JPH09185713A - Feature vector extraction device and information processor using the device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To attain fast conversion into a feature vector which is effective for the recognition and classification of an input vector by extracting plural feature quantities from multiplied vector information by its multiplicity in parallel and converting them as feature vectors. SOLUTION: An object vector input display means 1 inputs at least vector information to be processed, to the system. An object vector multiplying means 2 multiplies and duplicates the inputted vector information to be processed. Then a feature vector conversion output means 3 extracts plural feature quantities from the multiplied vector information to be processed by the multiplicity in parallel, converts them into new feature vectors, and output the vectors. In this case, the filter array 37 of the feature vector conversion output means 3 filters an image as the multiplied and duplicated object vectors differently plural times. This filter array 37 uses Gabor filters G, which extract directional spatial frequency structures from the image, as its constituent filters.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a feature vector extraction device for converting an input vector into a feature vector effective for recognition and classification, and an information processing device for performing recognition and classification using this feature vector extraction device. Regarding

[0002]

2. Description of the Related Art Normally, when recognizing or classifying various kinds of information such as images and signals, these are regarded as vector quantities, and recognition or classification is performed by calculating the degree of similarity with other vectors serving as comparison criteria. . Conventionally, a combination of a matched filter and a correlator, a joint transform correlator, a neural network, or the like has been used as a means for calculating the degree of similarity.

However, when images or signals having complicated characteristics are directly processed by these methods, information that does not contribute much to recognition or classification must be processed at the same time, which causes an error. In addition, there are many cases in which an error is generated due to a sharp reaction to some deformation.

Therefore, in order to solve these problems, various kinds of information such as images and signals are not directly processed, but features that greatly contribute to recognition and classification are extracted by preprocessing, and the extracted features are used for recognition. There have been many attempts to do so far. Taking an image as an example, these features are:
Typical examples are textures, structural features, colors, and temporal features. For texture, density histogram, co-occurrence matrix, difference statistics, etc. are calculated.For structural features such as edges, lines, and contours, Laplacian filter convolution, Hough transform, etc. , RGB space and H
With respect to the temporal characteristics by conversion into SV space or spectrum, etc., the characteristic amount is obtained by calculation of optical flow or conversion into wavelet.

However, the conversion into the feature quantity by these preprocessing takes an enormous amount of time, especially when a two-dimensional image is a target vector, and is not practical in a normal computer serial calculation. , Optical methods and parallel processing methods such as parallel computers have been adopted. Among them, as an optical method, conventionally, spatial frequency filtering using a lens is often used as a method for detecting an edge of a two-dimensional image or extracting a specific frequency component at high speed. Further, as a method of further developing this method and constructing a system compactly and inexpensively, there is a method disclosed in Japanese Patent Publication No. 3-66645. This method will be briefly described with reference to FIG. FIG.
(A) is a cross-sectional view of one lens, (b) is an overall cross-sectional view, and (c) is a perspective view showing the arrangement of a light shield with respect to the lens. In this method, both end surfaces are flat. A plurality of gradient index lenses 102 whose refractive index changes in the radial direction are arranged in parallel with each other, and a circular light shield 106 having a fixed diameter smaller than the lens outer diameter about the lens optical axis on the focal plane of each lens. By providing the above, it is intended to extract only the high frequency component from the input object image 109 and use it for the inspection of scratches and foreign matters on the transparent body. According to this method, even a large-area image can be filtered in parallel using individual small-diameter (inexpensive) lenses without using large-diameter (expensive) lenses. In addition, the gradient index lens used is flat on both end faces, and if a light-shielding body that is a filter is directly provided on this face, alignment and the like can be configured easily and compactly. .

However, in order to further improve the accuracy,
It is necessary to perform recognition and classification by using not only one kind of these feature amounts but also a plurality of feature amounts (that is, using a plurality of filters instead of a certain filter). As an example of this method, there is a "visual recognition system using a microlens array", Shunichi Kamemaru, video information (I), pages 65 to 70 (proposed in January 1993 issue. This method is shown in FIG.
Using a device such as the one shown in Fig. 1, we extract partial structural features of the target object by a multiple correlator in which multiple matched filters and correlators are arranged, and use this result to recognize it in a neural network. It is something to say. More specifically, the input image (letter K) is divided into two left and right sub-elements in the pre-processing, and then the plate 201 is duplicated for each reference object (6 in total, 3 in vertical and 2 in horizontal). (There is no detailed description, but it seems that this part is actually made on the plate by human means.) next,
In the main processing, the division pattern of the duplicated input image on the produced plate 201 is the reference object 20 produced in advance.
Matched filter array 2 made up of 2 vertical 3 x horizontal 2
03 are arranged on the Fourier transform plane by the number of divisions (2),
Micro-lens arrays 204 and 205 each consisting of two sets of twelve (3 × 2 × 2) micro-lenses are input to a multiple correlation system that performs Fourier transform and inverse Fourier transform in parallel. As a result, the duplicated division pattern of the input image and the correlation signal of the reference object are output in parallel to the output surface 206. Finally, in post-processing, the 12 correlation signals obtained in parallel are taken into the computer 208 by the CCD camera 207, binarized, and input to the back propagation input layer prepared as a neural network. , To recognize. It is reported that this method enables recognition of four characters D, K, O, and X and one blank.

[0007]

As described above, the method of Japanese Patent Publication No. 3-66645, which is a conventional example, has the merit that the system is compact and can be constructed at low cost.
There is only one type of feature quantity that can be extracted, and although it is sufficient for a specific purpose, it is insufficient for general recognition and classification of images with high accuracy.

On the other hand, the method by Kamemaru is to extract a plurality of feature values simultaneously and in parallel and to recognize them by inserting them into a neural network, so that high-speed and accurate recognition is possible to some extent. ing. However, since multiple correlations are taken by using the matched filter array of actual character sub-elements in the feature extraction unit, it reacts sensitively to some deformations of the matched filter described in the conventional example and causes an error. The drawbacks that arise remain.
In addition, although it is sufficient to recognize about 5 characters with a relatively simple structure as shown in the example, if the number of recognition targets is increased or an image having a complicated feature is processed, the number of divisions into sub-elements is increased. It is not very practical considering that the number of reference vectors must be increased and the number of matched filter arrays will be (number of divisions into subelements) x (number of reference objects), which will dramatically increase. .

The present invention has been made in view of the situation of the prior art as described above, and an object thereof is to provide a feature vector extraction device for rapidly converting an input vector into a feature vector effective for recognition and classification. Another object of the present invention is to provide an information processing device that performs recognition and classification with high accuracy using this feature vector extraction device.

[0010]

The feature vector extraction apparatus of the present invention which achieves the above-mentioned object is at least a target vector input display means for inputting vector information to be processed into the system, and this input vector display means. Target vector multiplexing means for multiple duplicating the vector information to be processed, and a plurality of feature quantities are extracted in parallel from the multiplexed vector information to be processed to correspond to the multiplicity, and a new feature is newly added. It is characterized in that it is composed of a feature vector conversion output means for converting and outputting as a vector.

With such a configuration, a plurality of feature quantities can be simultaneously extracted in parallel and can be output at high speed as feature vectors.

In this case, it is desirable to include a plurality of spatial frequency filters for multiplicity in the feature vector conversion output means so as to extract a plurality of spatial frequency structures of the vector information to be processed. Gabor as its spatial frequency filter for
A general image having a complicated spatial structure is obtained by extracting a directional spatial frequency structure using a filter, a directional bandpass spatial frequency filter, a wavelet filter, or the like. Even if it is not divided into subelements, it is possible to quickly and satisfactorily extract the features of the target vector information that are effective for the subsequent processing such as recognition and classification with a simple and minimal filter.

It is also possible to include a plurality of color filters for multiplicity in the feature vector conversion output means so as to extract features in the color space of the vector information to be processed. In that case, complex color Even with an image having a distribution, it is possible to quickly and satisfactorily extract the features of the target vector information effective for the subsequent processing such as recognition and classification with a simple and minimal filter.

It is also possible to include a plurality of moment detection filters for multiplicity in the feature vector conversion output means so as to extract the features of a plurality of moments of vector information to be processed. It is possible to quickly and satisfactorily extract the features of the target vector information that are effective for the subsequent processing such as recognition / classification with a simple and minimal filter, even for an image containing an object with.

By using these filters in combination, even an image in which features are more complicatedly combined can be recognized and detected with a simple and minimum filter.
It is possible to quickly and satisfactorily extract the feature of the target vector information that is effective for the subsequent processing such as classification.

Further, the information processing apparatus of the present invention includes at least a feature vector extracting device for inputting vector information to be compared and converting and outputting to a feature vector, and a feature vector output from the feature vector extracting device. Feature vector input means for displaying in the system, reference vector input means for inputting a vector serving as a reference for comparison, and calculating a similarity between the input feature vector and a vector serving as a reference for comparison. Similarity matching means,
An information processing apparatus configured by a similarity determination unit for determining the degree of similarity is characterized by being configured by any one of the above feature vector extraction devices as the feature vector extraction device. .

In this case, since the similarity is calculated using the feature vector or reference vector extracted / converted by the feature vector extraction device as described above, it is conventionally used when performing similar processing. Problems such as the combination of matched filters and correlators, joint transform correlators, etc. that cause an error by reacting sharply to deformation and noise, and the problem of insufficient dynamic range on the Fourier transform plane have been solved. It has become possible to provide an information processing device that recognizes and classifies quickly and accurately.

Another information processing apparatus of the present invention is a feature vector extracting apparatus for inputting target vector information and converting and outputting to a feature vector, and a feature vector output from the feature vector extracting apparatus in a system. In an information processing apparatus having a feature vector display unit for displaying on the screen and a competitive learning unit for recognition / classification, the competitive learning unit includes a similarity matching unit, a weight vector update display unit, and a winning vector determination. And a feature vector extracting device as any of the above feature vector extracting devices.

In this case, since the similarity can be calculated by the feature vector extracted and converted by the feature vector extraction device as described above and the winning element can be determined, the recognition and the classification are performed at high speed and with extremely high accuracy. It has become possible to provide a competitive learning type information processing device.

[0020]

BEST MODE FOR CARRYING OUT THE INVENTION Preferred first to ninth embodiments of a feature vector extraction device of the present invention and an information processing device using the same will be described below with reference to the drawings. [First Embodiment] As shown in the cross-sectional view of FIG. 1, the feature vector extraction apparatus of the present embodiment includes a target vector input display means 1 for inputting at least vector information to be processed into the system, and Target vector multiplexing means 2 for multiple duplicating the input vector information to be processed,
A feature vector conversion output means 3 for extracting a plurality of feature amounts in parallel from the multiplexed vector information to be processed by the degree of multiplicity and newly converting and outputting as a feature vector. is there. In this embodiment, a filter for extracting the spatial frequency structure is used as the feature amount.

More specifically, the target vector input / display means 1 uses Xe-Hg as the incoherent light source 11.
A lamp, a lens 12a and a mirror 12b having the same focal length, which are arranged in a confocal arrangement in order to efficiently convert the light of the lamp 11 into a parallel luminous flux, and the lamp 11 is arranged at the focal position, and a single color of this luminous flux. Of the incoherent light beam 15 which is substantially parallel to each other by the wavelength filter 13 and the mirror 14 for increasing the property, and further the image sensor 16 such as a CCD camera and the spatial light modulator which receives the signal from the image sensor 16 In the case of the present embodiment, which is the target vector, an image is displayed on the spatial light modulator 17 by the controller 18 and the driver 19 for generating the signal to be displayed on 17, and this information is transmitted to the light beam 15 as the spatial light modulator. It is input into the system by making it incident on 17. In the case of the present embodiment, as the spatial light modulator 17, a liquid crystal spatial light modulator of the electric address type and the transmissive read type whose transmittance changes according to the applied voltage is used.
Further, although the system is bent by the mirror 14, these are not essential and the system may be straightened by omitting the mirror 14 if there is enough space.

The target vector multiplex means 2 has the same multiplicity of the lens array 22 as the number of filters for extracting the image input by the target vector input display means 1 by the imaging lens 21 and the lens array 22. The data is duplicated and written in the spatial light modulator 23. At this time, the imaging lens 21 has a lens array 22 so that the reading surface of the spatial light modulator 17 coincides with the front focal position.
Are arranged so that the writing surface of the spatial light modulator 23 coincides with the focal point on the rear side thereof, and further, the optical axis of the imaging lens 21 and the optical axis of each lens of the lens array 22 are parallel. Further, as the spatial light modulator 23, a liquid crystal spatial light modulator of an optical address type and a reflective read type, whose reflectance changes according to the amount of incident light, is used.

The feature vector conversion output means 3 is substantially parallel to the laser as the coherent light source 31, the mirror 32 for bending the light beam of the laser 31 in the direction of the system, and the noise of the light beam of the laser 31 after being removed. A substantially parallel coherent light beam 34 generated by a beam expander 33 including a focusing lens 33a for converting into a coherent light beam, a spatial filter 33b, and a collimator lens 33c, and the light beam 34 on the reading surface of the spatial light modulator 23. A beam splitter 35 for making a substantially vertical incidence, and a multiplex image of the target vector which is multiplexed and duplicated by the target vector multiplexing means 2 and written in the spatial light modulator 23, and a read-out of the spatial light modulator 23. The front focal position is aligned with the surface and each optical axis is arranged so that it is parallel to the optical axis of the system. A lens array 22 and the Fourier transform lens array 36 which is composed of the same number of lenses,
The Fourier transform lens array 36 is arranged at the rear focal point in the same number as the Fourier transform lens array 36 so as to correspond one-to-one, and is used to perform a plurality of different filterings of an image which is a multiple duplication target vector. The filter array 37 is composed of a filter array 37 and a detector array 38, which is in the same number as the filter array 37 and is in close contact with or in the vicinity of the rear side so as to correspond one-to-one. In the present embodiment, as the detector array 38, photodetectors made of silicon arranged in an array are used, and thus the extracted feature amount is photoelectrically converted by each detector and has a current value proportional to the intensity. Further, a current-voltage converting / amplifying circuit using an operational amplifier or the like is connected to each of the detectors so that a voltage proportional to the characteristic amount is finally obtained.

With the above configuration, the multiplexed image of the target vector read by the substantially parallel coherent light beam 34 is Fourier transformed by the individual lenses of the Fourier transform lens array 36, and then the individual filters of the filter array 37. Different filtering (feature extraction) is performed in step S1, and the corresponding detector array 38 detects it as a feature amount and finally obtains it as a voltage value.

As each filter constituting this filter array 37, a Gabor filter G is used in this embodiment. The Gabor filter is for extracting a directional spatial frequency structure from an image, and in a broad sense, it is the same as a bandpass spatial frequency filter having a directional property. The description expression in the real space is, for example, G (x, y) = exp [−2π (x ² + y ² ) (u ₀ ² + v ₀ ² ) / σ ² ] × exp [j ² π (u ₀ x + v ₀ y) ] ... a (1), descriptive of the frequency space, G (u, v) = {σ 2 / (2u 0 v 0)} exp [-πσ 2/2 × {(u-u 0 ) ² / u ₀ ² + (v-v ₀ ) ² / v ₀ ² }] (2). In the equations (1) and (2), x and y are coordinate systems in the real space, u and v are coordinate systems in the Fourier transform plane (frequency plane), u ₀ and v ₀ are frequencies of vibration components, and σ Are variables or constants that respectively define Gaussian envelopes. 2 (a) and 2 (b), (1)
The outline of the Gabor filter represented by the equation and the equation (2) is shown. (A) of the figure is G in the real space based on the equation (1).
The abor filter, (b) in the figure is a Gabor filter in the frequency space based on the equation (2). Since this filter contains negative values in real space and only positive values in frequency space, it is better to handle in frequency space in order to process with light intensity that can only handle positive values directly. I understand. Further, it was found through experiments that it is appropriate to extract the characteristics of the spatial frequency in the range of several lines / mm to about 100 lines / mm. Therefore, in this embodiment, FIG.
, U = 0 direction (), v = 0 direction (),
In four directions of u = v direction () and u = -v direction (), the spatial frequency of u ₀ is 4√2 (= 5.7) lines / mm and 8√2 (= 11.3) lines / mm, 16√2 (=
22.6) lines / mm, 32√2 (= 45.3) lines / m
m, 64√2 (= 90.5) lines / mm, 20 kinds of filters were used. Therefore, the arrangement of the lens array or the like is 2 in which the cross-sectional direction of FIG. 1 is 4 and the direction perpendicular to this is 5
It was set to 0 multiplex. The actual filter was produced by outputting the result calculated by the formula (2) by a computer to the printer with a gray value and reducing the photograph. When calculating, σ =
And 2. When the focal length of each lens of the Fourier transform lens array 36 is 50 mm, one side of the input multiplexed image is 4 mm, and the center wavelength of the light flux is 633 nm, the centers of these Gabor filters are 90 μm from the optical axis, 0.179mm, 0.357mm, 0.7
Since it is only necessary to be 16 mm and 1.430 mm, the reduction ratio was determined in consideration of this.

The envelope of each of these filters increases as the spatial frequency of u ₀ increases, and the included frequency range increases. FIG. 4 schematically shows the filter array 37 used in this example. FIG.
An example of the Gabor filter output to the printer when the spatial frequency of u _{0 in} the u = v direction () is 32√2 (= 45.3) lines / mm is shown in (b) as an example.

From the above description, if the above configuration is adopted,
It is obvious that the Gabor filter, which is a kind of band-pass spatial frequency filter with a defined directivity, can instantaneously extract the spatial frequency structure of a plurality of frequencies in a plurality of directions from the image which is the target vector.

In this embodiment, the Xe-Hg lamp 11 is used to obtain the substantially parallel incoherent light beam 15.
Of course, any other incoherent light source can be used, and a substantially parallel light beam can be obtained by a coherent light source such as a laser and then made incoherent by a diffuser plate or the like. In that case, the substantially parallel coherent light flux 3
4 may be commonly used. Substantially parallel coherent light flux 3
Of course, 4 may be generated using another coherent light source such as an LD.

Further, in this embodiment, as can be seen from the sectional view of FIG. 1, homogeneous lenses are used as the lens array 22 and the Fourier transform lens array 36. A planar gradient index lens may be used, or a diffractive lens array may be used. In this case, there is an advantage that alignment becomes easy. Further, an aspherical surface may be used for the curved surface, or the gradient index lens may be provided with a curvature. In this case, the aberration can be easily corrected, and the information error becomes smaller.

Further, the filter array 37 and the detector array 38 are arranged separately as shown in FIG.
The light beam condensed by the lens 39 or the light beam converged to some extent may be detected by the detector array 38,
When the pitches of the filter array 37 and the detector array 38 are different from each other as shown in FIG. 5B, they may be arranged apart from each other, and then the lens 39 may change the magnification to form an image and the detector array 38 may detect the image. Any method may be used as long as the information can be efficiently transmitted to the detector array 38.

Each filter of the filter array 37 has a Gabo on the Fourier transform plane represented by the equation (2).
A band pass spatial frequency filter in which the half value width of the value of the r filter is binarized and the inside is a transparent region and the outside is an opaque region may be used. As an example, FIG. 3C shows an example in which the Gabor filter illustrated in FIG. 3B is binarized. In addition, W
The avelet filter W may be used.

W (a, b, x _n , y _n ) = 1 / (ab) ^1/2 × ∫∫f (x, y) h ^* {(x−x _n ) / a, (y−y _n ) / B} dxdy (3) If the filter array 37 is produced by the filter h (x, y) in which the variables a, b, x _n , and y _n in the equation (3) are appropriately changed, It is clear from the property of a general Wavelet filter that the above effect can be obtained, and that time series signals can be handled.

Further, as a modified example of the target vector input display means 1, the spatial light modulator 17 is a reflection readout type optical address type spatial light modulator 1 similar to the spatial light modulator 23.
FIG. 6 shows an example in which the object O is directly written using the imaging lens I ₁ by using the image forming lens I ₁ . In this example, a substantially parallel incoherent light beam 15 is generated in the same manner as in the above-described embodiment, and the light beam 15 is made to enter the reading surface of the spatial light modulator 17 by the beam splitter B ₁ substantially perpendicularly to display the target vector input. The target vector information written by the means 1 and recorded in the spatial light modulator 17 is read out as reflected light, and thereafter, a series of processing is performed as in the above embodiment. In this case, the speed is increased by directly writing with light without using an electrical address.

[Second Embodiment] As shown in the sectional view of FIG. 7A, the feature vector extraction apparatus of the present embodiment inputs a target vector for inputting at least vector information to be processed into the system. A display means 1, a target vector multiplexing means 2 for multiply duplicating the inputted vector information to be processed, and a plurality of parallel vector information from the multiplexed vector information to be processed by the degree of multiplexing. And a feature vector conversion output means 3 for extracting and converting the feature amount of the above and newly outputting as a feature vector. In this embodiment, a filter that extracts the feature amount in the color space is used as the feature amount.

More specifically, the target vector input display means 1 serves as an incoherent light source 11 for generating an incoherent light beam, a cold lamp often used in a liquid crystal panel display, and a color filter for increasing monochromaticity. Spatial light modulator 17
As in the first embodiment, the liquid crystal spatial light modulator of the electrical address type and the transmissive read type, and this spatial light modulator 17 are used.
Information is input into the system by a photographing device 16 such as a CCD camera for displaying an image which is a target vector, and a controller 18 and a driver 19 as in the first embodiment.

The object vector multiplexing means 2 arranges the image forming lens 21 and the lens array 22 for the image input by the object vector input / display means 1 to determine the number of filters to be extracted. The duplication is performed by the multiplicity of the lens arrays 22 set to be the same.

The feature vector conversion output means 3 is located in the vicinity of the rear focal position of each lens of the lens array 22 (the light beam may enter the aperture of each detector, and may be slightly shifted back and forth). The same number of detector arrays 38 as 22 and one-to-one correspondence with the detector array 38 are arranged on the lens array 22 side of the detector array 38 with the same number of one-to-one correspondence. And a filter array 37 for performing a plurality of different filterings.

In this embodiment, the filter array 37
Is a color filter, which is a multi-layer coating on the surface of a glass substrate so as to have transmission peaks at different wavelengths. In the case of the present embodiment, specifically, as shown in FIG. 7 (b), the multiplicity is set to 16 × 4 × 4, and filters having central peaks of transmission every 10 nm from 500 nm to 650 nm are arrayed. It was arranged and used. Further, as in the first embodiment, each detector of the detector array 38 is provided with a current-voltage converting / amplifying circuit composed of an operational amplifier or the like to output the characteristic amount as a voltage value.

From the above description, it is clear that with the above configuration, the features in the color space can be instantaneously extracted from the image which is the target vector by the plurality of color filters corresponding to the multiplicity.

[Third Embodiment] As shown in the cross-sectional view of FIG. 8A, the feature vector extraction apparatus of this embodiment has at least a target vector for inputting vector information to be processed into the system. The input display means 1, the target vector multiplexing means 2 for multiply duplicating the input vector information to be processed, and the multiplexed vector information to be processed from the vector information to be multiplexed in parallel. The feature vector conversion output means 3 is provided for extracting a plurality of feature quantities and newly converting and outputting them as feature vectors. In this embodiment, a filter that extracts the moment feature is used as the feature amount.

More specifically, the target vector input display means 1 and the target vector multiplexing means 2 are constructed in the same manner as in the second embodiment. The feature vector conversion output means 3 is arranged near the rear focal position of each lens of the lens array 22 in the same number as the lens array 22 so as to correspond one-to-one.
A filter array 37 for applying a plurality of different filterings to an image which is a multiple duplicated target vector
And a detector array 38 which is in close contact with the rear side or is arranged in the vicinity so as to correspond one-to-one with the same number as the filter array 37. In the present embodiment, the filter array 37 is a moment detection filter, and the characteristic of the moment M represented by the following equation is expressed as x ^p y
It is extracted by a filter with a transmittance distribution proportional to ^q .

M _pq = ∫∫f (x, y) x ^p y ^q dxdy (4) In the present embodiment, the multiplicity is set to 12 of 4 × 3, and the filter array 37 is configured as shown in FIG. As shown in (b), (p,
q) is (0,0), (0,1), (1,0), (1,
1), (0,2), (2,0), (1,2), (2,
1), (0, 3), and (3, 0) ten kinds of moment detection filters are manufactured and used by a method of reducing the printer output in the same manner as in the first embodiment (remaining 2). I didn't use them.) Further, as in the first embodiment, each detector of the detector array 38 is provided with a current-voltage converting / amplifying circuit composed of an operational amplifier or the like,
The feature amount is output as a voltage value.

From the above description, if the above configuration is adopted,
It is clear that the features of a plurality of moments can be instantaneously extracted from the image which is the target vector by the plurality of moment detection filters for the multiplicity.

[Fourth Embodiment] As shown in the sectional view of FIG. 9A, the feature vector extraction device of the present embodiment inputs at least a target vector for inputting vector information to be processed into the system. A display means 1, a target vector multiplexing means 2 for multiple duplicating the input vector information to be processed, and a plurality of parallel vector information to be processed from the multiplexed vector information to be processed. And a feature vector conversion output means 3 for extracting and converting the feature amount of the above and newly outputting as a feature vector. In this embodiment, the filter for extracting the spatial frequency structure shown in the first embodiment and the filter for extracting the feature amount in the color space shown in the second embodiment are used at the same time as the feature quantity. .

More specifically, the target vector input display means 1 has the same structure as that of the second embodiment, and the target vector multiplexing means 2 and the feature vector conversion output means 3 have the same structure as that of the first embodiment. , Lens array 2 of spatial light modulator 23
A filter array 37 'for color detection similar to that of the second embodiment is added to the second side. In this embodiment, as shown in FIG. 9B, four types of 2 × 2 filter arrays 37 'for color detection are prepared, and the respective color regions are similar to those in the first embodiment. 4x5 20 types of Gabs
The or filter 37 is applied. Therefore, the multiplicity of the lens array 22, the detector array 38, etc. is 8 × 10 = 80. The image input by the target vector input / display unit 1 is duplicated by the target vector multiplexing unit 2 by the multiplicity of 80, which is 4 × 5 by the color detection filter array 37 ′ added in the system. Each of the 20 regions is reproduced as an image of a different color for each region, and further, from the respective regions by the feature vector conversion output means 3, 4 × 5 20 types of Gabors similar to those in the first embodiment are used. A filter can extract a plurality of spatial frequency structures in a plurality of directions. That is, it is possible to instantly extract a plurality of spatial frequency structures of a plurality of color portions in a plurality of directions.

In this embodiment, the combination of the filter for extracting the spatial frequency structure shown in the first embodiment and the filter for extracting the feature quantity in the color space shown in the second embodiment is shown. Various modifications are conceivable, such as a combination of the moment detection filter shown in the third embodiment and these, or a combination with other similar filters.

[Fifth Embodiment] As shown in the sectional view of FIG. 10, the information processing apparatus according to the present embodiment has at least a target vector input display means 1, a target vector multiplexing means 2, and a feature vector conversion output means 3. And a feature vector extracting device of any of the first to fourth embodiments for inputting a vector to be compared and converting and outputting the feature vector, and the feature vector extracting device outputs the feature vector extracting device. Feature vector input means 4 for inputting a feature vector into the system, reference vector input means 5 for inputting a vector serving as a reference for comparison, and similarity between the input feature vector and a vector serving as a reference for comparison. And a similarity determination means 7 for determining the degree of similarity.

More specifically, the feature vector extraction device shown in the first embodiment is used.

The feature vector input means 4 also transfers a feature quantity vector relating to the spatial frequency of 20 elements of 4 × 5 obtained by the feature vector extraction device as a voltage value from the detector array 38, and a transfer section 41. The threshold value circuit 40 for binarizing this and writing it in the spatial light modulator 61 is composed of a driver 42 and a controller 43.

The reference vector input means 5 inputs a vector (image) serving as a reference for comparison in advance to the feature vector extraction device, and inputs the feature amount relating to the spatial frequency of 4 × 5 20 elements obtained from this feature vector input. 2 via means 4
A transfer unit 54 for digitizing and inputting, a memory 51 for recording this information as a reference vector, and a spatial light modulator 6 for reading information from this memory 51 at the time of identification.
It is composed of a driver 52 and a controller 53 for writing the data in No. 2.

Further, the similarity matching means 6 is a spatial light modulator 61 and a spatial light modulator 6 for displaying the feature quantity vector and the reference vector to be compared two-dimensionally.
2 and an imaging lens 63 installed between the two spatial light modulators 61 and 62 in order to put the two spatial light modulators 61 and 62 into an image forming relationship and to overlap elements of corresponding vectors. , A set of polarizers installed outside the spatial light modulators 61 and 62 for controlling the states of polarization of these spatial light modulators 61 and 62 and calculating negative exclusive OR or exclusive OR 64 and an analyzer 65, light source light for performing a series of processes, and an incoherent light beam 66 which is generated by the method shown in the first embodiment or the like and is incident from the polarizer 64 side, and two vector quantities. Condensing lens 6 for condensing a two-dimensional light beam, which is the calculation result of the negative exclusive OR or the exclusive OR, and converting the degree of similarity into the light intensity.
7 and a detector 68 for detecting the similarity converted into the light intensity.

More specifically, two spatial light modulators 61
Reference numerals 62 and 62 are transmission read-out types of a polarizer-free type made of TN liquid crystal, and used an electrically addressed spatial light modulator. In the TN liquid crystal, a voltage is applied to the portion corresponding to the vector component value 1 to allow the light flux to pass through without changing the polarization state, and no voltage is applied to the portion corresponding to the vector component value 0. The polarization state by 90 °
It has the property of changing and transmitting a light beam.

Therefore, if the polarizer 64 and the analyzer 65 are arranged in parallel outside the spatial light modulators 61 and 62, the NOR operation is calculated. In fact, when the values of the components of both vectors displayed on the spatial light modulators 61 and 62 are 1, both spatial light modulators 61 and 62 are
A voltage is applied to and is transmitted while the polarization direction remains unchanged, so that light is emitted from the analyzer 65. further,
When the value of the component of one vector is 1 and the value of the component of the other vector is 0, a voltage is applied to one spatial light modulator and the polarization direction remains unchanged, while a voltage is applied to the other spatial light modulator. Since the polarization direction is rotated by 90 ° and transmitted without being applied, the polarization direction is rotated by 90 ° in total,
No light is emitted from the analyzer 65. Furthermore, when the value of the component of any vector is 0, both spatial light modulators 6
No voltage is applied to 1 and 62, and both spatial light modulators 6
In 1 and 62, the polarization directions are rotated by 90 ° and transmitted, so that the total polarization direction is transmitted as it is, and light is emitted from the analyzer 65. Using a logic (bright true logic) in which 1 is used when light is emitted from the analyzer 65 and 0 is used when light is not emitted, the following table is obtained, which means that the NOT exclusive OR has just been calculated. . In this negative exclusive OR, since the output 1 is output only in the portion where the vector components match, it means that the degree of coincidence between the two vectors is accurately calculated.

[0054]

Needless to say, the exclusive OR can be realized by arranging the polarizer 64 and the analyzer 65 in crossed Nicols, or by using the opposite logic (dark true logic). In this case, the degree of coincidence in the portion where the output is 0 is high. It is clear that the above components enable the calculation of negative exclusive OR and exclusive OR, and thus the calculation of the degree of similarity can be performed accurately.

In the present apparatus, since the two spatial light modulators 61 and 62 used above have the same display pitch, the image forming relationship between them is made equal. An imaging lens 63 is installed between the two spatial light modulators 61 and 62. Thereby, the spatial light modulator 61,
It is possible to prevent the accuracy of calculation of the similarity from being lowered by the fact that the information input to 62 and displayed two-dimensionally with high density spreads due to the diffraction of light and overlaps with other than corresponding elements. Further, a condensing lens 67 for condensing a two-dimensional light flux which is the result of the calculation of the negative exclusive OR or the exclusive OR and converting the degree of similarity into the light intensity, and the similarity converted to the light intensity. The detector 68 for detecting the degree, specifically, a photodetector made of silicon and an operational amplifier for converting a current value output from this photodetector into a voltage value is two-dimensionally developed and the analyzer 15 The calculation result transmitted from is converted into a scalar value (voltage value) so that the degree of coincidence can be clearly determined at high speed. In the case of negative exclusive OR,
The higher the voltage value, the higher the degree of similarity between the two. on the other hand,
In the case of exclusive OR, conversely, the lower the voltage value, the higher the degree of similarity between the two.

Further, in the case of the negative exclusion logic, the similarity determining means 7 generates a current only when the voltage value from the detector 68 exceeds a preset threshold value.
And the lamp 7 that glows by the current from the threshold circuit 71
And 2. On the other hand, in the case of exclusive OR, on the contrary, the threshold circuit 71 may generate current when the voltage value is less than a certain threshold value.

With the above structure, the information processing apparatus of this embodiment is recognized as follows. The target image input by the target vector input / display unit 1 is converted into a feature vector by the feature vector extraction device, and input / displayed by the feature vector input unit 4 in the spatial light modulator 61 of the similarity matching unit 6. This information is read by the light beam 66 whose polarization direction is aligned by the polarizer 64. The read light flux carrying the information of the target vector has a component corresponding to the other spatial light modulator 62 to which the reference image sent from the reference vector input means 5 is input and displayed by the imaging lens 63. Stacked to match. This superimposed information further passes through an analyzer 65 arranged in parallel with the polarizer 64, so that the exclusive NOR of both images is calculated. This result is condensed by the condenser lens 67, photoelectrically converted by the photodetector 68, and further, the similarity determining means 7 turns on the lamp 72 as a similar object only when the set threshold value is exceeded, You can judge. Further, if different reference images are sequentially input / displayed from the reference vector input means 5, it is possible to recognize a plurality of reference images.

In the present embodiment, the similarity is calculated using the feature vector or reference vector extracted / converted by the above-mentioned feature vector extraction device.
Conventionally, there are problems such as a combination of a matched filter and a correlator that are used when performing similar processing, a problem such as a joint transform correlator that reacts sensitively to deformation and noise and causes an error, and Fourier transform. The problem of lack of a dynamic range on the conversion side has been solved, and it has become possible to provide an information processing apparatus that performs recognition and classification at high speed and with high accuracy.

As a modification of this embodiment, a plurality of similarity determining means 6 of this embodiment are arranged, and different reference vectors are inputted by the reference vector input means 5, and further, the feature vector input means 4 outputs the feature vectors. Information may be multiplexed and input, and processing may be performed in parallel. In this case, recognition is speeded up because the recorded reference vectors are not sequentially switched. For example, in the similarity determination means 6, a multiplexing optical system including a lens array in which a plurality of lenses similar to the target vector multiplexing means 2 are arranged two-dimensionally is used as the imaging lens 63, and the condensing lens 67 is used. The same lens array may be further used as the detector 68, and a detector array in which a plurality of detectors similarly composed of photodetectors and operational amplifiers are arranged in a two-dimensional manner may be used. As a result, the vector input / displayed in the spatial light modulator 61 is multiplexed and duplicated by the multiplicity of the lens array,
Further, the spatial light modulator 6 of the other spatial light modulator
A plurality of reference images input / displayed in 2 are superimposed so that the corresponding components match. The superimposed information is further condensed by the lens array having the multiplicity of the lens array and detected by the detector array.
Since a series of these processes are simultaneously performed in parallel, the similarity between one input vector and a plurality of reference vectors is calculated at high speed, and the relative processing speed is improved.

Further, the similarity determining means 7 may be composed of a computer with an interface for AD-converting and individually fetching each output voltage value from the detector 68, and software for determining the similarity. . When used in combination with the above parallelization, the software for similarity determination compares the voltage values obtained in parallel, and the reference vector showing the highest value is regarded as having a high similarity with the target vector. It may be produced so as to be recognized. In this case, some thresholds may be set and the same, similar, dissimilar, etc. may be further finely classified according to the thresholds. Further, if there is no similar object in the plurality of input / displayed reference vectors, different reference images may be sequentially input / displayed.

[Sixth Embodiment] As shown in FIG. 11, the information processing apparatus of the sixth embodiment uses the reference vector input means 5 of the fifth embodiment to perform the competitive learning by using the competitive learning means.
To the weight vector update display means 5'and the similarity determination means 7
Is an information processing apparatus that functions as a competitive learning type neural network for performing phase-preserving mapping and pattern recognition by changing the winning vector determining means 7 '. In this embodiment, the processing of the similarity matching means 6 is further changed to an inner product calculation processing, and the feature vector input means 4 and the feature vector conversion output means 3 in the feature vector extraction device are used to directly output the output feature vector. It has been changed to transmit in.

More specifically, the detector array 3 of the feature vector conversion output means 3 in the feature vector extraction device.
8, the path between the Fourier transform lens array 36 and the filter array 37 is bent at a substantially right angle by the mirror M ₁ , and the similarity matching is performed by the imaging lens 44 and the beam splitter 45 which are the feature vector input means 4. The information, which has passed through the filter array 37 and converted into the feature vector, is imaged / recorded on the writing surface of the spatial light modulator 61 of the means 6. At this time, as the spatial light modulator 61, an optical address type transmissive read type liquid crystal spatial light modulator whose transmittance changes according to the amount of light to be written is used. In addition, this feature vector is the spatial light modulator 6 in the transmission direction of the beam splitter 45.
Two-dimensional detector 55 set at the position corresponding to 1.
It is detected by, and converted into a voltage value by an operational amplifier or the like and output. This information is sent to the weight vector update display means 5'by the transfer unit 54.

The information written in the feature vector input means 4 is then sent to the competitive learning means. In this embodiment, a self-organizing feature map, which is a type of neural network, is used as this competitive learning means. First, the self-organizing feature map will be briefly described below. Self-organizing feature map (hereinafter S
Notated as OM. 12), as shown in FIG. 12, is composed of a layer ML of element groups arranged two-dimensionally (hereinafter referred to as map layer ML) and an input layer IL for inputting data. Although the map layer ML shows elements arranged two-dimensionally in FIG. 12, elements arranged one-dimensionally may be used. The input layer IL is coupled to all the elements of the map layer ML and can input data to all the elements of the map layer ML. The input data may be either a scalar or a vector, but here it is generally a vector X (n-dimensional). The element i of the map layer ML (i is the order on the map,
The number of all element layers is k. ) Are all weight vectors m _i (n
Dimension). The SOM algorithm determines the weight vector to be updated from the similarity between the input vector X and the weight vector m _{i of} each element <similarity matching>, and brings the weight vector m _i closer to the input vector X <Update>. Then, by repeating the actions of both, the weight vector m _i (1 ≦ i ≦ k) that reflects the distribution of the input vector X is generated. The specific expressions of <similarity matching> and <update> are shown below.

<Similarity Matching> <Update> m _i (t + 1) = m _i (t) + α (t) {X (t) −m _i (t)} iεN _c m _i (t + 1) = m _i (t) Others ... ( 6) where | X−m _i | is the Euclidean distance between X and m _i ,
c is the element with the smallest distance (victory element), N _c
Is the neighborhood of the winning element c in the map layer ML, α (t) is a positive constant, and t is the time. While repeating the update, N
The sizes of _c and α (t) are gradually reduced. Also, α
(T) can be selected so that it becomes smaller as it gets away from the winning element c. The inner product may be used for the similarity matching part as follows.

[0066] From the above, in this device, in the competitive learning process, (7)
It can be seen that the winning element is determined after the inner product similarity matching based on the expression is performed, and the weight vector based on the expression (6) is updated. More specifically, first, the feature vector displayed on the spatial light modulator 61 of the similarity matching means 6 is read by the light beam 66, and further, by the multiplex imaging optical system 63 including the imaging lens 63a and the lens array 63b. , The same number as the number of elements of the neural network are duplicated on the display surface of the spatial light modulator 62 by the multiplicity of the prepared lens array 63b. At this time, in the imaging lens 63a, the read surface of the spatial light modulator 61 is aligned with the front focal position thereof, and in the lens array 63b, the display surface of the spatial light modulator 62 is aligned with the rear focal position thereof. Furthermore, the optical axis of the imaging lens 63a and the optical axis of each lens of the lens array 63b are arranged to be parallel. Further, as the spatial light modulator 62, an electrically addressed and transmissive read type liquid crystal spatial light modulator whose transmittance changes according to an applied voltage is used.

As described above, the feature vector multiplexed by the number of elements of the neural network is then subjected to the spatial light modulator 6 by the signal from the weight vector update display means 5 '.
The information of the weight vector displayed in 2 is superposed on each other, and is further condensed by the lens array 67 'of the same number and the same array as the lens array 63b, and the inner product of each element is calculated. Similarly, the result is detected by the detector array 68 'having the same number and the same arrangement as the lens array 63b, further converted into a voltage value by an operational amplifier or the like, and sent to the victory vector determining means 7'.

Next, the process of calculating the inner product operation with the above configuration will be described with reference to FIG. The inner product calculation part of the equation (7) expands the input vector into a two-dimensional array having the number of elements N × N, and the weight vector has M × M sub-matrixes having the number of elements N × N (corresponding to the number of elements of the neural network) If it is expanded into an array, it can be rewritten as follows for each component.

[0069] To simplify the explanation, if N = 3 and M = 2, the portion of the similarity matching means 6 can be schematically written as shown in FIG. That is, the feature vector of 3 × 3 elements, which is input to the spatial light modulator 61, is generated by the multiplex imaging optical system 63 including the imaging lens 63a and the lens array 63b.
2 × 2 of the multiplicity of the lens array 63b is duplicated on the display surface of the spatial light modulator 62 ((A) of the same figure), and the information of the weight vector displayed on the spatial light modulator 62 ((FIG. B)) are respectively superposed, and the lens array 6
At 7 ', the light is individually collected and the inner product of each element is calculated. Specifically, when the processing result of each element is shown for the y ₂₁ component, the overlapping portion is multiplied and the condensing portion is added, Therefore, the inner product of the elements corresponding to the (2,1) component is calculated. The same can be calculated for other components, so in total, Therefore, it can be seen that the inner product of equation (8) N = 3 and M = 2 is certainly calculated. In the above optical system, the case of N = 3 and M = 2 has been described for the sake of simplicity. However, in the above optical system, the lens array, the arrangement of input vectors, the arrangement of weight vectors, etc. can be arbitrarily selected. It is clear that the equation is satisfied.

Next, the winning vector determining means 7'determines the element with the highest output as the winning element from the inner product calculation for each element output from the detector array 68 'of the similarity matching means 6. , And sends this information to the weight vector update display means 5 '. The weight vector update display means 5'updates the weight vector according to the update rule of the above equation (6). In this apparatus, the weight vector update display means 5'and the winning vector hand determination means 7'for competing learning are realized on the same computer.
Among them, the winning vector determining means 7'compares the result with the transfer section for AD-converting and individually fetching the individual output voltages from the detector array 68 'of the similarity matching means 6 and showing the highest voltage. It is made up of software that selects the ones as victory elements. The weight vector update display means 5'is a transfer unit 54 for AD-converting and fetching the feature vector information from the newly added two-dimensional detector 55, and the victory element information determined by the victory vector determination means 7 '. And software for updating the weight vector according to the update rule of equation (6) using the information of the feature vector from the two-dimensional detector 55, and the updated information of the weight vector is input to the spatial light modulator 62.
It is composed of a memory 51 for displaying, a controller 52 and a variber 53 (FIG. 10). In this case, it is more preferable to use the equation (6) after normalizing the feature vector. Further, in this competitive learning process, if the target vector is repeatedly input and the weight vector is sequentially updated, the weight vector becomes stable at an appropriate place, so learning may be stopped at that point.

In the recognition / classification process after learning, the vector to be identified is input with the weight vector fixed.
The input vector may be classified according to which of the detector arrays 68 'is the winning element. For example, FIG.
If the multiplicity is 6 × 6 and the number of elements is 36 as described above, each element may be classified into 36 types, or 30 and 6
It may be divided into two regions and classified into two types. It can be freely set by software according to the target. The weight vector is initially set to display a random data sequence.

In the present embodiment, as described above, since the feature vector extracted and converted by the feature vector extraction device as described above can be used to calculate the similarity and determine the winning element, It has become possible to provide a competitive learning type information processing device that performs recognition and classification at high speed and with extremely high accuracy.

[Seventh Embodiment] The information processing apparatus of this embodiment is also an information processing apparatus functioning as a competitive learning type neural network for performing phase preserving mapping and pattern recognition, as in the previous embodiment. The difference from the previous embodiment is that, as shown in FIG. 14, the feature vector input means 4 is an electric system instead of an optical system, and the spatial light modulator 61 of the similarity matching means 6 is an LED array. , And others are all configured the same and operate in the same manner. It should be noted that, along with the change of the feature vector input means 4, the transfer of the feature vector information to the weight vector update display means 5'is performed directly by the detector array 3
It has been changed to a transfer unit 54 that is sent from S8.

Specifically, the feature vector input means 4 inputs each output from the detector array 38 in parallel, and an interface for making the same number of LED arrays 61 as the detector array 38 emit light with a brightness proportional to this output. 41 and a driver-46.

In this embodiment, an LED is used instead of the liquid crystal spatial light modulator whose speed is limited to some extent in the state of the art, and since information is transmitted in parallel, a comparison with the sixth embodiment is made. The processing speed is further improved.
Further, in the present apparatus, in the competitive learning, the feature vector extracted and converted in advance is recorded in the feature vector input means 4 even if the target vector used for learning is not repeatedly input into the feature vector extraction device each time it is updated. It is also possible to provide a memory unit to store the data and read it from here. Further, if the characteristic vector information from the detector array 38 is normalized and is used in the weight vector update display means 5'and the similarity matching means 6, the matching accuracy is further improved.

[Eighth Embodiment] Like the previous embodiment, the information processing apparatus of the present embodiment is also an information processing apparatus that functions as a competitive learning type neural network for performing phase-preserving mapping and pattern recognition. The difference from the previous embodiment is that, as shown in FIG. 15, the feature vector extraction device uses the device of the fourth embodiment, and the inner product process of the similarity matching means 6 is the same as the negative exclusive OR as in the fifth embodiment. The point is that the degree of similarity is calculated instead of exclusive OR. Along with this, as the spatial light modulators 61 and 62 of the similarity matching means 6, as in the fifth embodiment, both of the TN liquid crystal transmission type without a polarizer and the electrical address type spatial light modulator are used. Further, a set of a polarizer 64 and an analyzer 65 placed outside of these spatial light modulators was arranged. With these configurations, the similarity matching unit 6 calculates a negative exclusive logical sum or an exclusive logical sum for each element, and calculates a more accurate similarity. Further, the feature vector input means 4 transfers the feature vector relating to the spatial frequency in each color space obtained by the feature vector extraction device as a voltage value from the detector array 38.
And binarize this information, and then convert this information to the spatial light modulator 61.
The threshold value circuit 40 for writing into the memory, the driver 42, the controller 43, and the transfer unit 54 for transferring the information of the feature amount vector to the weight vector updating means 5 '.

With the above configuration, in the present embodiment, the feature vector obtained by simultaneously extracting and converting the feature vectors of different categories such as color and spatial frequency by the above feature vector extracting device is used. Since it becomes possible to determine the winning element by calculating the degree of similarity using negative exclusive OR or exclusive OR, it is possible to provide a competitive learning type information processing device that recognizes and classifies more accurately at high speed. It became so. Further, in this apparatus, in the competitive learning, the feature vector extracted and converted in advance is recorded in the feature vector input means 4 without inputting the target vector used for learning into the feature vector extracting device repeatedly every time updating. It is also possible to provide a memory unit to store the data and read it from here.

[Ninth Embodiment] The information processing apparatus of this embodiment is also an information processing apparatus that functions as a competitive learning type neural network for performing phase-preserving mapping and pattern recognition, as in the previous embodiment. The difference from the previous embodiment is that, as shown in FIG. 16, the feature vector extraction device uses the device of the second embodiment to perform feature extraction in the color space, and further, the similarity matching means 6 calculates the degree of similarity. Is performed based on the interference action.

In the first half, the similarity matching means 6 first duplicates the feature vector input to the spatial light modulator 61 with the multiplicity of the number of elements, as in the previous embodiment. This is because the feature vector input to the spatial light modulator 61 is read out by the light beam 66, and is further copied on the writing surface of the spatial light modulator 61 'by the multiplex imaging optical system 63 including the imaging lens 63a and the lens array 63b. It is what makes me. At this time, the lens array 63b is arranged at the rear focal position of the spatial light modulator 6 so that the readout surface of the spatial light modulator 61 is aligned with the front focal position of the imaging lens 63a.
Further, the writing surfaces of 1'are aligned so that the optical axis of the imaging lens 63a and the optical axis of each lens of the lens array 63b are parallel to each other. In addition, the spatial light modulator 6
As 1 ', a liquid crystal spatial light modulator of an optical address type and a reflective read type, which modulates the phase of the read light according to the written light quantity, is used. On the other hand, in the intermediate portion, the spatial light modulator 61 'and the spatial light modulator 62 for displaying the weight vector (in the case of the present embodiment, an electrical address type reflection for modulating the phase of the read light by the applied voltage amount) are used. A read-out type liquid crystal spatial light modulator is used), the beam splitter 69, and the diffuser 60 are arranged as shown in the figure to form an interferometer. In this middle part,
The coherent light beam 66 'for reading enters the beam splitter 69 and is split into two light beams. Among them, one light beam is incident on the spatial light modulator 61 ′, from which the information of the feature vector duplicated with the multiplicity corresponding to the number of elements is read out after the information converted into the phase is reflected, and the beam splitter 69. Then, the other light flux enters the spatial light modulator 62, reads out the information in which the weight vector information is converted into the phase information, is reflected, and is reflected by the beam splitter 69.
Return to These two light beams are further combined by the beam splitter 6
They are combined at 9 and interfere with each other at the diffuser 60.
At this time, if the difference between the distance from the spatial light modulator 61 ′ to the diffuser 60 and the distance from the spatial light modulator 62 to the diffuser 60 is adjusted to be an integral multiple of the wavelength, the feature vector and the weight vector due to the interference action. Brightness and darkness are formed on the diffuser 60 according to the degree of coincidence. That is, the higher the degree of coincidence between the two, the brighter it becomes. If this information is collected by the lens array 67 'having the same arrangement as the lens array 63b, the degree of coincidence is calculated for each element of the neural network. Similar to the previous embodiment, this information is sent to the succeeding victory vector determining means 7'and the weight vector updating display means 5 ', and the same processing is performed.

With the above-mentioned configuration, also in this embodiment, the feature vector in the color space extracted / converted by the above-described feature vector extraction device is used, and the similarity is calculated by utilizing the interference action to obtain the winning element. Since it has become possible to determine, it has become possible to provide a competitive learning type information processing device that performs recognition and classification at high speed and with high accuracy.

In the above embodiments, a readily available liquid crystal device was used as the spatial light modulator, but other devices such as those utilizing the electro-optical effect or magneto-optical effect, organic compound compounds, etc. may be used. Of course good. As described above, if the lens array or the like is formed by integrally molding a homogeneous lens, a diffractive optical lens array, a gradient index lens array, or the like, alignment becomes easier and collation / recognition accuracy is improved.

Although the input vector is an image in the similarity calculating device and the information processing device in the above embodiments,
The present invention is not limited to this, and any signal or data obtained by another sensor such as a microphone, a concentration sensor, a flow rate sensor, or data calculated in a computer may be used.

Further, in the above embodiment, the one using negative exclusive OR or exclusive OR for similarity degree calculation basically binarizes information, but wants to handle it as multivalued. In such a case, coding may be performed by a method such as spatial coding and the same calculation may be performed. Further, the feature vector extraction device used in the fifth to ninth embodiments is not particularly unique to these embodiments, and any of them may be used, or a combination of them may be used. It is clear that the same effect can be obtained by using the feature amount of.

The above-described feature vector extraction device of the present invention and the information processing device using the same can be configured as follows, for example. [1] At least a target vector input display means for inputting vector information to be processed into the system, a target vector multiplexing means for multiple duplicating the input vector information to be processed, and The feature vector conversion output means for extracting a plurality of feature amounts in parallel from the multiplexed vector information to be processed by the multiplicity and newly converting and outputting the feature amount as a feature vector. A feature vector extraction device.

[2] The above [1], wherein the feature vector conversion output means includes a plurality of spatial frequency filters for multiplicity and extracts a plurality of spatial frequency structures of vector information to be processed. Feature vector extraction device.

[3] The feature vector extraction device according to [2], wherein the spatial frequency filter is a Gabor filter.

[4] The feature vector extraction device according to [2], wherein the spatial frequency filter is a directional bandpass spatial frequency filter.

[5] The feature vector extraction device according to [2], wherein the spatial frequency filter is a wavelet filter.

[6] The feature vector conversion output means includes a plurality of color filters for multiplicity, and the feature in the color space of the vector information to be processed is extracted. Feature vector extraction device.

[7] A plurality of moment detection filters for multiplicity are included in the feature vector conversion output means,
The feature vector extraction device according to the above [1], wherein the feature of a plurality of moments of vector information to be processed is extracted.

[8] At least a feature vector extraction device for inputting vector information to be compared and converting and outputting it to a feature vector, and a feature vector extraction device for displaying the feature vector output from the feature vector extraction device in the system. A feature vector input means, a reference vector input means for inputting a reference vector for comparison, a similarity matching means for calculating the similarity between the input feature vector and a reference vector for comparison, In the information processing apparatus including a similarity determination unit for determining the degree of sex, the above [1] is used as the feature vector extraction apparatus.
[7] An information processing apparatus comprising the feature vector extraction apparatus according to any one of [7] to [7].

[0092]

[9] A feature vector extraction device for inputting target vector information and converting and outputting to a feature vector, and a feature vector display means for displaying the feature vector output from the feature vector extraction device in the system. In an information processing apparatus having a competitive learning means for recognition / classification, the competitive learning means has a similarity matching means, a weight vector update display means, and a winning vector determination means, and further, the feature vector An information processing device comprising the feature vector extraction device according to any one of [1] to [7] as an extraction device.

[0093]

As described above, the feature vector extraction device of the present invention is provided with at least the target vector input display means for inputting the vector information to be processed into the system and the input vector information to be processed. Object vector multiplexing means for multiple duplication, and a plurality of feature quantities are extracted in parallel from the multiplexed vector information to be processed by the degree of multiplicity and converted as a new feature vector.
Since it is composed of the feature vector conversion output means for outputting, the feature quantity useful for the subsequent classification and recognition processing can be extracted at high speed.

Further, since the information processing apparatus of the present invention determines the similarity by using the feature vector extracted by such a feature vector extracting apparatus, the recognition and the classification can be performed with high performance. I can do it now.

[Brief description of the drawings]

FIG. 1 is a sectional view showing a schematic configuration of a feature vector extraction device according to a first embodiment of the present invention.

FIG. 2 is a diagram showing an outline of a Gabor filter.

FIG. 3 is a diagram showing a specific example of a filter used in the first embodiment.

FIG. 4 is a diagram schematically showing a filter array used in the first embodiment.

FIG. 5 is a sectional view of an essential part of a modified example of the first embodiment.

FIG. 6 is a sectional view showing the outline of another modification of the first embodiment.

FIG. 7 is a cross-sectional view showing a schematic configuration of a feature vector extraction device of a second embodiment and a view schematically showing a filter array.

FIG. 8 is a cross-sectional view showing a schematic configuration of a feature vector extraction device of a third embodiment and a diagram schematically showing a filter array.

FIG. 9 is a sectional view showing a schematic configuration of a feature vector extraction device of a fourth embodiment and a diagram schematically showing a filter array.

FIG. 10 is a sectional view showing a schematic configuration of an information processing apparatus according to a fifth embodiment.

FIG. 11 is a sectional view showing a schematic configuration of an information processing apparatus according to a sixth embodiment.

FIG. 12 is a diagram for explaining the outline of a self-organizing feature map.

FIG. 13 is a diagram for explaining the inner product calculation used in the sixth embodiment.

FIG. 14 is a sectional view showing a schematic configuration of an information processing apparatus of a seventh embodiment.

FIG. 15 is a sectional view showing a schematic configuration of an information processing apparatus of an eighth embodiment.

FIG. 16 is a sectional view showing a schematic configuration of an information processing apparatus according to a ninth embodiment.

FIG. 17 is a diagram for explaining a first conventional example.

FIG. 18 is a diagram for explaining a second conventional example.

[Explanation of symbols]

O ... Object I ₁ ... Imaging lens B ₁ ... Beam splitter M ₁ ... Mirror ML ... Map layer 1 L ... Input layer 1 ... Target vector input display means 2 ... Target vector multiplexing means 3 ... Feature vector conversion output means 4 ... Features Vector input means 5 ... Reference vector input means 5 '... Weight vector update display means 6 ... Similarity matching means 7 ... Similarity determination means 7' ... Victory vector determination means 11 ... Incoherent light source 12a ... Lens 12b ... Mirror 13 ... Wavelength Filter 14 ... Mirror 15 ... Incoherent light flux 16 ... Image sensor 17 ... Spatial light modulator 18 ... Controller 19 ... Driver 21 ... Imaging lens 22 ... Lens array 23 ... Spatial light modulator 31 ... Coherent light source 32 ... Mirror 33 ... Beam Expander 33a ... Focusing lens 33b ... Spatial Filter 33c ... Collimator lens 34 ... Coherent light flux 35 ... Beam splitter 36 ... Fourier transform lens array 37 ... Filter array 37 '... Filter array 38 ... Detector array 40 ... Threshold circuit 41 ... Transfer unit (interface) 42 ... Driver 43 ... Controller 44 Imaging lens 45 ... Beam splitter 46 ... Driver 51 ... Memory 52 ... Driver 53 ... Controller 54 ... Transfer unit 55 ... Two-dimensional detector 60 ... Diffuser 61 ... Spatial light modulator (LED array) 61 '... Spatial light modulator 62 Spatial light modulator 63 ... Imaging lens (multiple imaging optical system) 63a ... Imaging lens 63b ... Lens array 64 ... Polarizer 65 ... Analyzer 66 ... Incoherent light beam 66 '... Coherent light beam 67 ... Condensing Lens 67 '... lens array 68 ... detector 68' ... Detector array 69 ... beam splitter 71 ... threshold circuit 72 ... lamp

Claims (3)

[Claims] 1. At least a target vector input display means for inputting vector information to be processed into the system, and a target vector multiplexing means for multiple duplicating the input vector information to be processed. , Feature vector conversion output means for extracting a plurality of feature amounts in parallel from the multiplexed vector information to be processed and newly converting / outputting as a feature vector. A feature vector extraction device characterized by the above. 2. A feature vector extraction device for inputting at least vector information to be compared and converting and outputting to a feature vector, and a feature for displaying the feature vector output from this feature vector extraction device in the system. Vector input means, reference vector input means for inputting a vector to be a reference for comparison, similarity matching means for calculating similarity between the input feature vector and a vector to be a reference for comparison, similarity An information processing apparatus configured by a similarity determination unit for determining the height of the information processing apparatus, wherein the characteristic vector extracting apparatus according to claim 1 is used as the characteristic vector extracting apparatus. 3. A feature vector extraction device for inputting target vector information and converting and outputting to a feature vector, and a feature vector display means for displaying the feature vector output from this feature vector extraction device in the system. When,
In an information processing apparatus having a competitive learning means for recognition / classification, the competitive learning means has a similarity matching means, a weight vector update display means, and a winning vector determination means, and further, the feature vector An information processing device comprising the feature vector extracting device according to claim 1 as an extracting device.