JPH1125226A - On-line handwritten input character recognition method by fuzzy vector and device therefor and recording medium for executing the same method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method for recognizing an on-line handwritten input character with a further high accuracy. SOLUTION: A picture stroke 1 of a handwritten input character obtained by on-line is decomposed into a plurality of constituent vectors 3, each constituent vector 3 is converted into a fuzzy vector constituted of membership grades in a plurality of directions, and a fuzzy vector group being the set of the obtained fuzzy vectors is compared with the fuzzy vector group of each dictionary character preliminarily registered in a dictionary. Then, a character corresponding to the handwritten input character is recognized, based on the compared result.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to an online handwritten input character recognition method and apparatus for recognizing online handwritten input characters using a fuzzy associative memory, and a recording medium for executing the method.

[0002]

2. Description of the Related Art Handwritten input character recognition uses its algorithm to perform off-line character recognition for recognizing characters after writing such as OCR, and to capture handwriting information in real time when writing characters such as pen input. And on-line character recognition for identification.

In online character recognition, handwriting information of a character, that is, direction information, is collected, and this direction information is compared with direction information registered in a dictionary in advance to recognize a character most similar in the dictionary. Output as result.

FIG. 24 is an explanatory diagram of 16-direction coding of a stroke in conventional online handwritten input character recognition.

[0005] For example, the second stroke 1 of the "mountain" of the handwritten input character is converted into a 16-way code by the 16-way code shown in FIG. Character recognition is performed by collating the direction code string of the input handwritten input character with the direction code of the character registered in the Kanji dictionary (Journal of the Japan Society of Mechanical Engineers, Vol. 99 No. 926, pages 11-1).
4 (Jan. 1996).

[0006]

However, in the conventional online character recognition method, if the recognition target is a kanji character, a hiragana character, a katakana character, a number, an alphabetic character (large or small), or a symbol, characters having similar direction information are not recognized. There is a problem that recognition is difficult. Further, in the conventional online handwritten input character recognition method, a stroke is decomposed into 16 directions and converted into a 16-direction code. There was a limit that you could not issue.

The present invention provides an online handwritten input character recognition method for recognizing online handwritten input characters with higher accuracy.

[0008]

According to a first online handwritten input character recognition method of the present invention, an image stroke of a handwritten input character collected online is decomposed into a plurality of constituent vectors, and each constituent vector is divided into members in a plurality of directions. It is converted to a fuzzy vector composed of ship grades, and the obtained fuzzy vector family, which is a set of fuzzy vectors, is compared with the fuzzy vector family of each dictionary character registered in the dictionary in advance, and based on the comparison result, Characters corresponding to the handwritten input characters are recognized.

The second online handwritten input character recognition method according to the present invention is characterized in that a plurality of constituent vectors obtained by dividing a stroke of a handwritten input character collected online and each constituent vector of a transition stroke and / or a start and end stroke are provided. Is converted to a fuzzy vector consisting of membership grades in multiple directions, and the obtained fuzzy vector family, which is a set of fuzzy vectors, is compared with the fuzzy vector family of each dictionary character registered in the dictionary in advance. Characters corresponding to the handwritten input characters are recognized based on the result.

In the first and second methods, the following embodiments can be mentioned.

(1) The stroke of the handwritten input character is set to 3
It is divided into constituent vectors with two as the minimum unit.

(2) After extracting a character most similar to a handwritten input character from a dictionary, if the extracted character corresponds to a predetermined character, the characteristic point unique to the character is used to extract the character. The handwritten input character is identified, and the identification result is output.

(3) The stroke data of the handwritten input characters collected online and the corresponding dictionary characters are registered in the dictionary.

Further, the handwritten character recognition device of the present invention has a character input device, means for collecting characters handwritten by the character input device online, and decomposes a stroke of the handwritten input character into a plurality of constituent vectors. Means,
Means for converting each constituent vector into a fuzzy vector comprising membership grades in a plurality of directions, comparing the obtained fuzzy vector family, which is a set of fuzzy vectors, with the fuzzy vector family of each dictionary character registered in the dictionary in advance Means, and means for recognizing the handwritten input character based on the comparison result.

Further, a computer readable recording medium for executing the first method includes a step of decomposing an image stroke of a handwritten input character collected online into a plurality of component vectors, A procedure for converting into a fuzzy vector consisting of membership grades, a procedure for comparing a fuzzy vector family that is a set of obtained fuzzy vectors with a fuzzy vector family of each dictionary character registered in the dictionary in advance, and the comparison result And a program for executing a procedure for recognizing the handwritten input character based on the program.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is an explanatory diagram of a drawing stroke of a handwritten input character used in the present invention.

An image stroke 1 is a handwriting which is obtained by extracting a movement of a brush when the character is in contact with paper when writing a character as continuous direction information. The handwritten input characters include a plurality of drawings except for one-stroke characters such as numbers, hiragana, katakana, alphabets, kanji, symbols, etc. Each drawing has a handwriting of a brush (pen). The handwriting information expressing the movement of the brush for one stroke is called "stroke stroke", and one handwritten character has the same number of strokes as the number of character strokes. A character has a writing start position, and the character is formed while changing the direction from there. Since the way of changing the direction is peculiar to each character, expressing the image stroke 1 with the constituent vector 3 is very effective for character identification.

The constituent vector 3 represents the movement of the brush in the process of writing the handwritten input character, and the strokes of all the handwritten input characters can be decomposed into constituent vectors representing direction information. Since the total number of “component vectors” is the minimum unit that represents the handwriting of a handwritten character, in order to extract the handwriting information, at least the number of strokes of the handwritten input character is required. Considering that the number of decompositions is a power of 2 which is easy to digitally process and that the maximum number of strokes of kanji is 30 strokes (ran), it is preferable that the minimum unit is 32. The total number of constituent vectors is 3 with one handwritten input character.
When unified into two, a character having a plurality of strokes has (32 / number of strokes) component vectors per stroke. That is, the number of component vectors does not change depending on the length of all the images, and the number of component vectors per image is uniquely determined by the number of images. At this time, a decimal value expressing the number of component vectors is a problem. To avoid this, the following processing is performed by a program.

For example, in the case of three strokes of the handwritten input character "upper", when dividing into three integers such that the total is 32, if the number of constituent vectors per stroke is calculated to a decimal, 32/3 = 10. 667 (pieces). Next, 32
When ／ is multiplied by 1, 2 and 3 times and rounded down, (32/3) × 1 = 10.667 ≒ 11 (32/3) × 2 = 21.333 ≒ 21 (32/3) × 3 = 32 Therefore, the vector numbers of each image are distributed as follows: 1st image 1-11 (No.) 11 vector 2nd image 12-21 (No.) 10 vectors (21-11 = 10) 2nd image 22-32 (No. ) 11 vectors (32-21 = 11).

FIG. 2 is an explanatory diagram of a transition stroke used in the present invention.

The transition stroke is a constituent vector obtained by extracting the direction of the brush in the air away from the paper when writing a character. Many handwritten input characters can be recognized by strokes, but some characters cannot be recognized. Characters having the same stroke but different shapes are the target characters. This means that handwritten characters have features that cannot be completely expressed by strokes alone.
By combining the image stroke with the transition stroke, the recognition rate of the handwritten input character is improved as compared with the case where the image stroke is used alone. The transition stroke exists in a character having two or more strokes, and is non-handwriting direction information from the end point of each stroke to the start point of the next stroke. By combining the direction information of the transition stroke for characters having similar image strokes, the recognition rate is improved as compared with the case where only the image stroke is used. The number of transition strokes to be extracted can be represented by (number of strokes-1). In the case of “upper” and “lower” of the three strokes shown in FIG. 2, a component vector of a transition stroke from the end point of the first stroke to the start point of the second stroke: a From the end point of the first stroke to the start point of the second stroke The constituent vector of the transition stroke is b.

FIG. 3 is a diagram showing a difference in change of the transition stroke.

For example, in the case of "F" and "I", the constituent vectors of the first to third strokes are exactly the same. However, the direction of the first transition stroke a is the same, but the direction of the second transition stroke b is completely different, so that “F” and “I” can be distinguished by the difference of the second transition stroke. As described above, it is possible to determine a character by utilizing the property that the movement of the brush in the air varies depending on the character.

FIG. 4 is an explanatory diagram of the start and end strokes used in the present invention.

The start and end strokes are component vectors obtained by extracting direction information from the start point of the first stroke of the handwritten input character to the start and end points of each stroke. In the case of “upper” and “chi” shown in FIG. 4, the starting and ending stroke vectors from the starting point of the first stroke to the starting point of the second stroke: c The starting and ending strokes from the ending point of the first stroke to the ending point of the second stroke Component vector: d Component vector of starting and ending strokes from the end point of the first drawing to the starting point of the third drawing: e Component vector of starting and ending strokes from the ending point of the first drawing to the ending point of the third drawing: f This is effective for handwritten input characters that are difficult to distinguish by only two types of strokes and transition strokes. For example, "top" and "shi"
Characters with similar strokes and transition strokes are difficult to distinguish, but by adding start and end strokes, which are component vectors that extract direction information from the character start point to the start and end points of each image, the start and end strokes of both characters can be determined. Because they have quite different directions, a clear distinction is possible.

Next, an on-line handwritten input character recognition method using a fuzzy associative memory according to the present invention will be described.

The lines constituting the handwritten input characters written by the mouse 12 on the CRT 13 of the personal computer 10 shown in FIG. 5 appear to be continuous, but in fact, as shown in FIG. It is composed of Since these point data are all coordinate data, they include distance data between points, that is, information on the size of characters. Therefore, even if the characters are exactly the same, the character size varies depending on the case where the writer is different and the environment around when the character is written. FIG. 7 shows an example in which the character sizes are different. However, although the character size is different, the same character must naturally output the same recognition result in handwritten input character recognition. In order to derive the same recognition result even if the character size is different, it is necessary to convert the coordinate data into angle data. The method is described below.

A handwritten input character composed of a series of dots is
It divides into a plurality, in this example, 32, and replaces them with a total of 32 points. Next, the absolute angle of a straight line connecting the points is calculated. Based on the coordinate data of each point, ΔX and ΔY (see FIG. 8) are obtained, and a straight line connecting the two points is created. A straight line having this direction is a “component vector”, and equations (1) to (4) are used to derive the absolute angle θ.

Θ = arctan | Δy / Δx | (Δx> 0, Δy> 0) (1) θ = 180 ° −arctan | Δy / Δx | (Δx <0, Δy> 0) ((1) 2) θ = 180 ° + arctan | Δy / Δx | (Δx <0, Δy <0) (3) θ = 360 ° -arctan | Δy / Δx | (Δx> 0, Δy <0) ... (4) Next, the calculated angle θ is defined as (“0”: 0 ° direction, “
1 ": 90 ° direction," 2 ": 180 ° direction," 3 ": 2
(70 ° direction). The fuzzy vector represents one θ using membership grades in four directions (variables are represented by μ), and is obtained by equations (5) and (6) and Table 1.

Μ _A = 1−θ / 90 ° (5) μ _B = θ / 90 ° (6)

[Table 1] A function indicating the relationship between the membership grade and θ in these four directions is called a membership function and is shown in FIG. In the figure, the portions indicated by the dashed lines overlap, and are 0 ° to 360 °.
Are continuously expressed in the range. FIG. 10 shows an example in which four fuzzy vectors are actually obtained from the constituent vectors.
13 to FIG.

FIG. 10 shows the relationship between the configuration vector and the membership grade in the "0" direction. The shaded area represents the area where the membership function is zero.

Similarly, FIGS. 11 to 13 show the relationship between the constituent vectors and the membership grades in the "1", "2", and "3" directions, respectively. The hatched area indicates an area where the membership grade is 0.

FIGS. 14 and 15 are flowcharts showing the handwritten input character recognition system of the present invention.

Step 1 Start the program.

Step 2 It is determined whether or not the handwritten character to be inputted is already registered in the dictionary.

Step 3 If the dictionary has been registered, input the handwritten input characters as input characters.

Step 4 If the dictionary has not been registered, input the handwritten character as a dictionary registered character.

Step 5: A constituent vector is extracted from each of the image stroke, the transition stroke, and the start and end strokes of the handwritten input character.

Step 6 Each extracted constituent vector is converted into a fuzzy vector composed of four membership grades.

The absolute angle θ of each of the extracted constituent vectors is 4
The directional membership grades μ ₀ , μ ₁ , μ ₂ , and μ ₃ are determined by Equations 1 to 4, and the membership grade is determined by Equations 5 and 6 according to Table 1.

By the above processing, four membership grades μ ₀ , μ ₁ , μ ₂ , μ ₃ are output from one component vector, and these membership grades μ ₀ , μ ₁ ,
All the patterns of μ ₂ and μ ₃ are stored in the memory.

Step 7 Then, for the dictionary characters having the same number of strokes as the handwritten input characters, all the membership grades of the handwritten input characters and the dictionary characters obtained in step 6 are compared, and the input characters and the dictionary characters are compared. Is calculated.

When character recognition is performed, a method based on the area ratio of vectors when a fuzzy vector is displayed two-dimensionally is used. The similarity S is obtained by the following procedure. First, the similarity S _C of all the constituent vectors is determined by setting the membership grades of the constituent vectors of the input character to μ _T0 , μ _T1 , μ _T2 , μ.
_{T3, D0} membership grade configuration vector of the dictionary character μ, μ _D1, μ _D2, when the mu _D3 obtained by the following expression.

[0044]

(Equation 1) In the above equation, max: the larger value comparing the two values
min: the smaller of the two values.

Next, the similarity of each stroke is obtained by the following equation. The similarity _SOR of the image stroke is as follows when the similarity of the component vectors of the image stroke is S _COR .

[0046]

(Equation 2) Regarding the similarity _STR of the transition stroke, _assuming that the similarity of the constituent vectors of the transition stroke is S _CTR , the following is obtained.

[0047]

(Equation 3) Regarding the similarity S _SE between the start and end strokes, if the similarity between the constituent vectors of the start and end strokes is S _CSE ,

(Equation 4) Ask for.

The similarity S is obtained by weighted addition of the similarity of each stroke. The mixing ratio (W1: W2: W3) is shown in FIG.
As shown in FIG. 6, the weight of each stroke can be changed in ten steps.

[0049]

(Equation 5) Step 8 The similarity calculation in step 7 is performed for all dictionary characters, and the result is assigned to all dictionary characters.

Step 9 The one having the highest similarity in Step 8 is selected.

Step 10 The maximum similarity is extracted from the dictionary, and the dictionary character having the maximum similarity is output as an answer (first candidate).

Step 11 Characters output as first candidates in the procedure up to Step 10 are characters having almost the same stroke and easily recognized as erroneous, for example, "0" and "6", "Ko" and "Yu", etc. Is determined by a dedicated rule regardless of the similarity extraction result.

Step 12 FIGS. 17 to 21 are explanatory diagrams of the exclusive rule. FIG. 17 shows the case of “0” and “6”. If the height h2 of the start and end points is larger than ４ of the character height h1, it is identified as “6”, and if smaller than １ ／, it is identified as “0”. I do.

FIG. 18 shows the case of "U" and "Y". If the length w2 of the end point of the first stroke and the end point of the second stroke is larger than 1/4 of the character length w1, "YU" is displayed. , １ ／, it is identified as “co”.

FIG. 19 shows "P", "D", "b" and "W".
If the height h2 of the start point of the second stroke and the start point of the first stroke is larger than 1/5 of the character height h1, "b"
The height h3 between the end point of the first stroke and the end point of the second stroke is the character height h1.
If the height h4 of the ending point of the second stroke and the ending point of the first stroke is greater than 1/10 of the character height, it is identified as "W" if it is larger than 1/4 of the character height and "D" otherwise. .

FIG. 20 shows the case of "9" and "a". If the distance w1 (horizontal direction) from the start point to the end point is larger than +8 dots (+ as the right direction), "a", 8 dots If smaller, it is identified as "9". If the value obtained by dividing the vertical distance between the start point and the end point by the horizontal distance between the start point and the end point is greater than 15, the data is identified as “9”.

FIG. 21 shows the case of "n" and "h", where the value obtained by dividing the vertical distance h1 between the start point and the end point by the horizontal distance w1 between the start point and the end point is greater than 1.5. In this case, it is identified as “h”, and when it is smaller than 1.5, it is identified as “n”.

In this way, a wide range of character recognition is possible for kanji, hiragana, katakana, numbers, alphabets (large and small), and symbols.

Step 13 It is visually determined whether or not the handwritten input character matches the output result.

Step 14 If the handwritten input character does not exist as dictionary character data, it is determined whether or not to register it.

Step 15 When registering, a correct registered character is input from the keyboard.

Step 16 Input character data (fuzzy vector family) and registered characters are automatically written and registered in the dictionary file.

In the above embodiment, an example is shown in which the constituent vectors of the handwritten input character are converted into fuzzy vectors in four directions. However, the constituent vectors can be converted into fuzzy vectors in at least two directions. Absolute angle θ when using fuzzy vectors in two, three, and five directions
FIG. 25 shows a membership function indicating the relationship between and membership grade.

Further, in the above-described embodiment, an example in which the character recognition method is performed by comparing the similarity is shown, but the character recognition method may be performed by comparing the difference. An example is shown in FIG.

Up to step 6, the same method as the calculation based on the similarity is used.

Step 7 ': For a dictionary character having the same number of strokes as the input character, calculate the difference between each membership grade of the fuzzy vector family of the input character and each membership grade of the fuzzy vector family of the dictionary character for one character. I do. The difference is a different degree, and the closer to 0, the more similar. The dissimilarity includes a Hamming distance which is an integral of a difference between each membership grade, a Euclidean distance which is a distance per one membership grade, and DP matching.

Step 8 ': The differences between all the target dictionary characters are calculated, and the calculated differences are assigned to the dictionary characters.

Step 9 'A character having the minimum difference is selected from the dictionary characters.

Step 10 'The dictionary character having the minimum degree of difference is output as the answer (first candidate).

From step 11, the same method as the calculation method based on the similarity is used.

Next, FIG. 23 shows a configuration example of the character recognition device according to the present invention.

FIG. 23 is a block diagram showing a configuration of a computer for performing the processing of FIGS. 14 and 15 described above. FD or HD or an auxiliary memory) 11, an input device (mouse or pad and keyboard) 12 for inputting handwritten input characters, and a display device (CRT) 13 for displaying input characters and dictionary characters recognized as input characters And a main memory 14 for temporarily storing stroke data of handwritten input characters, stroke data of dictionary characters, and similarity of each dictionary character, and conversion of stroke data into fuzzy vector family data and calculation of similarity. CPU 15 to perform
And an input / output control unit 16 for controlling transmission of data between devices and instructions of programs. The procedure of character recognition in this character recognition device is as follows.

(1) Start the handwritten input character recognition program recorded on the recording medium 11.

(2) Handwritten characters are input by the mouse of the input device 12.

(3) The CPU 15 converts the stroke data of the handwritten input characters into fuzzy vectors, stores the converted fuzzy vector family and stroke count data in the memory 14, and displays the handwritten characters on the CRT 13.

(4) A fuzzy vector family of one character having the same stroke number is extracted from the dictionary character file and stored in the memory 14.

(5) The CPU 15 calculates the similarity between the two types of fuzzy vector families stored in the memory 14 and stores the dictionary characters and the similarity in the memory 14.
(Similarities are calculated for all dictionary characters having the same stroke count and stored in the memory 14.) (6) The dictionary characters stored in the memory 14 are sorted by the CPU 15 in descending order of similarity, and the dictionary characters having the highest similarity are sorted. Is displayed on the CRT as the recognition result of the handwritten input character.

[0078]

As described above, according to the present invention, an image stroke of a handwritten input character collected online is decomposed into a plurality of constituent vectors, and each constituent vector is converted into a fuzzy vector composed of membership grades in a plurality of directions. Converted, the obtained fuzzy vector family, which is a set of fuzzy vectors, is compared with the fuzzy vector family of each dictionary character registered in the dictionary in advance, and based on the comparison result, corresponding to the handwritten input character By recognizing the characters, the angle information of the handwriting can be continuously and finely expressed by the fuzzy vector family, so that the online handwritten input characters can be recognized with high accuracy. In addition, the recognition rate can be further improved by combining the identification with the transition stroke and the start and end strokes.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram of an image stroke.

FIG. 2 is an explanatory diagram of a transition stroke used in the present invention.

FIG. 3 is a diagram showing a difference between transition strokes.

FIG. 4 is an explanatory diagram of a start and end stroke used in the present invention.

FIG. 5 is an explanatory diagram showing an example of an input method of handwritten input characters.

FIG. 6 is an explanatory diagram showing that a handwritten input character is composed of a series of dots;

FIG. 7 is an explanatory diagram showing a case where the sizes of handwritten input characters are different.

FIG. 8 is an explanatory diagram of an absolute angle θ of a constituent vector.

FIG. 9 is an explanatory diagram of a membership function indicating a relationship between an absolute angle θ of a constituent vector and a membership grade.

FIG. 10 is an explanatory diagram showing a relationship between a configuration vector and a membership grade in the “0” direction.

FIG. 11 is an explanatory diagram showing a relationship between a configuration vector and a membership grade in the “1” direction.

FIG. 12 is an explanatory diagram showing a relationship between a configuration vector and a membership grade in the “2” direction.

FIG. 13 is an explanatory diagram showing a relationship between a configuration vector and a membership grade in the “3” direction.

FIG. 14 is a flowchart showing a handwritten input character recognition system of the present invention.

FIG. 15 is a flowchart showing a handwritten input character recognition system of the present invention.

FIG. 16 is an explanatory diagram of how to determine similarity.

FIG. 17 is an explanatory diagram of a dedicated rule.

FIG. 18 is an explanatory diagram of a dedicated rule.

FIG. 19 is an explanatory diagram of a dedicated rule.

FIG. 20 is an explanatory diagram of a dedicated rule.

FIG. 21 is an explanatory diagram of a dedicated rule.

FIG. 22 is a flowchart illustrating a character recognition method by calculating a degree of difference.

FIG. 23 is a block diagram illustrating a configuration of a computer for implementing the present invention.

FIG. 24 is an explanatory diagram of 16-direction coding of a stroke for conventional online handwritten input character recognition.

FIG. 25 is an explanatory diagram of a membership function showing a relationship between an absolute angle θ and a membership grade when fuzzy vectors in two, three, and five directions are used.

[Explanation of symbols]

1 stroke, 2 16 direction code, 3 configuration vector, 10 personal computer, 11 FD, HD or auxiliary memory, 12 input device, 13 CRT, 14 memory,
15 CPU, 16 I / O controller

Claims (7)

[Claims] An image stroke of a handwritten input character collected online is decomposed into a plurality of constituent vectors, each constituent vector is converted into a fuzzy vector including membership grades in a plurality of directions, and a set of the obtained fuzzy vectors is used. An online handwriting method comprising comparing a certain fuzzy vector family with a fuzzy vector family of each dictionary character registered in a dictionary in advance, and recognizing a character corresponding to the handwritten input character based on the comparison result. Input character recognition method. 2. A plurality of constituent vectors obtained by dividing an image stroke of a handwritten input character collected online and each constituent vector of a transition stroke and / or a start and end stroke into fuzzy vectors composed of membership grades in a plurality of directions. Converted, the obtained fuzzy vector family, which is a set of fuzzy vectors, is compared with the fuzzy vector family of each dictionary character registered in the dictionary in advance, and based on the comparison result, corresponding to the handwritten input character An online handwritten input character recognition method characterized by recognizing characters. 3. The online handwritten input character recognition method according to claim 1, wherein an image stroke of the handwritten input character is divided into constituent vectors having a minimum unit of 32 strokes. 4. After extracting a character that is most similar to a handwritten input character from a dictionary, if the extracted character corresponds to a preset character, the handwriting is performed using a characteristic point unique to the character. 4. The online handwritten input character recognition method according to claim 1, wherein input characters are identified. 5. The online handwritten input character recognition method according to claim 1, wherein stroke data of handwritten input characters collected online and dictionary characters corresponding thereto are registered in a dictionary. . 6. A character input device, means for collecting characters handwritten by the character input device online,
Means for decomposing a stroke of a handwritten input character into a plurality of constituent vectors, means for converting each constituent vector into a fuzzy vector comprising membership grades in a plurality of directions, a fuzzy vector family which is a set of obtained fuzzy vectors, and An online handwritten character recognition device comprising: means for comparing fuzzy vector families of respective dictionary characters registered in a dictionary; and means for recognizing the handwritten input character based on the comparison result. 7. A procedure for decomposing an image stroke of a handwritten input character collected online into a plurality of constituent vectors, a procedure for converting each constituent vector into a fuzzy vector including membership grades in a plurality of directions, and a step of converting the obtained fuzzy vector. A program for executing a procedure of comparing a set of fuzzy vector families with a fuzzy vector family of each dictionary character registered in the dictionary in advance, and a procedure of recognizing the handwritten input character based on the comparison result A computer-readable recording medium on which is recorded.