JPH04280385A - Neural network - Google Patents

Abstract

PURPOSE:To recognize a pattern to be checked even when correlative output values have variance by normalizing values corresponding to the optical intensity of correlative outputs according to values corresponding to the number of picture elements constituting the pattern to be checked or the same pattern and more than one reference patterns. CONSTITUTION:One pattern to be checked and plural reference pattern groups drawn at an image display device 16 are simultaneously read out by light flux 12 from a laser 11. The light flux 12 is made incident on a beam splitter 15 and split into light fluxes 17 and 18. The light flux 18 forms a congruent Fourier transformation pattern on a screen 31 by a Fourier transforming lens 21. This pattern is read in by a two-dimensional photoelectric conversion element 32 and drawn on a liquid crystal light value 35 by a liquid crystal driving circuit 33. These patterns are read out by light flux 37, and the optical intensity of correlative outputs through a Fourier transforming lens 41 onto a screen 42 is detected by a conversion element 43 and stored in a computer 51. On the other hand, the light flux 17 forms an image on a screen 53 by a lens 54, and the image is read in the computer 51 by a conversion element 52.

Description

[Detailed description of the invention]

0001

[Industrial Field of Application] The present invention relates to a pattern used for recognition of objects, characters, sounds, etc., and information retrieval, inference, association, etc.
This paper relates to a neural network that performs key recognition.

0002

2. Description of the Related Art In recent years, neural network technology has been attracting attention as a technology for effectively performing information processing, such as recognition, association, and inference, which are considered to be weak points of von Neumann computers. In particular, we use the error backpropagation (hereinafter referred to as BP) method, which is a learning method for neural networks, to generate a signal with usually multiple inputs and one output modeled on a unit in the network, that is, a neuron. By determining the coupling weight between the processing elements, flexible recognition that captures the characteristics between each input pattern can be performed.

[0003] A value corresponding to the correlation output light intensity for each reference pattern obtained by performing a correlation calculation between the test pattern to be optically recognized and classified and each reference pattern is input. Figure 2 shows a typical conventional neural network as described above.
FIG. That is, in a neural network, especially a neural network that performs learning using the error backpropagation (hereinafter referred to as BP) method, normally,
A set of units has a layered structure of three or more layers, of which the layer for inputting pattern signals is called an input layer, the layer for outputting recognition output signals is called an output layer, and the remaining layers are called intermediate layers or hidden layers. Here, the number of layers in the middle layer is 1, the number of units included in the input layer is 5, the number of units included in the middle layer is 3, and the number of units included in the output layer is 5. This is just one example of how to do this, and the number of intermediate layers and the number of units in each layer can take various values depending on requirements.

Each input value, that is, a value corresponding to the correlation output light intensity obtained from the optical correlation system, is outputted as an output from the input layer as it is, inputted to the intermediate layer, and then to the output layer, Output from the output layer. The input value to each unit of the intermediate layer and the input value to each unit of the output layer are calculated by multiplying the output value from each previous layer by the connection weight value corresponding to the connection, and calculating the value obtained by multiplying the output value from each previous layer by the connection weight value corresponding to the connection. If the total sum plus the bias value is x, then the input value is expressed as F(x). It is expressed as F(x)=1/[1+exp(-x)] and is a nonlinear function with a monotonous increasing tendency.

[0005] Furthermore, the coupling weight values and bias values between the input layer and the intermediate layer and between the output layer and the intermediate layer are independent values. In the configuration of a neural network as described above, a certain input pattern, that is, an output vector of an input layer unit (I1, I2, I3, I4,
By modifying the connection weight value between each unit so that the sum of squares of the error of each unit between the desired output from the output layer and the actual output when given I5) is small,
It is possible to obtain a neural network that outputs a desired output pattern in response to a certain input pattern. At this time, the BP method can calculate the load modification coefficient,
Several methods have been proposed to efficiently converge learning using the correction coefficients.

[0006] Conventionally, optical correlation systems for obtaining correlated output light intensity include a joint transform method, a matched filter method, a joint transform method using a lens array, etc. However, in either case, a spatial light modulator or photographic film is used as an image display device to optically display the test pattern to be recognized and classified and each reference pattern.

However, in a neural network that inputs a value corresponding to the optically obtained correlation output intensity as described above, the image display of the optical correlation system, which is the preprocessing of the neural network, is difficult. The test pattern is displayed at time T1 in the device.
The correlation output light intensity for each reference pattern obtained when displaying the pattern and the time T2 different from T1 on the image display device.
The correlation output light intensities for each reference pattern obtained when the same test pattern was displayed on the display had slightly different values. This is due to very large fluctuations in transmitted or reflected light from a spatial light modulator used as an image display device, or fluctuations due to speckle noise.

[0008] Furthermore, when a defective pattern of a test pattern is displayed on an image display device, the correlation output light intensity for each reference pattern obtained is This value is smaller than the correlation output light intensity for each reference pattern obtained when For this reason, new
The input pattern input to the physical network will be completely different when the test pattern is displayed and when the missing test pattern is displayed. The news that was trained only when displaying
In the ral network, erroneous recognition occurred regarding missing test patterns. Even if training is performed so that it can recognize missing test patterns, it is necessary to learn both the missing test patterns and the non-defective test patterns, which requires a lot of learning time. I was spending a lot of time.

It is also assumed that in a neural network with a fixed configuration, the number of categories that can recognize and classify defect-free test patterns is fixed. If you train it to recognize missing test patterns and try to attribute the missing test patterns to a certain category, the number of categories will be The number of categories decreased compared to the number of categories that could be recognized, and the recognition performance of the neural network deteriorated.

[Problem to be solved by the invention]

The present invention has been made to solve the above-mentioned problems. Even if there are variations in the correlation output light intensity, it is possible to perform recognition, and it is possible to recognize and classify missing test patterns by minimizing the decrease in the number of categories that can recognize and classify test patterns. The purpose of the present invention is to provide a neural network that has high performance and requires less learning time.

[0011]

[Means for Solving the Problems] The present invention has been made to solve the above-mentioned technical problems, and it is an object of the present invention to increase the number of layers in a neural network whose basic structure is hierarchical. The input layer has three or more layers, and the input layer receives the correlated output light intensity of each reference pattern obtained by performing correlation calculations between the test pattern to be optically recognized and classified and each of the plurality of reference patterns. Input the corresponding value normalized by the value corresponding to the number of pixels constituting the test pattern or the test pattern and one or more reference patterns, and calculate the number of pixels from the input layer to the output layer. A neural network is provided in which each connection weight value is determined by supervised learning.

[0011] The value corresponding to the number of pixels constituting the test pattern is preferably a value that corresponds to the intensity of transmitted light or reflected light from the image display device that displayed the test pattern. . Furthermore, the value corresponding to the number of pixels constituting the test pattern and one or more reference patterns is the value corresponding to the number of pixels constituting the test pattern and one or more of the reference patterns. A new value that is the sum of values according to the transmitted light or reflected light intensity with respect to the pattern.
ral network is preferred. Furthermore, it is preferable that the correlation output light intensity for each reference pattern is determined by a neural network that takes the peak value of the autocorrelation output or cross-correlation output between the test pattern and each reference pattern. be. The correlation output light intensity for each reference pattern is determined by the total light intensity of the light receiving range or the average light intensity depending on the spread of the autocorrelation output or cross-correlation output between the test pattern and each reference pattern. -ral network is preferred. Furthermore, the correlation output light intensity for each reference pattern is calculated based on the peak value of the autocorrelation output or cross-correlation output between the test pattern and each reference pattern, and the light receiving range according to the spread of the correlation output. A neural network based on taking the total amount of light or the average amount of light is suitable.

The correlated output light intensity for each reference pattern is 2
It is preferable that the linear range of the dimensional photoelectric conversion element or a part of the correlated output light intensity exceeds the maximum amount of light received by the two-dimensional photoelectric conversion element. Furthermore, the correlation calculation between the test pattern to be optically recognized and classified and each of the plurality of reference patterns is performed using a neural network obtained by using a filter based on a Fourier transform hologram. is suitable. Correlation calculations between the optically recognized and classified test pattern and each of the multiple reference patterns can be obtained by individually performing joint Fourier transform between the test pattern and each reference pattern. Either optically perform Fourier transform on the intensity patterns that have been detected individually again, or perform joint Fourier transform on the test pattern and each reference pattern all at once. It is preferable to use a neural network in which the intensity pattern obtained is optically subjected to Fourier transformation all at once. Furthermore, the correlation calculation between the test pattern to be optically recognized and classified and each of the multiple reference patterns is as follows:
The test pattern represented by the light transmittance distribution or the light reflectance distribution is overlapped with each reference pattern, irradiated with incoherent light, and the amount of light reflected or transmitted is measured. Neural nets such as these are preferred.

[0013]

[Operation] According to the configuration of the neural network of the present invention, a value corresponding to the correlation output light intensity obtained due to fluctuations in transmitted light or reflected light from an image display device of an optical correlation system is obtained. Even if there is variation, the value according to the correlation output light intensity is normalized by the value corresponding to the number of pixels forming the test pattern or the test pattern and one or more reference patterns. As a result, the input patterns of the neural network for the test pattern and the missing test pattern are almost the same, and there is no decrease in the number of categories that can recognize and classify the test pattern. The performance for recognition classification of missing test patterns is also improved. Further, there is no need to separately learn the missing test pattern, the learning time is short, and the missing test pattern can be recognized relatively easily and efficiently.

Next, the neural network of the present invention will be specifically explained using examples, but the present invention is not limited thereto.

[0015]

Embodiment 1 FIG. 1 is a schematic configuration diagram showing a specific configuration of an example of a neural network according to the present invention. Here, the part of the neural network made up of units is omitted as it is well known, and here we will first focus on the part of the optical correlation system that performs preprocessing of the neural network. A detailed explanation will be given of the luminous flux of the optical correlation system. Therefore, FIG. 1 shows a schematic block diagram of an optical correlation system for preprocessing a neural network.

In the schematic diagram of the optical correlation system shown in FIG. 1, the optical correlation system includes image output means 1, optical Fourier transform means 2, image output means 3, and optical Fourier transform means 4. , and a photodetector means 5.

Next, the above configuration will be briefly explained. One test pattern to be recognized and classified drawn on the image display device 16 and a plurality of reference pattern groups are simultaneously read out with the coherent light beam 12 emitted from the laser 11 and passed through the image display device 16. The light beam 12 then passes through a beam splitter.
15 and is divided into a beam of light 17 and a beam of light 18. The light beam 18 is screened by the Fourier transform lens 21.
A congruent Fourier transform pattern is created on the curve 31.

This congruent Fourier transform pattern is read by a two-dimensional photoelectric conversion element 32, drawn on an electrically addressed liquid crystal light valve 35 by an image processing and liquid crystal drive circuit 33, read out by a light beam 37, Rie conversion lens 41
Then, Fourier transform is performed again, and the correlated output light intensity on the screen 42 is detected by the two-dimensional photoelectric conversion element 43 and stored in the computer 51. On the other hand, the luminous flux 17 is
Images of the test pattern and a plurality of reference pattern groups are formed on the screen 53 by the lens 54.

The test pattern imaged on the screen 53
The light intensity of each of the reference pattern groups is read by the two-dimensional photoelectric conversion element 52 at substantially the same timing as the two-dimensional photoelectric conversion element 43, and is stored in the computer 51. Here, the memory in the computer 51 is
The correlated output light intensities of the tested pattern and each reference pattern, and the light intensities of the tested pattern and each of a plurality of reference pattern groups by the light beam 17 are calculated by the computer 51. However, what is stored in the memory is not limited to these values, but values corresponding to the obtained light intensity, such as CCD, which expresses the received light intensity in 256 8-bit gradations. It also stores items that receive light intensity linearly, such as photodetectors.

Next, a method of determining the correlation output light intensity between the test pattern and each reference pattern to obtain the correlation output will be explained. For example, if five character patterns A, E, N, T, and V are selected as the reference image group, and the character pattern to be recognized and classified is H, as shown in FIG.
The pattern you want to recognize and classify by placing two character patterns on concentric circles.
Suppose that the image is drawn on the image display device 16 at the center of concentric circles. In this case, the output of the two-dimensional photoelectric conversion element 43, that is,
The correlation output between the character pattern H and each reference pattern appears at each corresponding reference pattern position drawn on the image display device 16, and is detected as the correlation output light intensity as shown in FIG. be done. Here, the larger the peak value of the correlation output light intensity, the larger the circle mark. In addition, here, the reference patterns are 5 patterns A, E, N, T, and V, but this is to simplify the explanation, and the reference patterns and their numbers are A, E, N, T, and V. It is not limited to only E, N, T, and V, or only to five patterns.

The peak value of the correlation output light intensity obtained as described above is stored in the computer 51. Note that the autocorrelation appearing on the optical axis in FIG. 4 is omitted here because it is irrelevant.

Now, generally speaking, the correlation output consists of two patterns.
When the numbers are A(x, y) and B(x, y), it is expressed by the formula shown below. I(x', y')=∬A(x,y)B*(X'-X,Y
'-Y) dxdy (1) Here, * represents a complex conjugate amount. As can be seen from equation (1), the correlation output has a spread that is twice the size of the pattern in the one-dimensional direction.Therefore, in addition to the above-mentioned peak value, the correlation output is Even if we take the total amount of light in the light receiving range according to the spread of the correlation output, or the average amount of light,
It can be treated similarly as information representing the degree of correlation between patterns.

FIG. 5 schematically shows how the correlation light amount distribution looks, and the extent to which the two patterns overlap is shifted in the vertical direction of the paper. At this time, the correlated light amounts of patterns E and H reach a peak value when they completely overlap, as shown in Figure 5(a), and the magnitude of the patterns increases as shown in Figure 5(b). When shifted by that amount, one vertical bar of pattern H and one vertical bar of pattern E overlap, so that a correlation light amount smaller than the peak value is obtained. Figure 5(c) shows the correlation light amount distribution of pattern E and pattern H created in this way, where the large circle in the center is the peak amount, and the small circles at the top and bottom are the horizontal bars. The small horizontal circles indicate the reactions in the vertical bars. Therefore, information is obtained in which the characteristics of each pattern, which are not seen in the peak light amount, are better reflected in the total light amount or average light amount.

Therefore, when detecting the correlated output light intensity using a two-dimensional photoelectric conversion element, the peak value of the correlated output light intensity is received in a range exceeding the maximum light amount, and the weak portion of the correlated output light intensity is detected. By emphasizing the characteristics of each pattern that are not seen in the peak light amount, it is possible to obtain information that is better reflected.

Next, the correlation output light intensity between the test pattern and each reference pattern stored in the computer 51 is determined by the correlation output light intensity between the test pattern and each reference pattern by the light beam 17. The light intensity of each pattern group is normalized by the computer 51 and input to each unit of the input layer of the neural network.Here, the normalization method described above will be explained. do.

Standardization method a includes a method in which the correlation output light intensity of the test pattern and each reference pattern is normalized by the light intensity of the test pattern using the light beam 17. As normalization method b, the correlation output light intensity of each of the test pattern and each reference pattern is determined by the correlation output light intensity of one or more of the test pattern and each reference pattern. There is a method of standardizing each by the sum of values corresponding to the intensity of transmitted light or reflected light with respect to a reference pattern. Of these two normalization methods, the former normalization method a
This is a very effective means when recognition is performed using a physical network, and when the light passing through or reflecting the test pattern is fluctuating.

Suppose that the correlated output light intensities of the defective or nondefective test pattern H with respect to the reference patterns A, E, N, T, and V are I(ha), I(he), and I(hn ), I(h
t), I(hv), and the two-dimensional photoelectric conversion element 5
2. The amount of transmitted light or reflected light of the defective or defect-free test pattern H is defined as I(h). In the neural network of the previous application, each of the above-obtained correlated output light intensities I(ha), I(he), I(hn), I
(ht) and I(hv) were input as they were, supervised learning was performed, and pattern recognition was performed.

However, normally, the correlation output light intensity obtained when the test pattern is missing is higher than the correlation output light intensity obtained when the test pattern is not missing. The value will be lower depending on the amount. Therefore, if the above-mentioned low correlation output light intensity was also input to the neural network, in order to recognize the same pattern, a large number of correlation output lights must be used to take into account the case where the test pattern is missing. The intensity had to be input and learned, which required a lot of learning time. It is also assumed that in a neural network with a fixed configuration, the number of categories that can recognize and classify defect-free test patterns is fixed. Then, it is made to learn so that it can also recognize missing test patterns, and the missing test patterns are assigned to a certain category. As a result, the number of categories decreases compared to the number of categories that can recognize defect-free test patterns, and the performance of the neural network in pattern recognition deteriorates.

However, the neural network of the present invention
In this step, the standardized value expressed by the following formula (in normalization method a, the amount of transmitted light or reflected light of the test pattern is normalized) is input to each unit of the input layer, and the Performs guided learning and pattern recognition. I1(ha)=I(ha)/I(h)・・・・・・
・・・・・・・・・・・・(2) I1(he)=I(he
)/I( h)・・・・・・・・・・・・・・・・・・
(3) I1(hh)=I(hh)/I(h)...
・・・・・・・・・・・・・・・(4) I1(ht)=I
(ht)/I(h)・・・・・・・・・・・・・・・
...(5) I1(hv)=I(hv)/I(h)・
・・・・・・・・・・・・・・・・・・(6)

0030
] Next, why are the above (2), (3), (4), (5), and (6)
) The normalized value shown by the formula is applied to the neural network.
By inputting the missing pattern into the neural network, as mentioned above, when the missing pattern is recognized using a neural network,
Next, it will be explained why this is a very effective means when the light transmitted or reflected by the test pattern fluctuates and the obtained correlation output light intensity varies.

For example, the correlated output light intensity obtained from the defect-free test pattern H and the reference pattern A is In(ha)
The amount of transmitted light or reflected light of the defect-free test pattern H is In(h). Now, suppose that part of the test pattern H is missing. Then, the correlation output light intensity of the defective test pattern H that can be obtained at this time with respect to the reference pattern A is less than In(ha).
It is assumed that In(ha) is smaller. In other words, the correlation output light intensity changes by In(h
Assume that it is lower than a) by ΔIn(ha).

It is also assumed that the amount of transmitted light or reflected light of the defective test pattern H that can be obtained at this time is smaller than In(h) by ΔIn(h). That is, it is assumed that the amount of transmitted light or reflected light of the test pattern H is reduced by ΔIn(h) from In(h) by the amount of the defect in the test pattern H. From the above, the correlation output light intensity obtained when the test pattern H is defective is In(ha)-△In(ha)...
...(7) Also, the amount of transmitted light or reflected light of the defective test pattern H obtained when the test pattern H is defective is In(h) - △In(h)...・・・・・・・・・
...(8).

[In(ha)-△In(ha)]/[In(h)-
△In(h)]={In(ha)−△In(ha)}*
{1/In(h)}/[1-{△In(h)/In(h
)}]・・・・(9) Here, In(h)＞△In(h)・・・・・・・・・
・・・・・・・・・(10) always holds true, so (
9) The formula is {In(ha)-△In(ha)}*{1/In(h)
}*[1+{△In(h)/In(h)}]...(
11) It becomes. Considering equation (11), we find that {In(h
a)-△In(ha)}*{1/In(h)}...
(12) represents that the correlation output light intensity obtained when the test pattern H is defective is normalized by the amount of transmitted light or reflected light of the non-defective test pattern H.

Therefore, the equation (11) calculates the correlation output light intensity obtained when the test pattern H has no defects or defects. When normalized by the amount of transmitted light or reflected light of the test pattern H, the correlation output light intensity obtained when the test pattern H is defective is calculated by the amount of transmitted light or reflected light of the non-defective test pattern H. Compared to the normalized value, {In(ha)-△In(ha)}*{△In(h)}
/[In(h)2}...(13) This means that it will be larger. Here, if the correlated output light intensity obtained when the test pattern H is defective or not is always normalized by the amount of transmitted light or reflected light of the non-defective test pattern H, the ratio is , it is always the ratio of the correlated output light intensity when not normalized.

Therefore, as a result of normalization, the ratio of the correlated output light intensity when the test pattern H is defective and when it is not becomes smaller by the amount of the equation (13). become. Therefore, the difference in the correlation output light intensity when the test pattern H is missing and when it is not is almost eliminated by normalization, and is input to the neural network. Therefore, the number of categories that can be recognized and classified, which originally decreased due to the recognition of missing test patterns, will hardly decrease. In addition, the learning time required to form a neural network that can recognize missing test patterns can be reduced because there is no need to separately learn the test patterns. Recognition of defective test patterns can also be performed easily and efficiently.

The above description is the same for other test patterns or reference patterns, and is obvious, so a description thereof will be omitted here. Next, standardization method b will be explained. Standardization method b is a very effective means when recognizing a defective pattern using a neural network, and it is also a very effective means for recognizing defective patterns using a neural network. This takes into account the fluctuation of

The procedure for standardization method b is similar to standardization method a, in which the test pattern is set to H and the reference patterns are set to A, E, N,
When T and V, the correlated output light intensities of the defective or non-defective pattern H with respect to the reference patterns A, E, N, T, and V are I(ha), I(he), and I(hn), I(
ht), I(hv), and is obtained by the two-dimensional photoelectric conversion element 52. Defective or non-defective test pattern H
The amount of transmitted or reflected light is I(h). Also,
In standardization method b, the two-dimensional photoelectric conversion element 52 not only detects the test pattern but also the reference patterns A, E, N, T,
Ir(a) and Ir(e), which are the amount of transmitted or reflected light of V, respectively
), Ir(n), Ir(t), and Ir(v) are also detected and stored in the computer 51.

[0038] Using each value obtained by the optical correlation system as described above, standardization method b performs normalization expressed by the following formula, and performs supervised learning on the normalized value. , are input to each unit of the neural network input layer when performing pattern recognition. I2(ha)=I(ha)/{I(h)+I
r(a)}・・・・・・・・・・・・・・・・・・(14
) I2(he)=I(he)/{I(h)+
Ir(e)}・・・・・・・・・・・・・・・・・・(1
5) I2(hh)=I(hN)/{I(h)
+Ir(N)}・・・・・・・・・・・・・・・・・・(
16) I2(ht)=I(ht)/{I(h
)+Ir(t)}・・・・・・・・・・・・・・・・・・
(17) I2(hv)=I(hv)/{I(
h)+Ir(v)}・・・・・・・・・・・・・・・
・(18)

Then, why does the above (14), (15) (
16) By inputting the standardized values shown by equations (17) and (18) into the neural network, the missing pattern can be calculated.
Next, it will be explained why this method is very effective when a pattern is recognized using a neural network or when the transmitted or reflected light of a reference pattern is fluctuating. Further, here, in order to simplify the explanation, the sum of the test pattern and one transmitted or reflected light quantity is used, but the same treatment is possible even if there is one or more reference patterns.

Similar to the explanation for standardization method a, let In(ha) be the correlation output light intensity obtained between the defect-free test pattern H and the reference pattern A, and The amount of transmitted or reflected light of the defective test pattern H is assumed to be In(h). In addition, the test pattern H may be missing or the reference pattern A may be missing.
When the amount of transmitted or reflected light fluctuates, the correlation output light intensity of the defective test pattern H that can be obtained with respect to the reference pattern A is assumed to be smaller than In(ha) by ΔIn(ha). , Further, the amount of transmitted or reflected light of the defective test pattern H that can be obtained at this time is smaller than In(h) by ΔIn(h).

[0041] Furthermore, in standardization method b, fluctuations in the amount of transmitted or reflected light of the reference pattern are also taken into consideration.
Assume that the average value of the amount of transmitted or reflected light of pattern A is Iy(a), and the amount of transmitted or reflected light of reference pattern A is smaller by ΔIy(a). However, here, reference pattern A
The reference value for fluctuations in the amount of transmitted or reflected light is the average value I
Although y(a) is used here, this is for the purpose of simplifying the explanation, and the reference value is not limited to this. Using the values shown above and normalizing according to equation (14), we get

[In(ha)−ΔIn(ha)]/[{In(h)−
△In(h)}+{Iy(a)−△Iy(a)}] =
[In(ha)−△In(ha)]*[1/{In(h
)+Iy(a)}]/[1-{△In(h)+△Iy(
a)}/{In(h)+Iy(a)}]. ．． ．． ．． ．． ．． ．．
．． ．． ．． ．． ．． ．． ．． ．． ．． ．． ．． (19) Here, △In(h)+△Iy(a)<In(h)
+Iy(a), so equation (19) is [In(ha)−△In(ha)]*[1/{In(h
)+Iy(a)}]*[1+{△In(h)+△Iy(
a)}/{In(h)+Iy(a)}]. ．． ．． ．． ．． ．． ．．
．． ．． ．． ．． ．． ．． ．． ．． ．． ．． ．． (20) becomes. As can be seen from equation (20), even if the amount of transmitted or reflected light from the reference pattern fluctuates and △Iy(a) is positive, the term [In(ha) - △In(ha)] , decreases, and even if △Iy(a) is negative,
By increasing the term [In(ha)−ΔIn(ha)], fluctuations can be absorbed.

The effect of the missing test pattern is determined by standardization method a.
Since it is shown in , it will be omitted. Furthermore, the above also applies to other test patterns or reference patterns, which are normalized by the sum of transmitted or reflected light from one or more reference patterns and the test pattern. It is obvious that the same is true for both cases, so we will omit it here. In the above, we have described how to obtain the correlated output light intensity and how to standardize the correlated output light intensity. Even if we decide to
It is clear from the above explanation that there is a clear difference in the ability to recognize defective patterns and fluctuations in light intensity.

[0044]

Embodiment 2 This embodiment shows a neural network in which the portion that performs the correlation calculation uses a filter based on a Fourier transform hologram. This explains a neural network based on a correlation system using matched filters. FIG. 6 shows an optical system for reproducing a known matched filter. The matched filter 61 uses a pattern obtained by optically Fourier transforming a plurality of reference patterns in advance.
Multiple recording is used while changing the direction of plane wave irradiation for each reference pattern.

As shown in FIG. 6, the beam 12 emitted from the laser 11 is expanded to a suitable beam diameter by a beam expander 13, and the pattern to be recognized and classified drawn on the image display device 16 is displayed. - irradiate with light. Next, the light flux having the complex amplitude distribution of the pattern enters the beam splitter 15 and is divided into transmitted light and reflected light, and the transmitted light is Fourier transformed by the Fourier transform lens 21. , irradiates the multiple interference fringe pattern drawn on the matched filter 61. At this time, the reference pattern is selected from a reference pattern having the same spatial frequency spectrum as the pattern to be recognized and classified.
The diffracted light is emitted in the direction of the plane wave used when creating the beam.

When this diffracted light is focused onto the screen 31 by the condensing lens 22, one of the diffracted lights is focused on the reference pattern.
This becomes a cross-correlation pattern between the pattern and the pattern to be recognized and classified, and the other diffracted light becomes a convolution. Therefore, since the position where the cross-correlation output from each reference pattern appears is known in advance, the peak of the correlation output intensity at that position can be calculated.
The brightness value or total light intensity can be measured and
The data is stored in the memory in the data processor 51. Also, beam splitter
The reflected light from the screen 4 is reflected by the lens 54.
2, an image corresponding to the complex amplitude distribution of the pattern to be tested is formed.

The test pattern imaged on the screen 42
The light intensity of the light is read by the two-dimensional photoelectric conversion element 52 at substantially the same timing as that of the two-dimensional photoelectric conversion element 43, and is stored in the computer 51. Each value stored in the computer 51 is normalized by the computer 51 as in the first embodiment, and the obtained normalized values are respectively input to the input layer of the neural network. .

Note that with the method of Example 2, it is not possible to measure the light intensity of the transmitted or reflected light of each reference pattern, so normalization method b shown in Example 1 cannot be performed. do not have. Since the Fourier transform components of the respective reference patterns are not superimposed on the matched filter, the degree to which the contrast ratio of the correlation output decreases due to an increase in the number of reference patterns is reduced. Therefore, since the dynamic range of the obtained correlation output is less likely to decrease, the degree of separation of the correlation output light intensity or its standardized value increases, which improves the recognition rate when performing recognition with a neural network. increases. Similarly, the joint Fourier transform device shown in the specification of the patent application filed by the present applicant provides similar effects.

The neural network of the present invention recognizes each reference pattern of the joint Fourier transform and the pattern to be classified.
It can also be obtained by performing joint Fourier transforms with lens arrays individually and simultaneously in parallel. Furthermore, by using the neural network of the present invention, the light intensity of the transmitted or reflected light of each reference pattern and each test pattern can be detected by the same method as shown in Example 1. , a beam splitter transmits a light beam having a complex amplitude distribution after passing through an image display device.
This can be easily done by dividing using, etc. In addition, the correlation output light intensity based on the neural network obtained with this optical system is based on the standard shown in Examples 1 and 2 using the light intensity of each reference pattern and each test pattern. The data are standardized using methods a and b and input to the neural network.

When the neural network of the present invention is used, the Fourier transform components of each reference pattern overlap,
This eliminates the problem of lowering the contrast ratio of the cross-correlation strength between the reference pattern and the pattern to be recognized and classified. Furthermore, although both Examples 1 and 2 use coherent light, they are resistant to translational deviations of input patterns.

[0051]

Embodiment 3 FIG. 7 shows a known correlation optical system using incoherent light, and a case where the neural network of the present invention is applied to this will be explained. In this optical system, the image display device 16 and the reference pattern mask 62 are arranged close to each other, and the output of the product of the pattern drawn on the image display device 16 and the reference pattern is sent to the condenser lens array 71. The two-dimensional photoelectric conversion element 43 detects the output from each reference pattern focused by the two-dimensional photoelectric conversion element 43.
This is stored in memory in the computer 51. In addition, in this embodiment, in order to detect the reflected light of the test pattern,
A beam splitter 15 is installed between the image display device 16 and the reference pattern mask 62, and the transmitted light from the image display device 16 is reflected by the beam splitter 15.
4, the light intensity of the image formed on the screen 53 is detected by the two-dimensional photoelectric conversion element 52, and is stored in the computer 51.

In this embodiment, on the image display device 16,
Patterns to be identified and classified into pixels corresponding to the reference pattern mask 62 are displayed in an array. At this time, a cross-correlation peak between each reference pattern and the pattern to be identified and classified is detected. The obtained correlation output light intensity is normalized by the light intensity of the test pattern based on normalization method a, as in Examples 1 and 2, and is input to the input layer of the neural network. .

Although the optical correlation system as described above has no tolerance for translational deviation of the input pattern, it has a very simple structure and is easy to integrate.

[0054]

[Effects of the Invention] As explained above, the above-mentioned effects were obtained by the neural network of the present invention. Putting these together, we get the following remarkable technical effects. That is, first, even if the transmitted light or reflected light from the image display device of the optical correlation system fluctuates, and the value corresponding to the obtained correlation output light intensity varies, the test pattern can be recognized. It has been possible to provide a neural network that is easy to implement and that is also capable of recognizing missing test patterns.

Second, even when learning is performed on missing test patterns, a new method can be used that requires less learning time without reducing the number of categories that can recognize and classify the test patterns.
We were able to provide a virtual network.

[Brief explanation of the drawing]

FIG. 1 is a schematic diagram of an optical correlation system used for preprocessing of an optical neural network according to the present invention.

FIG. 2 is a configuration diagram showing the configuration of a typical conventional neural network.

FIG. 3 is a schematic diagram showing how a test pattern to be recognized and classified and each reference pattern are displayed on an optical correlation system image display device in the present invention.

4 is an explanatory diagram showing the state of correlation output when joint Fourier transform processing is performed on the test pattern and each reference pattern in FIG. 3 using an optical correlation system; FIG.

FIG. 5 is an explanatory diagram showing a distribution of correlated output light intensity in the present invention.

FIG. 6 is a schematic configuration diagram showing an optical correlation system using a matched filter in the present invention.

FIG. 7 is a schematic configuration diagram showing an optical correlation system using incoherent light, in which a test pattern, each reference pattern, and a Fourier transform lens are arranged in an array.

[Explanation of symbols]

1, 3 Image output means 2, 4 Optical Fourier transform means 11
Laser 12, 17, 18 luminous flux

Claims (10)

[Claims] Claim 1: In a neural network whose basic configuration is hierarchical, the number of layers is three or more, and the input layer receives a test pattern to be optically recognized and classified, and a plurality of individual patterns. A value corresponding to the correlation output light intensity for each reference pattern obtained by performing correlation calculation of the reference pattern is calculated between the test pattern or the test pattern and one or more reference patterns. 1. A neural network characterized in that a value normalized by a value corresponding to the number of pixels constituting the neural network is input, and each connection weight value from the input layer to the output layer is determined by supervised learning. 2. The value corresponding to the number of pixels constituting the test pattern is a value corresponding to the intensity of transmitted light or reflected light from the image display device that displayed the test pattern. The neural network according to claim 1, characterized in that: 3. The value corresponding to the number of pixels constituting the test pattern and one or more reference patterns is the same as that of the test pattern.
2. The method according to claim 1, wherein the value is the sum of values corresponding to transmitted light or reflected light intensity of the pattern and one or more of the reference patterns. -ral network. 4. The correlation output light intensity for each reference pattern is determined by taking the peak value of the autocorrelation output or cross-correlation output between the test pattern and each reference pattern. The neural network according to claim 1, characterized in that: 5. The above-mentioned correlation output light intensity for each reference pattern corresponds to the entire light receiving range according to the spread of the autocorrelation output or cross-correlation output between the test pattern and each reference pattern. 2. The neural network according to claim 1, wherein the neural network calculates the amount of light or the average amount of light. 6. The correlation output light intensity for each reference pattern is determined by the peak value of the autocorrelation output or cross-correlation output between the test pattern and each reference pattern and the correlation output. 2. The neural network according to claim 1, wherein the total light amount or the average light amount of the light receiving range is determined depending on the spread. 7. The above correlation output light intensity for each reference pattern is within the linear range of the two-dimensional photoelectric conversion element, or a portion of the correlation output light intensity is within the maximum light receiving area of the two-dimensional photoelectric conversion element. 2. The neural network according to claim 1, wherein the neural network is obtained in a range exceeding the amount. 8. The correlation calculation between the test pattern to be optically recognized and classified and each of the plurality of reference patterns is performed using a filter based on a Fourier transform hologram. The neural network according to claim 1. 9. The correlation calculation between the optically recognized and classified test pattern and each of the plurality of reference patterns is performed using a joint Fourier calculation of the test pattern and each reference pattern. The intensity patterns obtained by performing individual transformations may be optically Fourier transformed again, or a joint Fourier transformation of the test pattern and each reference pattern may be performed. 2. The neural network according to claim 1, wherein the intensity pattern obtained by collectively performing Fourier transform is optically subjected to Fourier transform again. 10. The correlation calculation between the test pattern to be optically recognized and classified and each of the plurality of reference patterns is performed using the test pattern expressed by the light transmittance distribution or the light reflectance distribution. The pattern and each reference pattern are overlapped and irradiated with incoherent light.
2. The neural network according to claim 1, wherein the neural network is operated by measuring the amount of reflected or transmitted light.