JPH04355732A - Detecting method for optical correlation output - Google Patents

Abstract

PURPOSE:To obtain high recognizing ability and to recognize many pattern with a small number of reference patterns by regarding information where the features of a pattern is reflected as one parameter for pattern recognition and making a difference in obtained correlation output light intensity large. CONSTITUTION:This method consists of optical correlation for obtaining the correlation output light intensity between the pattern to be detected and the respective reference patterns and a means which inputs a signal generated on the basis of the obtained result and recognizes and classifies the pattern. The former is, for example, a congruence Fourier transformation optical system. Namely, one image to be recognized and classified which is drawn on an image display device 16 and the reference pattern are read out with luminous flux 12 projected by a laser 11 and processed by Fourier transformation to form a Fourier-transformed pattern on a screen 31. This is read by a two-dimensional photoelectric converting element 32 and drawn on a light valve 35 by a liquid crystal driving circuit 33, and the image is processed by Fourier transformation again the correlation output light intensity on the screen 42 is divided into plural areas to detect pattern.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for detecting optical correlation outputs used in the field of optical information processing, particularly for pattern recognition and classification of images and other objects. That is, recognition associative processing;
The present invention relates to a method for detecting optical correlation outputs in classification processing, particularly in information processing in the field of optical measurement and image processing.

0002

[Prior Art] Optical pattern recognition and classification based on the correlation output light intensity between patterns obtained by an optical correlation system using a joint transform method or a matched filtering method. In the pattern recognition device, when the test pattern and the reference pattern are the same, the position where the peak light intensity of the correlation output, that is, the autocorrelation output, is obtained (hereinafter referred to as the correlation center position) is fixed. In an optical correlation system using the joint transform method, the peak of the autocorrelation output is shifted away from the optical axis, corresponding to the displayed relative distance between the test pattern and the reference pattern. - Ku appears. In addition, in an optical correlation system using the matched filtering method, to create a matched filter, the irradiation direction of the reference light is changed to expose a photographic plate, etc., but the peak of the autocorrelation output is appears in the irradiation direction of the reference light used for

Conventionally, as a pattern recognition classification method, there has been an autocorrelation type pattern recognition classification method. This involves preparing the same number of reference patterns as the number of patterns you want to recognize and classify, and calculating the autocorrelation output between the test pattern and each reference pattern when a certain test pattern is presented. The pattern with the largest peak value is recognized as the test pattern. However, as the number of patterns to be recognized and classified increases, the number of reference patterns increases.
As the number of patterns increases, the load on the memory device for storing reference patterns increases, making it difficult to recognize patterns efficiently.

Furthermore, as the number of reference patterns increases, in an optical correlation system using the joint transform method, the number of Fourier transforms of the reference pattern and the test pattern increases. The simultaneous interference fringes of the Fourier transform images of each pattern are superimposed on the Fourier transform surface of No. 1, deterioration of the interference fringes occurs, and the difference between the correlated output light intensities of each reference pattern is reduced. , which led to a decrease in the recognition rate. Furthermore, in an optical correlation system using the matched filtering method, there is a similar effect when recording interference fringes between the pattern and the reference light as when using the joint transform method, resulting in a decrease in the recognition rate. was inviting.

[0005] Therefore, we have proposed a cross-correlation type pattern recognition classification method. This is January 24, 1991
As explained in the specification of the Japanese application, a feature space is created based on the cross-correlation output of the test pattern with respect to each reference pattern, and patterns are recognized and classified. It is possible to recognize many test patterns using the reference pattern.

Here, in the cross-correlation type pattern recognition classification method, the method for detecting the correlated output light intensity is to detect the correlated output light intensity within a light receiving range having a certain finite spread around the correlation center position. The average value or peak value of is detected by a two-dimensional photoelectric conversion element such as a CCD, and one output is obtained for one correlation output.

FIGS. 2(a) and 2(b) are explanatory diagrams showing the conventional method of determining the light receiving range when detecting the correlated output light intensity. Here, the correlated optical output in FIG. 2(a) is a simple representation of the cross-correlated optical output between the letters L and I, and the blacked out area is the area where the correlated output optical intensity is strong. , the diagonally shaded range represents a weak portion of the correlated output light intensity. The outer dotted line indicates the range where the entire range of the correlation output pattern is received, and the inner dotted line indicates the range where the correlation output pattern spreads.
Each shows a range in which a portion of the correlated light output is received.

[0008] Generally, the correlation output is expressed by the following equation when two patterns are A(x, y) and B(x, y). I(x,y) =∬A(x,y)B*(x'-X,y'
-Y) dxdy...(1), the range in which the correlation output pattern expands is determined by increasing the vertical and horizontal sizes of the test pattern.
, W, where the vertical and horizontal sizes of the reference pattern are h and w, it has a spread of (H+h)/2 in the vertical direction and (W+w)/2 in the horizontal direction, centering on the correlation center position. It happens. Although only two types of light receiving ranges are shown in FIG. 2, the present invention is not limited to these.

However, in the method for detecting the correlation output light intensity as described above, by setting the light receiving range of the correlation output to be sufficiently larger than the width of the correlation output, shift-invariant detection of the correlation output light intensity is performed. Can be done. Therefore, flexible pattern recognition can be performed with respect to positional deviations of correlation outputs. However, on the other hand, if a cross-correlation output that reflects the characteristics of the pattern is obtained, such as the cross-correlation output between T and I or L and I shown in FIGS. The reason why the parts with strong correlation outputs are different in 2(a) and 2(b) is because the positions of the vertical lines composing the test patterns T and L are different.In other words, the characteristics of the patterns (This is a cross-correlation output that reflects the If detected, the test pattern T
, even though L are visually completely different.
As shown in FIGS. 2(a) and 2(b), the same output was obtained, and information reflecting the characteristics of the pattern that had been obtained was lost.

For the reasons mentioned above, it has been difficult to recognize patterns with high recognition ability. In addition, in an optical correlation system using the joint transform method, the test pattern and each reference pattern are connected to a liquid crystal light valve (hereinafter referred to as L).
(abbreviated as CLV), or in an optical correlation system using a matched filtering method, when the test pattern is displayed on the LCLV, the cross-correlation of the received light may be affected due to lack of contrast on the LCLV. It is not possible to make a large difference in the output light intensity, and the area to which the test pattern belongs in the feature space becomes narrow, resulting in a decrease in the number of recognizable patterns and the pattern recognition rate. was.

Furthermore, after processing the cross-correlation output and auto-correlation output, which are the analog values obtained as described above, by computer, a neural network having a basic structure with a hierarchical structure and three or more layers is created. - When inputting to each unit of the input layer of the test pattern, the difference in the detected correlation output light intensity cannot be taken very large, so the input vectors to each unit for each test pattern are similar. Therefore, in the feature space of the neural network, each test pattern with similar input vectors is
The area to which the pattern belongs becomes very small, making it difficult to obtain high performance in pattern recognition using neural networks.

[0012] Furthermore, the correlation output for each divided light-receiving area between a pattern belonging to a class to be identified and classified and a plurality of reference patterns is obtained multiple times, and a representative value of a set of a plurality of obtained correlation degrees is calculated. A membership function corresponding to the plurality of reference patterns for each class to be recognized and classified, created based on various quantities representing variations, and a correlation output between the test pattern and the plurality of reference patterns. The minimum membership value of the membership values or the average value of the membership values is determined to be the extent to which the test pattern belongs to the class to be recognized and classified. In the pattern recognition method using the characteristic fuzzy AND operation, for the same reason as mentioned above, the overlap of the membership functions of each pattern belonging to the class in each reference pattern is As a result, it has been difficult to obtain high performance in pattern recognition.

[0013]

[Problems to be Solved by the Invention] The present invention has been made to solve the above problems, and uses information reflecting the characteristics of a pattern as one parameter for pattern recognition. It is an object of the present invention to provide a method for detecting correlation output light intensities that can increase the difference in the obtained correlation output light intensities. The present invention also provides a method for detecting correlated output light intensity that can be incorporated into a shift-invariant optical system by taking the sum of the correlated output light intensities obtained for each divided region of each reference pattern. The purpose is to provide. Furthermore, in the present invention, when the obtained correlation output light intensity is input to a pattern recognition means using a neural network or a fuzzy AND operation, the area to which the test pattern belongs in the feature space can be determined. It is an object of the present invention to provide a method for detecting correlated output light intensity that can be used over a wide range of values. Another object of the present invention is to provide a method for detecting correlation output light intensity that has high recognition ability and can recognize many patterns with a small number of reference patterns.

[0014]

[Means for Solving the Problems] The present invention has been made to solve the above-mentioned technical problems, and provides an optical correlation output detection method for detecting a test pattern to be optically recognized and classified. and a plurality of reference patterns, and for the obtained correlation output for each reference pattern, a light receiving element divided into a plurality of regions detects the entire area of the correlation output. Alternatively, optical correlation is characterized in that a portion of the correlation output light intensity is detected and a signal created based on the correlation output light intensity obtained for each of the divided regions is input to a pattern recognition classification means. Provides a method for detecting output.

[0015] Correlation calculations between the test pattern to be optically recognized and classified and each of the plurality of reference patterns are performed using a software.
It is preferable to use a filter based on the Rie transform hologram. In addition, the correlation calculation between the correlation pattern between the test pattern to be optically recognized and classified and each of the plurality of reference patterns is performed using the joint frame of the test pattern and each of the plurality of reference patterns. - The intensity patterns obtained by individually performing the Fourier transform are optically transformed again individually, or the test pattern is combined with multiple reference patterns. It is preferable to optically perform Fourier transformation on the intensity pattern obtained by performing Fourier transformation all at once. Furthermore, the light-receiving element is divided into a plurality of regions, and one of the divided regions is
It is preferable to receive light within a finite range centered on the position where the peak of the autocorrelation output with respect to the reference pattern obtained when each reference pattern is displayed one by one as the test pattern appears. be. In addition, the light receiving element divided into a plurality of regions is configured to pass at least one reference pattern to each test pattern.
It is preferable to have a plurality of regions divided radially around the position where the peak of the autocorrelation output with respect to the reference pattern obtained when displayed as a pattern appears.

[0016] Furthermore, the signal created based on the correlated output light intensity obtained for each divided region is the peak value of the correlated output light intensity within each divided region for each reference pattern. Alternatively, it is preferable to take the total light amount, or the average value. Furthermore, the signal created based on the correlation output light intensity obtained for each divided area is labeled with the possible range of the correlation output light intensity for each divided area for each reference pattern. Preferably, the pattern is divided into several continuous ranges, and the pattern is encoded by labels attached to the ranges to which the correlated output light intensity obtained for each divided region belongs. Preferably, the pattern recognition classification means is a hierarchical neural network having three or more layers, and each connection weight value from the input layer to the output layer is determined by supervised learning. be. In addition, the pattern
The pattern recognition classification means obtains the correlation output for each divided light-receiving area multiple times between a pattern belonging to a class to be identified and classified and a plurality of reference patterns, and calculates a representative value of a set of the obtained correlation degrees. After creating a membership function corresponding to the plurality of reference patterns for each class to be recognized and classified based on various quantities showing the variation, the test pattern and the plurality of reference patterns are Take the membership value between the correlation output and the membership value assigned to each class to be recognized and classified, and calculate the smallest membership value of the membership values or the average value of the membership values. of,
It is preferable that the test pattern belongs to the class to be recognized and classified.

[0017]

[Operation] According to the optical correlation output detection method of the present invention,
LC that displays the test pattern and reference pattern when obtaining one output from one correlation output by detecting correlation output light intensity
Even when the LV contrast is low and the difference in cross-correlation output light intensity for pattern recognition is small, the cross-correlation output and autocorrelation between each test pattern and each of multiple reference patterns is Since the output is divided into a plurality of regions for detection, positional information regarding the presence or absence of feature amounts forming the test pattern can be incorporated. In the feature space when recognizing a pattern, the number of dimensions is generated equal to the number of divisions of the light receiving area, and the difference in the amount of light received within one dimension can be made large. Therefore, in the feature space, it is possible to secure an area to which many test patterns belong even with one reference pattern, so a large number of test patterns can be efficiently processed with a small number of reference patterns. Recognizable.

[0018] Furthermore, after obtaining one output from one cross-correlation output or one auto-correlation output and processing it by computer, it is When inputting to each unit of the input layer, the difference in correlated light output cannot be made very large, so an input vector that passes between the input vectors for each test pattern is obtained, and the neural network In the feature space of the neural network, even if the area to which each test pattern with similar input vectors belongs is very small, the features can be detected by dividing the correlation output and receiving the light. The number of dimensions in the space increases, and positional information related to the presence or absence of features forming the test pattern is included, and the difference in the amount of received light can be large in each dimension. A neural network that can easily secure the area to which the detection pattern belongs and has a high recognition ability regarding pattern recognition using neural networks.
You can get

In addition, in the pattern recognition method using the fuzzy AND operation, when one output is obtained from one correlation output to create a membership function, the class to be recognized and classified in the reference pattern is Even if the degree of overlap between the membership functions of each pattern belonging to a pattern is large, it is possible to perform pattern recognition with high recognition ability for the same reason as explained above regarding the operation of neural networks. can.

Next, the optical correlation output detection method of the present invention will be specifically explained using examples, but the present invention is not limited thereto.

[0021]

[Embodiment 1] A pattern recognition device to which the optical correlation output detection method of the present invention is applied can detect a test pattern and each reference pattern.
an optical correlation system for obtaining a correlation output light intensity with a pattern; and a pattern recognition and classification means for inputting a signal created based on the obtained correlation output light intensity and performing pattern recognition and classification. It consists of In this embodiment, first, an optical correlation system will be focused and explained.

The schematic diagram in FIG. 1 shows an example of an optical correlation system for obtaining optical cross-correlation and optical autocorrelation. In FIG. 1, optical correlations include image output means 1, optical Fourier transform means 2, image output means 3,
This is a joint Fourier transform optical system essentially consisting of an optical Fourier transform means 4 and a photodetector means 5.

Next, the configuration of FIG. 1 will be briefly explained. An image to be recognized and classified drawn on the image display device 16 and a group of reference patterns are read out by the coherent light beam emitted from the laser 11, and a congruent frame is displayed on the screen 31 by the Fourier transform lens 21. Create a transform pattern. The above joint Fourier transform pattern is converted into a two-dimensional photoelectric conversion element 3.
2, and the liquid crystal drive circuit 33 reads the electrically addressed liquid crystal light valve (hereinafter referred to as LCLV) 35.
drawn above, read out with light beam 37, Fourier transform lens 4
2, which has a light-receiving area in which the correlated output light intensity on the screen 42 is divided into a plurality of areas.
The correlated output light intensity detected by the dimensional photoelectric conversion element 50 and detected for each divided light receiving area is transmitted to the computer 51.
entered and recorded within.

Next, a method for detecting the intensity of correlated output light will be explained. FIGS. 3A and 3B show how the divided light-receiving areas detect the correlation output light intensity of cross-correlation.

In FIG. 3(a), the patterns drawn within the dotted lines are the test pattern "T" and the reference pattern "I".
, and the cross-correlation pattern between
The cross-correlation patterns between "L" and "I" are each simply shown, and the blacked out areas are the areas where the correlation output light intensity is strong, and the shaded areas are the areas where the correlation output light intensity is weak. It shows. Furthermore, the two-dimensional photoelectric conversion element 50 detects the correlated output light intensity inside the outermost portion indicated by the dotted line. In Fig. 3, the light receiving area covers the entire area of the correlation output pattern, but it is not necessary to cover the entire area, as long as the pattern can be recognized. , the light receiving area may be a part of the correlation output pattern.

The dotted lines within the light receiving area are the boundaries of the areas into which the light receiving area is divided, D(1), D(2), D(3).
, D(4), D(5) and D(1)', D(2)'
, D(3)', D(4)', and D(5)' indicate area names. In this embodiment, in order to simplify the explanation, the image is divided equally into four, but the number of divisions or the method of division is not limited thereto.

However, as the number of divisions increases, the amount of information to be processed increases when the value corresponding to the obtained correlation output light intensity is input to the recognition and classification means, which increases the burden on the recognition and classification means. , it becomes impossible to recognize the pattern efficiently, so it is necessary to select an appropriate number of divisions. Each divided light receiving area D(1) to D(4), D(1)' to D(4
)' each detect the correlated output light intensity within its region.

At this time, there are two ways to obtain the correlated output light intensity within a region: a method of detecting the average light amount and total light amount within each region, and a method of detecting the peak light amount within each region. There are two types, and when the average light amount and the total light amount are detected, that is, the light amount per pixel is detected, and the light intensity distribution of the correlation output is averaged, and each detected test pattern Although the difference in the correlation output light intensity for each time is reduced, fluctuations in the amount of detected light due to fluctuations in the amount of light, etc. are reduced, and the rate of pattern recognition errors due to fluctuations in the amount of light in the optical system, etc., can be reduced.

Furthermore, when a peak value is detected, fluctuations in the light intensity of the optical system are likely to affect the light intensity distribution, but the light intensity distribution of the correlation output is not averaged, and the Since the difference in the correlated output light intensity is expanded compared to the case where the average light amount is detected, it does not lead to a decrease in recognition ability. However, in the method of detecting the correlation output light intensity of the present invention, both the average value and the peak value are information amounts that reflect the characteristics of the test pattern, and are effective parameters for pattern recognition. becomes.

Now, let us consider the detected correlation output light intensity in FIGS. 3(a) and 3(b). Tested pattern
When the correlation output light intensity between pattern "T" and reference pattern "I" is detected as an average value, DA(1) = DA(2) DA(3) = DA(4)...・(2) becomes. Here, DA(1) to DA(4) are DA(1) to DA(
4) is the average value of the correlation output light intensity between the test pattern "T" and the reference pattern "I".

Equation (2) clearly expresses the left-right symmetry of the line segments that make up T, that is, the positional information of the line segments that make up T. Next, as shown in FIG. 3(b), the test pattern is
If it is "L" instead of "," then D'(1) ~ D
Average values DA' (1), DA' (2), DA' (3), DA of the correlated output light intensities detected in each light receiving area of '(2)'
'(4) all gave different outputs. this is,(
Unlike equation 2), it can be seen as information representing vertical and horizontal asymmetry from the center of the test pattern "L".

From the above, the correlation output pattern for the same reference pattern will be different due to the difference in the characteristics of the test pattern. It can be said that the distribution of output light intensity differs depending on the pattern to be tested. However, when displaying a pattern similar to "T" or "L" as a test pattern, (depending on the degree of similarity) those patterns
It is difficult to separate and recognize the Recognizing and classifying these similar patterns can be solved by adding one reference pattern that can classify similar patterns.

Here, average value detection has been explained using the correlation output light intensity detection method of the present invention, but the same can be said of peak detection. The peak value of each correlation output obtained for each divided light-receiving area reflects positional information that reflects the characteristics of each test pattern. However, since the correlation output obtained by peak detection is easily affected by fluctuations in the optical system, a stable optical system is required. Next, a method of detecting correlation output light intensity using more effective parameters for pattern recognition will be described. The detection method described here is such that when detecting a correlation output using a divided light-receiving element, one of the divided light-receiving areas receives light over a certain finite range centered on the correlation center position. .

FIG. 4 shows how correlated output light intensity is detected by the detection method of the present invention. In FIG. 4, the outer dotted line indicates the area that receives the entire spread of the correlation output pattern, and the inner dotted line indicates the boundary of the divided area, and D(1) to D( 5)
indicates the name of each divided light-receiving area. In particular, D(5
) is an area for detecting the correlation output light intensity near the correlation center position. The peak light amount appearing at the correlation center position is (1)
As can be seen from the equation, it represents the degree of overlap, ie, the degree of similarity, when the centers of the test pattern and reference pattern exactly overlap. When a correlation output is detected with a light receiving element whose light receiving area is divided radially from the correlation center position, such as the above-mentioned 4-split light receiving element, the peak light amount becomes a bias in each region including the peak. Put it away. However, Figure 4
By receiving the correlation output light intensity near the correlation center position, the amount of information representing the similarity between the test pattern and the reference pattern can be obtained. Therefore, compared to the case where the correlation output is detected by a light receiving element whose light receiving area is divided radially from the correlation center position, it is easier to secure the area to which the test pattern belongs in the feature space, and the recognition ability is further improved. .

[0035] Also. In FIG. 4, the number of divided light-receiving areas is D(
The number of divisions is 5 from 1) to D(5), but the number of divisions can be any number as long as there is always a region that can receive the amount of light in the vicinity of the correlation center position of the correlation pattern. I do not care. However, as the number of divisions increases as described above, when the value corresponding to the obtained correlation output light intensity is input to the recognition division means, the amount of information to be processed is large, so the burden on the recognition division means increases. Therefore, it is necessary to select an appropriate number of divisions. The above explained the analog detection method of the correlated output light intensity obtained from each divided light receiving element, but the detection method of the correlated output light intensity is not limited to analog processing method, but also digital Even if it has been processed,
High recognition ability can be obtained regarding pattern recognition. The digital detection method here refers to each reference pattern.
For each divided area, the possible range of the correlated output light intensity is divided into several consecutive ranges and labeled, and the range to which the correlated output light intensity obtained for each divided area belongs is labeled. It is expressed and encoded by attached labels.

Next, a digital method of detecting the correlated output light intensity will be explained. Here, we will explain the processing procedure when performing binary processing to simplify the explanation, but we will explain the possible correlation output light intensity for each divided area for each reference pattern. The number of divisions of the range when dividing the range may be two or more, and there is no problem even if multi-value conversion is performed. As a coding procedure, if the average value or peak value C1 of the correlation output light intensity is obtained for a certain divided light receiving area, for example D(1), the output is First, it is determined to which range the output light intensity belongs. Next, the correlated output light intensity is expressed by a label attached to the range to which it belongs. Further, the above-mentioned processing procedure is similar to other divided light-receiving areas and divided light-receiving areas for other reference patterns.

The binary output obtained in the above manner is expressed as a code. For example, as shown in FIG. 3A, assume that the following results are obtained: DA(1)=DA(2)=R(UP) DA(3)=DA(4)=R(LOW). R(UP) and R(LOW) indicate the respective range names when the possible range of the correlated output light intensity is divided into two. Here, if R(UP)=1 and R(LOW)=0, the coding of the correlation output light intensity of the test pattern "T" with respect to the reference pattern "I" is D(1). When arranged in the order of D(4), it can be expressed as (1100). When similar processing is performed on the correlation output light intensity between the reference pattern "I" and the test pattern "L" as shown in FIG. 3(b), the code becomes (1001 ), compared to (1100), it is possible to provide orthogonality between the test patterns, and it becomes easier to separate the areas to which the patterns belong in the feature space. In addition, when the number of test patterns is large, erroneous recognition of the test patterns can be prevented by increasing the number of divided regions of the light receiving element, dividing appropriate ranges, and increasing the number of reference patterns. By digitally processing the correlated output light intensity as described above, it is possible to perform pattern recognition that is flexible with respect to fluctuations in the light amount of the optical system and association of defective patterns.

Next, the reason will be briefly explained. FIG. 5 is an explanatory diagram showing the relationship between the correlation output light intensity and the output after digital processing. The horizontal axis is the magnitude of the correlation output light intensity, the vertical axis is the output after digital processing, and the thick solid line represents the relationship between the correlation output light intensity and the output after digital processing. Assume that the magnitude of the correlated output light intensity obtained at a certain point in time is at the position ■ in FIG. Assume that the next obtained correlated output light intensity is different from the position (3) due to fluctuations in the optical system or the like. At this time, the output after digital processing is 0 when the correlation output light intensity is between 0 and Th, and is always 0 if the correlation output light intensity is within this range. That is, the output is not easily affected by fluctuations in the amount of light. This also applies when it is desired to obtain an output of "1". However, if the correlated output light intensity is easily obtained around ■, if the value Th representing the boundary of the range is set as shown in Figure 5, a slight fluctuation in the light intensity will result in an incorrect output of "1". I end up giving out. With this, when dividing the possible range of the correlated output light intensity into multiple labeled ranges,
It is necessary to divide the area into appropriate areas where fluctuations in the amount of light will not cause alarm. In addition, when a missing test pattern is displayed, the correlation output becomes partially smaller by the amount of the missing test pattern.
Although it fluctuates, it always outputs the correct output as long as the fluctuation does not cause erroneous output. In other words, it can be said that there is flexibility in the association of missing patterns.

The digital detection method of correlated output light intensity has been described above. It should be noted that the digital detection method provides the same effect as the correlation output light intensity detection method in which the amount of light around the correlation center position is always received, so a description thereof will be omitted. In addition, in the method shown in this embodiment in which the light receiving area is divided radially from the correlation center position and the correlation output is received, or in the method in which one of the culture light receiving areas receives light within a finite range centered on the correlation center position, If the correlation output pattern shifts, the relative position of each divided light receiving area and the correlation output pattern will shift, and the shift invariance property will be lost. By taking the sum of the obtained correlation output light intensities, shift invariant correlation output light intensities can be detected.

[0040]

[Example 2] In the optical correlation calculation in Example 1, joint Fourier transform is performed on the test pattern and multiple reference patterns at once, and the obtained intensity pattern is re-transformed. The idea was to optically perform Fourier transformation all at once. However,
In the optical correlation system described above, interference fringes due to the Fourier transform patterns of the test pattern and a plurality of reference patterns appear superimposed on the screen 31, and the interference fringes are crushed. The information between each pattern was lost to some extent. In Example 2, in order to eliminate the weight of interference fringes, the test pattern and each of a plurality of reference patterns are individually subjected to joint Fourier transform, and the resulting intensity pattern is optically transformed again. An optical correlation system that performs correlation calculations by performing Fourier transforms individually will be described.

FIG. 6 is a block diagram showing an example of the optical correlation system as described above. The test pattern 73 brightly displays a part of the periphery of the test pattern, as shown in the explanatory diagram of a test pattern display example showing the display state of the test pattern in FIG. The image is multiplexed and formed on the writing surface of the spatial light modulator 72 by the lens array 66. At this time, the bright part of the test pattern and the surrounding brightly displayed parts are
The image is formed as a bright image. Therefore, the reference pattern in FIG.
As shown in the figure, the reference pattern has a transmittance distribution according to the reference pattern, and the area corresponding to the area where the image of the test pattern around it is formed. A reference pattern group mask 74 with a high transmittance distribution is connected to a spatial light modulator 72.
If it is placed near the writing surface of the
The paired patterns of the pattern and each reference pattern are transmitted to the spatial light modulator 7.
It is possible to write to 2.

The pair of the test pattern and each reference pattern written in the spatial light modulator 72 is read out by the light beam 61a, which is the readout light, and each pair of patterns is connected to the lens array. 67, and each Fourier transform pattern is written to the spatial light modulator 71.
It is read out by the light beam 61a. Each read Fourier transform pattern is Fourier transformed again by the lens array 68 to obtain correlated output light intensities for each test pattern and each reference pattern on the light receiving surface 65. Can be done. The correlation output light intensity obtained as described above is detected using each of the correlation output light intensity detection means shown in Example 1, and then used as a pattern for final pattern recognition and division. is input to the on-line recognition and division means.

As described above, according to the optical correlation system shown in the second embodiment, the test pattern and the plurality of reference patterns are displayed on the screen 31.
The interference fringes due to the Fourier transform pattern of the image are no longer superimposed, and deterioration of the contrast of the interference fringes can be avoided. As a result, it is possible to make a large difference in the correlation output light intensity, so it is possible to secure a wide area to which the test pattern belongs in the feature space, and it is possible to improve the pattern recognition ability. .

[0044]

Embodiment 3 Next, FIG. 9 shows the configuration of an example of means for obtaining optical correlation output light intensity according to the present invention. That is, an optical correlation system using a matched filter will be explained. FIG. 9 shows an optical system for reproducing a known matched filter. The matched filter 75 uses a pattern obtained by optically Fourier-transforming a plurality of reference patterns in advance and multiplexing them while changing the irradiation direction of a plane wave for each reference pattern.

As shown in FIG. 9, the beam 12 emitted from the laser 11 is expanded into a suitable beam system by the beam expander 13, and the pattern to be recognized and classified drawn on the image display device 16 is displayed. - irradiate with light. Next, the light beam having the complex amplitude distribution of the pattern is subjected to Fourier transformation by the Fourier transformation lens 21, and irradiates the multiple interference fringe pattern drawn on the matched filter 75. At this time, diffracted light is irradiated from a reference pattern having the same spatial frequency spectrum as the pattern to be recognized and classified in the direction of the plane wave used to create the reference pattern.

When this diffracted light is focused onto the screen 31 by the condensing lens 22, one of the diffracted lights is focused on the reference pattern.
This is the cross-correlation between the pattern and the pattern to be recognized and classified, and the other is the convolution. Therefore, each reference pattern
Since the position at which the cross-correlation outputs from the cross-correlation outputs appear is known in advance, if the correlation output light intensities thereof are detected by the two-dimensional photoelectric conversion element 50 using each of the correlation output light intensity detection means shown in Example 1, good. The output obtained as described above is then input to a pattern recognition and classification means where final pattern recognition and classification is performed.

As described above, according to the optical correlation system shown in Example 3, unlike the optical correlation system shown in Examples 1 and 2,
Adding a reference pattern is not easy since the photographic plate or the like must be exposed again to create a matched filter. However, since the optical correlation system shown in this embodiment has a very simple configuration, it can be easily implemented as an apparatus.

[0048]

[Embodiment 4] In Embodiment 4, each correlation output light intensity signal obtained after the detection according to the present invention is inputted to the correlation output light intensity, and a pattern recognition classification is performed. As an example of the pattern recognition and classification means, a pattern recognition and classification means using a neural network will be explained. In this example, the optical correlation system for obtaining the correlation output is not limited to the correlation systems shown in Examples 1 to 3.

FIG. 10 shows the process from the optical correlation system that optically performs correlation calculation between the test pattern and each reference pattern to the pattern recognition and classification means that performs recognition and classification of the test pattern. It is a block diagram which shows a series of flows. In FIG. 10, in the optical correlation system, correlation calculation between the test pattern and each reference pattern is optically performed, and the correlation output processing of the present invention is performed on the obtained correlation output light intensity. . Then, each output of the correlated output light intensity after being processed is inputted to each unit of the input layer of the neural network.

In the neural network in this example, the number of units included in the input layer is 5, the number of units included in the intermediate layer is 3, and the number of units included in the output layer is 5. , only one example is shown for ease of explanation, and the number of units in the intermediate layer and each layer can take various values depending on requirements. Further, as a learning method, back propagation (BP method) can be used.

[0051] Each output of each correlation output light intensity processed by the processing method of the present invention is outputted as is from the input layer, inputted to the intermediate layer and then to the output layer, and outputted from the output layer. Ru. The input value F(x) to each unit of the intermediate layer and the input value F(x) to each unit of the output value are obtained by multiplying the output value from the previous layer by the connection load value corresponding to the connection. If x is the sum of the values plus the bias value, it can be expressed by the following equation. F(x)=1/(1+exp(-x)) Also, the input layer
The coupling weight values and bias values between the intermediate layers and between the intermediate layer and the output layer are independent values.

Neural network configured as described above
In the diagram, given a certain input pattern, that is, the output vector (I1, I2, I3, I4, I5) of the input layer unit, the error of each unit from the desired output from the output layer is 2.
By modifying the connection weight values between each unit so that the sum of the products becomes small, a neural network that outputs a desired output pattern for a certain input pattern can be obtained. The BP method is a method for calculating a correction coefficient for the load at this time.

As shown in the specification of the previous application by the present inventors, in an optical correlation system, the average value or peak value within a finite range centered on the correlation center position of the obtained correlation output 1 from one correlation output pattern.
I only received one piece of information. Furthermore, the above detection method loses positional information that reflects the characteristics of the pattern. In particular, in average value detection, although a stable output can be obtained, the distribution of light intensity in the correlation output pattern representing the characteristics of the pattern is averaged, and it is difficult to obtain correlation output light intensities with large differences. Met.

For this reason, it is important to carefully select the reference pattern and light receiving area so that the input vectors are not similar. However, according to the detection method of the present invention, by dividing the receiving range of the correlation output light intensity between each reference pattern and the test pattern, it is possible to obtain information that reflects the characteristics of the test pattern. can. Furthermore, as can be seen from equation (1), the above information is related to the presence or absence of line segments, etc. that make up the pattern, so there is a large difference in the correlated output light intensity for each divided light-receiving area, and the Since the area to which the test pattern belongs can be widened,
Higher performance can be obtained in pattern recognition.

[0055]

[Embodiment 5] In Embodiment 5, after the detection method of the present invention is applied to the correlation output light intensity, signals of each correlation output light intensity obtained are input, and pattern recognition and identification are performed. As an example of pattern recognition and division means, a pattern recognition and division means using fuzzy AND operation will be described. The pattern recognition and classification means of this embodiment obtains the correlation output between the test pattern and a plurality of reference patterns multiple times for each divided light-receiving area, and calculates the average value of the obtained correlation degree and the standard. Create a membership function corresponding to the plurality of reference patterns for each test pattern to be recognized and divided based on the deviation,
A membership value between the correlation output obtained for each divided light-receiving area between the test pattern and the plurality of reference patterns and the membership assigned to each test pattern is calculated. The minimum membership value of the membership values is determined to be the extent of the pattern to be tested.

For example, "H", "E" are used as test patterns.
”, and there are “I” and “V” as reference patterns.Assume that the correlation output light intensity at this time is detected so as to divide the correlation pattern into left and right halves from the center as shown in FIG. In Fig. 11, the light receiving area on the left side is designated as D(1), and the light receiving area on the right side is designated as D(2).
The numbers are based on "E" and "I". FIG. 12 shows membership functions for each test pattern created by the detection method of the present invention.

In FIG. 12, the horizontal axis represents the degree of correlation, and the vertical axis represents the membership value. I(1) and I(2) are
Membership in D(1), D(2) for reference pattern "I", V(1), V(2) in D(1), D(2) for reference pattern "V" Represents membership. In addition, the trapezoidal shape of the membership is defined by the width of the standard deviation δ on the left and right sides of the average value obtained when correlation outputs for each divided light-receiving area are taken multiple times. The range from 3.delta. to 3.delta. is created so that the membership value decreases linearly from "1" to "0". Creation of membership in this example is as follows:
This is just one example and is not limited to the method described above.

The membership functions for the test pattern "H" are I(1) and I(2), V(1) and V(2).
The reason why they are the same is that the left and right symmetry of the test pattern "H" is manifested. (Actually, the membership will not be the same due to fluctuations in the amount of light, etc.) Also, in the membership function for the test pattern "E", I(1), I(2), V(1) The reason why and V(2) are different is because the left-right asymmetry of the test pattern "E" appears. As described above, membership functions for the test patterns "H" and "E" were obtained. Next, refer to the reference pattern "I" again,
Assume that the correlation output light intensity of the test pattern "H" is detected with respect to "V".

FIG. 13 shows a comparison between the membership function and the correlation output light intensity of the test pattern "H" with respect to the re-detected reference patterns "I" and "V". In FIG. 13, I'(1) and I'(2) are the light receiving area D(
1), D(2), and V'(
1), V'(2) is the reference pattern of the test pattern "H"
The graph shows the correlated output light intensities in the light receiving areas D(1) and D(2) with respect to the ring "V". Members of test pattern “H”
In the ship function, I'(1), I'(2), V'(
1), each membership value for V'(2) is "
1", and if the minimum membership value is taken, the "H"-likeness of the test pattern "H" becomes "1".On the other hand,
In the membership function of test pattern "E",
The respective membership values for I'(1), I'(2), V'(1), and V'(2) are "0", "0", and "0".
．． 8" and "0", and if the minimum membership value is taken, the test pattern "H" has a "E"-likeness of "0". Therefore, from the above results, the test pattern "H" is
It is recognized as "H". The fuzzy image shown in this example
In the pattern recognition method using the logical AND operation, as described in the specification of the previous application by the present inventor, since the difference in the correlation output light intensity obtained in the correlation output light intensity detection method is small, each The degree of overlap between the membership functions of each pattern belonging to the class to be recognized and divided in the reference pattern is large, and it is necessary to artificially correct the membership functions.

However, the difference in the correlated output light intensity obtained by the method of detecting the correlated output light intensity of the present invention increases,
The area to which the test pattern belongs can be widened in the feature space, making it possible to recognize patterns with high recognition ability. In this embodiment, the number of divisions of the light-receiving area has been explained as 2, but this is for the sake of simplifying the explanation, and as long as the processing load on the recognition division means does not increase, it may be divided into 2 or more. good.

[0061]

Effects of the Invention As explained above, the above-mentioned effects were obtained by the optical correlation output detection method of the present invention. Putting these together, we get the following remarkable technical effects. That is, first, it was possible to provide an optical correlation output detection method that can obtain information that reflects the characteristics of a pattern and can make a large difference in the obtained correlation output light intensities.

Second, by calculating the sum of the correlated output light intensities obtained for each divided area for each reference pattern,
We have been able to provide a method for detecting optical correlation output that allows a shift-invariant optical system to be assembled. Third, by performing digital processing on the correlated output light intensity obtained through detection, it is possible to convert the fluctuations of the optical system into a series, and it is also possible to further increase the difference in the correlated output light intensity. A method for detecting optical correlation output could be provided. Fourth, the correlated output light intensity processed by the present invention is applied to a neural network or fuzzy A
When inputting to a pattern recognition classification means using ND operation, the area to which the test pattern belongs in the feature space can be enlarged, and high recognition ability can be obtained in pattern recognition. We were able to provide a method for detecting optical correlation output.

[Brief explanation of drawings]

FIG. 1 is a block diagram of an optical correlation system using a joint Fourier transform method to obtain a correlation output.

FIG. 2 is an explanatory diagram showing how to take a light receiving area when detecting a conventional correlation output.

FIG. 3: 1 using the optical correlation output detection method of the present invention
FIG. 2 is a configuration explanatory diagram showing a state of detection using an example of a four-division light receiving element.

FIG. 4 is an explanatory diagram showing how the peak value of the autocorrelation pattern of pattern "E" is detected using a 5-divided light receiving element according to the optical correlation output detection method of the present invention. be.

FIG. 5: In the optical correlation output detection method of the present invention,
FIG. 3 is an explanatory diagram showing an output after digital processing with respect to correlation degree.

FIG. 6: In the optical correlation output detection method of the present invention,
FIG. 2 is a configuration diagram of an optical correlation system that individually performs correlation calculations between a test pattern and a reference pattern in order to obtain a correlation output.

7 is an explanatory diagram showing an example of display of a test pattern in the optical correlation system shown in FIG. 6; FIG.

[Fig. 8] In the optical correlation system shown in Fig. 6, reference pattern
FIG.

FIG. 9: In the optical correlation output detection method of the present invention,
FIG. 2 is a configuration diagram of an optical correlation system using a matched filtering method to obtain a correlation output.

[Fig. 10] A series of steps from the optical correlation system to the pattern recognition and classification means when a neural network is used as the pattern recognition and classification means in the optical correlation output detection method of the present invention. FIG. 2 is a configuration explanatory diagram showing the configuration of a neural network.

FIG. 11 is an explanatory diagram showing how a correlation output is detected when a light receiving area is divided into two in the optical correlation output detection method of the present invention.

FIG. 12: In the optical correlation output detection method of the present invention,
Reference patterns “I” and “E” for test patterns “H” and “E”
FIG. 3 is an explanatory diagram showing each membership function for V''.

FIG. 13 is an explanatory diagram showing how the obtained correlation output and membership function are compared in the optical correlation output detection 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,
61a, 61b luminous flux 13
Beam expander 14, 17, 18
Beam splitter 16
Image display device 21
, 41
Fourier transform lens 22
Condensing lens 31
Screen 32, 50
Two-dimensional photoelectric conversion element 33
Image processing and liquid crystal drive circuit 34
mirror 35
LCD light bulb 51

computer 65
Light receiving surface 67, 68
Lens array 71
, 72
Spatial light modulator 73
Test pattern 74
Reference pattern group mask

Claims (9)

[Claims] Claim 1: A method for detecting optical correlation output, comprising:
Correlation calculation is performed between the test pattern to be optically recognized and classified and each of the multiple reference patterns, and the received light divided into multiple regions is calculated based on the correlation output for each reference pattern obtained. A pattern recognition and classification means detects the correlation output light intensity of the entire area or a part of the area of the correlation output using the element, and generates a signal based on the correlation output light intensity obtained for each divided area. A method for detecting an optical correlation output, the method comprising inputting an optical correlation output. 2. A correlation calculation between the test pattern to be optically recognized and classified and each of the plurality of reference patterns is performed using a filter based on a Fourier transform hologram. 2. The method for detecting optical correlation output according to item 1. 3. The correlation calculation between the correlation pattern between the test pattern to be optically recognized and classified and each of the plurality of reference patterns is performed based on the correlation pattern between the test pattern and each of the plurality of reference patterns. joint group with
Intensity pattern obtained by performing Rie transform individually
Either optically perform the Fourier transformation of the images individually, or
Or, a joint frame of the test pattern and multiple reference patterns.
2. The method for detecting an optical correlation output according to claim 1, wherein the intensity pattern obtained by performing Fourier transform all at once is subjected to optical Fourier transform again. 4. The light receiving element divided into the plurality of regions:
One of the divided regions is centered at the position where the peak of the autocorrelation output for the reference pattern obtained when each reference pattern is displayed one by one as the test pattern. 4. The optical correlation output detection method according to claim 1, 2, or 3, wherein light is received in a finite range of . 5. The light receiving element divided into the plurality of regions includes:
At least a plurality of regions are divided radially around the position where the peak of the autocorrelation output with respect to the reference pattern, which is obtained when each reference pattern is displayed one by one as the test pattern, appears. The optical correlation output detection method according to any one of claims 1 to 3, characterized in that the method comprises: 6. A signal created based on the correlated output light intensity obtained for each divided region is a peak of the correlated output light intensity within each divided region for each reference pattern. Claims 1 to 3, characterized in that the value, the total amount of light, or the average value is taken.
4. The method for detecting optical correlation output according to any one of 4. 7. The signal created based on the correlated output light intensity obtained for each divided region is a signal that is generated based on the correlated output light intensity obtained for each divided region for each reference pattern. The range is divided into several consecutive labeled ranges, and a pattern is encoded by the label attached to the range to which the correlated output light intensity obtained for each divided region belongs. The optical correlation output detection method according to any one of claims 1 to 5. 8. The pattern recognition classification means is a hierarchical neural network having three or more layers, and determines each connection weight value from the input layer to the output layer by supervised learning. The optical correlation output detection method according to any one of claims 1 to 7. 9. The pattern recognition and classification means obtains correlation outputs for each divided light-receiving area between a pattern belonging to a class to be classified and a plurality of reference patterns a plurality of times, and After creating a membership function corresponding to the plurality of reference patterns for each class to be recognized and classified based on the representative value of the set of correlation degrees and various quantities indicating the dispersion,
The correlation output between the test pattern and the plurality of reference patterns and the membership value assigned to each class to be recognized and classified are taken, and the member with the smallest membership value is selected. - The optical system according to any one of claims 1 to 7, characterized in that the ship value or the average value of the membership values is set to the extent that the test pattern belongs to the class to be recognized and classified. How to detect correlation output.