JPH0981536A - Pattern recognition method and apparatus - Google Patents

Abstract

(57) [Abstract] [Purpose] To provide a pattern recognition method and device for directly extracting fuzzy rules from teacher data and realizing pattern recognition with high generalization ability. [Structure] Using the input / output data file 107, the data dividing unit 104 divides the input data into clusters,
Input / output data file 10 by rule extraction means 105
7 and the divided data file 108, a fuzzy rule is extracted for each cluster, the fuzzy rule is adjusted by the rule adjusting means 106, and the fuzzy rule inferring means 102 performs pattern recognition using the fuzzy rule file 109.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition method and apparatus, and more particularly to a pattern recognition method and apparatus suitable for performing pattern recognition by fuzzy inference.

[0002]

2. Description of the Related Art A neural network having a multilayer structure or fuzzy inference is used to construct a pattern recognition system for recognizing a character by inputting a feature amount of a character or diagnosing an abnormality of a device by inputting a state of a device. It is said to be suitable. Neural net is M. Eye. Tea Press (1986) Parallel Distributed Processing, pages 318-362 (Parallel Distributed
Processing, Vol. 1, MITPress, CambridgeMA, 1986,
As discussed in (pp. 318-362), a neural network can be constructed by learning using teacher data that is a set of input and output data, so there is no need to develop a problem-specific algorithm. Have an advantage

On the other hand, the use of fuzzy reasoning, which is based on fuzzy rules, has an advantage that a pattern recognition system can be easily constructed in a small-scale field in which expert knowledge is abundant. As a method of solving the problem that it is difficult to acquire a fuzzy rule by fuzzy reasoning, some methods have been proposed not by an expert but by a teacher data like a neural network. For example, Eye Triple E Transaction on Fuzzy Systems Vol. 3, No. 1, pp. 18 to 28 (IEEE
Trans. Fuzzy Systems, Vol. 3, No. 1, pp. 18-28,
In 1995), a method of approximating the existence region of data with a super-cuboid using teacher data is adopted. In this method, the minimum and maximum values of the teacher data belonging to each class are calculated to define a hyper-rectangle, and if there is an overlap in the hyper-rectangles between classes, the overlapped part is defined as a prohibited area, and the prohibited area is further defined. If there are data of these classes inside, fuzzy rules are extracted by defining hypercubes in the same way.

[0004]

When a neural network is used, the learning time is extremely slow, and there is a problem that the mechanism of the completed neural network is unknown. In addition, if the method of extracting fuzzy rules by the above-mentioned hypercube is adopted, learning is extremely fast and its movement can be analyzed by fuzzy rules, but the teaching data for extracting fuzzy rules and the performance of the obtained fuzzy rules are When the characteristics of the test data to be tested greatly differ, there is a problem that the performance of the test data, that is, the generalization ability is inferior to that of the neural network.

An object of the present invention is to solve the above-mentioned problems of neural networks and fuzzy inference, to directly extract fuzzy rules from teacher data consisting of a set of input data and output data, and to perform pattern recognition with high generalization ability. It is to provide a method and an apparatus capable of realizing

[0006]

According to the present invention, (a) input data of each class is divided (b) fuzzy rules are extracted by approximating a region using the divided data (c) teacher data It has a feature of tuning a fuzzy rule by adjusting the approximate region by using, and not only these independent means and methods, but also various combinations have respective features.

An example of means for solving the problem will be described below.

(B) Dividing the input data of each class Means for dividing the teacher data belonging to the same class into several clusters is provided according to the distribution of the data in the input space.

(B) The fuzzy rule is extracted by approximating the region using the divided data. The region where the teacher data included in each cluster is approximated,
A means for defining the approximated area as one fuzzy rule is provided.

(C) Tuning the fuzzy rule by adjusting the approximation region using the teacher data The approximation region of the fuzzy rule obtained in (b) is approximated so that the degree to which it is included in the class to which the rule belongs is large. Means are provided for adjusting the area.

[0011]

[Action]

(A) Means for dividing the input data of each class By dividing the teacher data of each class, the existence area of each class can be approximated more accurately.

(B) Means for extracting fuzzy rules By defining fuzzy rules corresponding to each cluster, the degree to which the input data belongs to the class corresponding to the cluster, that is, the degree of success can be calculated, and the fuzzy rules are satisfied. Pattern recognition can be performed by determining that the input data belongs to the most frequent class.

(C) Fuzzy Rule Adjusting Means By adjusting the approximate region so that the pattern recognition for the teacher data can be accurately performed, the recognition rate for not only the teacher data but also the test data can be improved.

[0014]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows a specific embodiment of the present invention. In the figure, 101 is fuzzy rule construction means, 102 is fuzzy rule inference means, 103 is storage means, 104 is data division means, 105 is rule extraction means, 106 is rule adjustment means, 107 is input / output data file, and 108 is A divided data file, and 109 is a fuzzy rule file. The input / output data file 107 stores teacher data composed of a set of input data and output data collected from a pattern recognition target. Data dividing means 10
Reference numeral 4 divides the data of the input / output data file 107 and stores the division result in the divided data file 108. The rule extraction means 105 extracts a fuzzy rule for each divided cluster using the input / output data file 107 and the divided data file 108, and the fuzzy rule file 10 is extracted.
9 is stored. The rule adjusting means 106 uses the input / output data file 107, the divided data file 108, and the fuzzy rule file 109 to adjust the fuzzy rule and stores the result in the fuzzy rule file 109. The obtained fuzzy rule is used by the fuzzy rule inference means 102 to infer an output for an unknown input. Here, the fuzzy rule construction means 101 and the fuzzy rule inference means 102 may be combined as shown in FIG. 1, but they may be separated and the fuzzy rule construction means 101 may extract the fuzzy rules offline and incorporate the obtained fuzzy rules. The fuzzy rule inference means 102 may form a pattern recognition system.

FIG. 2 shows a hardware configuration for implementing the present invention. In the figure, 10 is a central processing unit (CPU), 20
Is a memory, 30 is a keyboard, and 40 is a CRT terminal. A program for executing the data files 107, 108, 109 and the fuzzy rule construction means 101 and the fuzzy rule inference means 102 is stored in the memory 20, and is started by the user's instruction from the CRT terminal 40 and the fuzzy rule construction means 101 executes the fuzzy logic. Rule extraction and pattern recognition are performed by the fuzzy rule inference means 102. Note that the fuzzy rule inference means 102 may have a configuration incorporated in the system instead of the stand-alone configuration shown in FIG.

The structure of the fuzzy inference system will be shown before the method of realizing the data dividing unit 104, the rule extracting unit 105, and the rule adjusting unit 106 is described in order. Consider separating an m-dimensional input vector x into n classes. At this time, it is assumed that the class i (i = 1, ..., N) is divided into some clusters ij. Here, it is assumed that the cluster ij is the j-th cluster of class i.
The structure of the input / output data data file 107 at this time is shown in FIG. In the figure, 107-1, 107-2, ...
107-m corresponds to an m-dimensional input consisting of M, and 107-m
-O corresponds to a one-dimensional output consisting of M pieces. 107-1,
The value of the i-th element of 107-2, ..., 107-m indicates the m-th i-th input data, and the class to which this input data belongs is defined in the i-th element of 107-o. There is.
At this time, assuming that the number of classes is n, 102 is any number from 1 to n. When the output data 102 belongs to class i as n dimensions instead of one dimension, the i-th data may be 1 and the rest may be 0.

FIG. 4 shows the structure of the divided data file 108. The divided data file 108 is composed of the same M pieces of data as the input / output data file 107, and the pth data holds the cluster number of the pth data of the input / output data file 107. That is, p-th input data 1
I of cluster number ij corresponding to 07-1, ..., 107-m
And j are p-th output data files 107-o and p, respectively.
It can be obtained from the second divided data file 108.

At this time, it is assumed that the following rule Rij is defined for each cluster ij. However, c _ij is the center of cluster _ij .

[0019]

[Equation 1]

The membership function m _ij (x) for the input x in equation (1) is given by equations (2) to (4). Where _dij (x)
Is the weighted distance between x and c _ij , h _ij (x) is the adjusted distance,
α _ij (> 0) is the adjustment parameter of cluster ij, c _ij = (c _ij , ₁ ,
..., c _ij , _m ) where r _ij , k is the k-th radius of cluster _ij .

[0021]

[Equation 2]

[0022]

(Equation 3)

[0023]

(Equation 4)

FIG. 5 shows the file structure of the fuzzy rules R _ij in the fuzzy rule file 109. 109
−1 is m center coordinates c _ij , ₁ , ..., C _ij , m, and 109−2
Indicates a radius r _ij , ₁ , ..., R _ij , _m, and 109-3 indicates an adjustment parameter a _ij .

The configuration of the fuzzy system at this time is shown in FIG.
Shown in In the figure, 204, 205, and 206 are equations (1).
The fuzzy rules R ₁₁ , R _ij , and Rn ₁ indicated by
02 and 203 are inputs to the fuzzy rule R ₁₁ from inputs x ₁ , x ₂ and x _m , respectively.

The data of FIG. 5 is stored for the fuzzy rule R _ij 205, and using the input x and this data, equations (2) to (4) are calculated to calculate the value of the membership function. . At this time, when the membership function m _kl (x) is maximum, the input x is determined to belong to the class k. The exponential function in Expression (2) is for setting the output range of Expression (2) to [0, 1]. Therefore, when the input x is classified by the input of the exponential function of the equation (2), the minimum value of h _ij (x) should be obtained. This configuration is the simplest possible configuration. Each class must be divided into clusters in order to achieve a performance comparable to that of a neural network. For each cluster, the center c _ij and the radius r _{ij, k} can be estimated using the data belonging to that cluster. Then, in order to improve the generalization ability, the recognition rate of the teacher data is increased by allowing the correctly recognized data to be erroneously recognized and increasing or decreasing α _ij .

(B) Division of input data of each class A net having a linear combination stage at the output of FIG. 6 is called a radial basis function net. When comparing a radial basis function net and a multilayer neural network, a multilayer neural network is compared. Is said to be slower to learn but has higher generalization ability. This is an exponential function called a sigmoid function in a multilayer net, which is a global function whose output changes over the entire input space, whereas a radial basis function net uses a Gaussian function and is a local function. Because of the function. Therefore, in order to obtain the generalization ability equivalent to that of the multilayer neural network, the input area covered by the Gaussian function needs to be as large as possible. Therefore, when dividing data belonging to the same class into clusters, it is necessary to prevent the cluster data from becoming extremely small.

Although some clustering methods have been proposed, most of them use an iterative method, and most of them do not have a method of adjusting the number of data belonging to a cluster. Therefore, when clustering is performed by these methods, the number of data in a certain cluster is extremely small, and the data in a certain cluster may be extremely large. If the parameters of equation (2) are determined using a small amount of data, high generalization ability cannot be expected. Therefore, when the data is divided on each input axis so that the number of data contained in each cluster is about the same, the divided data is divided into two.
Try to divide by the axis with the least number of data in one cluster. That is, 1) Let N _{c be the} maximum allowable value of the number of data belonging to each cluster. Initially, each class consists of one cluster.

2) A cluster whose number of data belonging to the cluster exceeds N _c is defined as a cluster _ij , and the center c _ij is calculated by using the data included in the cluster ij as shown in equation (5). However, N _ij is the number of data included in cluster _ij . The data belonging to the cluster _ij is divided by x _k = c _{ij, k} . Where N _{ij, 1} and N _{ij, 2} are _defined as x _k ≧ c
Let _{ij, k} and x _k <c _{ij, k be} the number of data _items . At this time, k that minimizes | N _{ij, 1} −N _{ij, 2} | is the axis of division.

3) Repeat step 2) until there are no clusters whose number of data exceeds N _c .

For example, in FIG. 7, the input is two-dimensional and 210
Let ~ 215 be 6 pieces of teacher data. Then x ₂ = c
_When divided by _{ij, 2} , the number of data is divided into two and four clusters, but when divided by x ₁ = c _{ij, 1} , the number of data becomes 3 respectively, so x ₁ = c _{ij, 1} Split with.

[0032]

(Equation 5)

(B) Extraction of fuzzy rule The fuzzy rule R _ij given by the equation (1) is defined for each cluster. At this stage, the regions are approximated without considering the overlapping of regions of different classes. Overlapping of these regions will be eliminated later when adjusting α _ij .

First, the center c _{ij of the} cluster _ij is calculated using the equation (5). Then, the radius r _{ij, k} is estimated by the equation (6). Where θ is a parameter and a _{ij, k} is the center c of x _k
The average displacement from _{ij, k} , σ _{ij, k} is the variance of a _{ij, k and} is given by equation (7),
It is given in (8).

[0035]

(Equation 6)

[0036]

(Equation 7)

[0037]

(Equation 8)

However, if there is a deviation in the distribution of the cluster data around the center, the obtained elliptical region cannot faithfully approximate the distribution of the teacher data. For example, assume that one-dimensional data is distributed as shown in the upper diagram of FIG. Numerals 250 to 253 are teacher data, and the numbers above them indicate coordinates. When θ = 0, the center calculated by the equation (5) is located at x = 0 based on the data distribution. Then, r _{ij, 1} = 3 is obtained from the equations (6) to (8). As can be seen from the figure, the ellipse is offset from the center toward the positive side.

In order to avoid this, the average radius is calculated depending on whether the data is on the positive side or the negative side of the center, the center is moved so as to eliminate the deviation, and the new center is used. Recalculate radius. That is, first, the expressions (9) and (1
0) is calculated. Where a (p) _{ij, k} is the average positive offset of x _k from c _{ij, k} , a (n) _{ij, k} is the average negative offset of x _k from c _{ij, k} , N (p) _ij belongs to cluster ij and x _k −c _{ij, k} > 0
N (n) _ij is the number of data that satisfy c
It is the number of data satisfying _{ij, k} −x _k > 0. Then, the center is corrected as shown in equation (11).

[0040]

[Equation 9]

[0041]

(Equation 10)

[0042]

[Equation 11]

Using equations (11), (7) and equation (8), the radius r _{ij, k} is recalculated.

(C) Adjustment of fuzzy rules The fuzzy rules extracted as described above are defined without considering overlapping between classes. Therefore, the fuzzy rules are set so that the recognition rate is maximized with respect to the teacher data. Need to be adjusted. However, if the center and the radius are adjusted, the steepest descent method must be used and it takes time. Therefore, one parameter α _ij of the fuzzy rule R _ij will be adjusted. If α _ij is increased, the formula
The degree of success of the membership function given in (2) increases, and the degree of success decreases when it decreases. To illustrate the idea of adjustment, consider the case where one class is defined in one of the two classes shown in FIG. 261 is class 1 data, 262 is class 2 data, 263 is a class 1 membership function, and 264 and 265 are modified class 1 membership functions. The data 261 belonging to class 1 is erroneously recognized as class 2. If α ₁₁ is increased so that the membership function 263 falls between 264 and 265, the data 261 can be correctly recognized without misrecognizing the data 262. This can also be achieved by reducing α ₂₁ .

FIG. 10 shows a more complex example. 271,
275 is data belonging to class 2, 272, 273,
274 is data belonging to class 1. Also, 276 is a class 2 membership function, and 277 and 278 are modified class 2 membership functions. Data 271 is correctly classified as class 2, but data 272, 273, 274 is incorrectly classified as class 2. If α ₁₁ is increased or α ₂₁ is decreased, the data 271 is erroneously recognized first, but if erroneous recognition of the data 271 is permitted, the data 272, 273, 274 can be correctly recognized. That is, when α ₂₁ is decreased so that the class 2 membership function 276 is placed between 277 and 278, the data 271 is erroneously recognized, but the data 272, 273, and 274 are correctly recognized. This improves the recognition rate by two data. As described above, when the recognition rate is improved for each fuzzy rule R _ij , data that is correctly recognized is allowed to be erroneously recognized, and the optimum correction amount of α _ij is determined. In this way, α _ij is corrected until the misrecognition rate cannot be improved. It can be said that the method of allowing the misrecognition to improve the recognition rate as described above prevents the problem from falling into a minimal solution. Of course, the adjustment process is non-linear, and there is no guarantee that an optimal solution will always be obtained by this method.

[0046] determining the upper limit U _ij (l) and the lower limit L _ij (l) of the alpha _ij when allowing erroneous recognition of l-1 pieces of data limit was correctly recognized on the alpha _ij.
The input teacher data is a set X of data correctly classified by using the fuzzy rule {R _ij }, and a set Y of data that is misrecognized.
Split into. Then, with x (∈X) belonging to class i, equation (12)
Seeking something that satisfies. If the equation (12) is not satisfied, x is still correctly recognized even if α _ij is changed. If x further satisfies equation (13), there exists a lower limit L _ij (x) of equation (14) for correctly recognizing x.

[0047]

(Equation 12)

[0048]

(Equation 13)

[0049]

[Equation 14]

This will be supplemented with reference to FIG. 280,
281, 283 are membership functions of the clusters ij, ik, of respectively, and 282 is a modified membership function of the membership function 280. If the input x belongs to the class i, the membership function 280 is given as 280 and 28.
The input x is correctly recognized even if α _ij is decreased so that it falls between the two. On the other hand, if α _ij is further decreased and the slope of the membership function becomes larger than that of the membership function 282, the input x is erroneously recognized. If the equation (13) is not satisfied, that is, in the case of the equation (15), erroneous recognition does not occur in x even if α _ij is decreased.

[0051]

(Equation 15)

This will be described with reference to FIG. 290,
291, 292 are membership functions of the clusters ij, ik, of, respectively, and 293 is a modified membership function of the membership function 290. Assuming that the input x belongs to the class i, the degree of satisfaction of the membership function 291 at the input x is larger than that of the membership function 292.
The input x will not be misrecognized no matter how small α _ij is in the positive range.

From this, the lower limit L _ij (1) that does not erroneously recognize correctly recognized data is given by the equation (16). To simplify the discussion here, we assume that L _ij (x) is different for different x. At this time, one x is satisfied. Similarly, the lower limit that allows one correctly recognized data to be recognized incorrectly
L _ij (2) is given by equation (17) as the second minimum value in L _ij (x). Generally, the lower limit is given by equation (18).

[0054]

(Equation 16)

[0055]

[Equation 17]

[0056]

(Equation 18)

Similarly, the upper limit U _ij (l) can be obtained.
Select x (εX) belonging to class 0 (≠ i). If the cluster op has the minimum adjusted distance h _op (x), then equation (19) holds. The adjusted distance h _ij (x) is greater than h _op (x), so the upper limit of α _ij U _ij (x) in the range where x is recognized correctly
Is given by equation (20).

[0058]

[Equation 19]

[0059]

(Equation 20)

This will be described with reference to FIG. 301,
A membership function 302 is a membership function of the clusters of and ij. A membership function 300 is a membership function of a limit at which an input x belonging to a class o is correctly recognized. The input x is correctly recognized by the class o as long as the corresponding membership function is between 300 and 302 even if α _ij is changed.

From this, the upper limit U _ij (1) of α _ij that does not erroneously recognize correctly recognized data is given by equation (21). Here we assume that U _ij (x) is different for different x. From this, equation (21) is valid only for one x. Similarly, the upper limit of α _ij that allows misrecognition of one correctly recognized data
U _ij (2) is the second minimum value in U _ij (x) and is given by equation (22). Generally, it becomes as shown in Expression (23). From this α _ij
Is constrained as in equation (24).

[0062]

(Equation 21)

[0063]

(Equation 22)

[0064]

(Equation 23)

[0065]

(Equation 24)

Let (a, b) be an open section. If α _ij
When changed within the range of (L _ij (1), U _ij (1)), the correctly recognized data x (εX) remains correctly recognized. In addition, [a _ij ] is the closed interval and [U _ij (l−1), U
_ij (l)) or (L _ij (l), L _ij (l−1)] is varied, and misrecognition occurs in l−1 correctly recognized data x (εX).

[0067] For the improvement class i in the wrong data x of the recognition rate due to the change of α _ij (∈Y) or class i to belong to the class o (≠ i) the wrong data x (∈Y), by changing the α _ij Find out if it will be recognized correctly. First, consider the wrong data x (∈Y) that belongs to class i and belongs to class o (≠ i). If the equation (25) holds, this data is correctly recognized regardless of the value of h _ik (x) (k ≠ i). Where V _ij (x) is the lower limit of α _ij .

[0068]

(Equation 25)

This will be described with reference to FIG. 310,
311 is the membership function of each cluster ij, of
Reference numerals 12 and 313 indicate the range of the membership function that correctly recognizes the input x belonging to the class i. If α _ij is increased so that the membership function 310 falls between 312 and 313, the input x will be correctly recognized by the class i.

When α _ij is set to a value between the intervals (α _ij , U _ij (l)), Inc (l) is the number of erroneously recognized data that is correctly recognized. If V _ij (x) is included in the section (α _ij , U _ij (l)), Inc (l) is incremented by one. further,
It is defined as in Expression (26). α _ij to mαx (β _ij (l), U
_If it is set larger than _ij (l-1)), l-1 data are erroneously recognized, but Inc (l) data that was erroneously recognized is correctly recognized.

[0071]

(Equation 26)

Similarly, data x belonging to class o is assigned to class i
Suppose it was misrecognized by. Check if x is correctly recognized by reducing α _ij . To do so, the minimum adjusted distance of class o needs to be the second minimum among the n classes. That is, q in equation (27) needs to be o. Next, h _ij (x) is the smallest in class i, and the second smallest value in class i must be larger than the smallest adjusted distance in class o, as in equation (28). At this time, the data is correctly recognized if the equation (29) is satisfied. Where K
_ij (x) is the upper limit of α _ij .

[0073]

[Equation 27]

[0074]

[Equation 28]

[0075]

(Equation 29)

This will be described with reference to FIG. 320, 32
Reference numerals 1 and 322 denote membership functions of the clusters ij, of, and ik, and reference numerals 323 and 324 denote ranges of the membership function 320 for making the class x correctly recognize the input x. Membership function 320 is 323 and 324
Input x by adjusting α _ij so that
Will be correctly recognized by class o.

[0077] Dec (l) the alpha _ij the section _{(L ij (l), α} ij) data misrecognition when set to a number that is recognized correctly. In this case K _ij (x) is the interval _{(L ij (l), α} ij) is incremented _{by one} Dec (l) when contained.

Here, it is defined as in the equation (30). If α _ij is set smaller than min (γ _ij (l), L _ij (l-1)),
Although erroneous recognition occurs in l-1 correctly recognized data, Dec (l) data is correctly recognized.

[0079]

[Equation 30]

Change α _{ij For} l Inc (l) (l = 1, ..., 1M), l that satisfies the equation (31) is obtained. However, lM is a positive integer. Similarly, Dec (l) (l = 1, ...
, LM) is calculated to satisfy l (32).

[0081]

[Equation 31]

[0082]

(Equation 32)

When there are a plurality of l's satisfying the equation (31) or the equation (32), the smallest l is taken. First, equation (31) is
(32) Consider the case of greater than or equal to. If α _ij is larger than β _ij (l) in the interval (α _ij , U _ij (l)),
The increase in the number of correctly recognized data is Inc (l) -l + 1. Therefore, α _ij is set as in equation (33) in the interval [β _ij (l), U _ij (l)). Here, δ satisfies 0 ≦ δ <1.

[0084]

[Expression 33]

[0085] Similarly when the formula (31) is smaller than the formula (32), the alpha _ij of _{(L ij (l), γ} ij (l)) γ ij in the range of (l) smaller than the formula (34) To fix.

[0086]

(Equation 34)

Adjustment Method From the above discussion, the fuzzy rule may be adjusted as shown in FIG. That is, (step 400) an appropriate positive integer is set in the parameter 1M. At this time, lM-1 is α
_It is the maximum number of misrecognitions to allow when adjusting _ij .
Further, δ in the equations (33) and (34) is set to a value in the interval of [0, 1), and α _ij is set to the same positive initial value. (Step 410) For α _ij (i = 1, ..., n, j = 1, ...), L _ij (l), U _ij (l), Inc (l), Dec (l) , B _ij (l), and γ _ij (l) (l = 1, ..., 1M). Formula (31) or formula
Finding l that maximizes (32), equation (33) or equation (3
Change α _{ij according} to 4). (Step 420) Step 410 is ended when the recognition rate is not improved, otherwise step 410
Return to

FIG. 17 shows a pattern recognition system for recognizing a vehicle number as an example using the above fuzzy inference. In the figure, 801 is a plate cutting means, 802 is a character cutting means, and 803 is a feature extracting means. The details of each means are disclosed in the patents already filed. That is, regarding the means for detecting the vehicle and cutting out the plate,
JP-A-60-241387 (image processing device), JP-A-61-141087
(Image processing method and apparatus), JP-A-63-36383 (image processing apparatus), JP-A-63-153682 (character processing method and apparatus) for character extraction, and JP-A-62-293492 for feature extraction. The image of a running car, which is described in detail in (Character feature extraction method), is captured by an image recognition device on an industrial television. The image recognition device cuts out a license plate, cuts out characters one by one, and extracts the feature amount of the character. When recognizing numbers, the number of holes, the degree of bending, etc. are used as the feature amount. The feature amount is input to the fuzzy rule inference means 102, the degree of success for the numbers 0 to 9 is calculated using the equations (2) to (4), and it is determined that the number corresponds to the highest degree of success. Then, the identification result is output. The extraction and adjustment of the fuzzy rules are performed by the fuzzy rule construction means 101 in advance. That is, the data dividing unit 104 divides the teacher data into clusters, the rule extracting unit 105 extracts the fuzzy rule of the formula (1), and the rule adjusting unit 106 adjusts the fuzzy rule according to the procedure of FIG.

In order to evaluate this method, 810 pieces of teaching data for numbers 0 to 9 and 8 pieces of test data are used.
20 pieces were prepared and the fuzzy rule was decided as mentioned above.
At this time, the number of teacher data for each number is 81. lM = 10
When the upper limit Nc of the division is 70, the recognition rate of the test data is 99.63% and the adjustment of the fuzzy rule is 60M.
CPU time was 2.6 seconds on the IPS computer.

Comparing this with a neural network, the maximum recognition rate is 99.76% and the minimum is 98.90 when learning is performed 100 times by changing the initial value in the neural network.
%, And the average was 99.41%. Moreover, the learning time was 78.9 seconds per time with the same computer.

Therefore, with the method of the present invention, the performance near the maximum performance of the neural network can be realized in 1/30 of the calculation time. Not only numbers, but also kanji and hiragana can be recognized. At this time, the character image cut out instead of the feature amount may be divided into 8 × 7 or the like and input.

FIG. 18 shows a case where the fuzzy inference method of the present invention is applied to a pattern recognition system (blood cell classification device) for classifying white blood cells. White blood cells are classified according to the growth stage, and are classified into 5 to 7 types of mature cells and immature cells, respectively, and are classified into about 12 types of blood cells in total. Since the type is determined according to the growth stage, it is known that the boundary between adjacent types is ambiguous and it is very difficult to classify.

In the figure, 901 is a blood cell cutting means,
902 is a feature extraction means. The optically screened blood cell data is loaded into a blood cell classifier. The blood cell classifying device cuts out blood cells and extracts the characteristic amount of the blood cells such as the perimeter and area of the cell nucleus. The feature amount is input to the fuzzy rule inference means 102, the degree of success of about 12 types is calculated by using the equations (2) to (4), and it is determined that the feature degree belongs to the category corresponding to the highest degree of success. The identification result is output. The fuzzy rules are extracted and adjusted in advance by the fuzzy rule construction means 10
Do in 1. That is, the data dividing means 104
The teacher data is divided into clusters by the rule extraction means 10
In step 5, the fuzzy rule of the formula (1) is extracted, and the rule adjusting means 106 adjusts the fuzzy rule in the procedure shown in FIG.

The teacher data is set to 3 in order to evaluate this method.
097 pieces and 3100 pieces of test data were prepared, and the fuzzy rule was determined as described above. When lM = 10 and the upper limit Nc of division is 80, the recognition rate of the test data is 91.
58% adjustment of fuzzy rules with 60 MIPS calculator
CPU time was 41.6 seconds.

Comparing this with a neural network, the maximum recognition rate when the initial value was changed in the neural network and learning was performed 25 times was 90.46%, and the average was 87.44%.
Also, the learning time was 13 per session with the same computer.
3 minutes. Therefore, with the method of the present invention, it is possible to realize the performance exceeding the maximum performance of the neural network in a calculation time of 1/190.

[0096]

The effects of the embodiments of the present invention described above will be described.

(A) Dividing the input data of each class By dividing the input data of each class, the approximation of the area of each class can be made closer to the teacher data.

(B) Approximating a region with divided data to extract a fuzzy rule Approximating a region with divided data and using this as a fuzzy rule allows easy rule extraction. .

(C) Tuning the fuzzy rule by adjusting the approximation area using the teacher data By adjusting the approximate area so that the pattern recognition result for the teacher data is improved, the competitive relation between the fuzzy rules is adjusted. Can be adjusted, and the generalization ability can be improved as well as the recognition rate for teacher data.

[Brief description of drawings]

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

FIG. 2 is a block diagram showing a hardware configuration of the pattern recognition device shown in FIG.

FIG. 3 is an explanatory diagram showing a configuration of an input / output data file in the pattern recognition device shown in FIG.

4 is an explanatory diagram showing a configuration of a divided data file in the pattern recognition device shown in FIG.

5 is an explanatory diagram showing a configuration of a fuzzy rule file in the pattern recognition device shown in FIG.

FIG. 6 is an explanatory diagram showing a configuration of a fuzzy system.

FIG. 7 is an explanatory diagram showing how to divide data.

FIG. 8 is an explanatory diagram showing a state in which the center and radius of an ellipse are corrected when an existing region of input data is approximated to an ellipse.

FIG. 9 is an explanatory diagram showing an example of a specific method for adjusting a fuzzy rule.

FIG. 10 is an explanatory diagram showing another example of a specific method of adjusting a fuzzy rule.

FIG. 11 is an explanatory diagram showing a lower limit of parameter adjustment for fuzzy rule adjustment.

FIG. 12 is an explanatory diagram showing a lower limit of parameter adjustment for fuzzy rule adjustment.

FIG. 13 is an explanatory diagram showing an upper limit of parameter adjustment for fuzzy rule adjustment.

FIG. 14 is an explanatory diagram showing a state where erroneous recognition is eliminated by fuzzy rule adjustment.

FIG. 15 is an explanatory diagram showing a state in which erroneous recognition is eliminated by fuzzy rule adjustment.

FIG. 16 is an explanatory diagram showing a procedure for adjusting parameters for adjusting a fuzzy rule.

FIG. 17 is a block diagram showing a configuration of a pattern recognition system for recognizing a vehicle number to which the present invention is applied.

FIG. 18 is a block diagram showing the configuration of a pattern recognition system for classifying blood cells to which the present invention is applied.

[Explanation of symbols]

 10 CPU 20 Memory 30 Keyboard 40 CRT 101 Fuzzy Rule Constructing Means 102 Fuzzy Rule Inferring Means 103 Storage Means 104 Data Dividing Means 105 Rule Extracting Means 106 Rule Adjusting Means 107 Input / Output Data Files 108 Dividing Data Files 109 Fuzzy Rule Files 801 Plate Cutting Means 802 Character cutting means 803 Feature extracting means 901 Blood cell cutting means 902 Feature extracting means

Claims (18)

[Claims] 1. Fuzzy rule generation for generating fuzzy rules for pattern recognition by using input data and output data indicating which class the input data belongs to as teacher data and using input / output relationships of some teacher data. In the pattern recognition device having means, the fuzzy rule generating means has a data dividing means for dividing data into clusters, a rule extracting means for extracting fuzzy rules, and a rule adjusting means for adjusting the fuzzy rules, The data dividing means divides the data belonging to the same class into one or more clusters, and the rule extracting means is used to approximate the existing area of the input data belonging to the divided clusters with an elliptic area to correspond to the clusters. Extract the fuzzy rules, and use the rule adjustment means Pattern recognition apparatus characterized by adjusting the existing area of the fuzzy rules to the extent that belong becomes large. 2. The data dividing means has means for obtaining an average value for each axis of input data corresponding to each cluster, and means for obtaining the number of teacher data larger than the average value or less than the average value. The pattern recognition apparatus according to claim 1, wherein the cluster is divided along the axis having the smallest absolute value of the difference between the number of teacher data larger than the average value and the number of teacher data smaller than the average value. 3. The rule extracting means obtains an average value of displacement on the positive side and an average value of displacement on the negative side of the average value of each axis of the teacher data belonging to the cluster, 2. The pattern recognition apparatus according to claim 1, further comprising means for obtaining a radius, wherein the center is determined at a position without deviation based on a difference in average value of displacements. 4. The rule adjusting means is means for obtaining an upper limit and a lower limit of a slope of an elliptic region when l−1 (l is a positive integer) pieces of correctly separated data are erroneously recognized. And a means for calculating the inclination of the elliptical area for correctly recognizing the erroneously recognized data, and increasing or decreasing the inclination of one elliptic area so that the recognition rate is most improved. Item 1. The pattern recognition device according to item 1. 5. A fuzzy rule generation for generating fuzzy rules for pattern recognition by using input data and output data indicating which class the input data belongs to as teacher data, and using input / output relationships of some teacher data. In the pattern recognition device having means, the fuzzy rule generating means has a data dividing means for dividing data into clusters, a rule extracting means for extracting fuzzy rules, and a rule adjusting means for adjusting the fuzzy rules, The data dividing unit divides the data belonging to the same class into one or more clusters, and the rule extracting unit is used to approximate the existing region of the input data belonging to the divided clusters with an elliptic region to correspond to the clusters. Extract the fuzzy rules, and extract the data correctly separated by the rule adjusting means. Then, the upper and lower limits of the slope of the elliptic region when l−1 (l is a positive integer) data are erroneously recognized are calculated, and the slope of the elliptic region for correctly recognizing the misrecognized data is calculated. A pattern recognition device characterized by increasing or decreasing the inclination of one elliptical region so as to improve the recognition rate. 6. A device for recognizing a pattern using a fuzzy rule in which a membership function of the condition part of the fuzzy rule is defined in an elliptic region around the center of the cluster, and the degree of satisfaction of the condition part of each fuzzy rule with respect to input data is determined. A pattern recognition apparatus having a calculating means, and determining that an input belongs to a class to which a fuzzy rule having the highest satisfaction degree belongs, based on a calculation result of the satisfaction degree calculating means. 7. A device for pattern recognition using a fuzzy rule defined by a conditional point of a fuzzy rule and a weight of a distance from the representative point, and calculating a weighted distance between the input data and the representative point. A pattern recognition apparatus comprising means for calculating a weighted distance to a representative point of each fuzzy rule of an input, and determining that the input belongs to a class to which the fuzzy rule having the closest weighted distance belongs. 8. A method for generating fuzzy rules for pattern recognition by using input data and output data indicating which class the input data belongs to as teacher data, and using input / output relationships of some teacher data. , The step of dividing the data into clusters, the step of extracting the fuzzy rules, and the step of adjusting the fuzzy rules.In the step of dividing the clusters, the data belonging to the same class is divided into one or more clusters, In the step of extracting the fuzzy rule, the fuzzy rule corresponding to the cluster is extracted by approximating the existence area of the input data belonging to the divided cluster with the elliptic area, and the fuzzy rule corresponding to the class to which the teacher data belongs in the step of adjusting the fuzzy rule. Existence of fuzzy rules to increase the degree of belonging Fuzzy rule generation method for pattern recognition and adjusts the area. 9. In the step of dividing the cluster, an average value of input data corresponding to each cluster for each axis is obtained, and then the number of teacher data larger than the average value or less than the average value is obtained. 9. The fuzzy rule generation method for pattern recognition according to claim 8, wherein the cluster is divided along the axis having the smallest absolute value of the difference between the large number of teacher data and the smaller number of teacher data. 10. In the step of extracting the fuzzy rule, an average value of displacements on the positive side and an average value of negative sides of the average value of each axis of the teacher data belonging to the cluster are calculated, and the average value of each displacement 9. The fuzzy rule generation method for pattern recognition according to claim 8, wherein the center is determined at a position having no deviation based on the difference between the two. 11. In the step of adjusting the fuzzy rule, an upper limit and a lower limit of an inclination of an elliptic region when l−1 (l is a positive integer) of correctly separated data are erroneously recognized. 9. The method according to claim 8, wherein the slope of the elliptic region for obtaining the correct recognition of the erroneously recognized data is calculated, and the slope of one elliptic region is increased or decreased so as to improve the recognition rate most. Fuzzy Rule Generation Method for Pattern Recognition of Words. 12. A method for generating fuzzy rules for pattern recognition by using input data and output data indicating which class the input data belongs to as teacher data, and using input / output relationships of some teacher data. , The step of dividing the data into clusters, the step of extracting the fuzzy rules, and the step of adjusting the fuzzy rules, the step of dividing into clusters divides the data belonging to the same class into one or more clusters, In the step of extracting the fuzzy rule, the fuzzy rule corresponding to the cluster is extracted by approximating the existence area of the input data belonging to the divided cluster with an elliptic area,
In the step of adjusting the fuzzy rule, the upper and lower limits of the slope of the elliptic region when l−1 (l is a positive integer) of correctly separated data are erroneously recognized,
A fuzzy rule generation method for pattern recognition characterized by calculating the slope of an elliptical region for correctly recognizing misrecognized data and increasing or decreasing the slope of one elliptic region so as to improve the recognition rate most. . 13. A method for pattern recognition using a fuzzy rule in which a membership function of the conditional part of the fuzzy rule is defined in an elliptical region around the center of the cluster, wherein the conditional part of each fuzzy rule is satisfied for input data. A pattern recognition method characterized by calculating a degree and determining that an input belongs to a class to which a fuzzy rule having the highest success rate belongs. 14. A method for recognizing a pattern using a fuzzy rule defined by a representative point and a weight of a distance from the representative point in a conditional part of the fuzzy rule, wherein Calculate the weighted distance,
A pattern recognition method characterized by determining that an input belongs to a class to which a fuzzy rule having the closest weighted distance belongs. 15. A fuzzy rule for generating a fuzzy rule for recognizing a license plate of a vehicle by using input data and output data indicating which class the input data belongs to as teacher data and using input / output relations of some teacher data. In the pattern recognition device having rule generating means, the fuzzy rule generating means has a data dividing means for dividing data into clusters, a rule extracting means for extracting fuzzy rules, and a rule adjusting means for adjusting fuzzy rules. Then, the data dividing means divides the data belonging to the same class into one or more clusters, and the rule extracting means is used to approximate the existing areas of the input data belonging to the divided clusters with elliptic areas to form clusters. The corresponding fuzzy rule is extracted, and the teacher data is Pattern recognition apparatus characterized by a degree of belonging to a class to adjust the existing area of fuzzy rules so as to increase. 16. A pattern recognition device for recognizing a license plate using a fuzzy rule defined by a condition point of a fuzzy rule defined by a weight of a representative point and a distance from the representative point, with weighting of input data and a representative point. The weighted distance calculation means calculates a weighted distance to a representative point of each fuzzy rule of the input, and the input belongs to a class to which the fuzzy rule having the closest weighted distance belongs. A pattern recognition device, characterized in that 17. A pattern having fuzzy rule generating means for generating fuzzy rules for classifying blood cells by using input data and a class to which the input data belongs as output data and output data as teacher data. In the recognition device, the fuzzy rule generating means has a data dividing means for dividing the data into clusters, a rule extracting means for extracting the fuzzy rules, and a rule adjusting means for adjusting the fuzzy rules. The data belonging to the same class is divided into one or more clusters, and the rule extraction means is used to extract the fuzzy rule corresponding to the clusters by approximating the existence region of the input data belonging to the divided clusters with an elliptical region. The degree to which the teacher data belongs to the class to which the teacher data belongs by the rule adjusting means. Pattern recognition apparatus characterized by adjusting the existing area of fuzzy rules so as to increase. 18. A weighted distance between input data and a representative point in a pattern recognition device for classifying blood cells using a fuzzy rule defined by a condition part of a fuzzy rule defined by a representative point and a weight of the distance from the representative point. If the input belongs to a class to which the fuzzy rule having the closest weighted distance belongs, the weighted distance to the representative point of each fuzzy rule of the input is calculated by the weighted distance calculating means. A pattern recognition device characterized by making a judgment.