KR940001048B1 - On-line hand written character recognition method - Google Patents

Abstract

The on-line hand-written letter recognizing method in a method for recognizing hand-written Korean letter includes the steps of obtaining a character dot column by performing a pre-processing in a stroke unit without respect to Korean data from a tablet, giving a stroke type code according to the shape of the stroke to extract a characteristic code of the stroke, determining a candidate horizontal vowel and a candidate vertical vowel to sort in Korean prototype by determining the presence of a last phoneme, performing a group classification for searching for searching out initial phoneme, middle phoneme and last phoneme tree classifier to perform a detailed classification using a DP matching, thereby recognizing the letter. Thus, recognizing time is fast while occupying less storage area, and variation of the hand-writting motion apt to occur during writting the letter is absorbed for recognition.

Description

Online handwriting recognition method

1 is a block diagram schematically showing the configuration of a system for realizing the present invention.

2 is a flow chart for explaining the overall process of the present invention.

3 is a diagram showing code definitions in eight directions (DIR-0 to DIR-7) of straight lines forming a stroke.

4 is a diagram showing the definition of a stroke type code.

5 is a diagram showing an example of a code according to the definition of a stroke type code.

Figure 6 shows the feature code for 'ㅎ' consonants.

Figure 7 is a tree diagram of the initial, neutral, and final library.

8 is a view for explaining the elastic matching.

The present invention relates to an on-lind character recognition method, and more particularly, to a method of recognizing a Hangul character of a hand-written letter. Recently, the necessity for a handwritten character recognition system is increasing in many computer-related fields.

In particular, the development of a recognition method that is not limited to the writing method and writing speed of the writer is urgently required.

The reason for this is that the input method by the commonly used keyboard must remember the keyboard layout of the keyboard and requires additional training for the expert input. Because.

Online handwriting recognition is emerging as a convenient input method that can replace the keyboard input method, and in particular, it is increasing its importance as a core technology of a micro fan computer. Conventional techniques for recognizing unknown characters recorded by a stylus on a tablet, which is an on-line coordinate input device, have been previously normalized and recognized in character units without letter separation.

In this method of character recognition, a plurality of character recognition dictionaries must be provided in order to recognize a character, and the memory capacity must increase in proportion to the number of recognition dictionaries due to the large number of character recognition dictionaries. In addition, the conventional method takes a long time to recognize, has a sensitive reaction to the change of the writing movement that is easy to occur when writing the letter has a disadvantage that can be recognized by writing only in a predetermined stroke order. SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned problems of the prior art and to provide a method of recognizing a handwritten Korean character which reduces the number of memories required in hardware configuration and has a high recognition rate for characters.

In order to achieve the above object, a hardware system for realizing the present invention comprises a central processing unit performing overall processes for character recognition, a memory storing character data, a recognition library and a recognition program, and a tablet for generating online data. And at least an output device for indicating the recognition result.

As a technical means for achieving the object of the present invention in a system having a hardware configuration as described above, the method of the present invention comprises a pre-processing step of linearizing data obtained from a tablet to obtain a feature point sequence, and this preprocessing. A feature extraction step for determining the characteristics of the stroke using the feature point range obtained in the step, a type classification step for classifying the six types of Hangul using the characteristics of the stroke, and a character according to the characteristics of the stroke and the type of the determined Hangul. Recognition step of performing the recognition.

1 is a block diagram schematically showing a hardware structure for embodying the present invention, in which reference numeral 11 denotes a central processing unit for performing the steps according to the present invention, and reference numeral 12 denotes a character data and a recognition library and a recognition program. A memory for storing, 13 shows a tablet for generating data on characters recorded by a stylus pen, and 14 shows an output device showing character recognition results according to the present invention.

2 is a view showing an embodiment of the present invention, and the present invention will now be described in detail with reference to FIG.

(1) Pretreatment

The Hangul data input from the tablet 13 has a characteristic of being input by being divided into stroke units, and consonants and vowels are composed of the set of strokes, and one letter is composed of the consonants and vowels.

Therefore, to take advantage of these features, pretreatment is performed on a stroke basis.

The preprocessing of the Hangul Recognition Algorithm is a data smoothing that eliminates or reduces rapid fluctuations in data for sample sequences, and an algorithm that shortens only recognizable data.

In this algorithm, a piecewise-linear segment is obtained by eliminating the bleeding that occurs frequently in Korean by calculating the angle between the line segments that have two feature points. .

First, the slope m of a straight line connecting the starting point P ₁ and the k + 1 th P _{k + 1} is obtained, and then a tolerance region 2T having a width of T on both sides of the straight line is set.

The slope of the line connecting m = P ₁ and P _{k + 1} (y _{k + 1} -y ₁ ) / (x _{k + 1} -x ₁ ) (x _j , y _j ) = reference point for the piecewise-linear segment (x _{j + ik} , y _{j + ik} ) = When it is the ikth point from the reference point, each successive point (x _j + i _k , y _{j + ik} ), i = 1,2,3,... For, if this point enters the allowable area and the next point (x _{j + [i + 1] k} , y _j + _{[i + 1] k} ) enters the allowable area, then remove (x _{j + ik} , y _{j + ik} ). .

The Chebyshev approximation method is used to check which point (x _{j + ik,} y _j + _ik ) falls into the allowable area.

If this point satisfies the following formula (1), it enters the allowable area. If | m | ≤

m (x _{j + ik} -x _j )-(y _{j + ik} -y _j ) | ≤ (0.960 + 0.398 | m |) T

If | m |> 1,

[Equation 1]

If (X _{j + [n + 1] k} , y _{j + [n + 1] k} ) is the first point that does not fit in the allowable area, this point is the new reference point and m is also obtained as the reference point.

Where T controls the degree of data smoothing, and k can be regarded as an acceleration factor for curve fitting. T and k both control the extent of data thinning. Obtain the eight-way code (see Figure 3) of the line segments between the feature points, and if the direction codes of the adjacent line segments are the same, discard the feature points and combine them into one line segment.

And because most of the Hangul feature of Hangul occurs at the end of the stroke, the difference between the angles of the last two segments of each stroke is calculated to remove the hang, and if it is smaller than the threshold D ₁ , the last segment is discarded.

(2) Feature Extraction Process

Using the feature sequence in which the preprocessing process described above is performed, first, the stroke is classified as a straight stroke or a curved stroke.

If the stroke consists of two feature points, it is a straight stroke, otherwise it is divided into a curved stroke.

The straight stroke is assigned one of the stroke shape codes from 1 to 8 according to the 8-way code, and the curved stroke is the clockwise direction of rotation of the second segment relative to the first segment from the two adjacent straight segments that make up the stroke. ) And 16 stroke shapes defined in FIG. 4 according to the number of changes in the rotational direction of the linear segments constituting the stroke, and the initial rotational direction and the direction of the first segment of the stroke, defined in a counterclockwise direction (+). It is given one of the stroke-types.

In this case, if the line segment rotates from-direction to-direction continuously, it is indicated as one-direction so that the code becomes 8 or 12 depending on the direction of the first line segment.

Also, a stroke with a stroke that changes from-to + is marked with-+ and the code is 9 or 13, depending on the direction of the first segment of the stroke. An example of the code is in FIG.

The stroke code of the line connecting the end point of the stroke and the start point of the next stroke is assigned the eight-way code defined in FIG. Therefore, when one letter is composed of n strokes, the feature code consisting of n stroke shape codes and (n-1) stroke codes is as follows.

An example of a feature code representation for 'ㅎ' is shown in FIG. 6. (f ₁ , f ₂ , f ₃ ,…, f _2n-2 , f _2n-1 )

from f _index to index = odd → f _index = stroke shape code

index = even → f _index = Stroke code (0 to 7)

(3) Type Classification Process

Hangul has a feature that contains at least one neutral stroke within a letter. Therefore, it is necessary to find a neutral vertical stroke to classify the type of Hangul by analyzing the existence of verticality, and the structure of horizontal stroke and vertical stroke.

1) Horizontal and vertical vowel determination

The basic features of Hangul characters for Hangul type classification are as follows.

Iii) The vertical vowel is the rightmost of the strokes with the eight-way code defined in Figure 3 in Figure 3, and is longer than other strokes.

Ii) Horizontal vowels are the widest in most cases, although the eight-way code in the text has an exception among zero strokes (exception: letters written in vertical and horizontal strokes in ㅗ and ㅛ).

Based on these characteristics, the horizontal and vertical strokes are obtained using the following equation.

[Equation 2]

[Equation 3]

The meanings of the symbols used in the above formula are as follows.

The character image has a height H and all widths W including all black pixels as a minimum rectangle.

DIR _i represents the eight-direction code of the first line segment of the i-th stroke, and S _i (x) and S _i (y) represent the x and y coordinates of the first pixel of the i-th stroke, respectively.

SW _i represents the width of the i-th stroke.

Among the strokes satisfying equation (2), the horizontal stroke is determined at the leftmost stroke of the stroke's x-coordinate. If (3) is satisfied, it is determined by vertical stroke.

Obtain the horizontal and vertical vowels using equations (2) and (3) and verify that they are the correct horizontal and vertical vowels.

For example, if 'B' satisfies the following equation (4), the stroke determined by the vertical vowel is determined as the general stroke, and only the horizontal vowel is determined.

[Equation 4]

2) Determination of presence of species

The presence or absence of verticality in the case of vertical vowels is considered to be a finality if the stroke code at the start of the next stroke at the end of the vertical vowel is one of 2, 3 or 4.

If there is a horizontal vowel, the stroke that satisfies the expression (5) shown below by the first segment code of the strokes following the horizontal vowel is regarded as a stroke included in the neutral, and the stroke that does not satisfy the expression (5) is the final stroke. Decide

Where SH _h-mo is the height of the candidate horizontal collection stroke.

[Equation 5]

(4) recognition process

Recognition algorithm is composed of three elementary recognition parts of primary, neutral, and final.

In each phoneme recognition part, large classification is performed by each tree classifier consisting of tree, tree, and tree based on feature code extracted from pre-entered sample data, and by DP (Dynamic Programming) matching method. Recognize it with a detailed classification.

If one character is entered, the initial recognition recognizes the stroke type code and the inter-stroke code, which are the feature codes of the strokes up to the candidate horizontal vowels or the candidate vertical vowels obtained from the type classification, to the initial tree classifier. Recognized strokes are sent to the neutral tree classifier by sending the feature codes of the remaining strokes or the final strokes according to the Hangul type to the neutral tree classifier. Send it to the child to recognize the species.

As a result, the input data is recognized as one character by a combination of their phonemes recognized in the initial / neutral / final recognition part. This will be described in more detail as follows.

1) Major Category

Each of the initial / neutral / finality of the input data can reduce the recognition time by using a large classification that yields a set of several similar elements that are likely to be finally recognized.

Feature codes used in large classifications use stroke type codes and stroke-to-stroke codes.

The classifiers of the initial / neutral / single tree, which have previously obtained such feature codes for training handwriting data and stored them in a tree form, are searched and classified into candidate character sets having the same feature codes.

The order of searching for the phoneme tree classifier is as follows.

Step 1: Navigate the tree with the stroke shape code that characterizes the stroke.

Step 2: Navigate the tree with stroke-to-stroke code indicating the direction from the end of the stroke to the start of the next stroke.

Step 3: Repeat step 1 and step 2 when there are no feature code values or until there are no nodes to search in the tree.

The characters entered are searched in the above order and continue until no more nodes can be searched in the middle of the search, so that they are largely divided into candidate character sets with one or more phonemes at the last searched node. .

If there is no candidate character in the searched node, the result is misrecognized, and when a candidate character set having one phoneme is encountered, the phoneme is determined as the recognized phoneme.

If a candidate character set having two or more phonemes is obtained, the recognized phoneme is determined as a recognition phoneme through detailed classification using a DP matching method.

The elementary, primary, neutral, and triclassifiers are in Figure 7.

2) Detailed classification by DP matching

Cursive changes in shape depending on the speed and method of writing the same letter.

As a prototype matching method that can adapt to such a change, there is an elastic matching method.

This way, when matching an unknown character to a prototype, it does not attempt to match one-to-one for each point, but rather elastically.

By doing this, you can overcome slight differences in shape or the number of dots.

8 is an example of the number 3 elastic matching.

For detailed classification using the DP matching method, prototypes are consonant units, and each consonant may have one or more prototypes.

The DP matching method is to find the prototype with the smallest distance by calculating the distance (similarity) to each prototype for the unknown input.

Two features are extracted for each feature point of the feature sequence after preprocessing.

The first feature of the i th feature point V _i is the x coordinate x _i and the second feature is the y coordinate y _i .

The feature vector for a stroke with N feature points is as follows.

[Equation 6]

When matching a prototype with an unknown input character, the distance between the jth point of the kth prototype and the ith point of the unknown input character is calculated using the following formula:

d (i, j, k) = | y _i -y _ik | + | x _i -x _ik |

The distance of prototype k, where the matching distance between the prototype and the unknown input character is the smallest, is calculated as follows.

[Equation 7]

Where N is the number of feature points of the unknown input character, w is the function that maps the index of the unknown input character to the prototype index, and D _k is the distance between the kth prototype and the unknown input character. w has a boundary condition and a continuity condition to be a monotonic inceasing function.

The boundary condition of w is

w (1) = 1

w (N) = M _k

Where M _k is the number of feature points in the kth prototype.

In addition, the continuous condition of w is as follows.

w (i + 1) −w (i) = 0,1,2 for i = 1,... , N-1

Equation (7) is efficiently obtained by the dynamic programming method using the recursive relation as follows.

[Equation 8]

Where D (i, j, k) is the cumulative distance between the i th feature point of the unknown input character and the j th feature point of the prototype.

Starting from D (1,1; k) = d (1,1: k), we find D (N, M _k ; k) to be D _k .

Because of the characteristics of Hangul, 11,172 ((19 initial)) * (21 neutral) * (27 final or 27 unvoiced) characters are possible if you choose all the possibilities of Jamo. These features have to be stored in a computer, and it takes a long time to recognize, but if the recognition by phonemes according to the present invention, only for the first, neutral, and final kinds of 19, 21, 27 kinds of phonemes, respectively, and compared Recognition time is fast and requires less storage.

In addition, in the present invention, since the DP matching is used during the detailed classification, it is possible to absorb and recognize the variation of the writing movement which is easy to occur when writing the text.

Claims (7)

A method for recognizing handwritten Hangul characters, the method comprising: obtaining a feature sequence by performing preprocessing on a stroke basis for Hangul data provided from a tablet; The feature interpolation code distinguishes whether a character stroke is a straight stroke or a curved stroke, assigns a stroke shape code to the stroke according to the stroke shape, and stroke code for a line segment connecting the end point of the stroke and the start point of the next stroke. Extracting the feature code of the stroke; Analyzing the constituent structures of the horizontal stroke and the vertical stroke to determine candidate horizontal vowels and candidate vertical vowels, and determining the presence or absence of the finality based on the neutral horizontal stroke and the neutral vertical stroke and classifying them into Korean type; Characteristic recognition is performed by searching the initial tree classifiers, the neutral tree classifiers, and the final tree classifiers for the first, neutral, and final trees formed by the feature codes, and performing detailed classification using DP matching. On-line handwritten character recognition method comprising the step of. The on-line cursive of claim 1, wherein the stroke form is divided into the straight stroke when the stroke is composed of two feature points, and the curved stroke when the stroke is at least three or more feature points of the stroke. Character recognition method. The second line segment according to claim 1 or 3, wherein the straight line stroke is assigned one of the stroke shape codes corresponding to eight directions, and the curved line segment is based on a first segment of two adjacent straight line segments constituting a stroke. On-line characterized in that it is given one of the 16 stroke shape codes according to the number of changes in the rotation direction and the initial rotation direction and the direction of the first segment of the straight line segments constituting the stroke by defining the rotation direction of How to recognize handwritten characters. The method of claim 1, wherein the determining of the horizontal and vertical collections is performed. DIR _i = DIR-0 S _i (x) <Th ₁ SW _i ＞ Th ₂ W / H ＞ 1.5 Where H is the height of the character, W is the width of the character, DIR _i is the eight-direction code of the first segment of the i-th stroke, S _i (x) is the x-coordinate of the first pixel of the i-th stroke, and SW _i is the i-th And among the strokes satisfying the width of the stroke, Th ₁ and Th ₂ are the thresholds, determine the horizontal collection of the stroke with the leftmost x-coordinate of the stroke, S _i (y) <H * 2/5 H / W ＞ Th ₃ Wherein (S _i (y) is the y-coordinate of the first pixel of the i-th stroke, Th ₃ is the threshold value). The method of claim 1, wherein the determination of the presence or absence of the species indicates that when the vertical vowel is present, the species exists if the stroke code of the start point of the next stroke at the end of the vertical vowel is one of 2, 3, and 4; If the horizontal bar is present, the first segment code of the strokes following the horizontal bar is DIR _h-mo = 2 or DIR _h-mo = 3 SH _h-mo / SW _h-mo <Th ₄ Here, the stroke satisfying (where SH _h-mo is the height of the candidate horizontal collection stroke, DIR _h-mo is the direction code of the candidate collection, Th ₄ is the threshold value) is determined as the stroke included in the neutral, and And if it is not satisfied, judging by the final stroke, it determines that the finality exists. The method of claim 1, wherein the searching for the tree classifier comprises: searching the tree with the stroke shape code, searching the tree with the stroke code indicating the direction of the start point of the next stroke from the end point of the previous stroke, And repeating the steps until the stroke shape code and the stroke code, which are feature codes, or there are no nodes to search in the tree. The method of claim 1, wherein the major classification process sends the feature code for each of the strokes up to the candidate horizontal vowel and the candidate vertical vowel to the first tree classifier, recognizes the initial tree, and extends all or the final strokes of remaining strokes. Sending the feature code to the neutral tree classifier to recognize the neutrality, and if the species exists, sending the feature code of the remaining stroke to the seed tree classifier to recognize the species.