JPH07239939A - Image recognition device - Google Patents

Abstract

PURPOSE:To recognize an image even when the size of an image to be recognized is largely different from that of a teacher image and much noise is included while increasing the discrimination processing speed. CONSTITUTION:A characteristic part is extracted from teacher image information in the learning mode and given to a cerebellum modeled computer (CMAC), in which the characteristic part is learned, discrimination use image information is extracted in the discrimination mode and given to the CMAC, in which the image information most similar to the learned teacher information is discriminated and outputted. In the learning mode, image data 11 are given to a fast speed Fourier transformation (FFT) section 12, in which a prescribed frequency component is obtained, a filter 13 is used to extract a specific component and it is give to a CMAC unit 15, where the component is learned and an image code number is provided to the component. On the other hand, in the discrimination mode, the image data are given to a FFT section 17, from which a prescribed frequency component is obtained and given to the CMAC unit 15 via an expansion section 18 expanding the image and a discrimination section 19 makes discrimination. A code number of an image most similar among images learned in the learning mode is extracted.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image recognition device using a cerebellum model computer (CMAC).

[0002]

2. Description of the Related Art In recent years, neural networks (hereinafter referred to as N
There is a means for performing image recognition by using (N). This N
Since it is known that the image recognition apparatus using N requires a long learning time, the binary image is used as the preprocessing of the normal image, or the processing is performed by using a particularly high-speed CPU.

[0003]

The image recognition apparatus based on the above-mentioned NN starts learning with a known image as unknown, so that the learning time becomes extremely large and the recognition processing cannot be performed at high speed.

The present invention has been made in view of the above circumstances, and provides an image recognizing device capable of recognizing even if the recognition image is greatly deviated from the teacher image or there is a lot of noise while speeding up the determination process. The purpose is to do.

[0005]

According to the first aspect of the present invention, teacher image information is input,
The learning mode for learning the teacher image information and the determination mode for inputting the determination image information to be compared and determined with the image information learned in the learning mode are provided. Extract a part and attach an arbitrary code number to the image, input this code number and the extracted image into the cerebellum model computer to learn the teacher image information, and input the determination image information in the determination mode. And input to the cerebellum model computer,
The image information most similar to the learned teacher image information is determined and output.

A second aspect of the present invention is characterized in that the teacher image information and the determination image information are subjected to fast Fourier transform to obtain Cartesian coordinate and polar coordinate information at the output.

A third aspect of the invention is characterized in that the orthogonal coordinate information obtained from the determination image information is input to the cerebellum model computer via the image expansion / contraction unit.

In the fourth aspect of the invention, in addition to the teacher image information, the teacher image information that is rotated little by little is extracted and input to the cerebellum model computer together with the code number so that the teacher image information is learned.

A fifth aspect of the invention is characterized in that in the determination mode, the determination image information is input to the cerebellum model computer via the image rotation unit.

A sixth aspect of the invention provides an image expanding / contracting portion in the determination mode of the fourth aspect of the invention.

A seventh aspect of the invention is characterized in that the image information for determination is subjected to fast Fourier transform and then input to the image rotating section.

In the eighth invention, the teacher image information is subjected to fast Fourier transform, a power spectrum is obtained from the converted information, the maximum frequency of this power spectrum is standardized by a normalizing section, and the characteristic of the image is calculated from the standardized information. It is characterized by extracting a part.

A ninth aspect of the present invention is characterized in that the determination image information is subjected to fast Fourier transform, a power spectrum is obtained from the transformed information, and the maximum frequency of the power spectrum is standardized and extracted by the normalizing section. It is a thing.

[0014]

In the learning mode, the characteristic portion of the image is extracted from the teacher image information in advance. A code number is attached to the extracted image, and the code number and the extracted image are input to the cerebellum model computer to learn the teacher image information. Then, it is determined how similar the image information determined in the determination mode is to the learned image information. In the second invention, since the image may be rotated, polar coordinate information is obtained. In the third invention, the image is enlarged or reduced. In the fourth aspect of the present invention, the rotating state of the image that does not use the polar coordinate information is detected by learning the image that rotates little by little as the teacher image information in advance. In the fifth, sixth, and seventh inventions, in the determination mode, the image rotation unit rotates the determination image information by ± 90 ° and 180 °. 8th
In the invention and the ninth invention, by standardizing information,
Even if the position of the target from the camera is deviated or the illumination is changed, it can be recognized.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The embodiments of the present invention will be described below with reference to the drawings.
Will be described). CMAC is Cerebell
er Model Arithmetic Compu
Abbreviated ter, it is a formalization of a mathematical model of the information processing mechanism of many neurons in the cerebellar cortex. CMAC is defined by a series of mapping relationships as follows.

S → M → A → P where S is an input vector, M is a set of mossy fibers used to code S, A is a set of granule cells to which M is connected, and P is an output. It is a value.

FIG. 9 is a conceptual diagram of CMAC.
Indicates a set of input signals, and a control target value and a feedback signal are input to the set 1 of input signals. A load selection signal is transmitted from the input signal set 1, and this signal is input to a predetermined position in the load table 2. The output of Table 2 of the load is calculated and output by the total sum calculation unit 3. The output of the calculation unit 3 is deviated by the deviation detector 4 from the target value of the output which is the teacher signal, and the calculation unit 3 adjusts the load according to the deviation. Note that FIG. 9 also shows mapping-related positions.

CMAC takes advantage of the property that higher-order vectors can be generally expressed as a mapping to a lower-order space. In other words, a high-order vector can be expressed by using a plurality of low-order vectors, and the function required for the low-order vector is good at a low level. In general, a function is considered to be a mapping from a set of states defined by independent variables to a set of states of dependent variables and is expressed as f: C → E. This expression is "f is a relation that maps set C to set E"
And read. FIG. 10 shows this conceptual diagram. As can be seen from FIG. 10, for any state of the set C, one state in the set E can be obtained by the relation f. Also, multiple points in set C may be mapped to one point in set E.

Now, the input S = (s1, s2, s3 ... s
Let h be the operator as a function that maps n) to the output P, and h can be expressed as follows.

P = h (S) or P = h (s1, s2, s3 ... sn) This is illustrated in FIG. When the output is a vector P, it is expressed as an operator or a set of operators h as shown in FIG.

As described above, the operator can describe the action of one neuron or a group of neurons because the input is mapped by the function at the output. For example,
When describing the function of one neuron, considering the input to the neuron as a vector and the output as a scalar, it can be described as P = h (S). When considering a neuron group, if the input is considered as a vector, it can be described as P = H (S).

FIG. 1 shows a first embodiment in which the above-mentioned CMAC is applied to an image recognition apparatus. In FIG. 1, 11 is image data, and this image data 11 is introduced into an image recognition apparatus having two modes, a learning mode and a determination mode, and an image is first learned in the learning mode. In the learning mode, the fast Fourier transform unit (FFT unit) 12 and the filter 13 perform preprocessing of image data. The image data preprocessed by the FFT unit 12 and the filter 13 is a CM that constitutes the controller 14.
It is introduced into the AC unit 15. CMAC unit 15
Output and the arbitrary code number attached to the image are detected by the deviation detector 16 and the image is learned by the CMAC unit 15.

Next, it is determined how similar the image determined in the determination mode is to the learned image. The determination mode is introduced into the CMAC unit 15 after being processed by the pre-processing including the FFT unit 17 and the expansion / contraction unit 18. In FIG. 1, the CMAC unit is drawn separately in the learning mode and the determination mode, but they are shown separately for convenience of description, and actually show the same CMAC unit. CMAC unit 1
The similarity between the image learned as the image learned in 5 and the image to be determined is processed, the determination result is determined by the determination unit 19, and the code number of the most similar image is output.

The operation of the embodiment configured as described above will be described. First, the learning mode will be described. For example, the power spectrum (hereinafter referred to as PS) is obtained by the FFT unit 12 from a teacher image taken on a video. Only the characteristic PS of the image passes through the obtained PS by the filter 13. Here, an arbitrary code number is given to the image, and the image is learned by the CMAC unit 15 using PS and the code number.

Next, in the determination mode, the FF is changed from the determination image.
PS is calculated by the T unit 17. When the image is expanded / contracted, the expansion / contraction part 18 expands / contracts PS to expand the CMAC unit 1.
The image which is most similar to the image learned by the determination unit 19 is selected and the code number of the image is output. The controller 14 switches the learning mode and the determination mode, and depending on the type of the teacher image, the CMA
The number of C units 15 is increased or decreased.

Next, the details of the FFT units 12a, 12b and 17a, 17b, the filters 13a, 13b, the expansion / contraction unit 18, the CMAC unit 15 and the determination unit 19 will be described with reference to FIG. In FIG. 2, the FFT units 12a and 12b
And 17a and 17b determine the orthogonal coordinate component P (ωx, ωy) and the polar coordinate component P (ωr, ωθ) that are PS components of the image. The filters 13a and 13b have (i = 1, 2, ..., Z
-1, j = i + 1 ... z (│Pi (ωx, ωy) -Pj
Pass PS satisfying (ωx, ωy) │> quantization interval)). However, the number of teacher images = z, the spectra P1 (ωx, ωy) to Pz (ωx, ωy) of images 1 to z, and the quantum interval = CMAC are parameters.

The expansion / contraction part 18 is represented by P (ωx, ω)
y) is expanded / contracted to P (ωa, ωb), and P (ωa, ωb) is input to CMAC. Ωa = t · ωx, ωb = t
Ωy, t = expansion coefficient, P (ωa, ωb) = P (ωx,
ωy).

In the CMAC unit in the learning mode, one CMAC unit has a three-dimensional structure and three inputs and one input.
Output. The flow of processing is as follows. P (ωx, ωy) and P (ω determined by the filters 13a and 13b
r, and input to n number of Omegashita) from number 1 CMAC units, the output _{O 1 (x, y) ~O} n (x, y), O 1 (r,
obtain _{θ) ~O n (r, θ} ). The absolute value of the difference between this output and the code number is learned.

Next, in the CMAC unit in the determination mode, P (ωx, obtained by the FFT units 17a and 17b)
ωy) and P (ωr, ωθ), P (ω determined by the expansion / contraction unit 18
a, ωb) and outputs O _n (x, y) and O _n (r,
theta), obtaining a O _n (a, b).

The judging section 19 outputs outputs O _n (x, y), O
_The code number closest to the learning image is obtained from _n (r, θ) and O _n (a, b). Specifically, it is as follows. For this, the output O ₁ (x,
y), the frequency distribution is calculated, and the mode _values O _1max and O _1max
For the standard deviation σ ₁ . Similarly, the remaining O
_{2 (x, y) ~O n} (x, y) for, O _2max (x,
_{y) ~O nmax (x, y} ), σ 2 (x, y) ~σ n (x,
y) is calculated.

Similarly to the above, O _n (r, θ), O
_{For n} (a, b), O _1max (r, θ) to O
_nmax (r, θ) and σ ₁ (r, θ) to σ _n (r, θ), O
_{1max (a, b) ~O nmax} (a, b) and σ ₁ (a, b) ~
Find σ _n (a, b). Finally, (i = 1,2 ... n│
Among σ _n (x, y), σ _n (r, θ) and σ _n (a, b)), the mode O _i of the minimum value σ ₁ is set as the recognition result.

A case in which the image of the glove shown in FIG. 3A is recognized using the above embodiment will be described.

(1) Learning Mode Codes 60 and 1 are assigned to the gloved teacher images shown in FIG. 3A, respectively.
At 20 the CMAC unit is trained.

(2) Judgment Mode The judgment image of the glove of FIG. 3B is introduced into the CMAC unit and the output at that time is made into a frequency distribution, which is the solid line shown in FIG. Similarly, the dotted line in FIG. 4 is a frequency distribution of unlearned images. When both distributions are solid lines, a peak occurs in the code number 60 at the time of learning, and a dotted line has less peaks than the solid line. . Therefore, the image determination can be easily performed from the standard deviation of the mode 60 of the solid line, and the code number of the recognition result 60 is obtained.

FIG. 5 is a block diagram showing a second embodiment of the present invention. The same parts as those in FIGS. 1 and 2 are designated by the same reference numerals. The second embodiment of FIG. 5 is characterized in that the means for obtaining the polar coordinate component is omitted, and instead of this, in the learning mode, one image is rotated little by little to learn with the code number. The learning means for the image data 11 is the CMAC as in the above embodiment.
The image data 21 is obtained by rotating the image data 11 by θ degrees, which is performed by the unit 15a, and the PS of the image data 21 is also obtained by the FFT unit 12a and the filter 13a.
The PS and the code number (for example, 70) are learned by the CMAC unit 15b. This process is learned by the CMAC units 15c to 15n by obtaining PS for each image data 21 that is sequentially rotated by θ degrees. For example, the above process
0 degree / θ = n times. Now, assuming that the rotation angle is 10 degrees,
An image rotated by 90/10 = 9 times is learned for each image data, and each image is learned by 9 CMAC units.

As described above, one type of image data 11,
The CMAC unit is made to learn the image data 21 and new code number which are sequentially rotated by θ degrees after learning about 21. The filter 13a has the same characteristics for the type of image and any rotation.

In the judgment mode, PS of the image data 11 is expanded / contracted by the expansion / contraction unit 18 and input to the CMAC unit except that the polar coordinate component is obtained from the above embodiment, and the image is judged from the obtained output O _(n , _z). To do. First, as in the learning mode, FFT is performed on the image to obtain PS. The same PS is input to the CMAC units 15a to 15n, and the output O ₍₁ ,
_{1) to} O _(n , ₁₎ are calculated. This will find the underlined part of the following matrix equation.

[0038]

[Equation 1]

Next, in order to obtain the parts other than the underlined part of the equation (1), the PS expanded / contracted by Z times is used in the CMAC unit 15 as described above.
It can be obtained by inputting a to 15n.

In the expansion / contraction section 18, as shown in FIG.
_In the explanatory view when expanding and contracting to z at the expansion and contraction rate z, the relationship at this time is expressed by the following equation.

P _z (z · ωx, z · ωy) = P (ωx,
ωy) However, the expansion / contraction rate is 0.5 to 2.

The determination section 19 outputs the most reliable code number from the matrix of the above equation (1), and outputs zero if it is not reliable.

First, from the output O ₍₁ , ₁₎ , the most frequent value m ₍₁ , _{1) of the} histogram (hereinafter referred to as a support code) and the dangerous value σ ₍₁ ,
Ask for ₁₎ . This is based on the output O ₍₁ , ₁₎ from the decision unit 19 of FIG.
A histogram of the output O ₍₁ , ₁₎ shown in is created, and the supporting code m ₍₁ , ₁₎ of the histogram is obtained. Then, the dangerous value σ ₍₁ , ₁₎ for m ₍₁ , ₁₎ is calculated from the following equation.

[0044]

[Equation 2]

After the support code m _(n , ₁₎ and the dangerous value σ ₍₁ , ₁₎ are obtained from the output O _(n , ₁₎ as described above, the output O _(n , _{z) is} similarly obtained. Support code m _(n , _z) and dangerous value σ _(n ,
_z) is calculated. Next, the matrix of dangerous values σ _(m , _z) is obtained from the following equation, and σ _min with the smallest dangerous value is obtained.

[0046]

[Equation 3]

The supporting code m _(n , _{z) of the} judgment result outputs zero from σ _{min having} the smallest dangerous value.

Σ _min ≤S is a small dangerous value, and the determination result is the support code m _(n , _z) at that time.

When σ _min > S, the dangerous value is large, and the determination result is zero at this time. Note that S is experimentally determined by a threshold value for determining whether or not it can be determined.

Using the embodiment configured as described above, FIG.
After learning the images of the empty cans A, B, and C, the histogram of each dangerous value σ _min determined by the images similar to FIGS. 7A, 7B, and 7C is shown in FIG. In FIG. 8, the dangerous value σ of can A (FIG. 7A)
_min = 30.5, the dangerous value σ _{min of} can B (FIG. 7B) = 31.
0 and the dangerous value σ _{min of} can C (FIG. 7C) = 28.7, the histograms of can B and can C are shifted from can B by +10 in order.

In the above embodiment, since the FFT for obtaining the polar coordinate component is not required, the image judgment processing becomes fast and the judgment result with low reliability cannot be judged, so that the possibility of erroneous judgment becomes low. Further, the reliability of the determination result can be expressed by a numerical value, which facilitates the processing.

FIG. 13 is a block diagram of the overall construction showing the third embodiment of the present invention, and FIG. 14 is a construction explanatory view showing the detailed construction of the third embodiment, which is the same as FIG. 1, FIG. 2 and FIG. The parts are denoted by the same reference numerals. In the embodiment shown in FIGS. 13 and 14, the image rotation unit 21 is provided in the determination mode. This third embodiment also has two modes, that is, a learning mode and a determination mode, similar to the above-described embodiments, and an image is learned in the learning mode in advance, and how similar the image to be determined in the determination mode is to the learned image. To judge.

In the learning mode, as in the above-described embodiment, FFT is performed on the teacher image taken on the video to obtain PS. Next, a filter passes only the characteristic PS of the image. Here, attach an arbitrary code number to the image, and C with PS and this code number
The MAC unit 15 is made to learn.

In the judgment mode, the judgment image is FFTed and PS
Ask for. Next, rotate the PS to rotate the CMAC unit 1
Then, the judgment section 19 judges this output. The determination result is the code number of the most similar image among the learned images. Here, the rotation of PS is ± 90 ° and 180 °.

In each functional block, as in the above-described embodiment, the controller 14 switches the learning mode and the judgment mode and the CM depending on the type of the teacher image.
Increase or decrease the number of AC units n. FFT section 12a, 12
The operations of b and 17a are performed in the same manner as the embodiment of FIG.

In the image rotating unit 22, 0 °, 90 °, -90
The PS of 180 ° and 180 ° rotation is calculated as follows.

P ₀ (ωx, ωy) = P (ωx, ωy) P ₉₀ (ωx, ωy) = P (−ωx, ωy) P ₋₉₀ (ωx, ωy) = P (ωx, ωy) P ₁₈₀ ( ωx, ωy) = P (ωx, −ωy) The above four types of PS are CMAC units 15a, 15b ...
Used as input.

In the learning mode, one CMAC unit has a three-dimensional structure and three inputs and one output. The flow of processing is P (ωx, ω) obtained by the filters 13a and 13b.
y) sequentially input from CMAC unit No. 1 to No.
Output _{O 1 (x, y) ~O} n (x, y) obtained. Then, the absolute value of the difference between this output and the code number is learned.

In the CMAC unit in the judgment mode, P (ωx, ωy) [P ₀ , P ₉₀ , P _-90 ,
P ₁₈₀ ] are sequentially input to the CMAC unit and output O _0n ,
O _90n , O _−90n , and O ₁₈₀ are obtained. (However, n = CMAC
Number of Units) The determination unit 19 obtains the code number closest to the learning image from the outputs O _0n , O _90n , O _−90n , and O ₁₈₀ . As shown in FIG. 14, when the frequency distribution of the output O ₀₁ of the CMAC is obtained by the determination unit 19, the mode value O _1max and the standard deviation σ ₁ for this mode value O _1max are obtained. Similarly, the remaining outputs O _{02 to} O
For _0n, the mode _{O 2max (x, y) ~O} nmax (x,
y) and standard deviation σ ₁ (x, y) to σ _n (x, y). Similarly, the mode and the standard deviation of the outputs O _90n , O ₋₉₀ n, and O ₁₈₀ are obtained. Finally, the mode value O _i that is the minimum value of the standard deviation is set as the recognition result.

Using the embodiment of FIGS. 13 and 14, FIG.
The second embodiment is different from the second embodiment as follows. Second shown in FIG.
In an embodiment, the object rotation in the image is the polar FFT of the image.
It is necessary to perform FFT twice for each image, as determined by In addition, in the second embodiment of FIG. 5, any rotation can be dealt with, but in reality, the target lies sideways as a change in the posture of the target object (9
The examples shown in FIGS. 13 and 14 are limited to cases in which the rotation is 0 ° or the opposite (180 ° rotation). In this embodiment, therefore, one learning and one learning is performed without using the FFT in polar coordinates or the rotated teacher image several times.
It was made to be able to be recognized by FFT times. As a result, polar coordinates are not required, which has the advantage of speeding up the image determination processing.

FIG. 15 is a block diagram of the overall construction showing the fourth embodiment of the present invention, and FIG. 16 is a construction explanatory view showing the detailed construction of the fourth embodiment, and is shown in FIG. 1, FIG. 2, and FIG. The same parts are designated by the same reference numerals. In the embodiment shown in FIGS. 15 and 16, the FFT units 12, 17 and 12 are used.
Standardization units 23, 23a, 23b and 24 are provided to standardize the output of the FFT of the teacher image and the determination image information at a, 12b and 17a. This fourth embodiment also has two modes, that is, a learning mode and a determination mode, similar to the above-described embodiments, and the image is learned in the learning mode in advance, and how similar the image determined in the determination mode is to the learned image. To judge.

In the learning mode, as in the above-described embodiment, the FFT is performed on the teacher image information captured in the video, and the power spectrum PS
Ask for. Next, the maximum frequency of the PS is standardized by the standardization units 23, 23a, 23b and output. This output is then passed through only the characteristic PS of the image by filters 13, 13a and 13b. Here, an arbitrary code number is given to the image, and the CMAC unit 15 is made to learn with PS and this code number. In the determination mode, the image for determination is subjected to FFT to obtain PS. Next, the frequency having the maximum PS is standardized by the normalizing unit 24 and output as in the above. The output PS is input to the CMAC unit 15, and the output of the PS is determined by the determination unit 19. The judgment result is
It is the code number of the most similar image among the learned images.

In each functional block, as in the above-described embodiment, the controller 14 switches the learning mode and the judgment mode, and the CM depending on the type of the teacher image.
Increase or decrease the number of AC units n. The FFT unit 12,
17 and 12a, 12b, 17a, after obtaining P (ωx, ωy), which is the PS component of the image, the normalization units 23, 2
In 3a, 23b, and 24, the maximum PS frequency (ωx
max, ωymax) is standardized as follows.

Ωx = ωx / ωxmax Ωy = ωy / ωymax Pn (Ωx, Ωy) = P (ωx, ωy) / P (ωxma
x, ωymax) The outputs standardized by the normalization unit are filters 13, 13a,
13b, and the PS that satisfies the following equation is passed here. The PS that has passed through is the CMAC unit 15a, 15b ...
Entered in.

| Pni (Ωx, Ωy) -Pnj (Ωx,
Ωy) │> quantization interval (i = 1, 2 ... z-1, j = i + 1, ... z) where the number of teacher images = z and the spectrum Pn of images 1 to z
1 (Ωx, Ωy) to Pnz (Ωx, Ωy), quantization interval = CMAC parameter.

The above-mentioned CMAC unit has the same structure as that of the above-described embodiment, and as shown in FIG. 16, both the learning mode and the judgment mode have a three-dimensional structure and three inputs and one output. FIG.
In the learning mode shown, the filter 13a, 13b in the obtained Pn (Ωx, Ωy) type to n-th from 1st sequential CMAC units, output _{O 1 (x, y) ~O} n (x, y) and obtain. Then, the absolute value of the difference between this output and the code number is learned. In judgment mode, Pn (Ωx, Ωy) is sequentially C
The MAC unit is input and the output O _no (n = number of CMAC units) is obtained.

The judging section 19 is similar to the third embodiment in that
Request code number closest to the learning image from the output O _n.
Specifically, it is performed as follows. The frequency distribution is obtained for the output O ₀₁ of the CMAC units 15a, 15b, ... In the determination mode of FIG. 16, and the mode O _1max (x, y) and the standard deviation σ ₁ for this mode O _1max are obtained. Similarly,
For the remaining outputs O _{02 to} O _0n , the mode value O _2max (x,
y) to _Onmax (x, y) and standard deviation σ ₁ (x, y) to
Find σ _n (x, y). Finally, the mode value O _i that is the minimum value of the standard deviation is set as the recognition result.

As described above, the provision of the standardization section in the fourth embodiment has the advantage that it can be recognized even if the position from the target camera is displaced and that it can be recognized even if the illumination changes.

[0069]

As described above, according to the present invention,
In the first invention to the third invention, as compared with the case where the image determination process is NN, learning is performed, so that high-speed processing can be performed and determination can be performed easily, and the recognized image is significantly different from the teacher image, and noise is generated. You will be able to recognize at most. Fourth
Since the invention to the seventh invention do not use polar coordinate components, there are advantages that the image determination processing can be performed at a higher speed and the possibility of erroneous determination processing can be reduced. In the eighth invention and the ninth invention, by normalizing the information, it becomes possible to recognize even if the position of the target from the camera is displaced or the illumination is changed.

[Brief description of drawings]

FIG. 1 is a block diagram showing a first embodiment of the present invention.

FIG. 2 is a configuration explanatory view showing a detailed configuration of the first embodiment.

3A is an explanatory diagram showing a teacher image, and FIG. 3B is an explanatory diagram showing a determination image.

FIG. 4 is a frequency distribution characteristic diagram of an output O _{1 (x} , _y) .

FIG. 5 is a structural explanatory view showing a second embodiment of the present invention.

FIG. 6 is a diagram for explaining a stretchable portion.

7A, 7B and 7C are explanatory views showing images of empty cans.

FIG. 8 is a frequency distribution chart of empty cans shown in FIGS. 7A, 7B and 7C.

FIG. 9 is a conceptual diagram of CMAC.

FIG. 10 is a conceptual diagram showing a mapping relationship.

11 is an explanatory diagram showing an operator that maps to a scalar output P. FIG.

FIG. 12 is an explanatory diagram showing an operator that maps to a vector output P.

FIG. 13 is a block diagram showing a third embodiment of the present invention.

FIG. 14 is a structural explanatory view showing a detailed structure of a third embodiment.

FIG. 15 is a block diagram showing a fourth embodiment of the present invention.

FIG. 16 is a structural explanatory view showing a detailed structure of a fourth embodiment.

[Explanation of symbols]

11, 21 ... Image data 12, 12a, 12b, 17, 17a, 17b ... Fast Fourier transform section 13, 13a, 13b ... Filter 14 ... Controller 15, 15a-15n ... CMAC unit 16 ... Deviation detection section 18 ... Expansion / contraction section 19 ... Judgment part 22 ... Image rotation part 23, 23a, 23b, 24 ... Normalization part

Claims (9)

[Claims] 1. A learning mode in which teacher image information is input and learning is performed on the teacher image information, and a determination mode is input in which determination image information to be compared with image information learned in this learning mode is input. In the learning mode, the characteristic portion of the image is extracted from the teacher image information, an arbitrary code number is given to the image, and the code number and the extracted image are input to the cerebellum model computer to learn the teacher image information. In the determination mode, the input determination image information is extracted and input to the cerebellar model computer, and the image information most similar to the learned teacher image information is determined and output. Image recognition device. 2. The image recognition apparatus according to claim 1, wherein the orthogonal coordinate and polar coordinate information is obtained as an output from the teacher image information and the determination image information by using a fast Fourier transform. 3. The image recognition apparatus according to claim 1, wherein Cartesian coordinate information obtained from the determination image information is input to the cerebellum model computer via the image expansion / contraction unit. 4. In the learning mode, in addition to the teacher image information, the teacher image information that is rotated little by little is extracted and input to a cerebellum model computer together with a code number so that the teacher image information is learned. Claim 1
The image recognition device described. 5. The image recognition apparatus according to claim 1, wherein in the determination mode, the determination image information is input to the cerebellum model computer via the image rotation unit. 6. The image recognition apparatus according to claim 4, wherein the determination image information is input to the cerebellum model computer via the image expansion / contraction unit. 7. The image recognition apparatus according to claim 5, wherein the image information for judgment is subjected to fast Fourier transform and then inputted to the image rotation unit. 8. The fast image Fourier transform of the teacher image information is performed, a power spectrum is obtained from the transformed information, the maximum frequency of this power spectrum is standardized by a normalizing section, and the characteristic part of the image is extracted from the standardized information. The image recognition device according to claim 1, 2 or 4, wherein the image recognition device is extracted. 9. The image information for judgment is subjected to fast Fourier transform,
7. The image recognition apparatus according to claim 1, wherein the power spectrum is obtained from the converted information, and the maximum frequency of the power spectrum is standardized and extracted by the standardization unit.