JP3325934B2 - Optical pattern recognition and classification device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an optical pattern recognition / classification apparatus used in the field of optical information processing for recognizing and retrieving patterns of objects, characters, voices, and the like, and performing processing such as inference and association. .

[0002]

2. Description of the Related Art Conventionally, as a typical method for recognizing and classifying a pattern of an image or the like, there is a method using an optical correlation detecting means. This method is roughly classified into two types. Can be classified. One is to optically Fourier-transform the pattern, perform pattern matching on this Fourier transform plane, and obtain a correlation peak output according to the degree of the matching. A typical example is a matched filter method, A joint Fourier transform method as shown in FIG. 2 is known. These methods perform correlation detection using coherent light in order to perform an optical Fourier transform.

Here, a conventional joint Fourier transform method will be briefly described with reference to FIG. Image display device 14
In, the input pattern and the reference pattern are displayed side by side and in parallel, and the input pattern and the reference pattern are read out with the coherent light beam 12a emitted from the laser 11,
After passing through the Fourier transform lens 15, a joint Fourier transform image is formed on the screen 34. This image is taken by the imaging device 35
, Is drawn on the spatial light modulator 16, read out by the coherent light beam 12b, passes through the Fourier transform lens 20, and a correlation peak output is obtained.

Another method is a method of performing pattern matching in a real space, and can perform correlation detection using incoherent light. As a typical example of this method, there is a method called shadow correlation, in which an input pattern and a reference pattern are represented by a transmittance mask, the two masks are overlapped, and the amount of light transmitted through both masks is measured to obtain a correlation output. How to get.

[0005] Recognition and classification of a test pattern are performed using correlation outputs obtained by using these optical correlation detecting means. However, in these methods, since a correlation operation is performed on the entire input pattern, a similar pattern will have a similar correlation output. It is only possible to perform the process of correctly recognizing and classifying a missing pattern in the same way as a pattern with no missing, or to consider a missing pattern as a completely different pattern when the missing is large. However, even if the pattern can be correctly recognized and classified into the same desired class as the pattern having no defect, the similar pattern cannot be classified this time. Of course, it was not possible to specify the presence / absence of a pattern defect and the location of the defect.

On the other hand, there is a method of performing pattern recognition and classification using means such as a neural network. In the case of using this method, in identifying an image pattern, a sound pattern, and the like, first, a feature amount based on a plurality of parameters is extracted from an input pattern, and then a certain operation is performed on the set of the feature amounts to determine the identification ability. Is a general procedure. FIG. 3 is a configuration diagram showing a configuration of a conventional neural network. In FIG. 3, a plurality of feature values are input to each node of the input layer, passed through a connection that is multiplied and transmitted by a unique coefficient, passed through the intermediate layer, and output from each node of the output layer. With the number of parameters of this feature amount as n, a position vector having these features as elements is considered, and an n-dimensional space including this vector (which is a vector unique to the input pattern and is called an input vector) is defined as a feature space. I will call it. For example, in a two-dimensional image pattern, this is divided into N × M small areas,
The average value of the density in each region can be used as a feature value. The feature space in this case is an N × M dimensional space, and has a very large dimension in the case of a general image pattern.

In such a feature space, there are roughly two methods for identifying an input pattern. One method is to divide the feature space into respective classes or subclasses, and to identify the class by the class indicated by the region including the input vector in the feature space. Another method is to classify each class or subclass in the feature space. There is a method in which a reference position is prepared, and the input pattern and the norm (distance) between these references are identified by the shortest class. In the calculation of the norm in this case, since the degree of data variation may differ for each coordinate axis, the contribution of each coordinate axis is not calculated uniformly, and the contribution ratio is corrected using the variance value of the data on that coordinate axis. Can also be performed.

A typical example of the former method, that is, a class division method, is a multi-layer perceptron that is being actively studied at present and uses learning by an error back propagation method. However, when a pattern loss occurs in the same way as the method using the optical correlation operation, there is only a discriminating ability to determine whether or not the pattern can be correctly recognized by the class to be recognized and classified. Even if the missing pattern can be recognized and classified into a desired class by expanding the pattern, a problem has arisen that the similar pattern cannot be recognized and classified this time. Needless to say, it was possible to identify the presence / absence of a pattern defect and the location of the defect. As a typical example of the latter method, that is, the shortest norm identification method, a method called a Gaussian identification device has been proposed. However, even in this method,
There was the same problem as in the former.

[0009]

SUMMARY OF THE INVENTION Accordingly, the present invention has been made to solve the above-mentioned problems, and it is an object of the present invention to perform pattern recognition and classification, to identify the presence or absence of a defect, and to specify the location of a defect. It is an object of the present invention to provide an optical pattern recognition / classification device capable of performing the following.

[0010]

According to the present invention, in order to solve the above-mentioned technical problems, the present invention has the following features.
A feature amount extracting unit for extracting a quantity (hereinafter, referred to as a feature quantity) representing a feature of the input pattern (hereinafter, referred to as an input pattern) using an optical system that performs a correlation or an operation similar thereto; Means for discriminating the input pattern based on one or more characteristic amounts extracted by the means, a pattern to be recognized and classified (hereinafter referred to as a test pattern) at a specific position. A plurality of partial patterns (hereinafter, referred to as divided patterns), and a pattern dividing unit that uses an arbitrary one of the divided patterns as an input pattern. Based on the acquired feature amount, a determination is made for each of the divided patterns, and a result of the determination for each of the divided patterns is equal to or more than a predetermined number, and A discriminating means for discriminating a detection pattern; and the feature quantity extracting means displays a light source emitting a coherent light beam and one or a plurality of reference patterns in an amplitude transmittance distribution or a reflectance distribution, First image display means for giving the amplitude distribution to the light beam, a first Fourier transform lens for optically Fourier transforming the amplitude distribution of the light beam transmitted or reflected by the first image display means, and an input pattern A Fourier transform image creating means for creating a Fourier transform image, and a Fourier transform image of an input pattern arranged on the Fourier transform surface of the first Fourier transform lens and obtained by the Fourier transform image creating means. Display as distribution or reflectance distribution,
A spatial light modulator that further provides an amplitude distribution to the light beam, a second Fourier transform lens that optically further Fourier-transforms the amplitude distribution of the light beam transmitted or reflected by the spatial light modulator,
Light receiving means for detecting the amount of light of the image obtained by the second Fourier transform lens in accordance with the one or more reference patterns, and one or more outputs obtained by the light receiving means Is the above-mentioned feature amount.

According to the second aspect of the present invention, the Fourier transform image forming means includes: a light source for emitting a coherent light beam;
A second image display means for displaying the input pattern as an amplitude transmittance distribution or a reflectance distribution to further impart an amplitude distribution to the light flux, and an amplitude of the light flux transmitted or reflected by the second image display means. It is composed of a third Fourier transform lens for optically Fourier transforming the distribution, and imaging means for capturing a Fourier transform image of the input pattern obtained by the third Fourier transform lens.

According to the third aspect of the present invention, an optical system for performing correlation or an operation similar thereto with a quantity (hereinafter, referred to as a feature quantity) representing a feature of an input pattern (hereinafter, referred to as an input pattern) is used. An optical pattern recognition / classification apparatus comprising: a feature amount extracting unit for extracting the input pattern based on at least one feature amount extracted by the feature amount extracting unit; A pattern dividing unit that divides a desired pattern (hereinafter, referred to as a test pattern) into a plurality of partial patterns (hereinafter, referred to as divided patterns) at a specific position and uses any one of the divided patterns as an input pattern; ,
In addition, the determination unit performs a determination for each of the divided patterns based on the characteristic amount obtained for each of the divided patterns, and determines whether or not the result of the determination for each of the divided patterns matches a certain number or more. And a discriminating means for discriminating a test pattern: and the feature amount extracting means includes a coherent light source.
A light source that emits a flux, a third image display unit that displays the input pattern and the reference pattern side by side in parallel with a complex amplitude transmittance distribution or a reflectance distribution and gives the light flux a complex amplitude distribution, A fourth Fourier transform lens for optically Fourier transforming the complex amplitude distribution of the light beam transmitted or reflected by the third image display means, and a Fourier transform image obtained by the fourth Fourier transform lens, A fourth image display means for displaying the light flux as a complex amplitude distribution by displaying a rate distribution or a reflectance distribution, and optically Fourier transforming a complex amplitude distribution of the light flux transmitted or reflected by the fourth image display means A Fourier transform correlator (Join) comprising a fifth Fourier transform lens, and a light receiving means for receiving a pattern obtained by the fifth Fourier transform lens.
t Transform Correlator: hereinafter referred to as JTC), and the cross-correlation output between the input pattern and the reference pattern obtained by the light receiving means is used as the feature amount.

[0013]

In the configuration of the optical pattern recognition / classification apparatus of the present invention as described above, by obtaining a feature amount for at least one feature parameter for each divided pattern obtained by dividing the test pattern into a plurality of pieces at a specific position. Features of each divided pattern in the test pattern can be displayed with a small number of parameters. Based on the characteristic amount acquired for each of the divided patterns, a determination for each of the divided patterns is performed, and a determination result of each of the divided patterns is a predetermined number or more. During recognition and classification after learning, even if a test pattern has loss or distortion, the pattern can be correctly recognized and classified in the same way as when there is no loss, based on the determination result of the divided pattern with no loss. The location can be specified.

Further, in the feature extracting means, the coherent light beam emitted from the light source enters the first image display means. In the first image display means, one or a plurality of reference patterns are displayed with an amplitude transmittance distribution or a reflectance distribution. Therefore, the incident coherent light flux is modulated into the amplitude distribution of the reference pattern and transmitted light is transmitted. Alternatively, the light is emitted from the first image display device as reflected light. The modulated coherent beam passes through a first Fourier transform lens and
Create a Fourier transform image of the reference pattern on the focal plane.

On the other hand, the Fourier transform image of the input pattern is
It is created by the Fourier transform image creating means and is displayed on the spatial light modulator as an amplitude transmittance distribution or reflectance distribution. Since this spatial light modulator is arranged on the focal plane of the first Fourier transform lens, the Fourier transform image of the reference pattern enters the spatial light modulator and is further modulated by the amplitude distribution of the Fourier transform image of the input pattern. Then, the output of the product of the Fourier transform image of the reference pattern and the Fourier transform image of the input pattern is transmitted or reflected. Thereafter, the coherent light beam is again Fourier transformed by the second Fourier transform lens disposed behind the spatial light modulator by its own focal length, and the light amount of the obtained image is reduced to one or a plurality of reference patterns. The light is detected by the corresponding light receiving means. The detected output increases as the degree of coincidence between the Fourier transform image of the reference pattern and the Fourier transform image of the input pattern on the spatial light modulator increases. It can be said to be an output, and the acquisition of each component of a plurality of characteristic parameters of the input pattern can be performed simultaneously in parallel with a simple optical system at high speed.

Further, in the above-described Fourier transform image forming means of the optical system, the coherent light beam emitted from the light source is transmitted to the second image display means in which the input pattern is displayed as an amplitude transmittance distribution or a reflectance distribution. The incident and incident coherent light beam is modulated into the amplitude distribution of the input pattern and emitted as transmitted light or reflected light, and the modulated coherent light beam passes through the third Fourier transform lens and enters the Fourier transform of the input pattern at the focal plane. Create a conversion image,
Only by displaying the input pattern by the method of imaging the pattern by the imaging means, a Fourier transform image of the input pattern can be created optically at high speed and easily.

Further, in the feature amount extracting means, the coherent light beam emitted from the light source enters the third image display means. In the third image display means, since the input pattern and the reference pattern are displayed side by side in parallel with a transmittance distribution or a reflectance distribution having a complex amplitude, the incident coherent light flux is complex with the input pattern and the reference pattern. The light is modulated into an amplitude distribution and exits the third image display means as transmitted light or reflected light. The modulated coherent light beam passes through a fourth Fourier transform lens and forms a joint Fourier transform image of the input pattern and the reference pattern on the focal plane. This joint Fourier transform image is displayed on the fourth image display means as a transmittance distribution or a reflectance distribution of complex amplitude. Another coherent light beam emitted from the light source is
The incident coherent light flux incident on the fourth image display means is modulated into the complex amplitude distribution of the joint Fourier transform image,
The modulated coherent light beam emitted as transmitted light or reflected light passes through the fifth Fourier transform lens and creates a cross-correlation spot between the input pattern and the reference pattern on the focal plane. The light quantity of the cross-correlation spot is detected as a cross-correlation output by the light receiving means, and the acquisition of the characteristic amount can be performed at a high speed with a simple optical system.

Further, the discriminating means creates a membership function for each feature parameter of each divided pattern of the test pattern. It is possible to create an identification area for each test pattern whose degree of belonging is indicated by a membership function. In addition, by taking the characteristic parameter a plurality of times and using the representative value as a reference for creating the membership function, it is possible to create a region in the feature amount extracting means in consideration of the fluctuation of the output. After creation, the membership value of each component is obtained from the feature amount obtained for each divided pattern of the unknown test pattern, and the obtained minimum value or average value of the membership values is obtained for each divided pattern. It can be determined for each divided pattern how many test patterns belong to the identification area. Then, for all the divided patterns constituting the test pattern, to what extent they belong to which discrimination area is obtained, and if these are determined comprehensively, even if the test pattern has loss or distortion, the loss Based on the determination result of the divided pattern having no defect, the pattern can be correctly recognized and classified in the same manner as when there is no defect, and the defect location can be specified.

Further, a neural network is used as the discriminating means, and this neural network is prepared for each divided pattern, or one neural network is used in time series for each divided pattern, and the output layer of each neural network is used. By integrating the output states of the nodes and determining the test pattern formed by the divided pattern, even if the test pattern has loss or distortion, based on the determination result of the split pattern having no loss, The pattern can be correctly recognized and classified in the same manner as when there is no loss, and the position of the loss can be specified.

Next, the optical pattern recognition / classification apparatus of the present invention will be described in more detail with reference to the following examples, but the present invention is not limited thereto.

[0021]

FIG. 1 is a configuration diagram showing the configuration of an embodiment of an optical pattern recognition / classification apparatus according to the present invention. Well, FIG.
In the configuration of the optical pattern recognition / classification device described above, the optical pattern recognition / classification device essentially comprises an optical correlation calculation device 1 as a feature amount extraction unit, an identification / discrimination device 22 as a discrimination unit, and a pattern division device 18 as a pattern division unit. Hereinafter, this will be described in detail with reference to FIG.

In FIG. 1, a coherent light beam 12 emitted from a laser 11 is expanded to an appropriate light beam diameter by a beam expander 13, collimated, and incident on an image display device 14. The reference pattern stored in the computer 17 is drawn on the image display device 14 as a transmittance distribution of the intensity pattern, and the light beam 12 is transmitted to read the reference pattern. The light flux 12 passes through the Fourier transform lens 15 and forms a Fourier transform image of the reference pattern on the spatial light modulator 16 arranged on the focal plane of the Fourier transform lens 15. On the other hand, it is stored in the computer 17 or C
A test pattern image belonging to an identification class to be recognized and classified by an imaging device such as a CD camera from the outside is divided into a plurality of divided patterns at a specific position by a pattern dividing device 18, and one of the divided patterns is Fourier-divided. The converted pattern is input to the converted image forming device 19, and the spatial light modulator 16 displays the intensity pattern of the Fourier transform image of the divided pattern as a transmittance distribution.

FIG. 4 shows an example of a test pattern and a divided pattern. That is, the test pattern is divided at specific positions like the divided patterns a, b, c, and d. That is, the Fourier transform image of the reference pattern set is modulated by the intensity pattern of the Fourier transform image of each divided pattern when passing through the spatial light modulator 16 of FIG. The calculation of the product of the frequency components is performed. After this, the luminous flux 12
The light passes through the Fourier transform lens 20 and is again subjected to Fourier transform, and the light receiving device 21 outputs the result of the cross-correlation operation between the reference pattern and the divided pattern. Image display device 1
The reference patterns displayed in step 4 are sequentially changed, and the processing described above is performed to obtain an output of a cross-correlation operation for a plurality of reference patterns. The above-described processing is performed for all the divided patterns constituting the test pattern, and eventually, outputs for all the reference patterns of all the divided patterns constituting the test pattern are obtained.

These outputs are input to the identification determining device 22. First, it is determined to which identification class each divided pattern belongs. This determination is realized by checking in advance what range of values the plurality of outputs of the divided pattern of the test pattern indicate. As long as these ranges do not overlap among all outputs among a plurality of identification classes, a binary logic set is used to determine whether or not the output value of each input divided pattern for each reference pattern falls within the range checked in advance. It can be done from matching. Thereafter, the test pattern is obtained from the determination of all the divided patterns,
A decision is made as to which identification class it belongs to. For example, in this method, if the result of the determination for each divided pattern is the same as a majority of the number of divided patterns, a method of determining the test pattern to be the identification class to which the test pattern having the same majority is the belonging class. There is something like that. In this way, even if an unknown pattern has a defect, it is also possible to correctly classify the pattern into a desired identification class by using a divided pattern that does not include the defect, or to specify the position of the defect. The identification determination device 22 can realize the determination function by software using the computer 17, or can realize the same function by hardware using an electronic circuit.

Here, a liquid crystal panel or the like which is a general electric address type spatial light modulator can be used as the image display device 14 used. Here, as the spatial light modulator 16, a liquid crystal panel or the like, which is a general electric address type spatial light modulator, can be used. Further, as the light receiving device 21, an image sensor such as a photodiode or a CCD can be used. Alternatively, a process may be performed in which a circular mask or the like having an appropriate radius centered on the optical axis is arranged on the spatial light modulator 16 to cut off the divided pattern, the reference pattern, or both images near the optical axis. This allows
Low spatial frequency components, which are almost common in any pattern, are cut, and cross-correlation operations can be performed efficiently.

Furthermore, if a binary image of the intensity pattern of the Fourier transform image of the divided pattern is displayed on the spatial light modulator 16, the part of the Fourier transformed image of the reference pattern that matches the Fourier transformed image of the divided pattern is: Since the light passes through with little effect from the spatial light modulator, the light receiving device 21
Can be prevented from decreasing. As such a spatial light modulator, an electric address type using a ferroelectric liquid crystal can be used. In particular, a spatial light modulator using a ferroelectric liquid crystal is suitable not only for binarization of an image but also for high resolution and high contrast in principle. Further, if the binarization process is performed using an image processing device or the like, the same applies even when a general electric address type spatial light modulator is used.

Here, the pattern dividing device 18 includes a commercially available image processing device or the like having a function of dividing an input image into a plurality of portions and outputting an image separately for each portion, or an image display device. An optical method that can display an input image and output an image separately for each portion using a liquid crystal shutter or the like can be used.

[0028]

Embodiment 2 FIG. 5 is a configuration diagram showing the configuration of another embodiment of the optical pattern recognition / classification apparatus of the present invention. This embodiment is characterized in that a Fourier transform image of the divided pattern is optically obtained. In the configuration diagram of FIG. 5, the optical pattern recognition / classification device is essentially composed of an optical correlation calculation device 1 as a feature amount extraction unit, an identification determination device 22 as a determination unit, and a pattern division device 18 as a pattern division unit. Is configured.

In the optical pattern recognition / classification apparatus shown in FIG.
The coherent light beam 12 emitted from the laser 11 is expanded to an appropriate light beam diameter by a beam expander 13, collimated, and incident on a half mirror 31, where it is split into two light beams, a light beam 12a and a light beam 12b. Luminous flux 1
2a enters the image display device 32. Image display device 32
Are stored in the computer 17 or CC
A test pattern image belonging to an identification class to be recognized and classified by an imaging device such as a D camera from the outside, and one intensity pattern of the divided patterns divided by the pattern dividing device 18 is drawn as a transmittance distribution. As a result, the light flux 12a is transmitted and the divided pattern is read. Thereafter, the light flux 12a passes through the Fourier transform lens 33, and forms a Fourier transform image of the divided pattern on a screen 34 arranged on the focal plane. This image is taken by the imaging device 3
5 is taken.

On the other hand, the light beam 12b is deflected by the mirror 36 and then enters the image display device 14. The image display device 14 is drawn with a transmittance distribution of an intensity pattern of an image (hereinafter referred to as a reference pattern set) in which a plurality of reference patterns stored in a computer 17 as shown in FIG. The reference pattern set is read through. The light beam 12b transmits through the Fourier transform lens 15, and forms a Fourier transform image of the reference pattern set on the spatial light modulator 16 arranged on the focal plane of the Fourier transform lens 15. In the spatial light modulator 16, a Fourier transform image of the divided pattern imaged by the imaging device 35 is
Since the intensity pattern is displayed as the transmittance distribution, the light beam 12b is modulated by the intensity pattern of the Fourier transform image of the divided pattern when transmitted, and the product of the divided spatial pattern and the corresponding spatial frequency components of the reference pattern set is obtained. Is performed.

Thereafter, the light flux 12b is transmitted through the Fourier transform lens 20 and subjected to Fourier transform again. The light beam 12b is divided by the light receiving device 21 in which the light receiving surface is disposed at a position corresponding to each reference pattern in the reference pattern set. The result of the mutual additive operation with each reference pattern of the pattern set is output simultaneously and in parallel. Thereafter, the above-described processing is performed on all the divided patterns forming the test pattern, and eventually, outputs of all the divided patterns forming the test pattern for a plurality of reference patterns are obtained.

These outputs are input to the discrimination / judgment device 22 for recognizing and classifying the pattern to be inspected and for specifying a missing portion. These processes are the same as in the first embodiment, and a description thereof will be omitted.
As the image display devices 14 and 32 used here, a liquid crystal panel or the like which is a general electric address type spatial light modulator can be used. Here, a liquid crystal panel or the like, which is a general electric address type spatial light modulator, can be used as the spatial light modulator 16, and the Fourier transform pattern of the divided pattern is directly applied to the spatial light modulator 16. In this case, an optical address type spatial light modulator can be used.

Further, as the light receiving device 21, an image pickup device such as a photodiode or a CCD can be used. Further, as the imaging device 35, a CCD camera, an imaging tube, or the like can be used. In this embodiment, the light beam 12 is divided into two by the half mirror 31. However, two laser beams and two beam expanders are prepared, respectively, and the device is configured to separately generate two light beams 12a and 12b. There is no problem. Here, the pattern dividing device 18 displays the input image on a commercially available image processing device or the like or a display device having a function of dividing an input image into a plurality of portions and outputting images separately for each portion. An optical method that can output an image separately for each part using a liquid crystal shutter or the like can be used.

[0034]

Embodiment 3 FIG. 7 is a block diagram showing the configuration of still another embodiment of the optical pattern recognition / classification apparatus of the present invention. The present embodiment is characterized in that the JTC optical system is used for the feature amount extracting means.

In the configuration diagram of FIG. 7, the optical pattern recognition / classification device includes an optical correlation operation device 1 serving as a feature amount extracting means, an identification determining device 22 serving as a discriminating device, and a pattern dividing device 18 serving as a pattern dividing device. Consists essentially of

7, the coherent light beam 12 emitted from the laser 11 is expanded to an appropriate light beam diameter by a beam expander 13 and collimated, and the light beam 12a transmitted through the half mirror 31 is converted to an image display device 14.
Incident on. On the image display device 14, the divided pattern created by the pattern dividing device 18 as shown in FIG. 8 and the reference pattern are drawn side by side in parallel with the transmittance distribution of the intensity pattern. And the reference pattern are read. The light flux 12a transmits through the Fourier transform lens 15, and forms a joint Fourier transform image of the divided pattern and the reference pattern on a screen 34 arranged on the focal plane of the Fourier transform lens 15. This image is picked up by the image pickup device 35 and a joint Fourier transform image is displayed on the spatial light modulator 16.

On the other hand, the light beam 1 reflected by the half mirror 31
The light 2 b is deflected by the mirror 36, enters the spatial light modulator 16 displaying the joint Fourier transform image in a complex amplitude transmittance distribution, passes through the Fourier transform lens 20, and receives light received at the focal plane. A correlation peak is created on the device 21. The light quantity of the correlation peak obtained by the light receiving device 21 is a result of the correlation operation between the divided pattern and the reference pattern, and the reference pattern is sequentially changed to obtain a plurality of feature amounts as the correlation outputs of the divided pattern. Perform the above processing for all the divided patterns constituting the test pattern, after all,
An output is obtained for a plurality of reference patterns of all divided patterns constituting the test pattern.

These outputs are input to the discrimination / determination device 22 for recognizing and classifying the pattern to be inspected and for specifying a missing portion. Since these processes are the same as in the first embodiment, the description will be omitted. As the image display device 14 used here, a liquid crystal panel or the like which is a general electric address type spatial light modulator can be used. Here, as the spatial light modulator 16, a liquid crystal panel or the like which is a general electric address type spatial light modulator can be used, and a joint Fourier transform image is directly formed on the spatial light modulator 16. Then, an optical writing type spatial light modulator can be used. The light receiving device 21 includes a photodiode or a CC.
An imaging element such as D can be used.

Further, a CCD camera, an image pickup tube or the like can be used as the image pickup device 35. In this embodiment,
Although the light beam 12 is divided into two by the half mirror 31, there is no problem if two laser beams and two beam expanders are prepared and two light beams 12a and 12b are separately formed. Here, the pattern dividing device 18 displays an input image on a commercially available image processing device or the like having a function of dividing an input image into a plurality of portions and outputting images separately for each portion, or an image display device. Then, an optical device that can output an image separately for each portion using a liquid crystal shutter or the like can be used.

[0040]

Embodiment 4 In this embodiment, a fuzzy logic operation device is used as the identification determination device 22. In the present embodiment, the processes of the first to third embodiments are performed until the feature parameters are obtained using the feature amount extracting means such as the optical correlation calculation device as shown in FIGS. Therefore, only the part of the identification determination device 22 will be described.

This is based on a plurality of feature amounts extracted for each of the divided patterns of the test pattern belonging to the class to be recognized and obtained by the feature amount extracting means. After creating a membership function of the identification class to which the pattern belongs, a membership value is obtained for each feature parameter by comparing the feature amount of the divided pattern of the unknown test pattern with the above membership function, and the obtained plurality is obtained. Is determined based on the membership values of the test patterns, and the test pattern is determined from the determination results for all the divided patterns constituting the test pattern.

At first, the characteristic amount is obtained by the characteristic amount extracting means. In order to reduce the influence when the output of the characteristic amount extracting means is unstable, it is desirable to perform the plural times for each test pattern. . Therefore, the average value and the standard deviation amount, which are one of the representative values of the output values of the feature parameters, are calculated from the data of a plurality of times, and the membership function for the output intensity is set to a trapezoidal shape. The membership value is 1 up to the range, the membership value is three times the standard deviation, the membership value is 0, and the portion corresponding to the hypotenuse of the trapezoid is from the same to three times the standard deviation.

FIG. 9 shows a pattern A in which the output of the characteristic parameter 1 and the output of the characteristic parameter 2 are represented as membership values when a certain test pattern is A and the divided patterns are a, b, c, and d. 6 is a membership function graph of an identification class. In this way, membership functions for each division pattern are created for all the identification classes, and stored in the identification and classification device 22. Next, a division pattern of an unknown test pattern is shown in place of the pattern A, the output intensity of each component of the characteristic parameter is obtained, and the membership function of the identification class for each division pattern created in advance as described above is obtained. Collate.

For example, membership functions for each of the division patterns of the identification classes A, B, and C are shown in FIGS. 9, 10, and 11, respectively. At this time, it is assumed that the output of the feature parameter for each divided pattern of the unknown test pattern is the output of the s point for the reference pattern 1 and the output of the t point for the reference pattern 2, respectively. FIG. 12 shows this s point,
At point t, for the membership function of each division pattern,
It is a table showing how much membership value it has and how much it belongs to each identification class.

When an AND operation is performed for each divided pattern from the membership values shown in the table of FIG. 12, the minimum value of the membership values of the divided patterns of each test pattern is obtained. The membership values are 1, 0, 0, and 0.3 for the division patterns a, b, c, and d in class A, respectively, and are 0 for the division patterns a, b, c, and d in class B, respectively. , 0,
0, 0, and the divided patterns a, b,
1, 1, 1, 1 for c and d, respectively.

Thereafter, an unknown test pattern is determined from the membership values for each divided pattern. At this time, when the AND operation is performed again, the membership values for each identification class are 0, 0, and 0, respectively. 1 and thus the unknown pattern can be said to be 100% class C. Also, when there are a number of patterns in which the minimum value of the membership value is not 0, the degree to which class belongs can be estimated by using the minimum value.

Further, a loss or the like occurs in the unknown pattern. For example, the membership value of each of the divided patterns of class C becomes 0.3, 1, 1, and 1 and the division where the loss occurs occurs. In the case where the membership value of the pattern is reduced, in the method of determining an unknown test pattern, when the membership value 1 is a majority, the unknown pattern is classified into an identification class that is a majority in the divided pattern. If it is determined that the unknown pattern belongs, the unknown pattern belongs to the class C, and it can be said that there is a missing portion in the divided pattern a. In this way, even if there is a defect in the unknown pattern, it is possible to correctly classify it into a desired identification class and specify the location of the defect by using the divided pattern that does not include the defect. Can also be performed.

The arithmetic mean of the membership values shown in FIG. 12 can be used as a judgment material. That is, in this case, the classes A, B, and C for the unknown test pattern
Since they are about 0.4, 0.15 and 1 respectively, the degree of each pattern can be estimated. Also in this case, the arithmetic mean value is used as a judgment material in the same manner for each divided pattern,
The loss of the pattern can be identified. However, in the case of using the arithmetic mean, an answer may be obtained even if there is no feature value of a certain feature parameter.
It is meaningful as an auxiliary operation when it is difficult to obtain an answer in the D operation. In other words, the state significantly deviates from the state when the average value is obtained for some reason, the membership value is 0 for only one feature parameter component, and 1 for the other reference patterns.
In the case of, the AND operation cannot give a correct answer. In this case, a usage method such as using an arithmetic average value as the second candidate can be used.

In the present embodiment, the membership values are trapezoidal, but any membership fuzzy fuzzy set with no dent in the membership function can be used. Further, such an arithmetic device can be configured using a Neumann computer or an arithmetic device using hardware.

[0050]

Fifth Embodiment In the present embodiment, the above-described processing is performed up to the point where the characteristic amount is obtained by using a characteristic amount extracting means such as an optical correlation calculation device as shown in FIGS. Since it is the same as in Examples 1 to 3, only the part of the identification determination device 22 will be described.

The determination by the identification determination device 22 can be logically performed as in the first embodiment depending on the distribution of the identification classes in the divided pattern, but in the case of a plurality of identification classes, all of them can be logically determined. When the ranges that the feature parameters can take are stochastically overlapped, it is effective to make a determination for each divided pattern using a hierarchical neural network represented by a single-layer or multilayer perceptron. That is, in such a hierarchical neural network, the presentation of the characteristic parameter of the divided pattern of the test pattern belonging to each discrimination class and the desired output indicating the discrimination class at that time a plurality of times, each By performing on the division pattern, the connections in the neural network are organized so that the neural network outputs an identification class appropriate to the input division pattern, so only with the combination of the binarization logic, An identification class that cannot be completely separated can also be determined.

FIG. 13 is a block diagram showing an embodiment of the identification judging device 22 using a hierarchical neural network. In this embodiment, each neural network prepared for each of the divided patterns, that is, the divided patterns a, b, c, and d, classifies each divided pattern into an identification class. Is input to In this calculation means, if the result of the classification for each divided pattern is the same as the majority of the number of divided patterns, even if the unknown pattern has a defect, it is correctly classified into a desired identification class by the divided pattern containing no defect. In addition, it is possible to specify a defect location.

The identification determining device 22 includes a computer 17
The determination function can be realized by software using, or the same function can be realized by hardware by an electronic circuit. There is no particular limitation on the learning algorithm. For example, in the case of a multi-layered perceptron structure having three or more layers, an error back-propagation method based on a well-known generalized delta rule can be used.

The neural network also includes, for example, a network in which an identification class is created by a hypersphere in a feature space and a network in which a distance from a reference data point is weighted. There are functions and the like, and these may be used.

[0055]

As described above, the above-mentioned effects are obtained by the optical pattern recognition / classification apparatus of the present invention. Summarizing them has the following remarkable technical effects. That is, the test pattern is classified into identification classes for each of the divided patterns obtained by dividing the test pattern into a plurality of parts, and the recognition and classification of the test pattern are performed by integrating the results. Recognition and classification can be performed correctly, and the presence or absence of a defect can be identified and the location of the defect can be specified. That is, at the time of recognition and classification after learning, even if a test pattern has a loss or distortion, based on the determination result of a divided pattern having no loss, the pattern can be correctly recognized and classified in the same manner as when there is no loss. Thus, an optical pattern recognition / classification device capable of specifying a defect location can be provided.

[Brief description of the drawings]

FIG. 1 is a schematic configuration diagram showing the configuration of an embodiment of an optical pattern recognition / classification device of the present invention.

FIG. 2 is a schematic configuration diagram showing an example of a conventional optical pattern recognition and classification device using a joint Fourier transform method.

FIG. 3 is a schematic diagram showing an example of a conventional neural network for an optical pattern recognition / classification device.

FIG. 4 is an explanatory diagram showing an example of a test pattern and a division pattern of the test pattern according to the present invention.

FIG. 5 is a schematic configuration diagram showing the configuration of another embodiment of the optical pattern recognition / classification device of the present invention.

FIG. 6 is an explanatory diagram showing a reference pattern set displayed on a part of the optical pattern recognition / classification device of the present invention.

FIG. 7 is a configuration diagram showing the configuration of still another embodiment of the optical pattern recognition / classification device of the present invention.

FIG. 8 is an explanatory diagram showing a divided pattern and a reference pattern displayed on a part of the optical pattern recognition / classification device of the present invention.

FIG. 9 is an explanatory diagram showing a membership function created based on an output of a feature parameter of a division pattern of one identification class according to the present invention.

FIG. 10 is an explanatory diagram showing a membership function created based on the output of the characteristic parameter of the divided pattern of the test pattern of another identification class according to the present invention.

FIG. 11 is an explanatory diagram showing a membership function created based on an output of a feature parameter of a division pattern of still another identification class according to the present invention.

FIG. 12 is an explanatory diagram showing a membership value of an output of a characteristic parameter component for each divided pattern measured for an unknown test pattern according to the present invention with respect to a membership function of an identification class. is there.

FIG. 13 is an explanatory diagram showing an example of a neural network forming an example of the identification determination device of the present invention.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 Optical correlation calculation device 11 Laser 12, 12a, 12b Light flux 13 Beam expander 14, 32 Image display device 15, 20, 33 Fourier transform lens 16 Spatial light modulator 17 Computer 18 Pattern division device 19 Fourier transform image creation device 21 Light receiving device 22 Identification determining device 31 Half mirror 34 Screen 35 Imaging device 36 Mirror

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-57-204979 (JP, A) Toshiharu Takei et al., 31p-A-15 Pattern Recognition Method Using Optical Cross-Correlation and Fuzzy AND Operation, 1991 Proceedings of the 38th Spring Meeting of the Japan Society of Applied Physics, 3rd volume, pp. 865 (58) Field surveyed (Int. Cl. ⁷ , DB name) G02F 3/00 G06E 1/00-3/00 JICST file (JOIS)

Claims (3)

(57) [Claims] 1. An amount representing a feature of an input pattern (hereinafter, referred to as an input pattern) (hereinafter, referred to as a feature amount).
From a feature amount extracting means for extracting the input pattern using an optical system for performing a correlation or a similar operation, and means for determining the input pattern based on one or more feature amounts extracted by the feature amount extracting means. In the optical pattern recognition / classification device configured, a pattern to be recognized / classified (hereinafter referred to as a test pattern)
Is divided into a plurality of partial patterns (hereinafter, referred to as divided patterns) at a specific position, and any one of the divided patterns is used as an input pattern. A determination is made for each of the divided patterns based on the characteristic amount obtained for each of the patterns, and a determination of the test pattern is made based on whether or not the result of the determination for each of the divided patterns matches a certain number or more. The feature amount extracting means displays a light source that emits a coherent light beam and one or a plurality of reference patterns as an amplitude transmittance distribution or a reflectance distribution, and displays the amplitude distribution on the light beam. A first image display means for giving; a first Fourier transform lens for optically Fourier transforming an amplitude distribution of a light beam transmitted or reflected by the first image display means; Means for generating a Fourier transform image of the first Fourier transform lens, and a Fourier transform image of the input pattern obtained by the Fourier transform image creating means, the amplitude transmittance A spatial light modulator for displaying the light beam as a distribution or a reflectivity distribution and further giving an amplitude distribution to the light beam; and a second optically Fourier transforming the amplitude distribution of the light beam transmitted or reflected by the spatial light modulator.
And a light receiving means for detecting the amount of light of the image obtained by the second Fourier transform lens in accordance with the one or more reference patterns, and obtained by the light receiving means. The optical pattern recognition / classification device, wherein one or a plurality of outputs are used as the feature amount. 2. The image processing apparatus according to claim 1, wherein the Fourier-transformed image generating means displays a light source for emitting a coherent light beam and an input pattern as an amplitude transmittance distribution or a reflectance distribution, and further provides an amplitude distribution to the light beam. A third Fourier transform lens for optically further Fourier transforming the amplitude distribution of the transmitted or reflected light beam in the second image display means; and a third Fourier transform lens. 2. An optical pattern recognition / classification apparatus according to claim 1, further comprising an image pickup means for picking up a Fourier transform image of the input pattern. 3. A quantity representing a feature of an input pattern (hereinafter, referred to as an input pattern) (hereinafter, referred to as a feature quantity).
From a feature amount extracting means for extracting the input pattern using an optical system for performing a correlation or a similar operation, and means for determining the input pattern based on one or more feature amounts extracted by the feature amount extracting means. In the optical pattern recognition / classification device configured, a pattern to be recognized / classified (hereinafter referred to as a test pattern)
Is divided into a plurality of partial patterns (hereinafter, referred to as divided patterns) at a specific position, and any one of the divided patterns is used as an input pattern. A determination is made for each of the divided patterns based on the characteristic amount obtained for each of the patterns, and a determination of the test pattern is made based on whether or not the result of the determination for each of the divided patterns matches a certain number or more. Means: and the feature quantity extracting means emits a coherent light beam.
A light source to be radiated, a third image display means for displaying the input pattern and the reference pattern side by side in parallel with a complex amplitude transmittance distribution or a reflectance distribution to give a complex amplitude distribution to the light flux, and A fourth Fourier transform lens for optically Fourier transforming the complex amplitude distribution of the light beam transmitted or reflected by the image display means; and a Fourier transform image obtained by the fourth Fourier transform lens, the complex amplitude transmittance distribution Alternatively, a fourth image display means for displaying the light flux with a complex amplitude distribution by displaying it as a reflectance distribution, and a fourth image optically Fourier-transforming the complex amplitude distribution of the light flux transmitted or reflected by the fourth image display means. 5 joint Fourier transform correlator (Joint Transform Correlator: hereafter referred to as JTC) comprising light receiving means for receiving a pattern obtained by the fifth Fourier transform lens. Is referred to), the optical pattern recognition classifier of the cross-correlation output between the obtained input pattern and the reference pattern by the light receiving means, characterized in that the characteristic quantity of the.