JP3020680B2 - Pattern recognition classification method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition and classification method used in the field of optical information processing.

[0002]

2. Description of the Related Art Conventionally, there are a matched filter method and a joint transform method as optically performing a correlation operation. In the matched filter method, after performing a Fourier transform on a two-dimensional reference pattern, a reference wave is radiated, a so-called Fourier transform hologram is created, and a Fourier transform image of the test pattern is generated by using the hologram as a filter. Correlation calculation is performed by superimposing. In the joint transform method, a joint Fourier transform image of a test pattern and a reference pattern is recorded as an intensity pattern, and this is irradiated with a plane wave to perform a correlation operation. However, in both methods, the reference pattern group is prepared by the number of patterns to be recognized, and the pattern having the largest correlation output for the presentation of the test pattern, that is, the reference pattern group is obtained based on the autocorrelation output. In this method, the number of reference patterns to be stored increases as the number of reference patterns increases, so that the load on the two-dimensional or three-dimensional spatial light modulator increases, and the recognition device has The ability is also largely governed by the ability of the spatial light modulator and could not be used as a substantially better optical recognition associative device.

Therefore, the present applicant takes a correlation output of a pattern belonging to a class to be recognized and classified and a plurality of reference patterns a plurality of times, and based on a representative value of the obtained set of correlation degrees and various quantities indicating variations. The proposed pattern recognition classification method was provided. (See Japanese Patent Application No. 3-41516). With the pattern recognition and classification method of the prior application, the number of reference patterns, which was conventionally required by the same number as the number of patterns to be recognized, can be greatly reduced. Recognition method became possible.

In the pattern recognition / classification method, a reference pattern is determined by obtaining a correlation output in advance between a pattern belonging to a class to be recognized and classified and a pattern group prepared in advance as a reference pattern candidate. Do not approach. In other words, a set of reference patterns was created by selecting patterns that would fall apart. In this method, a correlation output must be obtained between all the patterns to be recognized and classified once and a group of patterns prepared in advance as reference pattern candidates in order to determine a reference pattern. It took time to decide. In addition, the number of patterns to be recognized and classified has increased,
How to select a reference pattern when it becomes necessary to add a reference pattern has not been particularly determined.

[0005]

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and it is an object of the present invention to provide an appropriate reference pattern creating means and a reasonable pattern recognition / classification method using an additional means. With the goal.

[0006]

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned technical problem, and provides a correlation output between a pattern belonging to a first class to be recognized and classified and a plurality of reference patterns. The first of a set of correlations obtained multiple times
The correlation value between a pattern belonging to a second class to be recognized and classified and the plurality of reference patterns is taken a plurality of times, and a second representative value of a set of obtained correlation degrees and the In a method of performing pattern recognition and classification based at least on an amount corresponding to a distance from a first representative value, that is, a distance in a feature space formed by coordinate axes of each of the plurality of reference patterns, or to perform recognition and classification In a pattern recognition and classification method in which a correlation output between a pattern belonging to a class and the plurality of reference patterns is an input value of a neural network having input neurons corresponding to the plurality of reference patterns, a pattern prepared in advance as a reference pattern candidate The pattern with the smallest correlation output obtained from all combinations of groups is identified as the reference pattern set. And adding a pattern having a low correlation output to the set of reference patterns to the set of reference patterns until the number of reference patterns becomes sufficient to recognize patterns belonging to the class to be recognized and classified. A pattern recognition / classification method is provided, wherein the plurality of reference patterns are created by means for creating a repeated reference pattern.

Further, a correlation output between a pattern belonging to a first class to be recognized and classified and a plurality of reference patterns is obtained a plurality of times, and at least a first representative value of the obtained set of correlation degrees is used as a reference, or The correlation output between the pattern belonging to the second class to be classified and the plurality of reference patterns is taken a plurality of times, and the distance between the second representative value of the obtained set of correlation degrees and the first representative value, that is, The distance in the feature space created by the coordinate axes of each of the plurality of reference patterns, in a method for pattern recognition and classification based on at least an amount corresponding to, or a pattern belonging to a class to be recognized and classified and the plurality of reference patterns, A correlation output of a neural network having input neurons corresponding to the plurality of reference patterns as input values. When a pattern belonging to a class to be classified is added, and recognition and classification are difficult with the plurality of reference patterns, a new reference pattern is prepared from a group of patterns prepared in advance as a reference pattern candidate. select by means adds a new reference pattern into a plurality of reference patterns, newly create a plurality of reference patterns, the correlation output between the added belongs to the class to be recognized classification pattern and the plurality of new reference pattern A pattern recognition and classification method is provided, wherein the plurality of reference patterns are created by using a reference pattern adding unit that takes a reference pattern.

The method of selecting a pattern having the lowest correlation output with respect to all of the reference pattern sets is to compare a pattern group prepared from the candidates with a total sum of correlation outputs of the reference pattern sets, It is preferred to choose the pattern with the lowest value. As for the method of selecting a pattern having the lowest correlation output with respect to all the sets of the reference patterns, it is preferable to compare the sum of squares of the correlation output with the set of the reference patterns and select a pattern having the lowest value. . Then, a method of selecting a pattern having the lowest correlation output with respect to all of the reference pattern sets is to compare the sum and square sum of the correlation outputs of the reference pattern set and each of the pattern groups prepared in advance as the reference pattern candidates. , It is preferable to select a pattern having a low value for both. The correlation output between the plurality of reference patterns and the test pattern is preferably a peak value of an autocorrelation output or a cross-correlation output between the test pattern and each of the plurality of reference patterns.

The correlation output between the plurality of reference patterns and the test pattern is the total light amount or the average light amount in the light receiving range according to the spread of the autocorrelation output or the cross-correlation output between the test pattern and each of the plurality of reference patterns. It is preferred that there be. Further, the correlation output between the plurality of reference patterns and the test pattern is a light receiving range corresponding to the peak value of the autocorrelation output or the cross-correlation output between the test pattern and each of the plurality of reference patterns, and the spread of the correlation output. Is preferably the total light amount or the average light amount. The correlation output between the plurality of reference patterns and the test pattern is obtained by standardizing the output value from the two-dimensional light receiving element by the transmitted light amount or the reflected light amount from the test pattern or the test pattern and the plurality of reference patterns. It is preferred to be an amount. The correlation output between the plurality of reference patterns and the test pattern is obtained when the linear range of the two-dimensional light receiving element or a part of the correlation output is obtained in a range exceeding the maximum light receiving amount of the two-dimensional photoelectric conversion element. It is suitable.

The correlation output between the plurality of reference patterns and the test pattern is obtained by using the above-mentioned reference pattern creation means and the above-mentioned reference pattern addition means obtained by using a filter created by using a Fourier transform hologram. It is suitable. The correlation output between the plurality of reference patterns and the test pattern is obtained by optically individually Fourier transforming the intensity pattern obtained by individually performing the joint Fourier transform of the test pattern and each of the reference patterns. Convert or
Alternatively, the intensity pattern obtained by performing joint Fourier transform of the test pattern and each reference pattern collectively by the Fourier transform lens array may be optically collectively Fourier transformed again by the Fourier transform lens array. It is suitable. The correlation output between the plurality of reference patterns and the test pattern is obtained by superimposing the test pattern represented by the light transmittance distribution or the light knitting rate distribution and each reference pattern, irradiating with incoherent light, and reflecting or transmitting the light. It is preferable to take a light amount.

The correlation output between the plurality of reference patterns and the test pattern includes a reference pattern or a Fourier transform intensity pattern of the reference pattern, and a Fourier transform surface of the test pattern, a mask representing a transmittance distribution or a reflectance distribution. Or, it is arranged as a spatial light modulator, and it is obtained by taking the amount of light of the test pattern transmitted or reflected, or the intensity pattern of the Fourier transform of the test pattern is transmitted to the Fourier transform surface of the reference pattern. A mask representing a rate distribution or a reflectance distribution, or arranged as a spatial light modulator,
It is preferable to simultaneously measure the amount of light of one or a plurality of reference patterns that have transmitted or reflected it. The pattern belonging to the class to be recognized and classified, and the correlation output of the plurality of reference patterns are taken a plurality of times, based on various amounts indicating the representative value and the variation of the set of the obtained degree of correlation,
After creating a membership function corresponding to the plurality of reference patterns for each class to be recognized, a correlation output between the test pattern and the plurality of reference patterns, and the membership function assigned to each class to be recognized and classified. It is preferable that the membership value of the test pattern is the smallest, or the average value of the membership values is the degree to which the test pattern belongs to the class to be recognized and classified.

A correlation output between a pattern belonging to a class to be recognized and classified and the plurality of reference patterns is used as an input value of an input neuron, and a correlation between the plurality of reference patterns and a pattern belonging to a first class to be recognized and classified is used. 1
And a weight corresponding to the distance between the first correlation output value and the second correlation output value between the first correlation output value, the plurality of reference patterns, and the pattern belonging to the second class to be recognized and classified. It is preferable to use the means for creating the reference pattern and the means for adding the reference pattern using a simple neural network. Further, a correlation output value between a pattern belonging to a class to be recognized and classified and the plurality of reference patterns is used as an input value of an input neuron, and an amount according to a distance in the feature space is used as an evaluation function in the feature space. It is preferable to use a means for creating the reference pattern and a means for adding the reference pattern by a neural network in which a class to be recognized and classified is formed by a neural network for forming an identification area surrounded by a hypersphere for each class.

The pattern recognition and classification method uses a correlation output value between the plurality of reference patterns and a pattern belonging to a class to be recognized and classified as an input value of an input neuron, and calculates a quantity according to a distance in a feature space. It is preferable that the class to be recognized and classified is a pattern recognition and classification method using a neural network that forms a recognition area surrounded by a hypersphere for each class, using as an evaluation function. Further, the pattern recognition and classification method includes a first correlation output value between a pattern belonging to a first class to be recognized and classified, the plurality of reference patterns, and a second correlation output value to be recognized and classified with the first correlation output value. Based on the amount corresponding to the distance between the pattern belonging to the class and the second correlation output value of the plurality of reference patterns, another class whose distance is closest to the correlation output value of a certain class in the feature space. It is preferable to use a pattern recognition method in which a closed area surrounded by a hypersphere whose radius is the distance to the midpoint of the distance to the correlation output value is an identification area for each class.

[0014]

According to the pattern recognition / classification method of the present invention, since the set of reference patterns is selected so that the spatial frequency components are different from each other as much as possible, the correlation output which is the correlation output between the reference pattern and the test pattern is used. The value varies greatly depending on the reference pattern and the test pattern, which facilitates recognition and classification. That is, when the correlation output between the reference pattern and the test pattern is obtained, the correlation output value between the test pattern and the reference pattern is plotted in the feature space as a result of several measurements, and a certain region is subjected to each test. The patterns will have, and their identification regions are each far enough apart to recognize and classify each other.

Next, the method of pattern recognition and classification by the reference pattern creation means and the addition means of the present invention will be described in detail with reference to examples, but the present invention is not limited thereto.

[0016]

Embodiment 1 FIG. 1 is a schematic configuration diagram showing the function of an optical correlation processing apparatus for showing an example of a pattern recognition / classification method by a reference pattern creating unit and an adding unit according to the present invention. In the optical arrangement shown in FIG. 1, the optical correlation processing device essentially comprises an image output unit 1, an optical Fourier transform unit 2, an image output unit 3, an optical Fourier transform unit 4, and a light detection unit 5. Since this is a known joint Fourier transform optical system, its configuration will be briefly described.

One pattern to be recognized and classified and a reference pattern group drawn on the image display device 16 are read out by a coherent light beam 12 emitted from a laser 11, and a joint Fourier transform pattern is created on a screen 31 by a Fourier transform lens 21. I do. The combined Fourier transform pattern is read by the two-dimensional photoelectric conversion element 32, drawn on an electric address type liquid crystal display (hereinafter abbreviated as CTV) 35 by the image processing device 33, read out by the light beam 37, and read by the Fourier transform lens 41. And the correlation intensity on the screen 51 is detected by the two-dimensional photoelectric conversion element 53. Also, Fourier transform lens 2
By arraying 1, 41 and making the image display device 16 simultaneously draw the test pattern and the plurality of reference patterns, the correlation strength between the test pattern and the plurality of reference patterns can be simultaneously detected in parallel. . Here,
The description will be made using the alphabet as the reference pattern and the test pattern. However, the reference pattern and the test pattern naturally indicate not only the alphabet but also the entire two-dimensional image pattern.

Hereinafter, a method of creating a reference pattern set using the optical correlation processing apparatus will be described. First,
As a reference pattern candidate, a group of patterns that can be a simple and possible reference pattern are prepared in advance. Then, correlation output values of all combinations of the patterns are obtained. In the combination, those having the lowest correlation output value are selected as a set of reference patterns. Next, an image pattern having a low correlation output value for both of these two reference patterns is added to the set as a new reference pattern. Repeat this,
A set of reference patterns is created in a number sufficient to recognize patterns belonging to the class to be recognized and classified.

Hereinafter, a method of performing pattern recognition and classification by reference pattern creating means and adding means according to the present invention will be specifically described. For example, E, K. When V, Y, M, and O are considered as reference pattern candidates, correlation output values are obtained for all combinations of the respective patterns. For example, patterns such as E and K, E and V, and K and Y are displayed on the image display device 16, and a correlation output value appearing on the screen 51 is detected by a two-dimensional photoelectric conversion element.
The combination having the lowest value is set as the first reference pattern set. Here, since the lowest correlation output value among all combinations of E, K, V, Y, M, and O is the combination of E and V, first, E and V are used as reference patterns. To be elected. Next, to further select a reference pattern,
An image pattern having a low correlation output value for both E and V previously selected is selected. Here, K and M have large correlation output values for both E and V, and Y has a small correlation output value for E but large for V. On the other hand, since O has a low correlation output value for both E and V, O is selected as the third reference pattern.

A method of performing recognition and classification using E and V among the obtained reference patterns will be described below. As a reference pattern set, two character patterns E and V, and a character pattern to be recognized and classified as H are arranged as E and H, V and H as shown in FIG. , One after another in chronological order. In this case, on the screen 51, the correlation output between the character pattern H and the reference character pattern E is a light intensity pattern as shown in FIG. E and H in FIG.
For example, when E and H completely overlap with each other, the correlation output intensity indicates a peak value at the center of the figure. When the vertical bars of E and H overlap, a vertical line on the left of the figure becomes Overlap, horizontal lines above and below the peak value in FIG.
The correlation pattern between E and H is represented as shown in FIG.

In FIG. 3, the autocorrelation pattern appearing on the optical axis is omitted here because it is not relevant here. Now, the peak value of the correlation output intensity obtained in this way is sent to the computer 52, stored, and similarly, the patterns to be recognized and classified are presented on the image display device 16 one after another.
The obtained peak value of the correlation strength is stored in the computer 52 one after another.

Further, it is desirable to perform these operations a plurality of times for a pattern to be recognized and classified. this is,
The peak value of the correlation output intensity fluctuates due to speckle noise due to coherent light, the temporal instability of the image display device 16 and the LCTV 35, and the input timing of the signal of the two-dimensional photoelectric conversion element 32. This is because accurate recognition classification cannot be performed if the recognition classification is performed using the correlation output value of Therefore, the average value and the standard deviation amount of the peak value for the pattern belonging to each class are calculated from the data of a plurality of times, and, for example, the range from the average value to the standard deviation is defined as a first correlation output value.
A correlation output value of three times the magnitude of the standard deviation is defined as 0, and a portion corresponding to the hypotenuse of the trapezoid is 3 times the magnitude of the standard deviation.
Up to double. FIG. 4 is a graph expressing the peak value of the correlation output with respect to E and V when the pattern to be recognized and classified obtained as described above is H, as a membership function. In this way, membership functions are created for all the classes to be recognized and classified and stored in the computer 52 together with information on the classes to be recognized and classified. However, since the standard deviation amount happens to be extremely small in a plurality of trials, it is necessary to add a slight correction.

The recognition classification of the unknown pattern is performed by taking the correlation output value between the membership function for each reference pattern obtained by the above-described method and the degree of correlation of the unknown pattern, and obtaining the smallest correlation output value of the correlation output value. The value or the average value is of a degree belonging to the class to be recognized and classified. The combination of the reference patterns as small as possible and having the recognition and classification ability can reduce the probability of erroneous recognition and reduce the recognition time. Leads to a reduction in

In FIG. 2, an unknown character pattern is presented in place of the character pattern H, and the peak value of the correlation output intensity for each reference pattern is fetched into the computer 52, and for each of the above-mentioned patterns H, W, N to be recognized and classified. Assume that the membership functions are as shown in FIGS. 4, 5 and 6, respectively. At this time, the peak value of the correlation output intensity of the unknown character pattern is
Assume that an output at point b is obtained for the point and the pattern V. FIG.
Is a graph showing how much the point a and the point b have a membership value for each pattern to be recognized and classified. In the above embodiment, only 1 and 0 were included, but there may be a value in a portion corresponding to the hypotenuse of the trapezoid. Generally, the membership value has an analog value. When the AND operation is performed fuzzy logic based on the membership values shown in FIG. 7, the minimum value of the membership values of the patterns to be recognized and classified is obtained. For unknown patterns, patterns H, W, and N are:
1, 0, 0. Therefore, the unknown pattern is 100
% Pattern H. On the other hand, when there are a number of patterns in which the minimum value of the membership value is not 0, the degree of being a certain pattern can be estimated with the value of the minimum value.

FIG. 8 shows recognition classification for reference patterns E and V created based on the average light amount in the light receiving range according to the spread of the correlation output obtained when an experiment was actually performed using the above-described joint Fourier transform method. The membership function of the desired pattern (15 patterns of the alphabet) is shown.
The average light amount in the light receiving range according to the spread of the correlation output will be described in detail later. For example, according to FIG. 3, it is the average light amount of the entire correlation pattern including the peak value.

In FIG. 8, the solid line represents the membership function for the pattern E, the broken line represents the membership function for the pattern V, and the horizontal axis represents the average light amount of the correlation output. However, there is one without one membership function, but this is because the other membership function did not fit within the horizontal scale,
It has been omitted. FIG. 9 shows the results obtained by experiments using the above-described joint Fourier transform method. The above-mentioned fuzzy logic obtained when the unknown pattern was presented three times each as 15 alphabetic patterns. A
It is a representation of the degree of recognition classification of each pattern derived from the ND operation result. The alphabet on the left is the input pattern, and the numbers in the first and second parentheses indicate the average light amounts of the measured correlations with respect to the reference patterns E and V, respectively. Further, the alphabet on the right side is a recognition character obtained by performing a fuzzy logical AND operation of the membership values shown in FIG.
It is doubled.

As described above, there are some confusing ones, but if the highest one is selected, all the patterns are identified by only two reference patterns.
This is a great feature not found in conventional correlation processing using general autocorrelation.

Here, it can be seen from FIG. 9 that the set of patterns T, L and I and the set of patterns K and M are particularly confusing, because the characteristics of each pattern are similar. Means that it is difficult to separate only by a reference pattern having a great feature of vertical and horizontal lines and V is an oblique line. Even if it is as it is, it can be used for recognition or classification by taking a high degree, but even if it is intended to remove this effect completely and use it for a complete recognition, It can be realized with very simple processing. That is, this can be realized by introducing one pattern in which membership functions between confusing patterns do not overlap. This is because since a fuzzy logical AND operation is performed, if at least one of the membership values is 0, it is recognized as a different pattern. Here, O may be used as a newly added reference pattern as described above. This is because T, L, I, which could not be separated by E and V
This is because the features of L and M can be separated by taking a correlation with O. If the correlation output is obtained from the reference pattern O and all the test patterns and added to the existing membership functions, when recognizing and classifying the unknown pattern, the existing reference pattern and the newly added reference pattern are used. If a correlation output is obtained and a membership function is created, recognition and classification can be performed by the above-described fuzzy logical AND operation.

In the above embodiment, not only when there is a confusing test pattern from the beginning but also when the number of test patterns increases and a reference pattern needs to be added,
Similarly, a reference pattern is added using the reference pattern O, and a membership circuit is formed for the increased test pattern by using a new reference pattern set formed by the conventional reference pattern and the newly added reference pattern. Just create it. If a membership function between the conventional test pattern and the reference pattern O is newly created, the unknown test pattern can be obtained by calculating the correlation output with the new reference pattern set by the membership function using the above fuzzy logic. Recognition classification can be performed by an AND operation.

In the above-described embodiment, the membership function is trapezoidal. However, the membership function may be any fuzzy fuzzy set having no depression. It doesn't matter, of course. In addition, in the present embodiment, the standard for creating the membership function is the standard deviation, and about three times the standard deviation is used as a measure of the spread of the membership function. However, it is needless to say that this spread may be determined. That is.

For example, if the area where the membership value becomes 1 is set to twice the standard deviation amount, approximately 97% of the detected degree of correlation falls into this area statistically. Can be detected at a rate of almost 100%. However, the spread of the membership function makes it difficult to separate from other classes. Now, there are various ways of obtaining the firing output intensity. The correlation output is generally
When the two patterns are A (x, y) and B (x, y), they are represented by the following equations. I (x ', y') = ∬A (x, y) B ^* (x'-x, y'-
y) dxdy where ^* represents the complex conjugate quantity.

It is clear from this equation that the correlation output has twice the size of the pattern. Also, since the distribution of the correlation output is different depending on A (x, y) and B (x, y), it is necessary to take the total light amount or the average light amount in the area considering the spread from the peak position. Means that information that better reflects the characteristics of each pattern not seen in the peak light amount is obtained. Also, both the peak value of the correlation output and the average light amount of the light receiving region according to the spread of the correlation pattern
When used as a membership function, two pieces of information having different properties are obtained for one reference pattern, so that the number of reference patterns can be further reduced.
Recognition classification can be performed more reliably from a lot of information.

Further, since the dynamic range of the two-dimensional light receiving element for detecting the correlation output is limited, it is practically impossible to receive the entire correlation light amount. Therefore, when the maximum value of the correlation light amount, that is, the peak value of the correlation output is adjusted to the maximum sensitivity of the two-dimensional light receiving element, the maximum effect can be improved when the correlation light amount is detected by the peak output. However, in the case where the total light amount or the average light amount in the light receiving range according to the spread of the correlation output is taken, the peak output of the correlation output is large, so that even if an output other than the peak of the correlation output is detected, a very large value is obtained. May not be possible.

In such a case, the maximum sensitivity of the two-dimensional light receiving element is set at a position lower than the peak value of the correlation light amount, and the light amount near the peak value of the correlation output is saturated. The dynamic range of the two-dimensional light receiving element can be effectively used by setting the light amount of the portion where the characteristics of the peripheral pattern is better reflected so as to exhibit a linear response. Therefore, it is very effective to obtain the total light amount or the average light amount in the light receiving range according to the spread of the correlation output.

Although a method of creating a membership function from correlation outputs and performing recognition and classification has been described here, there is another method using a neural network. In general, in order to recognize and classify a two-dimensional image pattern in a neural network, conventionally, input neurons are required by the number of pixels of the pattern, and a neural network formed by the input neuron is very complicated and huge. Recognizing and classifying a large number of image patterns was difficult because of the need for connections and intermediate neurons. Here, a correlation between the test pattern and the reference pattern is obtained by the optical correlation process, and the correlation is input as information of the test pattern to a neural network having input neurons corresponding to the number of reference patterns as input values. Thus, a large number of patterns can be recognized and classified with a neural network having a very small number of input neurons. Here, the neural network described above may be of any type as long as a correlation output is input to an input neuron.

That is, a first three-layer structure neural network, a first correlation output value between the plurality of reference patterns and a pattern belonging to a first class to be recognized and classified, and the first correlation output and the first correlation output A neural network that performs weighting based on an amount corresponding to a distance between a plurality of reference patterns and a second correlation output value of a pattern belonging to a second class to be recognized and classified, or an amount based on distance as an evaluation function, In the feature space, the class to be recognized and classified may be a neural network or the like that forms an identification area surrounded by a hypersphere for each class.

Further, a first correlation output value between a pattern belonging to the first class to be recognized and classified and obtained by the optical correlation processing device and the plurality of reference patterns separately from the neural network; In a feature space, the correlation output value and the plurality of reference patterns and a pattern belonging to the second class to be recognized and classified are based on various quantities indicating a distance corresponding to a distance between the second correlation output value. A pattern recognition and classification method in which a closed area surrounded by a hypersphere having a radius equal to the distance to the midpoint of the distance of the correlation output value of another class closest to the correlation output value of the class is used as an identification area for each class. It does not matter.

In this embodiment, the correlation optical system may be of any type. Various optical systems have also been proposed for the joint Fourier transform optical system, but the screen 31, the two-dimensional photoelectric conversion element 32,
Image display device 33, electric address type liquid crystal light valve 3
It goes without saying that what can be done by using an optical address type liquid crystal light valve instead of 5 is also possible. In the present invention, there is a difference in the use of the portion functioning as a spatial light modulator, but in principle, the same electrical address type and optical address type can all be used. It is. In addition, an analog type or a digital type can be used.

In the above embodiment, the joint Fourier transformer is used as the optical system for obtaining the correlation output. In this case, however, there is an advantage that the reference pattern can be easily added or rewritten. . Also, the image display device 16
By simultaneously presenting a plurality of reference patterns and one test pattern, a correlation output of the plurality of reference patterns and the test pattern can be obtained simultaneously and in parallel. However,
When presenting a plurality of reference patterns at the same time, the contrast ratio of the multiple interference fringe patterns on the screen 31 of FIG. 1 decreases with the increase in the number of reference patterns. Resulting in a reduced dynamic range, which effectively reduces the pattern discrimination ability. In the present embodiment, as shown in FIG. 2, the image display device 16 has described a method of processing one reference pattern and one test pattern in chronological order. An embodiment using a correlation processing method that can simultaneously detect correlation outputs of a plurality of reference patterns and a test pattern will be described below.

[0040]

Embodiment 2 FIG. 10 shows an optical system for reproducing a known matched filter. As the matched filter 61, a pattern obtained by optically Fourier transforming a plurality of reference patterns in advance and multiplex-recording them while changing the irradiation direction of the plane wave for each reference pattern is used.

As shown in FIG. 10, a light beam 12 emitted from a laser 11 is spread by a beam expander 13 into an appropriate light beam system, and irradiates a pattern to be recognized and classified drawn on an image display device 16. Next, the luminous flux having the complex amplitude distribution of the pattern is Fourier-transformed by the Fourier transform lens 21, and irradiates a multiple interference fringe pattern drawn on the matched filter. At this time, diffracted light is emitted from the reference pattern having the same spatial frequency spectrum as the pattern to be recognized and classified in the direction of the plane wave used when creating the reference pattern. When this diffracted light is condensed on the screen 31 by the condenser lens 22, one of the diffracted lights becomes a cross-correlation pattern between the reference pattern and the pattern to be recognized and classified, and the other becomes a convolution. Therefore, since the position where the cross-correlation output by each reference pattern appears is known in advance, the peak value of the correlation output intensity at that position or the total light amount can be measured.

Hereinafter, the correlation output intensity is similarly measured for other patterns to be recognized and classified, and the pattern recognition and classification method and the means for creating and adding the reference pattern set are the same as those in the above-described embodiment, and therefore will not be described. . Next, an example of an optical system whose integration is relatively simple will be simply described.

[0043]

Embodiment 3 An example shown in FIG. 11 is a known incoherent correlation optical system. The image display device 1 is used as an optical system.
6 and the reference pattern mask 62 are arranged close to each other, and the output of the product of the pattern drawn on the image display device 16 and the reference pattern is condensed on the screen 31 using the condensing lens array 71, The two-dimensional photoelectric conversion element 53 detects the output from each of the collected reference patterns. In the present embodiment, a pattern to be recognized and classified into pixels corresponding to the reference pattern mask 62 is displayed on the image display device 16.
They will be displayed in an array. In this case, a cross-correlation peak between each reference pattern and a pattern to be recognized and classified is detected.

The following operation can be performed in the same manner as in the first embodiment by displaying patterns to be recognized and classified one after another on the image display device and using their correlation outputs. For this reason,
A reference pattern creating unit, an adding unit, and a pattern recognition / classification method are omitted. In this case, there is no tolerance for the translational deviation of the input pattern, but the input pattern has a very simple configuration and is easily integrated. Next, an example of a shift-invariant correlation optical system that can completely cope with a translational deviation of a test pattern on a pattern display surface will be described.

[0045]

Embodiment 4 FIG. 12 shows a configuration of a silt invariant correlation optical system which can completely cope with a translational deviation of a pattern to be detected on a pattern display surface. In this optical system, a reference pattern drawn on an image display device 16 is read out by a light beam 21 output by a laser 11, a Fourier transform lens 21 is used to display a Fourier transform pattern of the reference pattern on a screen 31, and its Fourier transform is performed. The pattern intensity is expressed as a transmittance distribution or a reflectance distribution on the image display device 63 and used as a filter, or the reference pattern is expressed as a transmittance distribution or a reflectance distribution on the image display device 63 and used as a filter. When the test pattern drawn on the image display device 64 is transmitted or reflected by the Fourier transform lens through the image display device 63, the amount of light transmitted or reflected by the filter is equal to the Fourier transform pattern of the reference pattern or the light intensity of the reference pattern. It becomes a pseudo correlation output between the detection pattern and the Fourier transform pattern.

Further, the following may be performed. That is, the test pattern drawn on the image display device 16 is read out by the light beam 21 output by the laser 11, and the Fourier transform lens 21 displays the Fourier transform pattern of the test pattern on the screen 31. The intensity is expressed as a transmittance distribution or a reflectance distribution on the image display device 63, and is used as a filter.
When one or a plurality of reference patterns drawn in the above are transmitted or reflected by the Fourier transform lens through the image display device 63, the amount of light transmitted or reflected by the filter becomes the Fourier transform pattern of the one or more reference patterns. And a Fourier transform pattern of the test pattern. In particular, when a plurality of reference patterns are presented at the same time, a pseudo-correlation strength between the Fourier transform pattern of each reference pattern and the Fourier transform pattern of the test pattern can be obtained in a completely simultaneous and parallel manner.
It can be used as an optical system for a very fast and shift-invariant pattern recognition and classification method.

In the same manner, similarly, the test patterns are drawn one after another on the image display device 16 or 64, a pseudo correlation output with the reference pattern is obtained, and the first embodiment is used by using those correlation outputs.
In the same manner as in, recognition and classification can be performed. For this reason, the reference pattern creation unit and the addition unit, and the pattern recognition and classification method can be omitted.

[0048]

As described above, the above-described effects are obtained by the pattern recognition and classification method according to the reference pattern creating means and the adding means of the present invention. Summarizing them has the following remarkable technical effects. That is, the first
Next, the correlation output values of all combinations of patterns that are candidates for reference patterns are taken, the lowest set of the correlation output values is first set as the first reference pattern set, and
A pattern having a low correlation output is set as a new reference pattern for any of the reference patterns in the set, and a set of reference patterns is determined to a desired number one after another by this repetition, thereby greatly reducing the time required to create the reference pattern. Thus, a pattern recognition / classification method using a reasonable reference pattern creation unit can be provided.

Second, in order to select a reference pattern from patterns prepared in advance as reference pattern candidates irrespective of the number and type of test patterns, the selected reference pattern has generality. It is possible to provide a pattern recognition and classification method using a reference pattern creating unit that does not need to newly select a reference pattern for a set of test patterns of different types. Third, since the reference patterns are selected so as to have different characteristics as much as possible, the correlation output of the pattern to be recognized and classified has a large variation for each reference pattern, and a wide feature space is obtained. A pattern recognition / classification method using a reference pattern creating unit capable of obtaining more patterns to be recognized / classified even when the number of reference patterns is small can be provided.

Fourthly, when it is necessary to add a reference pattern due to the addition of a test pattern or the like, a new reference pattern candidate is prepared from the previously prepared pattern in the same manner as the reference pattern creation means. By adding a new reference pattern and using a set of reference patterns that have been used up to now and a new reference pattern, it is possible to obtain a correlation output for newly added test patterns, shorten the time, and add reasonable reference patterns. Could provide a pattern recognition and classification method using.

[Brief description of the drawings]

FIG. 1 is a schematic configuration diagram showing a joint Fourier transform optical system as an example used as an optical correlation processing device in a pattern recognition / classification method of the present invention.

FIG. 2 is a diagram showing a state in which a pattern to be recognized and a reference pattern are displayed when a joint Fourier transform optical system is used as an optical correlation processing device in the pattern recognition / classification method of the present invention.

FIG. 3 is a diagram showing a state of a correlation output when a joint Fourier transform optical system is used as an optical correlation processing device in the pattern recognition / classification method of the present invention.

FIG. 4 is a diagram showing a membership function created based on the degree of correlation of each reference pattern with respect to one pattern to be recognized and classified in the pattern recognition and classification method of the present invention.

FIG. 5 is a diagram showing a membership function created based on the degree of correlation of each reference pattern with another pattern to be recognized and classified in the pattern recognition and classification method of the present invention.

FIG. 6 is a diagram illustrating a membership function created based on the degree of correlation of each reference pattern with another pattern to be recognized and classified in the pattern recognition and classification method of the present invention.

FIG. 7 is a table showing a membership value of a reference pattern with respect to an unknown pattern in a pattern recognition / classification method of the present invention, the membership value having a pattern to be recognized / classified.

FIG. 8 is a diagram showing membership functions of 15 patterns to be recognized and classified for two experimentally obtained reference patterns in the pattern recognition and classification method of the present invention.

FIG. 9 is a diagram illustrating a recognition result for an unknown pattern using an experimentally obtained fuzzy AND operation and the accuracy of the recognition result in the pattern recognition and classification method of the present invention.

FIG. 10 is a schematic configuration diagram of an optical system using a matched filter as another example used as an optical correlation processing device in the pattern recognition / classification method of the present invention.

FIG. 11 is a schematic configuration diagram of an optical system as still another example used as an optical correlation processing device in the pattern recognition / classification method of the present invention.

FIG. 12 is a schematic configuration diagram of an optical system as still another example used as an optical correlation processing device in the pattern recognition / classification method of the present invention. .

[Explanation of symbols]

1, 3 Image output means 2, 4 Optical Fourier transform means 5 Light detection means 11 Laser 12, 37 Light flux 13 Beam expander 14, 17 Beam splitter 16, 63, 64 Image display device 21, 41 Fourier transform lens 22, 54 Condensing lens 31, 51 Screen 32, 53 Two-dimensional photoelectric conversion element 33 Image processing device 34 Mirror 35 Liquid crystal light valve 52 Computer 61 Matched filter 62 Reference pattern mask 71 Condensing lens array

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int. Cl. ⁷ , DB name) G06F 15/18 G02F 3/00 502 G06T 7/00 JICST file (JOIS)

Claims (18)

(57) [Claims] 1. A correlation output between a pattern belonging to a first class to be recognized and classified and a plurality of reference patterns is taken a plurality of times, and at least a first representative value of a set of obtained correlations is used as a reference or a recognition is performed. The correlation output between the pattern belonging to the second class to be classified and the plurality of reference patterns is taken a plurality of times, and the distance between the second representative value of the obtained set of correlation degrees and the first representative value, that is, The distance in the feature space created by the coordinate axes of each of the plurality of reference patterns, in a method for pattern recognition and classification based on at least an amount corresponding to, or a pattern belonging to a class to be recognized and classified and the plurality of reference patterns, A pattern recognition and classification method in which the correlation output of the above is used as an input value of a neural network having input neurons corresponding to the plurality of reference patterns. The pattern outputs having the smallest correlation output obtained from all combinations of the pattern groups prepared in advance as candidates for the pattern are set as a reference pattern set, and a pattern having a low correlation output for all of the reference pattern sets is set as a reference pattern set. Creating the plurality of reference patterns by means of a reference pattern creating unit that goes in addition to the set of reference patterns and repeats until the number of the reference patterns becomes a sufficient number to recognize a pattern belonging to a class to be recognized and classified. A pattern recognition classification method characterized by the following. 2. A correlation output between a pattern belonging to a first class to be recognized and classified and a plurality of reference patterns is obtained a plurality of times.
A correlation output between a pattern belonging to a second class to be recognized and classified and the plurality of reference patterns is obtained a plurality of times with at least a first representative value of the obtained set of correlations as a reference. Pattern recognition based on at least a quantity corresponding to a distance between a second representative value of the set of the reference patterns and the first representative value, that is, a distance in a feature space formed by coordinate axes of each of the plurality of reference patterns. In a classifying method, or in a pattern recognition / classifying method in which a correlation output between a pattern belonging to a class to be recognized and classified and the plurality of reference patterns is an input value of a neural network having input neurons corresponding to the plurality of reference patterns, When a pattern belonging to a class to be recognized and classified is added, and the plurality of reference patterns make recognition and classification difficult, a reference pattern A new reference pattern is selected from a group of patterns prepared in advance by the reference pattern creating means according to claim 1, and a new reference pattern is added to the plurality of reference patterns, and a plurality of new reference patterns are newly added. Generating a plurality of reference patterns by using a reference pattern adding unit that obtains a correlation output between a pattern belonging to the added class to be recognized and classified and the new plurality of reference patterns. Pattern recognition classification method. 3. A method of selecting a pattern having the lowest correlation output with respect to the entire set of reference patterns is performed by selecting a total of correlation outputs between each of a group of patterns prepared in advance as candidates for the reference pattern and the set of reference patterns. And selecting a pattern having the lowest value. 4. A method of selecting a pattern having the lowest correlation output with respect to all of the reference pattern sets is performed by selecting a pattern group prepared in advance as a reference pattern candidate and a sum of squares of correlation outputs of the reference pattern sets. And selecting a pattern having the lowest value. 5. A method of selecting a pattern having the lowest correlation output with respect to all of the sets of reference patterns is performed by selecting a total of correlation outputs between each of a group of patterns prepared in advance as candidates for the reference pattern and the set of reference patterns. 2. The pattern recognition and classification method according to claim 1, wherein the sum of squares is compared, and a pattern having a lower value in each of the sums is selected. 6. The generation of the reference pattern, wherein the correlation output between the plurality of reference patterns and the test pattern is a peak value of an autocorrelation output or a cross-correlation output between the test pattern and each of the plurality of reference patterns. And means for adding the reference pattern.
The pattern recognition classification method described in 1. 7. A correlation output between the plurality of reference patterns and the test pattern, the total light amount in a light receiving range corresponding to the spread of the auto-correlation output or the cross-correlation output between the test pattern and each of the plurality of reference patterns, or 3. The pattern recognition / classification method according to claim 1, wherein a means for creating the reference pattern having an average light amount and an means for adding the reference pattern are used. 8. A correlation output between the plurality of reference patterns and the test pattern, a peak value of an autocorrelation output or a cross-correlation output of the test pattern and each of the plurality of reference patterns,
2. The apparatus according to claim 1, wherein said reference pattern creation means and said reference pattern addition means, which are the total light amount or the average light amount in the light receiving range according to the spread of the correlation output, are used.
Alternatively, the pattern recognition / classification method described in 2. 9. A correlation output between the plurality of reference patterns and the test pattern is obtained by converting an output value from a two-dimensional light receiving element into a test pattern or the amount of light transmitted or reflected from the test pattern and the plurality of reference patterns. 3. The pattern recognition and classification method according to claim 1, wherein said reference pattern creating means and said reference pattern adding means are used as a standardized amount. 10. A correlation output between the plurality of reference patterns and a test pattern is obtained in a linear range of a two-dimensional light receiving element or in a range where a part of the correlation output exceeds a maximum light receiving amount of a two-dimensional photoelectric conversion element. 3. The pattern recognition and classification method according to claim 1, wherein said reference pattern creating means and said reference pattern adding means are used. 11. A correlation output between the plurality of reference patterns and the test pattern, the reference pattern creating means and the reference pattern adding means obtained by using a filter created by using a Fourier transform hologram. 3. The pattern recognition and classification method according to claim 1, wherein the method is used. 12. A correlation output between the plurality of reference patterns and the test pattern is obtained by converting an intensity pattern obtained by individually performing a joint Fourier transform of the test pattern and each of the reference patterns into an optical pattern. The intensity pattern obtained by performing the Fourier transform individually or collectively performing the joint Fourier transform of the test pattern and each reference pattern by the Fourier transform lens array is optically transformed again by the Fourier transform lens array. 3. The pattern recognition and classification method according to claim 1, further comprising a unit for creating the reference pattern by performing a Fourier transform at a time and a unit for adding the reference pattern. 13. A correlation output between the plurality of reference patterns and the test pattern is obtained by superimposing a test pattern represented by a light transmittance distribution or a light knitting rate distribution on each reference pattern and irradiating the test pattern with incoherent light. 3. The pattern recognition / classification method according to claim 1, further comprising a means for creating the reference pattern by taking an amount of reflected or transmitted light, and the additional means. 14. A correlation output between the plurality of reference patterns and the test pattern, the reference pattern or the intensity pattern of the Fourier transform of the reference pattern, the transmission pattern or the reflectance distribution on the Fourier transform plane of the test pattern. Mask to represent,
Or it is arranged as a spatial light modulator, and it is obtained by taking the amount of light of the test pattern that has transmitted or reflected it,
Alternatively, the intensity pattern of the Fourier transform of the test pattern is arranged on the Fourier transform surface of the reference pattern as a mask or a spatial light modulator representing a transmittance distribution or a reflectance distribution and transmitted or reflected. 3. The pattern recognition / classification method according to claim 1, wherein said reference pattern creating means and said reference pattern adding means for simultaneously obtaining light amounts of one or a plurality of reference patterns are used. 15. A correlation output between a pattern belonging to a class to be recognized and classified and the plurality of reference patterns is obtained a plurality of times, and a representative value of a set of a set of correlation degrees obtained and a quantity indicating a variation are set as a reference. After creating a membership function corresponding to the plurality of reference patterns for each class to be recognized, the correlation output between the test pattern and the plurality of reference patterns and the membership assigned to each class to be recognized and classified. Take the membership value with the function,
Means for creating the reference pattern and means for adding the reference pattern such that the membership value having the smallest membership value or the average value of the membership values is such that the test pattern belongs to the class to be recognized and classified. 3. The pattern recognition and classification method according to claim 1, wherein 16. A pattern belonging to a class to be recognized and classified and a pattern belonging to a first class to be recognized and classified by using a correlation output of the plurality of reference patterns as an input value of an input neuron. First
And a weight corresponding to the distance between the first correlation output value and the second correlation output value between the first correlation output value, the plurality of reference patterns, and the pattern belonging to the second class to be recognized and classified. 3. The pattern recognition and classification method according to claim 1, wherein said reference pattern creating means and said reference pattern adding means by a simple neural network are used. 17. A method according to claim 1, wherein a correlation output value between a pattern belonging to a class to be recognized and classified and said plurality of reference patterns is used as an input value of an input neuron, and a quantity corresponding to a distance in a feature space is used as an evaluation function. In the space, the class to be recognized and classified uses the means for creating the reference pattern and the means for adding the reference pattern by a neural network for forming an identification area surrounded by a hypersphere for each class. Item 3. The pattern recognition classification method according to item 1 or 2. 18. A pattern belonging to a first class to be recognized and classified, a first correlation output value between the plurality of reference patterns, and a first correlation output value and the first correlation output value and the plurality of reference patterns to be recognized and classified. In a feature space, a correlation output value of another class closest to a correlation output value of a certain class in a feature space based on an amount corresponding to a distance between a pattern belonging to the second class and a second correlation output value. Using the reference pattern creating means and the reference pattern adding means as a closed area surrounded by a hypersphere having a radius of the distance to the midpoint of the distance as an identification area for each class. The pattern recognition classification method according to claim 1.