JP3099203B2 - 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 patterns used for recognizing objects, characters, voices, etc. and for searching, inferring, associating information, and the like.
The present invention relates to an optical recognition / classification device for performing recognition and classification.

[0002]

2. Description of the Related Art In recent years, a neural network technique has attracted attention as a method of effectively performing information processing which is not good at a Neumann-type computer such as recognition, association, and inference. In particular, the error back propagation learning method (Error Back Propagatio
n: BP method) to determine the coupling weight between units (generally a multi-input, output signal processing element using a neuron as a model) in the network, thereby obtaining each input pattern.
Flexible recognition that captures the features of the

FIG. 2 is a configuration diagram showing the structure of a conventional neural network. In a neural network for learning by the BP method, a set of units is usually set to three.
A layer structure having at least two layers, in which a layer for inputting a pattern signal is an input layer, a layer for outputting a recognition output signal is an output layer,
The rest is called an intermediate layer or a hidden layer. The output value In (n: unit number) output from each unit of the input layer propagates to the intermediate layer unit, and further, the output value Cn (n: unit number) output from each unit of the intermediate layer Propagates to each unit in the output layer, and an output value On (n: unit number) is obtained from each unit in the output layer. (In FIG. 2, since the outputs of the input layer and the hidden layer unit are combined with a plurality of units, it looks like a multiple output. This is because the output from one unit is input to the plurality of units. Output value is 1
There is only one per unit. At this time, in principle, there is no connection between units included in the same layer, and there is no signal feedback from the layer closer to the output to the layer closer to the input layer.

In FIG. 2, the connection between the units is indicated by an arrow, and the connection load value between the unit of the input layer and the unit of the intermediate layer is denoted by Vij (i) with the corresponding unit number as a subscript. : Intermediate layer unit number,
j: an input layer unit number), Wij (i: output layer unit number, j: intermediate layer unit) in which the unit weight corresponding to the coupling load value between the unit of the intermediate layer and the output unit is added as a subscript Number). For example, the input layer unit 1 and the intermediate layer unit 1
Coupling load value and is expressed as V _11. The coupling load value of the intermediate layer unit 3 and the output layer unit 5 is represented as W _53.

Generally, the input / output characteristics of the intermediate layer unit and the output layer unit are expressed by the following equations.

(Equation 1) f (x) = 1 / [1 + exp (−x)] (3) where i and j are unit numbers corresponding to the above, respectively, and ξ and θ are each unit. Is the bias value at. The function shown in the equation (3) is called a sigmoid function.

In the above-described configuration of the neural network, when a certain input pattern, that is, an output vector (I ₁ I ₂ I ₃ I ₄ I ₅ ) of an input layer unit is given, a desired signal from the output layer is given. And the actual output O _i (here, i = 1, 2,..., 5) of the error of each unit.
A neural network that outputs a desired output pattern for a certain input pattern can be obtained by modifying the magnitude of the coupling load between the units so that the sum of squares is reduced. The BP is a method of calculating a correction coefficient of the load at this time (also referred to as a generalized δ rule). In order to efficiently converge the learning using the correction coefficient, several methods have been proposed. Have been.

However, when recognizing or associating an image or the like, if the number of pixels constituting the image is large and the number of input neurons is very large, the connection with the intermediate neurons becomes enormous, and the number of neurons becomes large. This has been a major obstacle in realizing a ral network. In addition, characters and images are not necessarily located at specific positions on the input surface, and if recognition or association is attempted regardless of the location, a text learned in a pattern whose position has been fixed once. -It cannot be handled by the Ral network. Therefore, there are proposals for presenting characters or learning at different positions in advance at the time of learning. However, it takes a long time to learn, and an extremely large number of intermediate neurons is required to secure invariance with respect to the position. It became a major obstacle in the transition.

[0008]

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and it is an optical pattern recognition which can secure position invariance with a very small number of input units. It is an object to provide a classification device.

[0009]

Means for Solving the Problems The present invention has been made to solve the above technical problem, even without low, and co <br/> Heeren DOO light source, a first pattern A first device for displaying, a first Fourier transform lens for optically Fourier transforming a complex amplitude distribution of a light beam emitted from the first device, and a Fourier transform of the first Fourier transform lens and is disposed on the surface, and a second device for displaying the second pattern, a light receiving element for detecting the amount of light amount or the reflected and transmitted through the second device, said first pattern, the second as the cross-correlation output with the pattern, based on the output from the retrieved light receiving element, the optical pattern recognition classifier and means for classifying recognize the test pattern
The means for recognizing and classifying the test pattern is
The pattern to be recognized and classified and each reference pattern or each reference
The cross-correlation output of the illumination pattern with the Fourier transform pattern
Multiple times, and reduce the representative value of the set of
Based on at least the above, for each class to be recognized and classified,
Fourier transform of the reference pattern or each of the reference patterns
After creating a membership function for the pattern,
The test pattern and each of the reference patterns or each of the reference patterns
The cross-correlation output with the Fourier transform pattern of the turn and
Membership assigned to the class to be recognized and classified
Take the membership value with the
Or the lowest membership value
The average value of the ship values was classified as the recognition pattern of the test pattern.
Characterized in that it belongs to a class
It is a strategic pattern recognition and classification device.

[0010] Even without low, exits and coherently light source, a first device for displaying a first pattern, the first device
Optically Fourier transform the complex amplitude distribution of the emitted light beam
A first Fourier transform lens for performing
The second lens is disposed on the Fourier transform plane of the Rier transform lens, and
A second device for displaying the pattern of
A light receiving element for detecting the amount of light passed or reflected,
Mutual phase between the first pattern and the second pattern
Function output based on the output from the photodetector taken out.
Means for recognizing and classifying the test pattern
In the biological pattern recognition and classification device,
The means for recognizing and classifying test patterns based on force
In the neural network, the reference pattern or
Prepare as many as the number of Fourier transform patterns of the reference pattern
The input neuron, the reference pattern or the
Fourier transform pattern of each reference pattern and the recognition classification
The cross-correlation output with the pattern belonging to the class
Input weights to determine connection weights between neurons
Thereafter, each reference pattern or a file of each reference pattern
-Cross-correlation output between the Lie transform pattern and the test pattern
Is input to the corresponding input neuron, and the
Recognition and classification of turns.

[0011]

[0012]

According to the optical recognition / classification apparatus of the present invention as described above, the comparison between the Fourier transform pattern of the pattern to be recognized and classified and the reference pattern or the Fourier transform pattern is performed in real time on the Fourier transform plane. The cross-correlation between each reference pattern and the Fourier transform pattern of each reference pattern can be detected by each light receiving element. Therefore, it is possible to compress the input information for the pixels constituting the pattern to be recognized and classified into information for the number of reference patterns. In addition, input the pattern you want to recognize and classify
Since the Fourier transform pattern of the pattern to be recognized and classified does not move with respect to the movement in the plane, the invariance with respect to the position can be completely secured. In this way, the compressed pattern information is mapped in the input space as many as the number of reference patterns, and the aggregate is represented by a membership function expressing the degree of belonging to a certain category. By performing collation with, it is possible to reduce the amount of information and perform recognition classification that ensures position invariance. If the above-mentioned compressed pattern information is input to the input layer of the neural network, the number of neurons can be remarkably reduced and the position invariance can be secured.

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

[0014]

[Embodiment 1] FIG. 1 is a schematic configuration diagram for explaining the configuration of an example of an optical pattern recognition / classification apparatus according to the present invention. FIG.
1, the optical pattern recognition / classification apparatus is essentially composed of an image output means 1, an optical Fourier transform means 2, an image output means 3, a light detection means 4, and a detection signal processing means 5. Is configured.

One of the patterns belonging to the class to be recognized and classified drawn on the pattern display device 15 is read out by the coherent light beam 12 output from the laser 11, and is read by the Fourier transform lens 21. The hood is displayed on the pattern display device 31.
A Rie transform pattern is formed. On the other hand, the computer 51
A reference pattern is displayed by a signal from the
The reference pattern serves as a spatial frequency filter. Therefore, a specific frequency component of the pattern belonging to the class to be recognized and classified is cut or attenuated, and the image is formed on the light receiving element 41 through the imaging lens 32. The signal received by the light receiving element 41 is sent to the computer 51. Next, in accordance with a signal from the computer 51, another reference pattern is successively presented on the pattern display device 31, and the test pattern detected by the light receiving element 41 together with the information of the reference pattern. The patterns are sequentially presented to the pattern display device 15, and the reference patterns are successively presented to the pattern display device 31 as described above, so that the patterns belonging to the class to be recognized and classified and the reference patterns are displayed. And the signal detected by the light receiving element 41 are stored in the computer 51 together.

In this way, the arrangement on the Fourier transform plane
In the obtained pattern display device 31 , since the cross-correlation processing between the Fourier transform pattern of the pattern belonging to the class to be recognized and classified and the reference pattern is performed, even if the test pattern moves in the input plane. Positional invariance can be ensured. Further, in general, since the number of pixels for displaying a pattern is very large, by compressing information into a small number of reference patterns, recognition and classification can be performed with a very small set of reference patterns. . Now, in the recognition process, a membership function corresponding to each reference pattern is created based on the reference patterns stored in the computer 51, the patterns belonging to the class to be recognized and classified, and their cross-correlation outputs.
This is performed by comparing the cross-correlation output between the test pattern and each reference pattern with the membership function. Here, an example of the creation of the membership function will be described below. For example, it is assumed that a Fourier transform pattern (a pattern incident on the pattern display device) of a pattern belonging to a class to be recognized and classified is shown in FIG. Note that the hatched portion indicates a portion where there is almost no light quantity.

Meanwhile, as a reference pattern, assuming that selects a two shown in FIG. 4 (a) and 4 (b), the pattern
The light beams transmitted through the display device 31 are respectively shown in FIG.
As shown in (a) and (b), a part of the Fourier transform pattern of a pattern belonging to a class to be recognized and classified is masked. That is, the pattern is output as a part of the pattern in which the spatial frequency components have been cut. FIG.
The shaded portion in the figure indicates a portion through which light does not pass. In this way, the spatial frequency component of the pattern to be recognized and classified is compared with the reference pattern. Now, sets of patterns belonging to the class to be recognized and classified and reference patterns are sequentially presented to the pattern display device 15 and the pattern display device 31, respectively, and data is input to the computer. However, it is desirable to perform these operations a plurality of times for each pattern to be recognized and classified.
This is because the cross-correlation output intensity fluctuates due to the effects of speckle noise due to coherent light, the temporal instability of the pattern display devices 15 and 31, the input timing of the signal of the light receiving element 41, and the like. This is because if recognition classification is performed using output values, accurate recognition classification cannot be performed. Therefore, the average value and the standard deviation value, which are one of the representative values of the cross-correlation output values for the patterns belonging to each class, are calculated from the data of a plurality of times, and the membership function for the cross-correlation output intensity for each reference pattern is trapezoidal. For example, a membership value of 1 from the average value to the range of the standard deviation is set to a membership value of 0, and a portion corresponding to the trapezoidal hypotenuse is set to a unity value of the standard deviation. Up to 3 times.

FIG. 6 is a graph showing the cross-correlation output for the reference pattern 1 and the reference pattern 2 when the pattern to be recognized and classified as described above is assumed to be an image A, as a membership function. It is. In this way, it is stored together with the information of all the classes to be recognized and classified. However, since the standard deviation value happens to be extremely small in a plurality of trials, it is necessary to add a slight correction.

Next, an unknown image pattern is used instead of the image A.
The computer 51 receives the cross-correlation output output intensity for each reference pattern and compares it with the membership function of the pattern to be recognized and classified. For example, suppose that the membership functions for the image patterns A, B, and C to be recognized and classified are those shown in FIGS. 6, 7, and 8, respectively. At this time, the correlation output intensity of the unknown image pattern is a point a with respect to the reference pattern 1,
It is assumed that the output at the point b is obtained with respect to the terminal 2. FIG. 9 shows this a
FIG. 9 is a graph showing the membership values of the points b and b with respect to the pattern to be recognized and classified. In the above case, we only had the values 1,0,
There may be a value in a portion corresponding to the oblique side of the trapezoid, and generally, the membership function value has an analog value.

If the AND operation is fuzzy-logically performed from the membership values shown in FIG. 9, the minimum value of the membership values of the patterns to be recognized and classified is obtained, and an unknown value is obtained. With respect to the pattern, the image patterns A, B, and C are 0, 0, and 1, respectively. Therefore, the unknown pattern can be said to be 100% image pattern C. On the other hand, when there are several patterns in which the minimum value of the membership value is not 0, the degree of a certain pattern can be estimated by using the values of the minimum values. Further, it can be used as a material for determining the arithmetic mean value of the membership values shown in FIG. That is, in this case, for an unknown pattern, the image pattern A,
Since B and C are 0.5, 0 and 1, respectively, the degree of each pattern can be estimated. However, when the arithmetic mean is used, an answer may be obtained even if the degree of correlation with a certain reference pattern does not match at all. Therefore, the fuzzy logical AND operation has fewer errors, but
It is meaningful as an auxiliary operation when it is difficult to obtain an answer by the operation. That is, for some reason, when the average value is significantly different from the state when the average value is obtained, the membership value is 0 for only one reference pattern and 1 for the other reference patterns. , The AND operation cannot give a correct answer. In this case, it is possible to use an arithmetic mean as the second candidate. In the above-described embodiment, the membership function is trapezoidal. However, the membership function may have any shape as long as the membership function does not have a depression and is a so-called convex fuzzy set. It goes without saying that it does not matter.

As described above, the method of creating a membership function for a pattern to be recognized and classified based on the cross-correlation output with each reference pattern and performing recognition and classification has been described. As a method, there is a method using a neural network described below. In general,
When recognizing an image or the like using a neural network, the number of constituent pixels of the image itself is very large.
At least 100 × 100), the number of input neurons becomes 10,000 or more. Further, if the number of images to be recognized and classified is large, the number of intermediate neurons required for the image becomes very large, and the number of interconnections becomes enormous. Furthermore, when recognizing an image, it is normal for the image to shift in the input plane,
It is necessary to ensure position invariance. However, in order to ensure positional invariance, the number of intermediate neurons must be much larger than the number of images to be recognized and classified. For this reason, the number of interconnections has been further increased, and it has been difficult to substantially recognize many images.

However, as described above, the cross-correlation with the reference pattern on the Fourier transform surface of the pattern to be inspected keeps the position invariance and the pattern to be inspected.
A very small number of input neurons can be obtained by compressing and expressing the enormous amount of information in the cross-correlation information with the reference pattern and assigning the same number of reference patterns to the prepared input neurons. Thus, many patterns can be recognized and classified. That is, as shown in FIG. 10, a cross-correlation output for the reference pattern 1 and the reference pattern 2 is inputted as an input neuron, so that an ordinary three-layer hierarchical neural network is used. After determining the load value of the connection between neurons, the cross-correlation output between the pattern to be inspected and each reference pattern is input to the corresponding input neuron. Will output the degree to which Ron belongs to the desired class.

It is needless to say that the neural network in the above embodiment may be any type as long as there is an input neuron. For example, a network that surrounds a hypersphere or a network that weights the distance from a reference data point.
The radial basis function, which is a function, may be used. Usually, the area near the optical axis of the Fourier transform pattern of the test pattern is large in light quantity at a low spatial frequency, and there is little information that is important as an information quantity. By performing processing such as masking the light receiving element 41 so as not to receive light, the SN ratio can be improved.

[0024]

[Embodiment 2] FIG. 11 is a schematic structural view of another optical pattern recognition / classification apparatus of the present invention. This embodiment is characterized in that the pattern for performing the cross-correlation calculation with the Fourier transform pattern of the test pattern is the Fourier transform pattern of the reference pattern. In the optical layout diagram of FIG. 11, the optical pattern recognition / classification device includes an image output means 1, an optical Fourier transform means 2, an image output means 3, a light detection means 4, and a detection signal processing means 5. It is essentially composed.

One pattern belonging to the class to be recognized and classified drawn on the pattern display device 15 is read out by the coherent light beam 12 emitted from the laser 11 and is read by the Fourier transform lens 21. A Fourier conversion pattern is created on the display unit 31. On the other hand, the reference pattern stored in the computer 51 is stored in the pattern display device 1.
6 is read by the beam 17 split by the beam splitter 14 and a Fourier transform pattern is created on the screen 33 by the Fourier transform lens 22. The Fourier transform pattern of the above-mentioned reference pattern is received by a two-dimensional light receiving element 42 such as a CCD and the pattern display device 31.
Draw on. In this case, the Fourier transform pattern of the reference pattern
Since the intensity pattern of the pattern is drawn as a transmittance distribution on the pattern display device 31, the modulation is performed by the intensity pattern of the Fourier transform pattern of the pattern belonging to the class to be recognized and classified. The product of the corresponding spatial frequency components of the pattern to be inspected and the reference pattern is output to the imaging lens 32.
As a result, the output of the sum of the products of the spatial frequency components (the cross-correlation output of the Fourier transform pattern of the test pattern and the Fourier transform pattern of the reference pattern) is output from the light receiving element 41. Obtainable. In this manner, it is the same as in the first embodiment that the reference patterns are successively presented on the pattern display device 16 and the data of the cross-correlation output is repeatedly obtained for the patterns to be recognized and classified. Omitted.

In this way, the arrangement on the Fourier transform plane
In the obtained pattern display device 31 , since the cross-correlation processing is performed between the Fourier transform pattern of the pattern belonging to the class to be recognized and classified and the Fourier transform pattern of the reference pattern, the test pattern is displayed on the input surface. Even if it moves, positional invariance can be ensured.
Furthermore, in general, since the number of pixels for displaying a pattern is very large, information is compressed into a cross-correlation output of a small number of reference patterns and a Fourier transform pattern, so that recognition can be performed with a very small set of reference patterns. It can be classified. Hereinafter, a method of performing recognition and classification, that is, a method of creating and comparing a membership function by the computer 51 and a method of inputting the input function to an input neuron of the neural network are also the same as those in the first embodiment, and will not be described.

Further, the pattern to be inspected can be recognized almost in real time. This is because the signal from the computer 51 is shown in FIG.
6, the reference pattern is input to the pattern display device 15, and different reference patterns are simultaneously displayed in parallel on the pattern display device 15 as shown in FIG. Is displayed on the pattern display device 16 only once. That is, the individual reference patterns drawn on the pattern display device 15 are Fourier-transformed simultaneously by the Fourier transform lens 21 and input to the pattern display device 31.

On the other hand, the test pattern is Fourier-transformed by the Fourier-transform lens 22, and its intensity pattern is
The pattern is drawn as a mask pattern on the pattern display device 31 via the two-dimensional light receiving element 42. Accordingly, the Fourier transform patterns of the individual reference patterns are modulated simultaneously and in parallel by this mask and input to the imaging lens 32. The imaging lens 32 is connected to the light receiving element 41 and the pattern display device 1.
If it is arranged at a position where 5 and 5 are conjugated,
On the light receiving element 41, the intensity pattern of the Fourier transform pattern of the test pattern is set at a position conjugate with the position of each reference pattern drawn on the pattern display device 15. The modulated reference pattern is displayed. The modulated individual reference patterns are detected by the light receiving element 41, so that the Fourier transform pattern of the reference patterns and the test pattern are simultaneously and in parallel.
The cross-correlation output with the Fourier transform pattern can be obtained. In this case, if a set of reference patterns is determined in advance and a photographic film or the like is used in place of the pattern display device 15, it is very easy and complete to present the reference patterns. Can be performed in parallel.

Incidentally, the reference pattern in this embodiment has a function of
The linear transformation pattern may be the intensity pattern itself, or the intensity pattern may be substantially binarized and presented to the pattern display device 31 as a light / dark pattern. When a binary pattern is displayed in a light-dark pattern, a spatial frequency component having a relatively high intensity of the reference pattern can be emphasized and output, so that the characteristics of the reference pattern can be further enhanced. A cross-correlation output with the emphasized test pattern can be obtained. Further, the pattern display device 1 according to the above embodiment is used.
Although 5, 16 and 31 are electric address type spatial light modulators such as transmission type liquid crystal television screens, they may be reflection type or optical address type spatial light modulators.
Usually, the light intensity near the optical axis of the Fourier transform pattern of the pattern to be inspected is large at a low spatial frequency and little important as the information amount. As in the first embodiment, the SN ratio can be improved by performing a process such as masking the light receiving element 41 so as not to receive light.

[0030]

Embodiment 3 FIG. 13 is a schematic structural view of an optical pattern recognition / classification apparatus showing another embodiment of the present invention. What is characteristic in this embodiment is that a plurality of patterns to be recognized and classified are displayed on one spatial light modulator, and each one of them is individually processed by a Fourier transform lens array. The point is that different reference patterns are presented for each of the Fourier transform patterns to be subjected to the Rier transform, and the cross-correlation outputs are taken in parallel.

The image display device 15 displays a plurality of identical patterns to which a class to be recognized and classified belongs.
The light is read out by the coherent light flux 12 emitted from the light source 1 and the pattern display device 3 is output by the Fourier transform lens array 23.
1, a plurality of Fourier transform patterns of a pattern to which the class to be recognized and classified belongs are created. On the other hand, the computer
Fourier transform lens array 2
When a separate reference pattern is drawn on the pattern display device 31 near each optical axis of No. 3, each reference pattern becomes
Since the patterns belonging to the class to be recognized and classified will act as separate spatial frequency filters,
A specific frequency component of the pattern to be inspected is cut or attenuated, and the pattern passes through the imaging lens array 34.
An image is formed on the light receiving element array 43.

The amount of light received by the light receiving element array 43 (that is, the cross-correlation output between the pattern belonging to the class to be recognized and classified and the reference pattern) is sent to the computer 51 to be recognized and classified. The information on the patterns belonging to the class, the information on each reference pattern, and their cross-correlation outputs are stored. Next, the same pattern belonging to the class to be recognized and classified differently is displayed in a plurality of pattern display devices 15.
The information of patterns belonging to the class to be recognized and classified, the information of each reference pattern, and their cross-correlation outputs are stored in the memory as described above.

In this way, the arrangement on the Fourier transform plane
Since the cross-correlation processing between the Fourier transform pattern of the test pattern and each of the reference patterns is performed in the pattern display device 31 that has been described, even if the test pattern moves on the input surface, the position of the pattern is not changed. Denaturation can be ensured. Further, in general, the number of pixels for displaying a pattern is very large, so that by compressing information into a small number of reference patterns, recognition and classification can be performed with a very small set of reference patterns. is there. Since then
The method of performing recognition and classification, that is, the method of creating and checking a membership function by the computer 51, and the method of inputting the input function to the input neuron of the neural network are the same as in the first embodiment, and will not be described.

In this case, since it is not necessary to sequentially present the reference patterns to the pattern display device 16 and take the cross-correlation output, it is possible to greatly reduce the time for incorporating the correlation output. it can. If a reference pattern for obtaining the cross-correlation output is determined in advance, the pattern display device 3
Instead of 1, a fixed mask pattern such as a film can be used. In this case, the time required to display the reference pattern can be reduced, so that a faster pattern recognition / classification apparatus can be provided. still,
Although the pattern display devices 15 and 31 in the above embodiments are electric address type spatial light modulators such as transmissive liquid crystal television screens, they may be reflection type or optical address type spatial light modulators. It may be a modulator. Usually, near the optical axis of the Fourier transform pattern of the test pattern, the light amount is large at a low spatial frequency and the amount of information is not important. By performing processing such as applying a mask to the light receiving element array 43 so as not to receive light, the SN ratio can be improved, but the same as in the first embodiment.

When a fixed mask pattern such as a photographic film is used, the Fourier transform pattern of the reference pattern in the second embodiment is prepared in another optical system. By arranging as a Fourier transform pattern mask of a separate reference pattern on the optical axis of the Fourier transform lens array, it is recognized as a Fourier transform pattern of the reference pattern of the second embodiment. Cross-correlation outputs with the pattern to be classified can be obtained in parallel and in real time.

In this case, similarly to the second embodiment, the Fourier transform pattern of the reference pattern may be the intensity pattern itself, or the intensity pattern may be substantially binarized to obtain a bright and dark image. May be used as a Fourier transform pattern mask. In the case where a binary pattern is converted into a light-dark pattern to form a Fourier transform pattern mask, a spatial frequency component having a relatively strong reference pattern can be emphasized and output. Therefore, the feature amount of the reference pattern is further emphasized. The cross-correlation output with the test pattern can be obtained.

[0037]

Embodiment 4 Next, FIG. 14 is a schematic configuration diagram showing still another example of the optical pattern recognition / classification apparatus of the present invention.
A characteristic of this embodiment is that a plurality of patterns belonging to a class to be recognized and classified are displayed on one spatial light modulator, and each one is individually Fourier-transformed lens array. , A Fourier transform pattern of a separate reference pattern is presented for each one of the Fourier transform patterns, and the cross-correlation outputs are simultaneously and in parallel. The point to take.

A plurality of identical patterns belonging to a class to be recognized and classified are displayed on the pattern display device 15, read out with the coherent light beam 12 emitted from the laser 11, and recognized on the pattern display device 31 by the Fourier transform lens array 23. Create multiple Fourier transform patterns for the pattern you want to classify. On the other hand, beam splitter 15
The light beam 17 split by the above-mentioned method irradiates a plurality of different reference patterns drawn on the pattern display device 16 by a signal from the computer 51. The luminous flux emitted from the pattern display device 16 is modulated by each reference pattern, is incident on the Fourier transform lens array 24, and creates a Fourier transform pattern corresponding to each reference pattern on the screen 33 . These are received by a two-dimensional light receiving element 42 such as a CCD, and are arranged on the pattern display device 31 such that the respective optical axes of the Fourier transform lens array 23 coincide with the center of the Fourier transform pattern of the reference pattern. In this way, since the Fourier transform pattern of each reference pattern acts as a separate spatial frequency filter for the pattern belonging to the class to be recognized and classified, a specific frequency component of the pattern to be recognized and classified is cut. Alternatively, the attenuated pattern is imaged on the light receiving element array 43 via the imaging lens array 34.

The amount of light received by the light receiving element array 43 (that is, the cross-correlation output between the pattern belonging to the class to be recognized and classified and the Fourier transform pattern of the reference pattern)
Is sent to the computer 51 and stores information of patterns belonging to the class to be recognized and classified, information of each reference pattern, and the cross-correlation output. Next, a plurality of the same patterns belonging to different classes to be recognized and classified,
The information is sequentially presented to the pattern display device 15, and the information of the pattern belonging to the class to be recognized and classified, the information of each reference pattern, and the cross-correlation output are stored as described above.

In this way, the arrangement on the Fourier transform plane
In the obtained pattern display device 31 , the cross-correlation processing is performed between the Fourier transform pattern of the pattern belonging to the class to be recognized and classified and the Fourier transform pattern of the reference pattern. Therefore, even if the test pattern moves within the input plane, position invariance can be ensured. Furthermore, in general, the number of pixels that display a pattern is
By compressing the information into a cross-correlation output of a small number of reference patterns with the Fourier transform pattern,
Recognition classification can be performed with a very small set of reference patterns. Hereinafter, a method of performing recognition and classification, that is, a method of creating and checking a membership function by the computer 51 and a method of inputting the input function to an input neuron of a neural network are the same as those in the first embodiment, and thus description thereof is omitted.

In this case, since it is not necessary to sequentially present the reference patterns to the pattern display device 16 and take the cross-correlation output, it is possible to greatly reduce the time for incorporating the correlation output. Can be. In the above embodiment, the pattern display devices 15 and 31 are electric address type spatial light modulators such as transmissive liquid crystal television screens, but may be reflection type or optical address type spatial light modulators. It may be a spatial light modulator. In addition, usually, near the optical axis of the Fourier transform pattern of the test pattern, the light amount is large at a low spatial frequency and the amount of information is not important. In order not to receive light, a process such as applying a mask to the light receiving element array 43 is performed.
The SN ratio can be improved as in the first embodiment. Further, similarly to the second embodiment, the reference pattern
The Fourier transform pattern of the pattern may be the intensity pattern itself, or the intensity pattern may be substantially binarized to be a light-dark pattern to form a Fourier transform pattern mask. it can. In the latter case, since the spatial frequency component of the reference pattern having relatively high intensity can be emphasized and output, the test pattern in which the feature amount of the reference pattern is further enhanced.
And a cross-correlation output with the target can be obtained. The spatial light modulator used here is, for example, a liquid crystal panel having a photoconductive layer provided on a nematic liquid crystal panel.
It is commercially available under the name of ht Valve (LCLV) from Fuse, USA. Also, devices using ferroelectric liquid crystals have been developed as having higher speed operation and higher resolution. Further, a transmissive element using Bi ₁₂ SiO ₂₀ (BSO) is also available. However, in the case of a transmissive element, a change in the optical system is necessary because the direction in which the readout light beam is incident is changed, but this is essentially the same as the case in which a reflective spatial light modulator is used. .

[0042]

As described above, the optical pattern recognition / classification apparatus of the present invention has the above-mentioned effects. Summarizing them has the following remarkable technical effects. First, the cross-correlation between the Fourier transform pattern of the pattern belonging to the class to be recognized and classified with an extremely simple configuration and the reference pattern or the Fourier transform pattern of the reference pattern. Since the quantity can be detected, an optical pattern recognition / classification apparatus capable of completely recognizing a pattern to be recognized / classified even when moved in a plane can be provided.

Second, a cross-correlation output between the Fourier transform pattern of the pattern belonging to the class to be recognized and classified and the reference pattern or the Fourier transform pattern of the reference pattern is detected. Therefore, by presenting a small number of reference patterns, the information of the pattern consisting of a very large number of pixels is compressed into an extremely small cross-correlation output, whereby the input of the neural network is reduced. An optical pattern recognition / classification apparatus capable of greatly reducing the number of neurons can be provided. Third, the cross-correlation output between the Fourier transform pattern of the pattern belonging to the class to be recognized and classified and the reference pattern or the Fourier transform pattern of the reference pattern can be detected. Therefore, by performing a fuzzy-logical AND operation of the membership function created based on the correlation output, it is possible to provide an optical pattern recognition / classification device capable of performing highly accurate recognition, classification and association. did it.

[Brief description of the drawings]

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

FIG. 2 is a schematic configuration diagram showing one example of a conventional three-layer neural network.

FIG. 3 is a diagram showing a pattern of a pattern to be recognized and classified according to the present invention.
FIG. 4 is an explanatory diagram showing an example of a Rier conversion pattern.

FIG. 4 is an explanatory diagram showing a distribution area of an example of a reference pattern used in the present invention.

FIG. 5 is a view showing a pattern of a pattern to be recognized and classified according to the present invention.
FIG. 9 is an explanatory diagram showing an example of a cross-correlation pattern between a Rier transform pattern and a reference pattern.

FIG. 6 is an explanatory diagram showing a membership function created based on a cross-correlation output between one pattern to be recognized and classified and a reference pattern according to the present invention.

FIG. 7 is an illustration showing a membership function created based on a cross-correlation output between another pattern to be recognized and classified and a reference pattern according to the present invention.

FIG. 8 is a diagram illustrating a membership function created based on a cross-correlation output between a pattern to be recognized and classified and a reference pattern according to another embodiment of the present invention.

FIG. 9 is an explanatory diagram showing a membership value of a correlation measured for an unknown pattern according to the present invention with respect to a pattern to be recognized and classified.

FIG. 10 is an explanatory diagram showing that a cross-correlation output between a reference pattern and a pattern to be recognized and classified according to the present invention is input to a neural network.

FIG. 11 is a schematic configuration diagram showing another configuration of the optical pattern recognition / classification device of the present invention.

FIG. 12 is an explanatory diagram showing a reference pattern displayed on a part of the optical pattern recognition / classification apparatus of the present invention.

FIG. 13 is a schematic configuration diagram showing still another configuration of the optical pattern recognition / classification device of the present invention.

FIG. 14 is a schematic configuration diagram showing another configuration of the optical pattern recognition / classification device of the present invention.

[Explanation of symbols]

 1, 3 Image output means 2 Optical Fourier transform means 4 Light detection means 5 Detection signal processing means 11 Laser 12, 17 Light flux 13 Beam expander 14 Beam splitter 15, 16, 31 Pattern Display device 18 mirror 21, 22 Fourier transform lens 23, 24 Fourier transform lens array 32 imaging lens 33 screen 34 light receiving element 41 two-dimensional light receiving element 43 light receiving element array 51 computer

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ identification code FI G06G 7/19 G06G 7/19 A (58) Field surveyed (Int.Cl. ⁷ , DB name) G02F 3/00 502 G02B 27 / 42 G03H 1/16 G06F 15/18 G06G 7/19 G06F 9/44 G06T 7/00 JICST file (JOIS)

Claims (2)

(57) [Claims] Even 1. A no less, and coherently light source, a first device for displaying a first pattern, the complex amplitude distribution of the light flux emitted from the first device, is optically Fourier transform a first Fourier transform lens for, and is arranged at the Fourier transform plane of the first Fourier transform lens, second
A second device for displaying the pattern of the above, a light receiving element for detecting the amount of light transmitted or reflected by the second device,
Said first pattern, as a cross-correlation output between the second pattern, based on the output from Installing the issued light receiving element, a light and a means for classifying recognize the test pattern
In the biological pattern recognition / classification device, the means for recognizing and classifying the test pattern includes:
The pattern you want to match and each reference pattern or each reference pattern
The cross-correlation output with the Fourier transform pattern of
The representative value of the obtained set of correlations at least
In the same way, for each class to be recognized and classified,
Or the Fourier transform pattern of each of the reference patterns
After creating a membership function corresponding to
Of each reference pattern or each reference pattern
Cross-correlation output with Fourier transform pattern and recognition classification
Membership function assigned to the class you want
The membership value of
Is also smaller, or the membership value
Of the test pattern to be recognized and classified
The optical pattern recognition / classification apparatus. 2. At least a coherent light source;
A first device for displaying a pattern, and an output from the first device.
Optically Fourier transform the complex amplitude distribution of the emitted light beam
A first Fourier transform lens for performing
The second lens is disposed on the Fourier transform plane of the Rier transform lens, and
A second device for displaying the pattern of
A light receiving element for detecting the amount of light passed or reflected,
Mutual phase between the first pattern and the second pattern
Function output based on the output from the photodetector taken out.
Means for recognizing and classifying the test pattern
In the biological pattern recognition / classification device, the means for recognizing and classifying the test pattern includes a neural network.
In the network, the reference pattern or the reference
Input patterns prepared for the number of Fourier transform patterns of the illumination pattern.
Each reference pattern or each reference is applied to a force neuron.
I want to classify the pattern by Fourier transform pattern
Input the cross-correlation output with the pattern belonging to the class.
Force to determine the connection weight value between neurons,
Fourier transformation of each reference pattern or each reference pattern
The cross-correlation output between the replacement pattern and the test pattern
Input to the input neuron, and
The above-mentioned optical pattern recognition characterized by classification.
Classifier.