JPH0632084B2 - Handwritten character recognition method by fuzzy reasoning - Google Patents

Description

The present invention relates to handwritten characters for recognizing various characters such as handwritten kanji, hiragana, katakana, kanji numbers, English letters and alphanumeric characters by fuzzy inference. It concerns the recognition method.

(Prior Art) When a handwritten character is input as input data to, for example, an electronic computer, it is extremely important to accurately recognize the handwritten character.

Therefore, various researches have been performed on the means for recognizing handwritten characters. Most of the above-mentioned conventional handwritten character recognition means extract features from the space-time axis of handwritten character input data.

(Problems to be Solved by the Invention) When recognizing a handwritten character, it is necessary to finally recognize it in real time. Therefore, how much processing time is required for each input character depends on the handwritten character. It becomes an important factor in determining the effectiveness as a recognition device.

However, the conventional handwritten character recognition means generally requires a large amount of calculation, so that the processing time is often long, and in order to solve this, there is a problem that a high-speed and expensive electronic computer must be used. It was

Therefore, in the present invention, the Fourier transform corresponding to the movement amount of the X, Y coordinate point sequence is performed in the stroke unit of the character in the handwriting process, the movement amount of the X, Y coordinate point sequence is treated as a frequency domain, and fuzzy inference is performed. It is a technical subject to be solved that the calculation amount is reduced by performing the above, and the processing time required for recognizing the input handwritten character is shortened.

(Means for Solving the Problem) The technical means for solving the above problem is a method for recognizing a handwritten character by fuzzy inference, in the process of handwriting a character in a character input means having a handwriting surface for handwriting a character, The same character is output from the same character input means as point sequence data corresponding to the X coordinate and the Y coordinate at predetermined time intervals,
The point sequence data corresponding to the handwritten characters output from the character input means is input to the input data normalization means, and the input data normalization means unifies the sizes of the characters handwritten by the character input means. , In order to make the writing speed of the input handwritten character constant, the interval of the point sequence of the handwritten character is made constant, and then the handwritten character normalized by the input data normalizing means is subjected to Fourier transform means. In the Fourier transform, Fourier transform corresponding to the X and Y movement amounts is performed, the intensity of the frequency is obtained, and the Fourier series data obtained by the Fourier transform in the Fourier transform means is converted into ambiguous handwritten character data by the fuzzy conversion means. After fuzzifying so that it can be handled, the rule generation means fuzzy pattern data Obtaining the pattern data from the standard pattern storage means for storing, generating a production rule for handwritten character recognition, after that the fuzzy inference means of the same number of strokes as the handwritten character input in the character input means. Between each of the pattern data stored in the standard pattern storage means and the fuzzy handwritten character data, fuzzy inference based on the production rule, to determine the highest confidence, This is to recognize the handwritten character in the process of outputting the character determined by the fuzzy inference means to have the highest certainty as standard character data corresponding to the input handwritten character from the recognized character output means.

(Operation) According to the handwritten character recognition method by fuzzy inference described above, the character input means inputs the character as point sequence data corresponding to the X coordinate and the Y coordinate at a predetermined time interval in the process of handwriting the character. Output to the data normalization means.

The input data normalizing means that inputs the point sequence data unifies the size of the input handwritten characters and also sets the interval of the point sequences of the handwritten characters in order to keep the writing speed of the input handwritten characters constant. Keep it constant. Then, the Fourier transform means performs a Fourier transform of the handwritten characters normalized by the input data normalization means in units of strokes corresponding to the X and Y movement amounts, obtains the frequency intensity, and further, in the fuzzy means, the Fourier transform means. The fuzzy conversion is performed so that the Fourier series data obtained by the Fourier transform can be treated as ambiguous handwritten character data.

On the other hand, the fuzzy inference means is a fuzzified input character that can be treated as ambiguous handwritten character data by the fuzzification means, and a pattern having the same number of strokes as the input character and stored in the standard pattern storage means. Fuzzy inference is performed on each of the data based on the production rules, and the one with the highest certainty is determined.
The recognized character output means outputs the character determined to have the highest certainty factor by the fuzzy inference means as standard character data corresponding to the input handwritten character.

(Embodiment) Next, an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a handwritten character recognition system. As shown in the figure, a tablet-shaped media graph 1 is used as a character input means, and characters handwritten on the media graph 1 are coordinate point sequence data corresponding to X and Y coordinates in the process of handwriting. Is input to the personal computer 2.

The media graph 1 has an effective reading range of 21
0 mm × 148 mm, resolution is about 0.1 mm, point reading error is ± 1 mm, effective reading height is 3 mm or less,
The point transfer speed is set to 35 points / second, and the point sequence data is set to be transferred to the personal computer 2 when the distance between the points becomes 1 mm or more.

When the point sequence data is transferred from the media graph 1 to the personal computer 2, the coordinate point sequence of each stroke of the handwritten character is accompanied by the stroke start and end information of the stroke according to the input order. Software, that is, transferred to the input data normalization unit 3.

Generally, the size of characters handwritten on the media graph 1 is different and the writing speed is also different. Therefore, the input data normalization unit 3 which is software of the personal computer 2 is also used.
Will normalize the character size and the writing speed for the input coordinate point sequence. Therefore, the size is normalized by shrinking or enlarging the input character so that it fits exactly in a 256-bit square, for example. However, the vertical and horizontal reduction rates or enlargement rates are the same. In the case of vertically or horizontally handwritten characters, there are blanks in the horizontal or vertical direction.However, since the space is set in a certain direction, the characters written vertically or horizontally are left-justified. To do. Further, in order to match the coordinate system of the square with the coordinate system of the display screen of the personal computer 2, the upper left is the origin and the Y coordinate is downward.

On the other hand, regarding the normalization of the writing speed, the above Media Graph 1
Based on the coordinate point sequence data input from, obtain a new coordinate point sequence such that the line length written in unit time becomes constant,
These new coordinate point sequence data are used as data for Fourier transform.

The Fourier transform in the Fourier transform unit 4 is performed for each stroke of the character written in the media graph 1 with respect to the amount of movement on the X axis and the amount of movement on the Y axis from the beginning of writing the stroke. Therefore, the given coordinate point sequence is converted into the movement amount on each axis. Figure 2 (A) shows
With respect to the character “NO”, the amount of movement on the X axis and the amount of movement on the Y axis are shown. By the way, when the Fourier transform is performed on the waveform as shown in FIG. 2 (A), since the start point and the end point do not match, the Fourier series for the discontinuous waveform is obtained. . For this reason, the Fourier series has a poor convergence rate. Therefore, the waveform is line-symmetrically folded at the end point as shown in FIG. 2 (B) so that the waveform becomes continuous, and the Fourier transform is performed on this waveform. Is.

By Fourier transform, each coefficient of f (t) = a0 / 2 + a1cosθt + b1sinθt + a2cos2θt + b2sin2θt + a3cos3θt + b3sin3θt ... Can be obtained. 3 (A) to 6 (A)
Shows the typical strokes, respectively.
(B) to FIG. 6 (B) show the amount of movement of each stroke on the X axis, and FIGS. 3 (C) to 6 (C) show the Fourier transform for the amount of movement on the X axis. It shows the value of each coefficient when. As described above, by folding back the waveform in line symmetry at the end position,
The bn terms (n = 1,2,3, ...) In the Fourier transform equation have small values, and therefore are not shown in the figure.

As shown in FIGS. 3 (C) to 6 (C), the coefficient a0 / 2 indicates the position of the center of gravity of the stroke, and a1 indicates the distance between the start point and the end point on the axis. , A2 represents the property of showing the bending condition on the axis. Note that a3 and a4 are a1,
Although it is considered that they have complementary meanings to a2, respectively, a3 and a4 are not used in the process of recognizing handwritten characters.

As described above, the length of each stroke and the intensity correspondence value of each frequency obtained by the Fourier transform are transferred to the fuzzification unit 5.

In general, the stroke length of a handwritten character or the strength of the frequency is different every time the same person writes the same character, and further varies depending on the person who writes.
Therefore, it must be considered that these data obtained from handwritten characters are not absolute, but indicate that they are close to the values. Therefore, it is appropriate to represent the above data using fuzzy values.
That is, an expression having ambiguity such as a very long stroke or an extremely short stroke length is used, and a similar expression is used for the frequency coefficient (strength). By using such ambiguous expressions, the production rules themselves for recognizing handwritten characters become easy to understand, and it is also included in this expression that some expressions close to it are included. You can have it.

Therefore, the fuzzy value related to the stroke length used in the fuzzification section 5 and its corresponding value are shown in FIG.
The fuzzy value for the frequency coefficient a0 / 2 and its corresponding value are shown in FIG. 8, the fuzzy value for the frequency coefficient a1 and its corresponding value are shown in FIG. 9, and the fuzzy value for the frequency coefficient a2 and its The corresponding values are shown in FIG. In addition,
In the personal computer 2, the fuzzy value is from O to F.
The hexadecimal numbers up to are used for convenience. This convenience value is also shown in FIGS. 7 to 10.

In addition, FIG. 11 shows the distribution state when the stroke length and frequency intensity are converted into fuzzy values for all the 14 Chinese kanji written by a certain person.

Generally, the stroke length is distributed to the larger one when the number of strokes is small and to the smaller one when the number of strokes is large.
As shown in FIG. 1, in the 14th stroke, it is already distributed in the smaller side. Further, a0, which represents the center of gravity of the stroke, has a substantially uniform distribution on both the X axis and the Y axis. A1 that indicates the distance between the start point and the end point is distributed with a slight inclination to the center. This is because the stroke length becomes shorter as the number of strokes becomes larger. Furthermore, a2, which indicates the degree of bending of the stroke, is inclined to the center. This is because there are few bending strokes.

Therefore, when the data input to the fuzzification section 5 is fuzzified and a fuzzy value is assigned, the fuzzification section 5 uses the values shown in FIGS. 7 to 11. However, the above data is not completely contained in the fuzzy value given to it, but has the possibility of being included in the fuzzy values in its vicinity. In fuzzy theory, the possibility of being included in a fuzzy value is called a membership value, and the relationship between the fuzzy value and the membership value is represented by a membership function. Membership functions are often represented by triangles. FIG. 12 shows the above example, and shows that the membership value at the fuzzy value given to the data is 1, and the membership value decreases at a rate of 0.1 as the distance from the value increases.

Next, the standard pattern portion 6 will be described.

In the standard pattern portion 6, a character handwritten as a standard character is stored in the form of a fuzzy value after being subjected to Fourier transform and fuzzy conversion. In addition, the rule generator 7
Then, the fuzzy data is extracted from the standard pattern portion 6, and a production rule is created for each standard character from this. This production rule is created for strokes and takes the form of "if conditional then conclusion". Further, the conditional statement is configured as a logical product of a plurality of conditions. The respective conditions define the conditions for each of the fuzzy data, that is, the stroke length and frequency strength. For example, it is assumed that the character "suspect" is input in a pattern as shown in FIG. 13 (A) and is stored in the standard pattern portion 6 as fuzzy data as shown in FIG. 13 (B). From this, the following production rules are created.

Rule “Suspect” 1: In the first stroke, the stroke length is considerably short, and when viewed in terms of the amount of movement of the X axis, the center of gravity of the stroke is considerably close to the left end, and the end point is to the right of the start point. The curve is horizontal, the center of gravity of the stroke is very close to the upper end, and the end point is considerably close to the start point. If the bend is vertical, generate the rule that this character is "suspect".

Rule “Suspicious” 2: In the second stroke, the stroke length is short, and the center of gravity of the stroke is very close to the left end, and the end point is close to the left side with respect to the start point when viewed in terms of the X-axis movement amount. The curve is slightly concave, and when viewed from the amount of movement of the Y-axis, the center of gravity of the stroke is considerably close to the upper end, and the end point is close to the start point. If is convexly and slightly bent, it produces the rule that this character is "suspect".

Next, the fuzzy inference unit 8 will be described.

In the fuzzy inference unit 8, the condition is satisfied and the condition is determined according to the condition. Therefore, first, the degree of satisfaction of the conditional sentence will be described using the term “confidence degree”.

As described above, the conditional sentence in the production rule is expressed as the logical product of the conditions. Therefore, it is necessary to determine the condition satisfaction, that is, the certainty factor of the condition and the certainty factor of the logical product of the conditions. Therefore, in the present embodiment, in consideration of ease of calculation, the confidence factor of each condition compares two membership functions and takes the smaller membership value at each fuzzy value, and takes the maximum value among them. It takes min-max (maximum of the minimum), and the confidence for the logical product of the conditions is min (minimum) in the confidence of the conditions. That is, the certainty factor is determined as follows. The description of the condition is "when A is A '", and A'is a fuzzy value given from the standard pattern. In addition, a fuzzy value of A ″ is obtained for A from the input character as well. For example, in the rule of “suspect” 2, “short stroke length” is a condition, but under this condition, A ′ is “short”. Yes, A is the stroke length. At this time, the stroke length indicates the length of the second stroke of the input character, and has a fuzzy value such as short or long. The confidence factor for this condition is obtained from the two fuzzy values, which uses the membership function indicated by the fuzzy value.

FIG. 14 and FIG. 15 are explanatory views for obtaining the certainty factor for the above conditions. The confidence about the condition is obtained from the membership function obtained from the standard pattern and the membership function obtained from the input character pattern, which is performed as follows. 2 for each fuzzy value
Compare the membership values of two membership functions,
The one with the smaller value is taken. Next, take the maximum membership value selected in this way. This is the certainty factor for the condition. Figures 14 and 15 show "Coagulation"
"Stroke length is short", which is one of the conditions of rule 2
It shows a method of obtaining the certainty factor for the condition of. Since the length of the second stroke is short in the standard pattern, the membership function is a triangle with the membership value 1 at the "short" position (4 in the figure). That is, the waveform on the left side of FIG. 14 is obtained. Here, it is assumed that the length of the second stroke in the input character is a little short. At this time, the membership function for the length of the second stroke of the input character is a triangle with the membership value 1 at the "slightly short" location (6 location in the figure). That is, the waveform on the right side of FIG. 14 is obtained. Next, the smaller membership value is selected according to the fuzzy value, and the waveform shown in FIG. 15 is obtained. Choose the largest membership value that is greater than this waveform. In the figure
Since it is 0.8, this is a certainty factor for the condition that the stroke length is short when a slightly shorter stroke is written (see Fig. 15).

Regarding the condition connected by the logical product, the smaller one of the certainty factors of the condition is set as the certainty factor of the condition connected by the logical product.

It is assumed that the characters in FIG. 16 (A) have been entered. At this time, the fuzzy value for the second stroke is as follows. The stroke length is a little short. Also, when viewed in terms of the amount of movement of the X axis, the center of gravity of the stroke is considerably close to the left end, the end point is considerably close to the left with respect to the starting point, and the end point is considerably close to the left with respect to the starting point. And the bend is slightly concave. Further, when viewed in terms of the amount of movement of the Y axis, the center of gravity of the stroke is considerably close to the upper end, and the end point is close to the start point and the bend is slightly convex. Therefore, if the rule of “suspect” 2 is applied, the certainty factor for each condition is 0.8 for the stroke length, 0.9 for the center of gravity of the stroke with the amount of movement of the X axis, 1.0 for the distance between the end point and the start point, and 0.9 for the bend. The center of gravity of the stroke is 0 when the axis moves.
9, the distance between the end point and the start point is 1.0, and the degree of bend is 0.9. Therefore, the logical product of the conditions, that is, the certainty factor for the conditional expression is 0.8, which is the smallest of these.

There are multiple production rules that lead to the same conclusion. In general, in fuzzy reasoning, the conclusion is also a fuzzy value, and the fuzzy value of each conclusion is corrected by the certainty factor obtained by the conditional statement, and if there are multiple things that lead to the same conclusion, the average is calculated. It is taken. However, in this embodiment, the conclusion is not a fuzzy value but a value of 0 or 1. Therefore, the certainty factor for the conclusion is the certainty factor of the conditional sentence. Also, if there are multiple identical conclusions,
Average confidence for each conclusion.

As an example of the above, it is assumed that the characters shown in FIG. 16 (A) have been input. If the fuzzified data corresponding to this is shown in FIG. 16 (B), the following certainty factors are obtained from the production rule for each stroke with the standard character "suspect".

The certainty factor for the first stroke is 1.0, the certainty factor for the second stroke is 0.8, and the certainty factors from the third stroke onward to the 14th stroke are 0.8, 0.8, 0.7, 0.8, 0.7, 0.9,
It becomes 0.9,0.8,0.8,0.8,0.9,0.9.

Therefore, since the average of these certainty factors is 0.83, the certainty factor of the standard character “suspect” for this input character is 0.83. The fuzzy inference unit applies the production rule to the input character for all the standard patterns having the same number of strokes as the input character, and calculates the certainty factor for each standard pattern of the input character. Then, the standard pattern having the highest certainty is output to the recognized character output unit 9 as a recognized character corresponding to the input character.

For example, if the characters in FIG. 16 (A) are input, the confidence level is 0.83 for the standard pattern “suspect” and the confidence level is 0.6 for “read”.
9, Confidence 0.66 against "wrong", Confidence against "theory"
A value such as 0.65 and a certainty factor of 0.65 for “acceptance” is obtained. Therefore, the input character is determined to be "suspicious" and is output to the recognized character output unit.

The characters inferred and concluded as described above are output from the recognized character output unit 9 as pattern signals corresponding to standard characters.

FIG. 17 is a character recognition process diagram for recognizing a handwritten character on the media graph 1 by the handwritten character recognition device by fuzzy inference configured as described above.

As shown in the figure, step 1 (hereinafter, S1, S2, S
3, ... S7. ), The X and Y coordinates indicating the position of the brush are input to the personal computer 2 as point sequence data at predetermined time intervals in accordance with the stroke order of the characters handwritten on the media graph 1. When the point sequence data corresponding to the handwritten character is input to the personal computer 2 in S2, the size of the input character is unified and the writing speed of the input character is normalized to be constant. In S3, a Fourier transform is performed on the amount of movement of the X coordinate and the amount of movement of the Y coordinate of each stroke of the normalized handwritten character, and then, in S4, each stroke of the normalized handwritten character. Each of the Fourier series a0 / 2, a1, a2 obtained by the respective Fourier transform with respect to the movement amount of the X coordinate and the movement amount of the Y coordinate is fuzzy.

In S5, the standard pattern portion is searched for fuzzified data of standard characters having the same number of strokes as the number of strokes of handwritten input characters, and a production rule is generated for each stroke based on the searched fuzzified data. In S6, based on the production rule, fuzzy inference is performed between the handwritten input character and the standard character retrieved from the standard pattern portion, and the highest confidence factor is determined, and then the highest confidence factor is determined in S7. The standard character determined to be high is output as a recognized character, and then the process proceeds to the next character recognition process.

(Effect of the Invention) As described above, according to the present invention, a character handwritten in the character input means is input as point sequence data corresponding to the X and Y coordinates, and the input data is Fourier-transformed in stroke units, The fuzzy value is used to express the coefficient up to the second frequency, and fuzzy inference is performed according to the production rules obtained from the standard pattern to recognize handwritten characters, so the amount of calculation is extremely small compared to conventional handwritten character recognition means. As a result, the processing time for recognizing handwritten characters can be shortened, and the recognition certainty factor of handwritten characters can be increased.

[Brief description of drawings]

The drawings relate to the embodiment of the present invention, FIG. 1 is a block diagram of a system configuration for recognition of handwritten characters, and FIG. 2A is a movement amount on the X axis and a Y axis for the character “NO”. Explanatory diagram showing the amount of movement, FIG. 2 (B) is a waveform diagram in which the waveform is folded back in line symmetry at the end point of the waveform in FIG. 2 (A), FIG. 3 (A), and FIG.
Drawing (A), Drawing 5 (A), and Drawing 6 (A), respectively, a stroke drawing which represented a typical stroke on coordinates, Drawing 3 (B), Drawing 4 (B), and Drawing 5 FIG. 6B and FIG. 6B are movement amount explanatory diagrams showing the movement amount of each stroke on the X-axis, FIG. 3C, FIG. 4C, and FIG.
FIGS. 6C and 6C are display diagrams showing respective coefficient values when the Fourier transform is performed on the movement amount on the X axis, and FIG. 7 is a fuzzy value regarding a stroke length of a handwritten character. And the correspondence diagram showing the corresponding values, FIG. 8 is a correspondence diagram showing the fuzzy values regarding the frequency coefficient a0 / 2 and the corresponding values, and FIG. 9 is the fuzzy value relating to the frequency coefficients a1 and their correspondence Correspondence diagram showing values, Fig. 10 is a fuzzy value related to frequency coefficient a2, and correspondence diagram showing the corresponding values. Fig. 11 is the stroke length and frequency intensity for all 14 educational Kanji characters. Fig. 12 is a membership function diagram, Fig. 13 (A) is the standard character "suspect" pattern diagram, and Fig. 13 (B) is the standard character "suspect" fuzzy data display diagram. , Fig. 14 shows the two membership functions. Memberships Function Figure, first
FIG. 5 is a membership function diagram with high certainty selected from the two membership functions shown in FIG.
FIG. 6 (A) is a pattern diagram of the input character “suspect”, FIG. 16 (B) is a fuzzy data display diagram of the input character “suspect”, and FIG. 17 is a character recognition process diagram. 1 …… Medea graph 2 …… Personal computer 3 …… Input data normalization unit 4 …… Fourier transform unit 5 …… Fuzzification unit 6 …… Standard pattern unit 7 …… Rule generation unit 8 …… Fuzzy inference unit 9 …… Recognition character output section

Claims (1)

[Claims] 1. In the process of handwriting a character in a character input means having a handwriting surface for handwriting a character, the same character is input as point sequence data corresponding to X and Y coordinates at predetermined time intervals. The point sequence data corresponding to the handwritten character output from the character input unit is input to the input data normalization unit, and the size of the character handwritten by the character input unit in the input data normalization unit. In order to make the writing speed of the input handwritten characters constant, after normalizing the intervals of the point sequences of the handwritten characters to be constant, the handwriting normalized by the input data normalization means Characters are X, Y stroke by stroke in Fourier transform means
The Fourier transform corresponding to the movement amount is performed, the intensity of the frequency is obtained, and the Fourier series data obtained by the Fourier transform in the Fourier transform means is fuzzy so that it can be treated as ambiguous handwritten character data. After the conversion, the rule generation means obtains the pattern data from the standard pattern storage means for storing the pattern data in which the standard characters are fuzzy, and generates the production rule for handwritten character recognition,
Thereafter, the fuzzy inference means performs the production between the pattern data stored in the standard pattern storage means having the same number of strokes as the handwritten characters input in the character input means and the fuzzy handwritten character data. Perform fuzzy inference based on rules, determine the one with the highest certainty, and recognize the character with the highest certainty determined by the fuzzy inference means as standard character data corresponding to the input handwritten character. A method for recognizing handwritten characters by fuzzy inference, which is characterized by outputting from an output means.