JPH07120324A - Color-classifying apparatus - Google Patents

Abstract

PURPOSE:To classify colors well even when a spectrum is changed, with a simple constitution of an apparatus at a low cost while enduring mechanical vibrations or the like without limiting a light source. CONSTITUTION:A rotary color filter 12 is prepared, which has a plurality of band-pass filters 12A-12E of different band regions. A control part 26 suitably inserts each of the band-pass filters 12A-12E between an object O and CCD 14 by means of a driving motor 24. A classifying/operating circuit 28 calculates a classification spectrum based on a statistic technique from a reflectance spectrum of the object O photographed by the CCD 14, and classifies the object O using the classification spectrum.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a color classification device for classifying objects by using colors.

[0002]

2. Description of the Related Art Conventionally, in the management of coating color and dyeing degree at various industrial production sites, or the color measurement of products, or the color measurement of a subject in medical or academic fields, etc. A color identification device for identifying the color of an object is used.

For example, in the color identification apparatus disclosed in Japanese Patent Laid-Open No. 3-267726, two classes are classified by statistically processing the reflection spectrum of the object. Specifically, the reflection spectrum of an object whose class is known is statistically processed by using Foley Sammon transform (FS transform) (Q.Tian, M.Barbaro et al.,
“Image classification by Foley-Sammon transfor
m ”, Optical Engineering, Vol.25, No7, 1986).

Here, the FS conversion is a method of classifying into two classes, and specifically, from S1 and S2 when two given objects are given, Fisher ratio R (di) = (di ^t S1 di) / (di ^t S2 di) ... (1) di: classification spectrum di ^t : classification spectrum (transpose) S1: interclass covariance matrix S2: intraclass covariance matrix R (di) is maximized Is to find the spectrum di. After that, the spectrum di for this classification
Is called a classification spectrum. This classification spectrum di is
Since it has the same number of dimensions as the spectrum of the object, it should be accurately expressed as di (.lambda.), But in this specification, it is expressed as di for simplicity. And Fisher rat
Two kinds of classification spectra for increasing io are obtained. Fisher
The classification spectrum di that maximizes the ratio is d1, and this d
The classification spectrum di that maximizes the Fisher ratio among the spectra orthogonal to 1 is d2. Two classes are classified by projecting each data on the space formed by the classification spectra d1 and d2. The classification spectra d1 and d2 are obtained from the following equation.

D1 = α1 S2 ^-1 Δ d2 = α2 S2 ^-1 [I- ( ^Δt S2 ^-2 Δ) / ( ^Δt S2 ^-3 Δ) S2 ^-1 ] Δ (2) where α1, α2 is a normalization coefficient, Δ is X1 -X2 (difference spectrum between class 1 and class 2), and I is an identity matrix.

The classification spectrum d1 thus obtained,
The inner product of the classification spectrum and the reflection spectrum of the object is obtained in order to project each data onto the space constituted by d2. The reflection spectrum of the object is f (λ) (however,
If λ = wavelength), the inner products t1 and t2 can be expressed by the following equations.

T1 = f (λ) d1 t2 = f (λ) d2 (3) Here, the symbol "." Represents an inner product operation.

In the device disclosed in the above publication, this inner product t1
, T2, the classification boundary is determined as shown in FIG. 7, and a filter having the characteristics of this classification spectrum is realized by using the diffraction grating 1 and the liquid crystal 2 as shown in FIG.

[0009]

However, the classification spectra d1 and d2 generally have a complicated shape as shown in FIG. 24, and since they take positive and negative values, the diffraction grating 1, the liquid crystal filter 2, etc. Mounting accuracy is also strictly required.
Therefore, if the mounting position of the device shifts due to mechanical vibration accompanying the movement of the device, the classification accuracy will be significantly reduced. There is also a problem that the diffraction grating itself is expensive. Therefore, there is a demand for a color classification device having a simple device configuration, low cost, and capable of withstanding mechanical vibration and the like.

Further, in the apparatus of the above publication, the light source is limited to some extent (lamp 3), so that it is not suitable for classification with respect to different light sources, and if the spectrum of the light source changes, good classification is performed. Difficult to do. When performing color classification in a factory or the like, the light source can be limited, but there is a demand for an apparatus that can perform good color classification even when the spectrum changes without limiting the light source.

The present invention has been made in view of the above points, has a simple device configuration, is low in cost, and can withstand mechanical vibration and the like, and its spectrum changes without limiting the light source. It is an object of the present invention to provide a color classification device that can perform good color classification even in cases such as cases.

[0012]

In order to achieve the above object, a color classification device according to the present invention is installed between an image pickup means for picking up an image of a reflection spectrum of an object and the object and the image pickup means. A plurality of band pass filters each having a different band and a reflection spectrum of the object imaged by the image pickup means are used to calculate a classification spectrum for classification using a statistical method, and the classification spectrum is used to It is characterized in that it comprises a classification means for classifying the objects.

[0013]

That is, according to the color classification apparatus of the present invention, a plurality of bandpass filters each having a different band are prepared, and each of the plurality of bandpass filters is provided between the object and the image pickup means. Deploy. Then, the classification means calculates a classification spectrum for classification using a statistical method from the reflection spectrum of the object imaged by the imaging means, and the object spectrum is classified using this classification spectrum.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Before explaining the embodiments of the present invention, in order to help understanding of the present invention, the principle of the present invention will be explained first. In the present invention, the filter for classification is not composed of a diffraction grating and a liquid crystal filter as in a conventional device, but a band that allows only a specific wavelength as shown in FIG. By using a filter as shown in (B) or (C) of the same drawing, which is a combination of a plurality of pass filters, a color classifying device having a simple and inexpensive structure is realized.

Further, in order to perform color classification under different light sources, the reflection spectrum of an appropriate reference plate is measured under the same conditions as when photographing an object, and the reflection spectrum of the object is used as the reference plate. The influence of the light source (illumination light) is removed by correcting the reflection spectrum of the above. That is, with λ as the wavelength, the reflection spectrum of the object is f (λ), and the reflection spectrum of the reference plate is s.
(Λ), the reflection spectrum of the illumination light is L (λ), and the sensitivity spectrum of the photographing system (transmission spectrum of the photographing lens, sensitivity spectrum of the image sensor, etc.) is M (λ), the photographing spectrum gi of the object (Λ) and the imaging spectrum gs (λ) of the reference plate are respectively gi (λ) = f (λ) × L (λ) × M (λ) gs (λ) = s (λ) × L (λ) The spectrum gi ′ (λ) of the object is expressed as × M (λ) and gi ′ (λ) = gi (λ) / gs (λ) = f (λ) / s (λ) (4) Can be represented.

Thus, the reflection spectrum L of the illumination light
The influence of (λ) can be removed, and if gi '(λ) is used, classification can be performed under different light sources. Further, when the brightness of the illumination light is further different, the signal gi ′ after removal is
It is sufficient to normalize the power of (λ).

Embodiments of the present invention will be described below with reference to the drawings. [First Embodiment] First, as a first embodiment of the present invention, a color classification apparatus for classifying objects of two classes will be described.

FIG. 1 is a diagram showing the configuration thereof. The color classification apparatus of the present embodiment includes an optical system 10 including a diaphragm and a lens, and a plurality of bandpass filters 1 as shown in FIG.
, C for capturing the images of the object O and the reference plate R. The rotary color filter 12 composed of 2A, 12B, ..., 12E.
CD 14, A / D converter 16, frame memory 18, monitor 20 for displaying the part being photographed, CCD drive driver 22, drive motor 24 for rotary color filter 12, CC
D drive driver 22 and rotary color filter drive motor 24
And the like, and a control unit 26 for sending a command to the classification operation circuit 28 and a classification operation circuit 28 for performing classification.
Composed of.

The rotary color filter 12 is shown in FIG.
As shown in FIG.
12E, each filter is shown in FIG.
It has the property of transmitting an arbitrary bandwidth as shown in. In this embodiment, the rotary color filter 12 is composed of five band pass filters for simplification of the drawing and description. The arrangement of the optical system 10 and the rotary color filter 12 may be reversed such that the rotary color filter 12 is arranged in front of the optical system 10.

As shown in FIG. 3, the classification operation circuit 28 includes a brightness component extraction unit 30 for extracting the brightness component of the object O, and a classification operation unit 32 for performing an operation for classification (FS conversion etc.). , And a classification determination unit 34 for performing learning and classification determination for classification determination.

Here, the luminance component extracting section 30 is shown in FIG.
As shown in FIG. 3, three measurement area extraction units 36A, 36 for extracting the measurement areas of the object O and the reference plate R of the photographed image.
B, 36W, three luminance component averaging units 38A, 38B, 38W for averaging measured luminance components, a luminance component memory "A" 40A for writing the luminance component of photographed class 1 or unknown data, photographed Luminance component memory "B" 40B for writing the luminance component of class 2 data, luminance component memory "W" 40W for writing the luminance component of the imaged reference plate R data, correction circuit 42 for correcting the influence of the light source, correction It has a brightness spectrum memory "dta" 44A for writing the data of class 1 or unknown class and a brightness spectrum memory "dtb" 44B for writing the data of corrected class 2.

Luminance component memories 40A, 40B, 40
As for W, it is possible to write as many luminance components as the number of band-pass filters (five in this embodiment) forming the rotary color filter 12.

The correction circuit 42 is composed of a divider 42 ₁ as shown in FIG. 5A or a remover 42 ₁ and a power normalization circuit 42 ₂ as shown in FIG. 5B. . In the following description of the present embodiment, it is assumed that the configuration shown in FIG.

Luminance spectrum memory 44A, 44B
Is capable of writing as many luminance components as the number N of samples of data to be photographed (each luminance component is composed of data of the number of filters).

On the other hand, the classification calculator 32, as shown in FIG. 5C, has a changeover switch "A" 46, a classification spectrum calculator 48 for obtaining a classification spectrum, and a classification spectrum d1 memory 5 for writing the classification spectrum d1.
0, classification spectrum d2 for writing classification spectrum d2
Memory 52, changeover switch "B" 54, integrator 5
6. The accumulator 58 is composed of an adder 58A and a latch 58B and performs cumulative addition.

Further, as shown in the figure, the classification judging section 34 includes a changeover switch "C" 60, a classification boundary determining section 62 for determining a classification boundary, and a classification boundary memory "cl" 64 for writing the determined classification boundary. , And a classification determination unit 66 that performs classification determination.

Next, the process of classifying the objects of two classes by using the color classifying apparatus having the above-mentioned configuration will be described. In this process, the learning mode for obtaining the classification boundary is first executed, and then the classification mode for performing the color classification of the data of unknown class is performed.

First, the learning mode will be described. This is to obtain a classification spectrum for classifying two classes of objects O as shown in FIG. First, the control unit 26 determines the direction and focal length of the optical system 10 as follows.
Adjust so that two classes of objects can be imaged simultaneously.
Focus adjustment is performed by a focus adjustment mechanism (not shown), and photometry is performed by a photometer (not shown) to set the aperture of the optical system 10 and the exposure time of the CCD 14.

Here, the position of the rotary color filter 12 is controlled so that the first band-pass filter (for example, 12A) of the rotary color filter 12 is used for photographing. Then, the first image is photographed by sending a photographing command to the CCD drive driver 22. The image data captured by the CCD 14 and A / D converted by the A / D converter 16 is transferred to and stored in the frame memory 18. Then, the classification calculation circuit 28 is caused to read the image data stored in the frame memory 18.

In the classification calculation circuit 28, the image data is first transferred to the luminance component extraction section 30. In the luminance component extraction unit 30, the measurement region extraction units 36A and 36B extract the classification target regions corresponding to class 1 and class 2 from the captured image data for each image, respectively. A luminance component is extracted for each pixel.
Then, the luminance component averaging units 38A and 38B detect the average value of the luminance in each region, and the luminance component memories 40A and 4B are detected.
Write to 0B. These are data da1 and db1.

Next, the rotary color filter 12 is rotated to the second
Take a second image with the filter (for example, 12B),
Similarly, the average value is written in the luminance component memories 40A and 40B. These are data da2 and db2.

Such an operation is performed up to the fifth filter (for example, 12E), and the luminance component memory "A" 40A
The data da3, da4, da5 are written in the same, and the data db3, db4, db5 are written in the luminance component memory "B" 40B. That is, with this series of operations, the luminance component memory "A"
40A includes data dai (where i = 1 to 5), and luminance component memory "B" 40B includes data dbi (i = 1 to 1).
Write 5).

Next, a reference plate R is placed in the vicinity of the object,
Similarly, images are taken with five types of filters, and the data dwi (i = 1 to 5) is written in the luminance component memory "W" 40W. After that, in the correction circuit 42, for the class 1, the luminance component memory “A” 40A and the luminance component memory “W” 40
Brightness component memory "B" from W, and for class 2
40B and luminance component memory “W” 40W, data is read out and correction is performed. In this correction, the data of the luminance component memory "A" 40A is _first divided by the data of the luminance component memory "W" 40W for each filter component by the divider 42 _{1 according to the} following equation.

[0034] ^{da m i '= da m i} / dwi ( where, i = 1~5, m = 1~N ) db m i' = db m i / dwi ( where, i = 1~5, m = 1 ~ N) (5) a By this calculation, the influence of different light sources (spectral characteristics) can be removed. Here, i is a filter number and m is a sample number.

Further, in the power normalization circuit 42 ₂ , the power value of the divided data is made constant,
The following calculation is carried out by the power value Ca ^m and Cb ^{m. da m i "= da m i} '/ Ca m ( where, i = 1~5, m = 1~N ) db m i" = db m i' / Cb m ( where, i = 1~5, m = 1~N) ... (5) b, where the power value Ca ^m and Cb ^m is

[0036]

[Equation 1] Or

[0037]

[Equation 2] Is. By this power normalization, it is possible to remove the influence when the brightness of the light source is different.

Da ^m i "and d thus obtained
b ^m i "is written as a brightness spectrum in the brightness spectrum memories" dta "44A and" dtb "44B. The above correction is performed by the number N of samples of the object, and the brightness spectrum memories" dta "44A and" dtb "44B. The luminance spectrum is written into the luminance spectrum memory “dta” in this case, as the object sample, the object itself may be exchanged or a different region of the same object may be used. Into 44A and "dtb" 44B, the luminance spectrum data of the sample number N of the object is written.

If the objects of the two classes cannot be imaged at the same time, the object and the reference plate are imaged and corrected for each class in the same manner as described above, and the brightness spectrum memories "dta" 44A and "dta" 44A are respectively obtained. Write each luminance spectrum in dtb "44B. This operation is performed for the number of samples N.

Next, in the classification calculation section 32, the changeover switch "A" 46 is changed over to the b side. Then, the brightness spectrum data relating to class 1 and class 2 is read from the brightness spectrum memories “dta” 44A and “dtb” 44B, respectively, and the classification spectrum calculation unit 48 uses the above-mentioned FS conversion to classify the classification spectrum d1i (however, i = 1
˜5) and d2i (i = 1 to 5) orthogonal thereto are obtained, and the classified spectra d1i and d2i are written in the classified spectra d1 memory 50 and d2 memory 52, respectively.

Next, the changeover switch "A" 46 is set to the a side, and the changeover switch "C" 60 of the classification judging section 34 is set.
To the b side. And the changeover switch "B"
54 is switched to the side a, and the brightness spectrum memory "dt
The "luminance spectrum data da ^m i from 44A" a, also reads the classification spectrum data d1i from the classification spectrum d1 memory 50, the inner product calculated by the multiplier 56 and the data calculator 58

[0042]

[Equation 3] The classification boundary determining unit 62 of the classification determining unit 34
Transfer to. Then, the brightness spectrum memory "dtb" 4
4B to the luminance spectrum data db ^m i "and the classified spectrum d1 memory 50 to the classified spectrum data d1 i.
Read out and do the inner product

[0043]

[Equation 4] And transfers the result to the classification boundary determining unit 62.

Next, the selector switch "B" 54 is switched to the b side to read out the brightness spectrum data da ^m i "from the brightness spectrum memory" dta "44A and the classification spectrum data d2 i from the classification spectrum d2 memory 52. , Inner product operation

[0045]

[Equation 5] And transfers the result to the classification boundary determining unit 62. Subsequently, the luminance spectrum data db ^m i "from the luminance spectrum memory" dtb "44B, also reads the classification spectrum data d2i from the classification spectrum d2 memory 52,
Inner product operation

[0046]

[Equation 6] And transfers the result to the classification boundary determining unit 62.

In this way, each class is processed by the number of samples, and the product division values thus obtained are used to determine the classification boundaries as shown in FIG. Write.

The above is the learning mode. Next, the classification mode will be described. In this classification mode, first, an object O of an unknown class to be classified as shown in FIG. 8 is photographed as in the learning mode, and the brightness spectrum dxi (however, i is stored in the brightness component memory “A” 40A).
= 1 to 5) is written. Subsequently, the reference plate R is similarly photographed under the same photographing condition as this, and the luminance component memory “W” 4
The brightness spectrum dwi (where i = 1 to 5) is written in 0W. Then, the data is read from the luminance component memories “A” 40A and “W” 40W, and the correction circuit 42 is read.
Correction by dxi '= dxi / dwi (where i = 1 to 5) (10), and the power normalization circuit 42 ₂ further normalizes the power value of the divided data.

[0049]

[Equation 7] Then, the brightness spectrum dxi "is written in the brightness spectrum memory" dta "44A.

Here, in the classification calculation section 32, the changeover switch "A" 46 is changed over to the a side, and in the classification judgment section 34, the changeover switch "C" 60 is changed over to the a side.
Then, the changeover switch “B” 54 of the classification calculation unit 32
Is first switched to the a side, and the brightness spectrum memory "dt
a "44A reads the luminance spectrum dxi" and the classified spectrum d1 memory 50 reads the classified spectrum data d1i, and the integrator 56 and the accumulator 58 calculate the inner product.

[0051]

[Equation 8] Then, tx1 is transferred to the classification determining unit 66 of the classification determining unit 34.

Next, the changeover switch "B" 54 is switched to the b side to read the luminance spectrum dxi "from the luminance spectrum memory" dta "44A and the classified spectrum data d2i from the classified spectrum d2 memory 52,
Inner product operation

[0053]

[Equation 9] Then, tx2 is transferred to the classification determining unit 66.

Then, the classification determining unit 66 reads out the classification boundary from the classification boundary memory "cl" 64, and judges from this data which side of the transferred inner product values tx1 and tx2 is on the classification boundary. Then, the classification result is output.

The operation up to this point is the classification mode. As described above, in the present embodiment, since the difference in the spectral characteristics of the light sources is corrected by the divider 42 ₁ and the difference in the brightness is corrected by the power normalization circuit 42 ₂ , good classification is performed for different light sources. It can be carried out. In this case, FIG.
Since the power normalizing circuit 42 ₂ is used as shown in (B) of No. ₂ , good classification can be performed even when the luminance of the light source changes. When the spectrum of the light source does not change but only the brightness changes, the division circuit 42 ₁ is unnecessary and only the power normalization circuit 42 ₂ is needed.

Since the rotary color filter 12 is used in a simple structure, it is inexpensive and robust against mechanical vibration.
Further, since it has the learning mode and the classification mode, it is possible to easily cope with different classification purposes.

Further, as shown in FIG. 9A, the classification operation section 32 is operated by the classification spectra d1i, d which have already been learned.
Classification spectra d1 and d2 memories 50 and 5 for storing 2i
A plurality of pairs of two pairs and a changeover switch "B" 54 for selecting them are provided, and different learned classification spectra are stored in the classification spectra d1 and d2 memories 50 and 52 of each set, respectively. By using the changeover switch “C ′” 68 for selecting a set, different classification purposes can be instantly dealt with.
Of course, the rotary filter 12 may be replaced.

In the present embodiment, the rotary color filter 12 is a circular filter 12 as shown in FIG.
A to 12E are arranged on the same circle, and the position of each filter is controlled so that each filter can be stopped. However, as shown in (C) of FIG.
By using the rotary color filter 12 in which the filters 12A to 12E are formed in an arc shape and arranged on the same circle,
There is no need to stop and control the position of each filter,
Since you can keep moving, you can perform classification processing faster. However, as a matter of course, in this case, it is necessary to synchronize the exposure timing of the CCD 14 and the rotation speed of the rotary color filter 12.

In this embodiment, a CCD is used as the image pickup device.
Although only one 14 is used, it goes without saying that a plurality of image pickup devices 14 and the optical path dividing means 70 may be used as shown in FIG.

Further, the classified result may be displayed as an image of a different color according to the classified class, or may be notified to the photographer by voice or the like. [Second Embodiment] Next, as a second embodiment of the present invention, a color classifying device for classifying multi-class (n-class) objects will be described. The color classification device of this embodiment is the same as the first embodiment described above.
The luminance component extraction unit 30 has the same configuration as that of the embodiment.
Only the internal structure of is different.

That is, as shown in FIG. 11, the luminance component extraction section 30 includes object measurement area extraction sections 36 _{1 to} 36 _n and 36 W for extracting the measurement area of the object O of the photographed image.
Luminance component averaging unit 38 for obtaining the average of the measured luminance components
_{1 to} 38 _n , 38 W and a brightness component memory "1" for writing the brightness component of the photographed class 1 or unknown class data
“40 ₁ , luminance component memory“ 2 ”for writing the luminance component of the captured class 2 data,” 40 ₂ , ..., Luminance component memory “n” for writing the luminance component of the captured class n data
"40 _n ", a brightness component memory "W" 40 W for writing the brightness component of the data of the photographed reference plate R. These brightness component memories "1" 40 ₁ to "n" 40 _n and brightness component memory The “W” 40W has a capacity capable of writing as many luminance components as the number (5) of filters forming the rotary color filter 12.

The brightness component extraction unit 30 further includes a correction circuit 42 for correcting the influence of the light source, a brightness spectrum memory "dt1" 44 ₁ for writing the corrected class 1 data, and a corrected class 2 data. Luminance spectrum memory “dt2” 44 ₂ for writing data, ..., Luminance spectrum memory “dt for writing corrected class n data
n "it has a 44 _n. These luminance spectrum memory" respectively _{dt1 "44 1 ~" dtn "} 44 n is the number of samples the luminance component (the luminance component only N of data to be captured filter number number of (Comprising data).

In the color classifying apparatus having such a configuration, similarly to the first embodiment described above, first, the learning mode for obtaining the classification boundary is executed, and then the classification for performing the color classification of the data of unknown class. Mode.

First, the learning mode will be described. For any two classes of objects in a multi-class object,
Imaging and correction are performed in the same manner as in the first embodiment, and the brightness spectra da ^m i ", db ^m i" are written in the brightness spectrum memories "dt1" 44 ₁ and "dt2" 44 ₂ . Then, as in the first embodiment, the classification spectrum d1i is obtained by using the FS conversion.
(However, i = 1 to 5) and d2i (i = 1 to 5) are obtained, and the classified spectra d1 and d2 memories 50 and 5 of FIG.
Write to 2. Subsequently, the remaining n-2 class objects are also photographed and corrected in the same manner as in the first embodiment, and the brightness spectrum ds is stored in the brightness spectrum memory “dts” 44 _s.
^m i "(where s = 3 to n, i = 1 to 5) is written. Here, s is a class number, m is a sample number, and i is a filter number. Luminance spectrum memory "dts" for objects up to
The classification boundary is found using the luminance spectrum ds ^m i "written in 44 _s and the classification spectra d1i and d2i. To find the classification boundary, the same operation as in the first embodiment is performed, and each class is classified. inner product of the classification spectra d1i the luminance spectrum ds ^m i "of the object ts1 (where, s =. 1 to
The inner product value ts2 (s = 1 to n) of n) and d2i is obtained and transferred to the classification boundary determining unit 62. Then, the transferred inner product value is determined by the classification boundary determining unit 62 as shown in FIG. 9B, and the classification boundary is determined and written in the classification boundary memory “cl” 64.

The above is the learning mode. Next, the classification mode will be described. That is, as in the first embodiment described above, the object O to be classified is photographed and corrected, and the brightness spectrum d is stored in the brightness spectrum memory “dt1” 44 _1.
Write xi. Then, similarly to the first embodiment,
The classification calculation unit 32 and the classification determination unit 34 perform classification determination.

The operation up to this point is called the classification mode. Even in the case of the above multiple classes, good classification can be performed using different light sources. In the second embodiment, the classification spectrum is obtained by FS conversion from the data of any two classes in the multi-class, but the classification spectrum may be obtained by using the information of all the classes. Good (this is described in more detail below as a third embodiment below).

The classification may be performed stepwise in several steps. For example, when classifying into 10 classes, first, classification is performed according to a classification spectrum that is divided into 2 classes of 5 classes, and a classification spectrum for further fine classification is selected and classified according to the classified result. You may do it. By carrying out in multiple stages in this way, the classification accuracy can be further improved.

[Third Embodiment] Next, as a third embodiment of the present invention, another color classifying apparatus for classifying multi-class (n-class) objects will be described. The color classification apparatus of this embodiment is different from the second embodiment in the configuration of the classification calculation unit 32.

That is, as shown in FIG. 12, the classification calculation section 32 includes a changeover switch "D" 72 and a conversion matrix calculation section 74 for obtaining a conversion matrix for projection onto the classification space.
And a conversion matrix memory 76 for writing the conversion matrix obtained by the conversion matrix calculation unit 74, and a conversion unit 78 for projecting into the classification space using this conversion matrix.

First, the classification method (H
TC) will be described. HTC (Hotelling trace cr
iterion) is a method for separating multi-class ones. Specifically, S1 when the reflection spectrum of an object is converted by the conversion matrix A is an interclass covariance matrix, and S2 is an intraclass covariance matrix. , HTC J = tr (S2 ^-1 S1) (13) (where tr (X) is the trace of the matrix X (sum of diagonal elements)) is to find a transformation matrix A that maximizes .

This A is obtained as follows. Here, dim is the m-th data (vector) of the i-th class, di is the average data (vector) of the i-th class, and d is the average data (vector) of all classes.
Further, I is a unit matrix, and Wi is a residual matrix (Wi = (di1− di di2− di) composed of data of the i-th class.
... dim- di)), Wt residual matrix consisting of all data (Wt = (d1 - d d2 - d ... dn -
d)), the eigenvalues matrix of eigenvector matrix of E (Wt ^t Wt), the Λ (Wt ^t Wt), matrix whitening S2 is U, eigenvector matrix of the V S2, and the eigenvalue matrix Γ to S2 .

First, the following eigenvalue problem is solved. In addition, S2
= Is a Wt Wt ^{t. S2 V = VΓ, V t V} = I ... (14) V t S2 V = Γ ... (15) Here, when multiplied by Wt from the left to the sides ^{(Wt t Wt) E = EΛ} ... (16) (Wt Wt ^t ) Wt E = (Wt E) Λ (17) By comparing this with the above equation (14), V = Wt E, Γ = Λ Next, S2 is whitened.

U ^t S2 U = I (18) However, U = Wt EΛ ^-1 (19) Similarly, for S1, U ^t S1 U = D (20) Then, the eigenvalue problem of D is solved.

[0074] ^{DΨ = ΨΘ Ψ t Ψ = I} ... (21) Ψ t DΨ = Θ ... (22) (Ψ t U t) S1 (ΨU) = Θ ... (23) which is, Z = (Ψ ^t U ^t ) X ... (24) is equivalent to S1 of Z when converted. By this conversion, S2 is whitened.

(U ^t Ψ ^t ) S 2 (U Ψ) = Ψ ^t I Ψ = Ψ ^t Ψ = I (25) A = Ψ ^t U ^t = Ψ ^t Λ ⁻¹ E ^t Wt ^t (26) For each class By multiplying the reflection spectrum of the target object by the conversion matrix A represented by the equation (26), it is possible to perform multi-class classification.

Next, color classification using the color classification device of the third embodiment will be described. Also in this embodiment, similarly to the first and second embodiments, the learning mode is first performed, and then the classification mode is performed.

First, the learning mode will be described. That is, multi-class objects are imaged and corrected in the same manner as in the first embodiment described above, and the brightness spectrum ds ^m i "(where s = 1 to n, i = 1 to 1) is stored in the brightness spectrum memory" dts "44 _s .
Write 5). Here, s is a class number, m is a sample number, and i is a filter number.

Next, the changeover switch "D" 72 of the classification operation unit 32 is changed over to the side b, the brightness spectrum ds ^m i "is read from the brightness spectrum memory" dts "44 _s , and the conversion matrix calculation unit 74 performs the above-mentioned operation. A conversion matrix is obtained using the HTC described above and written in the conversion matrix memory 76. This conversion matrix is referred to as A.

Next, the changeover switch "D" 72 is changed over to the a side and the changeover switch "C" 60 is changed over to the b side to read out the brightness spectrum ds ^m i "from the brightness spectrum memory" dts "44 _s. Conversion matrix memory 76
The conversion matrix A is read from the conversion matrix A, and the conversion unit 78 multiplies the respective luminance spectra ds ^m i ″ by the conversion matrix A.

[0080]

[Equation 10] In this multiplication, the number of dimensions of the luminance spectrum does not change,
In this case, it remains 5 dimensions, and this 5 dimensions data

[0081]

[Equation 11] Is transferred to the classification boundary determining unit 62 as an evaluation value for classification. In this equation (27), T represents transposition. The classification boundary determination unit 62 determines a classification boundary as shown in FIG. 9B and writes it in the classification boundary memory “cl” 64.

The above is the learning mode. Next, the classification mode will be described. Similar to the first embodiment described above, the object O to be classified is imaged and corrected, and the brightness spectrum dxi is written in the brightness spectrum memory “dt1” 44 ₁ . Then, similarly to the first embodiment, the changeover switch “D” 72 of the classification calculation section 32 is changed over to the a side, and the changeover switch “C” 60 of the classification determination section 34 is changed over to the a side to obtain the luminance spectrum. The brightness spectrum dxi "is read from the memory" dt1 "44 _{1, the} conversion matrix A is further read from the conversion matrix memory 76, and the brightness spectrum dxi" is multiplied by the conversion matrix A in the conversion unit 78, and the result (value) is obtained.
To the classification determining unit 66. That is

[0083]

[Equation 12] Become

[0084]

[Equation 13] Is sent to the classification determining unit 66. Then, the classification boundary is read out from the classification boundary memory “cl” 64, and the classification determination is made as to where the value transferred to the classification determination unit 66 is in the classification boundary of FIG. 9B.

The operation up to this point is referred to as a classification mode. As described above, by using HTC, it is possible to obtain a conversion matrix for optimal classification in the case of multiple classes.

Further, in the third embodiment, the classification is performed in five dimensions, but the dimensions may be appropriately reduced (for example, two dimensions) to determine the classification boundary. In the first to third embodiments described so far, the pass band of the band pass filter is set arbitrarily. In other words, you can use any filter that is already on the market,
The rotary color filter 12 can be manufactured at low cost.

[Fourth Embodiment] Next, the above-mentioned first to third
A fourth embodiment of the present invention will be described, which is different from the embodiment, in which a color classifying apparatus is configured to optimally obtain the band of the rotary color filter 12 according to an object to be classified.

The configuration of the color classification device of the fourth embodiment is the same as that of the first embodiment described above, and a processing method for calculating the optimum classification filter will be described here. FIG. 13A is a diagram showing a configuration of an apparatus for obtaining an optimization filter, which includes a spectrometer 80 for measuring a reflection spectrum of an object and a filter optimization calculation for finding an optimum filter. Circuit 82 and reference plate data memory 84
W, class 1 data memory 84A, class 2 data memory 84B, filter parameter memory 86A, classification spectrum d1 memory 86B, classification spectrum d2 memory 86C. In the fourth embodiment, the spectrum of the object O is photographed at fine intervals (for example, 5 nm) using the spectrometer 80, and a plurality of images (5
The optimum bandpass filter is determined.

First, as shown in FIG. 13A, the spectroscopic spectra of the reflection spectra of the object O of two classes and the reflection plate of the reference plate R of which the classes are known are photographed by the spectrometer 80,
Data memories 84A and 84B as data ds ^m (λ)
Write in. Here, s is a class number (s = 0 is a reference plate, s = 1 is class 1, s = 2 is class 2), and m is a sample number.

All the subsequent processing is performed in the filter optimizing operation circuit 82, which will be described with reference to the flow chart shown in FIG. First, the initial values of the parameters of the bandpass filter forming the rotary color filter 12 are set (step S1). This parameter is
As shown in FIG. 13B, the number of filters (k
Number of filters, the center wavelength (.lambda.i) of each filter, the bandwidth (wi) of each filter, and the transmittance (ti) of each filter. The total number of filters.times.3. Here, i is a filter number (i = 1 to k).

Then, first, one object O of class 1
The reflection spectrum d1 ¹ (λ) of is converted into data that has passed through a bandpass filter. Specifically, 1
About the first filter

[0092]

[Equation 14] Do. Then, the second sheet, the third sheet, ..., And the k-th sheet are carried out, and the above-mentioned operation is performed for the reflection spectrums of the other samples to obtain df1 ^m . Then, this operation is also performed for the reflection spectrum of the object O of class 2 and the reflection spectrum of the reference plate R, and df2 ^m
Find (i) and dfs (i). The above operation is the fourth
In the embodiment, this is called spectrum conversion (step S2).

The data df1 ^m (i) and df2 ^m (i) (where i = 1 to k, m = 1 to N) thus obtained are corrected with the data dfs (i) of the reference plate R ( Step S3),
FS conversion is performed to obtain a classification spectrum (step S4).

Then, from the obtained classification spectrum, Fi
The sher ratio is calculated to obtain the classification evaluation value C. Classification evaluation value C is S / N ratio of sensor other than this Fisher ratio
Or a value in consideration of the brightness of the object O, for example, the following formula.

[0095]

[Equation 15] Here, Ca, Cb and Cc are appropriate weighting factors.
Then, the classification evaluation value C and the filter parameter at that time are written in the filter parameter memory 86A (step S5).

Next, one of the parameters of the filter is changed (step S6), the spectrum conversion is performed again using the reflection spectrum of the object O (step S7), and the data of the reference plate R is used. Correct (Step S
8). Then, using this newly obtained data, F
S-transformation is performed to obtain the classification spectra d1 and d2 (step S9) and the classification evaluation value Cnew (step S9).
10).

At this time, the filter parameter memory 86
The classification evaluation value C written in A is read, and Cnew
And C are compared in size (step S11). As a result, if Cnew is larger than C, the classification evaluation value Cnew in the filter parameter memory 86A is updated to C, the filter parameters at this time are updated, and the values of the classification spectra d1 and d2 are respectively classified spectrum d1.
The data is written in the memory 86B and the classified spectrum d2 memory 86C (step S12). On the contrary, if Cnew is smaller than C, the parameter of the filter is returned to the state before the change (step S13).

Then, it is judged whether or not all the parameters of the filter have been changed (step S14). If the change has been completed, the processing is terminated, and if the change has not been completed, the procedure returns to the step S6. Change the parameter again.

The processing may be stopped when the classification evaluation value exceeds a predetermined value. As described above, the parameters of the bandpass filter of the rotary color filter 12 and the classification spectra d1 and d2 are obtained. Then, a bandpass filter having this characteristic is manufactured, and thereafter, the classification boundary can be obtained and the classification processing can be performed in the learning mode as in the first embodiment and the like. However, unlike the above-described first embodiment, the classification spectrum calculation unit 48 is unnecessary.

Further, in the above-mentioned flowchart, in step S11, the procedure simply proceeds to step S12 only when Cnew is larger than C. However, when Cnew is smaller than C, the procedure also proceeds to step S12 with a certain probability.
A global optimum value can be obtained by using the Simulated Anealing method.

Further, similar processing can be performed for classification of multiple classes (n classes). In this case, two arbitrary classes are selected and the FS conversion described above is used. Then, the classification evaluation value C is used to check the degree of separation between the two classes.
Fisher ratio, HTC indicating the separation degree of multi-class classification
The value, the S / N ratio snr of the sensor, the brightness df of the object, etc. are taken into consideration.

[0102]

[Equation 16] However, here, Ca, Cb, Cc, and Cd are appropriate weighting factors.

Further, instead of the FS conversion in which two arbitrary classes are selected, the conversion matrix as described in the third embodiment may be obtained and then the above classification evaluation value may be used. [Fifth Embodiment] Next, a fifth embodiment of the present invention will be described.
In the above-described fourth embodiment, the transmittance (t
i) was also optimized as a parameter. However, in consideration of the dynamic range at the time of shooting, it is preferable to determine the transmittance according to the brightness of the object. For example,
The dynamic range becomes larger and the S / N at the time of image pickup is improved by decreasing the transmittance in the band where the brightness of the object is high and increasing the transmittance in the band where the brightness of the object is low. Therefore, the value is set as shown in the following equation.

[0104]

[Equation 17] Where C is a constant, Wi is the bandwidth of the filter, and Lia
ve is the average brightness of the object in that band. This is schematically shown in FIG.

The exposure time may be changed as another means for improving the dynamic range. In this case, the shutter speed of the CCD is changed or the aperture of the optical system 10 is changed every time an image of each filter is captured.

When the rotary color filter 12 as shown in FIG. 2C is used, the size of the light-shielding portion (that is, each filter) is changed according to the exposure amount as shown in FIG. It may be changed by changing.

[Sixth Embodiment] The sixth embodiment of the present invention will be described below. Although the above-described embodiments deal with changing the light source, in the case of a limited light source,
Optimization can be performed more easily. Therefore, a color classifying device for classifying two types of light sources when the light source is limited will be described as a sixth embodiment of the present invention.

That is, when the light source is limited, since the correction by the reference plate R is unnecessary, the FS conversion can be performed in the dimension photographed by the spectrometer to calculate the ideal classification spectrum. A practical filter with a small number of dimensions can be calculated. FIG. 17A shows a configuration diagram up to the calculation of the entire filter.

The classification spectrum calculation unit 48 is similar to that of the first embodiment described above, and calculates the classification spectra d1 and d2 by FS conversion. However, the number of dimensions is only the number of dimensions of the spectrum photographed by the spectrometer 80, and in the filter calculation circuit 88, the classified spectra d1, d
Multiple bandpass filters to approximate 2 are calculated. For this calculation, the sequential processing as described in the fourth embodiment can be used. For example, the classification spectra d1 and d2 to be realized are approximated as a linear combination of appropriate bandpass filters. Classification spectrum d1, d
The spectra d1 'and d2' approximating 2 are d1 '= U1 (λ) + U2 (λ) + U3 (λ) + ... + Um (λ) d2' = V1 (λ) + V2 (λ) + V3 (λ) + ... + Vn (Λ) (33) As shown in the bandpass filter Ui (λ) (where i =
1 to m), Vj (λ) (where j = 1 to n),
Ui (λ), Vj that minimizes the next evaluation values C1 and C2
(Λ) is sequentially obtained.

C1 = | d1-d1 '| ² C2 = | d2-d2' | ² (34) As shown in the fourth embodiment, the evaluation value is further
The Fisher Ratio and S / N may be taken into consideration.

In this way, the classification spectra d1 and d
A plurality of bandpass filters that should approximate 2 are calculated, and as shown in FIG.
Configured as.

The classification operation circuit 28 is constructed as shown in FIG. First, since the luminance component extraction unit 30 does not need to correct the light source, the memories 40A and 40B and the correction circuit 42 are used.
, And the luminance spectrum is directly stored in the luminance spectrum memories “dta” 44A and “dtb” 44B. The classification calculation unit 32 is a simple cumulative addition circuit, and the cumulative addition values t1 and t2 are sent to the classification determination unit 34. The classification determination unit 34 has the same configuration as that of the first embodiment.

When the light source is limited as in the sixth embodiment, the circuit configuration becomes very simple as compared with the case where the light source is corrected. When the number of types of light sources is limited to several, a light source detection device 92 for detecting the light source 90 is provided as shown in FIG. 19, and the dedicated rotary color filter 12 is provided according to the type of the light source 90. A filter exchanging section 94 for exchanging the filter may be provided.

[Seventh Embodiment] Next, a seventh embodiment of the present invention will be described in which a neural network is used in the classification judging section 34 of FIG. 13 so as to favorably perform complicated classification judgment. The classification determination unit 34 outputs the class number of the object from the inner product values t1 and t2, and is shown in FIG.
As shown in FIG. 2, it can be configured by a two-input, one-output neural network 96 and a learning device 98 for learning the neural network 96.

In this neural network 96, the values of the inner product values t1 and t2 are the units 100A and 10 of the input layer.
Enter in 0B. These input layer units 100A, 1
00B, the input signal as it is, the intermediate unit 10
Distribute to 2A-102N. Intermediate unit 102A-
The unit 102N and the output layer unit 104 have a plurality of input terminals and one output terminal.

Each unit executes the processing of the following equation. y = f (Σw _i x _i + θ) (35) where x _i is the input to the terminal i, w _i is the weighting coefficient of the terminal i, and θ is the bias value. Further, f is a sigmoid function represented by the following equation.

F (x) = 1 / (1 + e ^−x ) ... (36) This neural network 96 has a weight memory 10
6, the weighting factors and bias values necessary for each class determination are the units 102A to 10A of the intermediate layer and the output layer.
2N and 104.

Then, in the learning mode, the generalized delta rule learning method devised by Lamelhardt et al. (“PDP model / cognitive science and search for neuron network” so that the class number is output from the neural network 96 by the learner 98). ”Chapter 8, DE Ramelhart, J. Am.
L. Learning is conducted by McClellan, PDP Research Group, translated by Shunichi Amari, Sangyo Tosho, 1989).

By using this neural network, classification judgment can be easily performed even when the classification boundary is complicated or when there are multiple classes. Further, the classification currently performed by humans can be easily realized by using a neural network.

When learning is carried out by the neural network, a preprocessing circuit for removing data not suitable for learning may be provided. In addition, because the neural network takes time to execute,
As shown in (B) of 0, a 2-input 1-output classification table 108 is created by using the neural network 96 obtained by learning, and the classification table 10 is executed at the time of execution.
8 can also be used. In this case, the table creation circuit 110 for creating the classification table 108 is used.

Further, although learning is performed again for different classification purposes, in this case, only the value of the weight memory 106 is different. Therefore, the learning may be performed for each classification purpose and stored in different weight memories, and at the time of execution, the weight memory or the classification table may be selected according to the classification purpose.

[Eighth Embodiment] In the above-described embodiments, the color classification device mainly performs the classification by the rotary color filter 12. In this case, when the object O moves, the image is picked up by each filter. The image is misaligned. In the first embodiment, the luminance component extraction unit 30 is configured to obtain the average of a certain range. Therefore, if this deviation is smaller than the average size, there is no practical problem. However, this deviation becomes a problem when an image of an object that moves at high speed is taken or when the object is small and a sufficient average range cannot be secured.
Therefore, a method of eliminating the influence of this deviation will be described as an eighth embodiment of the present invention.

That is, as shown in FIG. 21A, the luminance component extraction unit 30 has mask circuits 112A, 112B, 1
12W is provided. This mask circuit 112A, 112B,
As shown in FIG. 21 (B), 112W is one type of filter whose amplitude becomes smaller as it gets closer to the periphery of the two-dimensional image. When a shift occurs in the image, a different pattern enters from the surroundings in the peripheral portion of the image, and thus such a masking process can reduce the influence of the shift. As this mask, for example, an approximation of a Gaussian function is used. This method can obtain a great effect with simple processing.

Further, as shown in FIG. 22A, the image shift correction circuit 114 is used. This image shift correction circuit 11
Reference numeral 4 is for correcting the deviation of the images picked up by the filters 12A to 12E, and is specifically configured as shown in FIG. That is, the image shift correction circuit 11
4 is composed of a frame memory 116, a reference area memory 118 for storing a reference image used for correlation calculation, a correlation circuit 120 for performing correlation calculation, and a read control circuit 122. In such a configuration, the image captured through the filter is first stored in the frame memory 116, and a part thereof is stored in the reference area memory 118. Then, an image obtained through another subsequently captured image is stored in the frame memory 116, and the correlation circuit 120 performs a correlation calculation with the reference image stored in the reference area memory 118. This correlation calculation is performed by a method of comparing the sums of the absolute values of the differences or the like, and the amount of deviation between two consecutive filtered images is obtained. Then, based on this shift amount, the read control circuit 122 controls the read position from the frame memory so as to match the first captured image,
Misalignment between images can be corrected. This method is effective even when the object is small.

Further, by dividing the image into fine blocks, obtaining the shift amount for each block, and correcting this,
Even if one part in the image moves, the deviation can be corrected.

[0126]

As described in detail above, according to the present invention,
(EN) Provided is a color classification device which has a simple device configuration, is low in cost, can withstand mechanical vibration, etc., and can perform good color classification even when the spectrum changes without limiting the light source. it can.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration of a color classification device according to a first embodiment of the present invention.

FIG. 2A is a diagram showing characteristics of a plurality of bandpass filters used in a rotating color filter used in the color classification apparatus of the first embodiment, and FIGS. 2B and 2C are rotation diagrams. It is a figure which shows the structure of a color filter.

FIG. 3 is a block configuration diagram of a classification calculation circuit in FIG.

FIG. 4 is a diagram showing a configuration of a luminance component extraction unit in FIG.

5A and 5B are diagrams showing the configuration of the correction circuit in FIG. 4, respectively, and FIG. 5C is a diagram showing the configuration of a classification calculation unit and a classification determination unit in FIG.

FIG. 6 is a diagram showing objects of two classes.

FIG. 7 is a diagram showing classification boundaries determined in a learning mode by the color classification device of the first embodiment.

FIG. 8 is a diagram showing objects of unknown class to be classified.

FIG. 9A is a diagram showing another configuration example of the classification calculation section in the first embodiment, and FIG. 9B is determined in the color classification device of the second embodiment of the present invention. It is a figure which shows a classification boundary.

FIG. 10 is a diagram showing a modification of the color classification device according to the first embodiment.

FIG. 11 is a diagram showing a configuration of a luminance component extraction unit in the color classification device according to the second embodiment of the present invention.

FIG. 12 is a diagram showing configurations of a luminance component extraction unit and a classification determination unit in a color classification device according to a third embodiment of the present invention.

FIG. 13A is a diagram showing a configuration of an apparatus for obtaining an optimization filter in the fourth embodiment of the present invention,
FIG. 7B is a diagram relating to parameters of filters that form a rotary color filter.

14 is an operation flowchart of the filter optimization arithmetic circuit in FIG. 13 (A).

FIG. 15 is a diagram for explaining how to determine the transmittance of each filter in the fifth embodiment of the present invention.

FIG. 16 is a diagram showing a configuration of a rotary color filter used in a fifth embodiment.

FIG. 17A is a diagram showing a configuration of an apparatus for obtaining an optimization filter in the sixth embodiment of the present invention,
(B) is a diagram showing a configuration of a rotary color filter used in the sixth embodiment.

FIG. 18 is a diagram showing the configurations of a luminance component extraction unit, a classification calculation unit, and a classification determination unit in the color classification device of the sixth embodiment.

FIG. 19 is a diagram showing a modification of the color classification device of the sixth embodiment.

20A and 20B are diagrams showing a configuration example of a classification determination unit in the color classification device of the seventh embodiment of the present invention.

FIG. 21A is a diagram showing the configuration of a luminance component extraction unit in the color classification device of the eighth embodiment of the present invention, and FIG. 21B shows the characteristics of each mask circuit in FIG. It is a figure.

22A is a diagram showing a configuration of a color classification device of an eighth embodiment, and FIG. 22B is a diagram showing a configuration of an image shift correction circuit in FIG.

FIG. 23 is a diagram showing a configuration of a filter having a characteristic of a classification spectrum in a conventional color discrimination device.

FIG. 24 is a diagram for explaining a classification spectrum.

[Explanation of symbols]

10 ... Optical system, 12 ... Rotation color filter, 12A to 12E
... band pass filter, 14 ... CCD, 16 ... A / D converter, 18, 116 ... frame memory, 20 ... monitor,
22 ... CCD drive driver, 24 ... Rotation color filter drive motor, 26 ... Control section, 28 ... Sorting arithmetic circuit,
30 ... Luminance component extraction unit, 32 ... Classification operation unit, 34 ... Classification determination unit 34, 36A, 36B, 36W, 36 _{1 to} 36 _n
... measurement area extraction unit, 38A, 38B, 38W, 38 1 ~
38 _n ... Luminance component averaging unit, 40A, 40B, 40W,
40 ₁ to 40 _n ... Luminance component memory, 42 ... Correction circuit, 4
2 ₁ ... Divider, 42 ₂ ... Power normalization circuit, 44A, 4
4B, 44 _{1 to} 44 _n ... Luminance spectrum memory, 46,
54, 60, 68, 72 ... Changeover switch, 48 ... Classification spectrum calculation unit, 50, 52, 86B, 86C ... Classification spectrum memory, 56 ... Accumulator, 58 ... Cumulative calculation unit, 58A ... Adder, 58B ... Latch, 62 ... Classification boundary determining unit, 64 ... Classification boundary memory, 66 ... Classification determining unit, 7
0 ... Optical path splitting means, 74 ... Transform matrix calculating section, 76 ... Transform matrix memory, 78 ... Transform section, 80 ... Spectrometer, 82 ... Filter optimization arithmetic circuit, 84A, 84B, 84W ... Data memory, 86A ... Filter parameters Memory, 88 ... Filter arithmetic circuit, 90 ... Light source, 92 ... Light source detection device, 9
4 ... Filter exchange unit, 96 ... Neural network,
98 ... Learner, 100A, 100B, 102A-102
N, 104 ... Unit, 106 ... Weight memory, 108 ...
Classification table, 110 ... Table creation circuit, 112A,
112B, 112W ... Mask circuit, 114 ... Image shift correction circuit, 118 ... Reference area memory, 120 ... Correlation circuit, 122 ... Read control circuit, O ... Object, R ... Reference plate.

Claims (12)

[Claims] 1. An image pickup means for picking up a reflection spectrum of an object, a plurality of bandpass filters having different bands arranged between the object and the image pickup means, and an object picked up by the image pickup means. A color classification device, comprising: a classification means for calculating a classification spectrum for classification using a statistical method from the reflection spectrum of the object, and classifying the object using the classification spectrum. . 2. The color classification device according to claim 1, wherein the classification unit includes a normalization unit that normalizes spectral characteristics and luminance of a light source that illuminates the object. 3. The storage unit stores the reflection spectrum of an object of a known class imaged by the imaging unit, and the reflection spectrum of an object of a known class stored in the storage unit. It is provided with a calculating means for calculating the classification spectrum from the spectrum, and means for classifying an object whose class imaged by the imaging means is unknown by using the classification spectrum calculated by the calculating means. The color classification device according to claim 1. 4. The classifying means comprises a plurality of storage means for storing in advance reflection spectrums of different known classes, and a selection for appropriately selecting the storage means to output the stored reflection spectrums. An output unit, a calculation unit that calculates the classification spectrum from the reflection spectral spectrum output by the selection output unit, and a classification spectrum calculated by the calculation unit are used, and the class imaged by the imaging unit is unknown. The color classification device according to claim 1, further comprising: a means for classifying the objects. 5. The plurality of band pass filters are configured to have a band and an amplitude obtained from data obtained by imaging a reflection spectrum of an object of a known class in advance at predetermined wavelength intervals. The color classification device according to claim 1, wherein the color classification device is provided. 6. The color classification device according to claim 1, further comprising means for controlling the transmittance or the exposure time of each of the plurality of bandpass filters according to the brightness of the object. 7. When a light source for illuminating the object is limited, a reflection spectrum of the object whose class is known is imaged in advance, and a classification spectrum is calculated by statistically processing using Foley Sammon transformation. The color classification device according to claim 1, wherein the plurality of bandpass filters are configured to approximate the classification spectrum. 8. The plurality of bandpass filters comprises a plurality of sets of filter means each having a predetermined number of bandpass filters, and the color classification device selects a light source for illuminating the object and a type of the light source. A first light source detecting means for detecting, and one of the plurality of sets of filter means selectively arranged between the object and the image pickup means according to the type of the light source detected by the light source detecting means. Switching means and second switching means for appropriately inserting and arranging a predetermined number of bandpass filters of the filter means selectively arranged by the first switching means between the object and the image pickup means. The color classification device according to claim 1, further comprising: 9. The classification means calculates a classification spectrum for classification using a statistical method from the reflection spectrum of the object imaged by the imaging means, and calculates the classification spectrum and the reflection spectrum. The classification calculation means for calculating an inner product value, and the classification determination means including a neural network for classifying the object using the inner product value from the classification calculation means. Color classification device. 10. The classification determination means includes table creating means for creating a classification table by using a neural network learned by imaging an object whose class is known by the imaging means, and a class imaged by the imaging means. 10. The color classification device according to claim 9, further comprising: a unit that classifies an unknown object using a classification table created by the table creating unit. 11. The color classification device according to claim 1, further comprising a mask filter or an image shift correction circuit for correcting a shift between images picked up by the plurality of band pass filters. 12. The classification means, in classifying a plurality of classes, uses any two classes of the plurality of classes.
The color classification device according to claim 1, further comprising means for obtaining a classification spectrum by using the ley Sammon transform or the Hotelling trace criterion.