JP3279605B2 - Pattern identification method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for identifying an object, an image, a character, and the like, and for identifying a pattern used for information retrieval, inference, association, and the like.

[0002]

2. Description of the Related Art Conventionally, a neural network, a k-nearest neighbor method, and the like have been known as methods for identifying a pattern in which a Neumann-type computer is not good at processing images and sounds.

[0003] In the neural network, a number of recognition and classification methods have been proposed using an error back propagation learning method (hereinafter abbreviated as BP method), which is one of the learning methods.
FIG. 2 is a configuration diagram showing a configuration of the conventional neural network as described above. In the neural network for learning by BP method, usually, a layer structure set of three or more layers of units in the network, among them, the input layer a layer for performing pattern input, an output layer and a layer for outputting a recognition output signal , And the rest are referred to as intermediate layers or hidden layers. At this time, in principle, the units included in the same layer
It is assumed that there is no coupling and that there is no signal feedback from the layer near the output to the layer near the input.

[0004] Binding between the layers of the unit, it is indicated by arrows, V _ij which gave a corresponding unit number binding load value the unit and the intermediate layer unit of the input layer as a subscript
(I intermediate layer unit number, j is an is an input layer unit number), representing Ke with the coupling load between units of the intermediate layer and the units of the output layer as a subscript corresponding unit number W
_ij (i is an output layer unit number and j is an intermediate layer unit number). Output from each unit of the input layer output values I _n (n is a unit number), Medium
Propagates into the units during layer, further, the output values output from the respective units of the intermediate layer C _n (n is the unit number) is
Propagates to each unit in the output layer, the output value O _n from each unit of the output layer (n is the unit number) is obtained. (The output value of each unit is one for one unit.)

Generally, the input / output characteristics of the intermediate layer unit and the output layer unit are expressed by the following equations. C _i = f (ΣV _ij I _j + ξ _i ) (1) O _i = f (ΣW _ij C _j + θ _i ) (2) f (x) = 1 / (1 + exp (−x)) (3) i and j are the corresponding unit numbers as described above, and ξ and θ are the bias values in each unit. The function shown in (3) is called a sigmoid function.

In the above-described neural network configuration, an output from the output layer when an arbitrary input pattern, that is, an output vector (I ₁ , I ₂ , I ₃ , I ₄ , I ₅ ) of the input layer unit is given. O _i (where i = 1, 2,...
(5) By correcting the value of the coupling weight between the units so that the sum of squares of the error of each unit between the desired output (teacher signal) and the desired output (teacher signal) is reduced, a desired input pattern can be obtained for an arbitrary input pattern. A neural network that outputs an output pattern can be obtained. The BP method is a method of calculating a load correction coefficient at this time, and is also called a generalized δ rule.

However, in the above-described neural network, a number of typical patterns belonging to a class to be recognized are presented for learning, and even if an attempt is made to determine a connection weight value, a large number of patterns are required to converge the learning. Since iterative learning is performed twice, it takes time to construct a network, and the learning does not always converge. The learning often falls into a so-called local minimum, and a sufficient recognition result is often not given. In addition, since there is no free area in the feature space, there is a disadvantage that an unlearned input pattern is identified as one of the learned input patterns.

On the other hand, in the k-nearest neighbor method, an identification area is created in an input vector space representing the feature of the pattern, and the pattern is identified using this area. First , some easily distinguishable features of a pattern group to be identified are examined and quantified. The quantities representing those quantified features are displayed on one position vector. This vector is called the input vector.

For example, an input vector of a pattern C _Si obtained by extracting n features belonging to the identification class S is expressed as C _Si = (x ₁ , x ₂ ... X _n ) (4) where x _i represents the respective feature amounts. In this way, all the input vectors of the pattern group are created.

Next, an input vector belonging to the identification class s
C_SiAnd another k-th closest distance from that vector
Input vector C belonging to identification class T_TjThe midpoint of
Divide by a one-dimensional hyperplane. This is all input vectors
The area enclosed by the hyperplane for each class is
It becomes the identification area of the class. K must be other than 1
Thus, an overlapping area can be created. that's all
By then, preparation (learning) for identification is completed.FIG.
Is the input vector of n = 2 created by the above learning.
It is a conceptual diagram which shows an example of the identification area of space.

Thereafter, when a learned input pattern is given, the input pattern is identified in the class of the identification area containing the input vector. However, in this method, there is no problem that the learning does not converge or the learning takes time as in the above-described neural network, but the entire input vector space is taken into the discrimination area of an arbitrary class. Therefore, even if a different pattern is input from the input pattern that has been clearly learned, the same drawback as that of the neural network that the input pattern is always identified as one of the learned patterns. It has remained.

[0012]

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and the learning can be surely converged, the number of times of learning is only one, and the entire input vector space can be used. An object of the present invention is to provide a pattern identification method for identifying an unknown pattern that has not been learned even if an unknown pattern that has not been learned is input because an empty area is left without being filled with an identification area.

[0013]

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned technical problem, and has been developed in an n-dimensional input vector space having n data representing the characteristics of an input pattern as components. In the pattern identification method for identifying a pattern based on a positional relationship between the identification area and an input vector, the identification area is provided for each identification class to be identified . versus
The identification area is created by inputting an arbitrary input vector C _Si belonging to the first identification class and an input vector belonging to another identification class located at a k-th distance from the vector C _Si. Select the vector C _Ti and select the vector C
The _Si -C _Ti by calculating the distance between the vector C _Si point which internally divides at an arbitrary ratio, in the distance is the radius, the input vector C _Si region is in a hypersphere of n dimensions is the central R _i For each of the input vectors of the first identification class, a hypersphere region R _pi is created, the logical sum of the hypersphere regions R _{i is set} as the identification region of the first identification class, and the above operation is performed for each identification. There is provided a pattern identification method characterized by being performed for each desired class.

An identification area for identifying a pattern is created in an n-dimensional input vector space having n data representing the features of the input pattern as components, and the positional relationship between the identification area and the input vector is determined. based, in pattern identification method of identifying a pattern, the average of the vector of the input vectors belonging to one identification class is calculated, the vector of the average and C _S, in a close distance to the k-th from the vector C _S Vector C _T of the averages of different identification classes
Is calculated, and the distance between the vector C _S and a point that internally divides the vector C _S -C _T at an arbitrary ratio is calculated, and the distance is a radius, and the n-dimensional super is the center of the average vector C _S. ball
The above operation is performed for all the patterns to be identified as an identification area of the first identification class, and a similar identification area is created for each pattern.
To provide Rupa turn identification method.

A method in which the value of k is set to 1 is preferable. It is also preferable that the ratio of the internal dividing points is 1: 1. In addition, it is preferable that the data representing the characteristics of the input pattern is an output of a correlation operation between the input pattern and the reference pattern. The correlation calculation is preferably performed optically. It is preferable that the optical correlation operation uses means based on the similarity of the pattern on the optical Fourier transform plane.

The optical correlation calculating means based on the similarity of the pattern on the optical Fourier transform plane includes:
Modulate the complex amplitude distribution of the first light flux according to at least the coherent light source and the input pattern and the reference pattern (ie, display the input pattern and the reference pattern)
A first spatial light modulator; a first Fourier transform lens for optically Fourier transforming a complex amplitude distribution of the first light beam on the first spatial light modulator; A second spatial light modulator that modulates a complex amplitude distribution of the second light beam according to a Fourier-transformed light amount distribution of the light beam, and a complex amplitude distribution of the second light beam on the second spatial light modulator An optical correlation detection device comprising a second Fourier transform lens for optically Fourier-transforming the light beam and a light-receiving element for receiving a Fourier-transformed light amount distribution of the second light flux, wherein the first space On the optical modulator, a plurality of reference patterns for performing correlation detection are displayed simultaneously and in parallel, and near each of the reference patterns, an input pattern is displayed one by one for each reference pattern simultaneously and in parallel. The first Fourier transform lens is An array of individual lenses corresponding to each reference pattern and input pattern pair, and the second Fourier transform lens is provided on a second spatial light modulator generated by the first Fourier transform lens array. It is preferable that the lens array be an individual lens array corresponding to each Fourier transform pattern.

[0017]

In the above pattern recognition, assuming that n data are obtained from the input pattern, one of the input vectors belonging to the identification class S is represented by C _Si = (x ₁ , x ₂ ,..., X _n ) ( 4), and the vector C _Si is a vector belonging to a different identification class, and a vector having the k-th closest distance belongs to the identification class T, and C _Tj = (x ₁ ′, x ₂ ′) ,..., X _n ′) (5) If the two vectors C _Si and C _Tj are
The distance L from the input vector C _{Si of the} point that internally _divides C _Si −C _Tj into a: b is L = {(x ₁ −x ₁ ′) ² + (x ₂ −x ₂ ′) ² +. + ( _Xn − _xn ′) ² } ^1/2 × a / (a + b) (6)

[0018] Here, hypersphere area type having a radius L centering on the input vector C _Si is, (X ₁ over x _{1) ²} + (X ² over ^{_{x 2) 2 + ·· + (}} X n over x _n ) ² ≦ L ² (7) Here, X ₁ , X ₂ ,... X _n are variables of each component of the input vector space. For the other input vectors to which the identification class S belongs, the region equation (7) of the hypersphere is calculated between a vector belonging to each of the different identification classes and a vector having the k-th closest distance. The logical sum of the hypersphere region expression as shown in Expression (7) is taken, and the identification region of the identification class S is determined as the identification region condition of the identification class S. The above determination of the identification area is performed for all the classes to be identified, and the learning of the input pattern is completed. After learning, an input pattern to be identified is given, and if the input vector of the input pattern satisfies the identification area condition of the identification class S, the input pattern is identified as the identification class S. If the above-mentioned identification area conditions of all the identification classes are not satisfied, the input pattern is identified as an unknown input pattern.

In the creation of the identification area, when input vectors are relatively fixed for each class, for example,
By calculating the average vector of the input vectors belonging to the class S and the class T and setting them as the vectors C _Si and C _Tj , the above operation is performed. The identification area can be created in time, and the pattern can be identified. In the above operation, assuming that k = 1, a hypersphere is created with the vector closest to the vector C _S1 and the internal division ratio is set to a: b =
By making the ratio 1: 1 or the like, the overlapping of the identification areas can be reduced and the identification areas can be created.

The correlation output, which is the output of data representing the characteristics of the input pattern by the correlation operation between the input pattern and the reference pattern, indicates the similarity between the input pattern and the reference pattern, and indicates the similarity between the input pattern and the reference pattern. The similarity differs for each input pattern, and the same reference pattern does not change the similarity with respect to one input pattern. In addition, by increasing the number of reference patterns, the number of dimensions of the input pattern space also increases. Since identification is facilitated, by using the correlation output as an input vector component of the pattern identification method of the present invention, it is possible to easily create an identification area and identify a pattern.

The output of the above-described correlation operation can be obtained as a component of an input vector by obtaining a correlation output by an optical correlation operation. In addition, by using a joint Fourier transform method (JT method) or a matched filter method, which is a correlation calculation method based on the similarity of patterns on the Fourier transform plane, a correlation output can be used as a component of an input vector.

Furthermore, when the correlation detection based on the JT method is performed simultaneously and in parallel on a plurality of reference patterns using a spatial light modulator and a lens array, an input vector can be instantaneously created.

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

[0024]

Embodiment 1 In order to identify a pattern, first, data representing the characteristics of the pattern to be identified, such as images, characters, and sounds, is obtained. In this embodiment,
Two pieces of data are obtained from the input pattern, and this data is input to a computer, and the subsequent processing is performed on the computer. 1A and 1B are conceptual diagrams showing an input vector space in the present embodiment. Two input data 1 and 2 representing the characteristics of the input pattern become the dimensions of the space. As shown in FIG. 1A, input patterns belonging to the identification class A include input vectors C _A1 11 to C _A3 13,
The input pattern belonging to the identification class B is the input vector C
_{B1 14~C B3} 16, input patterns belonging to identify class C, as an input vector C _C1 17~C _C3 18, are plotted on a two-dimensional input vector space.

Next, learning of the recognition area of the identification class A is started. An input vector C _A1 11 identification Class A, the highest distance input vector other than the input vector C _A1 11 identification class A search close vector, it can be seen that C _B1 14 hits thereto. A half distance (distance between the midpoint of two vectors and C _A1 ) of the distance between the input vector C _A1 11 and the input vector C _B1 14 is calculated. A region R _A1 20 is created by a hypersphere (here, a circle) which is a vector C _A1 11. The same operation is performed on the vectors C _A2 12 and C _A3 13 belonging to the discrimination class A to form hypersphere regions R _A2 21 and R _A3 22. The logical sum of these hyperspheres is shown in FIG. The identification area _RA 24 shown in FIG. Identification Class B, also creation of C region, by the like operation, the identified region R _B 25, to learn the identification area R _C 26.

After the learning of the identification area is completed, pattern identification is performed. The input data 1 and 2 are obtained from the pattern to be identified, and the input vector of the input pattern to be identified is obtained in the input vector space. If the input vector of the pattern to be identified falls within the area of the identification area A, B or C created by the learning as described above, the pattern to be identified having the input vector to be examined is determined by the area where the vector exists. Are classified into classes. Figure
In 1b , if the input vector 31 does not belong to any class , such as the test input vector 32, which is identified as class B, the pattern of the test input vector 32 is a pattern that has not been learned. Identified to be.

[0027]

FIG. 4 shows an optical system for reproducing a known matched filter. The matched filter 46 converts a pattern obtained by optically Fourier transforming a plurality of reference patterns in advance,
A multiplexed recording is used while changing the irradiation direction of the plane wave for each reference pattern.

As shown in FIG. 4 , a light beam 42 emitted from a laser 41 is spread by a beam expander 43 into an appropriate light beam system, and is incident on a spatial light modulator 44 displaying an input pattern. Then, the light beam 42 subjected to the spatial modulation corresponding to the input pattern is converted to a Fourier transform lens 45.
And enters the matched filter 46 placed on the Fourier transform plane. The matched filter 46 is created based on the Fourier transform hologram of the reference pattern. If the correlation between the Fourier transform pattern of the reference pattern and the Fourier transform pattern of the input pattern during hologram recording is high, the hologram recording is performed. A light beam conjugate to the reference light beam is generated in the direction of the reference light beam at the time. If the reference light beam is a parallel light beam, a light beam that has been optically Fourier transformed again becomes a correlation pattern between the input pattern and the reference pattern. Therefore, the light beam 42 emitted from the matched filter 46 is converted to a Fourier transform lens 47.
And enters the photoelectric conversion element array 48, and this correlation pattern is detected.

Here, the position of the center of the correlation pattern is determined in advance by the direction of the reference light beam at the time of making the matching. Therefore, it is sufficient to arrange the photoelectric conversion element array in this portion. Further, the detection area of the correlation pattern can be adjusted by the area of the photoelectric conversion element,
If this area is reduced, a calculation result close to a normal inner product calculation is obtained, and if this area is increased, a calculation result is less likely to be affected by pattern deformation or shift.

The above photoelectric conversion element A obtained as the
Using the output of the ray as each component of the input vector, pattern identification can be performed by the operation shown in the first embodiment. At this time, if the direction of the reference light beam is changed and recorded in one matched filter of a plurality of reference patterns, those correlation patterns appear at different places on the intensity conversion element array 48, and the input vector Can be taken out simultaneously.

In addition, a joint Fourier transform (JT method) is used in an optical correlation operation method based on the similarity of a pattern on an optical Fourier transform surface.
There is. This method has a feature that it can be easily implemented using a spatial light modulator having the current specifications, and can be used in the pattern identification method of the present invention, like the matched filter method. However, at the same time, the ability to obtain a correlation output with a plurality of reference patterns is inferior to the matched filter method. Therefore, the present inventors have proposed an optical multiple correlation detection optical system in which correlation optical systems based on the JT method are arrayed (Japanese Patent Application No. 3-33423). The case where the pattern identification method described above is applied will be further described in the following embodiment.

[0032]

Embodiment 3 FIG. 5 is an explanatory diagram of an optical multiple correlation apparatus. Using the output of this device, the output due to the optical cross-correlation of the second embodiment can be detected in parallel, and can be used as the input vector component of the pattern recognition method of the present invention.

Light beam 5 emitted from coherent light source 51
The beam path 2 is expanded in beam diameter to an appropriate size by a beam expander 53, and the beam path is divided into two by a beam splitter 54. The light beam 52a that has passed through the beam splitter 54 is reflected by the total reflection mirror 55 and enters the readout surface of the optically addressed spatial light modulator 56. On the writing surface (that is, address surface) of the spatial light modulator 56, an image formed by forming a plurality of input patterns 57 and an image of a reference pattern group 58 are written. The writing optical system at this time is configured as follows. That is, the input pattern 57
Are illuminated with incoherent light or coherent light, and each lens 59a, 59b,.
に よ り forms an image on each writing surface.

On the other hand, each reference pattern of the reference pattern group 58 corresponds to each lens 59a, 5a of the lens array 59.
9b, arranged at positions corresponding to the respective input pattern images on the writing surface of the spatial light modulator 56, illuminated with incoherent light or coherent light, and spatially modulated by the imaging lens 60. The image is formed on the container 56. At this time, when exposing the input pattern 57 and the reference pattern group 58 at the same time, the respective optical paths are combined using the beam splitter 61. In FIG. 5 , in addition to this, a total reflection mirror 62 is used to bend the imaging optical system of the reference pattern.

FIG. 6 shows an example of the arrangement of the patterns written in the spatial light modulator 56. Each reference pattern 58 'and each input pattern 57' replicated by the lens array optics are arranged on the spatial light modulator 56 in pairs. Therefore, each pair of joints
By performing the calculation by the transform, the correlation calculation between each reference pattern and the input pattern can be performed at a time.

On the reading surface of the spatial light modulator 56, the reflectance distribution and / or the phase reflectance distribution having an intensity corresponding to the light intensity of the above-described write pattern (intensity reflectance distribution × phase reflectance distribution collectively). In the form of a complex amplitude reflection distribution). Therefore, the spatial light modulator 56
The light beam 52a incident on the spatial light modulator 56 has an intensity distribution or a phase distribution corresponding to these writing patterns or both (collectively referred to as a complex amplitude distribution in the form of intensity distribution × phase distribution). Emit.

Here, the reading surface of the spatial light modulator 56 and the writing surface of the spatial light modulator 63 are respectively the front focal plane and the rear focal plane of each lens 64a, 64b,. And the light flux 52a has a power spectrum pattern corresponding to each pair of the reference pattern and the input pattern, that is, a complex amplitude distribution corresponding to the pattern corresponding to each pair of the reference pattern and the input pattern. A light intensity pattern taking the intensity of the Fourier transform is formed. As a result, a complex amplitude distribution corresponding to the power spectrum pattern is formed on the reading surface of the spatial light modulator 63, and the light flux 5 reflected by the beam splitter 54 is formed on the reading surface of the spatial light modulator 63.
2b is further reflected by the beam splitter 65 and enters. Therefore, the light beam 52b emitted from the spatial light modulator 63
Has a complex amplitude distribution corresponding to a power spectrum pattern corresponding to each pair of a reference pattern pattern and an input pattern on the readout surface of the spatial light modulator 63, and the light beam 52b is formed by a lens 66a,
The light passes through the lens array 66 having the light receiving surfaces 67b. At this time, the reading surface and the light receiving surface of the spatial light modulator 63 correspond to each lens 6 of the lens array 66.
Are arranged on the front focal plane and the rear focal plane of each of 6a, 66b,..., And each lens is arranged corresponding to the power spectrum corresponding to each of the above-mentioned types. Therefore,
The light intensity distribution observed on the light receiving surface is a spatial light modulator 63
, The intensity distribution of the Fourier transform of the complex amplitude distribution corresponding to each power spectrum pattern on the readout surface, that is, the square pattern of the autocorrelation of the paired pattern of each reference pattern and the input pattern.

The square pattern of the autocorrelation of the paired pattern of each reference pattern and the input pattern is the square of the autocorrelation of each reference pattern and the square of the autocorrelation of the input pattern on the optical axis of each lens. Then, a square pattern of the correlation between the reference pattern and the input pattern appears at a relative position corresponding to the relative position between the reference pattern and the input pattern with respect to the optical axis. Therefore, since the position where the reference pattern and the input pattern are presented is known in advance, by observing the light intensity at the position where each correlation pattern appears on the light receiving surface 67, the correlation between each reference pattern and the input pattern can be determined. Can be detected.

It is to be noted that a method of arranging a photodiode array and a method of arranging an image pickup device such as an image pickup tube or a CCD on the light receiving surface 67 can be considered. Thus, by observing the light intensity at a desired position (ie, a position where a correlation pattern between each reference pattern and the input pattern appears), it is possible to detect a correlation with each reference pattern.

The output from the array light receiving surface 67 obtained as described above is used as each component of the input vector according to the present invention, and the pattern can be identified in the same manner as described in the first embodiment. . The optical multi-correlation detection optical system described here is described in the specification of Japanese Patent Application No. 3-33423 by the applicant of the present invention, including other embodiments.
Where the output of the described optics is
Since all of them can be used in the pattern identification method of the present invention, refer to those specifications.

[0041]

As described above, the above-described effects are obtained by the pattern identification method of the present invention. In summary, the following remarkable technical effects are obtained. That is, the learning is always converged by the pattern identification method of the present invention, and the number of times of learning only needs to be one. And a pattern identification method capable of identifying that is unknown.

[Brief description of the drawings]

FIG. 1A is a schematic configuration diagram illustrating the concept of an embodiment of a pattern identification method according to the present invention. FIG. 1B is a schematic configuration diagram showing the concept of one embodiment of the pattern identification method of the present invention.

FIG. 2 is a configuration diagram showing a configuration of a conventional typical neural network.

FIG. 3 is a conceptual diagram showing an example of an identification area according to a conventional k-nearest neighbor method.

FIG. 4 is a schematic configuration diagram illustrating functions of an optical system according to an embodiment of the pattern identification method of the present invention.

FIG. 5 is a schematic configuration diagram illustrating functions of an optical system according to another embodiment of the pattern identification method of the present invention.

FIG. 6 shows an arrangement of an input pattern and a reference pattern on a spatial light modulator used in the pattern identification method of the present invention.
It is explanatory drawing which shows an example.

[Explanation of symbols]

 11 to 19 input vector 20, 21, 22 hypersphere 24, 25, 26 identification area 31, 32 test input vector 41 laser 42, 52 luminous flux 43, 53 beam expander 44, 56, 63 spatial light modulator 45, 47 Fourier transform lens 46 Matched filter 48 Photoelectric conversion element array 51 Light source 54, 61, 65 Beam splitter 55, 62 Total reflection mirror 57 Input pattern 58 Reference pattern group 59, 64, 66 Lens array 60 Imaging lens 67 Light receiving surface

Continuation of the front page (56) References JP-A-59-32081 (JP, A) JP-A-3-225566 (JP, A) JP-A-57-64870 (JP, A) Tadayoshi Kurokawa, "Neural Net Work and its development support software, ”Interface, Japan, CQ Publishing Co., Ltd., October 1, 1989, Vol. 15, No. 10, pp. 239-250, Patent Office CSDB Document No .: CSNW199900640007 Yuzo Hirai, "Neural Network for Recognition and Memory", Journal of the Institute of Electronics, Information and Communication Engineers, Japan, The Institute of Electronics, Information and Communication Engineers, February 25, 1990, Vol. 73, No. 2, pp. 124-130, Patent Office CSDB Literature No .: CCNT199901403005 (58) Fields investigated (Int. Cl. ⁷ , DB name) G06N 1/00-7/00 G06G 7/60 G06K 9/00 G06T 7/00 G10L 15 / 00 JICST file (JOIS) CSDB (Japan Patent Office)

Claims (7)

(57) [Claims] 1. An identification area for identifying a pattern is created in an n-dimensional input vector space having n data representing components of an input pattern as components, and a positional relationship between the identification area and the input vector is determined. In the pattern identification method for identifying a pattern based on the identification area, the identification area is created one by one for each identification class to be identified, and the creation of the identification area is performed by an arbitrary one belonging to the first identification class. the input vector and C _Si, picks an input vector C _Ti belonging to another identification class in the closest distance from the vector C _Si, vector C _Si -
The distance between a point that internally divides C _Ti at an arbitrary ratio and the vector C _Si is calculated, and the distance is a radius, and the input vector C _Si
Is defined as a region R _i in an n-dimensional hypersphere centered at
The hypersphere region R _pi is created for all the input vectors of the identification class of, and the logical sum of the hypersphere regions R _{i is used} as the identification region of the first identification class, and the above operation is performed for each class to be identified. The pattern identification method, characterized in that: 2. An identification area for identifying a pattern is created in an n-dimensional input vector space having n pieces of data representing features of an input pattern as components, and a positional relationship between the identification area and the input vector is determined. In a pattern identification method for identifying a pattern based on an average vector of input vectors belonging to one identification class,
Le and C _S, the vector C _S mean vector C _T of another identifying class calculated on the closest distance from, the vector C _S -C _T and the point which internally divides at any ratio vector C _S Is calculated, and an n-dimensional hypersphere region whose distance is a radius and whose center is the average vector C _S is defined as an identification region of the first identification class. A pattern identification method characterized by performing the same for all patterns and creating a similar identification area for each pattern. 3. The method according to claim 2, wherein the ratio of the internally dividing points is 1: 1.
3. The pattern identification according to claim 1, wherein
Method. 4. The data representing the characteristics of the input pattern,
Uses output from correlation operation between input pattern and reference pattern
3. The method according to claim 1, wherein
Pattern identification method. 5. The method according to claim 1, wherein the correlation operation is performed optically.
The pattern identification method according to claim 4, wherein 6. The optical Fourier transform according to claim 1, wherein the optical correlation operation is an optical Fourier transform.
Using means based on the similarity of the patterns on the replacement surface
The pattern identification method according to claim 5, wherein: 7. A pattern on the optical Fourier transform plane.
The optical correlation calculation means based on the similarity between the input pattern and at least the input pattern and the reference
Modulate the complex amplitude distribution of the first light beam according to the pattern
(Ie, display input and reference patterns)
A first spatial light modulator and a first spatial light modulator
Optically Fourier transforming the complex amplitude distribution of the first light beam
A first Fourier transform lens for converting
A second light beam according to the Fourier-transformed light amount distribution of the light beam;
A second spatial light modulator for modulating the complex amplitude distribution of
Complex amplitude component of the second light beam on the second spatial light modulator
Second Fourier Transform for Optically Fourier Transforming Cloth
Interchangeable lens and its Fourier-transformed light amount of the second light flux
Optical correlation detection device comprising light receiving element for receiving distribution
, And the wherein the first spatial light modulator on a plurality of performing correlation detection
Are displayed simultaneously and in parallel, and each reference
One input pattern for each reference pattern near the pattern
Are displayed simultaneously and in parallel, and the first Fourier transform
Corresponds to each reference pattern and input pattern pair
An array of individual lenses, and the second lens
The Fourier transform lens is a first Fourier transform lens array.
Each Fourier transform on the second spatial light modulator caused by b
Specially designed to be an individual lens array corresponding to the pattern
The pattern identification method according to claim 6, wherein