KR19980025571A - Hangul Recognition System Using Regular Noise - Google Patents

Abstract

1. TECHNICAL FIELD OF THE INVENTION

Hangul Recognition System Using Regular Noise

2. The technical problem to be solved by the invention

By solving the problem of misrecognition using regular noise, we want to provide a Korean recognition system that has high recognition rate and further expands the number of characters and fonts.

3. Summary of Solution to Invention

A first step of classifying Korean into six types; After the consonant area other than the vowel is included in the vowel area to determine the phoneme area, the vowel area and the consonant area are learned according to the determined phoneme area to recognize the phoneme, and in case of type 5 characters, find the vertical position of the horizontal component. Separating the phoneme region and recognizing the phoneme through neural network learning; And a third step of recognizing the first step and the second step according to the recognition result.

4. Important uses of the invention

It is used for Hangul Recognition System to recognize printed Hangul.

Description

Hangul Recognition System Using Regular Noise

The present invention relates to a Hangul recognition system that greatly improves the phoneme recognition rate of Hangul using regular noise, and in particular, by greatly improving the low recognition rate problem of characters, which is the biggest problem of the conventional OCR. It is used for document archiving and intelligent information retrieval through automatic digitization of recognition documents, and it is widely applied to high speed communication, etc., and it can maintain high recognition rate regardless of increase of font number and number of characters. It's about the system.

Hangul has a number of characters and is difficult to recognize because it has the characteristics of forming a single character by combining initial, neutral, and Jongsung. Therefore, in recent years, there have been many attempts to recognize each element by phoneme. Among them, Hangeul is a phoneme in multi-letter multi-letter Korean document recognition using the hierarchical neural network of Prof. Jae-Wook Kwon, Sung-Bae Kim, and Jin-Hyung Kim in the Journal of the Korean Institute of Information Scientists and Engineers (B). According to the location of the, it is divided into six types and using the neural network recognition method, and reports relatively good results.

However, since the conventional method uses a fixed phoneme region as an input of a phoneme recognizer to recognize each phoneme, even parts other than the phoneme are used as inputs, which makes it difficult to learn neural networks, and also decreases the recognition rate. As the font and size of the character to be recognized vary, there is a problem that appears more prominently.

In addition, the conventional method has a problem in that it is difficult to learn the fixed vowel region because the position of the horizontal vowel is very diverse in the case of type 5 of the Korean type.

Finally, the conventional method attempts to recognize in two stages of type classification and self-perception, so that if an error occurs in any one of the steps, an error occurs as a whole and the recognition rate is greatly reduced.

The present invention devised to solve the above problems, unlike the conventional method of determining the minimum region including the corresponding phonemes in a given type as the phoneme region, takes advantage of the fact that the noise appears regularly according to the phoneme to be combined In order to include the consonant areas other than the vowel in the vowel area to determine the phoneme area, it also learns other phoneme types that are combined in neural network learning, and in case of type 5 characters, finds the vertical position of the horizontal component By using the algorithm that separates the phoneme domains based on the position, the position of the horizontal collection is fixed, and the two stages of recognition module play a complementary role by recognizing the two stages of type classification and self-recognition according to the recognition result. The purpose is to provide a Hangul recognition system with reduced errors.

That is, the present invention has a high recognition rate by solving various misrecognition problems which are not solved in the conventional recognition method using a neural network using regular noise, and further has an excellent recognition rate for the number of characters and fonts to be recognized. Its purpose is to provide a Hangul recognition system with extensibility.

1 is an overall processing diagram of a Hangul recognition system according to the present invention,

2 is a detailed processing diagram of the type classifier according to the present invention;

3A to 3F are detailed processing diagrams of the phoneme recognizer according to the present invention;

4A and 4B are explanatory diagrams of a vowel region;

In order to achieve the above object, the present invention, in the Hangul recognition system that classifies Hangul into six types having a certain phoneme region and recognizes Hangul using a phoneme recognizer for each phoneme region, A first step of classifying Hangul into six types according to the phoneme coupling type in order to take advantage of the fact that they appear regularly; After the consonant area other than the vowel is included in the vowel area to determine the phoneme area, the vowel area and the consonant area are learned according to the phoneme area determined at the time of neural network learning. A second step of recognizing the phoneme through neural network learning after fixing the position of the horizontal vowel using an algorithm for finding a vertical position of the horizontal component and separating the phoneme region based on the position; And the first and second steps according to the recognition result in order that the recognition module of the first step of classifying a type and the second step of recognizing a phoneme play a complementary role to reduce the overall error. And a third step of recognizing.

Hereinafter, with reference to the accompanying drawings will be described an embodiment according to the present invention;

1 is an overall processing diagram of a Hangul recognition system according to the present invention.

First, the preprocessing part corresponds to document scanning to size normalization in FIG. 1. Scan a text document with a 300 dpi scanner and save it in image format (.tiff) (1). This string image is segmented into 80x80 single character images (2). Subsequently, the size normalization of a single character image of 80x80 size to 40x40 size is performed (3).

The recognition process is largely divided into two stages: classification and self-awareness. First, the normalized text image is classified into six types according to the phoneme region of the Hangul using the type classifier (4). 4A and 4B described below represent six types of Korean characters, and FIG. 2 described below shows a detailed structure of a type classifier.

The output node numbers of the type classifier are sorted and stored in ascending order of their outputs, and only the output node numbers with the largest outputs are selected to enable the next class of self-identifier. The enable signal is output (5).

Each type of phoneme recognizer learns other phoneme types that are combined in neural network learning by determining the phoneme area by including consonant areas other than vowels in the vowel area in order to take advantage of the fact that noises appear regularly according to the phoneme to be combined. For example, in the case of the letter of type 5, the vertical position of the horizontal component is found and then the position of the horizontal collection is fixed by using an algorithm that separates the phoneme region based on the position (6 to 11). Each type of phoneme recognizer is shown in detail in FIGS. 3a to 3f described below, and each type of phoneme recognizer is composed of two or more phoneme recognizers to recognize each phoneme.

Thereafter, after taking the average of the output values selected from each type of the self-identifying device selected from each type of self-identifying device (12), it is determined whether the average output value is greater than 0.5 (13).

As a result of the determination, when the average output value is less than 0.5, the process transitions to the process of enabling the maximum value output node number by sorting the output value in order to consider the candidate type, and converting the next maximum value output node number to the maximum value output node number. Select to, i.e., put the candidate type back into the selected type and operate the self-identifying device of that type. At this time, when the candidate type of the classifier itself has a very low output value, the maximum value output node is replaced with the next maximum value output node only when the output value of the next maximum value output node is larger than 0.5 so as to use the first result as it is. . This operation is a way to solve the error of type classifier. The principle is explained in detail as follows.

In the general two-stage recognition method, if an error occurs in the previous step, an error occurs as a final result regardless of the next step. To prevent this error, backtracking is allowed according to the confidence of the result of the next step. In the Hangul recognition system according to the present invention, the type having the highest output value among the outputs of the type classification neural network is regarded as the type of the input character, and is inputted to the corresponding self-aware neural networks. However, if an error occurs in the previous step and recognizes it as an incorrect type, the areas used by the current self-aware neural network as inputs will be arbitrary phoneme areas mixed with several phonemes instead of the corresponding phoneme areas. Therefore, each phoneme recognizer recognizes a completely different image that it has never learned, so its outputs do not have any output at all. Therefore, if the output value of the node having the highest value does not exceed the first threshold (1), the current type is considered to be an incorrect type, and the candidate type is considered. Candidate type is the type with the highest output value after the current type in the previous classification stage. This process is repeated until the output values of the self-aware neural networks have a somewhat higher output value. If you are aware of all six types and you do not have a high output that you are sure of, then the answer is the one with the highest output.

So far, this has been a description of the case where the candidate type is considered and only the result of the classifier is considered, and the result of the type classifier can be considered here. When the candidate type is taken and the confidence level corresponding to the candidate type is lower than a certain level, no matter how the self-aware neural network attempts to consider the candidate type, the previous type may be considered as the answer considering the judgment of the classifier. The low certainty of the type classifier means that the output value of the node corresponding to the candidate type is lower than the second threshold (threshold 2). At this time, the confidence level of the classifier and the confidence level of the self-identifier can vary depending on where the greater weight is placed. Experimentally, the greater the weight of the self-aware neural network, the better the result. This may have this effect by making the value of the first threshold 1 greater than the value of the second threshold 2.

As a result of the determination, when the average output value is 0.5 or more, after mapping the selected output node number of each neural network output from each selected phoneme recognizer to the corresponding phoneme code (14), the phonemes are combined according to the phoneme code (15). ) The recognition process is completed by outputting the combined Hangul code.

2 is a detailed processing diagram of the type classifier according to the present invention.

All input features used in the type classifier use mesh vectors for the entire input image. The mesh vector is characterized by the number of black pixels contained in the window moving the mesh window of size N × M from the upper left of the input image. Therefore, if the size of the entire input image is X and Y, the dimension of the mesh vector becomes (X / N) x (Y / M). At this time, the size of the mesh window (mesh window) in the present invention is 5x5, the dimension of the mesh vector (mesh vector) is 64 dimensions.

In practice, each vector of a 64-dimensional mesh vector is normalized between 0 and 1 before being used as the input of the neural network.

Type classification neural network uses Multi-Layer-Perceptron (MLP), and its structure has 64 input nodes, 35 hidden nodes, and 6 output nodes.

3A to 3F are detailed processing diagrams of the phoneme recognizer according to the present invention.

All input features used in each type of recognizer use a mesh vector for the entire input image, and mesh window size according to each neural network used in each type of recognizer. Is different. At this time, all the self-identifiers use multi-layered perceptron (MLP) as a neural network.

What is important here is the input region of the autonomic recognizer neural network. When the width and length of the entire input area are 0 to 49, respectively, the phoneme areas of each type are indicated by coordinates in the drawings of each type. Unlike the conventional method, this area is the size including the consonant area in the case of the vowel area. This works as a way to solve the shortcomings of this recognition model.

Looking at the structure of the phoneme recognizer according to each type in detail as follows.

First, the first type of self-identification recognizer has an initial coordinate of (0, 2, 26, 34) when the coordinate of the input area is represented as (X1, Y1, X2, Y2), and the neutral coordinate is (4, 0, 39, 38). Therefore, the first type of phoneme recognizer is composed of two phoneme recognizers to recognize the first and the neutral, respectively, and the phoneme recognizer to recognize the initiality is a 3x3 mesh window and a 99-dimensional mesh vector. And input it into a multi-layer perceptron (MLP) having 99 input nodes, 20 hidden nodes, and 19 output nodes, and an autonomous recognizer for recognizing neutrality is a 3 × 3 mesh window. A 156-dimensional mesh vector is generated and input into a multi-layer perceptron (MLP) having 156 input nodes, 32 hidden nodes, and 9 output nodes.

In the second type of self-identification recognizer, when the coordinate of the input area is expressed as (X1, Y1, X2, Y2), the coordinate of the initial is (2, 1, 39, 29), and the neutral is (2, 2, 37, 39). Accordingly, the second type of phoneme recognizer is composed of two phoneme recognizers to recognize the first and the neutral, respectively, and the phoneme recognizer to recognize the initiality is a mesh window of 3x3 size and has a 130-dimensional mesh vector. To generate a multi-layer perceptron (MLP) having 130 input nodes, 25 hidden nodes, and 19 output nodes, and a self-identifying device for recognizing neutrality is a 3x3 mesh window. A 156-dimensional mesh vector is generated and input into a multi-layer perceptron (MLP) having 156 input nodes, 15 hidden nodes, and 5 output nodes.

In the third type of self-identification recognizer, when the coordinates of the input area are represented as (X1, Y1, X2, Y2), the initial coordinates are (1, 3, 29, 32), and the horizontal collection coordinates are (0, 3, 28). , 37), and the coordinates of the vertical collection are (2, 8, 39, 37). Accordingly, the third type of phoneme recognizer is composed of three phoneme recognizers to recognize the first consonants, the horizontal vowels and the vertical vowels, respectively, and the phoneme recognizer for recognizing the consonants is a three-dimensional mesh window of 3x3 size. Generates a mesh-like vector and inputs it into a multi-layer perceptron (MLP) having 100 input nodes, 45 hidden nodes, and 19 output nodes, and a self-identifying device for recognizing horizontal vowels has a 3x3 mesh type. Create a 120-dimensional mesh vector as a mesh and input it to a multi-layer perceptron (MLP) with 120 input nodes, 10 hidden nodes, and 3 output nodes, and recognizes vertical vowels. Phoneme recognizer is a 3x3 mesh window that creates a 130-dimensional mesh vector and has a multi-layer perceptor with 130 input nodes, 15 hidden nodes, and 5 output nodes. Enter in the loan (MLP).

In the fourth type of phoneme recognizer, when the coordinate of the input area is expressed as (X1, Y1, X2, Y2), the initial coordinate is (0, 1, 26, 26), and the neutral coordinate is (1, 1, 39, 26), and the final coordinates are (3, 25, 38, 39). Therefore, the fourth type of self-identification device is composed of three self-identification devices for recognizing the initial, neutral, and final species, respectively, and the self-identifying device for recognizing the first generation is a mesh window of 2x2 size and has a mesh of 182 dimensions. Generates a type vector and inputs it to a multi-layer perceptron (MLP) having 182 input nodes, 38 hidden nodes, and 19 output nodes, and a self-identifying device for recognizing neutrality is a 3 × 2 mesh window (mesh). window) creates a 169-dimensional mesh vector and inputs it to a multi-layer perceptron (MLP) with 169 input nodes, 35 hidden nodes, and 9 output nodes. A mesh size of 144 is created by a mesh window of size and input to a multi-layer perceptron (MLP) having 144 input nodes, 39 hidden nodes, and 26 output nodes. .

Characters of type 5 have very different horizontal collections, depending on the type of phonemes that are combined above and below the same vowel. For example, there are runes and soups. After separating two vowels with completely different vertical positions into a general fixed vowel area, the position and distribution of the strokes in the two areas are completely different even with the same vowel. That is, in the case of runes, TT is located at the bottom of the area, and in the middle, a part of the r consonants is occupied. In the case of bureau, TT is located at the top and a part of the base a is occupied. This makes the neural network harder to learn and reduces its recognition rate. In order to prevent this phenomenon, it is possible to make the neural network easier to learn and to recognize by using an algorithm that fixes the horizontal bar of horizontal vowels at the center of the vowel area.

The algorithm to find the vertical position of the horizontal collection in the text image of type 5 is as follows. The text image consists of 40 pixels of width and height respectively. First, the leftmost column of the text image is examined to find the first column of black pixels. In this case, if the noise of type 5 is no noise, the leftmost point of the horizontal collection is the point located to the leftmost of the character. The pixels in the horizontal row are examined at the vertical position where the black pixels are located to determine if the black pixels in the found columns are points or noise that are part of the horizontal collection. If it is part of the horizontal vowel, there are consecutive black pixels at the beginning, end, and middle of the horizontal row (if there is no noise, all the rows will be black pixels, but this may work as an algorithm). Will satisfy the conditions.

The vowel and the two consonant regions are separated as follows: initial lower coordinate = y + Δy1, final upper coordinate = y + Δy2, neutral upper coordinate = y + Δy1 ', neutral lower coordinate = y + Δy2' . At this time, the values of Δyl, Δy2, Δy1 ', and Δy2' are -1, 4, -12, and 15. When the regions are separated in this way, the size of the neutral region is always constant, but the size of the initial and final regions depends on the found vertical position, so the size is normalized. By doing as above, the vowel position of type 5 varies depending on letters or fonts, thus making it difficult to recognize.

The fifth type of self-identification recognizer has an initial coordinate of (0, 0, 39, Y-1) when the coordinate of the input area is represented by (X1, Y1, X2, Y2), and the neutral coordinate is (1, Y- 12, 38, Y + 15), and the final coordinates are (0, Y + 4, 39, 39). Thus, the fifth type of self-identification machine is composed of three self-identification devices to recognize the initial, neutral, and final species, respectively, and the self-identifying device for recognizing the primitive is normalized to 40x20 size and mesh-shaped windows of 3x2 size. window) to generate a 140-dimensional mesh vector and input it to a multi-layer perceptron (MLP) having 140 input nodes, 27 hidden nodes, and 19 output nodes, and a self-identifying device for recognizing neutrality is 38x28. Multi-layer perceptron (MLP) with 182 input nodes, 25 hidden nodes, and 5 output nodes by normalizing size to size and generating a 182-dimensional mesh vector with a 3x2 mesh window. ), And the phoneme recognizer to recognize the finality is normalized to 40x20 size and creates 140-dimensional mesh vector with 3x2 mesh window. It can force node and having the number of hidden nodes 35 and 24 can be input to the output node layer perceptron (MLP).

In the sixth type of phoneme recognizer, when the coordinates of the input area are expressed as (X1, Y1, X2, Y2), the initial coordinates are (2, 1, 27, 23), and the horizontal collection coordinates are (1, 13, 29). , 27), the vertical coordinates are (13, 2, 39, 27) and the final coordinates are (4, 24, 38, 39). Therefore, the sixth type of phoneme recognizer is composed of four phoneme recognizers for recognizing the consonants, the horizontal vowels, the vertical vowels, and the finality, respectively, and the recognizer for recognizing the consonants is a 3x3 mesh window. Creates a 72-dimensional mesh vector and inputs it into a multi-layer perceptron (MLP) having 72 input nodes, 20 hidden nodes, and 19 output nodes. A 54-dimensional mesh vector is created with a mesh window and input to a multi-layer perceptron (MLP) having 54 input nodes, 17 hidden nodes, and 3 output nodes, and recognizes a vertical collection. Multi-layer perceptor with 54 input nodes, 17 hidden nodes and 5 output nodes by generating a 54-dimensional mesh vector with a 3x3 mesh window. The self-identification device for entering the MLP, which recognizes the finality, creates a 72-dimensional mesh vector with a mesh window of 3x3 size, allowing 72 input nodes, 10 hidden nodes, and 9 Input to multi-layer perceptron (MLP) which has the number of output nodes.

4A and 4B are explanatory diagrams of a vowel region and show a form in which Korean characters are classified into six types having a certain phoneme region.

On the other hand, the Hangul recognition system according to the present invention showed a 99.65% recognition rate for the Myungjo and Gothic fonts, and maintained an average recognition rate of 98.9% even when including the Shinmyung, Middle and Korean fonts. In addition, as a result of recognizing the number of characters to be recognized by extending from 500 to 1405 characters, the difference in recognition rate was confirmed to be less than 0.1%.

The present invention described above is capable of various substitutions, modifications, and changes without departing from the spirit of the present invention for those skilled in the art to which the present invention pertains, and the above-described embodiments and accompanying It is not limited to the drawing.

The present invention as described above has the following effects.

First, there is an effect of resolving ambiguity and increasing extensibility due to the recognition of each phoneme and the expansion of the corresponding phoneme area. In the case of vowels, the recognition rate was improved by 73% compared to the conventional method. When the number of fonts was increased, the recognition rate was much lower than the conventional method, and when the number of characters was increased, the recognition rate was almost unchanged. .

Secondly, in the case of type 5 characters, the position can be fixed to shorten the learning time and increase the recognition rate. The most common misrecognition occurs in the previous studies of type 5 vowels. In the Hangul recognition system according to the present invention, the recognition rate can be improved by 53%.

Third, by considering the candidate type according to the neural network's confidence level, it is possible to solve the total error increase proportional to the product of each step error, which is the biggest disadvantage of the multi-level recognizer. In the conventional method, when an error occurs in the type classification neural network, the character is misrecognized regardless of the self-aware neural network. It is common for the whole error to appear as the product of the error in the previous step and the error in the next step. In the Hangul recognition system according to the present invention, considering the candidate type, the number of type errors is actually smaller than that of the type classification neural network. Appears.

Claims (11)

In the Hangul Recognition System, which classifies Hangul into six types with a certain phoneme region, and recognizes Hangul using a phoneme recognizer for each phoneme region, A first step of classifying Hangul into six types according to the phoneme combining type in order to take advantage of the fact that noises appear regularly according to the phoneme of the Hangul to be combined; After the consonant area other than the vowel is included in the vowel area to determine the phoneme area, the vowel area and the consonant area are learned according to the phoneme area determined at the time of neural network learning. A second step of recognizing the phoneme through neural network learning after fixing the position of the horizontal vowel using an algorithm for finding a vertical position of the horizontal component and separating the phoneme region based on the position; And Recognizing the first and second steps according to the recognition result in order that the recognition module of the first step of classifying the type and the second step of recognizing the phoneme play a complementary role to reduce the overall error. Hangul recognition system comprising a third step to. The method of claim 1, The first step is, Segmenting the string image into single character images after scanning the document consisting of the character string with a scanner and storing the image in an image format; Classifying the normalized text image into six types according to the phoneme region of the Hangul using a type classifier after normalizing the one text image to a predetermined size; And Sort and store the output node numbers of the type classifier in the order of their largest output values, and select an output node number having the largest output value to enable an enable signal for enabling the self-identifying identifier of the type. Hangul recognition system comprising the step of outputting. The method of claim 2, The type classifier, Generating a 64 dimensional mesh vector with a 5 × 5 mesh window; Normalizing each vector value of the 64-dimensional mesh vector between 0 and 1 before using it as an input of the neural network; And Multi-Layer-Perceptron (MLP), which has 64 input nodes, 35 hidden nodes, and 6 output nodes, receives and learns the normalized vector value and outputs the result. Hangul recognition system comprising a. The method of claim 2, The third step, Determining whether the average output value is greater than the first predetermined reference value after taking an average of the output values selected from each type of the self-identifiers and outputting from the enabled self-identifiers; As a result of the determination, when the average output value is smaller than the first predetermined reference value, the output value is sorted only when the output value of the next largest maximum output node is larger than the second predetermined reference value in order to consider the candidate type. Transitioning to enabling the number, selecting the next largest maximum value output node number as a new maximum value output node number to operate the corresponding phoneme recognizer; And As a result of the determination, when the average output value is greater than or equal to the second predetermined reference value, mapping an output node number output from each selected phoneme recognizer to a corresponding phoneme code, and then outputting a combination of phonemes according to the phoneme code; Hangul recognition system characterized in that. The method according to any one of claims 1 to 4, Phoneme recognizer for recognizing the first type of Hangul, When the coordinates of the initial area of the input area are (0, 2, 26, 34) and the neutral coordinates are (4, 0, 39, 38), it is composed of two autonomous recognizers to recognize the initial and neutral, respectively. In order to recognize the initial consonants, the autonomous recognizer generates a 99-dimensional mesh vector with a 3x3 mesh window and has a multi-layer with 99 input nodes, 20 hidden nodes, and 19 output nodes. The self-identifier for perceptron (MLP) and the recognition of neutrality is a 3x3 mesh window that generates 156-dimensional mesh vectors to generate 156 input nodes, 32 hidden nodes and 9 Hangul recognition system having a structure for inputting to the multi-layer perceptron (MLP) having the number of output nodes. The method of claim 5, Phoneme recognizer for recognizing the second type of Hangul, When the coordinates of the initial area of the input area are (2, 1, 39, 29) and the neutral coordinates are (2, 2, 37, 39), two self-identifiers are recognized to recognize the initial and neutral stars respectively. In order to recognize the initial consonants, the autonomous recognizer is a 3x3 mesh window that generates a 130-dimensional mesh vector and has a multi-layer with 130 input nodes, 25 hidden nodes, and 19 output nodes. The self-identifier for perceptron (MLP) and the recognition of neutrality is a 3x3 mesh window that generates a 156-dimensional mesh vector to generate 156 input nodes, 15 hidden nodes, and 5 Hangul recognition system having a structure for inputting to the multi-layer perceptron (MLP) having the number of output nodes. The method of claim 6, Phoneme recognizer for recognizing the third type of Hangul, The initial coordinate of the input area is (1, 3, 29, 32), the horizontal coordinate is (0, 3, 28, 37), and the vertical coordinate is (2, 8, 39, 37). In order to recognize the first consonants, the horizontal vowels and the vertical vowels, three self-identifiers are used. The self-identifiers for the recognition of the first consonants are generated by generating a 100-dimensional mesh vector with a 3x3 mesh window. A multi-layer perceptron (MLP) with 100 input nodes, 45 hidden nodes, and 19 output nodes, and a self-identifying device for recognizing horizontal vowels is a 3x3 mesh window. 3D mesh is generated by multi-layer perceptron (MLP) with 120 input nodes, 10 hidden nodes, and 3 output nodes. Create a 130-dimensional mesh vector with a mesh window 130 and the input node 15 can be hidden node with multi-layer perceptron (MLP) Hangul recognition system, characterized in that with a structure in which the input to obtain a five-number of output nodes. The method of claim 7, wherein Phoneme recognizer for recognizing the fourth type of Hangul, In the case where the initial coordinate of the input area is (0, 1, 26, 26), the neutral coordinate is (1, 1, 39, 26), and the final coordinate is (3, 25, 38, 39), It is composed of three self-identification devices to recognize the initial, neutral, and final species, respectively. The self-identification device for recognizing the first castle is a 2x2 mesh window that generates 182-dimensional mesh vector to create 182 input nodes. Input to multi-layer perceptron (MLP) with number, 38 hidden nodes, and 19 output nodes, and the autonomous recognizer for recognizing neutrality is a 3x2 mesh window, which is a 169-dimensional mesh vector Is generated and input into a multi-layer perceptron (MLP) having 169 input nodes, 35 hidden nodes, and 9 output nodes, and a self-identifying device for recognizing the finality is a 2x2 mesh window. Create a 144-dimensional mesh vector with 144 input nodes and 39 The number of hidden nodes with multi-layer system Hangul recognition, characterized in that with a structure in which the input perceptron (MLP), which will have the output node 26. The method of claim 8, In the case of type 5 characters, the process of fixing the position of the horizontal collection using an algorithm that finds the vertical position of the horizontal component and separates the phoneme region based on the position, From the leftmost column of the text image, the first column of black pixels is found, and then the horizontal position in the vertical position where the black pixels are positioned to determine whether the black pixels in the found columns are a point or noise that is part of the horizontal collection. Examine the pixels in the row, and if there are consecutive black pixels at the beginning, end, and middle of the horizontal row, recognize them as a horizontal collection, then the initial lower coordinate = y + Δy1, the final upper coordinate = y + Δy2, the neutral top The vowel and the two consonant areas are separated (Δyl, Δy2, Δy1 ', and Δy2') such as coordinates = y + Δy1 'and neutral bottom coordinates = y + Δy2' (-1, 4, -12, 15). Hangul recognition system, characterized in that to normalize. The method of claim 9, Phoneme recognizer for recognizing the fifth type of Hangul, The initial coordinate of the input area is (0, 0, 39, Y-1), the neutral coordinate is (1, Y-12, 38, Y + 15), and the final coordinate is (0, Y + 4, 39, 39), it consists of three self-identification devices to recognize the initial, neutral, and final species, respectively, and the self-identification device for recognizing the initials is normalized to 40x20 size and mesh window of 3x2 size (mesh window). ) Creates a 140-dimensional mesh vector and inputs it to a multi-layer perceptron (MLP) having 140 input nodes, 27 hidden nodes, and 19 output nodes, and a self-identifying device for detecting neutrality is 38x28 size. Multi-layer perceptron (MLP) with 182 input nodes, 25 hidden nodes, and 5 output nodes by normalizing the size and generating a 182-dimensional mesh vector with a 3x2 mesh window. And a phoneme recognizer to recognize the species, and normalizes the size to 40x20. Generates a 140-dimensional mesh vector with a high 3x2 mesh window and inputs it to a multi-layer perceptron (MLP) having 140 input nodes, 35 hidden nodes, and 24 output nodes. Hangul recognition system characterized in that having a. The method of claim 10, Phoneme recognizer for recognizing the sixth type of Hangul, The initial coordinates of the input area are (2, 1, 27, 23), the horizontal coordinates are (1, 13, 29, 27), the vertical coordinates are (13, 2, 39, 27), When the coordinates of the finality are (4, 24, 38, 39), four phoneme recognizers are used to recognize the first consonants, the horizontal vowels, the vertical vowels, and the final consonants, respectively. Create a 72-dimensional mesh vector with a mesh window and input it into a multi-layer perceptron (MLP) with 72 input nodes, 20 hidden nodes, and 19 output nodes, and recognizes horizontal vowels. The multi-layer perceptron (MLP) has 54 input nodes, 17 hidden nodes, and 3 output nodes by generating a 54-dimensional mesh vector with a 3x3 mesh window. Is a 3x3 mesh window. It generates a 54-dimensional mesh vector and inputs it into a multi-layer perceptron (MLP) having 54 input nodes, 17 hidden nodes, and 5 output nodes. Creates a 72-dimensional mesh vector with a mesh window and inputs it to a multi-layer perceptron (MLP) having 72 input nodes, 10 hidden nodes, and 9 output nodes. Hangul Recognition System.