JP3020629B2 - 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 the pattern of the input image has some ambiguity such as missing or deformation, it is possible to associate a stored image (reference image) corresponding to the ambiguity, and to perform character recognition. It is considered to be effective for voice recognition and other pattern recognition. However, this neural network requires an enormous number of neurons (units) in accordance with the number of pixels of the storage pattern and the number of patterns. Since wiring between neurons of the square of the number of Rons is required, it is difficult to store a large amount of information or many 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 a pattern identification apparatus capable of identifying an input image even when the input image is rotated or shifted, and capable of storing a large amount of reference images. Furthermore, the present invention extracts the features of the input information,
An object of the present invention is to provide a pattern discriminating apparatus in which learning time in a neural network is short, and learning is likely to be performed effectively.

[0006]

Means for Solving the Problems The present invention has been made to solve the above technical problems, and at least, has been achieved.
Receives a light intensity distribution pattern corresponding to an input pattern to be identified, approximately 180 degrees or 3 around the optical axis position.
In a fan-shaped two-dimensional space having a central angle of 60 degrees, a spatial small area group A is obtained by dividing this into two fan-shaped small areas having the same central angle, and outputs a value based on the amount of light in each small area. Light-receiving element group; output signal patterns by arranging respective outputs
In this case, the position in the signal pattern of the signal corresponding to the signal having the largest output or the signal having the smallest output among the small region groups is always a predetermined position, and Signal sequence rearranging means for shifting so that the order before and after the small region group is arranged in the two-dimensional space; one or a plurality of inputs can be received, and an input value based on the input is processed. Thus, a plurality of operation units (hereinafter, referred to as units) that output one output value, a connection that connects an output of each unit and an input of another unit or an input of the unit that has performed the output itself Means; each of said combining means comprises a neural network having a coupling coefficient for converting an output value associated therewith into an input value of a partner to be combined, wherein said neural network has a coupling coefficient. Output pattern - each output in the emission, respectively,
In a pattern discriminating apparatus for discriminating a pattern by inputting a corresponding unit of the neural network, one discriminator is used for learning a coupling coefficient of the neural network. A pattern belonging to a class (hereinafter referred to as a class) is rotated around the optical axis by a plurality of types of angles smaller than the central angle of the small area with respect to a reference presentation position, and presented. The pattern identification method is characterized by performing learning.

The output of the light receiving element group is obtained by adding an output value of an adjacent small area to an output value corresponding to one small area as an output value in the signal train rearrangement means. What is input is preferred.

In order to solve the above-mentioned problem, the present inventors have at least received a light amount distribution pattern corresponding to an input pattern to be identified at least approximately 180 around the optical axis position.
In a fan-shaped two-dimensional space having a central angle of 360 degrees or 360 degrees, a spatial small area group A obtained by dividing this into smaller fan-shaped sub-regions having the same central angle is provided. In the case of (1), it is assumed that the small areas at both ends are regarded as adjacent small areas, and a light receiving element for outputting a value based on the light quantity in each small area and the respective outputs are arranged to form an output signal pattern. In this case, the position in the signal pattern of the signal corresponding to the signal having the largest output or the signal having the smallest output among the small region groups is always a predetermined position, and the position before and after the position is determined. The order is such that the signal sequence rearrangement means for shifting and one or more inputs can be received so as to follow the order of arrangement of the small area groups in the two-dimensional space,
By performing an arithmetic operation on an input value based on the input, a plurality of operation units (hereinafter, referred to as units) for outputting one output value, an output of each unit, an input of another unit, or an output of the other unit are performed. And a neural network having a coupling coefficient for converting an output value related thereto into an input value of a partner to be coupled. A pattern identification device was proposed (filed on February 4, 1991).

The configuration of one example of this pattern identification device is shown in the schematic configuration diagram of FIG. FIG. 3A illustrates the light receiving area of the light receiving element group used here.
(B) is a schematic configuration diagram of an example of the optical autocorrelation pattern detection means used here. FIG. 4 is an explanatory diagram showing the output from the light receiving element group and the corresponding output from the signal sequence rearrangement device in the pattern identification device. That is, it shows the output from the small area group A when there is no rotation of the input pattern in the same input pattern and when there is a rotation, and the output from the corresponding signal sequence rearrangement device. (A) shows the output signal pattern from the small area group A, and (B) shows the output signal pattern from the small area group A when the input pattern is rotated. (C) shows the output signal pattern from the signal sequence rearrangement device corresponding to the output shown in (a), and (d) shows the output signal pattern.
The output signal pattern from the signal sequence rearrangement device corresponding to the output shown in (b) is shown.

In the pattern discriminating apparatus of FIG. 2, the presented input pattern 9 is converted into an auto-correlation pattern in an auto-correlation pattern detecting apparatus, and FIG.
The light enters a light receiving element group having a light receiving area corresponding to the small area group A6a shown in FIG. This autocorrelation pattern is
The light receiving area of FIG. 3A is indicated by a broken line. Therefore, each light receiving element outputs an output value corresponding to each small area 6An and corresponding to the amount of light incident on each small area. When these outputs are arranged in the order in which they are arranged in the light receiving element,
The signal train shown in FIG. Here, assuming that the input pattern is rotated by, for example, 90 degrees within the display surface, the output signal sequence from the light receiving element group is, as shown in FIG. The output is to 6a5 in (b), and similarly the output of 6a1 is to 6a6, and
The output of 6a5 shifts to 6a1, and the output of 6a6 shifts to 6a2. Therefore, the signal sequence rearrangement device 7 arranges, for example, the largest output at the beginning, arranges the outputs before and after the largest output in the order of the arrangement of the light receiving elements, and outputs the output of the light receiving element located at the end to the other. If it is arranged so as to be adjacent to the output of the light receiving element located at the end, the output from the signal sequence rearrangement device 7 with respect to the output shown in (a) as shown in (c) and (d) of FIG. Also, the output from the signal sequence rearrangement device 7 for the output shown in (b) is equivalent.

Therefore, in the above-mentioned pattern discriminating apparatus, if the rotation is invariable and the pattern on the light-receiving element does not exceed the range of the light-receiving element, it is recognized that the size is unchanged. Can be. Furthermore, if the input pattern is subjected to a shift invariant operation such as Fourier transform or autocorrelation as in this example, and input to the light receiving element, the input pattern is located at the position in the input screen. Invariant recognition can be performed. FIG. 3B shows an example of an optical system for optically performing the autocorrelation calculation. FIG.
In (b), the light beam 12 emitted from the light source 11 is
The beam diameter can be expanded to an appropriate size by the beam expander 13, and the beam path of the beam can be expanded by two beams by the beam splitter 14.
Divided into two. The light beam 12a reflected by the beam splitter 14 is reflected by the beam splitter 18 and enters the spatial light modulator 17.

Here, as the spatial light modulator 17, a liquid crystal light valve or the like in which a liquid crystal panel is combined with a photoconductive film is used. The input pattern 9 forms an image on the writing side of the spatial light modulator 17 (that is, the side on which the photoconductive film is disposed in the liquid crystal light valve) by using the imaging lens 16. Then, according to the light quantity distribution, the spatial light modulator 17
The local electric field applied to the spatial light modulator 17 changes, and the birefringence or optical rotation on the readout side of the spatial light modulator 17 (that is, the side where the liquid crystal is arranged in the liquid crystal light valve) changes according to the electric field intensity distribution. This change in birefringence or optical rotation can be extracted as a change in the phase of the incident light or as a change in the intensity of the incident light using an analyzer. Accordingly, the light flux 12a incident on the spatial light modulator 17
Is reflected with a phase distribution or intensity distribution corresponding to the input pattern 9, passes through the beam splitter 18, passes through the Fourier transform lens 19, and is incident on the writing side of the spatial light modulator 20. I do.

The spatial light modulator 20 includes a spatial light modulator 1
8 may be used, but the writing surface of the spatial light modulator 20 and the reading surface of the spatial light modulator 17 are disposed on the front focal plane and the rear focal plane of the Fourier transform lens 19, respectively. Thus, an optical Fourier transform pattern of the input pattern can be obtained on the writing surface of the spatial light modulator 20. Here, the light flux 12b transmitted through the beam splitter 14 is reflected by the front reflection mirror 5 and further reflected by the beam splitter 21 to form the spatial light modulator 2b.
0 is incident on the readout surface. And the spatial light modulator 17
Similarly, in the light beam 12b reflected on the reading surface of the spatial light modulator 20, a phase distribution or an intensity distribution corresponding to the light amount distribution on the writing surface is obtained.
2b is transmitted through the beam splitter 21 and passes through the Fourier transform lens 22 to enter the wedge type detector 23.

Here, the wedge type detector is shown in FIG.
A light receiving element having a wedge type light receiving area as shown in FIG. Therefore, the wedge type detector 23
Above, a Fourier transform pattern of the intensity pattern on the writing surface of the spatial light modulator 20 is observed, and this pattern has its autocorrelation pattern with respect to the input pattern. It becomes. However, in such a pattern discriminating apparatus, the above-mentioned problem has been solved. However, when the input pattern is rotated by an angle smaller than the center angle of the sector, the output pattern is reduced. Despite the slight rotation, it may change greatly, and it is necessary to further improve the invariance of the input pattern with respect to the rotation.

[0015]

According to the structure of the present invention, even if the size of the input pattern changes, the output from each small area does not change much if it is within the size of each small area. Input information that is invariant with respect to the size of the input pattern can be provided to the neural network. Further, according to the configuration including the signal sequence rearrangement means, when the change in the light amount distribution on the light receiving element is small within the small area, the input pattern rotates in the input plane. However, since the output pattern from the light receiving element group is only displaced in the absolute position and the relative positional relationship between the undulations of the output pattern is almost equal, the signal sequence of the small area of the maximum output or the minimum output is obtained. Is fixed by the signal sequence rearranging means, a constant input signal with respect to the in-plane rotation of the input pattern can be given to the neural network.

At this time, when the change of the light quantity distribution on the light receiving element becomes large within the small area, depending on which part of the light quantity distribution pattern the boundary of each small area is. , The fluctuation of the output pattern becomes large. At this time, the output data when the input pattern is rotated by an angle smaller than the central angle of the small area is taken as it is for a plurality of angles, and the new data is obtained by using those data.
By learning the ral network, it is possible to perform discrimination that can cope with all 360-degree rotations only by learning a small number of rotation pattern data in a small range. Further, regarding the output of the light receiving element group, an output value obtained by adding an output value of one determined adjacent small area to an output value corresponding to one small area,
By inputting the signal to the signal sequence rearranging means, even when there is a large light amount change in a small area, when the light amount distribution pattern of one part
The output value is the sum of the outputs of the two small areas. Therefore, the output pattern does not change drastically when the light quantity distribution pattern of the part falls within one small area.

The Fourier transform intensity pattern and the autocorrelation pattern of the input pattern to be identified are invariable even if the position where the input pattern is presented changes. By optically performing a Fourier transform or an autocorrelation operation with respect to the input signal and receiving the output light amount by the light receiving element group, it is invariant not only to rotation but also to shift of the input pattern. Identification can be performed.
Since the linear transformation intensity pattern and the autocorrelation pattern are point-symmetric patterns, the fan-shaped region can input a pattern equal to all patterns at 180 degrees.

Next, the pattern identification method of the present invention will be described in detail with reference to examples, but the present invention is not limited thereto.

[0019]

Embodiment 1 FIG. 1 is an explanatory diagram for explaining an identification process performed according to an example of a pattern identification method of the present invention. That is, if the input pattern is the same, there is no rotation of the input pattern, or if there is, the autocorrelation pattern on the small area group A and the output from the corresponding signal sequence rearrangement device. FIG. 5 is an explanatory diagram showing a state of an output signal pattern. That is, FIG. 1A shows the autocorrelation pattern of the input pattern, and shows the autocorrelation pattern when the light receiving region and the input pattern are not rotating. FIG. 1B shows an autocorrelation pattern when the light receiving region and the input pattern are rotated by 11 degrees. FIG. 1C shows an output signal pattern from the signal sequence rearrangement device corresponding to the pattern shown in FIG. FIG. 1D shows an output signal pattern from the signal sequence rearrangement device corresponding to the pattern shown in FIG.

The used signal sequence rearrangement device is basically the same as the conventional device, and is similar to the device shown in FIG. However, the pattern discriminating method of the present invention is characterized by its neural network learning method.

In this embodiment, for the sake of simplicity, the light receiving element of the wedge type detector, that is, the small area group A shown in FIG.
The number of small areas of 6a is eight. The pattern input to the light receiving element is an autocorrelation pattern of the input pattern, which can be obtained by the optical system shown in FIG. Here, since the autocorrelation pattern is used, the central angle of the sector of the small area group A is 18
0 degrees. The number of divisions of the light receiving area and whether to take Fourier transform or autocorrelation of the input pattern can be freely selected as needed. In the present embodiment, since the number of small regions is set to 8, the central angle of the sector of each small region is about 22.5 degrees. Therefore, for example, it is assumed that the input pattern is rotated in the input plane by 11 degrees which is about half of the central angle. Then, the autocorrelation pattern is also shown in FIG.
As shown in (b), the rotation is performed on the small area group A by the same angle. Comparing the output from the signal sequence rearrangement means at this time, the results are as shown in FIGS. 1C and 1D, and the output pattern from the signal sequence rearrangement means 7 despite the small rotation angle. Are very different.

The data numbers shown in FIGS. 1 (c) and 1 (d) indicate that the output of 6a6 in the small area is the largest for the small area shown in FIGS. 1 (a) and 1 (b). Starting from this, 1 is 6a6, 2 is 6a7, 3 is 6a8, and 4 is 6a.
1, 5 correspond to 6a2,. in this way,
Even if the rotation angle of the input pattern is small, the output from the signal sequence rearrangement means 7 may vary greatly. For example, the central angle α of this small area is divided by n, and the input is performed by α / n degrees. By rotating the pattern, taking output data n times, learning of a neural network is performed so as to identify the input pattern by the output data, and the center of the small area is obtained. To be able to identify even if rotation occurs in the corner,
If a neural network is organized, small area 1
With one rotation, the same output as that of the non-rotated one can be obtained by the operation of the signal sequence rearranging means 7, so that the invariant identification can be made for all the rotation angles. Note that the neural network that can be learned at this time is based on the error back propagation method (Error Back P
Hierarchical networks based on the propagation method and the Boltzmann machine are common, and various improved learning algorithms are being actively studied.

The output from the signal sequence rearranging means 7 is obtained by normalizing other output values according to the maximum value of the signal sequence and then inputting the output to the neural network. Can be used without leaving the dynamic range of

[0024]

Embodiment 2 Next, as shown in FIG. 5, the output from each small area in the small area group A is shown. The output for each small area 6an in the small area group A is shown. In the first embodiment, it has been described that the output from the signal sequence rearrangement means 7 may change greatly even with a slight rotation of the input pattern. However, this large change is extremely large. In such a case, completely different patterns must be stored as the same class in the neural network.
This increases the load on the network in order to cope with increasing the number of classes to be identified in the neural network.

Such a large fluctuation of the output due to a slight rotation occurs when a large change in the pattern on the small area group A occurs within a small area compared to the size of each small area. . In particular, when the region of the highest intensity in the pattern rides on the boundary of the small region, the intensity is divided into two regions sandwiching the boundary, and the other region of lower intensity is used. When the output of the signal sequence becomes larger, the position of the head of the signal sequence changes on the pattern in the small area group A.
It is possible that the output pattern from the server will be completely different. At this time, as shown in FIG. 5, by adding the outputs of the adjacent small areas (S1 to S8) to the signal sequence rearrangement means 7 as outputs from the small area group,
Since the output corresponding to the region having the strongest intensity in the pattern is always the largest, the position (region) on the pattern corresponding to the first signal in the signal sequence rearrangement means does not change. Significantly, the output from the signal sequence rearrangement means does not have a large change in rotation for the same input pattern.

Incidentally, what can be said in common to all the embodiments is that, by binarizing the output from the autocorrelation pattern detecting device 1 and then inputting it to the neural network 8, the learning is carried out. Since the convergence is performed more quickly, a pattern discriminating method for performing faster learning can be obtained.
Further, the signal sequence rearrangement means detects the maximum value or the minimum value from a group of signal sequences, and arranges the position of the value in the signal sequence as a determined position so that the order around the position is not changed. Since it can be changed, the device can be realized by software or hardware using a memory or a delay circuit.

Here, the wedge ring detector 23
Are photoelectric conversion elements having light receiving regions of respective shapes, the center of which is set on the optical axis of the Fourier transform optical system. 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 spatial light modulators 17 and 20 used here
For example, a liquid crystal panel provided with a photoconductive layer on a nematic liquid crystal panel is commercially available from Hughes, USA under the name of Liquid Crystal Light Valve (LCLV). Also, devices using ferroelectric liquid crystals have been developed as having higher speed operation and higher resolution. Furthermore, Bi
A transmission type element using ₁₂ SiO ₂₀ (BSO) is also available. However, in the case of a transmissive element, a change in the optical system is necessary because the direction in which the readout light beam is incident is changed, but this is essentially the same as the case in which a reflective spatial light modulator is used. .

The optical autocorrelation pattern detecting means 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 devices. That is,
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, thereby realizing the real time operation of the pattern identification device. This is an important feature in doing so. Secondly, the optical axis can be easily aligned, and therefore, it can be easily assembled and adjusted, and can be used in a bad environment.

[0030]

As described above, the above-described effects are obtained by the pattern identification device of the present invention. Summarizing them has the following remarkable technical effects. That is, first, a pattern that realizes identification of an input image by a neural network with an inexpensive device capable of processing at high speed.
A method for identifying the user can be provided.

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 characteristics of the input information to be identified, the learning time in the neural network can be shortened, and a pattern identification device having a high possibility of effective learning is provided. I was able to.

[Brief description of the drawings]

FIG. 1 shows an autocorrelation pattern with and without rotation of an input pattern in pattern identification performed by an optical pattern identification apparatus of the present invention, and a signal sequence rearrangement corresponding thereto. It is a schematic block diagram which shows the mode of the output from a means.

FIG. 2 is a configuration diagram showing a configuration of an example of a conventional pattern identification unit.

FIG. 3 is an explanatory diagram showing a light receiving region used in the present invention and a configuration diagram showing an example of an optical autocorrelation pattern detecting means using the same.

FIG. 4 is an explanatory diagram showing an output from a small area group, the same output when an input pattern is rotated, and an output from a signal sequence rearrangement means corresponding to each of them, used in the present invention.

FIG. 5 is an explanatory diagram showing an output from each small area performed in the present invention.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 autocorrelation pattern detection device 6a small region group A 6ai, 6bi small region 7 signal sequence rearranging device 8 neural network 9 input pattern to be identified 11 light source 12, 12a, 12b light flux 17, 20 space Optical modulator 19, 22 Fourier transform lens 23 Wedge ring detector

──────────────────────────────────────────────────続 き Continued on the front page (58) Fields investigated (Int. Cl. ⁷ , DB name) G06T 7/00 G06F 15/18 G02B 27/00

Claims (2)

(57) [Claims] 1. A fan-shaped two-dimensional space having a central angle of approximately 180 degrees or 360 degrees about an optical axis position for receiving at least a light amount distribution pattern corresponding to an input pattern to be identified. A light receiving element group having a spatial small area group A obtained by dividing this into fan-shaped small areas having the same central angle, and outputting a value based on the amount of light in each small area; an output signal pattern The position in the signal pattern of the signal corresponding to the largest output or the smallest in the small area group is always a predetermined position, and A signal sequence rearrangement means for shifting so as to follow the order of arrangement of the small area groups in the two-dimensional space, and one or more inputs can be received, and an input value based on the input is arithmetically processed. By doing A plurality of operation units (hereinafter, referred to as units) for outputting one output value, and coupling means for coupling an output of each unit with an input of another unit or an input of the unit which has performed the output. A neural network having a coupling coefficient for converting an output value related thereto into an input value of a partner to be coupled, and an output pattern from the signal sequence rearranging means. A pattern identification device for identifying a pattern by using each output in the pattern as an input of a corresponding unit of the neural network.
When learning the coupling coefficient of the ral network, a pattern belonging to one identification classification (hereinafter, referred to as a class) is represented by:
The pattern discriminating method according to claim 1, wherein the learning is performed by rotating around the optical axis by a plurality of types of angles smaller than the central angle of the small area with respect to a reference presentation position, and presenting the rotation. . 2. The signal train rearrangement means according to claim 1, wherein said output of said light receiving element group is obtained by adding an output value of an adjacent small area to an output value corresponding to one small area as an output value. 2. The pattern according to claim 1, wherein the input is performed within a field.
Identification method.