JPH05324925A - Fuzzy pattern recognizing device - Google Patents

Abstract

PURPOSE:To provide the fuzzy pattern recognizing device which can exactly recognize patterns even when a pattern formed with a line thicker than that of an ideal input pattern is inputted. CONSTITUTION:Concerning the pattern of an inputted recognizing object, a first processor 3 calculates the confidence degree of recognition by collating it with a pattern for collation. A second processor 9 calculates the total sum of the lengths of equal goodness-of-fit lines maximizing the goodness-of-fit of a three-dimensional membership function constituting the pattern for collation, decides the value of a parameter for deciding the fuzzy entropy of the three- dimensional membership function corresponding to the calculated value and prepares the pattern for collation corresponding to the value of the parameter.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a fuzzy pattern recognition apparatus having a function of accurately recognizing a pattern transformed from a basic pattern, such as a handwritten character for a typeface, by using a fuzzy theory.

[0002]

2. Description of the Related Art A conventional pattern recognition device generally recognizes a pattern whose shape is strictly defined.
A method using a neural network is known as a means for recognizing an ambiguous pattern that is not limited to a certain shape such as handwritten characters. In this, a neural network is formed by learning a pattern to be recognized in advance, and the input pattern is recognized thereby.

However, the method using this neural network requires a work of learning, and requires re-learning each time the number of pattern candidates to be identified increases. In addition, if the neural network is a black box and changes in the characteristics of the candidate of the pattern to be identified due to secular change of the input device that takes in the pattern signal of the recognition target, it is reflected in the operation of the device. It is difficult to do so, and there is a problem that erroneous recognition may occur.

On the other hand, in recent years, a method of recognizing an ambiguous pattern using a fuzzy theory has been proposed. This method is based on having a plurality of knowledge about the characteristics of the pattern and checking them one by one, such as "there is a vertical line near this portion". However, in the conventional pattern recognition method based on the fuzzy theory, it is necessary to acquire knowledge about pattern features (rule creation work) when developing a system including software, and pattern feature extraction is performed when actual pattern recognition is performed. An algorithm is needed. for that reason,
There is a problem that processing is complicated because many rules and membership functions of fuzzy sets are required.

[0005]

Therefore, as a solution to the problem of pattern recognition based on the conventional fuzzy theory, the inventor combines a preset fuzzy set membership function to collate it with a recognition target. We have proposed a pattern recognition method that prepares a basic pattern that should be used and compares it with the pattern to be recognized to calculate the certainty factor, which is an index for evaluating the pattern ( Japanese Patent Application No. 2-274355
issue). According to this method, a pattern to be collated with a recognition target is formed by combining membership functions indicating a fuzzy region having a boundary of an arbitrary shape (for example, an ellipse). It is possible to efficiently recognize patterns such as characters and figures without needing to do so. The matching pattern formed by combining the membership functions by this method is called a "fuzzy template", and this "fuzzy template" is also used in this specification.

In the above method, the fuzzy entropy of the three-dimensional membership function constituting the fuzzy template is fixed to a fixed value or an arbitrary setting method is set individually for each three-dimensional membership function. Has been adopted.

For example, even if a combination of only a three-dimensional membership function represented by the following equation (1), in which the length of the portion having the maximum fitness value (fitness 1) is unspecified, fuzzy There is an advantage that the template can be configured, that is, there is no mathematical constraint condition in the setting of the fuzzy template and the degree of freedom is high.

[0008] f = 2 exp [- {( x-A) 2 / a x 2 + (y-B) 2 / a y 2}] ... (1) In the above equation, a _x, a _y respectively x, It is a parameter proportional to the fuzzy entropy in the y direction.

Here, taking the number "8" as an example, the fuzzy template is expressed as follows.

An extension of the range of goodness of fit of a three-dimensional elliptic membership function with a goodness of fit of a circle having a center point of (6, 15) and a radius of 4.5 to [-1, 1] (the upper part of "8") Circle)

[0011]

[Equation 1]

When the range of the goodness of fit of a three-dimensional elliptic membership function having a goodness of fit of a circle having a center point of (6, 6) and a radius of 4.5 is [-1, 1] (of "8", Bottom circle)

[0013]

[Equation 2]

Here, the reason for expanding the range of the fitness to [-1, 1] is to give an evaluation value of- (minus) to an inappropriate input coordinate value. For example, a black image signal This is because, when is input, the total sum of the degree of conformity described later becomes a small value.

The fuzzy template of "8" is expressed by the following function created by the combination with the above.

F = max (f ₁ , f ₂ ) (4) With respect to the above equation, as shown in FIG. 2, 12 horizontal dots × 21 vertical dots are set on the xy coordinate plane. That is,
An area including the character "8" is divided into 12 areas in the x-axis direction and 21 areas in the y-axis direction, and each area (dot) is converted into a binary signal corresponding to the character shape. Here, the blank area is converted into a binary signal “0”, and the black area is converted into a binary signal “1”. As the size correction, the second black areas from the upper, lower, left, and right ends are placed in the second column from the upper, lower, left, and right ends, respectively.

At this time, assuming that the pattern shown in FIG. 8 is input, n = 42 out of 12 × 21 = 252 dots.
The binary signal “1” (black) is given to each of the two.

Therefore, by using the membership function f of the above equation (4), the sum T ₁ of the fitness of these n black dots and the average fitness S ₁ = T ₁ / n for each dot are obtained. .. In the example of FIG. 8, a total of goodness of fit T ₁ = 36.05 calculated by the coordinate values of n = 42 black dots and an average goodness of fit S ₁
＝ 36.05 / 42 ＝ 0.858.

Similarly, the sum T ₂ of the goodness of fit and the average goodness of fit S ₂ when the most ideal “8” pattern is input are obtained. In this case, the sum T of the suitability of 44 black dots
₂ = 43.59 and the average goodness of fit S ₂ = 43.59 / 44 = 0.991.

From the above results, the certainty factor is calculated by taking the weighted average of the ratio T ₁ / T ₂ of the total sum of the goodness of fit (total goodness of fit) and the ratio S ₁ / S ₂ of the average goodness of fit. in this case,
The confidence is as follows.

[0021]

[Equation 3]

The ratio T ₁ / T ₂ of the total goodness of fit exceeds 1 in some cases, but in that case, it is unconditionally set to 1. That is, min (T ₁ / T ₂ , 1).

Here, the fuzzy template is displayed as follows. That is, as shown in FIG. 2, a three-dimensional membership function base set plane (a coordinate plane formed by two coordinate axes other than the coordinate axis indicating the goodness of fit) is divided into a finite number of dots, and the center point of each dot is When the goodness of fit of the three-dimensional membership function obtained by the coordinate values is obtained,
A black circle with a radius proportional to the size of the goodness of fit for each dot is displayed in each dot. As a result, like the fuzzy template shown in FIG. 8, a place with a high degree of conformity is displayed in black, and a place with a low degree of conformity is displayed in white.

Next, as an example of pattern recognition, consider a case where one pattern "8" completely includes the other pattern "3" like the numbers "3" and "8". As shown in FIG. 9 (a), when an input pattern of "3" having the same thickness as the supposed ideal character of "3" is given, the fuzzy template of "8" shown in FIG. 9 (a). The certainty factor by collation with (39.5 / 51.3) × 0.5 + (0.98 / 0.98) × 0.5 = 0.89 is (39.5 / 39.5) ) × 0.5 + (0.98 / 0.98) × 0.5 = 1.0. That is, the certainty factor is higher in "3" than in "8", and a correct recognition result is obtained. This is due to the difference in the ratio of total goodness of fit.

However, when "3" written in a thick line is input as shown in FIG. 10 (a), the total matching degree of the input pattern becomes a large value, and it is collated with the fuzzy template of "8". Then, the ratio of total goodness of fit: min (63.3 / 51.3, 1) = 1.0 The ratio of average goodness of fit: min (0.90 / 0.98, 1) = 0.92, so the confidence in this case is 1.0 × 0.5 + 0.92 × 0.5 = 0.96, and in the collation with the fuzzy template of "3" in Fig. 10 (b), the ratio of total goodness of fit: min (63.3 / 39.5, 1) = 1.0 Ratio of goodness of fit: min (0.90 / 0.98) , 1) = 0.92, the confidence factor in this case is also 1.0 × 0.5 + 0.92 × 0.5 = 0.96. In other words, both “3” and “8” have the same certainty factor, which makes them indistinguishable.

This is because the above-mentioned fixed entropy method cannot deal with the change in the thickness of the character or number to be recognized because the fuzzy entropy has a constant value. On the other hand, in the arbitrary setting method, there arises a problem of how to set the parameters of the membership function according to changes in the thickness of letters and numbers.

Therefore, an object of the present invention is to accurately perform matching with a prepared fuzzy template even when a pattern formed by a line thicker than the ideal input pattern referred to for fuzzy template matching is inputted. An object of the present invention is to provide a fuzzy pattern recognition device capable of various pattern recognition.

[0028]

According to the present invention, in a device for pattern recognition based on a membership function representing a fuzzy set, a pattern input section for inputting a pattern to be recognized and one or more membership functions are generated. Then, by combining the membership functions, a matching pattern forming means for forming a matching pattern for recognizing a recognition target pattern input from the pattern input section, and a recognition target input from the pattern input section. Of the three-dimensional members forming the collation pattern, the arithmetic processing means calculating the certainty factor of recognition by comparing the pattern of No. 3 with the collation pattern formed by the collation pattern forming means. The total sum of the lengths of the equi-fitness lines having the maximum goodness of fit of the ship function is obtained, and the matching pattern is obtained according to the value. Determines the value of the parameter that determines the fuzzy entropy of all three-dimensional membership functions constituting down, characterized by creating a collation pattern on the value of the parameter.

[0029]

In the present invention, the total sum of the lengths of the equi-fitness lines having the maximum goodness of fit (fitness 1) of all three-dimensional membership functions constituting the matching pattern is obtained, and according to the value, The values of the parameters that determine the fuzzy entropy of all three-dimensional membership functions that make up the matching pattern (fuzzy template) are automatically determined for each fuzzy template. And the fuzzy template is created with the value of that parameter,
Even if the line thickness of the input pattern changes, the deterioration of the recognition accuracy can be compensated.

[0030]

1 is a block diagram showing an embodiment of a fuzzy pattern recognition apparatus according to the present invention. This device
It is configured to recognize characters and consists of the following components.

The image reading device 1 includes a light source for illuminating a recognition target in order to read an image of a typeface or handwritten characters,
A photoelectric converter that reads an image by reflected light of the image and converts the image into an electric signal is included. Alternatively, a pressure-sensitive handwritten character input device may be used.

The signal converter 2 converts the image signal read by the image reading device 1 into a binary signal of, for example, "0" (= "white") or "1" (= "black"), and , Is configured to correct the size of the read image. This correction (scaling) extends the shape up and down, for example, when the character has a shape crushed up and down.

The image reading apparatus 1 and the signal conversion section 2 constitute a pattern input section for inputting a pattern to be recognized.

The first processor 3 recognizes the read image (input pattern) which has been binarized and whose shape has been corrected by the signal converter 2. That is, when a pattern to be recognized is input, this is compared with a collation pattern (fuzzy template) stored in a fuzzy template storage unit 6 described later (collation) to calculate the recognition certainty factor of the input pattern. To do. Further, the pattern is classified into the pattern with the highest certainty as compared with other patterns with ambiguous boundaries, and the recognition result is output.

The pattern indicating the recognition result by the processor 3 is converted from a binarized signal to an analog signal by the signal conversion unit 4 and output by the output device 5 such as a CRT.

The fuzzy template storage unit 6 stores a fuzzy template serving as a reference for recognizing an input pattern. The fuzzy template parameters stored in the storage unit 6 can be manually adjusted and changed via a CRT used as a terminal device of a computer and a man-machine interface 7 such as a keyboard.

The three-dimensional membership function generator 8 generates a membership function required to form a fuzzy template.

The fuzzy template storage unit 6, the man-machine interface 7 and the three-dimensional membership function generator 8 constitute a collation pattern forming means.

When the operator creates a fuzzy template via the man-machine interface 7, the second processor 9 determines the fuzzy entropy according to the contour-linked entropy determination algorithm stored in the storage unit 10 described later. And sends it to the fuzzy template storage unit 6.

The contour interlocking entropy determination algorithm storage unit 10 is an algorithm for determining the fuzzy entropy according to the sum of the lengths of the equi-fitness lines having the maximum goodness of fit (goodness of fit 1) when the fuzzy template is created. Is stored.

In the above embodiment, the two processors 3 and 9 are used as the arithmetic processing means, but in the actual device (hardware), one CP is used.
U (main processor) can realize the functions of these two processors.

Next, the contour-linked entropy determination algorithm, which is a feature of the embodiment, will be described.

First, the fuzzy template is constructed 3
Let L be the sum of the lengths of the equi-fitness lines having the highest fitness of the dimension membership function (fitness 1). At this time, the parameter a _F that determines the fuzzy entropy is calculated by the following equation.

A _F = A−L / B (A, B: arbitrary constants) As an example, let A = 4, B = 50, that is, a _F = 4-L / 50. At this time, the length L ₈ of the equi-fitness line having the maximum goodness of fit (fitness 1) of the fuzzy template of “8” shown in FIG. 3A is calculated as L ₈ = 56.55, which is the length L ₃ of the equi-fitness line of the best fit (fitness 1) of the fuzzy template “3” shown in FIG. 3B.
Similarly, L ₃ = 42.41 is obtained by obtaining the length of the circumference. Therefore, the parameters a _F8 and a _F3 for determining the respective fuzzy entropies are a _F8 = 4-56.55 / 50≅2.87 a _F3 = 4-42.41 / 50≅3.15.

As a result, the shorter the contour, the larger the value of the fuzzy entropy, and the thicker the input pattern of "3" (FIGS. 10A and 3).
In comparison with the fuzzy template of "8", the ratio of total goodness of fit: min (63.3 / 51.3, 1) = 1.0 Ratio of goodness of fit: min (0.90 / 0.98, 1) = 0.92, the certainty factor is 1.0 × 0.5 + 0.92 × 0.5 = 0.96, but in the matching with the fuzzy template of "3" in Fig. 3 (b), the ratio of total goodness of fit: min (65.8 / 39.5, 1) = 1.0 The ratio of average goodness of fit: min (0.94 / 0.98, 1) = 0.96, so the certainty factor is 1.0 x 0.5 + 0.96 x 0.5 = 0.98.

That is, since a high matching degree is obtained on average in the collation with the fuzzy template of "3", the certainty factor of "3" becomes a larger value than the certainty factor of "8", which is correct. Can recognize patterns.

FIG. 4 shows a pattern to be recognized as the numeral "2" as another input pattern example. This is an example of a thick pattern represented by a closed contour with a blank in the middle. In the embodiment of the present invention, such white characters can be recognized.

5 to 7 show examples of patterns to be recognized as the numeral "3". In the fixed entropy method described above, it is impossible to distinguish between these input patterns from "3" and "8", but according to the present invention, "3" is recognized.

[0049]

As described above, according to the present invention, the fuzzy entropy of the three-dimensional membership function is determined according to the length of the equi-fitness line having the maximum fit of the fuzzy template. The input pattern can be accurately recognized without any limitation on the line thickness of the pattern to be recognized.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a fuzzy pattern recognition device according to the present invention.

FIG. 2 is a diagram showing coordinates of an input image.

FIG. 3 is a diagram showing a pattern matching result according to the present invention when an input pattern is thick.

FIG. 4 is a diagram showing an example of a thick input pattern of a numeral “2”.

FIG. 5 is a diagram showing an example of a thick input pattern of the numeral “3”.

FIG. 6 is a diagram showing another input pattern example of the numeral “3”.

FIG. 7 is a diagram showing another input pattern example of the numeral “3”.

FIG. 8 is a diagram showing an example of collation using a fuzzy template corresponding to an input pattern.

FIG. 9 is a diagram showing a pattern matching result when the input pattern is normal.

FIG. 10 is a diagram showing a pattern matching result by a conventional method when an input pattern is thick.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Image reading device, 2 ... Signal conversion part, 3 ... 1st processor, 4 ... Signal conversion part, 5 ... Output device, 6 ... Fuzzy template storage part, 7 ... Man-machine interface, 8
A three-dimensional membership function generator, a second processor, a contour interlocking entropy determination algorithm storage unit.

Claims (1)

[Claims] 1. An apparatus for pattern recognition based on a membership function representing a fuzzy set, a pattern input section for inputting a pattern to be recognized, and one or more of the membership functions are generated and the membership functions are combined. As a result, a matching pattern forming unit that forms a matching pattern for recognizing a recognition target pattern input from the pattern input unit, and a recognition target pattern input from the pattern input unit by the matching pattern formation unit And a calculation processing unit that calculates a certainty factor of recognition by comparing with the matching pattern formed by the above-mentioned. Obtain the sum of the lengths of the maximum equal-fitness lines, and according to the value, all 3
A fuzzy pattern recognition device characterized in that a value of a parameter that determines fuzzy entropy of a dimension membership function is determined and a matching pattern is created based on the value of the parameter.