JPH07271917A - Method for compiling handwritten character recognition dictionary and device therefor - Google Patents

Abstract

PURPOSE:To enable a user to additionally register the feature vectors with high character recognizing efficiency in a character recognition dictionary which is used when a feature vector is calculated out of an input character pattern based on the expansion of orthogonal functions of the discrete Fourier transform, the discrete cosine transform, etc., and then the handwritten characters are recognized by the calculated feature vector. CONSTITUTION:The feature vector of an input pattern is calculated (102), and the 'closeness' is decided between the feature vector included in the feature space and A representative point registered in the dictionary (105). If both are close to each other, the registered representative point is corrected so as to increase the 'closeness' to the calculated feature vector (106). If both are not close to each other, the calculated feature vector is additionally registered as a new key point (107).

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 creating a dictionary used in handwritten character recognition.

[0002]

2. Description of the Related Art As one of methods for recognizing handwritten characters, a feature amount of an input handwritten character is calculated, and which input handwritten character is determined according to the calculated feature amount. There is a way. In this method, for each character, the characteristic amount that should be the representative of that character is obtained in advance and stored in the dictionary, and when recognizing the handwritten character, the one that is closest to the input characteristic amount of the character is calculated. The feature amount stored in the dictionary is searched, and the character corresponding to the searched feature amount is used as the recognition result. Since the method based on the feature amount can perform character recognition without referring to the stroke order or the number of strokes, it can be applied to both online and offline handwritten character input devices, and can also be applied to reading printed characters. In addition, in the method based on the feature amount, it is possible to obtain information on the number of strokes in the case of online handwritten character recognition. Therefore, a dictionary is organized for each number of strokes, and the dictionary is applied according to the number of strokes of the input character pattern. By searching only the stroke number dictionary, the recognition time can be shortened and the recognition rate can be further improved.

In the present specification, a pattern or figure which is a target of character recognition such as a handwritten character or a printed character is called an input character pattern. Typically, the input character pattern is represented as a set of lines input by the writer as a handwritten character using a tablet or the like. A character corresponds to a character code on a one-to-one basis. As a result of performing character recognition on the input character pattern, the input character pattern is identified as a specific character.

When handwritten character recognition is performed using the feature amount, the scalar amount is not used as the feature amount, but
It is common to use vector quantities. Here, the feature amount represented by the vector amount is called a feature vector.

In the method using the feature vector, after preprocessing the inputted input character pattern, for example, orthogonal function expansion such as discrete Fourier transform or discrete cosine transform is applied to obtain and calculate coefficients of each degree. A feature vector is obtained by sequentially arranging the respective coefficients. This feature vector is an n-dimensional space (this space is called a feature space)
It represents a point inside. n is an integer of 1 or more.
On the other hand, in the dictionary, a feature vector representing each character is registered as a representative point for each character. Therefore, the representative point is also a point in the feature space. Then, when performing handwritten character recognition, the representative point having the smallest distance from the feature vector to the input character pattern that has been input is searched from the dictionary, and the character corresponding to the found representative point is output as the recognition result. It Therefore, for example, in order for the character "Kou" to be recognized, the representative point for the character "Kou" must be registered in the dictionary in advance.

By the way, when performing handwritten character recognition,
Since writing habits vary considerably among authors, it is important to create a dictionary that maintains a high recognition rate regardless of writer variations. Therefore, it is common to collect a large number of input character pattern samples written by many writers for the same character and create a dictionary based on the collected sample. In this case, as a specific dictionary creation method, (1) average the feature vectors of the collected input character patterns for each character, and use the averaged feature vector as the representative point, that is, for each character. Method of setting one representative point, (2) Calculate the feature vector of each sampled input character pattern, and register all the calculated feature vectors as representative points, that is, set a large number of representative points for each character. Method, (3) Calculate the feature vector for each sampled input character pattern, calculate the average for each character, and also calculate the statistic that represents the variation and the correlation between the feature quantities (feature vector component values) and calculate the average. There is a method of registering not only a representative point but also a statistic, that is, one representative point and a statistical parameter are set for each character.

Further, when a specific user continuously inputs handwritten characters, it is desirable to be able to learn a dictionary. Here, learning means registering a feature vector based on the handwritten character of the specific user as a representative point in the dictionary. By carrying out this learning, the writing habit of the specific user is reflected in the dictionary, and the recognition rate of handwritten character recognition for the specific user can be further improved.

[0008]

The above-mentioned conventional dictionary creating method in handwritten character recognition has the following problems. In the method of setting one representative point for each character shown in (1) above, since the variation due to the writer or the writing style is difficult to be reflected in the representative point, erroneous recognition tends to increase. In the method of setting a representative point for each input character pattern shown in (2) above, since the amount of information required to express one character is large, the size of the dictionary becomes extremely large. It takes a long time to search. In the method of attaching the statistical parameter shown in (3) above, in order to modify this statistical parameter, all the feature vectors used in the calculation of the statistical parameter before modification are required, so in fact Unable to learn a dictionary for a particular user.

An object of the present invention is to provide a method and apparatus for creating a dictionary in which variations due to writers and writing methods can be sufficiently expressed and a user can register his or her own handwritten character without increasing the dictionary size. It is in.

[0010]

A handwritten character recognition dictionary creating method of the present invention calculates a feature vector for an input input character pattern and identifies a character corresponding to the input character pattern according to the feature vector. A method of creating a dictionary which is used in character recognition for recognizing the input character pattern and holds a representative feature vector corresponding to the character for each character as a representative point, in which a new input is made to the dictionary. A first step of inputting the new input character pattern and calculating a feature vector for the new input character pattern when it is requested to register a character pattern; and searching the dictionary for the feature vector Is close to the representative point of the character corresponding to the feature vector, and if it is close, the pair is moved so as to approach the feature vector. To correct the representative point of the character, in other cases, and a second step of adding to the dictionary as a representative point of the character to the corresponding said feature vector.

The handwritten character recognition dictionary creating apparatus of the present invention is
Used in character recognition to calculate a feature vector for an input input character pattern, identify a character corresponding to the input character pattern according to the feature vector, and recognize the input character pattern, and for each character A dictionary creating apparatus that holds a representative feature vector corresponding to the character as a representative point, wherein a new input character pattern is input, and a feature vector for the new input character pattern is calculated as a feature vector calculating unit, Determination means for determining whether or not the feature vector is close to the representative point of the character corresponding to the feature vector by searching a dictionary, and if the result of the determination by the determination means is close, approach the feature vector Correct the representative point of the corresponding character, otherwise use the feature vector as the representative point of the corresponding character. And an update means that you want to add to the dictionary.

[0012]

The representative point already registered in the dictionary is compared with the feature vector of the new input character pattern, and if the two are "close" in the feature space, the feature vector of the new input character pattern is approached. If the representative points that have already been registered are corrected and both are not close, the feature vector of the new input character pattern is additionally registered as a new representative point, so there is no need to increase the number of representative points extremely. It is possible to create a dictionary with a high recognition rate. Therefore, it is possible to create a dictionary that can be additionally registered by the user and can sufficiently express the variation due to the writer and the writing style without increasing the dictionary size.

In the present invention, "close" or "not close"
There are various possible methods for determining whether or not, but from the viewpoint of the amount of calculation and the recognition rate of the created dictionary, for example,
(1) Compare the Euclidean norm of the difference vector between the representative point and the feature vector with a given threshold value, or (2) Calculate the absolute value of the component of each dimension of the difference vector between the representative point and the feature vector. It is desirable to use a method such as comparing the maximum one of them with a given threshold. In addition, when modifying the representative point so as to approach the feature vector, various methods are conceivable. However, a method that considers how many of the already-input character patterns the representative point corresponds to is used. It is desirable to use. As a method of adding the number of corresponding input character patterns, for example, a weight is added to the representative points and the average of the representative points and the feature vector is used as the corrected representative points.

Further, in order to further improve the recognition rate when the dictionary created as described above is used, the dictionary can be improved. To improve the dictionary, for example, after the dictionary is created, character recognition is performed using the created dictionary for the input character pattern used to create the dictionary, and if incorrect recognition is performed, the input character pattern This is performed by additionally registering the feature vector in a dictionary newly created as a representative point.

[0015]

Embodiments of the present invention will now be described with reference to the drawings. FIG. 1 (a) is a block diagram showing the configuration of a handwritten character recognition dictionary creating apparatus according to an embodiment of the present invention.

This handwritten character recognition dictionary creating device creates a dictionary 25 for storing representative points for each character, and includes a handwritten character input / display device 10 for inputting handwritten characters of a writer as an input character pattern, A feature vector calculation unit 12 that calculates a feature vector by performing preprocessing or orthogonal function expansion on an input input character pattern, and a feature vector storage unit 22 that stores the feature vector calculated for each input character pattern. And a comparing unit 23 that compares the calculated feature vector with the representative point stored in the dictionary 25, and an updating unit 24 that updates the dictionary 25 according to the comparison result of the comparing unit 23. The feature vector calculated by the feature vector calculation unit 21 is directly sent to the comparison unit 23, and at the same time, sent to the feature vector storage unit 22.
The feature vector calculation unit 21 and the feature vector storage unit 22 are both configured to receive the feature vector.

The handwritten character input / display device 10 is, for example,
The liquid crystal display panel and the input tablet are integrated as the input area 14, and the input character pattern 1 is pressed by pressing the pen 11 to write characters.
3 is input, and at the same time, the input character pattern 13
Is displayed as it is in the input area. Further, in the lower right corner of the input area 14 in the figure, the learning character 13
Is displayed. Here, the learning character is a character to be learned by the dictionary 25. In the illustrated example, the hiragana "a" is the learning character, and corresponding to this, the writer inputs "a" as the input character pattern.

The comparison unit 23 includes the feature vector calculation unit 21.
It receives a feature vector from the feature vector storage unit 23.
Dictionary 2 for the character corresponding to the feature vector
5 is searched, and the substitution of the character stored in the dictionary 25
The closest value to the feature vector received among the points
Select surface points, select selected representative points and received feature vectors
Compare with Toll to see if they are close to each other or not.
Set. Both the representative point and the feature vector are in n-dimensional space (feature sky
It is expressed as a vector within
Then, (1) the Euclidean norm of the difference vector between the two is
Given threshold k₁Whether or not it is below (2)
Absolute value of each of n components (each dimension component) of the cuttle
And given the value of the largest of these absolute values
Threshold k ₂Can be used if
It Euclidean norm of difference vector is threshold k₁When
The criterion is said to be selected in the n-dimensional feature space.
Radius k around the representative point₁On or within the hypersphere of
It is the same as whether there is a feature vector in the part. Also,
The maximum absolute value of the components of the difference vector is the threshold value k₂With
The condition is that each side is centered on the representative point in the feature space.
Each is parallel to the coordinate axis and the length of one side is 2 · k.₂And
Whether there is a feature vector on the surface or inside the hypercube
It is the same value as Kauka. Threshold k₁, k₂Is fixed to a certain value
You can specify it to distinguish numbers, kanji, English letters, etc.
May be changed according to
Adaptively by evaluating the recognition rate of character recognition by the created dictionary
You may decide.

Besides, (3) whether or not a feature vector exists on the surface or inside of the hypercube in the feature space centered on the selected representative point, or (4) in the feature space centered on the representative point. It is also possible to determine whether or not a feature vector exists on the surface or inside of the hyperellipsoid of. These correspond to the weighted components of each dimension of the feature vector under the conditions (1) and (2).

An example of the structure of the dictionary 25 is shown in FIG. In the dictionary 25, for each representative point to be registered, the character code of the character corresponding to the representative point, the vector value of the representative point, and the number of times the representative point is determined based on the number of feature vectors And are stored. A plurality of representative points can exist for one character. In the example shown in the figure, the JIS kuten code is used as the character code, and the character "A" (kuten code: 0
402), two representative points are registered.

The updating unit 24 updates the dictionary 25 based on the judgment result of the comparing unit 23, and the feature vector input to the comparing unit 23 is "close" to the representative point selected from the dictionary 25. If it is determined, the value of the representative point is corrected so as to approach the input feature vector, and if it is determined that it is "not close", the input feature vector is set as the representative point for the corresponding character in the dictionary 25. Add to. As a method of correcting the representative point when it is determined to be “close”, for example, the dictionary 25 is searched to check how many feature vectors the representative point is determined based on,
In the case of being based on m feature vectors (when the "number of times" column of the representative point is m), the representative point and the input feature vector are divided into 1: m points. The representative point is moved, and the dictionary 25 is rewritten with the moved value.
Then, there is a method of adding 1 to the value in the "number of times" column of the representative point. In addition to this, there is also a method of always approaching a fixed value.

On the other hand, when the comparison unit 23 determines that the feature vector is not “near”, the updating unit 24 registers the input feature vector in the dictionary 25 as a representative point and
1 is set in the "number of times" column of the newly registered representative point. Incidentally. If no representative point of the character corresponding to the input feature vector is registered in the dictionary 25, the number of times is set to 1 and the feature vector is registered as the representative point for the character.

Next, a procedure for creating a dictionary using this handwritten character recognition dictionary creating apparatus will be described with reference to the flowchart of FIG.

When an input character pattern corresponding to a learned character is input via the handwritten character input / display device 10 (step 101), the feature vector calculation unit 21 calculates the feature vector of the input character pattern (step 1).
02). The calculated feature vector is sent to the comparison unit 23 and also stored in the feature vector storage unit 22. It is determined whether or not the character corresponding to the input character pattern is the first character to be registered, that is, a character that has not been registered in the dictionary 25 yet (step 103). Is step 107
Move to. This determination is performed by the comparison unit 23 searching the dictionary 25 and checking whether or not there is a representative point for the learning character corresponding to the input character pattern.

If it is determined in step 103 that the character is not the first one, that is, if it is a character that has already been registered, the comparing unit 23 calculates the representative point of the character corresponding to the calculated feature vector. The representative point closest to the feature vector is searched from the dictionary 25 (step 10).
4). Then, the "closeness" between the searched representative point and the calculated feature vector is obtained and compared with the reference value to determine whether they are "close" or "not close" (step 10).
5). The criteria for this determination are as described above.

When "closeness" ≤reference value, that is, when it is determined that the value is "close", the updating unit 24 sets the value of the representative point searched so as to approach the calculated feature vector as described above. (Step 106), and the process ends. On the other hand, when “closeness”> reference value, that is, when it is determined that “not close” or when it is determined that the character is an unregistered character in step 103, the updating unit 24 calculates. The feature vector is newly registered in the dictionary 25 as a representative point of the character to be registered (step 10).
7), the process ends.

With respect to the input character patterns sequentially input to the handwritten character input / display device, steps 101 to
By successively executing the processing of step 107, learning of the dictionary 25 proceeds, and a dictionary based on a series of input character patterns is completed.

By the way, in the above-mentioned dictionary creation process,
The determination result of “close” or “not close” in the comparison unit 23 depends on the representative point that has been modified before the determination. Therefore, even when the same set of input character patterns is used, the contents of the created dictionary differ depending on the input order. That is, a dictionary that is not optimized for the set of input character patterns is created. Further, in order to correctly recognize characters between similar characters, the shape of the boundary between the regions corresponding to these characters in the feature space must be correctly reflected in the dictionary. However, in the case of the dictionary created as described above, the shape of the boundary in the feature space may not be sufficiently reflected. Therefore, as shown in the flowchart of FIG. 2B, it is desirable to improve the dictionary 25 by using the feature vector stored in the feature vector storage unit 22. The procedure for improving the dictionary 25 will be described below.

Using the created dictionary 25, character recognition is executed for all feature vectors stored in the feature vector storage unit 25, that is, for all input character patterns. As a result of character recognition, an input character pattern is recognized as a character other than the character that the input character pattern originally represents, and if an erroneous recognition occurs, the feature vector that caused the erroneous recognition is originally represented. Registered as a representative point of existing characters (step 11)
1). By repeating this step 111 until the end condition is satisfied (step 112), the dictionary 25 is successively improved. Examples of the termination condition include whether step 111 has been executed a predetermined number of times, or whether the frequency of occurrence of erroneous recognition has fallen below a predetermined value.

The present embodiment will be further described below with reference to specific examples. Note that, for convenience of illustration of the feature space, the feature space is assumed to be two-dimensional. In reality, the feature space is a higher dimensional space.

In FIG. 3, the judgment standard in the comparison unit 23 is
Hypersphere with radius k ₁ centered on a representative point (circle in the example shown)
In the case where a feature vector exists on the surface (circle) or inside of, the case will be described as "close". Figure 3
As shown in (a), a radius k ₁ centered on the representative point 32
When the representative point 32 is located within the circle (the circle indicated by the solid line in the figure),
The comparison unit 23 determines that the distance is “close”, and as a result, as shown in FIG. 3B, the representative point moves so as to approach the feature vector 33 and becomes the representative point 32 ′. The position of the representative point before the movement is indicated by "+", and the circle before the movement is indicated by a chain line. On the other hand, as shown in FIG.
When the feature vector 33 exists outside the circle with the radius k ₁ centered at, the feature vector 33 is added as a new representative point 34, as shown in FIG.
The original representative point 32 is preserved as it is.

In FIG. 4, the judgment standard in the comparison unit 23 is
A case where a feature vector exists on the surface (side) or inside of a hypercube (square in the example shown) having one side of 2 · k ₂ around the representative point will be described as “close”. is there. As shown in FIG. 4A, the representative point 42 is placed in the above-described square (the solid line square in the figure) centered on the representative point 42.
4 is located, the comparison unit 23 determines that it is “close”, and as a result, as shown in FIG. 4B, the representative point moves so as to approach the feature vector 43 and becomes the representative point 42 ′. The position of the representative point before the movement is indicated by "+", and the square before the movement is indicated by a chain line. On the other hand, as shown in FIG. 4C, when the feature vector 43 exists outside the square centered on the representative point 42 and having one side of 2 · k ₂ ,
As shown in (d), the feature vector 43 is added as a new representative point 44. The original representative point 42 is preserved as it is.

5 (a) to 5 (d) are diagrams for explaining the process of improving the dictionary. Here, the numbers "2", the kanji "Otsu", and the alphabet "Z" are taken up as characters that are very similar to each other. Further, the description will be given assuming that the feature space is two-dimensional.

When the dictionary 25 is created, the feature space 51
In the figure, it is assumed that the representative point 52 for "2", the representative point 53 for "Otsu" and the representative point 54 for "Z" are arranged as shown in FIG. 5 (a). In the character recognition using the feature vector, the character recognition is performed according to the representative point closest to the feature vector of the input character pattern. Therefore, the vertical bisector between the representative points in the feature space 51 corresponds to each character. It becomes the boundary of the area. In FIG. 5 (a), Y
This is the border 55 of the character shape. The area of each character is indicated by "2", "Otsu", and "Z", respectively.

By the way, it is assumed that the feature vector 56 (indicated by a mark x in the figure) in the area of the character "2" actually corresponds to "Z". Therefore, if the character recognition of this feature vector 56 is performed using this dictionary, it will be misrecognized. As a result, as shown in FIG. 5B, the feature vector 56 becomes the second representative point 54a for "Z". As a result, at least the boundary between "Z" and "2" is shifted toward the representative point 52 side of the character "2" and becomes the boundary 55a. However, it is assumed that the feature vector 57 on the "Z" side of the boundary 55a actually corresponds to the character "2". Then, similarly, as shown in FIG. 5 (c), this feature vector 57 is 2 for "2".
The second representative point 52a is formed, and the boundary between "2" and "Z" is shifted to the representative point 54 side of "Z" to become the boundary 55b.
Similarly, it is confirmed whether or not erroneous recognition occurs with respect to all input character patterns, and in the case of erroneous recognition, the feature vector is newly added as a representative point, which is indicated by a black circle in FIG. 5 (d). The representative points are arranged as described above, and the shape of the boundary 55 is suitable for correctly performing character recognition between these similar characters.

Although the embodiments of the present invention have been described by taking the creation of a dictionary for recognition of handwritten characters as an example, the present invention can be applied to the creation of a dictionary for recognition of print characters and the like.

[0037]

As described above, according to the present invention, the representative points already registered in the dictionary are compared with the feature vector of the new input character pattern. If the representative points that have already been registered are corrected so that they approach the feature vector of the pattern, and if they are not “near”, the feature vector of the new input character pattern is additionally registered as a new representative point. It is possible to create a dictionary that can be additionally registered, and that can sufficiently express variations due to writers and writing methods without increasing the dictionary size.

After the dictionary is created, character recognition is performed using the dictionary created for the input character pattern used to create the dictionary, and if misrecognition occurs, a feature vector for the input character pattern is obtained. Is additionally registered in the dictionary newly created as the representative point, the dictionary is improved, and the recognition rate when character recognition is performed using this dictionary can be further improved.

[Brief description of drawings]

FIG. 1A is a block diagram showing a configuration of a handwritten character recognition dictionary creating apparatus according to an embodiment of the present invention, and FIG. 1B is a diagram showing an example of a configuration of a dictionary.

2 (a) and 2 (b) are flowcharts for explaining a process of creating a dictionary.

3A to 3D are diagrams illustrating an example of updating a representative point.

FIG. 4A to FIG. 4D are diagrams illustrating an example of updating representative points.

5 (a) to 5 (d) are diagrams for explaining an actual example of dictionary improvement.

[Explanation of symbols]

10 handwritten character input display device 11 pen 12 learning character 13 input character pattern 14 input area 21 feature vector calculation unit 22 feature vector storage unit 23 comparison unit 24 update unit 25 dictionary 31,41,51 feature space 32,32 ', 34, 42, 42 ', 44, 52 to 54, 52a,
54a Representative point 33,56,57 Feature vector 55,55a, 55b Boundary 101-107,111,112 steps

Claims (7)

[Claims] 1. A character recognition method for calculating a feature vector for an input input character pattern, identifying a character corresponding to the input character pattern according to the feature vector, and recognizing the input character pattern. , A method of creating a dictionary that holds, for each character, a representative feature vector corresponding to the character as a representative point, wherein when a new input character pattern is requested to be registered in the dictionary, A different input character pattern to calculate a feature vector for the new input character pattern, and searching the dictionary to determine whether the feature vector is close to the representative point of the character corresponding to the feature vector. If it is close, the representative point of the corresponding character is corrected so as to come close to the feature vector. Handwriting recognition dictionary generating method and a second step of adding to the dictionary vectors as the representative point of the corresponding character. 2. The handwritten character according to claim 1, wherein the determination of the closeness is performed by comparing a Euclidean norm of a difference vector between the representative point and the feature vector with a given threshold value. How to create a recognition dictionary. 3. The determination of the closeness is performed by comparing a maximum value of absolute values of components of each dimension of a difference vector between the representative point and the feature vector with a given threshold value. The method for creating a handwritten character recognition dictionary according to claim 1, which is performed by. 4. The method according to claim 1 is executed to create a dictionary, and then the created dictionary is used for the input character pattern used to create the dictionary. A handwritten character recognition dictionary creating method in which character recognition is performed as a result, and in the case of erroneous recognition, a feature vector for the input character pattern is additionally registered in the dictionary as a new representative point. 5. A character recognition method for calculating a feature vector for an input input character pattern, identifying a character corresponding to the input character pattern according to the feature vector, and recognizing the input character pattern. A device for creating a dictionary that holds, for each character, a representative feature vector corresponding to the character as a representative point, wherein a new input character pattern is input, and the feature vector for the new input character pattern is set as a feature vector. Calculation means, determination means for searching the dictionary to determine whether the feature vector is close to the representative point of the character corresponding to the feature vector, and if the result of the determination by the determination means is close to the feature, The representative point of the corresponding character is corrected so as to approach the vector, and in other cases, the feature vector of the corresponding character is changed. A handwriting character recognition dictionary creation device having update means for adding to the dictionary as a representative point. 6. The handwritten character recognition dictionary according to claim 5, wherein the determination means makes a determination by comparing a Euclidean norm of a difference vector between the representative point and the feature vector with a given threshold value. Creation device. 7. The determination means compares the maximum value of the absolute values of the components of each dimension of the difference vector between the representative point and the feature vector with a given threshold value. The handwritten character recognition dictionary creating device according to claim 6.