JP2834894B2 - Optical 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 an optical pattern recognition and classification method used in the field of optical information processing.
That is, the present invention relates to a recognition associating process and a classification process, and more particularly, to an optical pattern recognition and classification method by arithmetic processing of information processing in the optical measurement field and image processing field.

[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. The former method performs a Fourier transform on a two-dimensional reference pattern, irradiates a reference wave, creates a so-called Fourier transform hologram, and uses the hologram as a filter to superimpose the Fourier transform image of the test pattern to obtain a correlation. It performs the operation. In the latter 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 identified and classified, and the pattern having the maximum correlation output with respect to the presentation of the test pattern, that is, the autocorrelation output. Since the result obtained based on the method is a detection result, the number of reference patterns to be stored increases with an increase in the number of reference patterns, and the load on the two-dimensional or three-dimensional spatial light modulator increases. The ability as a discriminating and classifying device is also largely governed by the capability of a spatial light modulator, and cannot be used as a substantially excellent optical discriminating associative device.

Further, the inventors have introduced a feedback system for changing the amount of light applied to the reference pattern based on the degree of correlation obtained by the latter method, thereby increasing the number of reference patterns and reducing the number of test patterns. Although it has been clarified that the efficiency of identification and association can be improved, it is desired to further increase the number of test patterns that can be substantially recognized and classified. In addition, many recognition and classification methods using neural networks have been proposed, but if it is necessary to add a new pattern that needs to be recognized and classified after learning, the weight of the connection must be re-learned and created. It was not rational because it took a lot of time to learn.

[0005]

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and it is possible to discriminate and classify a large number of test patterns with a very small number of reference patterns to be recorded. It is an object of the present invention to provide a method for optical pattern recognition and classification that can be performed.

[0006]

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned technical problem. In an optical pattern recognition and classification method, a pattern belonging to each class to be identified and classified and a plurality of reference patterns are provided. Each of the correlation outputs with the pattern is measured a plurality of times , and a member corresponding to the plurality of reference patterns is classified for each class to be recognized and classified based on a parameter indicating a representative value and a variation of a set of obtained correlation degrees.
A ship function is created, and the size of the overlapped area of the membership functions between the classes to be recognized and classified with respect to the same reference pattern, or the largest area in the overlapped area of the membership functions . the degree of overlap of the membership values to determine the combination of the reference pattern to decrease the degree of overlap of the then test pattern
And the reference pattern determined by the above combination
For each class for which the correlation output is taken and the correlation output is to be identified and classified
Each of the above membership functions assigned to
Take the membership value of each
Membership value with minimum membership value
Of the test pattern by the arithmetic mean of the
And an optical pattern recognition / classification method characterized in that it is determined whether the class belongs to a class to be recognized / classified.

A method of reducing the degree of the overlap is to calculate the number of regions having an area where the size of the overlapped area is larger than a predetermined size, and to calculate the number of regions from the reference pattern having the smallest number. A reference pattern is added in ascending order to form a set of reference patterns, and in all two combinations of classes to be recognized and classified, the number of reference patterns included in the set of reference patterns and the predetermined size The number of overlapping regions of the membership function for the same reference pattern having an area larger than the number of overlapping regions is obtained, and a set of reference patterns at a stage where this number is smaller than a predetermined number is determined. What is achieved by doing so is preferable.
Further, a method of reducing the degree of overlap, the maximum membership values in the heavy <br/> were each other becomes region of, and calculates the number of areas with a large membership value than a predetermined magnitude A reference pattern is added in ascending order from the reference pattern having the smallest number as a set of reference patterns, and in any combination of any two of the classes to be recognized and classified, the reference patterns included in the set of reference patterns are The number is obtained when the number and the number of overlapping areas of the membership function for the same reference pattern having a membership value larger than the predetermined size are obtained, and this number is larger than the predetermined number. What is realized by setting a set of reference patterns at a smaller stage is preferable.

The membership function preferably defines a convex fuzzy set. Then, the correlation output between the plurality of reference patterns is suitably those taking the peak value of the autocorrelation output or the cross-correlation output between each reference pattern of multiple <br/> number of the the test pattern. Also, the correlation output between the plurality of reference patterns, which takes a total light or the average amount of received light range according to the spread of the autocorrelation output or the cross-correlation output between該被test pattern and the plurality of the reference pattern Is preferred. Also, the correlation output between the plurality of reference patterns, the total amount of the light receiving range in response to the spread of the correlation output between the peak value of the autocorrelation output or the cross-correlation output between each of the plurality of reference patterns of the with the test pattern or it is suitable that averaged light quantity. Further, the correlation output with the plurality of reference patterns is obtained by converting the output value itself from the two-dimensional light receiving element or the output value from the two-dimensional light receiving element into the test pattern or the amount of transmitted light from the test pattern and the plural reference patterns. Alternatively, an amount standardized by the amount of reflected light is preferable. It is preferable that the correlation output with the plurality of reference patterns be obtained in a linear range of the two-dimensional light receiving element or in a range where a part of the correlation output exceeds the maximum light receiving amount of the two-dimensional photoelectric conversion element.

The correlation output with a plurality of reference patterns is preferably obtained using a filter created using a Fourier transform hologram. Further, the correlation output with a plurality of reference patterns is obtained by optically individually Fourier transforming again the intensity pattern obtained by individually performing the joint Fourier transform of the test pattern and each reference pattern, or It is preferable that the intensity pattern obtained by collectively performing the joint Fourier transform of the detection pattern and each reference pattern is collectively optically collectively subjected to the Fourier transform again. The correlation output with the plurality of reference patterns is obtained by irradiating the test pattern represented by the light transmittance or the light reflectance distribution and each reference pattern with incoherent light, and taking the amount of light reflected or transmitted. Are preferred.

[0010]

According to the above-described optical pattern recognition / classification method, the features included in each reference pattern are shared by obtaining the degree of correlation between the patterns belonging to some classes to be recognized and classified and the reference patterns. The degree of correlation is obtained for each pattern to be recognized and classified by the number of reference patterns.This set of correlations takes into account the temporal instability of the spatial light modulator and the fluctuation of the correlation due to spec noise. Expressed as a membership function.

Here, in order to optimally classify a class to be recognized and classified by a small number of reference patterns, a set of reference patterns is set so that the overlap of membership functions of the classes to be recognized and classified for the same reference pattern is reduced. This is realized by selecting and selecting membership functions corresponding to these few reference patterns. Next, when the test pattern is classified and classified, the degree of matching with the class to be recognized and classified is determined by comparing the degree of coincidence with the class to be recognized and classified by the minimum value or the average value of the membership values. You can get more.

Next, the optical pattern recognition / classification method of the present invention will be specifically described with reference to examples, but the present invention is not limited thereto.

[0013]

FIG. 1 is a schematic diagram showing the function of an optical correlation processing apparatus according to an example of an optical pattern recognition / classification method of the present invention. In the optical layout diagram of FIG. 1, the optical correlation processing device essentially comprises an image output means 1, an optical Fourier transform means 2, a pattern output means 3, an optical Fourier transform means 4, and a light detection means 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 formed on a screen 31 by a Fourier transform lens 21. . This joint Fourier transform pattern is read by a two-dimensional photoelectric conversion element 32, drawn on an electric address type liquid crystal light valve (hereinafter abbreviated as LCLV) 35 by a liquid crystal drive circuit 33, read out by a light beam 37, and subjected to a Fourier transform lens 41. , The Fourier transform is performed again, and the correlation intensity on the screen 42 is detected by the two-dimensional photoelectric conversion element 43.

In the method of the present invention, after the correlation between the pattern to be recognized and classified and each reference pattern is detected by the above-mentioned optical system, the correlation between each reference pattern and each pattern to be recognized is calculated by the computer 51. , A membership function corresponding to the reference pattern is created, and recognition and classification of an unknown image are performed by collation with the membership function.

Here, an example of a method for creating the membership function will be described below. For example, three character patterns E, V, and X are selected as a reference pattern group, and a character pattern to be recognized and classified is set to H, as shown in FIG.
It is assumed that two character patterns are drawn on the image display device 16 at equal intervals on a concentric circle, and a character pattern to be recognized and classified at the center of the concentric circle. In this case, the output of the two-dimensional photoelectric conversion element 43, that is, the correlation output between the character pattern H and each reference character pattern appears at the corresponding reference character position drawn on the pattern display device, as shown in FIG. Is detected as light intensity.

Here, the larger the peak value of the correlation output intensity, the larger the circle. Because, H
This is because the two parallel lines and the horizontal bar constituting the above react with the horizontal bar and the vertical bar of the components of each reference character pattern, resulting in a difference in the correlation output intensity as illustrated. In FIG. 3, the autocorrelation pattern appearing on the optical axis is omitted here because it has no relation here.

The correlation output intensity obtained in this way is sent to the computer 51 and stored in the computer 51. Similarly, patterns to be identified and classified are successively presented on the image display device 16 to obtain the peak value of the obtained correlation intensity. Are further stored in the computer 51 one after another. Further, these operations are performed plural times for each pattern to be identified and classified. This is because the peak value of the correlation output intensity fluctuates due to the influence of the spec noise due to the coherent light, the temporal instability of the image display device 16 and the LCLV 35, and the input timing of the signal of the two-dimensional photoelectric conversion element 32. This is because accurate identification and classification cannot be performed if the recognition and classification are performed using the correlation output values of the times. Therefore, the average value and the standard deviation amount of the peak values for the patterns belonging to each class are calculated from the data of a plurality of times, and the membership function of the peak values of the correlation output intensities for the patterns of E, V, and X is trapezoidal. The membership value of 1 from the average value to the range of the standard deviation is defined as a membership value of 1, and the value of the trapezoidal hypotenuse is equal to the size of the standard deviation. Up to 3 times.

FIG. 4 is a graph showing the peak value of the correlation output for E, V, and X as a membership function when the pattern to be identified and classified as described above is H. In this way, membership functions are created for all the classes to be recognized and classified, and stored in the computer 51 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.

In the recognition of the unknown pattern, the membership value between the membership function for each reference pattern obtained by the above method and the degree of correlation of the unknown pattern is taken, and the smallest membership value among the membership values is obtained. Alternatively, since the average value is 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 separation capability can reduce the probability of erroneous recognition and reduce the recognition time. It leads to reduction.

Hereinafter, a method for creating such a reference pattern set will be described. It is assumed that the membership functions for the patterns H, N, and W to be recognized and classified as described above are those shown in FIGS. 4, 5, and 6, respectively. Here, the following operation is performed in order to classify rationally with a set of reference patterns as small as possible. That is, for the same reference pattern, the size of the overlapping area of the membership functions of the patterns to be recognized and classified is expressed as the degree of overlap,
The number of regions having this area larger than a predetermined size is calculated, and reference patterns are added in ascending order of the smallest number to form a set of reference patterns. Next, in any combination of any two of the classes to be recognized and classified, the number of reference patterns included in the selected set of reference patterns and the membership for the same reference pattern having an overlapping area larger than a predetermined size The number in the case where the number of areas where functions overlap is found is determined, and the addition of the reference pattern is stopped when this number becomes smaller than the desired number. Hereinafter, the method will be specifically described.

For example, for three patterns H, N, and W, there are three combinations of two patterns, and for each of them, the area where the membership function of each reference pattern overlaps is calculated. And the table shown in FIG. Here, the larger the overlapping area of the membership functions of each reference pattern is, the more the corresponding two patterns cannot be identified. Therefore, for example, if it is attempted to almost completely distinguish between all patterns, a pattern having an area much larger than 0 becomes inconvenient. Therefore, for those where the overlapping area of the membership functions of each reference pattern is larger than 0.5,
When it is assumed that misrecognition is likely to occur and the number of regions having an area of 0.5 or more is counted, the reference patterns E, V, and X are 1, 2, and 3, respectively. Therefore, first, the pattern E having the smallest degree of overlap is selected, and then the pattern V having the smallest degree of overlap is selected. At this time, the number of sets of the reference patterns is two, and the area of the overlap of the membership functions composed of two combinations of the patterns to be recognized and classified with respect to the reference patterns E and V is:
Since all the numbers greater than 0.5 become 1,
The set of reference patterns that can be identified for these three patterns is two patterns E and V, and the pattern X becomes unnecessary.

It is also possible to select a reasonable set of reference patterns based on other criteria. That is, first, for the same reference pattern, the maximum membership value in the overlapping area of the membership functions of the pattern to be recognized and classified is represented as the degree of overlap, and a value whose magnitude is larger than a predetermined magnitude is expressed as a degree of overlap. The number of areas having the reference patterns is calculated, and reference patterns are added in ascending order from the smallest number to form a set of reference patterns. Next, in any combination of any two of the classes to be recognized and classified, the number of reference patterns included in the selected set of reference patterns and the members for the same reference pattern having a degree of overlap greater than a predetermined size The number when the number of the overlapped areas of the ship functions coincides with each other is determined, and when this number becomes smaller than the desired number, the addition of the reference pattern is stopped. Hereinafter, specifically,
The method is shown below.

For example, for three patterns H, N, and W, there are three combinations of two patterns, and for each of the three patterns, the maximum value in the area where the membership functions of the respective reference patterns overlap is obtained. When the membership value is calculated, a table shown in FIG. 8 is obtained. Here, the larger the maximum membership in the area where the membership function of each reference pattern overlaps, the larger the corresponding 2
This means that it is not possible to distinguish between the two patterns. Thus, for example, if one wants to almost completely discriminate between all the patterns, one in which the above-mentioned maximum membership value is considerably larger than 0 would be inconvenient. Therefore,
If the maximum membership value in the overlapping area of the membership functions of each reference pattern is larger than 0.5, it is assumed that misrecognition is likely to occur, and 0.5
Counting the region having the above area, the reference pattern E,
V, X are 1, 2, and 3, respectively. Therefore, first, the pattern E having the smallest degree of overlap is selected, and then the pattern V having the smallest degree of overlap is selected. At this time, the number of sets of reference patterns is two, and two of the patterns to be recognized and classified with respect to the reference patterns E and V are two.
One of the largest membership value definitive in overlapping areas of the membership functions consisting of the combination is, the number of larger than 0.5, since all become 1, a set of identifiable reference pattern for the three patterns , Patterns E and V, and the pattern X becomes unnecessary.

As described above, when the scale for evaluating the degree of overlap of the membership functions is set to the overlapping area, in this embodiment, a range larger than the standard deviation forms the membership function. Since the probability that the degree of correlation of the test pattern is observed in this range is statistically about 33%, the overlapping area is represented by such a stochastic distribution. Since it is reflected to some extent, it has the advantage that it is easy to adapt when recognizing an actual unknown pattern. On the other hand, when the scale for evaluating the degree of overlap of the membership functions is set to the maximum value of the membership values in the overlapping area, the worst value in the recognition process can be assumed.
This is a method in which in the recognition process, a membership value between a correlation degree of an unknown pattern and a membership function of each reference pattern is obtained, and the minimum or average value is used as a recognition result. For patterns E and V, patterns H and W to be identified and classified may have a membership value of 0.2 at the worst.
In this case, two patterns cannot be recognized. However, such a case is extremely rare and hardly occurs in actual recognition.

Next, the reason why the above method is effective will be described. That is, for example, N reference patterns that do not overlap as much as possible on the Fourier transform surface are prepared, and M patterns to be recognized and classified are M. In addition, the standard deviation of the observed degree of correlation is 3
Assuming that the area of double spread is R and the dynamic range of the detector is D, the number of non-overlapping areas, that is, the number B of orthogonal areas, is about B = D / R. Here, assuming that the spatial frequencies of the patterns to be recognized and classified are various, it is possible to discuss stochastically to some extent. That is, for one wants to recognize classification class, the number G when put N reference patterns B number of regions, B ^N, and the relative class to be recognized classification into M, G = B
^{(N + M)} . On the other hand, in order to be completely recognizable, M classes can be realized by reserving separate areas for each of the numbers in all cases.

[0027] In other words, the number of the case, the _G P _M. Therefore, the probability P becomes P = _G P _M / _G. Now, let's say that there are 1000 classes to be recognized and classified.
Assuming that 10 or 20 orthogonal regions can be secured, the table shown in FIG. 9 can be obtained by examining how much reference pattern is required. It is clear from this table that the number of reference patterns that can almost completely identify 1000 classes is that the number of orthogonal regions is one.
0, in the case of 20, each 8 or more, can <br/> be referred to as 7 or more. Therefore, as described above, by adding reference patterns while calculating the degree of overlap of membership functions, it is possible to almost certainly derive a combination of reference patterns that separates all classes.

A method of performing recognition and classification using the set of reference patterns E and V obtained as described above will be described in detail below. 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 taken into the computer 51 and collated with the membership function of the pattern to be recognized and classified. I do. For example, assume that the membership functions for the patterns H, W, and N to be recognized and classified are those shown in FIGS. 4, 5, and 6, respectively.
At this time, it is assumed that the peak value of the correlation output intensity of the unknown character pattern is output at point a for pattern E and at point b for pattern V. FIG. 10 is a graph showing the degree of membership of the points a and b with respect to each pattern to be recognized and classified. In the above example,
Although it has only the values of 1 and 0, it may have a value in a portion corresponding to the hypotenuse of the trapezoid. In general, the membership value has an analog value.

By performing a fuzzy logical AND operation from the membership values shown in FIG. 10, the minimum value of the membership values of the patterns to be recognized and classified is obtained. H, W,
N is 1, 0, 0, respectively. Therefore, the unknown pattern can be said to be 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 a certain pattern can be estimated by using the values of the minimum values. In addition, the arithmetic mean of the membership values shown in FIG. 10 can be used as a judgment material. That is, in this case, for the unknown pattern, the patterns H, W, and H are respectively
Since they are 1, 0.5, and 0, the degree of each pattern can be estimated. However, when the arithmetic mean is used, an answer may be given even if the degree of correlation with a certain reference pattern does not match at all, so that a fuzzy logical AND
The operation has less error, but is meaningful as an auxiliary operation when it is difficult to obtain an answer by an AND operation.
In other words, if for some reason the average value is significantly different from the state when the average value is taken, the membership value is 0 for only one reference pattern and 1 for the other reference patterns,
An AND operation cannot provide a correct answer.
In this case, it is possible to use an arithmetic mean as a second candidate.

In the above-described embodiment, the membership function is trapezoidal. However, any membership function can be used as long as it has a so-called convex fuzzy set with no depression. ,I do not care. Also, the standard for creating the membership function in this embodiment is taken as the standard deviation,
Although about three times this is used as a measure of the spread of the membership function, this spread may be determined in any way. For example,
If the area where the membership value becomes 1 is set to be twice the standard deviation, statistically, about 97% of the detected degree of correlation is obtained.
But since entering this area, it is possible to detect the class to be recognized classification at a ratio of almost 100%, for the membership function spreads, correspondingly, will Dzu pleasure to the separation of other classes.

The correlation output intensity can be determined in the same manner by obtaining the total light amount in the light receiving range according to the spread of the correlation output or the average light amount other than the peak value. The correlation output generally represents the two patterns as A (x,
y) and B (x, y) are represented by the following equations. I (x ′, y ′) = ∬A (x, y) B ^* (x′−x, y′−y) dxdy Here, ^* indicates a complex conjugate amount. From this equation, it can be seen that the correlation output has twice the size of the pattern. Therefore, from the peak value, it is possible to obtain the total light amount in the area in consideration of the spread or the average light amount. FIG. 11 schematically shows the distribution of the correlation light amount. The overlapping state of the two patterns is expressed by being shifted in the vertical direction on the paper. In this case, for example, the correlation light amount between the pattern E and the pattern H is a
As shown in the figure, when they completely overlap, the peak value is reached,
As shown in FIG. 11B, when the pattern is displaced by the size of the pattern, one vertical bar of the pattern H and the vertical bar of the pattern E overlap with each other.

FIG. 11C shows the correlation light amount distribution of the pattern E and the pattern H created in this manner. A large circle at the center indicates the peak light amount, and small circles at the top and bottom indicate the horizontal bar. The reaction is indicated by a small circle next to the vertical bar. Therefore, during the total light amount or the average light amount, the feature of each pattern not seen in the peak light amount is as follows.
Information that is better reflected can be obtained.

FIG. 12 is a diagram showing the identification and classification of the 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 in an actual experiment using the above-described joint Fourier transform. The membership function of the desired pattern (15 alphabetic patterns) is shown. The solid line is a membership function for pattern E, and the dashed line is
The horizontal axis represents the membership function for the pattern V, and the horizontal axis represents the average light amount of the correlation output. However, 1
Some have only one membership function, but this is omitted because the other membership function did not fit within the horizontal scale. FIG. 13 shows the results obtained by experiments using the joint Fourier transform method. Unknown pattern of the alphabet 15
Each of the patterns represents the degree of identification and classification of each pattern derived from the fuzzy logical AND operation result obtained when the pattern is presented three times. The alphabet on the left side is the input pattern, and the numerical values in the first and second parentheses indicate the average light amount of the correlation measured for the reference patterns E and V, respectively. The alphabet on the right is a fuzzy logical AN of the membership values in FIG.
D is the recognition character obtained by performing the D operation, and the numerical value is AND
The calculation result is multiplied by 100.

Although there are several such confusing ones, if the highest one is selected, all the patterns are identified by only two reference patterns. This is not seen in the conventional correlation processing using general autocorrelation. In this experiment,
For simplicity, the creation of the reference pattern set is shown in FIG.
As shown in FIG. 8, since the reference patterns E and V which can completely separate the alphabets H, N and W are used, the separation of the other test patterns is slightly deteriorated. FIG. 13 shows that the pattern T,
The set of L and I and the set of patterns K and M are particularly confusing, because the characteristics of each pattern are similar. It can be used to perform or classify. However, even when this effect is completely removed and the system is intended to be used for performing full recognition, it can be realized by a very simple process. That is, this can be realized by introducing one pattern in which the membership functions between confusing patterns do not overlap. This is because if the membership value is at least 0, it is recognized as a different pattern in order to perform a fuzzy logical AND operation.

Further, it is needless to say that almost all separated classes can be recognized by using the method of creating a reference pattern set described in this embodiment. Further, even when the number of classes and patterns to be identified and classified increases, it is possible to easily cope with the case. This is because it is only necessary to create a membership function for the increased pattern. In this case, an operation such as changing all the weights of all the connections connecting the neurons by additional learning as in a normal neural network is completely unnecessary. Further, in the neural network, when the number of classes and patterns to be identified and classified increases, there is a problem that the learning time becomes extremely long due to an increase in the number of neurons. In the pattern identification and classification method of the present invention, such a problem is caused. Does not occur. However, in this case, the increased pattern or class may be quite close to the already stored pattern or class. Avoid such a situation by increasing the number of reference patterns with some margin in advance. Can be.

When used for applications such as association,
The original pattern can be associated with the defect pattern by extending the degree of expansion of the membership function to a value larger than that in consideration of the fluctuation of the optical system, and in particular, by extending the correlation function to a side having a lower correlation strength. Conversely, when associating a main pattern with a pattern to which extra information is added, the original pattern can be associated by extending the extent of the membership function to the side with higher correlation strength. However, in this case, in order to identify a defect pattern or a pattern to which extra information is added, the spread of the membership functions becomes large, the degree of overlap between the membership functions becomes large, and an orthogonal region is formed. Since the number is substantially reduced, the number of patterns that can be recognized and classified is reduced. In order to avoid this, the entire light quantity of the test pattern or the test pattern and the reference pattern may be received, and the detected correlation light quantity may be normalized by this light quantity.

For example, the beam splitter 17 is provided between the image display device 16 and the Fourier transform lens 21 in FIG.
Then, the total amount of light of the test pattern or the test pattern and the reference pattern is received, and the output is sent to the computer 51. The other optical systems are the same as in FIG. 1 and are not shown in FIG. By using the normalized light quantity obtained as described above and normalizing the correlation light quantity of the test pattern with respect to each reference pattern, in the normalization based on the light quantity of the test pattern, for example, for a certain test pattern, correlation amount for each reference pattern is a X _N, this test pattern is the decrease in the correlation amount for each reference pattern in the case of defects, and had a delta _N, the case without normalization, X While the _N - [delta _N, Doing normalized, if the amount of the test pattern in the absence of defects in the test pattern a, a decrease in amount of light when there is a defect as D, normalized The correlation light quantity is X _N / A
And _{(X N -Δ N) / (} A-D) , and the Out little defect, since it is _{A> D, (X N -Δ} N) / A * (1 + D
/ A), and the effect of the loss can be reduced by D / A compared to the case where the data is not standardized.

Therefore, the number of classes to be recognized and classified can be increased accordingly. In the case of normalizing with the total light amount of the test pattern and the reference pattern, it is possible to reduce both the temporal fluctuation and the instability factor of the spatial light modulator and the influence of the defect to some extent. The reason is that if B is the light amount of the reference pattern and E is the variation due to the fluctuation of the spatial light modulator, B> E.
Normalized correlation quantity when without and defective test pattern, _{respectively, X N / (A + B} ), (X N -
Delta _N) / and since (A-D + B-E ), among defect is small, not greater variation of the fluctuation, A + B> D +
Since E is obtained, the effect of the defect and the effect of the fluctuation of the spatial light modulator can be reduced by (D + E) / (A + B) in the same manner as the above-described normalization based on the light amount of the test pattern.

The present invention can also be used for purposes such as determining whether or not there is a defect. In this case, the case is opposite to the case of association, and when the spread of the membership function created in consideration of the fluctuation of the optical system is exceeded, for example, when the standard deviation due to the fluctuation of the optical system exceeds 3 times, Can be used in such a way as to be regarded as having a defect.

In the above embodiment, the reference pattern 2
Although at least 15 patterns could be identified, very roughly, considering the capability of the method of the present invention, for example, the dynamic range of the detected correlation output is 1: 100 and 1: 1. Assuming that the range in which the membership values of the two membership functions are not 0 has a spread of 10, 1 for N reference patterns
This means that an orthogonal area of 0 ^N is created. It suffices if a pattern to be identified and classified is put in this. Thus, a very small number of reference patterns can distinguish and classify a very large number of different patterns. Furthermore, since the fuzzy AND operation is used, it is not necessary to assign a membership function to a completely orthogonal region when a decision is made that allows ambiguity, and the number of patterns that can be identified further increases. Can be.

If 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 are used as the membership function, information having two different properties can be obtained for one reference pattern. . Therefore, the number of reference patterns can be further reduced,
Reliable recognition classification can be performed 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 all of the correlation light amount. Therefore, if the maximum value of the correlation light amount, that is, the peak amount 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, when taking the total light amount or the average light amount in the light receiving range according to the spread of the correlation output, 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 too large value may be obtained. It 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 correlation output peak value is saturated, so that the peripheral pattern is not affected. By setting the light amount of the part that reflects the characteristics better so as to show a linear response, the dynamic range of the two-dimensional light receiving element can be used effectively, so that the correlation output can be expanded. This is very effective when taking the total light amount or the average light amount in the corresponding light receiving range.

In this embodiment, a joint Fourier transformer is used as an optical system for obtaining a correlation output. In this case, however, there is an advantage that addition or rewriting of a reference pattern is easy. Furthermore, in order to improve the classification and classification ability,
It is better not to be a joint Fourier transform correlator. The reason for this is
When the reference pattern group and the test pattern are simultaneously Fourier-transformed, the contrast ratio of the multiple interference fringe patterns on the screen 31 in FIG. 1 decreases with an increase in the number of reference patterns, so that the obtained correlation output decreases. This is because the dynamic range to which the membership function is assigned is reduced, and the pattern discrimination ability is reduced substantially. Therefore, the correlation processing method for avoiding the decrease in the correlation output is more efficient. Next, an embodiment using such a correlation processing method will be described.

[0044]

Embodiment 2 FIG. 15 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 the reference waves for each reference pattern while changing the irradiation direction of the plane wave is used. As shown in FIG. 15, a light beam 12 emitted from a laser 11 is expanded to an appropriate light beam diameter by a beam expander 13, and irradiates a pattern to be identified and classified drawn on an image display device. 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. Next, the luminous flux having the complex amplitude distribution of the pattern is subjected to Fourier transform by the Fourier transform lens 21, and irradiates a multiple interference fringe pattern drawn on the matched filter. At this time,
From a reference pattern having the same spatial frequency spectrum as the pattern to be identified and classified, diffracted light is emitted in the direction of the plane wave used when the reference pattern was created. When this diffracted light is condensed on the screen 31 by the condenser lens 22, one diffracted light becomes a cross-correlation pattern between the reference pattern and the pattern to be identified 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 or the total light amount at that position can be measured. Hereinafter, the process of similarly measuring the correlation output intensity for the other patterns to be identified and classified and creating the membership function is the same as in the first embodiment, and a description thereof will be omitted.

In this case, since the Fourier components of the respective reference patterns are not superimposed on the matched filter,
As the reference pattern is increased, the contrast ratio of the correlation output decreases less. Accordingly, since the dynamic range to which the membership function is assigned is hardly reduced, an extremely large number of patterns can be identified with a small number of reference patterns. It should be noted that the matched filter in the above-described embodiment is often formed by a Fourier transform hologram, but it is needless to say that the same can be performed by the Fourier transform hologram itself.

The same effect can be obtained by other types of joint Fourier transform devices. That is,
In brief, the joint Fourier transform of each reference pattern of the joint Fourier transform and the pattern to be recognized and classified is performed individually and simultaneously in parallel using a lens array.
This eliminates the problem that the Fourier transform components of the respective reference patterns overlap with each other and lower the contrast ratio of the cross-correlation strength between the reference pattern and the pattern to be recognized and classified. Further, unlike the matched filter, the number of reference patterns can be increased or rewritten, so that more flexible processing can be performed. For example, in recognition, when a reference pattern set in advance cannot be completely separated, a request for rewriting a part of the reference pattern can be answered. Since the first and second embodiments are based on a coherent optical system, they have a feature that they are resistant to translational deviation of an input pattern. This is because the correlation output position moves correspondingly to the translation shift of the input pattern.

Next, an example of an optical system that is simple and relatively easy to integrate will be described.

Third Embodiment The example shown in FIG. 16 is a known incoherent correlation optical system. As an optical system, the image display device 16 and the reference pattern mask 62 are arranged close to each other, and the output of the product of the pattern drawn on the image display device 16 and the reference pattern is output to the screen 31 using the condenser lens array 71. The output from each of the reference patterns collected on the upper side and collected by the two-dimensional photoelectric conversion element 43 is detected. On the image display device 16, patterns to be identified and classified are displayed in an array for pixels corresponding to the reference pattern mask 62. In this case, a cross-correlation peak between each reference pattern and the pattern to be identified and classified is detected. In the operation described below, other patterns to be identified and classified are sequentially displayed on the image display device, and these output values are used for each of the patterns to be identified and classified based on membership of the number of each reference pattern. By creating a function, the same operations as in the first and second embodiments can be performed. However, in this case, there is no tolerance for the translational deviation of the input pattern, but there is a characteristic that the configuration is very simple and the integration is easy.

The optical correlation processing system used in the above embodiment may be of any type. Various optical systems have also been proposed for the joint Fourier transform correlation optical system, but the screen 31 and the two-dimensional photoelectric conversion element 3 shown in FIG.
2. The same can be achieved by using an optically-addressed liquid crystal light valve instead of the pattern processing device 33 and the electrically-addressed liquid crystal light valve 35.

In the present invention, there is a difference in the specification of the part functioning as a spatial light modulator, but in principle, all of them have the same electrical address type and optical address type. Can be used. Further, an analog type or a digital type can be used. Most of the devices described below belong to analog devices, but devices using ferroelectric materials correspond to digital devices. Examples of the electric address type include PLZT and KD in addition to the liquid crystal light valve described above.
Those to which an electro-optical effect is added, such as P and BSO, are often used. Also in the case of the optical address type, a material in which a photoconductive layer is combined with the same material as the electric address type is generally used. However, in a crystal having a photovoltaic effect such as BSO or BaTiO ₃ , a light-induced refractive index change occurs due to spontaneous polarization according to the input light intensity. Therefore, these spatial light modulators can be either a transmission type or a reflection type. It can also be configured as However, in the case where the read light completely cancels the information of the write light in the optical address type, the wavelength range of the read light and the write light is separated, and the read light does not affect the write information. It is necessary to take measures such as doing so. When an electric address type is used, a two-dimensional photoelectric conversion element for obtaining the input information and a driving circuit therefor are required, but there is an advantage that the signal is easily processed.

[0050]

As described above, the above-described effects are obtained by the optical pattern recognition / classification method of the present invention. Summarizing them has the following remarkable technical effects. That is, first, the time-series fluctuations of the spatial light modulator and the time-series fluctuations such as speckle noise are included in the membership function of the degree of correlation created based on the statistical amounts, so that the analog function is obtained. An optical pattern recognition and classification method that can remove the uncertainty of the judgment in the processing can be provided. Second, an optical pattern capable of performing extremely accurate recognition and classification by taking the membership value of the correlation function of each reference pattern and the membership function of the pattern to be recognized and performing a fuzzy logical AND operation. A recognition classification method could be provided.

Third, by determining an optimal combination of a set of reference patterns, an optical pattern that can quickly and accurately recognize and classify a large number of patterns by only presenting a small number of reference patterns. A recognition classification method could be provided. Fourth, an optical pattern recognition / classification method that can be easily executed even when a new pattern to be recognized / classified is added, by creating a membership function of the degree of correlation of each reference pattern with respect to this pattern. Could be provided.

Fifth, even if the recognition result is confusing,
By adding a small number of reference patterns and newly creating a membership function for this pattern, it was possible to provide an optical pattern recognition and classification method that can be easily and completely recognized. Sixth, an optical pattern recognition and classification method capable of performing flexible processing by adjusting the extent of the membership function separately from the fluctuation of the optical system according to the use of recognition, classification, association, and the like. Could be provided.

Seventh, in the case of discriminating and classifying a pattern to which missing or extra information has been added, the correlation output is normalized by the amount of transmitted or reflected light of the test pattern and the reference pattern, thereby reducing the discriminating and classifying ability. An optical pattern recognition and classification method with less occurrence could be provided.

[Brief description of the drawings]

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

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

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 optical 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 optical 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 optical pattern recognition and classification method of the present invention.

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

FIG. 7 is a table showing an overlapping area of membership functions of a combination of patterns as one method of creating an optimal reference pattern set in the optical pattern recognition and classification method of the present invention.

FIG. 8 shows the maximum membership value in the overlapping area of the membership function of each pattern combination as another method of creating an optimal reference pattern set in the optical pattern recognition and classification method of the present invention. This is the table shown.

FIG. 9 is a table stochastically calculating the number of reference patterns required to identify 1000 patterns in the optical pattern recognition / classification method of the present invention.

FIG. 10 is a diagram illustrating a membership value of a correlation degree measured for an unknown pattern with respect to a pattern to be recognized and classified in the optical pattern recognition and classification method of the present invention.

FIG. 11 is a diagram schematically showing a cross-correlation pattern used in the optical pattern recognition / classification method of the present invention.

FIG. 12 is a diagram illustrating membership values of 15 patterns to be recognized and classified for two reference patterns experimentally obtained in the optical pattern recognition and classification method of the present invention.

FIG. 13 is a diagram showing a recognition result for an unknown pattern using an experimentally obtained fuzzy AND operation and the accuracy of the recognition result in the optical pattern recognition and classification method of the present invention.

FIG. 14 is a diagram schematically showing a part of an optical system used for normalizing a detected degree of correlation in the optical pattern recognition / classification method of the present invention.

FIG. 15 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 optical pattern recognition / classification method of the present invention.

FIG. 16 is a schematic configuration diagram of still another example of an optical system used as an optical correlation processing device in the optical pattern recognition / classification method of the present invention.

[Explanation of symbols]

 1, 3 pattern 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 pattern display device 21, 41 Fourier transform lens 22 condensing lens 31, 42 Screen 32, 43 Two-dimensional photoelectric conversion element 33 Liquid crystal drive circuit 34 Mirror 35 Liquid crystal light valve 51 Computer 61 Matched filter 62 Reference pattern mask 71 Condensing lens array

──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Hideki Yamazaki 585 Tomimachi, Funabashi-shi, Chiba Sumitomo Cement Co., Ltd. New Business Headquarters (58) Field surveyed (Int.Cl. ⁶ , DB name) G06F 7 / 60

Claims (12)

(57) [Claims] 1. A optical pattern recognition classification method, the correlation output each belongs patterns and a plurality of reference patterns for each class to be identified classification multiple measurements, a representative value and the variation of the set of correlation obtained Based on the parameters shown, a membership function corresponding to the plurality of reference patterns is created for each class to be recognized and classified, and membership among the classes to be recognized and classified for the same reference pattern the degree of overlap of the maximum membership value in size or membership function area overlapping comrades area of overlap of functions comrades,
The combination of the reference pattern that reduces the degree of the overlap is determined, and then the combination is determined by the combination of the test pattern and the above-described combination.
The correlation output with the reference pattern is taken, and the correlation output is assigned to each class to be identified and classified.
Each membership function is checked against each
Membership value, and the highest
Is smaller than the membership value or the membership value
According to the averaging value, the test pattern
An optical pattern recognition / classification method, characterized in that it is determined whether the optical pattern belongs to a class . 2. The method for reducing the degree of overlap is as follows.
The size of the overlapping area is calculated as the number of regions having an area larger than a predetermined size, and the reference patterns are set by adding reference patterns in ascending order from the reference pattern having the smallest number. And in any combination of any two of the classes you want to recognize and classify,
Determine the number when the number of reference patterns included in the set of reference patterns and the number of overlapping regions of the membership function for the same reference pattern having an area larger than the predetermined size coincide with each other, 2. The optical pattern recognition and classification method according to claim 1, wherein the number of reference patterns is set at a stage where the number is smaller than a predetermined number. 3. The method for reducing the degree of overlap is as follows.
The maximum membership value in the overlapping area is calculated as the number of areas having a membership value larger than a predetermined size, and the reference patterns are added in ascending order from the reference pattern having the smallest number. A set of reference patterns, and in all two combinations of classes to be recognized and classified, the number of reference patterns included in the set of reference patterns and a membership value larger than the predetermined size. The number is determined when the number of overlapping regions of the membership function for the same reference pattern coincides with each other, and the number is set as a reference pattern set at a stage where the number is smaller than a predetermined number. 2. The method for recognizing and classifying an optical pattern according to claim 1, wherein: 4. The method according to claim 3, wherein said membership function defines a convex fuzzy set. 5. A correlation output between the plurality of reference patterns and a peak value of an auto-correlation output or a cross-correlation output between the test pattern and each of the plurality of reference patterns. The optical pattern recognition and classification method according to 1. 6. A correlation output between the plurality of reference patterns and a total light amount or an average in a light receiving range according to a spread of an auto-correlation output or a cross-correlation output between the test pattern and each of the plurality of reference patterns. The optical pattern recognition and classification method according to any one of claims 1 to 4, wherein the light amount is measured . 7. A correlation output of said plurality of reference patterns corresponding to the spread of the correlation output between the peak value of the autocorrelation output or the cross-correlation output between the test pattern and the plurality of the reference pattern and the total amount or the average amount of received light range
Optical pattern recognition classification method according to any one of the preceding claims, characterized in that that. 8. A correlation output between the plurality of reference patterns and the output value itself from the two-dimensional light receiving element or the output value from the two-dimensional light receiving element is obtained by converting the output value from the two-dimensional light receiving element into a test pattern or a test pattern and a plurality of reference patterns The optical pattern recognition and classification method according to any one of claims 1 to 7, wherein the amount is standardized by a transmitted light amount or a reflected light amount. 9. A correlation output with the plurality of reference patterns, wherein the correlation output is obtained in a linear range of the two-dimensional light receiving element or in a range where a part of the correlation output exceeds the maximum light receiving amount of the two-dimensional photoelectric conversion element. The optical pattern recognition and classification method according to any one of claims 1 to 8, wherein 10. The optical pattern recognition according to claim 1, wherein the correlation output with the plurality of reference patterns is obtained using a filter created using a Fourier transform hologram. Classification method. 11. A correlation output between a plurality of reference patterns, wherein an intensity pattern obtained by individually performing joint Fourier transform of a test pattern and each reference pattern is optically individually again Fourier transformed again. The intensity pattern obtained by collectively performing a joint Fourier transform of a test pattern and each reference pattern is obtained by optically collectively performing a Fourier transform again. 10. The method for optical pattern recognition and classification according to any one of claims 9 to 9. 12. A correlation output between the plurality of reference patterns is obtained by irradiating a test pattern represented by a light transmittance or a light reflectance distribution and each reference pattern with incoherent light, and reflecting or transmitting the same. The optical pattern recognition and classification method according to any one of claims 1 to 9, wherein the optical pattern recognition and classification method is obtained by taking an amount of light.