JP2997403B2 - Handwritten character recognition method and apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a method and apparatus for recognizing handwritten characters, and more particularly to a method and apparatus for recognizing handwritten Hangul characters.

[0002]

2. Description of the Related Art Optical character recognition (OCR: Optical)
“Character Recognition” refers to a technology for recognizing the contents of a portion corresponding to a character in a character image input through a scanner. Such an O
The CR technology is being researched and developed roughly classified into a case of recognizing printed characters and a case of recognizing handwritten characters. Although the technology for recognizing printed characters continues to evolve rapidly, the recognition of handwritten characters is an excellent handwritten character recognition device due to the unspecified number of handwritten people and the variety of handwritten characters due to irregular handwriting habits. Was difficult to appear.

[0003] A typical technique for recognizing handwritten characters is a template match (Template Matchin).
g) Method, NNC (Nearest Neighbor)
Classification method, structural method,
There are an artificial neural network method, a method using a concealed Markov model, and the like. However, most of these methods are methods for digits or European handwriting with a small number of characters to be recognized, and it is difficult to apply them directly to Hangul. Therefore, the recognition of handwritten kanji and Arabic numerals, which are currently being actively researched, employs a method combining various recognition techniques ((1) TH Hildebrandta).
nd W. Liu, Optical Recognit
ion of Handwriting Chines
e Characters: Advances Sin
ce 1980, PatternRecognition
n, Vol. 26, no. 2,1983,205-22
5 (2) E. Cohen, J .; J. Hull and
S. N. Srihari, Understanding
Handwriting text in aStr
uctured Environment: Deter
mining ZIP Codes from Add
lesses, in Character and Ha
ndwriting Recognition: Ex
Panding Frontiers, P.S. S. P W
ang ed. , World Scientific
Publishing Company, 1991,
21-264).

In the NNC method, after arbitrarily obtaining a representative value for a finite number of characters to be recognized, when recognizing an input character pattern, a representative value most similar to a feature value extracted from the input character pattern is searched for and applied. This is a pattern recognition method for obtaining a character or a character set. In general, the NNC method is easier to implement and apply than other pattern recognition methods, and has superior performance in terms of processing speed. Therefore, the NNC method is effectively used when the number of characters to be recognized is large, such as Hangul or Kanji. obtain.

[0005] Most of the Hangul recognition methods and devices that have been published so far are based on squared fonts, hand-separated phonemes, so that the first, middle, and final voices are not connected, or the first, middle, and middle voices.
Saengmul body created by separating the final voice up and down (Saem-
mul characters). In addition, all of the common Hangul 2350 characters are not targeted for recognition, and 100 to 1000 characters of Hangul, which are usually mainly used, are selected and recognized only as targets, making it difficult to apply to a general environment.

[0006] Since the number of characters in the common Hangul is 2350 and the number of characters to be recognized is too many, and the characters handwritten by many handwritten persons are various, a recognition device having excellent performance is provided by the conventionally proposed method. Therefore, a new recognition method has been required.

[0007]

SUMMARY OF THE INVENTION An object of the present invention is to provide a method and an apparatus for recognizing handwritten Hangul characters having excellent recognition performance in order to solve the problems of the prior art.

[0008]

According to the present invention, there is provided a method for inputting a character image having a size of I × J, comprising the steps of: Calculating a feature vector comprising coordinates in a statistical manner; searching for a representative vector most similar to the calculated feature vector in a set of representative vectors corresponding to the character to be recognized; Determining a character to be recognized and a candidate character group including the character, and determining the candidate character through a neural network configured using a finite number of feature vectors for a character pattern of the candidate character group. and a step of outputting as a final recognition result one character of the group, prior to
The method for calculating the feature vector is as follows.
Split video into M vertical strips and horizontally
Divided into N strips and non-linear into cells of size M x N
Shape division (I ≧ M, J ≧ N, and I, J, M, N are 0
Larger integer) and for each divided cell
The vertical component value FV and the horizontal component value FH are calculated by the following equation: Fv = Bv / Bp FH = BH / Bp (where Bp is the total number of black pixels in the input image, and Bv
Is the number of black pixels belonging to the vertical component in the cell
Where BH is a black picture belonging to the horizontal component existing in the cell.
Calculate 2MN feature values
And the pattern of rotating the input video by 45 °
The same method as above is applied to another 2MN
Calculating the feature values, the vertical, horizontal, diagonal lines and
And 4MN feature values for reverse oblique line components
Providing as a feature vector of the turn .

An apparatus according to the present invention includes an input unit for inputting a handwritten character image as a binary character image, a preprocessing unit for separating the input character image and outputting the characters one by one, and representing each recognition target character as a representative. A candidate character group storage unit storing a candidate character group consisting of candidate characters including the recognition target character, a representative vector storage unit storing all feature vectors corresponding to all the recognition target characters, and the preprocessing unit , A feature vector of the character pattern input through the above-described process, a representative vector most similar to the obtained feature vector is searched from the representative vector storage unit, and a recognition target character for the searched representative vector and a candidate character for the recognition target character A large classification unit for searching a group from the candidate character storage unit, a neural network storage unit including a neural network for each candidate character group stored in the candidate character group storage unit, After recognizing one character as a final recognition result from the candidate character group searched from the large classification unit through the neural network stored in the neural network storage unit, a control signal is provided to the preprocessing unit and And a detailed classifying unit that supplies characters to be recognized until all the characters in the input character image are recognized.
Divide an input image of J size into M vertical strips
Split into N strips in the horizontal direction, M × N size
A non-linear division (I ≧ M, J ≧ N, I,
J, M, and N are integers greater than 0).
The vertical component value FV and the horizontal
The direction component value FH is calculated by the following equation: Fv = Bv / Bp FH = BH / Bp (where Bp is the number of all black pixels in the input image, and Bv
Is the number of black pixels belonging to the vertical component in the cell
Where BH is a black picture belonging to the horizontal component existing in the cell.
Calculate 2MN feature values
A feature value calculation unit to rotate the input image by 45 °.
Apply the same method to the turn for the other
A rotation feature value calculation unit for calculating 2MN feature values of
4MN for vertical, horizontal, oblique and reverse oblique components
Feature value of the input character pattern
And a feature vector providing unit .

[0010]

According to the method of the present invention, a group of characters to be recognized is extracted from a handwritten character pattern input using a template matching method.
That is, by forming a candidate character group and outputting one character of the candidate character group as a final recognition result through a neural network having an excellent ability to perform the recognition target character group, unrestricted character recognition in Hangul handwritten recognition is achieved. It is possible.

[0011]

Embodiments of the present invention will be described below in detail with reference to the accompanying drawings.

According to the method of the present invention, a character image is input in an I × J size, and a feature vector composed of a finite number of coordinates as a feature for the pattern of the input character image is calculated by a statistical method and recognized. The representative vector set corresponding to the target character is searched for a representative vector most similar to the calculated feature vector, and a character corresponding to the searched representative vector and a candidate character group including the character are determined to form a candidate character group. One character in the obtained candidate character group is recognized as a final recognition result through the corresponding neural network.

FIG. 1 is a block diagram of a handwritten character recognition device according to the present invention. The apparatus of FIG. 1 includes an input unit 10 for converting a handwritten character image into a binary image through a scanner and inputting the input image, a preprocessing unit 20 for separating the binary image input through the input unit 10 and outputting one character at a time, A candidate character group storage unit 30 storing a candidate character group corresponding to a target character, a representative vector storage unit 40 storing each representative feature vector corresponding to each recognition target character, and the pre-processing unit 20 are input. The characteristic vector of the character pattern is obtained, a representative vector most similar to the obtained characteristic vector is searched from the representative vector storage unit 40, and a character to be recognized for the searched representative vector and a candidate character group for the character are determined by the candidate character. A neural network storage unit 60 including a neural network for all candidate character groups stored in the storage unit 30 and the neural network stored in the neural network storage unit 60 are used. After recognizing one character in the candidate character group searched from the large classification unit 50 as a final recognition result, a control signal is provided to the preprocessing unit 20 to determine the next character to be recognized. And a detailed classification unit 70 for supplying the information in a system. This process is repeated until all the characters in the input document have been recognized.

The input unit 10 is a video scanner or a CCD.
An image pickup device such as a camera converts an image into a digital pattern of a binary image which can be read by a computer, and inputs the digital pattern. At this time, the character area is digitized to a black point, and the background is digitized to a white point.

The pre-processing unit 20 removes noise components other than the character pattern from the input pattern generated by the input unit 10,
Separate each Hangul character and output one character at a time. When composing a document in order to simplify the configuration of the preprocessing unit, a picture or an explanatory note excluding a handwritten character area is input to the input unit 1.
It is desirable to perform the printing process using a color that cannot be detected by the image pickup device 0.

The large classifying section 50 obtains a feature vector of the input character pattern input from the preprocessing section 20 in the following manner.

A vector composed of a finite number of coordinates is calculated by a statistical method as a large classification feature for a character pattern. In order to obtain a feature vector, first, an input image of I × J size is divided into cells of M × N size (I ≧ M, J ≧
N, and I, J, M, and N are integers greater than 0), and apply a non-linear method to solve the problem due to the character diversity of many handwriting users (Y. Yamashiti).
a, K. Higuchi, Y .; Yamada an
d Y. Haga, Classification
of Handprinted Kanji Char
actors bythe structured s
egment matching method, pa
tterrn recognition letter
s, Vol. 1, 1983, 475-479). That is, the input video is divided into M vertical strips,
Equalize the number of black pixels in each strip. When the image is horizontally divided into N strips in the same manner, the input image is divided into MN cells. For each cell, two feature values, a vertical component value FV and a horizontal component value FH, are calculated as follows.

Fv = Bv / Bp FH = BH / Bp (where Bp is the number of all black pixels in the input video, and Bv
Is the number of black pixels belonging to the vertical component existing in the cell, and BH is the number of black pixels belonging to the horizontal component existing in the cell. Here, the vertical and horizontal components are vertical and horizontal runs (R
(UN: continuous black pixels in a scan line) means a set of runs whose length is larger than the critical value T. The critical value T is dynamically set according to the thickness W of the given character image,
For example, the thickness is set to 2.5 times the thickness of the character image. The thickness W of the character image is given by the following equation.

W = Bp / (Bp−Wp) (where Bp is the number of black pixels in the input image, and Wp is a black pixel of three adjacent pixels in a 2 × 2 window arranged side by side with the pixels) 2A to 2C show an execution example of obtaining 2MN feature values from an input character image by the method described above. Further, the same method as above is applied to a pattern obtained by rotating the input image by 45 °, and another 2MN feature values are calculated. The horizontal and vertical components extracted from the rotated image correspond to oblique and reverse oblique components from the original image. As a result, a feature vector including 4MN values is configured.

As described above, the representative vector storage unit 40 stores characters to be recognized, for example, 23 in the case of Hangul handwriting.
The feature vectors for each of the 50 representative characters are obtained and stored in the same manner. That is, each of these can be obtained by averaging the feature vectors of the sample patterns for the recognition target characters, and the more representative the training sample, the more reliable the representative vector can be obtained.

In the major classifying unit 50, the representative vector storage unit 40
In the set of representative vectors stored in, search for the most similar feature vector obtained from the input character. The similarity Si between the feature vector V extracted from the input character pattern and an arbitrary representative vector Mi in the representative vector set is measured by the distance in the space where the feature vector is distributed, and its formula is as follows. .

[0022] Here, V = v1, v2, v3,..., Vm, Mi = mi
1, mi2, mi3,..., Min, n = 4 MN, 1 ≦
i ≦ 2350. Therefore, as the value of Si decreases, the similarity between the two vectors is considered to be high, and a representative vector having the highest similarity is searched for.

The candidate character group storage unit 30 stores candidate character groups for the characters to be recognized. The candidate character groups are determined as follows. A chaotic pair set (mis-recognition type, confu) obtained by performing Template Matching on a training sample to form a candidate character group
Use the Ssion Matrix. That is, Ci
Characters Cj other than

[0024]

[Outside 1]

By collecting information on characters for each character, arranging the characters in order of frequency, and selecting some of the higher-order characters, a candidate character group for Ci can be constructed. As a result, the candidate character group configured in this way is stored in the candidate character group storage unit 30. In the embodiment of the present invention, the correct recognition rate for the candidate character group for each of the recognition target characters, that is, the possibility that the recognition target character for the given sample character pattern is included in the candidate character group found by template matching. It was kept at 99% or more, and the size of each character group was limited so as not to exceed 20 characters.

As described above, the large classifying section 50 finds the feature vector of the input character pattern, searches the representative vector storage section 40 for the most similar representative vector, and searches for the character corresponding to the searched representative vector and the character. A process of searching for a candidate character group including the character string with reference to the candidate character group storage unit 30 is performed.

The neural network unit 60 stores neural networks in order to finally search for one character from a group of candidate characters by the large classification unit. Each neural network is configured as follows.

As a feature of the neural network for the character pattern, a vector composed of a finite number of coordinates is calculated by a statistical method. In order to extract the features used in the neural network, the character image is nonlinearly divided into N × M cells in both the horizontal and vertical directions as in the case of the feature vector of the large classification unit 50, and then the black in each cell is extracted. A cell feature value is calculated by calculating the ratio between the horizontal run and the vertical run to which the pixel belongs. An example of feature extraction for a neural network is shown in FIG. First, the character image is nonlinearly divided into N × M cells for each of the horizontal and vertical methods, and then each black pixel Px, y (0 ≦ x ≦ W,
0 ≦ y ≦ H; W is the width of the character image, H is the height of the character image), and the length RLHx, y of the horizontal run to which the horizontal run belongs and the length RLVx, y of the vertical run are obtained. Next, each P
With respect to x and y, the horizontal contribution DCHx, y and the vertical contribution DCVx, y are obtained as follows.

DCHx, y = RLHx, y / (RLH
x, y + RLVx, y) DCVx, y = RLVx, y / (RLHx, y + RLV)
x, y) Finally, the average value of DCHx, y, DCVx, y for all the black pixels present in the N × M cells that have been nonlinearly divided is obtained, and a 2 (N × M) -dimensional feature vector is calculated. Also, the method described so far is applied to white pixels,
Further, another 2 (N × M) -dimensional feature vector is calculated to finally obtain a 4 (N × M) -dimensional feature vector.

The neural network structure adopted in the present invention is the same as in FIG. Here, the input node is 4 (N × M) corresponding to the number of feature values extracted from the input character video, the output layer is made up of nodes as many as the number of characters to be recognized, and the number of intermediate nodes is the number of output nodes × 2 + 1 Consists of Here, one error correction node was used for each of the input layer and the intermediate layer. FIG. 6 shows an execution example of recognizing handwritten Hangul characters in a neural network as shown in FIG.

The neural network unit 60 comprises the neural network described above. The number of nerves constituting the neural network unit is the same as the number of candidate character groups in the candidate character group storage unit. The same is true. The neural network unit 60 can increase the degree of recognition through learning on samples belonging to the code character group.

The detailed classification unit 70 finally searches one character from the candidate character group corresponding to the input character pattern searched from the large classification unit 50 through the neural network unit 60 to recognize handwritten characters. After the recognition, a control signal is provided to the pre-processing unit 20 so that the next character pattern to be recognized is classified into the large classification unit 50.
To supply.

In the present invention, a handwritten character corresponding to a set of completed Hangul 2350 characters or a subset thereof is recognized. A recognition system was constructed for Hangul 990 characters commonly used for performance experiments of the method and apparatus of the present invention. The components of the configured system are 99
Zero 7 × 7 × 4 dimensional representative vectors, 990 candidate character groups consisting of 5.7 candidate characters on average, and an average size of (7 × 7 × 4) × 12.4 × 5.7 There are 990 neural networks, an input unit preprocessing unit, a large classification unit, and a detailed classification unit. The constructed system was transferred to Pohang U.
diversity of science and
2nd International Conference on Handwritten Character Data (DW Kim, SY Bang, Hangeul Handwritten Image Database PE92, Document Analysis and Recognition, published in Technology, Tsukuba, Japan, October 1993, (pp. 470-473), 693,000 characters corresponding to about 70% of the total 99000 characters were trained, and when the remaining 29,000 characters were recognized, the correct recognition rate was about 90%.

[0034]

According to the method and apparatus for recognizing handwritten Hangul characters according to the present invention, a large amount of handwritten information can be obtained by automatically recognizing Hangul characters created on arbitrary paper using an arbitrary handwriting tool. Tasks such as automatically inputting data into a computer and automatically sorting a large amount of mail can be performed. Therefore, it is possible to automate the input operation of the contents of the existing document and the newly generated information that depend on the manual operation, and to generate many benefits in terms of time, human power, and money.

[Brief description of the drawings]

FIG. 1 is a block diagram of a handwritten character recognition system according to the present invention.

FIGS. 2A to 2C are diagrams for explaining feature extraction from an input character image in the large classification unit shown in FIG. 1;

FIG. 3 is a diagram for explaining a method of forming a candidate character group of a recognition target character in the large classification unit and the detailed classification unit shown in FIG. 1;

FIG. 4 is a diagram for describing extraction of horizontal and vertical features from an input character image in a detailed classification unit.

FIG. 5 is a diagram for explaining a neural network configuration of a neural network storage unit.

FIG. 6 is a diagram illustrating an example of a result of performing handwritten character recognition according to the present invention.

[Explanation of symbols]

 Reference Signs List 10: input unit, 20: preprocessing unit, 30: candidate character group storage unit, 40: representative vector storage unit, 50: large classification unit, 60: neural network storage unit, 70: detailed classification unit

Continuation of the front page (72) Inventor Li So-gyu 663-15 Dong-dong, Mapo-gu, Seoul, Republic of Korea (56) References JP-A-4-286087 (JP, A) JP-A-2-59883 (JP, A) JP-A-5-324838 (JP, A) JP-A-3-182986 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06K 9/00-9 / 82

Claims (7)

(57) [Claims] 1. A step of inputting a character image having an I × J size, a step of calculating a feature vector including a finite number of coordinates as a characteristic of the pattern of the input character image by a statistical method, Searching for a representative vector most similar to the calculated feature vector in a representative vector set corresponding to the target character; and obtaining a recognition target character corresponding to the searched representative vector and a candidate character group including the character. And outputting one character of the obtained candidate character group as a final recognition result through a neural network configured using a feature vector including a finite number of coordinates for the character pattern of the candidate character group. The method for calculating the feature vector includes: an input image of I × J size is vertically stripped into M strips;
Divided into N strips in the horizontal direction and M × N
Non-linear division into cells of size (I ≧ M, J ≧ N,
I, J, M, and N are integers greater than 0), and the vertical component value FV and the horizontal
The direction component value FH is calculated by the following equation: Fv = Bv / Bp FH = BH / Bp (where Bp is the number of all black pixels in the input image, and Bv
Is the number of black pixels belonging to the vertical component in the cell
Where BH is a black picture belonging to the horizontal component existing in the cell.
Get 2MN number of feature values determined by a number of elementary)
And also for the pattern obtained by rotating the input image by 45 °.
Applying the same method as above, further 2MN feature values are obtained.
Calculating and 4MN for the vertical, horizontal, diagonal and reverse diagonal components
Feature values as the feature vector of the input character pattern.
Providing handwritten character recognition.
Method. 2. The handwritten character recognition method according to claim 1 , wherein the vertical and horizontal components are a set of runs having a length larger than a threshold value T among runs in the vertical and horizontal directions. 3. The handwritten character recognition method according to claim 2 , wherein the threshold value T is dynamically obtained according to the thickness W of the character image to be recognized. 4. The thickness of the character image is expressed by the following equation: W = Bp / (Bp−Wp) (where Bp is the number of black pixels in the input image, and Wp is 2 × 4. The handwritten character recognition method according to claim 3 , wherein the number of black pixels included in three adjacent pixels in two windows is obtained. From wherein said candidate character group is a character Ci, the character the distance to the representative vector of the feature vector Cj to be within the character it is erroneously recognized as close to Ci than the distance to the representative vector of the feature vector Cj 2. The method according to claim 1, further comprising the steps of: collecting information of the following, arranging the characters in order of frequency, and deciding a number of higher-order characters as a candidate character group for Ci. Handwritten character recognition method. 6. A feature vector for calculating an output of the neural network includes a step of non-linearly dividing a character image into N × M in both horizontal and vertical directions, and each black pixel Px, y (0 ≦ 0). x ≦ W, 0 ≦ y ≦ H;
W is the width of the character image, H is the height of the character image) and the length RLHx, y of the horizontal run to which the character image belongs and the length RLVx, y of the vertical run.
And the horizontal contribution DC at each Px, y
Hx, y and the vertical contribution DCVx, y are determined by the following equation: DCHx, y = RLHx, y / (RLHx, y + RLVx, y) DCVx, y = RLVx, y / (RLHx, y + RLVx, y) And finally calculating a 2 (N × M) -dimensional feature vector by calculating an average value of DCHx, y and DCVx, y for all black pixels present in the non-linearly divided N × M cells. The above process is applied to white pixels, and another 2 (N × M) -dimensional feature vector is calculated.
2. The handwritten character recognition method according to claim 1, wherein an (N × M) -dimensional feature vector is obtained. 7. An input unit for inputting a handwritten character image as a binary character image, a preprocessing unit for separating the input character image and outputting one character at a time; A candidate character group storage unit storing a candidate character group consisting of candidate characters including characters; a representative vector storage unit storing all feature vectors corresponding to all the recognition target characters; The representative vector most similar to the determined feature vector is searched from the representative vector storage unit, and a character to be recognized for the searched representative vector and a candidate character group for the recognized character are determined by the candidate vector. A large classification unit searched from a character storage unit; a neural network storage unit including a neural network for each candidate character group stored in the candidate character group storage unit; After recognizing one character of the candidate character group searched from the large classification unit through the neural network stored in the unit as a final recognition result, a control signal is provided to the preprocessing unit to perform next recognition. And a detailed classifying unit that supplies the characters until all the characters in the input character image are recognized. The large classifying unit converts the input image of I × J size into M vertical images. strip
Divided into N strips in the horizontal direction and M × N
Non-linear division into cells of size (I ≧ M, J ≧ N,
(I, J, M, N are integers greater than 0)
And the vertical component value FV and the horizontal
The direction component value FH is calculated by the following equation: Fv = Bv / Bp FH = BH / Bp (where Bp is the number of all black pixels in the input image, and Bv
Is the number of black pixels belonging to the vertical component in the cell
In, BH black picture that belong to the horizontal components present in the cell
Get 2MN number of feature values determined by a number of elementary)
Feature value calculation unit, and a pattern obtained by rotating the input image by 45 °.
Applying the same method as above, further 2MN feature values are obtained.
A rotation feature value calculation unit to calculate , and 4MN for the vertical, horizontal, oblique and reverse oblique components
Feature values as the feature vector of the input character pattern.
And a feature vector providing unit to be provided.
Handwritten character recognition device.