JPH0443316B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Application of the Invention] The present invention relates to an online handwritten character recognition method that sequentially performs recognition processing using information about the process of handwriting characters.

[Background of the invention]

Conventional online handwritten character recognition methods can be roughly divided into the following four methods.

The first method considers the change in motion of the pen point when handwriting characters as a set of one-dimensional waveforms that are decomposed into orthogonal coordinate components, approximates this one-dimensional waveform by orthogonal function expansion, and uses the coefficients of the orthogonal function. This is a method for recognizing characters.

The second method approximates each stroke that makes up a character as a connection of directionally quantized vectors in eight directions, classifies each of these approximated strokes into several basic strokes, and creates characters from the combination of basic strokes. This is a method to recognize

The third method classifies each stroke that makes up a character into several basic strokes, creates a feature table that describes the character using stroke end points, intersection points, etc., and compares the input character with this feature table. This method recognizes input characters by comparison.

The fourth method is a method proposed in Japanese Patent Publication No. 57-6151, "Online Recognition Processing Method for Handwritten Characters," in which the coordinates of the starting point, the coordinates of the ending point, and the coordinates of the starting point and ending point of each stroke of the input character are calculated. Extract the coordinates of the midpoint of the center position, use each point as a feature point, determine the sum of distances from each feature point in a standard character prepared in advance, and create a standard character whose sum of distances is the minimum value. This is a method that recognizes the above input characters.

Each of these conventional online handwritten character recognition systems has the following problems. In other words, in the first method, the accuracy of approximation using orthogonal functions is not necessarily good for characters whose main components are straight lines for kanji and katakana, and the recognition rate decreases because the topological shape of the characters cannot be grasped. was inviting.

In the second method, recognition rate decreases due to classification errors in basic strokes that occur when each stroke constituting an input character is classified into basic strokes. Furthermore, since detailed descriptions are required for all characters to be recognized, this description requires a great deal of effort.

In the third method, as with the second method, the recognition rate decreases due to basic stroke classification errors, and it takes a lot of time and effort to create a feature table that describes in detail all the characters to be recognized. Requires memory requirements.

The fourth method is a method proposed to improve the problems of the first to third methods, but it has the following problems.

(1) Each stroke is approximated by three points: a start point, a middle point, and an end point, but as shown in FIG. 1, the approximation accuracy is not necessarily good for strokes that have bending points. A stroke like a in FIG. 1 becomes like b in three-point approximation.
This is because the midpoint of a stroke does not necessarily correspond to the bending point of the stroke.

(2) If a stroke includes a splash, the coordinates of the end point will not only change depending on how you write it, but also the coordinates of the midpoint will change due to the influence of the splash. It will change. This is because the midpoint is extracted as the center point of the stroke line segment.

Variations in the midpoint and endpoint feature points depending on the author and writing style are a factor in lowering the recognition rate.

(3) Regardless of the shape of the stroke, the starting point,
Since the stroke is approximated by three points, the midpoint and the end point, no information on the shape of the stroke can be obtained. Therefore, the candidate characters are a standard character group selected based on the stroke number information of the input character, and the number of candidate characters is large, and the matching process takes a lot of time.

(4) Even strokes that can be approximated by a single straight line, that is, strokes that can be sufficiently approximated by only two points, the start point and the end point, are approximated by three points as described above, so there is redundancy in the feature points, and the standard The memory capacity for storing characters is increasing.

[Purpose of the invention]

An object of the present invention is to eliminate the above-mentioned problems and provide an online handwritten character recognition method suitable for recognizing characters such as katakana and kanji that are mainly composed of linear components.

[Summary of the invention]

In order to achieve the above object, the present invention extracts the start point coordinates, end point coordinates of each stroke of an input character, and if the stroke has a bending point, extracts the bending point coordinates and treats each point as a feature point. , select as candidate characters a group of standard characters that have the same number of strokes as the input character and each stroke has the same number of feature points as the input character from among standard characters prepared in advance;
A method characterized by determining the sum of distances between each feature point of the input character and each standard character serving as the candidate character, and recognizing the standard character for which the sum of the distances is the minimum value as the input character. It is.

[Embodiments of the invention]

FIG. 2 shows a functional block diagram of one embodiment of the invention. In the figure, numeral 1 is a character information input device, a so-called tablet, 2 is a preprocessing section, 3 is a feature point extraction section, 4 is an inter-pattern distance calculation section, 5 is a minimum distance detection section, 6 is an output terminal, and 7 is a standard This is a pattern (standard character) memory section.

The principle of the present invention is as follows. first,
Input characters from the character information input device 1 are subjected to coordinate transformation of each writing point in the preprocessing unit 2 so that the center of gravity of the input character becomes the origin (hereinafter referred to as position normalization). Further, the size is normalized so that the average value of the distance between each writing point and the center of gravity becomes a constant value. Furthermore, the preprocessing unit 2 detects the number of strokes of the input character. This number of strokes is determined from among the X-axis and Y-axis coordinate values of each writing point obtained from the character information input device 1 and the Z-axis information about whether or not the input pen is pressed against the input surface of the character information input device 1. Z
Based on the axis information, for example, the change from crimping (Z=1) to detachment (Z=0) is obtained by counting the change over one character. Information on this number of strokes 20
1 is candidate character selection information for the input character.

After this preprocessing, the feature point extraction unit 3 extracts each stroke of the input character from the start point, end point, and if there is a bending point in the stroke.
One of the bending points is extracted as a feature. The starting point is the point at which the stroke begins, and the end point is the point at the end of the stroke. Also, the bending point is
The maximum distance H between the straight line connecting the start point and end point and the writing point within the stroke is H/with respect to the distance L between the start point and end point.
When there is a writing point for which L is equal to or greater than a certain value, this corresponds to the writing point that provides the maximum distance H.

Therefore, for a nearly straight stroke, two points, the starting point and the ending point, are extracted, and for a curved stroke, three points, the starting point, the bending point, and the ending point, are extracted as feature points.

This feature point extraction unit outputs feature point number information 301 of each stroke as second candidate character selection information for the input character.

The inter-pattern distance calculation section 4 calculates the inter-pattern distance between the input character described above and a pattern that has been pre-processed and feature points extracted in the same manner as the input character and stored in the standard pattern memory section. At this time, the target standard pattern has the same number of strokes as the input character, and the number of feature points of each stroke is the same as the input character, based on the stroke number information 201 and the feature point number information 301 of each stroke. Select the feature points that are equal to the feature points of each stroke corresponding to the stroke order, and use these as the target.

Therefore, if the recognition target category is represented by θ and the distance between patterns is d〓, then d〓 is obtained as a relaxation of the distance between corresponding feature points in stroke order. The distance between the feature points may be any amount that represents the distance between the feature point A of the input character and the feature point B of the corresponding standard pattern, such as the Euclidean norm between A and B, the city block distance, etc. is used.

The minimum distance detection unit 5 detects the minimum value from the inter-pattern distance d〓 calculated in this way, recognizes the category θ indicating the minimum value as an input character, and outputs the code corresponding to the category θ to the output terminal 6.
Output to.

FIG. 3 shows a specific embodiment of the present invention.
1 is a tablet, and 12 is an input pen, which corresponds to the character information input device 1 shown in FIG. 13 is a tablet interface section, 81 is a microprocessor,
82 is a random access memory (hereinafter referred to as RAM), and 83 is a read-only memory (hereinafter referred to as ROM).
), 61 is an output interface section, and 6 is an output terminal.

The microprocessor 81 operates according to the above-mentioned principle.
The program to run is
It is stored in ROM83.

Information about each writing point of characters written on the tablet 11 with the input pen 12 is taken into the microprocessor 81 via the tablet interface section 13. Then, each piece of writing point information corresponding to one character is stored in a predetermined area of the RAM 82.

Then, for each writing point information for this one character,
First, preprocessing is performed. Since this preprocessing is not directly related to the present invention, it will be omitted here. Details are described in Japanese Patent Publication No. 57-6151. Moreover, the outline was described in the above-mentioned explanation of the principle.

Each stroke information of the preprocessed input character is again
It is stored in a predetermined area of RAM82.

After this preprocessing, feature points of each stroke are extracted. As explained in the above principle, the characteristic points of each stroke are the starting point,
The two points are the end point, and in the case of a curved stroke, the three points are the start point, the bend point, and the end point.

Figure 4 shows the principle of extracting bending points. If the writing points of a certain stroke are P ₁ to _{P 10} in stroke order, P ₁ corresponds to the starting point and P ₁₀ corresponds to the ending point.

Extraction of bending points is performed as follows.

(1) Distance h′ _i between the straight line SE connecting the starting point S=P ₁ and the ending point E=P ₁₀ and each writing point Pi (i=2 to 9) in the stroke
seek.

(2) Detect the maximum value H′ of distance h′ _i .

(3) Find the distance L' between the starting point S and the ending point E.

(4) When H′/L′ is greater than the threshold TH, the pen point that gives H′ is the inflection point.

Here, if the coordinate values of each writing point Pi are (xi, yi), then the distance h′ _i between the straight line SE and each writing point Pi is determined as follows.

Line SE is Ax + Bx + C = 0 (1) However, A = (y ₁ - y ₁₀ ) B = - (x ₁ - y ₁₀ ) C = (x ₁ - x ₁₀ ) y ₁ - (y ₁ - y ₁₀ ) x ₁ , h′ _i is Find it as.

Further, the distance L' between the starting point S and the ending point E is determined as L'=√( ₁ − ₁₀ ) ² +( ₁ − ₁₀ ) ² =√ ² + ² (3).

Therefore, if H=h' ₅ , H/L' becomes H'/L'=|Ax ₅ +By ₅ +C|/A ² +B ² (4).

Then, when H′/L′ is greater than the threshold value TH,
The writing point Ps is extracted as an inflection point and used as a feature point.
Otherwise, it is determined that this stroke does not have a bending point, and the two points, the starting point and the ending point, are extracted as features.

FIG. 5 shows a specific processing flow.

Here, in order to reduce the amount of processing for calculating h' _i and L' mentioned above, instead of h' _i , only the numerator on the right side of equation (2) is calculated as h _i = |Axi + Byi + C (5), The maximum value H is detected. Also, instead of L', L=A ² +B ² (6) is calculated, and H'/L' is obtained as H/L.
As is clear from equation (4), this H/L is H′/
It is equivalent to L′.

In (1) of Figure 5, a straight line connecting the starting point and ending point
Calculate the coefficients A and BC of the SE linear equation (1).

Then, the distance h _i ((5)
Equation) is calculated in (2). Note that if the number of pen points of a stroke is K, then i is 1 to K-1. Then, in (3), the maximum value H of h _i is determined. In (4), the distance L between the starting point and the ending point
Find (formula (6)).

The ratio H/L of H and L obtained in this way is the threshold value TH
Use (5) to determine whether the value is greater than or not. If the determination result in (5) is negative, it is determined that there is no bending point within the stroke, and the number of feature points P(s) is set to 2 in (6). Note that s is a stroke number, which is assigned according to stroke order. Then, as the feature point of the sth stroke,
P _s,1 and P _s,2 are the starting point or writing point P ₁ and the ending point or writing point
It is stored in a predetermined area of the RAM 82 in correspondence with the coordinate value of P _k . Note that Pa, b represents the b-th feature point of the a-th stroke. If the result in step (5) is positive, it is determined that a bending point exists within the stroke, and in step (7), the number of feature points P(s) is set to 3. In (9), as the feature point of the sth stroke,
P _s,1 , P _s,2 , P _s,3 are made to correspond to the coordinate values of the starting point, the writing point Pi that gives H, and the ending point, and are stored in a predetermined area of the RAM 82 .

The above procedure is repeated for each stroke of one character.

The feature points of the input characters extracted in this way are used in the inter-pattern distance calculation section 4 to calculate the standard pattern and the inter-pattern distance.

At this time, the target standard pattern has the same number of strokes N of the input character, and the number of feature points P(s) (s=1 to N) of each stroke is equal to that of the input character.

Here, if the recognition target is the clear sound of katakana characters, it can be classified as shown in FIG. 6 according to the pattern of the number of strokes and the number of feature points P(s). For example, for "A", the number of strokes is 2, the number of feature points is 3 because the first stroke (in stroke order) has an inflection point, and 2 because the second stroke does not have an inflection point. In FIG. 6, the pattern of the number of feature points P(s) is expressed as (2, 2).

Similarly, for "Se", the number of strokes is 2, and the number of feature points P(s) pattern is (3, 3).

Therefore, if standard patterns are classified and stored in the standard pattern memory section 7 in advance as shown in FIG. can be limited.

FIG. 7 shows a processing flow of inter-pattern distance calculation and minimum distance detection. In this figure, it is assumed that each standard pattern is classified and stored according to the number of strokes, and the pattern information of the number of feature points P(s) of each stroke is given and stored together with the coordinate values of the feature points.

In FIG. 7, (1) corresponds to selection of candidate characters corresponding to input characters, (2) corresponds to inter-pattern distance calculation, and (3) corresponds to minimum distance detection.

First, the minimum value dmin of the distance between patterns is initially set to a certain value β. This β may be a value larger than the distance between patterns that is normally obtained. Then, in (1), among the standard patterns equal to the number of strokes N of the input character, the number of feature points P(s) of each stroke of the input character and the number of feature points of each stroke of the standard pattern P〓 ^m (s)(s =1~N) and compare s=
For 1 to N, select a standard pattern such that P(s)=P〓 ^m (s). This is done by repeating the judgment in (11) several times. If the result in (11) is negative, perform the next standard pattern. Next, the inter-pattern distance from the selected standard pattern is calculated in (2). If the inter-pattern distance for the standard pattern category θ _n is dθ _n , first, in (21), d〓 _n is initially set to 0, and in (22), the distance d between the corresponding feature points is
Find (Psi, P〓 ^m si) and add it to d〓 _n .

Note that Psi indicates the coordinate value (x _si , y _si ) of the i-th feature point of the s-th stroke, and s=1
˜n, i=1˜P(s), P(s)=2 or 3.

Therefore, in (2), the distance between patterns d〓 _n is d〓 _n = _N 〓 ^s=1 _P(s) 〓 ^i-1 d(Psi, P〓 ^m si) (7) d〓 _n = _N Calculate 〓 ^s=1 _P(S) 〓 ⁱ⁼¹ √( _si −〓 ^m _si ) ² + ( _si −〓 ^m _si ) ² (8).

Note that in equation (8), the distance between feature points is expressed as the Caulkrid norm, but as shown in the equation below, it can be expressed as the square of the Caulkid norm (equation (9)) or the city block distance (equation (10)). Needless to say, it's a good thing to do.

d〓 _n = _N 〓 ^s=1 _P(s) 〓 ⁱ⁼¹ {(x _si −x〓 ^m _si ) ² + (y _si −y〓 ^m _si ) ² }(9) d〓 _n = _N 〓 ^{s =1} _P(s) 〓 ⁱ⁼¹ {｜x _si −x〓 ^m _si |+｜y _si −y〓 ^m _si |} (10) The inter-pattern distance d〓 _n obtained in this way is (31)
In , it is determined whether the distance between patterns is smaller than the minimum value d _nio . If the judgment result is negative, use the method described above for the next standard pattern.
Perform processes (1) and (2).

Further, when the determination result is positive, in (32), d _nio is replaced with dθ _n , and the identification number m of the standard pattern is stored in a predetermined area of the RAM 82, which is referred to as PC. This PC can be anything that identifies standard patterns;
In FIG. 7, the numbers are a series of standard patterns within the same number of strokes, but it goes without saying that the numbers may be, for example, JIS codes indicating the categories of the standard patterns.

The above processes (1), (2), and (3) are repeated in sequence. If the number of standard patterns equal to the number of strokes N of input characters is M, then processes (1), (2), and (3) are repeated M times. As a result, the standard pattern corresponding to the minimum distance d _nio is recognized as an input character, and the corresponding code is output.

〔Effect of the invention〕

As explained above, in the present invention, if each stroke of an input character has a starting point, an ending point, and a bending point, the bending point is extracted as a feature point.
A nearly straight stroke is close to each other at two points, the start point and the end point, and a stroke with a bend point is close to each other at three points, the start point, the bend point, and the end point. Therefore, input characters can be more faithfully approximated with fewer feature points. Thereby, the memory capacity of the standard pattern can be reduced. For example, for the 46 katakana characters shown in Figure 6, the total number of feature points in the present invention is 240, but
In the conventional case where each stroke is uniformly approximated by three points, the number of points is 321. Therefore, in the present invention, the memory capacity of the standard pattern can be reduced by about 25% compared to the conventional method. Further, even in a stroke with springs, the coordinate values of the end point itself vary depending on the springs, but the bending point is not affected by the springs, and the stability of the feature point is good. (In the conventional three-point approximation method, 1/2 of the length of the wing results in a change in the coordinate value of the midpoint.) Also, depending on the number of feature points of each extracted stroke, it is possible to determine whether the stroke is an almost straight line or not. By using this information, it is possible to reduce the number of candidate characters for input characters to less than 1/2 compared to the conventional method, and to reduce the total amount of calculation for inter-pattern distance calculations to 1/2 or less. This speeds up the recognition process.

[Brief explanation of drawings]

FIG. 1 is a diagram explaining a conventional feature point extraction method, FIG. 2 is a diagram showing a functional block diagram of one embodiment of the present invention, FIG. 3 is a diagram showing a specific embodiment of the present invention, and FIG. The figure is a diagram explaining the feature point extraction method of the present invention, FIGS. 5 and 7 are diagrams showing the processing flow of the present invention, and FIG. 6 is a diagram explaining candidate character classification of the present invention. 1...Character information input device, 2...Preprocessing section, 3
...Feature point extraction unit, 4...Inter-pattern distance calculation unit, 5...Minimum distance detection unit, 7...Standard pattern memory unit, 81...Microprocessor, 82...
...RAM, 83...ROM.

Claims (1)

[Claims] 1. Inputting characters while writing with a character writing device,
In an online handwritten character recognition method that recognizes input characters while tracing the strokes of the input characters, the start and end points are connected based on the start point coordinates, end point coordinates, and pen point coordinate series of each stroke of the input character. The maximum value H of the distance between the straight line SE and each writing point coordinate value is extracted, and the ratio (H/L) between the length L of the straight line SE and the maximum value H is larger than a preset threshold. Then, the coordinates of the inflection point extracted as the inflection point are the writing points that give the maximum value H, and the coordinates of the inflection point are taken as the feature points of the stroke. Select a group of standard characters with the same number of feature points as the input character as candidate characters, determine the sum of the distances between the feature points of the input character and the feature points of each standard character serving as the candidate character, and select the minimum distance. An online handwritten character recognition method characterized by recognizing a standard character serving as a value as the input character.