JPH09114980A - High speed quality evaluation system for extruded granulated goods - Google Patents

Abstract

PROBLEM TO BE SOLVED: To automatically evaluate the quality of noodle-like molded goods at a high speed and to classify the goods by evaluating the quality of the noodle-like molded goods extruded from grid-like holes from the surface shapes. SOLUTION: A hierarchizing means 3 generates hierarchical structures for respective spatial frequency bands different in spatial frequencies, by a two-dimensional discrete wavelet-conversion of the image-picked-up signal of a video camera 2. A background separation/feature extraction means 4 separates the background of the noodle-like goods 1 and extracts image data of the noodle- like goods 1. An outline extraction means 5 executes an inverse wavelet conversion based on the high-order layer of the hierarchical structures so as to extract the outline of the noodle-like goods 1. An image re-composting means 6 executes the inverse wavelet conversion to image elements in the extracted outline to the resolution of the lowest layer of the hierarchical structure so as to obtain reproduced image data obtained by detaching respective noodle-like goods 1. A quality estimation means 7 having a multilayered perceptron-type neural network evaluates the surface state of the noodle-like goods 1 on the re-constituted image.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high-speed quality product of an extrusion-granulated product for high-speed evaluation of the quality of the noodle-shaped product which is extruded from a lattice-like hole and granulated by cutting into pellets. Regarding the evaluation method.

[0002]

2. Description of the Related Art In the field of visual inspection for determining the quality of manufactured objects, when the shape of an object to be inspected is known, such as a rigid body, a so-called model is used because of its geometrical characteristics. Pattern classification is possible by matching or the like. However, when a non-rigid article whose shape is not always constant is to be inspected, it is difficult to use that shape as an inspection parameter, and it is necessary to use another method such as texture analysis.

Further, in this type of inspection field, it is a necessary condition that both rigid and non-rigid bodies can be processed in real time and at high speed without affecting the inspection system. In particular, in the noodle-shaped molded product extruded from the lattice-shaped holes, a non-rigid body having a substantially circular cross section is curved with an arbitrary curvature, and exists in a state of being in contact with each other in an arbitrary direction or overlapping each other.

In such quality inspection of the target object, a skilled worker evaluates the quality visually, that is, without contact. That is, the degree of quality of this kind of object has a strong correlation with the roughness of the surface, and in the above visual evaluation, the quality is estimated from this surface state. As described above, conventionally, the quality evaluation of the noodle-shaped molded product has been performed by a skilled worker who visually observes the surface state of the noodle-shaped molded product, and there is no known case where such a work is automated. On the other hand, in the field of image coding processing,
A wavelet transform capable of separating an image into frequency bands is known. In the field of the present invention, the wavelet transform is used to extract or enhance the surface state of a two-dimensional image. No scheme has been proposed. Further, a steerable filter is known as a device capable of extracting a high frequency component having a contour in an arbitrary direction with a small number of filters.
In the related art, there has not been proposed a method of separating objects that are in contact with each other at high speed by using a stationary filter.

Further, in order to inspect cracks in an object, a method has been proposed in which cracks are measured by a hierarchical method. This is for extracting various cracks by binarizing them, and is capable of accurately measuring the width of cracks and the like. Regarding the above wavelet conversion, S. Gee. Murray, "Theory For Multi-Resolution Signal Decomposition: The Wavelet Representation" (SG Ma
llet "A Theory for Multire
solution Signal Decomposi
tion: The Wavelet Represen
station ”IEEE Trans. Pttern
Anal. & Mach. Intel. , PAMI-
11, 7, pp. 674-693, July 198
9), and for the stationary filter, see W. tea. Freeman and E. H. Andersen "The Design and Youth of Stable Filter" (WT Freeman &
E. FIG. H. Andelson "The Design a
nd Use of Steerable Filter
r "(IEEE Trans. Pattern Ana
l. & Mach. Intel. , PAMI-13,
9, pp. 891-906, Sep. 1991), and Japanese Patent Application Laid-Open No. 3-138775 for further inspection of the cracks.
There is the disclosure of the publication.

[0006]

In the above-mentioned conventional techniques, a technique for evaluating the surface condition of a target object has not been developed, nor has a technique for separating contour data at high speed using a stable filter been developed. . Further, in the case where the crack is measured by layering the image data, the width of the crack can be accurately estimated, but the surface condition of the target object cannot be evaluated.

The present invention quantitatively evaluates the degree of smoothness of the surface of the target object, and judges the degree of cracks and irregularities on the surface of the target object on an object-by-object basis based on the operator's evaluation criteria learned in advance. However, it is possible to classify by quality. In particular, some noodle-shaped molded products extruded through lattice-shaped holes may appear to be clearly separated on the image, but they are in contact with each other, and are straight or have any curvature. Since it is generally curved at, and its size also varies, it is necessary to forcibly separate the object on the image.

Therefore, it is not possible to apply a model matching method for colliding this kind of object with a typical model. Although the quality of an object can be estimated from the surface state of the object, the processing using the high-pass filter that extracts cracks and the like takes time and is not suitable for high-speed processing.

[0009] Further, when the varieties to be evaluated have changed,
The filter coefficient has to be changed each time and the coping ability is lacking. Furthermore, in order to save processing time and suppress unnecessary noise, it is necessary to exclude the target object from outside the region, but in the conventional filter processing, processing time is saved and real space and frequency space are handled at the same time. Is also difficult.

The relationship between the quality of the target object and its surface state varies little by little depending on the type of object, and there is no linear relationship between the quality and the surface state. And, there is no formalization of this relationship, and it currently depends on the experience of the worker.
For this reason, it is not appropriate to apply the conventional linear pattern classification method to the quality of an object that is extruded and granulated from a lattice-shaped hole.

That is, the object of the present invention is to solve the above-mentioned problems of the prior art and to automatically evaluate the quality of noodle-shaped molded products extruded through holes such as a lattice at high speed by image processing and classify them. It provides a system.

[0012]

In order to achieve the above object, the present invention firstly takes an image of an object to be evaluated (noodle-shaped molded article) with a video camera, and the image data thereof is layered by wavelet conversion. Turn into. Background separation and feature extraction are performed based on the frequency components of the hierarchical images.

Further, the contour of the object is extracted based on the hierarchized upper layer image, the objects which are in contact with each other are separated on the image, and the brightness distribution of the feature-extracted image is obtained for each object. The quality is estimated based on this. That is, according to the first aspect of the invention, as shown in FIG. 1, a video camera 2 for capturing an image of the noodle-shaped molded product 1 and converting the surface state thereof into two-dimensional image data.
A layering means 3 for creating a layered structure for each spatial frequency band having different spatial frequencies by two-dimensional discrete wavelet transform of the image pickup signal of the video camera 2, and the noodle-shaped molded product 1 using the layered structure. 2. Background separation / feature extraction means 4 for separating the background of the noodle-shaped molded product 1 to extract the image data of the noodle-shaped molded product 1, and the noodle-shaped molding by performing an inverse wavelet transformation based on the upper layer of the layered structure. The contour extracting means 5 for extracting the contour of the product 1 and the pixels in the contour extracted by the contour extracting means 5 are subjected to the inverse wavelet transform up to the resolution of the lowermost layer of the hierarchical structure to obtain the noodle-shaped molded product 1. An image recombining means 6 for obtaining reproduced image data separated from each other, and a surface state of the noodle-shaped molded product 1 for the image reconstructed by the image recombining means 6 are evaluated. And a quality estimating unit 7 having a multilayer perceptron neural network, and evaluating the quality of the quality of noodle-like molded article is extruded from the grid-shaped hole from the surface state.

The object to be inspected in the above invention is
It is not limited to noodle-shaped molded products extruded from lattice-shaped holes, but the extruded noodle-shaped molded products are chopped into pellets, and other objects whose quality can be evaluated from the surface condition. Can be similarly applied to.

[0015]

BEST MODE FOR CARRYING OUT THE INVENTION A video camera 2 has a CCD image pickup device or an image pickup tube and another photoelectric conversion device, and is a noodle-shaped molded product 1 formed by extrusion through holes such as a lattice.
Is imaged and its surface state is converted into two-dimensional image data. The layering means 3 creates a layered structure for each spatial frequency band having different spatial frequencies by subjecting the image pickup signal of the video camera 2 to two-dimensional discrete wavelet conversion.

The layering by the wavelet transform is executed as follows. Here, for simplification of description, one-dimensional description will be given. That is, the input signal from the video camera 2 is i (x), and the output having the low frequency component is O
_L (x), if an output with a high-frequency component and _{O H (x), O L} (x), O H (x) is

[0017]

(Equation 1) ・ ・ ・ ・ ・ ・ ・ (1) However,

(Equation 2) Is the wavelet transform coefficient,

(Equation 3) Is the low frequency characteristic,

(Equation 4)

Has high frequency characteristics. FIG. 2 is a schematic diagram for explaining the structure of the layering means by the wavelet transform, where the number of layers is 3, 1 is the original image, and 2 is
a, 2b are horizontal down samplers, 3a, 3b, 3c,
3d is a vertical down sampler, 4 is the first layer, 5a, 5b
Is a horizontal down sampler, 6a, 6b, 6c and 6d are vertical down samplers, and 7 is a second layer.

In the figure, assuming that the number of pixels of the original image 1 is N, this original image 1 is horizontally downsampled by using low-frequency wavelet and high-frequency wavelet transform coefficients, respectively.
a, 2b, and then sampled each image in the same manner as in the vertical down samplers 3a, 3b.
The number of pixels is N / 2 by sampling with b, 3c, and 3d.
A first layer consisting of four images is obtained.

The horizontal low-frequency image and the vertical low-frequency image of the first layer image are wavelet-converted and passed through the horizontal downsamplers 5a and 5b. Similarly, the wavelet conversion is performed and the vertical downsampler is also subjected to the wavelet conversion. By sampling at 6a, 6b, 6c, and 6d, the second layer 7 composed of four images with the number of pixels of N / 2 ² is obtained.

FIG. 3 is a schematic diagram for explaining the structure of the image recombining means by the inverse wavelet transform, in which the vertical upsamplers 8a and 8b are used for the respective second layers 7 of the four images described in FIG. , 8c, 8d, and the inverse wavelet transform is performed with the low-frequency and high-frequency wavelet transform coefficients and added by the adders 9a, 9b.

The two inversely transformed sheets are further passed through horizontal upsamplers 10a and 10b, which are inversely wavelet transformed by wavelet transform coefficients of low frequency and high frequency and added by an adder 11 to be recombined. Obtain the layer 12. Then, the two images of the first layer of re-synthesis are sampled by the vertical upsamplers 13a, 13b, 13c, 13d and similarly inverse wavelet transformed, two images are added by the adders 14a, 14b.

The two converted images are sampled by the horizontal upsamplers 15a and 15b, and the two inverse wavelet-converted samples are added by the adder 16 and recombined image 1
Get 7. FIG. 4 is an explanatory diagram of frequency components included in each layer of the layered structure. The images obtained by the wavelet transform have independent frequency bands, and the number of layers with the original image as the lowest layer is In the case of 3, the frequency components of each image are as shown in the figure.

In the figure, HL ^j is the horizontal frequency component in the layer j in the high frequency region, the vertical frequency component in the low frequency region,
ω _x indicates a horizontal frequency band, and ω _y indicates a vertical frequency band.
The background separation / feature extraction means 4 separates the background of the noodle-shaped molded product 1 using the layered structure and extracts the image data of the noodle-shaped molded product 1. The highest resolution of the layered structure is the lowest resolution, and the lower resolution is the higher resolution. Further, the low frequency image in a certain layer J (corresponding to O _L in the equation (1)) does not have the high frequency component that the image in the layer lower than that layer has.

Therefore, when the images are recombined, the image corresponding to the required frequency component of the required layer J is validated,
Information other than the required frequency band is extracted by disabling the others. In the layer J where the cracks and irregularities on the surface of the target object are sufficiently attenuated in the low frequency image,
The low frequency image I _LL ^j (x, y) is set to I _LL ^j (x, y) = 0, when I _LL ^j (x, y) <TH (2) by the threshold value TH, and the lower order , I _LL ^j-1 (x, y) = 0, when I _LL ^j (x, y) <TH, and I _LL ^j (x / 2, y / 2) <TH (3 ).

By performing recomposition while performing this conversion,
Separates the surface of the target object from the background and shadow. Then the hierarchy
The image of j is associated with the frequency band included in the image by I
_HL ^j, I_LH ^j, I_HH ^j, I_LL ^jAnd

Bandwidths of each image ω _HL ^j , ω _LH ^j , ω _HH
^j and ω _LL ^j are horizontal frequency band ω _x and vertical frequency band ω
_{Using y} , ω _HL ^j = ((1/2) ^j π ≤ ω _x <(1/2) ^j-1 π, 0 ≤ ω _y <(1/2) ^j π) ω _LH ^j = (0 ≤ ω _x <(1/2) ^j π, (1/2) ^j π ≤ ω _y <(1/2) ^j-1 π) ω _HH ^j = ((1/2) ^j π ≤ ω _x <( 1/2) ^j-1 π, (1/2) ^j π ≤ ω _y <(1/2) ^j-1 π) ω _LL ^j = (0 ≤ ω _x <(1/2) ^j π, 0 ≤ ω _y <(1/2) ^j π) (4) Further, ω _LL ^j has a relationship of ω _LL ^j-1 = ω _HL ^j + ω _LH ^j + ω _HH ^j + ω _LL ^j (5) with the image on the upper layer.

Therefore, the spatial region to be removed of the image having the frequency band to be removed is set to 0, and the features due to cracks and irregularities are emphasized by recombining. The contour extracting means 5 extracts the contour of the noodle-shaped molded product 1 by performing the inverse wavelet transformation based on the upper layer of the hierarchical structure, and the image reconstructing means 6 is extracted by the contour extracting means 5. Inverse wavelet conversion is performed on the pixels within the contour to the resolution of the lowermost layer of the hierarchical structure to obtain reproduced image data in which each of the noodle-shaped molded products 1 is separated.

FIG. 5 is an explanatory view of a constitutional example of a stay rub filter, in which 20a to 20c have θ of 0 ° and 60, respectively.
Are two-dimensional filters of 20 ° and 120 °, 21a to 21c are multipliers, 22 is an adder, and 23 is a lookup table. Further, FIG. 6 is a conceptual diagram of a window and inclination estimation according to the inclination of the edge, and FIG. 7 is a conceptual diagram of a window perpendicular to the edge.

Here, a stable filter for emphasizing the edge in the direction θ is obtained by using the Nth derivative G _N of the Gaussian function G (x, y), [ _GN ⁽ θ ⁾ ] = [Σ ^(M) _{(r = 1)] k r} (θ) and ^{_{[G N (θr)] ····}} (6).

Here, [G _N ⁽ θ ⁾ ],
_{^{[Σ (M) (r =}} 1)], [G N (θr)] , respectively

(Equation 5) Means For example, in the case of N = 2, the stationary filter has the configuration shown in FIG. First, the following processing is performed in the layer J. [G _N ⁽ θ ⁾ ] and its Hilbert transform pair [H
_N ⁽ θ ⁾ ], the orientation θm (0 ° ≦ θm ≦ 360 °
The energy of (discretely selected value of) is calculated by the following equation.

[E ⁽ θ m ⁾ ] = ((I _LL ^J * [G ⁽ θ m ⁾ ]) ² + (I _LL ^J * [H ⁽ θ m ⁾ ] ² ) ^1/2 (7) Where * indicates convolution integral, [E ⁽ θm ⁾ ],
[G ⁽ θm ⁾ ] and [H ⁽ θm ⁾ ] are respectively

(Equation 6) Means

In the fan-shaped region WH having the central axis of the direction θm shown in FIG. 6, the energy concentration S (θm)
Is determined by the following equation. S (θm) = ∫ _WH [E ⁽ θm ⁾ ] dxdy (8) This θm at which S is the maximum is the estimated edge direction θe.

With respect to the direction θ⊥ perpendicular to θe, in the region WV having the direction θ⊥ shown in FIG.

The spatial phase [φ ⁽ θe ⁾ ] of each pixel is [φ ⁽ θe ⁾ ] = tan ⁻¹ (I * [H ⁽ θe ⁾ ] / I * [G ⁽ θe ⁾ ] ... ( 9) where [φ ⁽ θe ⁾ ] and [H ⁽ θe
⁾ ] And [G ⁽ θe ⁾ ] are respectively

(Equation 7) Means

The calculation result is χ / 2 rad. And 3χ / 2ra
d. If it is near, it is regarded as a step-like edge. Also, 0
rad. And χrad. If it is near, it is regarded as a pulse edge. At the edge point, if it is a stepped edge, it is a contour. If the edge is pulsed, [E ⁽ θe
⁾ ] Is equal to or greater than a certain threshold TH2, S (θm) is sufficiently large, and the distance to the nearest edge existing in the direction θ⊥ is L or more, the contour is defined. Here, L is a threshold value of the distance set according to the shape of the target.

Here, [E ⁽ θe ⁾ ] is

(Equation 8)

Means If there are contour points on both sides of a region WH with an arbitrary orientation θ centered on an arbitrary image and the average value of the orientations of the contours on each side is θ ± Δθ, the image is regarded as an edge. Regarding θ ⊥ perpendicular to the direction θ of this edge, the center edge is defined as the contour.

Next, a layer j (0≤j <
In J), the above ( ^{1) to} (3) are executed at the coordinates (x, y) where I _LL ^{j + 1} (x / 2, y / 2) is the contour. This is repeated until the lower layer (j = 0). The quality estimating means 7
The quality of the noodle-shaped molded product extruded from the lattice-shaped holes using the multilayer perceptron type neural network for evaluating the surface condition of the noodle-shaped granulated product 1 for the image reconstructed by the image reconstructing means 6 is good or bad. Is evaluated from its surface condition.

That is, the quality estimating means 7 uses a perceptron type neural network to estimate and classify the quality for each surface condition of the noodle-shaped molded product which is the target object. The input of the quality estimating means 7 is the luminance distribution for each object, and the output is the quality rank. First, the number (rank number) for classifying the quality is set in advance. Then, the input image frame is presented to an operator who has conventionally performed visual classification, and
Let the pixel average quality be classified.

Next, the average luminance distribution of the object area in one image is input, and this classification result is used as a teacher signal. Do this learning uniformly from the worst to the best. Using the learning result thus obtained, the quality of the noodle-shaped molded product extruded through the lattice-shaped holes is automatically evaluated based on its surface condition.

[0042]

Embodiments of the present invention will be described below in detail with reference to the drawings. In this embodiment, a high-speed quality evaluation system will be described in which an organic rubber chemical is used as a target object. This organic rubber chemical is extruded from a large number of lattice-shaped holes into a noodle-shaped molded product having a diameter of about 1 mm, which is magnified and imaged by a video camera.

FIGS. 8 and 9 are system configuration diagrams for explaining one embodiment of the high-speed quality evaluation system for extruded granulated products according to the present invention, both of which are divided for convenience of illustration and circle A in FIG. , Circle B and circle C are combined to form one drawing. 8 and 9
In the figure, 30 is a layering unit (wavelet layering unit) using a wavelet transform for layering image data of a target object captured by a video camera, 31 is a background separation / feature extraction unit, and 33 is an image. Resynthesis / luminance distribution calculator, 3
Reference numeral 4 is a neural network quality estimation unit.

The input image is decomposed into multiple resolutions by the wavelet conversion in the wavelet layering unit 30. Here, the input image has 512 × 512 pixels, and the highest layer has 64 × 64 pixels. Also, the resolution is 25
There are 6 gradations. The image signal enlarged and picked up by the video camera is first layered by the wavelet layering unit 30, but in this embodiment, the layers are three layers except the original image.

Each image I in each layer_HL ^j, I_LH ^j, I_HH ^jYou
And the top level I_LL ^jIs the background separation / feature extraction unit 31.
(J = 1 to 3). Background separation / feature extraction unit 31
Distinguishes the object from the background, and does not need to know the area of the object
Frequency components are removed. This result is mask bit MB
^j(MB^j _LL, MB ^j _HH, MB^j _LH, MB^j _HL, J = 1
Is output to the image recomposition / luminance distribution calculation unit as
You.

The contour extraction unit 32 extracts the contour of the object.
Is issued and the result is the mask bit MC^j(J = 1 to 3)
Is output to the image recomposition / luminance distribution calculation unit 33.
In the image recomposition / luminance distribution calculation unit 33, the mask bit MB
^jAnd mask bit MC ^jIf is "1"
Then, the corresponding frequency component of the corresponding pixel is removed
You.

Further, from the result obtained by this image recomposition, a frequency distribution with brightness as a variable is calculated. The calculation result is input to the neural network quality estimation unit 34,
The quality of the object is estimated from the relationship between the brightness distribution learned in advance and the quality. Hereinafter, each unit in the above system configuration will be further described. 10 and 11 are block diagrams of the background separation / feature extraction unit, where 31a and 31b are 1
Hierarchical recomposition unit, 31c, 31d, 31e are comparators, 31
f, 31g and 31h are AND circuits, 31i is a rejection band selector, and 31j to 31s are OR circuits.

In the same figure, inputting a layered image
Then, the one-layer recombining units 31a and 31b^j _LL(J ≧
1) is required. FIG. 12 shows an example of the configuration of the one-layer recomposition unit.
31d is a circuit diagram to be described.₁, 31d_ThreeIs horizontal h
Filter, 31d_Two, 31d_FourIs a horizontal g filter, 31
d_Five, 31d₆, 31d₁₁Is an adder, 31d₇, 31d
₈Is a horizontal upsampler, 31d ₉Is a vertical h fill
31d_TenIs a vertical g filter, 31d₁₂Is vertical up
It is a sampler.

In I ^J _LL , a pixel whose luminance is lower than the threshold value TH is recognized as a background, and a background flag B ^J _LL (J = 1 to 3) is set if it is a background even in a lower resolution image. This background flag B ^J _LL and band elimination selector 31i
From the band selected by the flag MB ^J _LL (J = 1 to
3) is set. In this case, the entire area of the image corresponding to the selected band is masked.

FIG. 13 is a circuit diagram for explaining an example of the structure of the contour extracting section, in which 32a, 32c and 32f are layer selectors.
32b is a one-layer recombining unit, 32d is a logical product circuit, and 32e.
Is a one-layer contour extraction unit, and 32g is a layer management unit that outputs a layer signal for selecting a layer. In this contour extracting unit, first, the hierarchy selectors 32a, 32c, 32f
Is the third layer in the layer selection signal Layer from the layer management unit 32g.
The layer is specified.

As a result, I ³ _ll , I ³ _LH , I ³ _Hl ,
I ³ _HH is input to the one-layer recomposition unit 32b. I ^j-1 _LL, that is, I ² _LL, is output from the one-layer re-synthesis unit 32b. On the other hand, the input of the layer selector 32c is always 1 in the current layer by the layer selection signal Layer, so that I ³
_ll is output as it is to the first-layer contour extraction unit 32e. 1
The extracted contour is output from the hierarchical contour extraction unit 32e as a flag MC ³ for each pixel.

When the processing of this layer is completed, the layer management unit 32g sends the layer selection signal Layer to the layer selectors 32a, 32a.
2c and 32f are applied to select the second layer. I ² _LH , I ² _Hl , I ² _HH and I ² _ll synthesized in the previous hierarchy are input to the input section of the one-layer re-synthesis section 32b. Further, only the pixels corresponding to the contours extracted in the previous layer are input to the one-layer contour extraction unit 32e through the AND circuit 32d.

Then, the processing up to the first layer is repeated in the same manner as described above, and the flags MC ² and MC ¹ are output. 14, FIG. 15 and FIG. 16 are circuit diagrams for explaining an example of the configuration of the one-layer contour extraction unit, which is divided into three for convenience of illustration. In each of the figures, 321a is a real part of the stationary filter, 321b is the same imaginary part, and 322a _1- .
322a _m is a square root arithmetic circuit of the square of the first input and the square of the second input, 322b _{1 to} 322b _m is a tan ⁻¹ arithmetic circuit of the first input and the second input, 323a and 323b are in the direction θ.
A selector 324, an average value calculation circuit of θ in the window WH, 325 a local maximum energy detection circuit in the window WV, 326 a phase determination circuit, 327a _{1 to} 327a.
₃ is a threshold value TH1, TH2, TH3 determination circuit, 3
28a ₁ and 328a ₂ are AND circuits, 328a ₃ is an OR circuit, 329 is an edge search circuit within the distance L, 330 is a contour detection circuit within the window WH, 331 is a center determination circuit within the window WV, and 332 is a logic The product circuit 333 is an OR circuit.

FIG. 17 is a circuit diagram for explaining an example of the structure of the real part of the stallable filter of the one-layer contour extraction circuit shown in FIG. In the figure, 3210a, 3210b,
3210c has orientations θ of 0 °, 60 °, and 120 °, respectively.
Two-dimensional filters 3211a, 3211b ... 32
Reference numeral 11m is a calculation unit, and its operation is as described with reference to FIGS. 5, 6, and 7.

FIG. 18 shows the structure of the image recomposition / luminance distribution calculation unit.
The drawings are 3310a, 3311a to 3311c,
3312a to 3312c and 3313a to 3313c are discussed.
A laminating circuit, 3314 is a layer selector, and 3315 is one layer recombining.
Completion unit, 3316 is hierarchy management unit, 3317 is brightness distribution calculation
Department. This image re-synthesis / luminance distribution calculation unit is hierarchical
Image and flag MB^J, MC ^JIs entered and the hierarchy pipe
Flag MB according to the layer selection signal from the processing unit 3316^J, M
C ^JThe set pixel has zero brightness.

First, the layer selector 3314 selects the third layer in response to the layer selection signal Layer from the layer management unit 3316, and the image of the third layer is input to the one-layer recomposition unit 3315. Next, the layer selector 3314 selects the second layer,
The image of the second layer is input to the one-layer recomposition unit 3315.
Hereinafter, the first layer is repeated until the first layer re-synthesizing unit 3315.
Output a feature extraction image.

This feature-extracted image is processed by the brightness distribution calculation unit 33.
It is input to 17, and the appearance frequency of each gradation of luminance 256 is calculated. The calculated result is given to the quality estimating means as a luminance distribution having gradation as a variable, and evaluated by the neuro network. Since the configuration and operation of the neuro network itself are known, detailed description thereof is omitted here. In the present embodiment, learning is performed by using the visual evaluation result by a skilled worker who was engaged in the quality evaluation of the noodle-shaped molded product as teacher data.

During operation of this system, the quality distribution is evaluated using the neural network constructed as a result of the above learning, with the luminance distribution output from the image recomposition / luminance distribution calculating section as an input. The evaluation operation is as described in the embodiment of the invention. FIG. 19 is a schematic diagram illustrating the overall structure of a high-speed quality evaluation method for extruded granules according to the present invention, in which 100 is an extrusion granulator, 101 is a raw material supply pipe, 102 is an additive A supply pipe, and 103 is Additive B
A supply pipe, 104 is a molded product conveying device, 105 is a video camera, 106 is a high-speed quality evaluation device, 107 is a fuzzy arithmetic device, 108 is a control panel, and 109 and 110 are electromagnetic valves.

The extrusion granulator 100 has a granulation mechanism for kneading raw materials and additives to form granules, and a large number of holes arranged in a lattice, and the kneaded product is extruded through the lattice holes. A noodle-shaped molded product is manufactured by. The noodle-shaped molded product extruded from the extrusion granulator 100 is conveyed to the next step by the granule conveying device 104. The molded product conveying device 104 has a function of shredding the noodle-shaped molded product into an appropriate length to form a pellet, or a function of aligning the noodle-shaped molded product or the pellet-shaped molded product under a video camera. It can also be done.

The video camera 105 is the molded product conveying apparatus 10.
The granulated product conveyed in 4 is enlarged and imaged, and its image signal (image data) is given to the high-speed quality evaluation device 106. The high-speed quality evaluation device 106 non-quantitatively estimates and executes a quality evaluation equivalent to the evaluation of a skilled worker in a non-contact manner based on the configuration described above. The quality evaluation value estimated by the high-speed quality evaluation device 106 is given to the fuzzy calculation device 107, and the flow rate of the additive is calculated by a so-called ambiguous process, and this is calculated.
8 is output.

The control board 108 controls the opening and closing of the electromagnetic valves 109 and 110 based on the input from the fuzzy arithmetic unit 107 to control the amount of the additive and also controls the kneading conditions of the extrusion granulator 100. The control panel 108 also controls the transfer of the granulated product transfer device 104. FIG. 20 is a photocopy diagram illustrating an example of the surface state of a good noodle-shaped molded product and an explanatory diagram of the luminance distribution obtained by image processing. (A) is a photocopy diagram of the noodle-shaped molded product. , (B) are the one luminance distribution chart.

As shown in (a) of the same figure, no cracks or irregularities were visually recognized on the surface of the noodle-shaped molded product evaluated as a good product, and the brightness obtained by the image processing is shown in (b). It is distributed as follows. FIG. 21 is a photocopy diagram illustrating an example of the surface state of a defective noodle-shaped molded product and an explanatory diagram of the luminance distribution obtained by image processing. (A) is a photocopy of the noodle-shaped molded product. FIG. 1B is a single luminance distribution chart.

As shown in FIG. 9A, many cracks and irregularities were visually recognized on the surface of the noodle-shaped molded product evaluated as a defective product, and the brightness obtained by the image processing was (b).
It is distributed as shown in. Based on such a difference in luminance distribution, the high-speed quality evaluation device 106 having the quality estimation means having the multilayer perceptron type neural network estimates the quality of the noodle-shaped molded product conveyed by the granulated product conveying device 104. .

As a result, quality evaluation can be performed at high speed without requiring an operator to perform quality evaluation, and the evaluation result can be fed back to the amount of additives supplied to the extrusion granulator or the granulation conditions. As a result, a high quality granulated product can be manufactured. Note that, in FIG. 19, the noodle-shaped molded product is imaged of its surface state while being conveyed by the molded product conveying device 104 or in a state where the conveyance is stopped, but instead of this, the extrusion granulator 100.
The field of view of the video camera may be directed to the outlet of the molded product, and the noodle-shaped molded product immediately after molding may be imaged.

Further, the present invention is not limited to the noodle-shaped molded product described in the above embodiment, but a noodle-shaped object is further cut into pellets, and the quality of quality is judged from the surface condition. It goes without saying that it can be applied to the evaluation of all possible objects.

[0066]

As described above, according to the present invention,
A worker who has quantitatively evaluated the degree of smoothness of the surface of the target object, especially the noodle-shaped object extruded and granulated from the lattice-shaped holes, and the degree of cracks and irregularities on the surface of the target object, which has been learned in advance in object units. It is possible to provide a high-speed quality evaluation method for extruded granulated products, which can be evaluated based on the evaluation standard of No. 1, and automatically evaluated and classified for each quality at high speed.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a basic configuration of a high-speed quality evaluation method for an extrusion granulated product according to the present invention.

FIG. 2 is a schematic diagram illustrating a configuration of a layering unit by wavelet conversion.

FIG. 3 is a schematic diagram illustrating a configuration of an image recombining unit by inverse wavelet transform.

FIG. 4 is an explanatory diagram of frequency components included in each layer of the hierarchical structure.

FIG. 5 is an explanatory diagram of a configuration example of a stallable filter.

FIG. 6 is a conceptual diagram of a window according to an edge inclination and inclination estimation.

FIG. 7 is a conceptual diagram of a window perpendicular to an edge.

FIG. 8 is a partial view of a system configuration for explaining an example of a high-speed quality evaluation method for an extrusion granulated product according to the present invention.

FIG. 9 is a partial view connected to FIG. 8 of a system configuration for explaining an example of a high-speed quality evaluation method for an extrusion granulated product according to the present invention.

FIG. 10 is a partial configuration diagram of a background separation / feature extraction unit.

FIG. 11 is a partial configuration diagram of the background separation / feature extraction unit connected to FIG. 10;

FIG. 12 is a circuit diagram illustrating a configuration example of a one-layer recomposition unit.

FIG. 13 is a circuit diagram illustrating a configuration example of a contour extraction unit.

FIG. 14 is a partial circuit diagram illustrating a configuration example of a one-layer contour extraction unit.

FIG. 15 is a diagram illustrating a configuration example of a one-layer contour extraction unit.
It is a partial circuit diagram connected to.

FIG. 16 is a diagram illustrating a configuration example of a one-layer contour extraction unit.
It is a partial circuit diagram connected to.

FIG. 17 is a circuit diagram illustrating a configuration example of a real part of a sustainable filter of the one-layer contour extraction circuit in FIG.

FIG. 18 is a configuration diagram of an image recomposition / luminance distribution calculation unit.

FIG. 19 is a schematic diagram illustrating the overall configuration of an example of a high-speed quality evaluation method for extruded granulated products according to the present invention.

FIG. 20 is a photocopy diagram illustrating an example of the surface state of a good noodle-shaped granulated product and an explanatory diagram of a luminance distribution obtained by image processing.

FIG. 21 is a photocopy diagram illustrating an example of a surface state of a defective noodle-shaped granulated product and an explanatory diagram of a luminance distribution obtained by image processing.

[Explanation of symbols]

 1 noodle-like granulated product 2 video camera 3 layering means 4 background separation / feature extraction means 5 contour extraction means 6 image recomposition means 7 quality estimation means

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Norihisa Taniguchi 1099-1 Kami Ube Sanmon, Ube City, Yamaguchi Prefecture (72) Inventor Toshiaki Sakai 11 316, Ohnokita, Hirabu-cho, Kumage-gun, Yamaguchi Prefecture

Claims (1)

[Claims] 1. A high-speed quality evaluation method for an extruded granulated product, which comprises evaluating the quality of the noodle-shaped molded product obtained by shredging into pellets after extrusion molding into noodles and granulating the product, from the surface condition thereof. A video camera that images a noodle-shaped molded product and converts the surface state of the noodle-shaped molded product into two-dimensional image data, and a two-dimensional discrete wavelet transform of the image signal of the video camera to create a hierarchical structure for each spatial frequency band with different spatial frequencies. And a background separation / feature extraction unit that separates the background of the noodle-shaped molded product by using the hierarchical structure and extracts image data of the noodle-shaped molded product, and an upper layer of the hierarchical structure. Contour extraction means for extracting the contour of the noodle-shaped granulated product 1 by performing an inverse wavelet transformation on the basis of Image recombining means for obtaining reconstructed image data by separating each of the noodle-shaped molded products by performing inverse wavelet conversion up to the resolution of the lowest layer of the hierarchical structure for pixels in the frame, and reconstructing by the image recombining means. A quality estimation means having a multilayer perceptron type neural network for evaluating the surface condition of the noodle-shaped molded product with respect to the formed image, and the quality of the noodle-shaped molded product extruded from the lattice-shaped holes is judged from the surface condition. A high-speed quality evaluation method for extruded granulated products characterized by evaluation.