JPH06139362A - Pattern discriminating device - Google Patents

Abstract

PURPOSE:To enable the discrimination of input patterns forming complicated distribution by properly using the discrimination with the use of a distance and the discrimination with the use of a neural circuit. CONSTITUTION:Distances are determined for every category between an input pattern 400 and a pattern 600 for decision prepared every category by a distance calculation means 200. If the minimum value of the distances of all the categories is less than a first threshold, the processing in the distance calculation means 200 is executed, and if the minimum value is more than the first threshold, the processing in the neural circuit 300 is executed. A means selecting this processing is a first decision means 100. The discrimination with the use of a distance and the discrimination with a neural circuit 300 which are obtained for the input pattern 400 in this way are selectively adopted in accordance with the distance of the input pattern 400 from the center of distribution. Namely, the discrimination with the use of the distance and discrimination with the neural circuit 300 are properly used.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern discriminating apparatus, and more particularly to a pattern capable of discriminating a pattern having a complicated distribution by using discrimination using a distance between a standard pattern and an input pattern and discrimination by a neural circuit. Regarding an identification device.

[0002]

2. Description of the Related Art As a conventional pattern identification method, a method using a distance between a standard pattern and an input pattern is known (see, for example, "Introduction to Character Recognition" (Telecommunications Association)).

Here, a method using the distance will be described. The standard patterns of N categories are M ′ _i (i = 1,
2, ..., N), x ′ is an unknown pattern, and each is defined by an n-dimensional space vector, M ′ _i = (m _i1 , m _i2 , ..., m _in ) ^T (1) x ′ = (x ₁ , x ₂ , ..., X _n ) ^T (2), where T is a symbol representing transposition. The distance d _i between the standard pattern M ′ _i and the unknown pattern x ′ is represented by d _i = D (M ′ _i , x ′) (3). As the D, Euclidean distance D _y , city block distance D _c , Mahalanobis distance D _M, etc. are known, but any may be used.

In order to identify the input pattern, the distance distance d i for each i is calculated, the category for the minimum distance d _i among N _i is calculated, and this is used as the identification result. Has been taken.

In addition to the above distance calculation, a neural circuit is one of the methods recently adopted for pattern identification (for example, MIT Press, Paral).
(See lelDistributed Processing).

FIG. 5 is a schematic view of an example of a conventional apparatus for performing discrimination using a neural circuit.

The neural circuit 4 is a neural circuit 4 having a three-layer structure consisting of an input layer, an intermediate layer and an output layer, and there is a connection only from the element in the lower layer to the element in the upper layer. There are no interlaced connections and connections between elements in the same layer. In the figure, the element 47 of the m-th intermediate layer is (weight vector w _'m), the load 44 in the element 47 of the n-th output layer (load vector u' _n). Element 47 of the intermediate layer
Has a threshold value of 42 (σ _m ) but the output layer element 47 has a threshold value of 4
It is assumed that 5 (ρ _n ) is set. Further, the input pattern 101 input to the neural circuit 4 and the output vector 401 having the output 46 of the neural circuit 4 as an element,
Teacher signal vector 402 for indicating desired output
There is.

The input / output relationship in each layer in the neural circuit 4 is expressed by equations (4) and (5). The discrimination rate of the intermediate layer element is Y _m = f (w ′ _m · x ′ + σ _m ) (4), and the discrimination rate of the output layer element is Z _n = f (u ′ _n · y ′ + ρ _n ) ( 5) is represented. Here, f is a monotonically increasing saturated nonlinear function, “·” is an inner product between vectors, and vector y ′ is a vector having y _m as an element. Learning is input pattern 10
In order to minimize the squared error between the output vector 401 having the output 46 from the output layer obtained by processing 1 according to equations (4) and (5) and the teacher signal vector 402, each element 4
7. Adjust the loads 41 and 44 of 7. After the adjustment of the load is completed, the input pattern 10 can be identified in order to identify the input pattern 101.
A method is often adopted in which the processing of equations (4) and (5) is performed on 1 to find the category corresponding to the maximum output Z _n among N pieces and use this as the identification result. This method has an advantage that a high degree of discrimination is possible because a piecewise linear discrimination function is formed inside the neural circuit 4.

[0009]

However, the conventional pattern identification method by distance calculation has advantages such as easy processing and easy expansion of the number of standard patterns, but has a slightly complicated distribution. There is a disadvantage that it is difficult to identify the pattern. In order to solve this, there is a method of converting the standard pattern M ′ _i into multiple templates, but if the number of templates exceeds the optimum value, patterns that were correctly identified so far cannot be identified due to the newly added template. There is a limit to the method itself.

Further, the conventional method for identifying patterns by the neural circuit shown in FIG. 5 has a disadvantage that the learning time becomes longer as the number of patterns increases.

Further, under the present circumstances in which an optimal neural circuit design method has not been established, a method of performing a discrimination experiment by changing the configuration of the neural circuit and learning parameters and adopting the neural circuit having the highest discrimination rate (for example, , IEICE Technical Report PRU88-58, pp79-86), and the time and labor required for the experiment is extremely large.

As described above, the conventional pattern identification using the distance makes it easy to create a standard pattern, but it cannot identify a complicated distribution.
Discrimination using a neural circuit is suitable for discriminating a complicated distribution, but has a problem that the load of creating a neural circuit becomes large.

The present invention has been made in view of the above points,
It is an object of the present invention to provide a technique for easily configuring a pattern identification device capable of identifying a pattern having a complicated distribution.

[0014]

FIG. 1 is a diagram for explaining the principle of the present invention. The present invention provides an input pattern 400
Between the input pattern 400 and the standard pattern, the distance calculation means 200 for obtaining the distance representing the closeness between the input pattern 400 and the standard pattern 500 for each standard pattern, and the input pattern 400 from a plurality of neural circuit elements, and each input pattern 4
The element for obtaining one output by applying a saturation function to the value obtained by multiplying the load corresponding to the element of 00 and arranging the elements is arranged in layers, and the identifying means by the neural circuit 300 in which the layers are multilayered is included. In the pattern identification device, the distance calculation means 200
The distance is calculated for each category between the input pattern 400 and the determination pattern 600 prepared for each category according to, and if the minimum value of the distances of all the categories is less than the first threshold value set in advance, the distance calculation means 200 The first determining means 1 for executing the process and selecting the process to execute the process in the neural circuit 300 if the minimum value is equal to or larger than the first threshold value.
00, and the distance calculation means 20 by the first determination means 100.
When the processing of 0 is executed, the minimum value of the distance is less than or equal to the predetermined second threshold value, and the difference between the next smallest value and the minimum value of the processing is greater than or equal to the predetermined third threshold value. When the two conditions are satisfied, the category corresponding to the minimum value is used as the identification result, and when the two conditions are not satisfied, the second determination means 201 and the first determination means 1 that reject the identification result.
When the processing in the neural circuit 300 is executed according to the determination result of 00, the maximum value of the output of the output device of the multilayered neural circuit 300 is equal to or greater than a predetermined fourth threshold value, and the maximum value and the next maximum value When the two conditions that the difference is greater than or equal to the predetermined fifth threshold value are satisfied, the category corresponding to the element that outputs the maximum value is set as the identification result, and when the two conditions are not satisfied, the identification result is rejected. 3 determination means 301.

[0015]

The present invention selectively employs the discrimination using the distance obtained for the input pattern and the discrimination using the neural circuit according to the distance of the input pattern from the center of the distribution.
This makes it possible to distinguish between the identification using the distance and the identification using the neural circuit, so that the identification using the distance is in charge of identifying the pattern existing in the portion relatively close to the center of the distribution. However, for other patterns, since the discrimination using the neural network is performed, it is possible to discriminate the complicated distribution. In addition, since the number of patterns in charge of the neural circuit is reduced by sharing the duties, the time required for learning is shortened and the load of neural circuit creation is reduced as compared with the case where all neural circuits are identified.

[0016]

Embodiments of the present invention will be specifically described below with reference to the drawings. FIG. 2 is a block diagram showing the configuration of the pattern identification apparatus according to an embodiment of the present invention. The pattern identifying apparatus shown in FIG. 1 uses the distance and the neural circuit separately for identification.

The pattern identifying apparatus of FIG. 2 determines whether to identify using the distance between the determination pattern output from the determination pattern memory unit 102 and the input pattern 101 or the neural circuit 4. Identification method determination unit 1,
When performing discrimination based on the distance based on the determination result of the identification method determination unit 1, the distance between the standard pattern input from the standard pattern memory unit 202 and the input pattern 101 is calculated,
The distance calculation unit 2 that outputs the output vector 201 and the output vector 201 output from the distance calculation unit 2 are compared with a preset threshold value to determine the identification result, and the identification result is stored in the identification result memory unit 6. On the other hand, the determination unit 3 that outputs, on the other hand, when the identification method determination unit 1 determines to perform the identification by the neural circuit 4, the determination is performed using the neural circuit 4 that outputs the output vector 401 and the output vector 401 from the neural circuit 4. And a determination unit 5 for outputting the identification result to the identification result memory unit 6, and a control unit 7 for controlling each of the above units. In the figure, the signal indicated by the broken line is a control signal from the control unit 7.

The identification method determination unit 1 is a main element of the present invention, and determines whether the identification processing is performed on the input pattern 101 by the distance calculation unit 2 or the neural circuit 4. . Here, details of the operation of the identification method determination unit 1 will be described.

FIG. 3 is a diagram for explaining the operation of the identification method determination unit according to the embodiment of the present invention. 2, those parts which are the same as those corresponding parts in FIG. 2 are designated by the same reference numerals, and a description thereof will be omitted.

The identification method determination unit 1 is composed of a distance calculation circuit 11, a minimum value detection circuit 12, a comparison circuit 13, a switch 14, a terminal 15 and a terminal 16.

Among them, the distance calculation circuit 11 is arranged to determine the vector x'of the input pattern 101 and the judgment pattern memory 10
The distance g _i = D from the vector S ′ _i (i = 1, 2, ..., N) of the judgment pattern of each category read from No. 2
Calculate (S ' _i , x').

The minimum value detection circuit 12 is a distance calculation circuit 11
The minimum value is detected from the N distances g _i calculated in
This is the minimum distance value g _min .

The comparison circuit 13 compares a predetermined threshold value T ₁ and the minimum distance value g _min, and the minimum distance value g _min is the threshold value T ₁
If it is smaller (g _min <T ₁ ), the switch 14 is tilted to the terminal 15 side, and if the minimum distance value is greater than or equal to the threshold value T ₁ (g _min
min ≧ T ₁ ) and tilt the switch 14 to the terminal 16 side. The terminal 15 is connected to the distance calculation unit 2 and the terminal 16 is connected to the neural circuit 4, and when a signal flows to the terminal 15,
When the distance calculation unit 2 executes the next process and the signal flows to the terminal 16, the neural circuit 4 executes the next process.

When a signal flows to the terminal 15, the distance calculation unit 2 reads the vector x ′ of the input pattern 101 and the standard pattern M ′ (i = 1, 2, ..., N) of each category read from the standard pattern memory unit 202. ) With the distance d _i ,
A vector having the distance d _i as an element is set as an output vector 201. The determination unit 3 detects the minimum value d _min and the next small value d _2-min from the elements of the output vector 201, compares them with the predetermined thresholds θ ₁ and θ _2, and compares the following two conditions d _min ≦ θ ₁ (6) d _2-min −d _min ≧ θ ₂ (7) If both are satisfied, the category corresponding to the minimum value d _min is sent to the identification result memory unit 6 as the identification result. If at least one of the equations (6) and (7) is not satisfied, it is rejected and sent to the identification result memory unit 6.

On the other hand, when a signal flows through the terminal 16 and the neural network 4 is used for identification, the neural circuit 4 uses a vector whose element is the output of the uppermost layer element obtained by processing the input pattern 101. As an output vector Z′401.
Here, the neural network 4 is one of the presented learning patterns.
It is assumed that the learning method that has passed through the terminal 16 side in the identification method determination unit 1 is sufficiently learned using the teacher signal vector t′402.

The decision section 5 determines the output vector Z'401.
Maximum value g from prime_maxAnd the next maximum value g _2-maxAnd detect
Predetermined threshold η₁, Η₂Compared with the following two conditions g_max ≧ η₁ (8) g_max-G_2-max≧ η₂ If both (9) are satisfied, the maximum value g_maxCorresponding to
Is sent to the identification result memory unit 6 as the identification result. formula
When at least one of (8) and (9) is not satisfied
Is rejected and sent to the identification result memory unit 6. This
The input pattern 101 is identified as follows.
It

FIG. 4 is a flow chart showing the flow of the operation of the embodiment of the present invention.

Step 1: The input pattern 101 is input to the pattern identification device.

Step 2: The distance calculation circuit 11 of the identification method determination unit 1 reads the determination pattern from the determination pattern memory unit 102 and calculates the distance between the vector of the determination pattern of each category and the input pattern 101. It is sent to the minimum value detection circuit 12.

Step 3: The minimum value detection circuit 12 determines the minimum distance value g _min from the distance input from the distance calculation circuit 11.
Is detected and transmitted to the comparison circuit 13.

Step 4: The comparison circuit 13 compares the minimum distance value g _min with the predetermined threshold value T ₁ to obtain the minimum distance value g _{min.
If min} is greater than or equal to the threshold T ₁ , the process proceeds to step 5, and if the minimum distance value g _min is smaller than the threshold T ₁ , the process proceeds to step 6.

Step 5: The distance calculator 2 calculates the distance between the vector of the input pattern 101 and the standard pattern read from the standard pattern memory 202, compares it with a predetermined threshold value, and when a predetermined condition is satisfied, Outputs the category corresponding to the minimum distance value to the identification result memory unit 6 as the identification result.

Step 6: The neural network 4 obtains the discrimination result by applying the weight corresponding to the element of the input pattern by the teacher signal vector 402 and operating the saturation type function on the summed value, and discriminates the discrimination result. The result is output to the memory unit 6.

All the learning / identification processing described above is performed by a signal from the control unit 7.

In this embodiment, the learning rule is assumed to be backpropagation learning, but any learning rule may be used as long as it is a method of using the output vector 401 and the teacher signal vector 402. In addition, in the present embodiment, the identification method determination unit 1
Has shown the case of making a determination to separate two types of identification methods, but it is clear that the identification method determination unit 1 of the present invention can be applied to the case of three or more types, which increases the number of identification methods. Therefore, it can be easily expanded.

Although the present invention has been specifically described based on the embodiments, it is needless to say that the present invention is not limited to the above embodiments and various modifications can be made without departing from the spirit of the invention. Yes.

[0037]

As described above, according to the present invention, an input pattern having a complicated distribution can be identified by properly using the identification using the distance and the identification using the neural circuit.

Since the distance calculation means and the neural circuit share the identification, the load on the neural circuit is small,
Makes it easier to create neural circuits.

[Brief description of drawings]

FIG. 1 is a diagram for explaining the principle of the present invention.

FIG. 2 is a block diagram showing a configuration of a pattern identification device according to an exemplary embodiment of the present invention.

FIG. 3 is a diagram for explaining an operation of an identification method determination unit according to an exemplary embodiment of the present invention.

FIG. 4 is a flowchart showing a flow of operation of the embodiment of the present invention.

FIG. 5 is a schematic view of an example of a device for performing discrimination using a conventional neural circuit.

[Explanation of symbols]

 1 Identification Method Judgment Section 2 Distance Calculation Section 3 Judgment Section 4 Neural Circuit 5 Judgment Section 6 Identification Result Memory Section 7 Control Section 11 Distance Calculation Circuit 12 Minimum Value Detection Circuit 13 Comparison Circuit 14 Switch 15, 16 Terminal 41 Load 42 Intermediate Layer Threshold 43 Output of intermediate layer 44 Weight 45 Output layer threshold 46 Output layer output 47 Element 101 Input pattern 201 Output vector 202 Standard pattern memory unit 401 Output vector 402 Teacher signal vector

 ─────────────────────────────────────────────────── ─── Continued Front Page (72) Inventor Katsunori Shimohara 1-1-6 Uchisaiwaicho, Chiyoda-ku, Tokyo Nihon Telegraph and Telephone Corporation

Claims (1)

[Claims] 1. Between an input pattern and a standard pattern,
Distance calculating means for obtaining the distance representing the closeness between the input pattern and the standard pattern for each standard pattern, and taking in the input pattern from a plurality of neural circuit elements, multiplying the load corresponding to the element of each input pattern In a pattern identification device including an element for obtaining a single output by operating a saturated function on the summed value in a layered manner, and an identification means by a neural circuit in which the layers are multilayered, the distance calculation means , The distance is calculated for each category between the input pattern and the determination pattern prepared for each category, and the minimum value of the distance of all categories is set to a predetermined first value.
If it is less than the threshold value of, the processing in the distance calculation means is executed, and if the minimum value is greater than or equal to the first threshold value, the first determination means selects processing to execute the processing in the neural circuit. When the processing of the distance calculating means is executed by the first determining means, the minimum value of the distance is less than or equal to a second threshold value set in advance, and the next small value and the minimum value of the processing are When the two conditions in which the difference between the two is greater than or equal to a third threshold value set in advance are satisfied, the category corresponding to the minimum value is set as the identification result, and
When the two conditions are not satisfied, a second determination unit that rejects the identification result, and an output of the output device of the multilayered neural circuit when the processing in the neural circuit is executed by the first determination unit Is greater than or equal to a predetermined fourth threshold value, and the difference between the maximum value and the next maximum value is greater than or equal to a predetermined fifth threshold value 2
When one of the two conditions is satisfied, a category corresponding to the element that outputs the maximum value is used as a discrimination result, and when the two conditions are not satisfied, a third determination unit that rejects the discrimination result is provided. Pattern identification device.