JPH08101889A - On-line handwritten character recognition method - Google Patents

Abstract

PURPOSE: To provide the on-line handwritten character recognition method with which a one-stroke character can be recognized as well. CONSTITUTION: When a stroke contains (n) pieces of coordinate points, feature points are extracted by comparing |Δθi | calculated by an expression I with a constant, the feature points and an ON/OFF information expression II for a pen at each feature point are transformed to a feature amount expression III for a polar coordinate system, and the character is recognized by calculating D (X,Y) of an expression V by DP matching, etc., with a template expression IV.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pen input device used in an electronic notebook, a word processor, etc., and more particularly, in such a pen input device, online handwritten character recognition for accurately recognizing not only regular strokes but also continuous characters with large deformation. Regarding the method.

[0002]

2. Description of the Related Art Conventionally, when a character is input to an electronic notebook, a word processor, a computer, etc., a keyboard or the like has mainly been used. However, a certain amount of skill is required for inputting with a keyboard, and as the size of electronic notebooks, computers, etc. becomes smaller, it is no longer possible to secure the area required to secure the keys of the keyboard, etc. In view of this, online handwriting input by handwriting using a tablet or the like has been increasingly used.

Conventionally proposed online handwritten character recognition methods are roughly classified into a structure analysis method and a pattern matching method.

In the structural analysis method, input characters are
Described with a basic stroke prepared in advance consisting of several types of straight lines. The set of basic strokes is identified according to the identification theory. In this structural analysis method, characters are described by basic strokes prepared in advance. Therefore, when recognizing continuous characters with large stroke deformation, it is difficult to extract the basic stroke.

In the pattern matching method, a standard pattern of a character to be written is prepared in advance, considering that the character consists of a set of strokes. Then, the strokes forming the standard pattern and the input pattern are associated with each other. The distance determined between the associated strokes is obtained, and the inter-pattern distance is obtained as the sum of these distances. The standard pattern with the smallest distance between patterns is recognized as the input character. The pattern matching method has a problem that the strokes of the standard pattern and the input pattern are associated with each other incorrectly when recognizing an input pattern having a large deformation. Even if the strokes are correctly matched, the pattern distance calculated with the correct standard pattern is larger than the pattern distance calculated with other (wrong) standard patterns. However, there is a problem in that recognition may not be performed correctly as a result.

[0006]

As described above, in the conventional method, since each stroke forming a character is described only by a basic stroke, or each stroke is characterized by a plurality of coordinates, a continuous character is generated. As described above, when a plurality of strokes are connected and handwritten like one stroke, there is a problem that strokes are not correctly associated with each other, resulting in erroneous recognition.

Therefore, it is an object of the present invention to provide an online handwritten character recognition method capable of recognizing not only regular characters but also continuous characters.

An object of the invention described in claim 2 is to provide an online handwritten character recognition method capable of recognizing not only regular characters but also continuous characters and capable of high-speed and accurate character recognition.

An object of the invention described in claim 3 is to provide an online handwritten character recognition method capable of correctly recognizing not only regular characters but also continuous characters.

An object of the invention described in claim 4 is to provide an online handwritten character recognition method capable of correctly recognizing not only regular characters but also continuous characters.

An object of the invention described in claim 5 is to provide an online handwritten character recognition method capable of recognizing not only regular characters but also continuous characters at high speed and accurately.

[0012]

The invention according to claim 1 is for recognizing a handwritten character on the basis of time-series data, which is provided from the handwritten character data output means and represents the movement of the coordinates of the handwriting of the handwritten character. An online handwritten character recognition method, the step of extracting feature points from given time-series data, and a plurality of extracted feature point data,
Recognizing handwritten characters by comparing with a template prepared in advance. The step of extracting the characteristic points is to select a plurality of adjacent predetermined data from the given time series data and a line connecting the start point and the end point of the selected predetermined plurality of data. Minutes and a step of obtaining an angle formed by a line segment connecting two adjacent points in a plurality of selected predetermined data, and a predetermined angle between the obtained angle and the predetermined threshold angle. A step of determining whether the relationship is established, a step of extracting a predetermined one of the selected data as a feature point when it is determined that the predetermined relationship is established, and a given time series data In contrast, until the selection of a plurality of data is completed, the step of selecting and the step of obtaining the angle,
It includes a step of repeating the determining step and the extracting step.

The invention as defined in claim 2 is as set forth in claim 1.
Online handwriting recognition method, selecting step
Is the given n (n ≧ 3) time series data,
Adjacent k + 2 (k = 1 to n-2) datax
(T₀) ~x(T_{k + 1}) Is selected. The book
Alphabet underlined in the description
Represents a vector. The process of finding the angle is selected
Start point of k + 2 datax(T₀) And the final pointx
(T_{k + 1}),andx(T_i),x(T_i-1) To
hand,

[0014]

[Equation 9]

The angle Δθ _i (i = 1, 2,
,, k + 1) is included. The determining step includes the step of determining whether or not each Δθ _i satisfies | Δθ _i | ≧ θ ^* .

The invention according to claim 3 is an on-line handwritten character recognition method for recognizing a handwritten character by receiving a handwriting information string representing a handwriting of a handwritten character written by using the writing means from the handwritten character data output means. Is. This method extracts a plurality of feature points from the writing information sequence, and the feature amount point sequence for each segment connecting adjacent feature points.

[0017]

[Equation 10]

And the step of determining. x (d _i ) represents the coordinate vector of the i-th feature point, and p (t _i ) represents the on / off information of the writing means at the i-th feature point. This method further includes the feature amount used for recognition for each feature point.

[0019]

[Equation 11]

The step of determining is included. θ (f _i ) is the total of the direction change of the handwriting up to the i-th feature point, and p _i is the on / off information of the writing means between the i−1th feature point and the i-th feature point. Is. This method further includes the calculated feature amount X and the template Y registered in advance.

[0021]

(Equation 12)

Between and

[0023]

(Equation 13)

The inter-pattern distance is obtained by calculating, and the template corresponding to the minimum inter-pattern distance is
The step of selecting as a recognized handwritten character is included.

The invention according to claim 4 is the online handwritten character recognition method according to claim 3, wherein the angle change of the handwriting at the i-th feature point, the i-1th feature point and the i-th feature point.
The amount of movement of the handwriting to and from the i-th feature point and the length of the handwriting up to the i-th feature point, respectively.

[0026]

[Equation 14]

According to

[0028]

(Equation 15)

It is characterized in that the feature amount is obtained from θ (f _i ) and p _i defined in accordance with

The invention described in claim 5 is the online handwritten character recognition method according to claim 4, wherein in the step of obtaining the minimum value of the distance between patterns by the dynamic programming matching method,

[0031]

[Equation 16]

It is characterized in that the amount of calculation of the inter-pattern distance is reduced by pruning the path that becomes

[0033]

According to the first aspect of the present invention, the characteristic points are extracted from the time-series data from the handwritten character data output means, and the plurality of extracted characteristic point data are compared with the template prepared in advance. As a result, handwritten characters are recognized. When extracting characteristic points, a plurality of adjacent predetermined data are selected from the given time series data, and the line segment connecting the start point and the end point of the selected data and the selected data And the angle formed by the line segment connecting two adjacent points is calculated. If a predetermined relationship is established between the calculated angle and a predetermined threshold angle, for example, if the absolute value of the calculated angle is greater than or equal to the threshold angle, then the predetermined data among the selected data is determined. An object, for example, one that is located at the beginning in time series and one that is located in front in time series of the two adjacent points are extracted as feature points. Given time series data,
Until the selection of a plurality of data is completed, the step of selecting, the step of obtaining the angle, the step of determining, and the step of extracting are repeated. By this processing, one continuous writing information string including the characteristics of the handwriting and the comparison, including the off stroke, is obtained.

According to the invention of claim 2, claim 1
In the online handwriting recognition method described,
Start point of k + 2 datax(T₀) And the final pointx
(T_{k + 1})andx(T_i),x(T_i-1) Corresponding to
Angle Δθ_i(I = 1, 2, ..., K + 1)
Δθ_iIs | Δθ_i│ ≧ θ^*Features depending on whether or not
It is determined whether or not to extract points. By this method
Unnecessary data to accurately represent the characteristics of input characters
Is removed, and only feature points suitable for recognition can be extracted.

According to the third aspect of the invention, a plurality of feature points are extracted from the writing information sequence, and the feature amount (θ (f _i ), p _i ) used for recognition is obtained for each feature point. To be Further, the inter-pattern distance is obtained between the obtained feature amount and the template registered in advance, and the template corresponding to the minimum inter-pattern distance is selected as the recognized handwritten character. Since it is possible to perform pattern matching that considers on / off of writing means for one continuous writing information string including off strokes, not only regular writing but also recognition of continuously deformed continuous characters is more accurate. I can do it.

According to the fourth aspect of the invention, the angle change of the handwriting at the i-th feature point, the movement amount of the handwriting between the (i-1) th feature point and the i-th feature point, the i-th feature point The feature point or the length of the handwriting is obtained, and the feature amount is obtained based on these. By performing pattern matching using the feature amount thus obtained, the recognition rate of correct characters is improved.

According to the fifth aspect of the present invention, the amount of calculation of the inter-pattern distance can be reduced by pruning the path satisfying a predetermined condition in the step of obtaining the minimum value of the inter-pattern distance by the dynamic programming matching method. . Therefore, the calculation load of pattern matching can be reduced and the recognition speed can be improved.

[0038]

DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing the configuration of an online handwritten character recognition apparatus for realizing an embodiment of the present invention. Referring to FIG. 1, online handwritten character recognition device 30
Is a tablet 40 for outputting writing information as time series data, a stroke detection unit 42 for detecting a stroke from the output of the tablet 40, and a feature extraction unit 44 for extracting a feature point from the detected stroke.
A matching unit 46 for performing matching between the extracted feature points and a plurality of pre-registered templates; a template storage unit 48 for pre-registering a template used for matching; And a display unit 50 for displaying the matching result.

The tablet 40 is an example of handwritten character data output means. Note that the handwritten character data output means is not limited to a tablet, and the movement of the handwriting of the handwritten character can be output as a coordinate string, and the pen on / off (during stroke or a blank section between strokes) can be output. ), Any information can be output as long as it can output the information.

The processing performed by each unit will be described below with reference to the following mathematical formulas.

[0041]

[Equation 17]

The data obtained from the tablet 40 shown in FIG. 1 is defined as x (t _i ) in the equation (1). Where x
(T _i ) indicates two-dimensional coordinate data at time t _i .
The time t _i is the i-th sampling timing in the case where the time when the first (0th) data of the handwriting is obtained is t ₀ and the sampling is performed at a constant timing thereafter. In Expression (1), p (t _i ) represents the on / off state of the stylus pen at time t _i . For example, when the pen is on, it is a logical 1, and when the pen is off, it is a logical 0.

The outline of the feature point extraction processing performed by the stroke detection unit 42 and the feature extraction unit 44 in FIG. 1 is as follows.

(1) Step 1 Referring to FIG. 2, three coordinate data x (t ₀ ) and x (t
₁ ), x (t ₂ ) are selected. These three points are three consecutive points in the obtained handwriting data. From the coordinate data of the three points shown in FIG. 2, the angle Δθi is obtained according to the equation (2). In the case of FIG. 2, two of Δθ ₁ and Δθ ₂ are obtained.

If θ ^* is set to a predetermined threshold value (0 to 180 °), and if the equation (3) is satisfied, the step 2 described later will be performed.
Otherwise, let x (t ₀ ) and x (t ₁ ) be feature points, and set x (t ₀ ): = x (t ₁ ) and return to step 1.

(2) Step 2 Subsequently, referring to FIG. 3, four coordinate data x (t ₀ ),
,, x (t ₃ ) is given. Also in that case, three Δθ _i (i = 1, 2, 3) are obtained according to the equation (4). If equation (5) is i =
If all of 1, 2, and 3 are satisfied, the subsequent step 3
Otherwise, that is, if any one (Δθ _{3 in} the case of FIG. 4) is larger than θ ^* , as shown in FIG. 4, characterize x (t ₀ ), x (t ₂ ). A point, x (t
₀ ): = x (t ₂ ) and the process returns to step 1.

(3) Step 3 Hereinafter, the same processing as in Step 1 and Step 2 is performed while increasing the number of coordinate data to be processed by one.

(4) When step k k + 2 coordinate data x (t ₀ ), x (t _{k + 1} ) are given, Δθ _i (i =
1, ..., K + 1) are calculated. If expression (7) is satisfied for all i (i = 1, ..., K + 1), go to step k + 1, otherwise, set x (t ₀ ) and x (t _k ) as feature points, and set x (t ₀ ): = x (t _k ) and return to step 1.

(5) Ending Step When the above processing is completed up to the coordinate point number M of the input stroke, the extraction processing ends. Note that M is the formula (1)
Also indicates the number of coordinate points of the input stroke.

FIG. 5 shows an example in which the characteristic points of "ka" are extracted by the above procedure. In FIG. 5, “●” indicates a feature point, and “◯” indicates coordinate data other than the feature point. In FIG. 5, strokes for which the on / off information of the pen is on are indicated by solid lines, and strokes for which the pen is off are indicated by dotted lines.
The term “off stroke” used in this specification refers to a pseudo stroke that connects the end point coordinate of a certain stroke and the start point coordinate of the next stroke. For example, in FIG. 5, line segments connecting x (t ₅ ) and x (t ₆ ) and between x (t ₇ ) and x (t ₈ ) are off-strokes.

The threshold angle θ ^* is 0 as described above.
A value of ˜180 ° can be taken, and an empirically optimum value may be selected. In the case of the experiment described later, θ ^*
Is set to 10 °. The threshold angle θ ^* is common to the equations (3), (5) and (7). of course,
It is conceivable to change this threshold angle in each step, but it is easier to make it common.

FIG. 16 and FIG. 17 show the control of software for realizing the above-mentioned feature point extraction method in the form of flowcharts, the details of which will be described later.

From the feature points extracted as described above,
Regarding the process to obtain the feature amount of pattern matching,
This will be described below with reference to the following equations (8) to (17).

[0054]

(Equation 18)

First, it is assumed that the characteristic point sequence, which is a set of the coordinates of the characteristic points extracted as described above and the on / off information of the pen at the coordinates, is represented by the equation (8). i =
Δθ _i (i = 1, 1, ...
..., M) is calculated. Note that x ¹ (t _i ) is x (t _i )
The x-axis component of x, and x ² (t _i ) is its y component.

Similarly, the value represented by the equation (10) is also calculated. In the present specification, a variable in which a variable is overlined as shown on the left side of Expression (10) will be referred to by adding (overbar) after the variable name. For example, the left side of equation (10) is called Δf _i (overbar).

Δf _i (overbar) is normalized according to equation (11) to obtain Δf _i . According to equation (11), Δf _i
Indicates the distance between x (t _i-1 ) and x (t _i ) when the total length of the handwriting (the total movement distance of the pen) is 1.

When f _i is calculated according to equation (12), f _i
i represents the length of the handwriting up to the characteristic point x (t _i ) (sum of pen movement distances).

The feature amount used for character recognition in this embodiment is represented by the equation (13). Θ in equation (13)
(F _i ) is defined by equation (14), and p _i is equation (1
5). “Δθ _{k on} the right side of equation (14)
(F _k) "is the length of stroke up feature point x (t _k) in the case of the f _k as shown in equation (12), represent the [Delta] [theta] _k at that point x (t _k) There is.
That is, θ (f _i ) on the left side of Expression (14) is k = 1 to k.
It can be considered to be a total angle change _obtained by accumulating Δθ _k calculated at each feature point up to = i.

Further, according to the equation (15), two feature points ( x (t _i-1 ), p (t _i-1 )), ( x (t _i ), p
The state of the pen in the section (t _i )) indicates that “the state of the pen is from time t _{i−1 to} time t _i ”.
The processing represented by the formulas (8) to (14) is considered to be processing for coordinate conversion of the feature point data from the xy coordinate system to the polar coordinate system, but the pen information accompanying it is converted to the polar coordinate system. It is equal to the pen information in the section between two feature points.

The feature quantity used for recognition is defined as described above. Similarly, it is assumed that the feature amount of the template registered in advance is represented by Expression (16). Two feature quantities X _i = (θ (f _i ), p _i ) and Y _j = (η
For (g _j ), q _j ) and ρ are defined by equation (17). Here, d used on the right side of Expression (17)
(P _ik , q _jk ) is a “weight” parameter for reflecting the state (on / off) of the pen in the inter-pattern distance. By introducing such a "weight" parameter, it is possible to perform more accurate pattern matching in consideration of character deformation. Characters can be recognized by summing the distances between the feature amounts calculated according to Expression (17) for the entire handwriting.

By the way, in order to recognize the character, it is necessary to obtain the sum of the equation (17) for the entire handwriting, but there is the following problem here. The number of obtained feature point data is M as shown in the equation (13), and the number of template feature amounts is N as shown in the equation (16). Generally, M ≠ N. In this case, the sum of equation (17) cannot be simply obtained. Hereinafter, the method of obtaining the sum will be described with reference to the following equations (18) to (20).

[0063]

[Formula 19]

[0064]

(Equation 20)

[0065]

[Equation 21]

A method of obtaining the sum of equation (17) in such a case will be described below. Two feature vector sequences X and Y to be associated are defined as follows.

X = X ₁ , X ₂ , ..., X _M Y = Y ₁ , Y ₂ , ..., Y _{N where} X _i and Y _j are feature vectors defined by equations (13) and (16), respectively. Suppose there is.

FIG. 6 is a two-dimensional illustration of associating these two sequences with each other so that the inter-pattern distance is minimized. With reference to FIG. 6, the corresponding conversion function for both X and Y vector sequences is the grid point c = on the ij plane.
It is represented by a sequence of points (i, j). If this point sequence is F, then it becomes as follows.

F = c ₁ , c ₂ , ..., C _k , ..., C _K , where c _k = (i _k , j _k ).

When the distance between the two feature vectors X _ik and Y _jk is represented by d (c _k ), the total sum D (F) of the distances along the point sequence F is represented by the equation (20a). However, the formula (2
In 0a), w _k is the path length in the k-th correspondence.

It is shown that the smaller the value of the equation (20a), the better the correspondence between the feature vector sequences X and Y.

Here, the equation (20a) is changed to the equation (18),
(19), then i₁= J₁= 1, i _K= M, j_K= N
Minimize with respect to F under the constraint of equation (17a)
Think about it. Expression (18) is the sub index i_k, J
_kIs a condition of monotonicity. Equation (19) is the same as
It is the condition of continuity for these sub-indexes.
Equation (17a) should satisfy these sub-indexes
It is a boundary condition.

In this equation (20a), d (c _k ), w
_{If k} is _defined as equations (20b) and (20c), equation (20)
Is obtained. Equation (20) defined in this way is
The distance measure used for handwritten character recognition.

In calculating the value D (X, Y) shown in equation (20) as the distance measure used for recognition, d (p _ik , q)
The value of _jk ) is empirical, but the idea for determining it is as follows.

(1) Consider the case where the registered template is written with the correct number of strokes. When the input character is written in a regular type or continuous character, the part of the input character where each pen is off matches the part where the pen of the template belonging to the same category as the input character is off. Therefore p _i = q _j
When “= up”, Δθ _i , Δη _i , and Δ are the same as in other cases.
f _i and Δg _j are not important. The “category” refers to each character to be recognized. For example, when the input is "ka", the template belonging to the same category as the input character refers to a template for recognizing "ka".

(2) Consider the case where the registered template is written with the correct number of strokes and the input characters are written in continuous characters. In this case, a part of the input character where the pen is turned on may coincide with a part where the pen of the template belonging to the same category as the input character is turned off. Therefore, when p _i = “down” and q _j = “up”, a very large penalty is not applied. That is, in this case d
The value of (p _ik , q _jk ) is made small.

(3) On the other hand, p _i = “up” and q _j = “d”
If "own", the penalty is increased.

Various values can be considered as the value of this weight parameter. In the case of the experimental example described later, the following d (p
_i , q _j ) are used.

[0079]

[Table 1]

When considered in this way, in equation (17)
The defined ρ is the input feature (θ (f _i), P_i) And Ten
Plate characteristics (η (g_j), Q_j) Between
Can be thought of as the distance difference that takes into account the on and off of
You. Actually, ρ is the input feature as shown in equation (17).
Direction difference between stroke and template feature × weight parameter
Demanded as a meter.

A DP (dynamic programming) matching method is one method for calculating the inter-pattern distance D (X, Y) between the input X and the template Y according to the equation (20). An example of applying the DP matching method to the equation (20) is shown in the following equation (21).

[0082]

[Equation 22]

In the third equation of the equation (21), D (i _K , j _K ) obtained according to the second equation is the D expressed by the equation (20).
It is represented as (X, Y).

When an input is given, the inter-pattern distance is calculated between the template registered in advance and the feature quantity obtained from the feature points extracted by the above-mentioned method, using the equation (21). Then, the obtained minimum inter-pattern distance may be selected as the recognition character.

FIG. 7 shows an example of the template and the feature vector of the input data. 8 to 12 show examples of matching between "ka" written by one person and "ka" written by another person. It should be pointed out here that FIGS. 9-12 are in fact one continuous view as shown in FIG. The numerical values shown in FIGS. 9 to 12 are calculated by the DP matching method according to the equation (21).
It should also be noted that there are small surface errors due to rounding errors in the calculation process.

According to the above, matching can be performed between the input character and the template in consideration of the off stroke. Therefore, compared to the conventional method, there is an effect that more consecutive characters can be recognized.

On the other hand, in order to improve the recognition accuracy, it seems that the number of templates should be increased. However, if the inter-pattern distance is calculated using equation (21) for many templates, it is considered that the calculation load becomes large and the recognition speed decreases. Therefore, it is desired to speed up the calculation of the distance between patterns. The following are possible measures to speed up the calculation process.

(1) The operation of writing a character is slower than the operation speed of a computer. Therefore, while obtaining the coordinate values from the tablet, the formula (2
The process can be performed faster if the calculation of 1) is started.

(2) Let M be the number of feature points. You
Then, according to the procedure of Expression (11), Δf _iNormalize
And all Δf_iUntil (i = 1, ..., M) is obtained
Therefore, the calculation of the equation (21) cannot be performed. Sanawa
If the writing operation of chi characters is not completed, the matching operation is effectively
Cannot be performed, and calculation is done in parallel with writing the letters.
You cannot get the benefit of doing.

In order to deal with this problem, in the present embodiment, the DP matching path is pruned as follows, and the calculation load is reduced to reduce the calculation load. This calculation method will be described with reference to the following formula.

[0091]

(Equation 23)

First, the value D ^* (X, X,
Y) is defined. The numerator on the right side of formula (22) is formula (23)
Is defined by Here, i _k and j _k are sub-indexes as described above.

Now, for two unnormalized feature vectors represented by the above equations (24) and (25), M = N, f _i (overbar) = g _j (overbar), p _i = q _It is assumed that _j is established and Expression (26) is established. Equation (26) means that the angular difference between the two unnormalized feature vectors (Equations (24) and (25)) is a constant value γ. Where i =
Let j = k and K = M = N. By substituting these values and equation (23) into equation (22) and rearranging, equation (27) is obtained.
That is, in this case, the left side of the equation (27) becomes equal to the constant γ and is scale invariant.

However, the equation (22) obtained in this way
However, calculation cannot be performed until Δf _i is obtained. Therefore,
In order to reduce the amount of calculation by pruning many nodes (i _k , j _k ) in equation (21), the following equation (22)
To use.

The second in the square root of the denominator in the equation (22)
The item (formula (28)) is obtained from a template registered in advance. Therefore, the equation (28) can be calculated in advance. Also, because of physical limitations such as the size of the tablet, there is a natural upper limit to the size of the input handwriting relative to the character size of the template. Therefore, it is possible to obtain, as an empirical value, a constant α ^* such that equation (29) holds for all templates. This constant α ^* is theoretically a large value. However, it can be made considerably smaller by providing a character input frame. For example, the size of the input character can be limited to twice the character size of the template. In the experimental example described below, α ^* = 2
Is used.

Further, considering the expression (27), there is an empirical value γ ^* for which the expression (30) holds for all templates and inputs. As the interpretation of equation (27),
It is considered that the characters are rotated by the angle γ as a whole regardless of the scale. Therefore, γ ^{* on} the right side of the equation (30) is an upper limit value that limits rotation and exists between 0 ° and 180 °. In the experimental example described below, γ ^* = 25 is used.

Expressions (29) and (3) are added to expressions (22) and (23).
By combining 0), the inequality of equation (31) is obtained. Equation (31) should hold for all templates and inputs. The part on the right side of the equation (31) is irrelevant to the input and can be calculated in advance if the experience values α ^* and γ ^* are given. Therefore, the value represented by the equation (23) is calculated by the DP matching method represented by the equation (21) similarly to the equation (20), and if there is a node for which the relation of the equation (31) does not hold, the node and the subsequent nodes are calculated. You can trim your path. This reduces the amount of calculation in the DP matching method and improves the recognition speed.

FIG. 13 to FIG. 15 show the calculation process when the paths are pruned in this way. 14 and 15
It should be pointed out here that they are combined as shown in FIG. As shown in FIGS. 14 and 15, by making the shortest route to the node not to calculate the path after the node that does not satisfy the equation (31), the case shown in FIGS. It can be seen that a considerable amount of calculation can be reduced by comparison.

Next, a flow chart in the case of implementing the processing for extracting the characteristic points by the software executed on the computer according to the method of the present invention will be described with reference to FIGS. 16 and 17.

Referring to FIG. 16, first in step 59,
Substitute 0 for i ₀ and select x (t ₀ ) as a feature point.

Subsequently, in step 60, the index i ₀
Is determined to be a constant M-2 or less. X in step 61 from all the coordinates points M of strokes this expression is entered unless satisfied is that the processing is completed
(T _M ) is selected as the feature point, and the whole process ends.

When the inequality in step 60 is satisfied, i ₀ +1 is substituted for the variable k in step 62.

Then, at step 64, p (t _i0 ) (time t
A determination is made as to whether the pen on / off information at _i0 ) is "up". If it is up, control is shown in FIG.
7 to step 82. Otherwise, control proceeds to step 65. The process of FIG. 17 will be described later.

At step 65, it is judged if k is M or more. If M or more, control proceeds to step 82 in FIG. 17, otherwise control proceeds to step 66.

At step 66, similarly, p (t _k )
Is up. If it is up, control is shown in FIG.
No. 7, step 82, otherwise control proceeds to step 68.

In step 68, the angle (x(T_{k + 1}),x
(T_i0)) Is calculated. This calculation is similar to equation (9)
Can be calculated.

Subsequently, in step 70, i ₀ +1 is substituted for i and the control proceeds to step 72 in FIG.

In step 72, it is determined whether i≤ (k + 1) is satisfied. If this inequality does not hold, control proceeds to step 73, and if it holds, control proceeds to step 74.

In step 73, k is incremented by 1, and the control returns to step 65 in FIG.

On the other hand, in step 74, the angles ( x (t _i ), x
(T _i-1 )) is calculated. Then, at step 76, Δθ _i is calculated according to the relationship shown in equation (2).

At step 78, it is determined whether or not | Δθ _i | <θ ^* holds. If this inequality does not hold, control proceeds to step 82, otherwise control proceeds to step 80.

If the inequality in step 78 is satisfied, i is incremented by 1 in step 80, and the control returns to step 72. Hereinafter, the processing of steps 72 to 78 is repeated.

On the other hand, when the process proceeds to step 82, x (t _k ) is selected as the feature point.

Subsequently, in step 84, k is substituted for i ₀ , and the control returns to step 60 in FIG.

The feature points are extracted from the input data as described above. Then, the equation (2
A flowchart of the calculation method of 0), that is, the method of obtaining Expression (21) will be described. In addition, prior to this calculation (M +
1) × (N + 1) two-dimensional array D (M, N) is prepared. Referring to FIG. 18, first, step 9
0 is substituted for D (0,0) at 0, and D (0,1) to D
∞ is inserted into (0, N). In fact, D (0,
The maximum number that can be handled by the computer is substituted into 1) to D (0, N). In the processing from step 94 below, the value of each column is calculated for each row, and after the calculation for one row is completed, the processing for the next row is performed. At step 94, a determination is made as to whether i is M or less. If this inequality does not hold, it means that the processing has been completed for all the rows, so the calculation processing ends. If the inequality holds, control proceeds to step 96.

At step 96, ∞ is substituted for D (i, 0). Control continues to step 98. Steps 98-1
The process of 02 is a calculation for the element of each column of the i-th row.

At step 98, it is judged whether j is N or less. If this inequality does not hold, it means that calculation of all column elements for this row has been completed, and the control advances to step 104. Otherwise, control proceeds to step 100.

In step 100, a process of calculating D (i, j) according to the equation (21) is performed.

Then, in step 102, j is incremented and the control returns to step 98. Meanwhile, step 104
Then, i is incremented by 1 and the process returns to step 94.

As described above, the calculation by the DP matching of the equation (21) is performed.

[0121]

As described above, according to the invention described in claim 1, one continuous writing information string including an off-stroke can be obtained. By comparing this writing information string with a template, handwritten characters can be obtained. Is recognized. Since the character recognition is not based only on strokes, it is possible to provide an online handwritten character recognition method that can recognize not only regular characters but also continuous characters.

Further, according to the second aspect of the present invention, unnecessary data for accurately expressing the characteristics of the input character is removed, and only the characteristic points suitable for recognition can be extracted. Since the comparison with the template can be performed using only the feature points suitable for recognition, it is possible to provide an online handwritten character recognition method that requires a small amount of calculation and enables accurate character recognition at high speed.

According to the third aspect of the present invention, one continuous writing information sequence including the off stroke can be compared with the template in consideration of on / off of the writing means. Therefore, not only regular writing but also continuous characters with large deformation can be recognized correctly.

According to the fourth aspect of the invention, since pattern matching is performed by using the feature amount that accurately represents the feature of the handwritten character, the recognition rate of the handwritten character is improved, and not only the regular character but also the continuous character is improved. An online handwritten character recognition method that can be correctly recognized can be provided.

According to the invention described in claim 5, the amount of calculation in the dynamic programming method can be reduced. Therefore, the calculation load of pattern matching can be reduced and the recognition speed can be improved. As a result, it is possible to provide an online handwritten character recognition method capable of recognizing not only regular characters but also continuous characters at high speed and accurately.

[Brief description of drawings]

FIG. 1 is a block diagram of an on-line handwritten character recognition apparatus for performing a method according to an embodiment of the present invention.

FIG. 2 is a schematic diagram showing a case where feature points are extracted from three points.

FIG. 3 is a schematic diagram showing a case where feature points are extracted for four points.

FIG. 4 is a schematic diagram showing a case where feature points are extracted from four points.

FIG. 5 is a schematic diagram showing an example of input data and feature points regarding “ka”.

FIG. 6 is a schematic diagram showing a matching path of two feature vector sequences X and Y.

FIG. 7 is a diagram showing a template and an input pattern in comparison.

FIG. 8 is a diagram schematically showing the relationship between the diagrams shown in FIGS. 9 to 12;

FIG. 9 is a diagram showing a result when the present invention is implemented using a DP matching method.

FIG. 10 is a diagram showing a result when the present invention is implemented using a DP matching method.

FIG. 11 is a diagram showing a result when the present invention is implemented using a DP matching method.

FIG. 12 is a diagram showing a result when the present invention is carried out using a DP matching method.

13 is a schematic diagram showing the relationship between FIG. 14 and FIG.

FIG. 14 is a diagram showing a result when the present invention is performed using a DP matching method and path pruning is performed.

FIG. 15 is a diagram showing a result when the present invention is performed using a DP matching method and path pruning is performed.

FIG. 16 is a flowchart showing a feature point extraction process.

FIG. 17 is a flowchart showing a feature point extraction process.

FIG. 18 is a flowchart of DP matching.

[Explanation of symbols]

 30 Online Handwritten Character Recognition Device 40 Tablet 42 Stroke Detection Section 44 Feature Extraction Section 46 Matching Section 48 Template Storage Section 50 Display Section

 ─────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Takashi Matsumoto 1-6-1, Nishi-Waseda, Shinjuku-ku, Tokyo Within Waseda University (72) Inventor Akira Tsuruta 22-22 Nagaike-cho, Abeno-ku, Osaka-shi, Osaka Within Sharp Corporation

Claims (5)

[Claims] 1. An online handwritten character recognition method for recognizing a handwritten character based on time-series data, which is provided from a handwritten character data output means and represents movement of coordinates of a handwriting of the handwritten character, at a given time. A step of extracting feature points from the series data, and a step of recognizing handwritten characters by comparing a plurality of feature point data extracted by the step of extracting the feature points with a template prepared in advance. Including extracting the characteristic points, selecting a plurality of adjacent predetermined data from the given time-series data, and selecting a start point and a final point of the selected predetermined plurality of data. Determining the angle formed by the line segment connecting the points and the line segment connecting two adjacent points in the selected plurality of predetermined data; A step of determining whether a predetermined relationship is established between the predetermined angle and a predetermined threshold angle, and a plurality of the selected predetermined plurality if it is determined that the predetermined relationship is established. Extracting a predetermined one of the data as a feature point, with respect to the given time-series data, until the selection of the plurality of data is completed, the selecting step,
An online handwritten character recognition method including the steps of obtaining the angle, the determining step, and the extracting step. 2. The selected n steps are given.
Among (n ≧ 3) time series data, adjacent k + 2 data
(K = 1 to n-2 natural number) datax(T₀) ~x
(T_{k + 1}) Is selected, the step of obtaining the angle includes the step of selecting the selected k + 2 data.
Start point of datax(T₀) And the final pointx(T_{k + 1}),and
pointx(T_i),x(T_i-1) With respect toAngle defined by_i(I = 1, 2, ..., k + 1)
Including the step of determining, the step of determining includes a predetermined threshold angle θ^*Against
And each Δθ_iIs | Δθ _i│ ≧ θ^*Determine whether or not
The online handwritten character according to claim 1, including a step of
Recognition method. 3. An online handwritten character recognition method for recognizing a handwritten character by receiving a handwritten character string representing a handwriting of a handwritten character written by using a handwriting character from a handwritten character data output means, the handwritten character recognition method. By extracting a plurality of feature points from the sequence, the feature point sequence for each segment connecting adjacent feature points Each of the x (t _i ) is a coordinate vector of the i-th feature point,
The p (t _i ) represents the on / off information of the writing means at the i-th feature point, and the method further includes, for each feature point, a feature amount X (3) Θ (f _i ) is the total change in the direction of the handwriting up to the i-th feature point, and p _i is the i-th feature point to the i-th feature point. This is on / off information of the writing means, and the method further includes the obtained feature amount X and the template Y registered in advance. Between and, An online handwritten character recognition method including a step of calculating a distance between patterns by calculating and selecting a template corresponding to the minimum distance between patterns as a recognized handwritten character. 4. A change in handwriting angle at the i-th feature point, a movement amount of the handwriting between the (i-1) th feature point and the i-th feature point, a length of the handwriting up to the i-th feature point, Respectively, And further obtain The online handwritten character recognition method according to claim 3, wherein the feature amount is obtained from θ (f _i ) and p _i defined according to the above. 5. The step of obtaining the minimum value of the inter-pattern distance by a dynamic programming matching method, wherein: The online handwritten character recognition method according to claim 4, wherein the amount of calculation of the distance between patterns is reduced by pruning the path that becomes