vowel extraction method extracts the vertical vowel presence area, then extracts the vertical line, determines the left and right projection images, and then extracts the vertical vowel. Extraction of fragmented stroke is a technique of detecting inner and outer contours and extracting fragmented stroke information by using a contour tracking algorithm.

Next, the method for recognizing the Hangul of low-quality characters including Jyoyoung includes the steps of separating phonemes from the input text image, and comparing the stroke thickness and the change value between the strokes constituting the phonemes. Extracting mesh eyes from the phoneme, extracting the stroke change feature in the eye, and recognizing the phoneme by comparing the similarity between the stroke change feature vector and the reference feature vector, and combining the recognized phonemes It is composed of steps.

In order to extract the mesh from the isolated phoneme, we extract the vertical and horizontal strokes of the phoneme, and compare the position and size with the neighboring strokes and repeat the process until the stroke thickness and the change value are over a certain value. .

The similarity with the reference feature vector is compared using the stroke change in the mesh as the feature vector. In this process, in order to make the similarity, which is the recognition criterion, insensitive to the change of stroke position, weight is given to the area to be compared.

As a result, it extracts the feature of the stroke change from the mesh of the nonlinear interval automatically formed according to the change of the stroke, obtains the distance difference with the neighboring areas, and determines the similarity based on the distance difference, and recognizes and combines the characters.

As described above, when looking at the technology disclosed in the prior patent, each phoneme is independently recognized and the phoneme recognition result is combined, so that if one phoneme among the phonemes constituting the letter is misrecognized, a character recognition error will inevitably occur. The problem still remains.

Accordingly, the present invention has been made to solve the above-described problems, the purpose of which is to recognize the Hangul character image present in the document image obtained through an image input device such as a scanner or camera, The present invention provides a method for improving a phoneme-based character recognizer to prevent misrecognition.

According to an embodiment of the present invention, there is provided a method of improving a phoneme-based character recognizer. The method of classifying a phoneme combination type for a text image included in a document image may include: A phoneme classification step of classifying the phonemes included in the character image based on the result of the judgment of the presence or absence of the species, and the phoneme classification step based on the result of the judgment of the species presence judgment step Recognizes the consonants, neutral vowels and / or final consonants classified in, wherein the recognition of the vowels is performed based on the recognition reliability of the consonants and / or the final consonants. Comprising a phoneme combining step of combining the recognition results of the consonants, neutral vowels and / or consonants It characterized.

In addition, the present invention for achieving the above object is characterized by having a recording medium recording a program implementing the method of improving the phoneme-based character recognizer performance.

1 is a view showing a conventional Hangul character recognition method process,

2 is a block diagram for performing a method for improving a phoneme-based character recognizer according to the present invention.

3 is a view showing in detail the recognition unit shown in FIG.

4 is a view connecting two vectors according to an embodiment of the present invention.

5 is a detailed flowchart illustrating a method of improving a phoneme-based character recognizer according to the present invention.

<Description of the symbols for the main parts of the drawings>

10: character image input unit 20: phoneme combination type classification unit

30: final judgment unit 40: image separation unit

50: recognition unit 60: coupling unit

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

2 is a block diagram for performing a method for improving a phoneme-based character recognizer according to the present invention. The character image input unit 10, the phoneme combination type classification unit 20, the finality determination unit 30, and an image are shown. The separator 40, the recognition unit 50, and the coupling unit 60 are included.

The text image input unit 10 inputs the Hangul text image existing in the document image acquired from the image input device such as the scanner S1 or the camera S2 to the phoneme combination type classification unit 20.

The phoneme combination type classification unit 20 classifies the combination type of phonemes constituting the Hangul character image input from the text image input unit 10 and provides the result to the finality determination unit 30.

The finality determination unit 30 determines whether the finality of the phonemes classified and provided by the phoneme combination type classification unit 20 is provided to the image separation unit 40.

The image separator 40 separates the image of the consonant consonant and the neutral vowel when there is no finality according to the finality judgment of the finality determination unit 30, and has a finality according to the finality determination of the finality determination unit 30. In this case, the image of the initial consonant and the neutral vowel and the final consonant are separated and provided to the recognition unit 50.

As shown in FIG. 3, the recognition unit 50 includes an initial consonant recognition unit 51 and a final consonant recognition unit for recognizing the primary consonants, the final consonants, and the neutral vowels separated by the image separator 40. 52 and a neutral vowel recognition unit 53 are provided.

The first consonant recognizer 51 calculates a feature vector of the first consonant, recognizes the first consonant using the first consonant, calculates the recognition reliability of each first consonant, and selects and combines the first consonant having the highest reliability as the recognition result. It provides to the part 60.

In other words, the recognition reliability of each consonant is denoted by x _i , and the vector having the component is denoted by the consonant reliability vector X. I is an index of individual consonants, i may be "a" when 1, "2" when 2, and "ㅎ" when 19 is the last index, which is the same as Equation 1.

Where Nx is the number of consonants.

The final consonant recognizer 52 calculates a feature vector of the final consonant, recognizes the final consonant using the final consonant, calculates the recognition reliability of each final consonant, and selects and combines the final consonant having the highest reliability as a recognition result. It provides to the part 60.

That is, the recognition reliability of each final consonant is represented by z _i , and the vector having the component is represented by the final consonant reliability vector Z. Where i represents the index of the individual final consonants in the same way as the initial consonants.

Where Nz is the number of final consonants.

The neutral vowel recognition unit 53 calculates a feature vector of the neutral vowel, and varies recognition depending on the presence or absence of the final vowel. First, in the absence of the final trait, the consonant consonant reliability vector and the neutral vowel feature vector are connected. As shown in FIG. 4, the vowels are connected, the neutral vowels are recognized using the connected vector, the recognition reliability of each neutral vowel is calculated, and the neutral vowels having the highest reliability are selected and provided to the combiner 60.

Here, the neutral vowel feature vector is represented by F , and f _i is an i- th feature value, which is a positive value greater than zero.

In addition, the initial consonant recognition result is used as a feature value in recognizing a neutral vowel, and generates a recognition reliability vector G by Equation 3. Where G is to be.

That is, the recognition result reliability (x _i ) of the initial consonants is normalized so that their sum is 1 (divided by the sum of total reliability values), and then multiplied by the maximum value of the neutral vowel feature values (max f), thereby making the initial consonant reliability The effective section of is equal to the valid section of the neutral vowel feature.

In this way, the reliability of the initial consonant and the feature value of the neutral vowel are normalized to obtain a new feature vector that is not biased against the reliability value or the feature value. The reliability g _i of the newly changed consonants is a component of the reliability vector G. Then, the vectors G and F are concatenated to calculate the recognition reliability of each neutral vowel, and the neutral vowel having the highest confidence is selected.

Next, when there is a finality, the initial consonant reliability vector, the final consonant reliability vector, and the neutral vowel feature vector are connected, and the connection is connected as shown in FIG. 4, and the neutral vowel is recognized using the connected vector. In addition, the recognition reliability of each neutral vowel is calculated, and the neutral vowel having the highest reliability is selected and provided to the combiner 60.

That is, it recognizes primary and final consonants with a relatively high recognition rate and uses this information to recognize neutral vowels. The input of the neutral vowel recognition is the concatenation of the initial consonant reliability vector G , the final consonant reliability vector H (calculated in the same way as G ), and the neutral vowel feature vector F. Thus, the recognition reliability of each neutral vowel is calculated, and the neutral vowel having the highest reliability is selected.

The component of the final consonant reliability vector H is shown in Equation 4. Where H is to be.

Here, z _i represents the recognition reliability of the i-th final consonant, Nz is the number of the final consonants, f _j ≥ 0, Denotes the largest component of the feature vector of the neutral vowel, wherein h _i is the changed reliability and the i th component of the final consonant reliability vector H.

The combiner 60 combines the phonemes respectively recognized by the recognizer 50 to perform final character recognition.

Referring to the flowchart of FIG. 5, the method of improving the phoneme-based character recognizer according to the present invention having the above-described configuration will be described in detail.

First, the character image input unit 10 that obtains a document image from an image input device (for example, a scanner S1 or a camera S2, etc.) is configured to display a Hangul character image existing in the input document image. Enter in step 20 (step 501).

The phoneme combination type classification unit 20 classifies the combination types of the phonemes constituting the Hangul character image input from the text image input unit 10 and provides them to the finality determination unit 30 (step 502).

The finality determination unit 30 determines whether the finality of the phonemes classified and provided by the phoneme combination type classification unit 20 is present (step 503).

If there is no finality of the phoneme in the determination step 503, after separating the initial consonant and the image of the neutral vowel in the image separation unit 40, the initial consonant is transmitted to the initial consonant recognition unit 51 in the recognition unit 50. The neutral vowel is provided to the neutral vowel recognition unit 53 in the recognition unit 50 (step 504).

On the other hand, if there is a finality of the phoneme in the determination step 503, after separating the image of the initial consonant and the neutral vowel and the final consonant in the image separation unit 40, the initial consonant is transmitted to the initial consonant recognition unit 51. The neutral vowel is provided to the neutral vowel recognition unit 53, and the final consonant is provided to the final consonant recognition unit 52 (step 505).

The consonant consonant recognizer 51 calculates the feature vector of the consonant consonant provided separately from the image separator 40, recognizes the consonant consonant using the consonant consonant, calculates the recognition reliability of each consonant consonant, Branch is selected as a recognition result of the consonant consonants provided to the coupling unit (60). In other words, the recognition reliability of each consonant is denoted by x _i , and the vector having the component is denoted by the consonant reliability vector X. I is an index of individual consonants, i may be "a" for 1, "2" for 2, and "ㅎ" for 19, which is the same as Equation 1 above. (Step 506).

The final consonant recognizer 52 calculates a feature vector of the final consonant provided separately from the image separator 40, recognizes the final consonant using the same, calculates the recognition reliability of each final consonant, The branch is selected as a recognition result of the final consonant and provided to the coupling unit 60. That is, the recognition reliability of each final consonant is represented by z _i , and the vector having the component is represented by the final consonant reliability vector Z. Here i represents the index of the individual final consonants in the same way as the initial consonants, which is the same as Equation 2 (step 507).

The neutral vowel recognition unit 53 calculates a feature vector of the neutral vowel separated from the image separator 40 and checks whether there is a finality in recognition of whether the recognition is different according to the presence or absence of the final vowel (step 508).

If there is no finality in the check step 508, the consonant reliability vector and the neutral vowel feature vector are concatenated, and the concatenation is concatenated (step 509) and neutral using the concatenated vector, as shown in FIG. The vowel is recognized (step 510), the recognition reliability of each neutral vowel is calculated (step 511), and the neutral vowel having the highest reliability is selected and provided to the combiner 60 (step 512).

Here, the neutral vowel feature vector is represented by F , f _i is a feature value of the i- th neutron vowel and is a positive value greater than 0, and the initial consonant recognition result is used as a feature value in recognizing the neutral vowel, The recognition reliability vector G is generated by Equation 3. That is, the recognition result reliability (x _i ) of the consonant consonants is normalized to be their sum 1 (divided by the sum of the total reliability values), and then multiplied by the maximum value of the neutral vowel feature values (max f), The effective section is made equal to the valid section of the neutral vowel feature.

In this way, the reliability of the initial consonant and the characteristic value of the neutral vowel are normalized to obtain a new feature vector that is not biased against the reliability value or the feature value. The reliability g _i of the newly changed consonants is a component of the reliability vector G. Then, the vectors G and F are concatenated to calculate the recognition reliability of each neutral vowel, and the neutral vowel having the highest confidence is selected.

If there is a finality in the check step 508, the initial consonant reliability vector and the final consonant reliability vector and the neutral vowel feature vector are connected, and the connection is connected as shown in FIG. 4 (step 513). Recognize the neutral vowels (step 514), calculate the recognition reliability of each neutral vowels (step 515), and select the neutral vowels having the highest reliability and provide them to the combiner 60 (step 516).

That is, it recognizes primary and final consonants with a relatively high recognition rate and uses this information to recognize neutral vowels. The input of the neutral vowel recognition is concatenated by the initial consonant reliability vector G and the final consonant reliability vector H (calculated in the same way as G ) and the neutral vowel feature vector F. Therefore, the recognition reliability of each neutral vowel is calculated, and the neutral vowel having the highest reliability is selected, and the component of the final consonant recognition reliability vector H is expressed by Equation 4 described above.

The combiner 60 combines the phonemes respectively recognized by the recognizer 50 to perform final character recognition (step 517).

In the present invention described above, in recognizing a Hangul character image existing in a document image acquired through an image input device such as a scanner or a camera, the recognition of a character due to a vowel misrecognition can be prevented, and thus the surplus of neutral vowel recognition information is prevented. To increase the recognition rate of the neutral vowel, and consequently to increase the character recognition rate. In order to recognize the neutral vowels that are relatively difficult to recognize, the vowel recognition rate can be increased by using the initial consonant and the final consonant recognition information for the recognition of the neutral vowel, and the sums of the recognition reliability values of the initial consonant and the final consonant are set to 1 first. By normalizing and then multiplying the maximum values of the neutral vowel feature values to normalize the consonant recognition confidence values and the vowel feature values distributed in different effective intervals, the recognition of the neutral vowel is not biased against the confidence value of the consonant or the feature value of the vowel. As a result, a stable recognition rate is obtained.

Claims (8)

In the phoneme-based character recognition method, Classifying the phoneme combination type for the text image included in the document image; Determining whether there is a finality in the text image based on the classified phoneme combination type; A phoneme classification step of classifying phonemes included in the text image based on the determination result of the finality determination step; Recognize the consonants, the neutral vowels and / or the consonants classified in the phoneme classification step based on the result of the determination of the presence or absence of the final consonants, wherein the recognition of the neutral vowels is based on the recognition reliability of the consonants and / or the final consonants. A phoneme recognition step, executed based on And a phoneme combination step of combining the recognition results of the consonant, the consonant and / or the consonant which are recognized in the phoneme recognition step. The method of claim 1, In the phoneme recognition step, As a result of the determination of the presence or absence of the finality, if the finality is not included in the text image, the recognition of the neutral vowel is performed based on the recognition reliability of the initial consonant. If the character image includes the finality, the recognition of the neutral vowels is performed based on the initial consonant and the recognition reliability of the final consonant. . The method of claim 2, In the phoneme recognition step, Recognition reliability for the initial consonant, the neutral vowel and the final consonant are represented by vectors X and Z respectively expressed by the following equations , Equation 1 Equation 2 Here, Nx and Nz represent the number of recognizable leading consonants and final consonants, respectively, and x _i and z _j represent the recognition reliability of the i th initial consonant j th final consonant, respectively. How to improve. The method of claim 3, wherein In the phoneme recognition step, As a result of the determination of the presence or absence of the finality, the recognition of the neutral vowels is performed based on the recognition reliability vector G of the initial consonant expressed by the following equation. Equation 3 Wherein f _j represents a recognition reliability of an i- th neutral vowel among the recognizable neutral vowels. The method of claim 4, wherein Recognition of the neutral vowel is an operator in which the initial consonant reliability vector G and the neutral vowel reliability vector F are defined by the following equation. Connected by Run based on Here, Ny is a method of improving the phoneme-based character recognizer, characterized in that the number of recognizable neutral vowels. The method of claim 4, wherein In the phoneme recognition step, If the character image includes the finality, the recognition of the neutral vowels is performed in the recognition reliability vector G of the initial consonant and the recognition reliability vector H of the final consonant represented by Equation 4 below. Based on Equation 4 Method for improving the phoneme-based character recognizer, characterized in that executed. The method of claim 6, The recognition of the neutral vowel is a vector in which the consonant reliability vector G, the neutral vowel reliability vector F, and the final consonant reliability vector H are connected. Characteristic-based character recognizer performance improvement method, characterized in that executed on the basis of. A recording medium on which a program that implements the method of any one of claims 1 to 7 is recorded.