JP3116055B2 - Optical pattern identification device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION The present invention relates to a neural network.
More particularly, the present invention relates to a pattern identification device using a neural network to which a preprocessing device for input information is added.

[0002]

2. Description of the Related Art In recent years, as a method of identifying a pattern of an input image (that is, a test image), a method using a neural network has been often proposed. According to the method using the neural network, even if there is an ambiguity such as a slight lack or deformation of the pattern of the input image, it is possible to associate a storage image (reference image) corresponding to the ambiguity and perform character recognition. It is considered to be effective for voice recognition and other pattern recognition. However, in this neural network, an enormous number of neurons are required depending on the number of pixels of the storage pattern and the number of patterns.
In general, wiring is required between the squares of the number of neurons, so that it is difficult to store a large amount of information or a large number of images at a time.

Further, if the size of the pattern to be inspected changes, or if the rotation or position in the screen changes, correct identification cannot be obtained, so that the rotated pattern or the pattern having a different size is not obtained. Therefore, the number of patterns that can be stored has to be reduced. Storing a large number of images in this way leads to an increase in the number of neurons, and as a result, causes problems such as an increase in apparatus scale, an increase in processing time, and an increase in apparatus cost.

Further, if the pattern of the input information is directly used as the input of the neural network, even the extra information which does not represent the characteristics of the input information is input to the neural network. The number of neurons and the number of connections in the neural network are increased more than necessary, and the characteristics are unclear.
In the worst case, the learning is not performed effectively.

[Problems to be solved by the invention]

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problem, and has an object to provide a pattern identification device capable of realizing an input image identification by a neural network with a high-speed and inexpensive device. The purpose is to provide. It is another object of the present invention to provide an optical pattern discriminating apparatus that can discriminate even when a rotation or a shift of an input image exists, and that can store a large amount of reference images. A further object of the present invention is to provide a pattern identification device that extracts a feature of input information, thereby shortening a learning time in a neural network and having a high possibility of effective learning. I do.

[0006]

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned technical problems, and at least a plurality of inputs can be received and an input value based on the plurality of inputs is processed. Accordingly, a plurality of operation units (hereinafter, referred to as neurons) for outputting one output value, a coupling means for coupling an output of each of the neurons, and an input of another neuron or a neuron which has performed the output, are provided. Each of the combining means has a pattern identification device for performing pattern identification using a neural network having a coupling coefficient for converting an output value related to the input value into an input value. An autocorrelation pattern detection device for obtaining a pattern or for obtaining a square pattern of the autocorrelation of the input pattern, One, the output based on the autocorrelation pattern obtained as input of the neural network, further
In the case the autocorrelation pattern is a two-dimensional pattern, a second fan-shaped with a substantially 180 ° to more central angle around the peak position of the autocorrelation pattern
In the three-dimensional space, the spatial small area group A is divided into fan-shaped small areas having the same central angle, and in the two-dimensional fan-shaped space, these are divided by concentric arcs centered on the peak position. , Assuming a small region group B divided into small regions, and integrating the value of the pattern of the autocorrelation of the input pattern or the value of the pattern of the square of the autocorrelation within the small region from each of the small regions. as the output of, among these, the integral value by the small region groups a, or the integral value by the small region group B, or the optical pattern recognition apparatus characterized by the both the input of the neural network so
is there.

[0007] The autocorrelation pattern detecting device for obtaining a square pattern of the autocorrelation of an input pattern comprises at least a coherent light source and a first spatial light modulator or spatial light for coherently displaying an input pattern to be identified. A modulator, a first Fourier transform lens for optically Fourier transforming the first spatial light modulator or the coherent pattern displayed on the screen of the spatial light modulator, and a Fourier transform obtained by this. A second spatial light modulator or a spatial light modulator that modulates the intensity distribution or the complex amplitude distribution of the output coherent light according to the intensity distribution based on the complex amplitude distribution, and the second spatial light modulator or the spatial light modulator It is preferable to have a second Fourier transform lens for optically Fourier transforming again the output coherent light from.

In order to output the respective integrated values, a light receiving element having a shape corresponding to each small area is arranged on a Fourier transform plane by the second Fourier transform lens .
Good. The first and / or second spatial light modulator or a spatial light modulator device, the two-dimensional image sensor as an input unit, receives the output video signal of the image sensor, the electrically addressed spatial light modulator Then, it is preferable to display a pattern formed based on the intensity distribution of incident light incident on the two-dimensional image sensor. The input putter
Autocorrelation path for obtaining the square of the pattern of emission of the autocorrelation
Turn detecting apparatus, the first spatial light modulator or spatial light modulator, the input pattern to be identifies, displays double in relation translation, correlation between the two patterns the display Is the square pattern of the autocorrelation
It is suitable. When the integrated value of the small area group A is output, the central angle is set to approximately 180 degrees, and the arrangement pattern of the output signal from each small area is changed to the small area having the largest output or the smallest area. The position in the output signal pattern of the output signal from the region is always at a predetermined position, and the order of output from the small regions before and after the position is shifted so as to be the same as the spatial arrangement of the small region group. comprising a device for relocation of a signal sequence to be, also the rearranged signal sequence pattern is input to the neural network
It is preferred that. In addition, it is preferable that the output from the autocorrelation pattern detection device is binarized using an appropriate threshold level and then input to the neural network.

[0009]

According to the configuration of the present invention, the autocorrelation function of the input pattern is such that the direction and length of the line segment contained in the pattern are integrated around the origin. Further, when there is a periodic structure in the input pattern, information on the period and direction is also included. Therefore, taking the autocorrelation of the input pattern can be regarded as extraction of one effective characteristic, and by using this information as the input of the neural network, sufficient identification can be performed with a small amount of input information. Can be performed. Further, in the case of the configuration having the small area group A, the output from each small area does not change even if the size of the input pattern changes. Invariant input information can be provided to the neural network.

Further, in the configuration having the small area group B, even if the input pattern rotates in the input plane, the output from each small area does not change. For in-plane rotation, invariant input information can be provided to the neural network. Further, when the above-mentioned configuration is provided with the small area group A and the signal sequence rearrangement device, the maximum output or minimum output area is obtained even if the input pattern rotates in the input plane. Are constant in the in-plane rotation of the input pattern, and are kept unchanged even if the size of the input pattern changes. An input signal can be input to a neural network.

As means for obtaining the square pattern of the autocorrelation of the input pattern to be identified, a coherent light source,
When using optical means using the first and second spatial light modulators or spatial light modulators and the first and second Fourier transform lenses, first the first spatial light modulator or spatial light modulator Input pattern displayed coherently on the spatial light modulator
Fourier transform of the complex amplitude distribution based on the light is obtained on the light receiving surface of the second spatial light modulator or the spatial light modulator by the first Fourier transform lens, and the light on the light receiving surface is obtained. Depending on the intensity distribution, the complex amplitude distribution of the output light of the second spatial light modulator or the spatial light modulator has a squared pattern of the Fourier transform of the input pattern. This second
Output light from the spatial light modulator or spatial light modulator of
A complex amplitude distribution corresponding to the Fourier transform of the complex amplitude distribution of the output light is formed on the light receiving element placed on the Fourier transform surface through the second Fourier transform lens. This complex amplitude distribution corresponds to the autocorrelation pattern of the input pattern, but what is actually observed by the light receiving element is
Because of the light intensity distribution, a square pattern of the autocorrelation of the input pattern is obtained as the output of the light receiving element.

However, at this time, if the contrast of the light quantity distribution of the output light from the used spatial light modulator or spatial light modulator is low, DC noise may be added to the autocorrelation pattern. In such a case, when two input patterns are displayed side by side on the first spatial light modulator or the spatial light modulator so as to move in parallel with each other, the second pattern is obtained. On the spatial light modulator, in addition to the squared pattern of the autocorrelation of the entire input pattern, two display patterns
The square pattern of the correlation between the patterns is obtained. The squared pattern of the correlation between these two display patterns is the same as the squared pattern of the autocorrelation because the two patterns are the same pattern. It does not contain. If the output from the autocorrelation pattern detection device is binarized and then input to the neural network, the learning will be focused more quickly.
Faster learning can be performed.

Next, the optical pattern discriminating apparatus of the present invention will be specifically described with reference to examples, but the present invention is not limited thereto.

[0014]

[Embodiment 1] FIG. 1 is a schematic configuration diagram showing a specific configuration of an example of a pattern identification device of the present invention.

That is, an input pattern such as a character is divided into, for example, small pixels and input to the autocorrelation pattern detection device 1 as pixel data. Here, means for inputting the input pattern to the autocorrelation pattern detection device 1 is not shown, but it is known that the input is performed using, for example, an imaging device using a CCD or an imaging tube. Means. When the autocorrelation pattern detection device 1 uses optical means, the input means forms an input pattern on the writing surface of the optically addressed spatial light modulator. Can also be realized.

In the autocorrelation pattern detection device 1, the calculation of the autocorrelation of the input pattern and the input of the data of the calculation result are performed.
Conversion to the ral network input data format is performed. Here, the calculation of the autocorrelation of the input pattern in the autocorrelation pattern detecting device 1 can be easily realized by an arithmetic circuit or a computer program. In particular, when the number of pixels in the input pattern is large and a large amount of time is required for the integration operation, the operation result of the Fourier transform of the square of the Fourier transform becomes equivalent to the autocorrelation. When the autocorrelation calculation is performed by
A higher-speed autocorrelation operation can be performed by using the Rier transform (FET) algorithm.

The autocorrelation calculation can be performed optically. Speaking of conventional optical correlation calculation,
Matched filtering and joint transform methods are known, but previous studies focused solely on detecting peaks in the correlation output. However, by taking the output pattern linearly,
Alternatively, an autocorrelation pattern can be obtained by taking the output pattern not only at the peak but also over a finite area. However, when observing the optically obtained pattern as an intensity pattern as usual, the pattern obtained here is exactly the square pattern of the autocorrelation. is there.

The result of the autocorrelation operation is obtained by discarding the position information of the line segment constituting the input pattern and extracting only the information on its length and direction, and extracting the information on the periodic structure. Compared to the case where the input pattern is directly represented by pixel information by raster scanning, a large amount of information can be compressed. In particular, since the result of the autocorrelation calculation results in rotationally symmetric data, it is particularly effective to represent the calculation result data in an r-θ coordinate system (that is, a polar coordinate system).

By using the data whose information amount has been greatly reduced in this manner as an input to the neural network 8, an effective pattern identification can be performed even with a small neural network. Can do it. Also,
Since unnecessary information (that is, information that does not represent a feature) is reduced, rapid convergence of learning is obtained.

Although the signal sequence rearrangement device 7 shown by a broken line in FIG. 1 is not essential to the present invention, by providing it, the storage pattern in the neural network can be further improved. Can be more. The signal sequence rearrangement device 7 will be described with reference to the drawings in the description of another embodiment described later. Here, the neural network 8
Are realized by using an electric processing circuit or by using a program on a general-purpose processing circuit. However, these circuits tend to be large in size due to the nature of a neural network. Hampering the construction of a full-fledged neural network. Since the construction of the circuit scale increases in proportion to the square of the input information amount, in the present invention, the reduction of the input information amount has a very great effect. Incidentally, a method of optically realizing this neural network is also being studied at present. In the optical neural network as well, the reduction in the amount of input information can reduce the load on the spatial light modulator, and
This has great effects such as speeding up learning.

[0021]

Embodiment 2 When the pattern of the autocorrelation is a two-dimensional pattern, as described above, the data of the autocorrelation calculation result is stored in the rotating coordinate system (ie, the r-θ coordinate system). In a two-dimensional fan-shaped space having a center angle of about 180 degrees or more around the peak position of the autocorrelation pattern, the fan-shaped two-dimensional space having the same center angle is used. A spatial small area group A divided into small areas and a small area group B divided into small areas by concentric arcs centered on the peak position in the fan-shaped two-dimensional space are assumed. Of the autocorrelation value or the squared value of the autocorrelation in the small area of each of the small areas as an output from each of the small areas. B, or both, are input to the neural network. Are each at r-theta coordinate system, theta corresponding to theta direction integral output during a certain time in the r direction integration output and r constant, more effectively de - can take advantage of data

FIG. 2 shows that when the coordinate system is taken as described above,
FIG. 4 is an explanatory diagram showing a detection area of an autocorrelation pattern. That is, the case where each of the small area groups A and B has an area of 180 degrees is shown. Here, the small area group A6a
Is composed of fan-shaped small regions 6a1, 6a2,... 6an, and the central angles of the fan-shaped regions are equal in all the small regions. The small region group B6b includes concentric band-shaped small regions 6b1, 6b2,..., 6bn. here,
Since the correlation output may be particularly large near the center,
A dead zone (that is, the correlation output of that portion is not added to the output of each region) may be used.

Considering such a region, since the autocorrelation pattern is a point-symmetric pattern, all the autocorrelation information is contained in each of two small region groups each extending 180 degrees. Will be. In the small area group A6a, if the radius of the fan shape is sufficiently large, the input pattern to be identified is determined.
Even if the size of the area changes, the output from each small area 6ai does not change, and the size-invariant identification can be performed. Further, in the small area group B, even if the input pattern to be identified is rotated in the input plane, the output of each small area 6bi does not change, so that rotation invariant identification can be performed.

[0024]

Third Embodiment In the latter half of the first embodiment, it has been described that the autocorrelation operation of the present invention can also be performed optically. This optical means includes at least a coherent light source,
First to coherently display the input pattern to be identified
Of a spatial light modulator or a spatial light modulator and a coherent display pattern displayed on the first spatial light modulator or the spatial light modulator. A second spatial light modulator that modulates the intensity distribution or the complex amplitude distribution of the output coherent light in accordance with the intensity distribution based on the complex amplitude distribution resulting from the Fourier transform by the Fourier transform lens or The spatial light modulator includes a second spatial light modulator or a second Fourier transform lens for optically Fourier transforming the output coherent light from the spatial light modulator again. Then, the optical autocorrelation operation can be performed in quasi-real time (however, as described above, to be precise, the observed light amount corresponds to the square of the autocorrelation). Also, here, as described in the second embodiment, in order to output the integrated value in each small area, it is preferable to arrange a light receiving element having a shape corresponding to each small area.

FIG. 3 shows an optical auto-correlation pattern provided with the optical auto-correlation calculating means and the light receiving element having the above construction.
FIG. 2 is a configuration diagram illustrating an example of an application detection unit. For example, He
A beam 12 emitted from a coherent light source 11, such as a Ne laser, has its beam diameter expanded to an appropriate size by a beam expander 13, and a beam 12a and a beam 12b by a beam splitter 14. Divided into Luminous flux 1
2a is further reflected by the beam splitter 18;
The light enters the readout surface of the spatial light modulator 17.

The spatial light modulator 17 changes the intensity reflectance distribution of the reading surface or the refractive index distribution for the reading light (these are collectively referred to as a complex amplitude reflectance distribution) in accordance with the intensity distribution of the input light. It is. Therefore, the light flux 12a reflected by the spatial light modulator 17 (that is, the light flux output from the spatial light modulator 17) has an intensity distribution and a phase distribution corresponding to the input light intensity distribution to the spatial light modulator 17 (these are represented by (Collectively referred to as a complex amplitude distribution).
Therefore, as an input to the spatial light modulator 17, the input pattern 9 is illuminated, and an image is formed on the input surface (writing surface) of the spatial light modulator 17 by an imaging lens. A light beam 12a having a complex amplitude distribution corresponding to the negative direction is obtained on an output surface (that is, a reading surface) of the spatial light modulator 17.

The light flux 12a further passes through a beam splitter 18, passes through a Fourier transform lens 19, and is incident on a writing surface of a spatial light modulator 20. The spatial light modulator 20 has the same configuration as the spatial light modulator 17 described above. At this time, the reading surface of the spatial light modulator 17 and the writing surface of the spatial light modulator 20 are each a free surface.
The focal plane is set to be the front focal plane and the rear focal plane of the Lie transform lens 19. Then, a light intensity distribution corresponding to the square of the Fourier transform of the complex amplitude distribution of the light flux 12a on the readout surface of the spatial light modulator 17 is obtained. Here, the light beam 12b reflected by the total reflection mirror 15 and further reflected by the beam splitter 21 enters the read surface of the spatial light modulator 20, and the light beam on the write surface With the complex amplitude distribution corresponding to the intensity distribution, the light is emitted from the spatial light modulator 20, transmitted through the beam splitter 21, passed through the Fourier transform lens 22, and passed through the wedge ring detector 2.
3 is incident. At this time, the reading surface of the spatial light modulator 20 and the light receiving surface of the wedge ring detector 23 are located on the front focal plane and the rear focal plane of the Fourier transform lens 22, respectively. A light amount distribution corresponding to the square of the Fourier transform of the complex amplitude distribution of the light beam 12b on the readout surface of the spatial light modulator 20 is obtained on the light receiving surface of the spatial light modulator 20.

Here, the wedge ring detector 23
Is a photoelectric conversion element having a light receiving area having the same shape as the autocorrelation pattern detection area shown in FIG. 2, and its center is set on the optical axis of the Fourier transform optical system in FIG. . Thus, on the wedge ring detector 23,
Light intensity pattern corresponding to the square of the autocorrelation of the input pattern
Is obtained. Therefore, by making the output the input of the neural network 8, the input pattern to be identified can be obtained.
Can be identified.

The light receiving element used here may be a two-dimensional image sensor having a rectangular light receiving area instead of such a wedge ring detector. However, as described in the second embodiment, a wedge ring detector is used. However, the characteristics of the autocorrelation pattern can be effectively extracted. Further, it is easy to recognize that the size or the rotation is unchanged.

The spatial light modulators 17, 20 used here
For example, a liquid crystal panel with a photoconductive layer provided on a nematic liquid crystal panel is called Liquid Crystal Valve (LCLV).
It is commercially available from Hughes, USA. Also, devices using ferroelectric liquid crystals have been developed as having higher speed operation and higher resolution. Further, Bi ₁₂ Si
A transmissive element using O ₂₀ (BSO) is also available. Correctly, in the case of a transmissive element, a change in the optical system is necessary because the direction in which the readout light beam enters is changed, but it is essentially the same as in the case of using a reflective spatial light modulator. is there.

The optical system used in the present embodiment is more suitable for the pattern discriminating apparatus of the present invention in the following points as compared with other optical correlation calculating apparatuses. First, a spatial light modulator having a relatively low resolution can be used. This is because a spatial light modulator operating in real time has a relatively low resolution.
This is an important feature in realizing the real-time operation of the identification device. Second, the optical axis alignment is easy, and
It can be easily assembled and adjusted, and can be used in a bad environment.

[0032]

Embodiment 4 Spatial light modulators 17 and 20 used in Embodiment 3
Can be replaced by the spatial light modulator shown in the fourth embodiment. In other words, this spatial light modulator uses a two-dimensional image sensor as an input unit, and in response to this signal, the electric address type spatial light modulator uses the two-dimensional image sensor to obtain a pattern based on the incident light intensity distribution incident on the two-dimensional image sensor. -Is displayed.

FIG. 4 is a block diagram showing an example of such a spatial light modulator. The input light beam (that is, the writing light beam) is received by the CCD image sensor 31. here,
A CCD is used as an image sensor. This is a camera tube,
A photodiode array can also be used. CCD
An output signal from the image sensor 31 passes through an image receiving circuit 32 and a data processing circuit 33, and is output by a liquid crystal panel driving circuit 34.
When the liquid crystal panel 35 is driven, the C
An intensity transmittance distribution, a refractive index distribution, or both of them (collectively referred to as a complex amplitude transmittance distribution) are formed based on the distribution of the amount of light incident on the CD imaging element 31.

Therefore, by making the coherent light incident on the liquid crystal panel 35, the emitted light has a complex amplitude distribution based on the light intensity pattern incident on the CCD image pickup device 31 on the liquid crystal panel 35. Will have. still,
Here, the liquid crystal panel 35 is used as the most general electric address type spatial light modulator. In addition, a liquid crystal panel in which electrodes are arranged in a matrix on PLZT made of transparent ceramics has been developed.

By the way, the data processing circuit 3 in the present embodiment
3 is useful in that, for example, binarization or other filtering and image synthesis can be easily performed on an input image by connecting it to a circuit itself or a computer. Further, as described above, the spatial light modulator used here can use a CCD image sensor, a liquid crystal panel, and a circuit group associated therewith. Thereby,
A spatial light modulator can be easily and inexpensively manufactured.

[0036]

Fifth Embodiment In the third embodiment, an example of the optical autocorrelation pattern detecting means has been described. In such an optical autocorrelation pattern detecting means, the dynamic range of the spatial light modulator is small. In this case, that is, when the extinction ratio is low, a large amount of light is concentrated on the optical axis of the detection surface, and the detection of the autocorrelation pattern may be inaccurate. this is,
On the display surface of the spatial light modulator, the amount of light leaked from a place where the transmittance should originally be 0 is concentrated on the optical axis on the detection surface which is the Fourier transform surface. This is because In such a case, the input pattern to be identified is displayed on the first spatial light modulator or the spatial light modulator in a positional relationship of two parallel movements, and this is displayed. It is preferable to treat a pattern corresponding to the square of the correlation between two patterns as a square pattern of the autocorrelation.

FIG. 5 is an explanatory diagram for explaining the display of the input pattern to be identified and the intensity pattern of the autocorrelation having the above configuration. Here, “a” in FIG. 5 is an output light intensity distribution from the spatial light modulator 17 obtained on the readout surface of the spatial light modulator 17 in FIG. 3 or an output light intensity on the liquid crystal panel 35 in FIG. In the input pattern display area 41, two intensity distributions corresponding to the input pattern 42 are output at positions mutually parallel to each other. These intensity distributions are based on the input pattern to the autocorrelation pattern detection device 1, and the two input patterns are displayed at positions parallel to each other.

Such an input is applied to the spatial light modulator 1 of FIG.
In No. 7, two imaging lenses are arranged with respect to the input pattern, and two input patterns are imaged on the writing surface of the spatial light modulator 17, thereby easily realizing. Can be. In the spatial light modulator shown in FIG.
As before, two imaging lenses may be used, and C
The captured image of the input pattern may be synthesized by the data processing circuit 33 by the CD image sensor 31 and displayed on the liquid crystal panel 35 twice. FIG.
The second spatial light modulator (ie, spatial light modulator 2 shown in FIG. 3)
0) or the spatial light modulator (that is, the spatial light modulator 20 in FIG. 3 replaced with the spatial light modulator shown in FIG. 4) to show the input or output light intensity distribution. this is,
An interference fringe pattern is added to the square pattern (ie, intensity pattern) of the Fourier transform of one pattern of the input pattern 42.
Are superimposed.

The period of the interference fringe pattern is inversely proportional to the distance between the two input patterns 42 shown in FIG. 5A, and the direction of the periodic structure is two input patterns shown in FIG. 5A. 42 corresponding to the relative directions. The intensity pattern obtained by optically Fourier transforming the intensity pattern 43 of the Fourier transform of the input pattern shown in FIG.
In the pattern shown in FIG. 5C, this is the intensity pattern 44 of the 0th-order autocorrelation and the intensity pattern 4 of the ± 1st-order autocorrelation.
5, and this intensity pattern is obtained on the wedge ring detector 23 in FIG. Where 0
The next auto-correlation intensity pattern 44 is the auto-correlation of the entire pattern shown in FIG. 5A, and the ± 1 order auto-correlation intensity pattern 45 is the two auto-correlation patterns shown in FIG. 5A. Putter
This is an autocorrelation pattern by the mutual 42.

At this time, if the modulation degree of the second spatial light modulator or the spatial light modulator is low, all of the light leaked there will be on the optical axis on the output surface (ie, 0 in FIG. 5C). (The central part of the intensity pattern 44 of the next autocorrelation). Therefore, it is impossible to accurately observe the zero-order autocorrelation intensity pattern 44.
If the output is observed by aligning the center of the wedge ring detector in FIG. 3 with any center of the intensity pattern 45 of the ± 1 order autocorrelation, the spatial light modulation or the modulation degree of the spatial light modulator can be obtained. In this case, accurate recognition can be performed while avoiding the adverse effect when the value is low.

[0041]

Sixth Embodiment In the second embodiment, if an autocorrelation pattern detection area as shown in FIG. 2 is set, it is possible to identify rotation invariance or size invariance. However, the input pattern to be identified
If the screen has both rotation in the input plane and a change in size, it is difficult to cope with it only by setting these areas. Therefore, in this case, the central angle of the entire sector corresponding to the small area A shown in FIG. 2 is set to approximately 180 degrees, and the arrangement pattern of the output signals from each of the small areas is within the small area group. The position of the largest output or the smallest output is the predetermined position, and the order before and after is the same as the spatial arrangement of the small area group.
It is preferable that a rearrangement device for a signal sequence to be shifted is provided, so that the rearranged signal sequence pattern is input to a neural network. In this way, since the small region group A is originally invariable in size, by adding rotation invariance to the small region group A, accurate identification can be performed even when both the size and the rotation change.

FIG. 6 is a graph illustrating the operation of the apparatus for rearranging the signal sequence. That is, the same input pattern
7 is a graph showing an output from the small area group A and a corresponding output from the signal sequence rearrangement device when there is no rotation of the input pattern and when there is a rotation of the input pattern. 6A shows the output from the small area group A, b shows the output from the small area group A when the input pattern is rotated, and c shows the signal sequence corresponding to the output shown in a. D, indicating the output from the relocation device
Indicates an output from the signal sequence rearrangement device corresponding to the output indicated by b. That is, the output from the small area group A when the input pattern to be identified is at a certain reference in-plane rotational position is shown. The number of small areas is eight, and each of the small areas 6a
1 to the small area 6a8.

FIG. 6A is a graph showing the output intensity from each of the small areas. In contrast, FIG.
Shows the output from each small area when the same input pattern as in the case of a is rotated by a certain angle in the input plane. Comparing FIGS. a and b, the output of 6a3 in FIG.
The output of 6a5 in FIG. 6B is the output of 6a5 in FIG. 6B, the output of 6a7 in FIG. 6A is the output of 6a6 in FIG.
It can be seen that the output is cyclically shifted by two small areas. Therefore, the signal pattern rearrangement device cyclically shifts and rearranges the outputs from these small areas with reference to the position of the maximum value or the position of the minimum value, thereby obtaining an input pattern.
The output signal can be the same even if the motor rotates in the plane. FIGS. 6c and 6d show the outputs from the small area group A in FIGS. 6a and 6b, respectively, rearranged by the signal sequence rearrangement apparatus based on the maximum value. The signals before and after that are cyclically shifted in the order of the small areas. Then, originally, the outputs of FIGS. 6A and 6B are cyclic shifts of the same pattern, so that the rearranged signal sequences of FIGS. 6C and 6D are equivalent.

Therefore, even if the input pattern rotates in the input plane, substantially the same input can be given to the neural network 8. Further, since the output from the small area group A is originally invariant to the change in the size of the input pattern, even if the change in the size of the input pattern and the rotation in the input plane occur simultaneously. Accurate identification can be performed without increasing the number of storage patterns in the neural network.

In all of the above-described embodiments, the output from the autocorrelation pattern detector 1 is binarized and then input to the neural network 8, so that the learning can be performed more quickly. Can converge. Therefore, it is possible to obtain a pattern discriminating apparatus that performs faster learning.

[0046]

As described above, the above-described effects can be obtained by the optical pattern discriminating apparatus of the present invention.
Summarizing them has the following remarkable technical effects. That is, firstly, an inexpensive apparatus capable of processing an input image by a neural network at high speed can be realized.

Second, even when there is a rotation or a shift of the input image to be identified, the identification can be easily performed.
An optical pattern identification device capable of storing a large number of reference images could be provided.

Third, by extracting the features of the input information to be identified, the learning time in the neural network can be shortened, and a pattern recognition device having a high possibility of effective learning is provided. We were able to.

[Brief description of the drawings]

FIG. 1 is a schematic diagram showing an example of an optical system for realizing an optical pattern identification device according to the present invention.

FIG. 2 is an explanatory diagram showing an autocorrelation pattern detection area used in the present invention.

FIG. 3 is a block diagram showing an example of an optical autocorrelation pattern detecting means used in the present invention.

FIG. 4 is a configuration diagram illustrating an example of a spatial light modulator used in the present invention.

FIG. 5 is an explanatory diagram illustrating an intensity pattern of an autocorrelation when two input patterns to be identified are displayed according to the present invention.

FIG. 6 shows the output from the small area group A and the signal sequence corresponding thereto with and without the rotation of the input pattern for the same input pattern to be identified, which is performed in the present invention. It is an explanatory view showing an output from an arrangement device.

[Explanation of symbols]

REFERENCE SIGNS LIST 1 autocorrelation pattern detection device 6a small region group A 6ai, 6bi small region 6b small region group B 7 signal sequence rearrangement device 8 neural network 9 input pattern to be identified 11 light source 12, 12a, 12b Light beam 17, 20 Spatial light modulator 19, 22 Fourier transform lens 23 Wedge ring detector 31 CCD image sensor 32 Image receiving circuit 33 Data processing circuit 34 Liquid crystal panel drive circuit 35 Liquid crystal panel 41 Input pattern display area 42 Input pattern 43 Intensity pattern of Fourier transform of input pattern 44 Intensity pattern of 0th-order autocorrelation 45 Intensity pattern of ± 1st-order autocorrelation

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-2-293985 (JP, A) JP-A-64-46888 (JP, A) JP-A-2-293827 (JP, A) JP-A 64-64 33686 (JP, A) JP-A-64-26285 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06F 15/18 G06T 7/00

Claims (7)

(57) [Claims] 1. At least a plurality of inputs can be accepted;
By calculating the input values of the plurality of inputs ,
A plurality of operation units (hereinafter referred to as neurons) for outputting one output value, and coupling means for coupling an output of each of the neurons with an input of another neuron or a neuron that has performed the output; Each coupling means is a pattern identification device for identifying a pattern using a neural network having a coupling coefficient for converting an output value related to the input value into an input value. for obtaining, or with a self-correlation pattern detector for obtaining a square of the pattern of the autocorrelation of the input pattern, and, to the output based on the autocorrelation pattern obtained with the input of the neural network, further wherein When the autocorrelation pattern of is a two-dimensional pattern
Is centered on the peak position of the autocorrelation pattern
Fan-shaped two-dimensional with a central angle of approximately 180 degrees or more
In space, divide this into smaller fan-shaped areas with equal central angles
In the fan-shaped two-dimensional space,
This is reduced by a concentric arc centered on the peak position.
Assuming a small area group B divided into areas, the input pattern
Of the autocorrelation pattern or the square of the autocorrelation
The integral of the turn value within that sub-region is calculated for each sub-region.
Out of these, the small area group A
Integrated value, or the integrated value of the small area group B, or
The optical pattern identification device, wherein both of them are input to a neural network . 2. The pattern of the square of the autocorrelation of the input pattern.
The autocorrelation pattern detection device for obtaining the
At least, the coherent light source and the input pattern to be identified
Spatial light modulator or space that displays coherently
A light modulator and the first spatial light modulator or a screen of the spatial light modulator.
Optically Fourier transform the coherent pattern displayed in
A first Fourier transform lens for transforming, and a complex amplitude distribution obtained by Fourier transform by the first Fourier transform lens.
The intensity distribution of the output coherent light according to the intensity distribution or
Second spatial light modulator or a spatial light modulating the complex amplitude distribution
A modulator and an output from said second spatial light modulator or spatial light modulator
Second optical element for optically Fourier transforming coherent light again
2. The lens according to claim 1, further comprising a Fourier transform lens.
An optical pattern identification device as described. 3. The method according to claim 1, wherein each of the integral values is output.
The light receiving element having a shape corresponding to the region is formed by the second Fourier
Be placed on the Fourier transform plane by the transform lens
The optical pattern according to claim 1, wherein:
Identification device. 4. The first and / or second spatial light modulator.
Alternatively, the spatial light modulator uses a two-dimensional image sensor as an input unit.
The output video signal of this image sensor is
To the two-dimensional image sensor.
A sensor formed based on the intensity distribution of incident light
3. A display for displaying a turn.
Or the optical pattern identification device according to 3. 5. The square pattern of the autocorrelation of the input pattern.
The autocorrelation pattern detection device for obtaining the
The spatial light modulator or spatial light modulator
The force pattern is displayed twice in a translational relationship.
The square of the correlation between the two indicated patterns is
5. A square pattern according to claim 2, wherein:
An optical pattern identification device according to any of the preceding claims. 6. An integrated value of the small area group A is output.
In this case, the central angle is set to approximately 180 degrees,
Change the arrangement pattern of the output signals
Output signal pattern of the output signal from the
The position during the turn is always the predetermined position,
The order of output from the small areas before and after that is the space of the small area group.
Reallocation of the signal sequence, shifting it to be similar to the
And the rearranged signal train pattern
Is input to the neural network.
The optical pattern according to any one of claims 1 to 5,
Identification device. 7. An output from the auto-correlation pattern detection device is adapted.
After binarization using the appropriate threshold level,
7. The method according to claim 1, wherein the information is input to a network.
An optical pattern identification device according to any of the preceding claims.