JP3338575B2 - Online handwritten character recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an online handwritten character recognition apparatus for recognizing handwritten characters using spatially asymmetric direction codes.

[0002]

2. Description of the Related Art An online handwritten character recognition device recognizes a character in real time by directly using, for example, information on a moving direction of an electronic pen on a tablet. There are known a method of generating a code sequence of stroke directions (8 directions or 16 directions) and examining its characteristics, a method of characterizing the coordinate position of each stroke, that is, an end point, a refraction point, and the like.

The former recognition method is disclosed in, for example,
There is an online handwritten character input device described in JP-A-8-101377. In this device, a handwritten character input from a tablet is converted into a character stroke element decomposed into eight-direction components, and a character stroke code corresponding to a combination of the character stroke elements is created. The handwritten character is recognized by comparing the code with the character stroke code in the dictionary.

[0004]

However, since the direction of the stroke of the handwritten character is mainly from left to right and from top to bottom, the direction code of the stroke is different from the prior art. If the direction is set, there is a difference in the appearance frequency of the direction code. In addition, in the direction of low appearance frequency, there is a tendency for the writer to fluctuate greatly, and it is difficult to cope with such deformation of input characters. There is a problem that the algorithm becomes complicated.

An object of the present invention is to set a direction code to be spatially asymmetric in consideration of an appearance frequency in a stroke direction,
It is an object of the present invention to provide an online handwritten character recognition device that can reduce the types of codes to prevent an increase in the size of a memory or dictionary and that is resistant to character deformation.

[0006]

In order to achieve the above object, according to the present invention, a handwritten character input unit for directly inputting a handwritten character as a time-series coordinate point group is provided. A preprocessing unit that normalizes the size, converts each stroke constituting the character to a vector sequence connected in the writing order by approximating each of the strokes constituting the character, and replacing the direction of each vector of the stroke with a spatially asymmetric direction code, A feature extraction unit that extracts a sequence of direction codes as a feature amount of a stroke;
Character recognition for recognizing the input character by comparing a dictionary in which the stroke characteristic amount of each character is registered in advance with the stroke characteristic amount of the input character and the stroke characteristic amount of the character in the dictionary. And a part.

According to the second aspect of the present invention, a handwritten character input unit for directly inputting a handwritten character as a time-series coordinate point group, and normalizing the size of the input character so that each stroke constituting the character is broken. A preprocessing unit that approximates and converts the vector sequence into a vector sequence connected in writing order, and replaces the direction of each vector of the stroke with a spatially asymmetrical direction code and extracts the direction code sequence as a first feature value of the stroke. A second feature extraction unit that extracts a sequence of values obtained by quantizing the length of each stroke vector based on a predetermined reference value as a second feature amount of the stroke; A sentence for recognizing the input character by comparing a dictionary in which the stroke characteristic amount of each character is registered in advance with the stroke characteristic amount of the input character and the stroke characteristic amount of the character in the dictionary. It is characterized in that a recognition unit.

According to a third aspect of the present invention, the predetermined reference value is calculated based on a sum of lengths of vectors constituting a character.

According to a fourth aspect of the present invention, for a character composed of a plurality of strokes, there is provided a first stroke sequence calculating section for obtaining a first vector sequence connecting the starting points of the strokes in the writing order. The second feature extraction unit performs the first feature extraction.
The first and second feature amounts are extracted from the vector sequence.

According to the fifth aspect of the present invention, for a character composed of a plurality of strokes, a first stroke sequence calculating unit for obtaining a first vector sequence connecting the starting points of each stroke in writing order, and a circumscribed rectangle of each stroke. A second stroke sequence calculating unit for obtaining a second vector sequence connecting the centers in the writing order, wherein the first and second feature extracting units include:
It is characterized in that the first and second feature amounts are extracted from two vector sequences.

According to a sixth aspect of the present invention, there is provided a direction code optimizing section for comparing and analyzing the writing habits of the writer with the characters in the dictionary in advance and creating an optimum directional code for each writer.
The feature extraction unit is characterized in that a feature amount of a stroke is extracted based on the created direction code.

[0012]

The direction code according to the present invention is spatially asymmetric. That is, one direction code is a range of 90 degrees with a boundary vertically upward and horizontally leftward from the center,
A code in six directions created by dividing the other area into five equal parts is used. As a result, characters with large distortion can be recognized without increasing the dictionary size.

[0013]

DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be specifically described below with reference to the drawings. <Embodiment 1> FIG. 1 shows the structure of Embodiment 1 of the present invention.
A handwritten character input unit 1 for directly writing handwritten characters and inputting a handwriting as a time-series coordinate point group, a preprocessing unit 2 for normalizing the input characters, and approximating and vectorizing strokes, and a vector direction A feature extraction unit 3 that extracts a stroke as a feature amount, a dictionary 4 in which stroke feature amounts are registered for each character, and a feature amount of an input character and a character in the dictionary are compared. It comprises a character recognition unit 5 for judging whether or not, and a recognized character output unit 6.

A handwritten character input unit 1 such as a tablet converts a character into a time-series coordinate point sequence for each stroke, and a preprocessing unit 2 first determines the size of the input character by using the long side of a circumscribed rectangle. Is normalized to be a predetermined constant value r. Next, a polygonal line approximation is performed by a conventionally known method. Subsequently, the adjacent points are connected in the writing order, and the stroke is converted into a vector sequence, and the preprocessing ends. FIG.
The input handwritten character, (a), and the result of the preprocessing are shown (b).

The feature extraction unit 3 calculates a direction code of a vector as a feature of a stroke represented by a vector sequence. In a conventional method, an isotropic eight-direction code as shown in FIG. 8 is often used. In the present invention, an asymmetric direction code as shown in FIG. 9 is used. In other words, the direction code according to the present invention has one direction code (code 2) that covers a range of 90 degrees vertically upward and horizontally leftward from the center.
Is assigned), and the other areas (for 270 degrees) are divided into five equal parts, and codes 1, 0, 5, 4, and 3 are respectively assigned to the areas for 54 degrees. . As a result, the number of direction codes is reduced to six as compared with the conventional direction code, so that the memory can be reduced accordingly.

FIG. 10A shows a first example of a "loop", and FIG. 10B shows a second example of a "loop".
FIG. 11A shows a first example of “splash”, and FIG. 11B shows a second example of “splash”. FIG. 10, 1
As shown in FIG. 1, in the conventional direction code, the stroke has different values due to the deformation. FIG.
In 0 (a), the code of the stroke (A, B, C, D) is “5317”, and in (b), the code of “521”
0 ”. Therefore, at the time of collation, if both try to recognize, either register the direction code in the dictionary,
It is necessary to perform complicated processing in the matching algorithm.

On the other hand, if the code of the present invention shown in FIG. 9 is used, there is no difference in the characteristic amount due to the degree of deformation shown in FIGS. 10 and 11 (both in FIGS. 10 (a) and 10 (b)). The direction code “4205”, and FIGS. 11A and 11B both become the direction code “42”), increase in the size of the dictionary and complexity of the matching algorithm can be prevented.

Using the characteristic amount extracted by the characteristic extracting unit 3, the character recognizing unit 4 matches the characteristic amount of the character registered in the dictionary 5 for each stroke. First,
Examine the difference in the number of direction codes between the stroke of the input character and the stroke of the character in the dictionary.
Perform matching for each code. The matching is performed by comparing the first code of the input stroke with the code of the stroke in the dictionary, and if the codes match, the code for matching is updated and repeated. At this time, a combination that changes the order of the codes is not performed.

FIG. 12 shows an example of matching. (A)
Shows an example of an input character, and (b) shows an example of a character in a dictionary. The input character is composed of a first stroke and a second stroke, and the second stroke is composed of seven vectors. The characters in the dictionary are composed of a first stroke and a second stroke, and the second stroke is composed of five vectors. (C) shows the correspondence between the input character and the direction code of the character in the dictionary. The code corresponds to the first stroke, and the code “0310” for the second stroke.
Is supported.

If there is no suitable code in all strokes, the characters in the dictionary are replaced and the same collation is performed. If a matching code exists, a matching score is calculated. The number of matched code sets is a, the number of non-matched codes of stroke codes of the input characters is b, and the number of mismatched codes of stroke codes at the time of character recognition in the dictionary is b. When c is set, the matching score s is set as s = (b + c−2 × a) / (b + c + 2 × a). (D) shows a calculation example of the matching score s in the character recognition unit 4, and in this example, -0.4286 is obtained. The matching score is calculated for characters in the dictionary in which the matched code exists, and the character having the smallest value is output to the recognized character output unit 6 as a character recognition result. If the character does not match any character in the dictionary, the input character is not recognized.

<Embodiment 2> FIG. 2 shows the configuration of Embodiment 2 of the present invention. In the second embodiment, a feature extraction unit 7 is added to the configuration of the first embodiment, which extracts a value obtained by quantizing the length of a vector to be an integral multiple of a predetermined constant as a stroke feature amount. Have been. Therefore, compared to the first embodiment, the feature amount of the stroke used for matching is increased by one. Other components are the same as those of the first embodiment.

The feature extraction unit 7 calculates the length of the vector, and calculates the length of the vector by using the value of 1/10 of the normalization constant r used in the preprocessing unit 2 as a reference value. A value obtained by dividing and quantizing the first decimal place is rounded to a vector feature value. FIG. 13 shows a matching example of the same characters as in the first embodiment. In (c), the feature extraction unit 7
Then, the quantization length is extracted, and the quantization length for each character is registered in the dictionary together with the above-described direction code.

The formula for calculating the matching score is changed as follows. First, as in the first embodiment, a matching score s for the direction code is obtained. Next, as shown in FIG. 13D, a set of vectors “5,
0, 3, 1, 0 "is extracted, and the quantization length of the stroke vector of the input character is" 2, 4, 10, 4, 2 "
The absolute value of the difference between the value and the quantization length of the vector of the stroke of the character in the dictionary “2, 6, 11, 5, 3”, and the sum of the respective quantization lengths are obtained. Let d be the sum of the absolute values of the differences in quantization length, and e be the sum of all the sums of the quantization lengths. At this time, the matching score t is set to t = s + d / e. As in the first embodiment, a character having the smallest matching score is output as a recognition result. The use of the feature amount relating to the length of the vector is effective when characters such as “earth” and “shi” have the same stroke direction but different lengths.

<Embodiment 3> FIG. 3 shows the configuration of Embodiment 3, which is configured by adding a length reference value calculation unit 8 to the configuration of Embodiment 2. The length reference value calculation unit 8 calculates the lengths of all the vectors constituting the input character, calculates one tenth of the length as a reference value, and passes the value to the feature extraction unit 7. In handwritten characters, the variation in length often increases, especially in short strokes. In this embodiment, unlike the above-described second embodiment, the reference value for quantization of the length of a vector is set to 10% of the sum of the lengths of all vectors constituting a character.
By using one-half and increasing the reference value for quantization, the effect of length variations can be reduced.

<Embodiment 4> FIG. 4 shows the configuration of Embodiment 4. In the fourth embodiment, a stroke sequence calculator 9 is added to the configuration of the second embodiment. In the case of a character having a plurality of strokes, the stroke sequence calculator 9 creates a vector sequence in which the starting points of the strokes are connected in the writing order. FIG. 14A shows a vector sequence (21, 22, 23) between the starting points of the character "Bold", and FIG. 14B shows a vector sequence (31, 32, 3) between the starting points of the character "Dog".
3) is shown. Then, the feature extraction units 3 and 7 obtain the feature amount of the vector sequence connecting the starting points, and use it for matching.

That is, the character recognizing unit 4 regards a vector string connecting the starting points as a stroke, and calculates a matching score t 'calculated for the stroke in the same manner as the matching score t described in the second embodiment. . Then, as a matching score u, a sum u = t + t ′ of a score t ′ related to the stroke and a score t ′ calculated by regarding a vector string connecting the starting points as a stroke is calculated, and the character having the smallest value is recognized as a recognition result. Output as As described above, in the present embodiment, since the feature amount related to the positional relationship between the strokes of the character is used, “thick”
This is effective when recognizing a set of characters that cannot be identified only by the direction and length, such as “and“ dog ”.

<Embodiment 5> FIG. 5 shows the structure of a fifth embodiment. The present embodiment is configured by adding a stroke sequence calculator 10 to the configuration of the fourth embodiment. In the case of a character having a plurality of strokes, the stroke sequence calculation unit 10 creates a vector sequence connecting the centers of the circumscribed rectangles of the strokes in the stroke writing order, in the same manner as the vector sequence connecting the starting points in the fourth embodiment. Use for matching.

FIG. 15A shows a first example of the character "kawa", and FIG. 15B shows a first example of the character "kawa".
In the case of a character including a stroke having a large variation of the starting point position, such as the second stroke of “river”, a vector sequence connecting the starting points of the fourth embodiment is introduced. Column matching may result in incorrect recognition. That is, FIG.
In (a), the vector sequence at the starting point of the stroke is (4)
1, 42), and its direction code is “0, 0”. In FIG. 3B, the vector sequence is (51, 52), and its direction code is “5, 0”.

Therefore, in this embodiment, a feature quantity relating to the positional relationship of the center of the circumscribed rectangle of the character stroke is added.
By using a plurality of vector sequences related to the positional relationship between strokes, the above-described adverse effects can be reduced. In (a), 43 is the circumscribed rectangle of the first stroke, 44 is the center of the circumscribed rectangle, 45
Are the centers of the circumscribed rectangles of the second stroke;
7 is a vector sequence connecting the centers. The same applies to (b), and the direction codes of (a) and (b) are both “0, 5”.

<Embodiment 6> It is characterized in that a spatially asymmetric direction code used in the present invention can be freely set. Therefore, it is also possible to create a unique direction code for each scribe. FIG. 6 shows a configuration of the sixth embodiment. It is configured by adding a direction code optimizing unit 11 to the configuration of the fifth embodiment. The direction code optimizing unit 11 determines the optimal direction code by comparing and analyzing the writing habit with the characters in the dictionary from the characters written by the writer in advance in order to create the direction code for each writer.

In FIG. 6, the flow of data relating to the dictionary 5 and the direction code optimizing unit 11 at the time of creating a directional code for each writer is indicated by a broken line. To create a writer-specific directional code, a writer is first presented with a certain number of characters to be written. The direction code of the stroke is extracted by the feature amount extracting unit 3, the direction code optimizing unit 11 calculates the average of the direction codes, and compares the average with the standard direction code registered in the dictionary 5.

As shown in the character example of the upper right of FIG.
If the written character is significantly distorted, the standard direction code will deviate from the standard character in the dictionary. When such a deviation is detected by the comparison in the direction code optimizing unit 11, the boundary of the direction code is gradually changed in the direction of the arrow as shown in FIG. Repeat until the character code is almost equal.
The direction code obtained as a result is registered in the dictionary 5 as a direction code unique to the writer. At the time of actually performing the character recognition, the data flow relating to the dictionary 5 and the direction code optimizing unit 11 is changed to the solid line side in FIG. The feature quantity extraction in the quantity extraction unit 3 is executed.

As described above, in this embodiment, since the directional code for each writer is created, it is possible to recognize a character which is significantly deformed as compared with a standard character, and the number of models registered in the dictionary. There is no need to increase. In addition, since it adapts to the writer's habit, the recognition rate is also improved.

[0034]

As described above, according to the first aspect of the present invention, in the recognition of online handwritten characters,
Since the direction of each vector of a stroke constituting a character is replaced with a spatially asymmetric direction code and a row of direction codes is extracted as a feature amount of the stroke, the distortion is large without increasing the dictionary size. Characters can be recognized.

According to the second aspect of the present invention, the character code having the same direction code can be identified by using the direction code of the vector constituting each stroke and the length of the vector together as the feature amount of the stroke. It becomes possible.

According to the third aspect of the present invention, since the reference value for the quantization of the length of the vector is the sum of the character vectors, it is possible to absorb the variation in the length that often appears in a relatively short stroke. Can be.

According to the fourth aspect of the present invention, in a character composed of a plurality of strokes, a characteristic quantity relating to the positional relationship of the starting point of the stroke is introduced, so that characters which cannot be distinguished only by the direction and the length are identified. be able to.

According to the fifth aspect of the present invention, in a character composed of a plurality of strokes, a characteristic quantity relating to a positional relation of a center of a circumscribed rectangle of a stroke is introduced, so a characteristic quantity relating to a positional relation of a starting point is introduced. The adverse effects of this can be mitigated.

According to the sixth aspect of the present invention, by creating a directional code for extracting a stroke feature amount for each writer, a character deformed due to the writer's character habit can be converted into a dictionary size. It can be recognized without increasing.

[Brief description of the drawings]

FIG. 1 shows a configuration of a first exemplary embodiment of the present invention.

FIG. 2 shows a configuration of a second exemplary embodiment of the present invention.

FIG. 3 shows a configuration of a third embodiment of the present invention.

FIG. 4 shows a configuration of a fourth embodiment of the present invention.

FIG. 5 shows a configuration of a fifth embodiment of the present invention.

FIG. 6 shows a configuration of a sixth embodiment of the present invention.

FIG. 7A shows an input handwritten character, and FIG.
Indicates the result of preprocessing.

FIG. 8 shows an isotropic eight-way code.

FIG. 9 illustrates an asymmetric directional code of the present invention.

FIG. 10A shows a first example of a “loop”,
(B) shows a second example of “loop”.

FIG. 11A shows a first example of “splash”,
(B) shows a second example of “splash”.

FIG. 12 is a diagram illustrating an example of matching.

FIG. 13 is a diagram illustrating another example of matching.

14A illustrates a vector sequence between start points of a character “fat”, and FIG. 14B illustrates a vector sequence between start points of a character “dog”.

FIG. 15A shows a first example of a character “kawa”.
(B) shows a first example of the character “kawa”.

FIG. 16 shows an example of a character rising to the right.

FIG. 17 is a diagram illustrating a modification of a direction code.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Character input part 2 Preprocessing part 3 Feature extraction part 4 Character recognition part 5 Dictionary 6 Recognized character output part

Claims (6)

(57) [Claims] 1. A handwritten character input unit in which handwritten characters are directly input as a time-series coordinate point group, a size of the input characters is normalized, and each stroke constituting the characters is approximated by a broken line and connected in writing order. A pre-processing unit that converts the direction of each vector of the stroke into a spatially asymmetric direction code, and a feature extracting unit that extracts the sequence of the direction code as a feature amount of the stroke; And a character recognition unit that recognizes the input character by comparing a dictionary in which the characteristic amount of the stroke is registered with a characteristic amount of the stroke of the input character and a characteristic amount of the stroke of the character in the dictionary. An online handwritten character recognition device, comprising: 2. A handwritten character input unit in which a handwritten character is directly input as a time-series coordinate point group, a size of the input character is normalized, and each stroke constituting the character is approximated by a broken line and connected in writing order. A first processing unit for converting the direction of each vector of the stroke into a spatially asymmetric direction code and extracting the direction code sequence as a first feature amount of the stroke; Department and
A second feature extraction unit that extracts a sequence of values obtained by quantizing the length of each vector of the stroke based on a predetermined reference value as a second feature amount of the stroke; A registered dictionary, and a character recognition unit that recognizes the input character by comparing the feature amount of the stroke of the input character with the feature amount of the stroke of the character in the dictionary. Online handwritten character recognition device. 3. The online handwritten character recognition device according to claim 2, wherein the predetermined reference value is calculated based on a sum of lengths of vectors constituting the character. 4. A first stroke sequence calculation unit for obtaining a first vector sequence connecting a starting point of each stroke in writing order with respect to a character composed of a plurality of strokes, wherein the first and second feature extraction units are provided. 3. The online handwritten character recognition device according to claim 2, wherein the first and second feature amounts are extracted from the first vector sequence. 5. A first stroke sequence calculation unit for obtaining a first vector sequence in which the starting points of each stroke are connected in writing order for a character composed of a plurality of strokes, and the center of a circumscribed rectangle of each stroke is connected in writing order. A second stroke sequence calculating unit for obtaining a second vector sequence; wherein the first and second feature extracting units extract the first and second feature amounts from the first and second vector sequences; 3. The online handwritten character recognition device according to claim 2, wherein: 6. A direction code optimizing unit for comparing and analyzing a writing habit of a writer with characters in the dictionary in advance to create an optimum directional code for each writer, and wherein the feature extracting unit 3. The online handwritten character recognition device according to claim 1, wherein a stroke feature amount is extracted based on the obtained direction code.