JPH09138785A - Pattern matching method and device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To precisely identify a pattern which is not learned so as to output it in a pattern matching device using a neural network. SOLUTION: A learning signal holding part 50 generates a boundary signal around an initial learning input signal belonging to the same category based on the initial learning input signal written into an initial learning input signal memory 56 by a pattern boundary setting unit 54 and writes the signal into a learning input signal memory 51 with the initial learning input signal. A pattern boundary setting unit 54 generates a boundary teacher signal which corresponds to the boundary signal and by which the unit output of an output layer 20 becomes zero or a small value from the initial teacher signal written into an initial teacher signal memory 58, and writes the boundary teacher signal into a teacher signal memory with the initial teacher signal. The initial learning input signal and the boundary signal are learned by using data written into the learning input signal memory 51 and the teacher signal memory 52.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern matching apparatus, which uses a neural network to recognize handwritten characters, voice recognition, or an information signal as a pattern, and a final pattern corresponding to the detected pattern. The present invention relates to a pattern matching device that teaches an output and thereafter automatically executes a matching process.

[0002]

2. Description of the Related Art Conventionally, a learning operation of recognizing handwritten characters and inputting a pattern to be identified so as to obtain a control output corresponding to various information signals and correcting the corresponding output by comparison with a teacher pattern is repeated. After that, a pattern matching device is known in which after that, an arbitrary identification pattern is input to automatically determine the pattern. Such a pattern matching device is generally configured to use a neural network configured by connecting a multi-input single-output type element (unit) with another unit by a coupling element.

The neural network used in this type of pattern matching device is an information processing technique that imitates the information processing function of the cerebrum of a living organism, and is useful because it is easy to apply and can be applied to nonlinear problems. It has characteristics. Further, since the method is different from the conventional Neumann-type computer, it is expected to be applied to the problem which has been difficult in the Neumann-type computer.

The components of the neural network are shown in FIG.
As shown in (1), it is composed of a unit 1 corresponding to a nerve cell of an organism and a coupling element 2 corresponding to a nerve fiber, which connects an input portion and an output portion of this unit 1. Usually, effective information processing is not performed only by using a single unit, so a plurality of units are appropriately combined using a coupling element to form a network.

Each unit is coupled to a plurality of other units through a coupling element 2, and the output of the other unit is multiplied by a weighting factor assigned to each coupling element, as shown in the following formula 1, In addition, the value of an input / output function having the difference between this input value and a given threshold value as an independent variable is output and becomes an input to another unit via a coupling element for output.

(Equation 1) However, O _i : output value of the i-th unit, f: input / output function w _ji : strength of coupling (coupling coefficient) from the j-th unit to the i-th unit, o _j : j-th unit Is the output value of the unit, θ _i is the threshold value of the i-th unit. Also, depending on the definition, Equation 1

(Equation 2) In some cases,

At this time, the input / output function f is shown in FIG.
A saturation type function called a sigmoid function as shown in is often used. This function imitates the function of a nerve cell of an organism, and is a monotonically increasing function that takes a value from 0 to 1, and when the input value is equal to the threshold value (or when the input value is the threshold value and the absolute value is It becomes 0.5 when the signs are equal.

Next, useful information processing can be performed by constructing a network using the units having the above-mentioned functions. As a typical network, one as shown in FIG. 7 is considered. There is. FIG. 7 (1) shows a unit in which 1 constitutes an input layer in a hierarchical network,
Units 2 and 3 are first and second intermediate layers, respectively, and unit 4 is an output layer, and units 5, 6, and 7 are coupling elements. Such a hierarchical network has a hierarchical structure. Units in different layers are connected to each other, but units in the same layer are not connected, and information is unidirectionally transmitted from the input section to the output section. Transmitted. It is, so to speak, a feed-forward type network. This model is known as a cerebellum model, and at the same time, it is an important network that is used in many fields such as pattern matching and control.

In this model, a pattern to be identified, a time series signal, etc. are given to the input layer. For example, in the case of application to pattern matching such as recognition of handwritten characters, the input layer is given a handwritten character or a feature pattern after some kind of feature extraction processing is performed on the handwritten character. On the other hand, each unit in the output layer is set so as to give a judgment result for the category to be learned. That is, if the output of the unit 1 is, for example, 0.5 or more, the neural network is set so that it can be considered that the given pattern is identified as the category 1 pattern.

FIG. 7 (2) shows a mutual coupling type network, where 1 is a unit and 2 is a coupling element. In this way, all units are mutually connected in the mutual connection type network. In addition to this, there are networks that have feedback, and networks that partially have coupling between the same layers.

Next, in order to allow the network to have useful functions as described above, learning is performed by giving learning data to the network. Since there are many types of learning methods and networks, there are many learning methods. Here, an example in which a learning method using a backpropagation algorithm, which is the most typical learning method at present, is applied to the three-layer hierarchical neural network shown in FIG.

In learning, first, the coupling coefficient and the threshold value are set to small values, the output when a learning pattern is given to the input layer is calculated, and the result and a valid output for the pattern are designated. This is performed by changing the coupling coefficient and the threshold value so that the output of the output layer approaches the teacher signal for the given learning pattern as compared with the teacher pattern. The detailed steps of learning are as follows.

(Step 1) The coupling coefficient matrices {W _ij }, {V _jk } and the threshold vectors {θ _j }, {γ _k } that determine the state of the network are initialized with small random values. (Step 2) Let the first pattern be a learning pattern. (Step 3) Put the value of the learning pattern in the output {I _i } of the input layer unit, and use the coupling coefficient matrix {W _ij } from the input layer to the intermediate layer and the threshold θ _j of the intermediate layer unit j. , prompts U _j to the intermediate layer unit j, obtaining the output H _j of the hidden unit j the input U _j and sigmoid function f shown in FIG. 6 (2).

(Equation 3)

(Step 4) Output of the intermediate layer unit {H
_j }, the coupling coefficient matrix {V _jk } from the intermediate layer to the output layer and the threshold value γ _k of the output layer unit k, the input S _k to the output layer unit _k is obtained, and the input S _k and The output O _k of the output layer unit k is obtained by the sigmoid function f.

(Equation 4) (Step 5) From the difference between the teacher signal T _K corresponding to the learning pattern and the output O _k of the output layer, the coupling coefficient connected to the unit k of the output layer and the error δ _k with respect to the threshold of the output layer unit k are _obtained .

(Equation 5) (Step 6) From the error δ _k , the coupling coefficient matrix {V _jk } from the _hidden layer to the output layer, and the output H _j of the hidden layer, the error δ between the coupling coefficient connected to the hidden layer unit j and the threshold value of the hidden unit. ask for _j .

[0014] coupling coefficient between the (Step 7) a product of the output H _j and constants α error [delta] _k and the intermediate layer unit j in the output layer unit k obtained in step 5 with the conventional intermediate layer output layer To correct the coupling coefficient V _jk connected from the intermediate layer unit j to the output layer unit k. Further, by adding the product of the error δ _k and the constant β to the threshold value of the conventional output layer unit k, the threshold value γ _k of the output layer unit k is added.
To correct.

(Equation 6) (Step 8) The error .delta..sub.j in the intermediate layer unit j and the product of the output _I.sub.i of the input layer unit _i and the constant .alpha. Modify the coupling coefficient W _ij from unit i to intermediate layer unit j. Further, the threshold value θ _j of the intermediate layer unit _j is corrected by adding the product of the error δ _j and the constant β to the threshold value of the conventional intermediate layer unit j.

(Equation 7)

(Step 9) The next pattern is used as a learning pattern. (Step 10) Return to Step 3 until the learning pattern ends. (Step 11) Update the number of learning iterations. (Step 12) If the number of repetitions of learning is less than or equal to the limit number, the process returns to step 2.

By such learning, the neural network acquires the pattern matching ability and can identify the patterns already learned and the unlearned patterns belonging to the learned category.

[0017]

However, in the conventional pattern matching apparatus using the neural network processing, when a pattern belonging to a category that has not been learned is identified, it is erroneously determined as one of the learned categories and output. There are often cases. This problem is generally a response to unlearned data.

That is, in the conventional pattern matching device, when discriminating between categories A and B as shown in FIG. 9, patterns belonging to these two categories A and B (marked by "○" or "x") are used. ) Is used for learning. As a result of this learning, a boundary line D like a broken line shown in FIG. 9 is set. Then, depending on which side of the boundary line D the input pattern is, it is determined which category the pattern belongs to. Therefore, it can be correctly determined by identifying the data belonging to A or B, but when a pattern that does not belong to these categories is input, this pattern is forcibly determined to be one of the categories.

Incidentally, the reason for the erroneous determination is that in the case of the conventional input / output function, when a value considerably larger or smaller than the threshold value is inputted, a value of 1 or 0 is output regardless of the value. It is considered that this is because the so-called saturated type function is used, so that it only has the effect of setting the boundary surface between the learned categories.

The present invention focuses on the above-mentioned conventional problems, and aims at enabling an unlearned pattern to be properly identified and output by a pattern matching apparatus using neural network processing, and in particular, Even when the added value of the input signal multiplied by the weighting coefficient in each unit is not only near the threshold value but also greatly apart from the threshold value, an appropriate output signal is output, thereby improving the pattern matching determination accuracy. It is an object to provide a pattern matching device that can be improved.

[0021]

The present invention is described in, for example, FIG.
Data that gives the peripheral boundaries of categories A and B (shown by “⊚”) are combined from the learning data, and when these data are given to the input layer, the output of all cells (units) in the output layer This is based on the knowledge that the value should be learned so that the value becomes “0” or a small value. That is, the teacher signal for the input signal corresponding to the boundary pattern has a small value over all cells. By this, the pattern located at the boundary of both categories is used as a learning pattern, and as a result, it is recognized that the range shown by the solid line surrounding the learning data is the range of the category to be identified. The false recognition rate is greatly reduced.

Therefore, according to the present invention, a plurality of learning input patterns are input to the input unit group, and the output pattern of the output unit group corresponding to the input learning input pattern is compared with a predetermined teacher pattern. And the input coefficient for learning is repeated by repeating the operation of correcting the threshold value of the unit forming each unit group and the weighting factor of the coupling element connecting each unit group based on this error. After learning, the pattern matching method of inputting an arbitrary pattern to the input unit group and comparing the output pattern of the output unit group with the output pattern corresponding to the learned input pattern belongs to the same category. A boundary pattern around the plurality of learning input patterns is created based on the learning input pattern, After setting the teacher signal output from each unit of the output unit group corresponding to the field pattern to 0 or a small value and learning the boundary pattern together with the learning input pattern, an arbitrary pattern is input to the input unit group. By inputting, the above-mentioned object is achieved.

As the boundary pattern, a circle, an ellipse, or a rectangular area including all the learning input patterns belonging to the same category may be set, and a value outside this area may be set as the boundary pattern. Further, the boundary pattern is a maximum boundary value obtained by adding a predetermined value to the maximum value of each element of a plurality of learning input patterns belonging to the same category, and a predetermined value is subtracted from the minimum value of each element. By setting so as to include the minimum boundary value, it is possible to easily create a boundary pattern.

Further, the pattern matching apparatus according to the present invention is such that an input unit group to which a pattern to be identified is inputted, an output unit group to output a pattern corresponding to the input pattern, the input unit group and the output unit. At least one or more intermediate unit groups interposed between the groups, and a coupling element connecting these unit groups,

Learning signal holding for holding a plurality of learning input patterns to be input to the input unit group and a teacher pattern created corresponding to these learning input patterns and being compared with the output pattern of the output unit group And each unit of each unit group adds the output of the preceding unit multiplied by the weighting factor assigned to each coupling element, and the difference between this added value and the set threshold value is independently calculated. An operation result based on an input / output function as a variable is output, and based on an error between the output pattern of the output unit group obtained by inputting the learning input pattern into the input unit group and the corresponding teacher pattern, After the correction of the weighting factor of the coupling element and the threshold value is repeated to learn the learning input pattern, an arbitrary pattern is input to the input unit group. In the pattern matching device configured to perform pattern matching by means of boundary signal synthesis, the learning signal holding unit generates a boundary pattern around learning input patterns belonging to the same category based on the plurality of learning input patterns. And a boundary teacher signal synthesizing unit for setting the teacher signal corresponding to the boundary pattern and the output of each unit of the output unit group to 0 or a small value. The small value of the teacher signal corresponding to the boundary pattern is, for example, a value of about 0.1.

[0026]

According to the above construction, when the boundary pattern around the outside of the category is input to the input unit group, the output of each unit of the output unit group becomes "0" or a small value. By learning, as a result, the output of the output unit can be set to a small value even for a pattern far from the category. Then, by learning so that the output of the output unit with respect to the input of the pattern belonging to the category becomes large, the output corresponding to the input arbitrary pattern is compared with the learned output, so that the pattern becomes the category. Whether or not they belong can be easily and accurately determined, and erroneous determination can be eliminated. In addition,
The boundary pattern is synthesized based on the learning patterns belonging to each category, but this synthesizing method may be performed individually according to the identification target.

[0027]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing the overall configuration of the pattern matching device according to the embodiment. As shown in the figure, this device has an image receiving device 10 that captures recognition target images such as characters and various signal waveforms for which pattern matching is to be performed. The image receiving device 10 can directly display an image to be recognized, and outputs the image display signal to the pattern input unit 12. The pattern input unit 12 divides the image display area of the image receiving device 10 into a two-dimensional matrix, and separates individual image signals corresponding to each divided area according to the magnitude of the signal amount, for example, the intensity or density of the signal. It is supposed to be output to the conversion processing unit 14. As a result, the image signal input to the image receiving device 10 is input to the conversion processing unit 14 as a pattern signal.

The conversion processing unit 14 to which the output signal of the pattern input unit 12 is input is an input layer 16 as a unit group to which a pattern to be identified is input, and a unit group to output a pattern corresponding to the input pattern. Output layer 20 and an intermediate layer 18 as an intermediate unit group interposed between the input layer 16 and the output layer 20 and connected to these layers 16 and 20 by a plurality of coupling elements. . The input layer 16, the intermediate layer 18, and the output layer 20 each have a plurality of multi-input / single-output type elements (units) corresponding to the nerve cells of the organism, and the units of each layer are connected to correspond to nerve fibers. It is configured to use a neural network configured by connecting to units in other layers via elements.

That is, as shown in the example of the intermediate layer 18 in FIG. 2, each unit 22 constituting the intermediate layer 18 is
Each unit of the input layer 16 is coupled through the coupling element 23, and the output signal I of each unit of the input layer 16 is input as a value obtained by multiplying the weighting coefficient W assigned to each coupling element 23. ing. Then, the intermediate layer unit 22 outputs the output signal obtained by performing a predetermined calculation to the coupling element 2
4 to each unit 25 of the output layer 20. In addition, the unit 25 of the output layer 20 performs the same operation as the intermediate layer unit 22 and outputs the operation result via the coupling element 27.

As described above, the circuit configuration of the conversion processing unit 14 including the input layer 16, the intermediate layer 18, and the output layer 20 is as shown in FIG. 3, and the output calculation circuit 40 of the input layer 16 and the intermediate layer 18 are the same. The input calculation circuit 42, the output calculation circuit 44 of the intermediate layer 18, the input calculation circuit 46 of the output layer 20, and the output calculation circuit 48 of the output layer 20 are connected in series.

The input layer output calculation circuit 40 obtains the output of each unit of the input layer 16 based on the input pattern. In this example, an input / output function for calculating an output value is generally used as a linear function. It is a function. For this reason,
Each unit of the input layer 16 output the same values as the corresponding output value of the pattern input section 12 (I ₁ ~I _n). Also,
The input calculation circuit 42 of the intermediate layer 18 uses the output I of the input layer output calculation circuit 40 and the coupling coefficient W of the coupling element 23 that couples the input layer 16 and the intermediate layer 18 to each unit of the intermediate layer 18. The input value U to 22 is calculated, and the output calculation circuit 44 of the intermediate layer 18 calculates the output H of each unit 22 of the intermediate layer 18 from the input value U obtained by the input calculation circuit 42 and the input / output function described later. calculate. Then, the input calculation circuit 46 of the output layer 20 outputs the output H of the intermediate layer output calculation circuit 44.
And the coupling coefficient V of the coupling element 24 coupling the intermediate layer 18 and the output layer 20 to each unit 2 of the output layer 20.
5, the input S supplied to 5 is calculated, and the output layer output calculation circuit 4
Reference numeral 8 obtains the output O of the output layer 20 based on the calculated value S of the output layer input calculation circuit 46 and the input / output function.

Output of conversion processing unit 14, that is, output layer 2
The output O of each unit 0 of 0 is basically input to the pattern output unit 26. Then, the pattern input unit 26
Stores the pattern (vector) obtained by the output value O of each unit 25 in the memory 28, compares it with the pattern, reads the output pattern matching the output signal of the output layer 20 from the memory 28, and displays it on the display device 30. To do.

However, in the present embodiment, the learning function by back propagation is provided, and therefore, the error detection section 32 for inputting the output signal from the output layer 20 is provided at the subsequent stage of the conversion processing section 14. It is provided. A teacher device 34 is connected to the error detection unit 32, and a teacher signal corresponding to a learning input pattern (input signal) given to the image receiving device 10 is transmitted when learning for learning the pattern matching ability of the pattern matching device. It is provided from the teacher device 34.

A learning signal holding section 50 is connected to the image receiving apparatus 10 and the teacher apparatus 34. This learning signal holding unit 5
0 is a learning input signal memory 51, a teacher signal memory 52,
Pattern boundary setter 54, initial learning input signal memory 5
6, an initial teacher signal memory 58. Then, the pattern boundary setter 54 creates a boundary pattern based on the data (input pattern for learning) written in the initial learning input signal memory 56 and inputs this boundary pattern into the initial learning input. The teacher signal (boundary teacher signal) corresponding to the boundary pattern is created based on the initial teacher signal written in the learning input signal memory 51 and the initial teacher signal memory 58 together with the data written in the signal memory. The signal and the boundary teacher signal are written in the teacher signal memory 52. Further, at the time of learning of the pattern matching device, an input signal for learning (input pattern) is given from the learning input memory 51 to the image receiving device 10, and a teacher signal corresponding to this input signal is given from the teacher signal memory 52 to the teacher device 34. To be

The pattern boundary setter 54 defines boundaries around a group of initial learning input signals belonging to the same category. As shown in FIG. 4, the maximum and minimum values of the initial learning input signals are input. It has a value calculator 60. The maximum value / minimum value calculation unit obtains the maximum value and the minimum value of each element of the initial learning input signal for each category. And
The maximum value and the minimum value calculated by the maximum value / minimum value calculation unit 60 are input to the alternative value calculation unit 62. The substitute value calculation unit 62 obtains a value obtained by adding a predetermined value to the input maximum value (maximum boundary value) and a value obtained by subtracting the predetermined value from the input minimum value (minimum boundary value). And sends it to the boundary signal synthesizer 64. These alternative value calculation units 62
The predetermined value to be added or subtracted by can be varied depending on the distance between the categories, and is determined by simulation or the like. Then, the alternative value calculation unit 6
The value obtained by 2 is sent to the boundary signal synthesis unit 64.

The boundary signal synthesizing unit 64 creates a boundary signal (boundary pattern) that defines the boundary of the category by replacing any element of the initial learning input signal with the value obtained by the alternative value calculating unit 62, and performs the initial learning. The input signal and the learning input signal memory 50 are written. In addition, the pattern boundary setter 54
The boundary teacher signal synthesizer 66 synthesizes the boundary teacher signal corresponding to the boundary signal obtained by the boundary signal synthesizer 64, based on the input initial teacher signal. The boundary teacher signal calculated by the boundary signal teacher signal synthesizer 66 is supplied to each unit 2 of the output layer 20 of the conversion processor 14 when the boundary pattern obtained by the boundary signal synthesizer 64 is given to the image receiving device 10.
The output value of 5 is "0" or a small small value, which is "0" in principle.

In the embodiment constructed as described above,
When the pattern (input signal) to be identified is given to the image receiving device 10, the pattern input unit 12 divides the image display region of the image receiving device 10 into a two-dimensional matrix, and a signal corresponding to the image signal corresponding to each divided region. Is output to the conversion processing unit 14. Since the input / output function of the input layer output calculation circuit 40 provided in each unit constituting the input layer 16 of the conversion processing unit 14 is a linear function, the output value of the pattern input unit 12 input to each unit. Output a value I _i equal to.

Each unit 22 of the intermediate layer 18 corresponds to the input layer 1
The value I _i output from each unit of 6 is received via the coupling element 23. Then, each unit 22 of the intermediate layer 18
The input calculation circuit 42 calculates the input U _j by the following formula 8 in the same manner as the formula 3 described above, and sends it to the output calculation circuit 44.

(Equation 8)

However, U _j is the input value of the j-th unit 22 of the intermediate layer 18, and W _ij is the combination of the i-th unit of the input layer 16 and the j-th unit of the intermediate layer 18. A coupling coefficient representing the strength (weighting coefficient of the coupling element 23),
I _i is the output of the i-th unit of the input layer 16.
Further, θ _j is a threshold value of the j-th unit 22 of the intermediate layer 18.

When the input calculation circuit 42 receives the input value obtained by the above equation 8, each unit 22 calculates the output value H _j using the input / output function. As the input / output function used for this calculation, the saturation type sigmoid function shown in FIG. 5B may be used.
The input / output function f shown in is used.

(Equation 9)

This input / output function f produces a maximum value output when the absolute value of the input value is equal to the threshold value, and a positive output which does not exceed the maximum value for input values different from the threshold value. Where c is a coefficient, θ is a threshold value, and is a constant or variable. Then, x is an input value.

That is, the output calculation circuit 44 of the intermediate layer 18
Is calculated based on the input value U calculated by the above equation 8 based on the above equation 9, and this calculated value is used as the output value H _j to the next layer of each unit 22 in the intermediate layer 18.
This is

(Equation 10) Is calculated as Where H _{j is} the output value of the j-th unit, U _{j is} the input value of the j-th unit, and θ _j is the threshold value of the j-th unit.

Further, the intermediate layer 18 thus obtained
The output H _i from the input layer is input to the output layer 20 via the coupling element 24. In the input calculation circuit 46 in the input part to the output layer 20, first, the output H _i of the intermediate layer 18 and the intermediate layer 18 are input. And the coupling coefficient V of the coupling element 24 connected to the output layer 20 is used to calculate the input S supplied to each unit 25 of the output layer 20. This arithmetic processing is similar to the above-mentioned formula 8,

[Equation 11]

However, V _jk is a coupling coefficient (weighting coefficient) of the coupling element 24 connecting the j-th unit 22 of the intermediate layer 18 and the k-th unit 25 of the output layer 20, and γ _k is the output layer. The threshold of the kth unit 25, S _k, is the input value to the kth unit 25 of the output layer 20.

Next, each unit 25 of the output layer 20
The output value O is calculated on the basis of the input / output function f of the mathematical expression 9. That is, in the output layer 20, the output calculation circuit 48 performs an operation based on the input value S calculated by the above equation 11 based on the above equation 9 to output the outputs of the plurality of units 25 constituting the output layer 20. The value is calculated (output signal O). This is

(Equation 12) Calculate by

The output signal O thus obtained is output from the pattern output section 26 for each unit 25 of the output layer 20, and an output pattern matching the output signal of the output layer 20 is stored in the memory by the pattern output section 26. The pattern read out from the display device 28 and input to the image receiving device 10 is displayed on the display device 30 or the like as a corresponding pattern.

Learning by back propagation of the pattern matching device is performed as follows. The initial learning input signal memory 56 is written with an initial learning input signal (initial learning input pattern) for the pattern matching device to learn, and the initial teacher signal memory 58 is written with the input signal written in the initial learning input signal memory 56. A teacher signal (teacher pattern) corresponding to is written. The signals written in the memories 56 and 58 are read by the pattern boundary setting unit 54. The pattern boundary setter 54 creates a boundary signal and a boundary signal teacher signal from the read input signal and teacher signal, and uses these created signals together with the initial learning input signal and the initial teacher signal as a learning input signal and a teacher signal. The learning input signal memory 51 and the teacher signal memory 52 are written.

That is, in the pattern boundary setting unit 54, the maximum value / minimum value calculation unit 60 has the initial learning input signal memory 56.
The initial learning input signal written in is read, the maximum value and the minimum value of the input signal for each category are calculated, and the calculated values are input to the alternative value calculation unit 62. For example, when the initial learning input signal is represented by a two-dimensional vector (two-dimensional coordinates x, y), the maximum / minimum value calculation unit 60 causes the maximum and minimum values x and y of each category to be x _q max. , X _q min, y _q max,
Find y _q min. However, the subscript q indicates the category to which the initial learning input signal belongs, and becomes a different numerical value or symbol for each different category.

The substitute value calculation unit 62 adds coordinate values x _q max + φ, y _q max + ψ that are obtained by adding a predetermined value to the input maximum value.
And a coordinate value x _q min−χ, y _q min−ω obtained by subtracting a predetermined value from the input minimum value, and the obtained coordinate value is input to the boundary signal synthesis unit 64. The boundary signal synthesizing unit 64 replaces the coordinate value of any one of the initial learning input signals written in the initial learning input signal memory 56 with the coordinate value obtained by the alternative value calculating unit 62 to determine the boundary of the category. (X _q max + φ, y _qi ), (x _qi , y _q min
-Ω) and the like are created and written in the learning input signal memory 50 together with the initial learning input signal. The boundary teacher signal synthesis unit 66 of the pattern boundary setting unit 54 synthesizes the teacher signal corresponding to the boundary signal obtained by the boundary signal synthesis unit 64 based on the input initial teacher signal, and the synthesized boundary signal. The teacher signal for writing is written in the teacher signal memory 52 together with the initial teacher signal. Also, the coupling coefficient and the threshold value of each coupling element are set to small values by uniform random numbers.

In this way, when the learning signal composed of the learning input signal and the teacher signal is created and the coupling coefficient and the threshold value are set, the first learning input signal (learning input pattern) is sent to the image receiving device 10. give. The learning input signal provided to the image receiving device 10 is transmitted to the pattern input unit 1 as described above.
2 is input to the conversion processing unit 14, the output signal O is output from each unit 25 of the output layer 20, and the error detection unit 3
2 is input. A teacher signal (teacher pattern) corresponding to the first learning input signal is input to the error detector 32 via the teacher device 34. The error detection unit 32 compares the output O of each unit 25 of the output layer 20 with the teacher signal given via the teacher device 34, obtains the error between the two, and outputs the error to the display device 30 or the like. Then, the threshold value of each unit and the coupling coefficient of each coupling element are changed and corrected so that the error between the output of the output layer 20 and the teacher signal is minimized according to a predetermined algorithm of back propagation.

When the learning for the first learning input signal is completed in this way, the next learning input signal is input to the image receiving device 10 and similarly learned. Then, this learning is performed for all learning input signals written in the learning input signal memory 51, and the same applies to the boundary signal obtained by the boundary signal combining unit 64 of the boundary setting unit 54.
However, the teacher signal of the boundary signal (the teacher signal for boundary signal) obtained by the boundary signal synthesis unit 64 is set so that the output O of each unit 25 of the output layer 20 becomes “0” or a small value. Then, the learning input signal (learning input pattern) is learned by performing the above learning until the error between the output value and the teacher signal becomes equal to or less than a predetermined value or until the learning number reaches a predetermined number.

[0052]

EXAMPLE A pattern matching apparatus constructed in this manner was used to identify a pattern represented by a two-dimensional vector.

FIG. 5 shows a two-dimensional vector used for learning. The learning pattern belonging to category 1 is the symbol "*"
, And the learning patterns belonging to category 2 are indicated by the symbol “+”. Table 1 shows pattern data (coordinates) belonging to the categories 1 and 2.

[0054]

[Table 1]

The maximum value x ₁ max of x belonging to category 1 is 0.75, the minimum value x ₁ min is 0.25, and the maximum value y of y is y.
₁ max is 0.75 and the minimum value y ₁ min is 0.25. Therefore, in order to obtain the boundary signal (boundary pattern) of category 1, if φ and ψ added to these maximum values and χ and ω subtracted from the minimum value are respectively 0.15,

(Equation 13)

[Equation 14]

(Equation 15)

(Equation 16) Is obtained.

Therefore, when the boundary signal is obtained from the data in Table 1 and the mathematical expressions 13 to 16, the boundary signals for the data A are a ₁ (0.9, 0.75) and a ₂ (0.1, 0). ．
_{75), a 3 (0.75,0.9)} , a 4 (0.75,
0.1). In the same manner, boundary signals were similarly obtained for the data B to I of category 1 and the data A ′ to I ′ of category 2. An example of the boundary signal thus obtained is shown by a circle in FIG.

Then, after learning the pattern using these boundary signals and the initial learning input signal shown in Table 1 to acquire the pattern discrimination ability, pattern matching is performed using the unlearned test data. Was tested. Table 2 shows the test data and test results. Also,
The data for assay is shown in FIG.

[0058]

[Table 2]

Comparing these results, the discrimination result by the boundary pattern learning according to the present embodiment is a very good judgment result. That is, in the conventional method in which the boundary signal is not learned, the output value for the unlearned data pattern that does not belong to the category to be identified is large and the misrecognition rate is high. It has an extremely low false positive rate. For example, if an output of 0.7 or more is obtained, it is assumed that it is determined that the data belongs to the set category, and the conventional method is applied to data that does not actually belong to any of the set categories 1 and 2. Then, except for the data numbers 2 and 3, the other patterns are identified as belonging to the category 1 or 2. On the other hand, when the pattern matching device according to the present embodiment is used, the pattern On the other hand, the output value is 0.36 or less, and it is determined that the output value does not belong to the learned category. Therefore, the correctness rate of the data that does not belong to the learned category is 100%, which is higher than the conventional method.

In the above embodiment, the boundary pattern is created by using the maximum value of the elements of the initial learning input signal plus a predetermined value and the minimum value of the elements minus the predetermined value. However, it is also possible to obtain an area such as a circle including the initial learning input signal belonging to the same category and use the points outside the area as the boundary pattern.

[0061]

As described above, according to the present invention,
By learning the output of each unit of the output unit group to be “0” or a small value when a boundary pattern around it, which is outside the category, is input to the input unit group, the result is that The output of the output unit can be set to a small value even for patterns that are far from the category, and it can be easily and accurately determined whether or not the input pattern belongs to the category, and erroneous determination is eliminated. It becomes possible.

[Brief description of the drawings]

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

FIG. 2 is an explanatory diagram of a unit according to the embodiment of the present invention.

FIG. 3 is a circuit configuration example of a conversion processing unit according to the embodiment of the present invention.

FIG. 4 is a configuration block diagram of a pattern boundary setting unit provided in a learning signal holding unit according to the embodiment of the present invention.

FIG. 5 is a diagram in which learning data, boundary data, and data used for testing are plotted on xy coordinates.

FIG. 6 shows a schematic diagram of a neural network element (unit) and a sigmoid function which is a kind of input / output function used in this unit.

FIG. 7 is a schematic diagram of a network structure example, (1) is a hierarchical network, and (2) is a mutual connection type network.

FIG. 8 is a schematic image of a three-layer hierarchical network.

FIG. 9 is an explanatory diagram of conventional pattern matching.

[Explanation of symbols]

 10 image receiving device 12 pattern input unit 14 conversion processing unit 16 input layer 18 intermediate layer 20 output layer 22, 25 units 23, 24, 27 coupling element 26 pattern output unit 28 memory 30 display device 32 error detection unit 34 teacher device 50 learning signal Holding unit 54 Pattern boundary setting unit 60 Maximum value / Minimum value calculating unit 64 Boundary signal combining unit 66 Boundary teacher signal combining unit

Claims (3)

[Claims] 1. A plurality of learning input patterns are input to the input unit group, and an output pattern of the output unit group corresponding to the input learning input pattern is compared with a predetermined teacher pattern to obtain an error between them. , The learning input pattern is learned by repeating the operation of correcting the threshold value of the unit forming each unit group and the weighting factor of the coupling element connecting each unit group based on this error. After that, in the pattern matching method of inputting an arbitrary pattern to the input unit group and comparing the output pattern of the output unit group with the output pattern corresponding to the learned input pattern, the plurality of learnings belonging to the same category Create a boundary pattern around the input pattern for learning based on the learning input pattern, and correspond to this boundary pattern. The teacher signal output from each unit of the output unit group is set to 0 or a small value, the boundary pattern is learned together with the learning input pattern, and then an arbitrary pattern is input to the input unit group. Characteristic pattern matching method. 2. The boundary pattern is predetermined from a maximum boundary value obtained by adding a predetermined value to the maximum value of each element of the plurality of learning input patterns belonging to the same category, and the minimum value of each element. 2. The pattern matching method according to claim 1, further comprising a minimum boundary value obtained by subtracting the calculated value. 3. An input unit group to which a pattern to be identified is input, an output unit group that outputs a pattern corresponding to the input pattern, and at least an interposition between the input unit group and the output unit group. One or more intermediate unit groups, a coupling element that connects these unit groups, a plurality of learning input patterns to be input to the input unit group, and the output created corresponding to these learning input patterns A learning signal holding unit that holds an output pattern of the unit group and a teacher pattern to be compared, and each unit of each unit group adds the output of the preceding unit multiplied by the weighting factor assigned to each coupling element. And outputs a calculation result based on an input / output function in which the difference between the added value and the set threshold value is an independent variable, and the learning input pattern Based on an error between the output pattern of the output unit group and the corresponding teacher pattern obtained by inputting to the input unit group, the learning is performed by repeatedly correcting the weighting factor of the coupling element and the threshold value. In the pattern matching device configured to perform the pattern matching by inputting an arbitrary pattern to the input unit group after learning the input pattern for learning, the learning signal holding unit is based on the plurality of learning input patterns. , A boundary signal synthesizing unit for generating a boundary pattern around a learning input pattern belonging to the same category, and a teacher signal corresponding to the boundary pattern and the output of each unit of the output unit group is 0 or a small value. And a boundary teacher signal synthesizing unit.