JP2686231B2 - Handwritten character online recognition method and apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for recognizing characters handwritten on a tablet online, and is particularly suitable for improving the recognition rate for characters continuously written without cutting the brush. Online handwriting character recognition method and apparatus.

[0002]

2. Description of the Related Art As a conventional main method for recognizing each character input from a tablet or the like online, there are a structure analysis method and a method using pattern matching. In the structural analysis method, the shape of each part of the input pattern is replaced with a code to represent it, and character recognition is performed by matching with each category pattern of the already coded dictionary.
This structure analysis method can perform high-speed character recognition, and a high recognition rate can be obtained by setting restrictions on the number of strokes and stroke order.

There are various types of methods using pattern matching, such as a method that considers statistical fluctuations of a pattern, and a method that is based on a deformation model. Among the pattern matchings, the one that employs DP matching or backtracking is effective for recognizing continuous writing characters by applying it to the entire input pattern, but the problem is that the amount of calculation is too large.

[0004]

A handwritten character input device such as a tablet is designed to input handwritten characters in a predetermined area in units of one character. However, it is rare that the pen is separated from the tablet for each handwriting and the handwriting is properly input, and in many cases, one character is usually written continuously with one or two strokes.

In the above-described structural analysis method, the stroke number restriction can be avoided to some extent by registering the stroke number variation pattern in the dictionary. However, since enormous dictionary capacity is required to register all the patterns of continuous writing characters in the dictionary, there is a limit to improving the recognition rate of continuous writing characters by registering the stroke number variation pattern.

The pattern matching method of performing DP matching or backtracking is effective for recognition of continuously written characters as described above. However, since the processing that attempts to correspond to the next candidate of correspondence is developed, it requires a huge amount of calculation, and the recognition speed becomes slow, so it is not practical to recognize online when handwriting input. There is.

An object of the present invention is to provide a handwritten character online recognition method and a device therefor capable of recognizing continuously written characters at high speed and at a high rate.

[0008]

In order to achieve the above purpose, tablet
Features from the coordinate data string of handwritten input characters output from
Extract the point sequence into the input pattern feature point sequence and use the dictionary pattern.
Between the feature point sequence of the input pattern and the feature point sequence of the input pattern
Applying a matching process to the feature point sequence of the input pattern and the
The degree of similarity with the feature point sequence of the dictionary pattern is calculated, and
Handwriting character ON to recognize the input pattern from similarity
In the line recognition method, the matching process is advanced.
Multiple features of either pattern separated by a predetermined distance
The points consecutively corresponded to one feature point in the other pattern
In this case, the association of multiple consecutive feature points is correct.
It is achieved by canceling it as not. The above purpose is
In addition, a plurality of feature points separated by a predetermined distance in one pattern
Do not continuously correspond to one feature point of the other pattern
If not, it is judged that the correspondence between the feature points is correct and
This is achieved by proceeding with the subsequent matching process between the feature points .

[0009]

[Function] When finding correspondences between feature point sequences, all combinations
Take the corresponding measures to obtain the respective evaluation values and
If you take the method of adopting a high one, the calculation time
Becomes huge and not practical. Therefore, in the present invention
Is a feature point of one of the patterns
Multiple consecutive features that correspond to multiple consecutive feature points
If the distance between the points is greater than or equal to the specified distance, such pair
Canceled and canceled the response because it was impossible.
Evaluate by taking the correspondence from the feature point to the previous feature point
Is not performed. By doing this, it is efficient
Correspondence between feature point sequences can be made to shorten calculation time
Can be achieved. In addition, the predetermined distance of one pattern
Multiple feature points that are distant from each other
If the feature points do not correspond, the correspondence between the feature points
Is determined to be correct, and the subsequent matching process between the feature points
By advancing the reason and not making a backtrack in the matching process,
It can be further shortened.

[0010]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a flowchart showing a main part of a processing procedure of a handwritten character online recognition method according to an embodiment of the present invention. This flowchart calculates a feature point matching unit 1 that makes a correspondence between a dictionary pattern feature point sequence and a handwritten input pattern feature point sequence, and a similarity between two feature point sequences associated by the matching unit 1. Similarity calculation unit 2
Consists of

The feature point matching unit 1 has a corresponding point selection processing step 3 and a corresponding point reconditioning examination processing step 4
And the corresponding point determination processing step 6 and the corresponding point determination processing step 6, and the similarity calculation unit 2 includes a partial similarity calculation processing step 7 and a weighted average calculation processing step 8.

Using a stylus pen or the like (not shown),
For example, when the character "white" is input by handwriting in a predetermined area 30 of a tablet (not shown) as shown in FIG. 2, the handwriting of the handwriting and the coordinate point sequence data according to the writing order are input from the tablet. Is output. Therefore, using appropriate publicly known means, coordinate point sequence data P ₀ (X ₀ , Y ₀ ), P ₁ (X ₁ , Y ₁ ), P ₂ (X ₂ , Y ₂ ), ... are extracted. FIG. 3 shows an example of feature points extracted from “white” in FIG. 2 by black dots. When this handwritten input pattern feature point sequence is input to the feature point matching unit 1, the matching unit 1
As will be described later in detail, the handwritten input pattern feature point sequence and the feature point sequence Q ₀ (U ₀ , V ₀ ), Q ₁ (U ₁ , V ₁ ), Q ₂ (U ₂ , V ₂ ), ... Correspondence between the respective feature points is output to the similarity calculation unit 2. The matching unit 1 outputs the following data string as the association processing result.

[Ψ ₀ , Ψ ₀ ], [ψ ₁ , Ψ ₁ ], [ψ ₂ , Ψ ₂ ], ... Here, [ψ, Ψ] is the ψ-th point of the feature point sequence of the handwriting input pattern. On the other hand, it means that point Ψ of the feature point sequence of the dictionary pattern corresponds.

According to the present embodiment, the handwriting input pattern feature point sequence of FIG. 3 and the dictionary pattern "white" feature point sequence of FIG. 4 are associated as follows, for example.

[0,0], [1,1], [2,2], [3,3], [4,4], [6,5],
[7,6], [8,7], [9,8], [10,9], [11,10] Similarity calculation is performed on the data string that associates the feature point string of the handwritten input pattern with the dictionary pattern. When input to the unit 2, the similarity calculation unit 2 calculates the similarity between both patterns as described later in detail. The processing by the feature point matching unit 1 and the similarity calculation unit 2 is performed for all the dictionary patterns for each character prepared as a dictionary for one handwritten input pattern, and the highest similarity is obtained for the handwritten input pattern. The character pattern in the dictionary is selected as the first candidate character, and the character having the second or lower similarity is a candidate character according to the rank.

The matching rule executed in the corresponding point selection processing step 3 will be described below. Note that the correspondence relationship output to the similarity calculation unit 2 is a correspondence relationship determined as will be described later, but in this step 3, the correspondence [ψ ′, Ψ ′] at the previous stage is made.

[0017] First of all, is between two of the feature point sequence of the first point be deemed to always be corresponding. That is, [ψ ₀ ', ψ ₀ '] = [0,0]. After associating the first points, the corresponding points are determined according to the following rules. If P _I and Q _J are associated with each other and [ψ _k ′, ψ _k ′] = [I, J], the next corresponding data [ψ _{k + 1} ′,
Ψ _{k + 1} '] selects one from the following three combination candidates [I + 1, J + 1], [I, J + 1], [I + 1, J]. In the situation shown in FIG.

[0018]

(Equation 1)

E (i + 1, J + 1), E (I, J) obtained by substituting these three pairs into E (i, j) represented by
This is performed by selecting the set having the largest value out of +1) and E (I + 1, J).

Here, d _in (i-1): distance between points P _i-1 and P _i d _dic (j-1): distance between points Q _j-1 and Q _j θ _ij : vector [P _i-1 , P _i ] and vector [Q _j-1 ,
Q _i ] is an angle N: a normalized size. The normalized size N is the length of one side of the square when the input character pattern is scaled up or down in a fixed square frame at equal horizontal and vertical magnifications to make the sizes uniform.

As described above, in this embodiment, the value E (i, j) of the product of the evaluation value of the distance difference and the evaluation value of the direction difference is adopted as the selection criterion, and only when both the distance and the direction are short. It is said that there is correspondence of feature points. Since the direction difference is expressed not by the angle θ itself but by its cosine value cos, the evaluation value of the direction difference does not become a large value for minute fluctuations, and if it exceeds a certain range, it becomes a large direction. Since there is a difference, an appropriate corresponding point can be selected.

In step 3, when the processing for taking correspondence is sequentially performed according to the above rule, as shown in FIG. 6, for example,
In some cases, one feature point a ₀ of one pattern A corresponds to a plurality of consecutive feature points b ₀ , b ₁ , b ₂ , ... Of the other pattern B. Also, for example, as shown in FIG.
One pattern a plurality of feature points b ₀ consecutive of the other patterns A to the feature point b ₀ with B ', b _1', ... with the corresponding, the plurality of last feature point a ₀ the pattern B
Also occurs when it corresponds to a plurality of consecutive feature points b ₀ , b ₁ , ... Therefore, the process proceeds to the corresponding point re-retrieving condition examination process in step 4, and it is determined whether the following conditions (1) and (2) are satisfied, and it is determined whether the corresponding points are correctly selected.

[0023] [Condition (1)] there is no corresponding plurality of points with respect to points b _0, the value of L _m = L _B -L _A exceeds the threshold value L _S.

Here, L _A : the distance between the feature point a ₀ and the feature point immediately before it L _B : the feature immediately before the feature point b ₀ from the last feature point of the plurality of feature points corresponding to the feature point a ₀ corresponding plurality of points are present with respect to the sum [condition (2)] point b ₀ of the distance between the feature points adjacent to the point, L _m = L _B - (before L _m) - (b ₀ and its The value of (distance from immediately preceding feature point) exceeds the threshold L _S.

[0025] Here, (prior to _{L m) = L B '-L} A' L A ': distance L _B feature points b ₀ and feature points of the immediately preceding': a plurality corresponding to the feature points b ₀ Sum of distances between adjacent characteristic points from the characteristic point immediately before the first characteristic point to the last characteristic point If neither of the above conditions (1) and (2) is satisfied, corresponding point selection processing is performed. It is determined that the correspondence obtained in step 3 is correct, the process returns to step 3 again, and the next corresponding point selection process is continued.

When the above condition (1) is satisfied, the process proceeds to step 5 for re-corresponding points, and the correspondence between the feature point b ₀ and the feature point a ₀ is left as shown in the lower diagram of FIG. Then, the other correspondences are canceled, and the correspondence between the feature point next to a ₀ and the feature point b ₀ is created, and the process returns to step 3.

Even when the above condition (2) is satisfied, the process proceeds to the corresponding point re-acquisition processing step 5, and after the correspondence between the feature point b ₀ and the feature point a ₀ , as shown in the lower diagram of FIG. The correspondence between the feature point next to a ₀ and the feature point b ₀ is created, and the process returns to step 3.

If the correspondence between the final point of the handwriting input pattern feature point sequence and the final point of the dictionary pattern feature point sequence is selected in step 3, the corresponding point determination process from step 3 to step 6 is performed. Proceed to. In this step 6,
Each correspondence [ψ ₀ ', ψ
_{Of 0} '], [ψ ₁ ', ψ ₁ '], ..., only the correspondence according to the following process is fixed, and the correspondence between other corresponding points is not fixed.
Then, only the determined correspondences are extracted to obtain correspondence data strings [ψ ₀ , Ψ ₀ ], [ψ ₁ , Ψ ₁ ], .... FIG. 8 is a diagram showing processing by the following confirmation processes 1 to 6. In the figure, the relationship connected by the dotted line is the correspondence relationship obtained in step 3, and the relationship connected by the solid line is the confirmed correspondence.

[Confirmation Process 1] Correspondence between the first points and the last points is confirmed. [Confirmation Process 2] The correspondence between the stroke start point of one pattern and the stroke end point of the other pattern is erased. [Confirmation Process 3] Correspondence between stroke end points is confirmed. [Fixing Process 4] If there is any other unfixed correspondence with respect to the fixed correspondence point, it is deleted. [Determination process 5] A one-to-one correspondence is determined. [Confirmation process 6] The correspondence that has not been confirmed is searched for and confirmed later. Then, if there are some other correspondences with respect to the confirmed point, they are deleted. Repeat this until there are no undetermined correspondences.

The corresponding data strings [ψ ₀ , ψ ₀ ], [ψ ₁ , ψ ₁ ], ... Established according to the above-described deciding process are output from the feature point matching unit 1 to the similarity calculation unit 2. In this similarity calculation unit 2, first, the partial similarity S _n is calculated in the partial similarity calculation processing step 7, and then the weighted average of the partial similarity S _n , that is, the similarity is calculated in the weighted average calculation processing step 8. SR is calculated.

The processing in the partial similarity calculation processing step 7 will be described. The partial similarity S _n is the corresponding pair [ψ _n , Ψ _n ]
It is the similarity between the patterns in the section sandwiched between and [ψ _{n + 1} , ψ _{n + 1} ]. First, sd _in (i) _j and sd _dic (i) _j are defined as in the following _equations 2 and 3. Here, MAXS is a constant.

[0032]

(Equation 2)

[0033]

(Equation 3)

Next, as shown in the following Expression 4, the work variables are substituted.

[0035]

(Equation 4)

Then, using the working variables, the following equation 5 is calculated. Each operation symbol indicates the same operation as the operation symbol in C language.

[0037]

(Equation 5)

When R _dic > R _in , the processing of the following _equation 6 is performed,

[0039]

(Equation 6)

When R _{dic ≤R in} , the processing of the following _equation 7 is performed.

[0041]

(Equation 7)

S =
The partial similarity S _n is obtained by the following equation 8 by repeating until ψ _{n + 1} or t = ψ _{n + 1} .

[0043]

(Equation 8)

After the partial similarity S _n is obtained in the partial similarity calculation processing step 7, the weighted average value SR of each partial similarity S _n is then obtained by the following equation 9. This weighted average value SR is the similarity between the handwritten input pattern and the dictionary pattern.

[0045]

(Equation 9)

Here, WS _n is the weighted average weighting with respect to the partial similarity S _n determined by the following expression 10.

[0047]

(Equation 10)

According to the above-described embodiment, the correspondence between the feature point sequence of the handwritten input pattern and the feature point sequence of the dictionary pattern is determined while allowing expansion and contraction only within a certain range, and the partial similarity of this correspondence relationship is established. Since the similarity is calculated from the weighted average of the degrees, the recognition rate is high and high-speed online processing is possible.

[0049]

According to the present invention, it is possible to recognize a handwritten character pattern at a high recognition rate and at a high speed.

[Brief description of the drawings]

FIG. 1 is a flowchart showing a main part of a processing procedure of a handwritten character online recognition method according to an embodiment of the present invention.

FIG. 2 is a diagram showing an example of a handwritten character “white”.

FIG. 3 is a diagram showing a sequence of characteristic points extracted from “white” in FIG.

FIG. 4 is a diagram showing a feature point sequence of a dictionary pattern “white”.

FIG. 5 is an explanatory diagram of a method of obtaining corresponding points and an evaluation method.

FIG. 6 is an explanatory diagram of re-taking corresponding points when a plurality of characteristic points correspond to one characteristic point.

FIG. 7 is an explanatory diagram of retake of corresponding points when a plurality of feature points correspond to one feature point.

FIG. 8 is an explanatory diagram of corresponding point determination processing.

[Explanation of symbols]

1 ... Feature point matching unit, 2 ... Similarity degree calculation unit, 3 ... Corresponding point selection process, 4 ... Corresponding point re-conditioning examination process, 5 ... Corresponding point re-charging process, 6 ... Corresponding point determination process, 7 ... Partial similarity degree calculation Processing, 8 ... Weighted average calculation processing, P _i , Q _j , a _i ,
b _i ... feature points.

Claims (8)

(57) [Claims] 1. A handwritten input sentence output from a tablet
Input pattern by extracting feature point sequence from character coordinate data sequence
The feature point sequence is the feature point sequence of the dictionary pattern and the input pattern.
Corresponding to the sequence of feature points of the
Between the characteristic point sequence of the pattern and the characteristic point sequence of the dictionary pattern
Between the input patterns of the input pattern
In the online recognition method of handwritten characters for recognition,
The process of matching
Multiple feature points that are separated by a certain distance in a row
If a single feature point of
An online recognition method for handwritten characters, which is characterized by canceling the correspondence of the feature points of the character as incorrect . 2. The method according to claim 1, wherein one of the patterns is
Multiple feature points that are a fixed distance apart from each other
If it does not correspond to one feature point, the pair of feature points
Correspondence between feature points after judging that the response is correct
An online recognition method for handwritten characters, which is characterized by advancing the attachment process . 3. A handwritten input sentence output from a tablet
A characteristic point sequence P _i (i = 0, 1, ...) Is extracted from the character coordinate data sequence.
Output the input pattern feature point sequence and use the features of the dictionary pattern
Point sequence Q _j (j = 0, 1, ...) And the input pattern feature point sequence
Between the input pattern feature point sequence and the previous
The similarity between the dictionary pattern feature point sequences is calculated, and the similarity is calculated.
Handwritten character online that recognizes the input pattern from
Input pattern features associated with each other
The following correspondence between the point P _i and the dictionary pattern feature point Q _j is determined by three _{methods of} [P _{i + 1} , Q _{j + 1} ], [P _i , Q _{j + 1} ], and [P _{i + 1} , Q _j ] . When selecting from the set, the distance between both feature points
The product of the evaluation value of the separation and the evaluation value of the direction difference from the previous feature point
Select only the pair with the best value of and check the correspondence of the other two pairs
In addition, the best value set is as follows: E (i, j) = {2N− (| P _i −Q _j |)} · cos θ _ij where N: normalized size θ _ij : vector [ An online recognition method for handwritten characters, characterized in that the value of the angle formed by P _i-1 , P _i ] and the vector [Q _j-1 , Q _j ] is the largest among the three sets. 4. The distance | P _i between two points according to claim 3.
Instead of −Q _j |, P _i = (X _i , Y _i ), Q _j = (U _j ,
V _j ), the handwritten character online recognition method is characterized in that E (i, j) is calculated using | X _i −U _j | + | Y _i −V _j |. 5. When one feature point of one pattern corresponds to a plurality of feature points of the other pattern by repeating either claim 3 or claim 4, the last of the plurality of feature points When the difference between the sum of the distances from the feature points to the feature point immediately before the first feature point and the distance between the one feature point and the feature point immediately before the feature point is equal to or more than a predetermined distance, An online recognition method for handwritten characters, characterized in that the correspondence other than the correspondence between the plurality of feature points and the first feature point is canceled. 6. Correspondence between both patterns obtained by repeating any one of claims 3 to 5 [decision process 1] Correspondence between first points and last points is decided. [Confirmation Process 2] The correspondence between the stroke start point of one pattern and the stroke end point of the other pattern is erased. [Confirmation Process 3] Correspondence between stroke end points is confirmed. [Fixing Process 4] If there is any other unfixed correspondence with respect to the fixed correspondence point, it is deleted. [Determination process 5] A one-to-one correspondence is determined. [Confirmation process 6] The correspondence that has not been confirmed is searched for and confirmed later. Then, if there are some other correspondences with respect to the confirmed point, they are deleted. Repeat this until there are no undetermined correspondences. An online recognition method for handwritten characters, characterized by being determined according to the process of. 7. The pattern according to claim 6, wherein a partial similarity between corresponding feature point sequences of both patterns is calculated based on the determined correspondence between the feature point sequences, and the partial similarity is weighted and averaged. An online recognition method for handwritten characters, which is characterized by calculating the similarity of the characters. 8. A handwritten character online recognition device comprising a tablet for outputting input handwritten characters as a coordinate data string, and a processing device for recognizing the handwritten characters by performing pattern matching with the dictionary pattern by fetching the coordinate data string. The handwriting character online recognition device according to claim 1, wherein the processing device comprises means for executing the method according to any one of claims 1 to 7.