JPH0991382A - Method and device for on-line handwritten character recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain high recognizing capability while sufficiently suppressing dictionary capacity and a processing time by deciding recognition candidate character categories as to an input pattern in order from a character category regarding a standard pattern which is short in pattern-to-pattern distance. SOLUTION: A stroke correspondence determination device 7 determines one-to-one stroke correspondence after variations in stroke number and stroke order are absorbed between an input pattern to be recognized that is sent out of an input pattern storage device 6 and respective standard pattern in respective character categories sent out of a standard pattern storage device 5. A stroke affine transformation operation determining device 8 determines optimum affine transformation operation. An inter-pattern distance calculating device 10 calculates an inter-pattern distance based upon the sum of the distances by corresponding stroke couples. Then an inter-pattern distance rearranging device 11 sequences and outputs the recognition candidate categories of the input pattern in the increasing order of the distances by the character categories.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention compares an input pattern of characters written in an arbitrary stroke number and stroke order with a standard pattern of the correct stroke number and stroke order prepared for each character category, and determines which input pattern The present invention relates to an on-line handwritten character recognition method and apparatus for automatically determining whether a character category is met.

[0002]

2. Description of the Related Art Conventionally, as an online handwritten character recognition technique for improving resistance to handwriting deformation, there is a technique in which expected handwriting deformation, that is, for example, inclination variation or specific continuous writing is registered in advance. It was

After that, both the input pattern and the standard pattern are converted into one-stroke writing on the assumption that the stroke order is correct, and the overlap between the feature points is maximized while allowing expansion and contraction in the time axis direction by using the dynamic programming method. Techniques for pattern matching have been proposed.

Further, as a standard pattern, not only the average shape of a large number of handwritten character patterns belonging to each character category, but also information about handwriting deformation such as a covariance matrix of position coordinates of feature points forming a character is stored. , A technique that uses a statistical discriminant measure between input patterns has been proposed.

[0005]

However, the above-mentioned conventional handwritten character recognition technology in which expected handwriting deformations are registered in advance has an essential limit that it cannot deal with unexpected deformations at all. It was

Further, the conventional handwriting character recognition technique using dynamic programming has a problem that a large handwriting deformation cannot be absorbed only by expansion and contraction in the time axis direction, and the processing time becomes enormous.

Further, in the conventional handwritten character recognition technique using the statistical discriminant scale, there is a problem that a huge amount of handwritten character patterns must be collected in order to obtain stable statistical information in advance, and the standard pattern As a result, since not only the average value but also information such as a covariance matrix is stored, there is a problem that the dictionary capacity of the standard pattern becomes extremely large.

As a means for solving the above problems, there is a technique described in Japanese Patent Laid-Open No. 63-206881.
The outline of this technique is as follows. First, as the standard pattern of each character category, only the average position coordinates of each characteristic point of a large number of handwritten character patterns written in the correct stroke number and stroke order are stored, and handwritten characters written in any stroke number and stroke order are stored. Also forms an input pattern in which the position coordinates of each feature point are stored, and a one-to-one correspondence between feature points is determined between the input pattern and the standard pattern of each character category. Next, for each feature point of the standard pattern, a local affine transformation that maximizes the overlap with the corresponding feature point group in the input pattern including the feature point group in the standard pattern included in a constant neighborhood is performed. Stable pattern matching is performed after repeatedly applying the standard pattern to transform the standard pattern.

This technique provides a theoretical solution in that any handwritten deformation can be absorbed by iterative operations of local affine transformation around each feature point and that the standard pattern dictionary capacity is small. On the other hand, on the other hand, it requires trial and error to determine the optimal size of a certain neighborhood to which local affine transformation is applied and to control the number of iterations to prevent excessive absorption of handwriting deformation, and the optimal local for each feature point. However, the processing time is enormous due to the iterative affine transformation, which leaves a big problem in practical use.

As described above, in the conventional online handwritten character recognition technique for increasing the resistance to handwriting deformation, large or unexpected handwriting deformation can be absorbed, and the dictionary capacity and processing time are minimized. That
No essential solution has been obtained that satisfies both.

The present invention has been made under the circumstances as described above, and an object of the present invention is to provide a standard pattern in which the correct number of strokes and the average shape in the stroke order are stored even for an input pattern with large handwriting deformation. It is an object of the present invention to provide an on-line handwritten character recognition method and apparatus that perform pattern matching with stable absorption of handwriting deformation between the two, and have high recognition ability while sufficiently suppressing the dictionary capacity and processing time.

[0012]

According to a first aspect of the present invention, for each character category, each stroke forming a character is represented by a fixed number of characteristic points, and the position coordinates of the characteristic points are stored in advance as a standard pattern. The input pattern of input handwritten characters is represented by the characteristic points, the standard patterns are extracted for each character category, and variations in the number of strokes and stroke order are absorbed between the extracted standard patterns and the input patterns. 1: 1 stroke correspondence is determined, and for each stroke of the input pattern, an affine transformation operation is performed on the corresponding stroke of the input pattern such that the overlap with the corresponding stroke of the extracted standard pattern is maximized. For each stroke of the input pattern based on the determined affine transformation operation. A shaped input pattern is generated by performing conversion, and between the extracted standard pattern and the shaped input pattern, the inter-pattern distance is calculated based on the sum of the distances for each corresponding stroke pair, and all standard patterns and all Based on the inter-pattern distance between the generated standard input pattern and the generated input pattern, the character categories related to the standard pattern having the smaller inter-pattern distance are sequentially set as the recognition candidate character categories for the input pattern. An online handwriting recognition method.

According to the present invention, it is possible to significantly improve the recognition accuracy when a large or unexpected handwriting deformation is included. In particular, since the uniform affine transformation is performed not for the entire pattern but for each feature point, but for each stroke, a stable shaping operation is realized without being too hard or too soft. In addition to improving the resistance to such handwriting transformation, the dictionary size can be reduced,
It is also clear that the processing time can be suppressed sufficiently.

According to a second aspect of the present invention, when determining the affine transformation operation, for each stroke of the input pattern, the affine transformation is performed for a plurality of stroke groups including the stroke and an appropriate number of strokes preceding and following the stroke order. The affine transformation operation that maximizes the overlap with the corresponding plural stroke groups in the standard pattern is determined.

According to this aspect, the degree of handwriting deformation included in the character data to be recognized is optimized by optimizing the affine transformation operation by including an appropriate number of strokes that precede and follow each stroke for each stroke. It is possible to control the deformation absorption capacity in accordance with.

According to a third aspect of the present invention, when calculating the inter-pattern distance, the inter-pattern distance is calculated by adding the sum of the distances for each stroke pair corresponding between the input pattern and the standard pattern. To do.

According to this aspect, it is possible to prevent erroneous recognition due to the reduction in the pattern distance difference between similar characters due to excessive shaping.

According to a fourth aspect of the present invention, when the number of strokes is made uniform between the extracted standard pattern and the input pattern, the one having the smaller number of strokes is arranged.

According to a fifth aspect of the present invention, when the respective strokes are made to correspond one-to-one between the extracted standard pattern and the input pattern, the number of characteristic points is made uniform for each stroke pair. .

According to a sixth aspect of the present invention, each stroke forming a character is represented by a certain number of characteristic points for each character category, and the position coordinates of the characteristic points are stored in advance as a standard pattern. A plurality of character patterns are represented by feature points for each character category, and their average position coordinates are stored in advance as a standard pattern.

According to this aspect, it is particularly easy to generate the standard pattern.

According to a seventh aspect of the present invention, when extracting the characteristic points, a fixed number of points that increases with the stroke length including the start point and the end point of the stroke is extracted.

According to an eighth aspect of the present invention, character information input means for inputting an input pattern of handwritten characters, feature point extracting means for expressing each stroke constituting the input pattern by a certain number of feature points, and a character For each category, each stroke forming a character is represented by a fixed number of characteristic points, and the standard pattern storage means in which the position coordinates of the characteristic points are stored in advance as a standard pattern, and the standard pattern is extracted for each character category, Stroke correspondence determining means for determining a one-to-one stroke correspondence that absorbs variations in the number of strokes and stroke order between the extracted standard pattern and the input pattern, and the extracted standard for each stroke of the input pattern. Corresponding strokes of the input pattern that have maximum overlap with corresponding strokes of the pattern Stroke affine transformation operation determining means for determining an affine transformation operation for the stroke, and shaping input pattern generation means for performing affine transformation to generate a shaping input pattern based on the determined affine transformation operation for each stroke of the input pattern. , Between the extracted standard pattern and the shaping input pattern, inter-pattern distance calculation means for calculating the inter-pattern distance based on the sum of the distances for each corresponding stroke pair, all standard patterns and all standard patterns On the basis of the inter-pattern distance between each of the generated shaped input patterns and the generated pattern, the inter-pattern distance rearranging means is set as a recognition candidate character category for the input pattern in order from the character category of the standard pattern having a small inter-pattern distance. Online handwriting characterized by It is a character recognition device.

According to a ninth aspect of the invention, the stroke affine transformation operation determining means determines each stroke of the input pattern when determining the affine transformation operation.
Affine transformation is applied to a plurality of stroke groups including the stroke and an appropriate number of strokes that move back and forth in the stroke order,
The affine transformation operation that maximizes the overlap with the corresponding plural stroke groups in the standard pattern is determined.

According to a tenth aspect of the present invention, the inter-pattern distance calculating means calculates a sum of distances for each stroke pair corresponding between the input pattern and the standard pattern when calculating the inter-pattern distance. In addition, the inter-pattern distance is calculated.

According to the eleventh aspect of the present invention, the stroke correspondence determining means aligns the stroke number between the extracted standard pattern and the input pattern with the one having the smaller stroke number.

According to a twelfth aspect of the present invention, the stroke correspondence determining means makes each stroke one to one between the extracted standard pattern and the input pattern.
The number of characteristic points is made uniform for each pair of strokes.

In the thirteenth aspect of the present invention, the feature point extracting means extracts, when extracting the feature points, a fixed number of points which increases with a stroke length including a start point and an end point of a stroke.

According to a fourteenth aspect of the present invention, the standard pattern is also input through the character information input means,
The feature points are extracted by the feature point extraction means.

According to a fifteenth aspect of the present invention, a plurality of character patterns are input for each character category via the character information inputting means, and a plurality of character patterns corresponding to the plurality of character patterns are provided after the feature point extracting means. It further comprises feature point average position coordinate calculation means for calculating average position coordinates of a plurality of feature points for the stroke.

According to this aspect, it is particularly easy to generate the standard pattern.

[0032]

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

FIG. 1 is a block diagram showing the configuration of an embodiment of the online handwritten character recognition device of the present invention.

In FIG. 1, 1 is a character information input device, 2
Is a preprocessing device, 3 is a feature point extraction device, 4 is a feature point average position coordinate calculation device, 5 is a standard pattern storage device, 6 is an input pattern storage device, 7 is a stroke correspondence determination device, and 8 is a stroke affine transformation operation determination. Device, 9 a shaping input pattern generation device, 10 an inter-pattern distance calculation device, 11
Is an inter-pattern distance rearrangement device.

The operation of each device will be specifically described below.

The character information input device 1 is preferably composed of an existing data tablet, and the time series data of the position coordinate of the writing point is input for each stroke for the handwriting of one character pattern drawn by a dedicated pen on the data tablet. Then, it is sent to the pretreatment device 2. The character information input device 1 is used not only when inputting a large number of handwritten character patterns belonging to each character category when creating a standard pattern, but also when inputting an input pattern.

The pre-processing device 2 receives the writing point position coordinate sequence of the one-character pattern sent from the character information input device 1,
Processing such as noise removal and normalization of position and size is performed based on a known technique, and then sent to the feature point extraction processing device 3. This preprocessing device 2 is also used for both input when creating a standard pattern and input of an input pattern.

The feature point extraction device 3 has a fixed number of points of the writing point position coordinate sequence of one character pattern sent from the preprocessing device 2, which increases from the writing point sequence forming each stroke, including the start point and the end point, with the stroke length. The writing points closest to the positions obtained by equally dividing the stroke by that amount are sequentially extracted as feature points, and they are stored as a feature point sequence of one character pattern. Here, regarding the selection of the total number p of characteristic points according to the stroke length, for example, p = 2, 3, 5, 9, 17, ... May be discretely increased as the stroke length increases. For example, p = 2 is equivalent to extracting only the start point and end point of the stroke, and p = 3 is the start point,
This is equivalent to extracting the end point and the midpoint of the stroke.
FIG. 2 shows an example in which feature points are extracted from writing points for the character category “I”. Regarding the feature point sequence for each stroke of one character obtained here, those relating to the standard pattern are sent to the feature point average position coordinate calculation device 4, and those relating to the input pattern are sent to the input pattern storage device 6.

The feature point average position coordinate calculation device 4 sends a feature point sequence sent from the feature point extraction device 3 to a correct number of strokes and a large number of character patterns written in writing order for each character category to be recognized. Is used to calculate the average position coordinates for each feature point of each stroke that constitutes the character category, and is sent to the standard pattern storage device 5 as the feature point average position coordinates.

The standard pattern storage device 5 determines the average position coordinates of the feature points forming each stroke sent from the feature point average position coordinate calculation device 4 for each character category to be recognized according to the correct stroke number and stroke order. The series is stored as a standard pattern of the character category. Here, when each standard stroke constituting the standard pattern of each character category is denoted by R, R can be expressed as a series of average position vectors of the characteristic points of the standard stroke as shown in Expression (1). , The standard pattern is the individual standard strokes stored in the correct stroke order.

[0041]

## EQU1 ## R = (r ₁ , r ₂ , ..., R _i , ..., R _p ) ... (1) where p: total number of feature points determined from standard stroke length r _i = (r _ix , r _iy ) T: Average position vector of i-th feature point of standard stroke A set of standard patterns of all character categories expressed by average position coordinates of feature points forming each stroke is sent to the stroke correspondence determining device 7. To be done.

The input pattern storage device 6 stores, for an arbitrary number of strokes and a character pattern written in the stroke order, to be recognized,
A series of position coordinates of feature points forming each stroke sent from the feature point extraction device 3 in accordance with the number of strokes and the stroke order is stored as an input pattern of the character category. Here, if each input stroke forming the input pattern is denoted by S, S can be expressed as a series of position vectors of the feature points of the input stroke as shown in Expression (2), and the input pattern is These input strokes are stored in the correct stroke order.

[0043]

[Number 2] _{S = (s 1, s 2} , ···, s j, ···, s p ') ··· (2) However, p' the total number of feature points is determined from the input stroke length s _j = ( s _jx , s _iy ) T: Position vector of the j-th feature point of the input stroke The input pattern represented by the position coordinates of the feature points forming each stroke is sent to the stroke correspondence determining device 7.

The stroke correspondence determining device 7 determines the number of strokes and the stroke order between the input pattern to be recognized sent from the input pattern storage device 6 and the standard pattern of each character category sent from the standard pattern storage device 5. A one-to-one stroke correspondence that absorbs the fluctuation is determined. Here, if the number of strokes of the input pattern and the standard pattern is described as M, and the one with a small number of strokes is described as N (M ≧ N), this apparatus aligns both patterns with the pattern of the smaller stroke number N and sets N pairs. Stroke correspondence is determined. Such an optimal one-to-one
Known techniques can be used to determine the stroke correspondence of, for example, Proceedings of the Third International
Conference on Document Analysis and Recognition
(Montreal, Canada; August 14-16, 1995; IEEE COMPUT
Wakahara et al.'S On-Line Cursive Kan published in ER SOCIETY)
ji Character Recognition as StrokeCorrespondence P
In the technique shown in a paper entitled roblem, stroke correspondence is determined by a two-step process. In the first stage processing, all strokes of the N-image pattern and N strokes in the M-image pattern are processed.
The optimum N pairs of stroke correspondences that minimize the sum of stroke distances with the strokes of the book are determined by using the discrete combination optimization method. By this processing, the fluctuation of the stroke order is absorbed. Next, in the process of the second stage, the non-corresponding (M−N) strokes in the M image pattern are overlapped with the N image pattern on any of the two already-corresponding strokes preceding and following in the stroke order. Selectively combine to improve. By this processing, the variation in the number of strokes is absorbed, and both patterns are arranged in the number N of strokes.

By the stroke correspondence determining operation, both patterns become patterns with the number of strokes N, and N pairs of stroke correspondences are obtained. For the input pattern and the standard pattern arranged in N strokes, the characteristic point sequence representation of N strokes forming the input pattern is written in the order of strokes.
_{S 1, S 2, ···,} S i, ···, denoted as S _N, the feature point string representation of the N strokes constituting the standard pattern _{R 1, R 2, ···,} R i ,..., denoted as R _N, R _i (1
≦ i ≦ N) is associated with S _i (1 ≦ i ≦ N). However, as shown in FIGS. 3A and 3B, it is assumed that the number of characteristic points is made uniform for each stroke pair. In this way, the characteristic point sequence representation of the input pattern and the standard pattern in which the number of strokes, the stroke order, and the number of characteristic points are uniform is sent to the stroke affine transformation operation determining device 8.

Stroke affine transformation operation determining device 8
Is a feature point sequence representation of an input pattern and a standard pattern in which the number of strokes, the stroke order, and the number of feature points, which are sent from the stroke correspondence determining device 7, are used, and a plurality of stroke groups including the stroke for each stroke forming the input pattern. By performing the affine transformation operation on, the optimum affine transformation operation that maximizes the overlap with the corresponding multiple stroke groups in the standard pattern is determined. I-th stroke S _i of input pattern (1 ≦ i ≦ N)
The optimum affine transformation operation corresponding to is determined by the following procedure.

First, by taking advantage of the large correlation of handwriting deformation between successive strokes, a plurality of stroke groups including the stroke n ₁ preceding the stroke in the stroke order and the stroke n ₂ succeeding in the stroke order are extracted. However,
n ₁ ≧ 0, n ₂ ≧ 0, and for example, n ₁ = 1 and n ₂ = 1
Alternatively, n ₁ = 1 and n ₂ = 0 are used. That is, S _i
For (1 ≦ i ≦ N), u = max (1, i−n ₁ ).
, And v = min (N, i + n ₂ ), a plurality of stroke groups S _u , S _{u + 1} , ..., S _i in the stroke order are written.
..., S _v-1 , S _v are extracted. This will be described with an example. For example, assume that the character category is "cry". In this case, the stroke is S _{1 as} shown in FIG.
To S ₈ . Now, it is assumed that the stroke S ₂ is targeted as the i-th stroke S _i . In this case, extracted, for example _{n 1 = 1, n 2 =} 1, a plurality stroke group S _1, including stroke S ₁ and subsequent stroke S ₃ that is preceding, S _2, S _3. Also, the i-th
If the stroke S ₈ are subject as a stroke _{S i, n 1 = 1,} n 2 = 1 for subsequent stroke exists even prior multiple strokes S ₇ including only the stroke S ₇ that, Extract S ₈ .

Let q be the number of all the feature points included in these, and the position vectors of these q feature points are connected in the order of writing in order as shown in equation (3).

[0049]

[Equation 3] C ₁ = (s ′ ₁ , s ′ ₂ , ..., S ′ _k , ..., s ′ _q ) (3) For example, as described above, the stroke is the i-th stroke S _i. When S ₂ is the target, and the characteristic points are configured as shown in FIG. 5, C ₁ becomes as shown in Expression (4).

C ₁ = (s ₁ , s ₂ , ..., S ₁₉ ) (4) On the other hand, in the standard pattern, a plurality of stroke groups S _u , ...
S _i , ..., S _v correspond to R _u , ..., R _i ,.
, R _v , and the total number of feature points contained in them is the same q. A series of position vectors of these q feature points in the stroke order is written as shown in Expression (5).

[0051]

Equation 4] _{C 2 = (r '1,} r' 2, ··, r 'k, ··, r' q) ··· (5) Next, the feature point according to a i stroke S _i of the input pattern An affine transformation operation defined by equation (6) is performed on each feature point s ′ _k included in the column C ₁ to maximize the overlap with the feature point sequence C ₂ by the transformation on the feature point sequence C ₁ , that is, the equation Affine transformation operations A _i and b _i that minimize the objective function Ψ _i shown in (7) are determined.

[0052]

S ″ _k = A _i s ′ _k + b _i (1 ≦ k ≦ q) (6) where A _{i is} a matrix of 2 rows and 2 columns and represents expansion, contraction, rotation, and distortion b _i : Representing translational movement with a two-dimensional vector

## EQU6 ## Ψ _i = Σ _k ‖s ″ _k −r ′ _k ‖ ² = ‖A _i s ′ _k + b _i −r ′ _k ‖ ² (7) (1 ≦ i ≦ N) where Σ _k : sum of k = 1, ..., q ‖ ‖: Euclidean norm of vector The minimization problem of the objective function Ψ _i is reduced to a solution of simultaneous linear equations for each component of A _i and b _i by a known technique. Then
This can find a solution at high speed.

In this way, the optimum affine transformation operation for each stroke S _i of the input pattern is determined, and a set of those stroke affine transformation operations {A _i , b _i }.
(1 ≦ i ≦ N) and the characteristic point sequence representation of the input pattern and the standard pattern are sent to the shaping input pattern generation device 9.

The shaping input pattern generation device 9 uses the set of stroke affine transformation operations for each stroke of the input pattern sent from the stroke affine transformation operation determination device 8 and the feature point sequence representation of the input pattern and standard pattern, A stroke affine transformation operation is performed on each stroke of the input pattern to generate a shaped input pattern. Here, the feature point sequences of the i-th stroke S _i of the input pattern and the corresponding i-th stroke R _i of the standard pattern are described by equations (8) and (9). However, the number of feature points is p
, And both strokes are aligned as described above.

[0055]

[Equation 7] _{S i = (s 1, s} 2, ···, s k, ···, s p) ··· (8) R i = (r 1, r 2, ···, r k , ..., r _p ) (9) The shaping input pattern is generated for each stroke unit of the input pattern. I-th stroke S _i (1
The shaping operation for ≦ i ≦ N) is performed by the following equation (10) for each feature point of the stroke using the stroke affine transformation operations A _i and b _i .

[0056]

S ″ _k = A _i s ′ _k + b _i (1 ≦ k ≦ p) (10) A series of s ″ _k (1 ≦ k ≦ p) thus obtained is , A shaping input stroke S ″ _i . The shaping input pattern is a shaping input stroke sequence {S ″ _i } (1 ≦
It is expressed as i ≦ N). The feature point sequence representation of the shaping input pattern and the standard pattern is calculated by the inter-pattern distance calculation device 10.
Sent to

FIG. 6 is a diagram showing an example of processing for the character category "cry". Here, FIG. 6A is a diagram showing an input pattern, FIG. 6B is a diagram showing a standard pattern, and FIG. 6C is a diagram showing a shaped input pattern. FIG. 6 (d) is a diagram in which these patterns are superimposed, “•” is the point of the input pattern, “◎” is the point of the standard pattern, “○” is the point of the shaped input pattern, and “●” is the standard. The points where the pattern and the shaping input pattern overlap are shown. It should be noted that interpolated point sequences are shown between the characteristic points in the stroke.

Further, FIG. 7 is a diagram showing an example of processing for the character category "permanent".

The inter-pattern distance calculation device 10 uses the feature point sequence representation of the shaping input pattern and the standard pattern sent from the shaping input pattern generation device 9 to make a corresponding stroke between the shaping input pattern and the standard pattern. The inter-pattern distance is calculated based on the sum of the distances for each pair. As described above, the shaping input stroke S ″ _i (1 ≦ i ≦ N)
Corresponds to the i-th stroke R _i of the standard pattern. First, the inter-stroke distance d (S ″ _i , R _i ) between the corresponding stroke pair S ″ _i and R _i is calculated by the equation (11).

[0060]

## EQU9 ## d (S ″ _i , R _i ) = (1 / p) Σ _k ‖s ″ _k −r _k ‖ ² ... (11) where Σ _k : k = 1, ..., p Then, based on the sum of the distances between the strokes, the pattern distance D between the shaping input pattern and the standard pattern is calculated by the formula (1
Calculated in 2).

[0061]

[Equation 10] D = (M / N) Σ _i d (S ″ _i , R _i ) ... (12) where Σ _i : sum of i = 1, ..., N M / N: number of strokes The normalization multiplier (≧ 1) due to the difference or the degree of overlap before shaping may be calculated and calculated using the equation (13) in order to be evaluated and added in the same manner.

[0062]

[Equation 11] D = (M / N) Σ _i [d (S ″ _i , R _i ) + d (S _i , R _i )] (13) When using the equation (13), excessive shaping is performed. By this, it is possible to prevent erroneous recognition due to a decrease in the distance between patterns between similar characters.
Is sent to the inter-pattern distance rearrangement device 11 and temporarily stored.

The operation in each of the above-mentioned devices 7 to 10 is
The input pattern and the standard patterns of all the character categories to be recognized are sequentially repeated, and the calculated inter-pattern distances are sequentially transmitted to the inter-pattern distance rearrangement device 11 and accumulated.

The inter-pattern distance rearrangement device 11 rearranges the inter-pattern distances D calculated for all the character categories to be recognized in ascending order, and for each applicable character category in the ascending order of the distance D. The input patterns are ordered and output as recognition candidate categories.

[0065]

As described above, in the present invention, first,
For each of the character categories to be recognized, the average position coordinates of the characteristic points of each stroke are calculated in advance using the correct number of strokes and the multiple character pattern written in the stroke order, and this is stored as a standard character pattern. , A character pattern of unknown category written in an arbitrary number of strokes and stroke order is represented by an input pattern consisting of feature points of each stroke. Then, the optimum one-to-one stroke correspondence in which the number of strokes is aligned between the input pattern and the standard pattern of each character category is determined. Furthermore,
An optimal affine transformation operation is performed on each stroke of the input pattern so that the overlap between the corresponding strokes increases, and a shaped input pattern is generated, and pattern matching is performed with the standard pattern. Therefore, by the above processing, it is possible to significantly improve the recognition accuracy when a large or unexpected handwriting deformation is included. In particular, since the uniform affine transformation is performed not for the entire pattern but for each feature point, but for each stroke, a stable shaping operation is realized without being too hard or too soft. Further, it is apparent that not only the resistance to handwriting deformation is improved, but also the dictionary size can be reduced and the processing time can be sufficiently suppressed. Furthermore, by optimizing the affine transformation operation for each stroke, including the appropriate number of strokes before and after in the stroke order, the ability to absorb deformation according to the degree of handwriting deformation included in the character data to be recognized. It is possible to control

[Brief description of drawings]

FIG. 1 is a configuration block diagram of an embodiment of an online handwritten character recognition device of the present invention.

FIG. 2 is a diagram showing an example in the case of extracting feature points from writing points.

FIG. 3 is a diagram illustrating an example of a process of aligning the number of characteristic points for each stroke pair.

FIG. 4 is a diagram for explaining a process in a stroke affine transformation operation determining device.

FIG. 5 is a diagram for explaining processing in a stroke affine transformation operation determining device.

FIG. 6 is a diagram illustrating an example of processing for a character category “cry”.

FIG. 7 is a diagram illustrating an example of processing for a character category “naga”.

[Explanation of symbols]

 1 Character Information Input Device 2 Pre-Processing Device 3 Feature Point Extraction Device 4 Feature Point Average Position Coordinates Calculation Device 5 Standard Pattern Storage Device 6 Input Pattern Storage Device 7 Stroke Correspondence Determination Device 8 Stroke Affine Transformation Operation Determination Device 9 Shaped Input Pattern Generation Device 10 inter-pattern distance calculation device 11 inter-pattern distance rearrangement device

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Kazumi Odaka 3-19-2 Nishishinjuku, Shinjuku-ku, Tokyo Inside Nippon Telegraph and Telephone Corporation

Claims (15)

[Claims] 1. Strokes forming a character are represented by a fixed number of characteristic points for each character category, and the position coordinates of the characteristic points are stored in advance as a standard pattern, and an input pattern of an input handwritten character is stored. Expressed by the characteristic points, the standard pattern is extracted for each character category, and a one-to-one stroke correspondence that absorbs variations in the number of strokes and stroke order is determined between the extracted standard pattern and the input pattern, and the input For each stroke of the pattern, determine the affine transformation operation for the corresponding stroke of the input pattern such that the overlap with the corresponding stroke of the extracted standard pattern is maximum, and for each stroke of the input pattern, Generates a shaped input pattern by performing affine transformation based on the determined affine transformation operation , Between the extracted standard pattern and the shaping input pattern, the inter-pattern distance is calculated based on the sum of the distances of the corresponding stroke pairs, and the shaping is generated for all the standard patterns and all the standard patterns. An online handwritten character recognition method, characterized in that, based on a distance between patterns with an input pattern, character categories related to a standard pattern having a smaller distance between patterns are sequentially set as recognition candidate character categories for the input pattern. 2. When deciding the affine transformation operation, for each stroke of the input pattern, affine transformation is performed on a plurality of stroke groups including the stroke and an appropriate number of strokes preceding and succeeding in the stroke order, and the affine transformation is performed. The on-line handwritten character recognition method according to claim 1, wherein an affine transformation operation that maximizes the overlap with a corresponding plurality of stroke groups is determined. 3. When the distance between patterns is calculated,
The online handwritten character recognition method according to claim 1, wherein the inter-pattern distance is calculated by adding a sum of distances for each stroke pair corresponding between the input pattern and the standard pattern. 4. The on-line handwritten character recognition method according to claim 1, wherein when the number of strokes is made uniform between the extracted standard pattern and the input pattern, the number of strokes is made smaller. 5. The feature points according to claim 1, wherein when the strokes are made to correspond one-to-one between the extracted standard pattern and the input pattern, the feature points are made uniform for each stroke pair. Online handwriting recognition method. 6. When each stroke forming a character is represented by a fixed number of characteristic points for each character category and the position coordinates of the characteristic points are stored in advance as a standard pattern,
The online handwritten character recognition method according to claim 1, wherein a plurality of character patterns are represented by feature points for each character category, and the average position coordinates of them are stored in advance as a standard pattern. 7. The on-line handwritten character recognition method according to claim 1, wherein when extracting the characteristic points, a fixed number of points that increases with a stroke length including a start point and an end point of the stroke is extracted. 8. A character information input means for inputting an input pattern of handwritten characters, a feature point extraction means for expressing each stroke constituting the input pattern by a fixed number of feature points, and a character for each character category. Each stroke is represented by a fixed number of characteristic points, the position coordinates of the characteristic points are stored in advance as a standard pattern standard pattern storage means, the standard pattern is extracted for each character category, the standard pattern extracted and Stroke correspondence determining means for determining one-to-one stroke correspondence that absorbs variations in stroke count and stroke order between the input pattern and strokes corresponding to the extracted standard pattern for each stroke of the input pattern. Affine transformation for the corresponding strokes of the input pattern so that the overlap is maximized Stroke affine transformation operation determining means for determining a work, shaping input pattern generation means for performing affine transformation on the basis of the determined affine transformation operation for each stroke of the input pattern to generate a shaping input pattern, and the extraction Between the standard pattern and the shaped input pattern, the inter-pattern distance calculating means for calculating the inter-pattern distance based on the sum of the distances of the corresponding stroke pairs, and for all the standard patterns and all the standard patterns, respectively. Based on the inter-pattern distance with the generated shaped input pattern, the inter-pattern distance rearranging means that makes the recognition candidate character category for the input pattern in order from the character category related to the standard pattern having the smaller inter-pattern distance, An online handwriting recognition device. 9. The stroke affine transformation operation determining means, when determining the affine transformation operation, for each stroke of the input pattern, regarding a plurality of stroke groups including the stroke and an appropriate number of strokes preceding and following the stroke order. 9. The on-line handwritten character recognition device according to claim 8, wherein an affine transformation operation is performed to determine an affine transformation operation that maximizes the overlap with a corresponding plurality of stroke groups in the standard pattern. 10. The inter-pattern distance calculation means calculates the inter-pattern distance by adding the sum of the distances for each stroke pair corresponding between the input pattern and the standard pattern when calculating the inter-pattern distance. The on-line handwritten character recognition device according to claim 8, which is calculated. 11. The online handwriting according to claim 8, wherein the stroke correspondence determining means aligns the stroke number between the extracted standard pattern and the input pattern in a direction having a smaller stroke number. Character recognizer. 12. The stroke correspondence determining means, between the extracted standard pattern and the input pattern,
9. The on-line handwritten character recognition device according to claim 8, wherein the number of characteristic points is made uniform for each stroke pair when each stroke is made to correspond one-to-one. 13. The online system according to claim 8, wherein the feature point extracting means extracts a fixed number of points that increases with a stroke length including a start point and an end point of a stroke when the feature points are extracted. Handwritten character recognition device. 14. The on-line handwritten character recognition device according to claim 8, wherein the standard pattern is also input through the character information input means and the feature points are extracted by the feature point extraction means. 15. A plurality of character patterns are input for each character category via the character information inputting means, and a plurality of features for a plurality of strokes corresponding to the plurality of character patterns are provided after the feature point extracting means. The online handwritten character recognition device according to claim 14, further comprising feature point average position coordinate calculation means for calculating average position coordinates of points.