JPH05280960A - Defect inspection device - Google Patents

Abstract

PURPOSE:To obtain an appropriate defect identifying algoris in a short time by learning with learned data used in a defect inspector device for identifying a category of a shape of the defect. CONSTITUTION:A defect inspector comprises a neural network which obtains characteristic amounts including at least length and width of a defect and identifies a category of a shape of the defect with the characteristic amounts as inputs. The neural network is designed in a three-layered structure, wherein the number of units on an input layer is matched with the number of dimensions of a characteristic vector having characteristic amounts containing at least length, width and area as components, the number of units on an intermediate layer is matched with the number of identified planes on a characteristic space, and the number of units on an output layer is matched with the number of categories of a shape, respectively. The neural network can be further used for identifying defect classes. An supervisor signal is provided with certainty obtained by a membership function, and the certainty determined by this membership function may be added to identification results.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a defect inspection apparatus for identifying the shape and grade of a defect by using an image signal obtained by scanning an inspection object.

[0002]

2. Description of the Related Art A defect inspection apparatus is known which optically scans the surface of a steel plate, plastic film, paper or the like to detect scratches on the surface or internal defects. Here, conventionally, an image signal obtained by scanning an inspection object is compared with a preset threshold value, and when the image signal exceeds or is smaller than this threshold value, it is determined to be a defect (for example, See Japanese Patent Application No. 3-60754, etc.).

There has also been proposed an apparatus capable of discriminating the shape and type of the detected defect. For example, a defect discrimination logic is constructed based on the feature amount obtained from the defect, and the defect type is discriminated by this.

[0004]

However, in this case, it is necessary for a human to construct a discriminant logic for obtaining a discriminant result expected from a defect feature amount by trial and error, and in order to construct an appropriate logic. Requires a lot of work.

[0005]

SUMMARY OF THE INVENTION It is therefore an object of the present invention to provide a defect inspection apparatus capable of obtaining an appropriate defect identification algorithm within a short time by performing learning using learning data. ..

[0006]

According to the present invention, an object of the present invention is to extract a region including a defect of an inspection target by using a differential image obtained by subjecting an image signal obtained by scanning the inspection target to spatial filtering processing including a differential filter. In the defect inspection apparatus for identifying the shape category, feature quantity measurement processing means for obtaining a feature quantity including at least the length, width, and area of the defect, and identifying the shape category of the defect by inputting the feature quantity It is achieved by a defect inspection device characterized by comprising a neural network.

That is, a neural network is used to identify the shape of a defect. Here, this neural network has a three-layer structure, and the number of units in the input layer is made equal to the number of dimensions of a feature vector having a feature amount including a length, width, and area as a component, and the number of units in the middle layer is characterized. The number of units in the output layer is made to coincide with the number of shape categories, and the number of identification planes in space is made to coincide.

Further, a neural network is used for classifying defects. Also in this case, the neural network is 3
It is a layered structure, and the number of units in the input layer is at least the dimension number of the feature vector that has the feature amount including defect area and concentration information as the component, and the number of units in the intermediate layer is the identification plane in the feature space of this feature vector. The number of units in the output layer matches the number of categories in the class.

The teacher signal given at the time of learning of the neural network may be given by the certainty factor obtained by the membership function, and the certainty factor determined by the membership function may be added to the identification result of the shape or grade category.

[0010]

FIG. 1 is a block diagram of a first embodiment of the present invention,
2 is a detailed block diagram of a part thereof, FIG. 3 is an explanatory diagram of an example of a method of determining the threshold TH1, FIG. 4 is an explanatory diagram of an example of a method of determining the threshold TH2, and FIG. 5 is an explanatory diagram of the threshold TH3.

In FIG. 1, reference numeral 10 is an inspection object such as a steel plate, paper, and a plastic film, and the inspection object 10 is sent from a supply roll 12 to a winding roll 14. The winding roll 14 is driven by a winding motor 16. While the inspection object 10 is being fed, the image on the surface is read by the image detecting means 18 of the flying spot type.

The image detecting means 18 is a laser light source 2.
The scanning beam L composed of the laser light emitted from 0 is scanned (main scanning) in the width direction of the inspection object 10 at a constant speed by the rotating mirror (polygonal mirror) 24 rotated by the motor 22 while the inspection object is inspected. The light reflected by the surface of 10 is received by a pair of light receivers 28 (28).
a, 28b) to receive light. That is, the light-receiving rod 26 is disposed close to and parallel to the main scanning line 30 of the scanning beam L, and when the reflected light is received, the light-receiving rod 26 is totally reflected by the inner surface of the light-receiving rod 26 and guided to both ends thereof, and the photomultiplier (photoelectron The amount of light received is detected by a light receiver 28 such as a multiplier.

The image signal output from each photodetector 28 is amplified by a preamplifier and a main amplifier, and waveform-shaped into analog image signals a ₁ and a ₂ . In each of the signals a ₁ and a ₂ , signals corresponding to successive different main scanning lines 30 appear at regular time intervals in the time axis direction.

The levels of the signals a ₁ and a ₂ change such that the level decreases when the scanning beam L is far from the scanning position on the main scanning line 30, and conversely the level increases when the scanning beam L approaches the scanning position. Therefore, in this embodiment, both signals a ₁ and a ₂
Are added by the adding means 32, and the influence due to the change of the scanning position on the main scanning line 30 is removed to become the signal a ₃ . The output level of the signal a ₃ is greatly curved due to a change in the incident angle of the scanning beam L on the surface of the inspection object 10, unevenness or texture of the surface of the inspection object 10, attenuation in the light guide rod 26, and the like. There is.

The added signal a ₃ is converted into a digital signal a ₄ in the A / D conversion means 34. Eg 2
It is converted into a density signal a ₄ of 56 gradations. Then, it is stored in a line memory (not shown).

The A / D-converted signal a ₄ is density-converted by the density converting means 36, is input to the differential filter 38, and is spatially filtered. The density converting means 36 converts the signal a ₄ into an appropriate density distribution within a range of 256 gradations by using a predetermined conversion table. The differential filter 38 superimposes a spatial filter having a predetermined weighting coefficient on, for example, a 3 × 3 pixel area centered on the pixel of interest, calculates the product of the corresponding pixels, and outputs the sum of these.

The density converting means 36 and the differential filter 38 have been described in detail in Japanese Patent Application No. 3-60754, and will not be repeated here. The density converting means 36 is omitted and the signal a ₄ is directly differentiated by the differential filter 38.
You may enter in. The signal a ₅ spatially filtered by the differential filter 38 has the low-frequency component in the signal a ₄ removed to enhance the contour of the defect. This signal a ₅ forms the differential image A ₀ .

Reference numeral 40 denotes a defective address detecting means, which compares the signal a ₅ with a threshold value TH0, and a ₅ > TH0.
Determining a defect at the time of (or a ₅ <TH0). When this defect is detected, the address Ad of this defect is obtained and stored. This address Ad is obtained from the rotation angle of the speed detector (pulse generator) 42 and the motor 22. The method of determining the signal a ₅ as a defect is not limited to the method described here.

The address Ad of the defect is obtained as described above. In this process, the presence or absence of the defect may be simply determined, so the threshold TH0 used here is higher than the threshold TH3 described later, that is, the signal level of the defect. It is set to a level close to to prevent erroneous detection due to noise.

[0020]

[Binarization processing means] Output signal a of the differential filter 38
_The image formed by _5, that is, the differential image A ₀ is input to the binarization processing means 42. As shown in FIG. 2, this means 42 obtains a third threshold TH3 from the first threshold TH1 and the second threshold TH2, and binarizes the signal a ₅ using the third threshold TH3. It is a thing.

The first threshold value TH1 is determined from the overall density distribution of this differential image, and can be determined using, for example, the density histogram shown in FIG. This Figure 3
If the distribution width of density is x, the frequency threshold N (TH1) is calculated by using N (TH1) = (total frequency) / f (x).

Where f (x) is a function of the distribution width x,
It is determined by a quadratic function of x or the like. The frequency threshold N (TH
If the density TH1 of the histogram 1) is used as the first threshold value, this is a threshold value that reflects the density distribution of the entire differential image. The value of N (TH1) may be obtained experimentally in advance and treated as a constant in the process of defect detection processing.

The second threshold value TH2 is obtained from the average density value of the area near the target pixel. For example, as shown in FIG. 4, the density d of a 3 × 3 pixel centered on the pixel of interest
Can be obtained by obtaining the total sum Σd _n of and dividing this by the number of pixels 9.

The third threshold TH3 is determined using these first and second thresholds TH1 and TH2. For example, with k as a constant, TH3 = TH1-k (TH1-TH2) is determined. This constant k is experimentally determined in advance in consideration of the degree of texture.

In the binarization processing means 42, the differential image A
The third threshold TH3 determined for each pixel of ₀ is set to the differential image A
By comparing with the density of each pixel of ₀ (comparing means 4
4) Obtain the binary image A ₁ . This binary image A ₁ is an image in which the area of the differential image A ₀ (signal a ₅ ) in which the pixel density is TH3 or more is, for example, “1”, and the other area is, for example, “0”. It includes not only the defective area but also the area determined to be defective due to noise.

The operation of the binarization processing means 42 will be further described with reference to FIG. FIG. 5A shows the overall threshold TH1.
On the other hand, the density average value TH2 in the vicinity of the pixel of interest is shown in an overlapping manner, and the difference between them (TH1−TH2) is shown. FIG. 5B shows a curve k obtained by multiplying this difference from TH1 by a constant k.
The TH3 obtained by subtracting (TH1-TH2) is shown. Figure 5
(C) shows the relationship between the differential image A _{0 and} this third threshold value TH3.

As is apparent from this (C), the third threshold value TH3 binarizing the differential image A ₀ is to follow the global change of the image density varies and minute of the differential image A ₀ The change in concentration can also be taken out. The constant k used here
Is closer to 0, the effect of TH1 is greater, and the closer to 1, the effect of TH2 is greater. Therefore, it is possible to change the binarization process depending on the process, such as setting k close to 0 in a dirty process and k close to 1 in a clean process.

[0028]

[Vertical Shrinkage Processing] In FIG. 2, reference numeral 46 is a vertical shrinkage processing means. This means 46 contracts the binary image A ₁ in the vertical direction, that is, the sub-scanning direction (the feeding direction of the inspection object 10),
The connection between the defective portion and the background noise is removed without impairing the shape of the linear defect in the vertical direction.

FIG. 6 is a diagram for explaining the algorithm of this means 46, and FIG. 7 is a diagram showing images before and after the processing.
In this processing, the upper and lower two neighboring pixels X of the processing target pixel X ₀ are
Using the values of ₁ and X ₂ , the processed pixel X ₀ ′ is converted into X ₀ ,
Set to 0 when either X ₁ or X ₂ is 0, and set to 1 when all are 1. Here, "1" indicates a defective pixel and "0"
Is a background pixel. This process is performed for all pixels in raster scan order. As a result, as shown in FIG. 7, only the vertically continuous region remains, and the horizontally extended region and dot noise are erased.

The vertical contraction processing may be omitted when the value of k of the binarization processing means is close to 0 or when many defects occur not only in the vertical direction but also in the horizontal direction. Further, in the above description, it is assumed that defects that are continuous in the vertical direction are left, but when defects that are continuous in the horizontal direction (main scanning direction) are left, the left and right sides of the upper and lower two neighborhoods X ₁ and X ₂ in FIG. Similar processing may be performed using pixels in the vicinity of two.

[0031]

[Defective Area Extracting Means] In FIGS. 1 and 2, reference numeral 48 is a defective area extracting means. This means 48 extracts a continuous area containing the address Ad of the defect obtained by the defect address detecting means 40 from the binary image A ₁ obtained by the binarization process. Then, the area of the binary image A ₁ which does not include the address Ad is removed because it is not a defect.

FIG. 8 is a diagram for explaining the algorithm of this process, and FIG. 9 is a diagram showing images before and after this process. This process is performed as follows. First, the following processing is performed with the pixel designated by the defective address Ad by the defective address detecting means 40, that is, the pixel (defective pixel) in which the defective signal is generated, as the point of interest.

The coordinates of this point of interest are recorded and the corresponding pixel is labeled. A defective pixel in the binary image A ₁ which is not yet labeled is found in the search order shown in FIG. When found, the point of interest is moved to the pixel and the process returns to. If not found, return the point of interest to the previous point of interest,
Return to.

The procedures from to are repeated until no unlabeled pixels are found in the area. This process is performed for all defective signals due to the address (Ad) of the area other than the already labeled area. This result 2
Of the plurality of areas of the value image A ₁ , all the areas including the defect signal can be extracted and the other areas can be removed (FIG. 9).

[0035]

[Shape Correction] Next, shape correction processing is performed by the shape correction means 50 (FIG. 2) on the region including the defect thus extracted. This process is performed in order to facilitate the process and improve the recognition accuracy when recognizing the defect shape, type, grade, etc. as described later on the extracted area.
Depending on the output format of the defect area extraction means 48, this means 5
The processing by 0 may be omitted.

The shape correction means 50 includes an expansion process 52, a filling process 54, and a contraction process 56 as shown in FIG. 10, 11 and 12 show expansion processing 52,
FIGS. 13A and 13B are explanatory diagrams of algorithms of the hole filling process 54 and the contraction process 56, and FIGS.

[0037]

[Expansion process] The expansion process 52 is a process of adjoining fragmented
Connect several defect areas or connect discontinuous contours
It is a thing. That is, as shown in FIG.
Element X ₀ For example, the pixel X in the vicinity of 8 around the₁ ~ X₈ With
From the relationship, the processed pixel X₀′, X₀ ~ X₈ Which of
1 when X is 1, X₀ ~ X₈ Changes to 0 when all are 0
Replace.

Here, "1" indicates a defective pixel,
“0” is a background pixel. This process is repeated for each pixel in raster scan order. It is desirable to repeat this process twice consecutively. As a result, the unprocessed image shown in FIG. 13A can be combined into one continuous area as shown in FIG. 13B.

[0039]

[Blank Filling Process] The hole filling process 54 fills a hole in an image when the hole is present in the image as shown in FIG. That is, as shown in FIG. 11, the same "0" as the background
The area formed by the pixels of # 1 and # 2 that is not continuous with the edge of the image is determined as a hole area, and this hole area is changed to the pixel of "1" which is the same as the defect. By this processing, the hole shown in FIG. 13B is filled in as shown in FIG.

[0040]

[Shrinking Process] The shrinking process 56 is to restore the image of the defect expanded and enlarged by the expansion process 52 to the original size. The algorithm of the processing is centered on the target pixel X ₀ as shown in FIG. 12, for example eight neighboring pixels X ₁ ~
From the relationship with X ₈ , the processed pixel X ₀ ′ is converted into X _{0 to} X ₈
Is converted to 0 when any one of them is 0, and is converted to 1 when all of X _{0 to} X ₈ are 1.

Here, "1" indicates a defective pixel and "0".
Indicates a background pixel. This process is repeated for each pixel in raster scan order. This process is repeated the same number of times as the expansion process 52, that is, twice. As a result, as shown in (D) of FIG. 13, an image (defect image A ₂ ) in which only an area (defect area) including a shape-corrected defect is extracted is output.

[0042]

[Image measurement processing] The image A ₂ in which only the defect area is extracted as described above is input to the image measurement processing means 58, divided into areas by the area division processing 60, and further, by the feature amount measurement processing 62. The feature amount of each divided defect is obtained.

The area division processing 60 is performed by, for example, the defect image A ₂
A plurality of defect areas are divided by examining the distribution of the defect areas by using the peripheral distribution information projected in the x-axis direction and the y-axis direction.

[0044]

[Feature amount measurement process] The feature amount measurement process 62 obtains a feature amount such as area, length, width, direction, density average value, polarity (positive or negative) for each of the divided defect areas. 14 and 15 show the concept of these feature quantities. 14
(A) indicates the area of the defect D, and the number of pixels included in the defect D is measured. The length is obtained from the absolute maximum length of the defect D as shown in FIG. The width is obtained from the maximum width in the direction perpendicular to the maximum length obtained in (B) as shown in FIG. 14 (C).

By comparing the vertical fillet diameter and the horizontal fillet diameter as shown in FIG. 14D, the direction is vertical if (vertical fillet diameter) ≧ (horizontal fillet diameter) (vertical fillet diameter) <(horizontal fillet diameter) If it is (fillet diameter), it is judged to be horizontal.

The density average value is obtained as follows. FIG. 15 shows the concept of measuring this average density value. First, the defect and the background are binarized to extract a defective region and fill it, that is, the defect image A ₂ is used as a mask image. Then, a logical operation is performed with the original image. As a result, a grayscale image is obtained in which only the defective portion is cut out from the original image. Here, as the original image, an image based on the output signal a4 of the A / D conversion 34 in FIG. 1 or a differential image A ₀ can be used. The defect density average value of this defect portion and the background density average value of the other background portions are obtained, and the difference between them is used as the density average value. That is, (density average value) = | (defect density average value) − (background density average value) |

For the polarity, the average density value of these defects and the background is used, and if (defect density average value) − (background density average value) ≧ 0, white, (defect density average value) − (background density average value) ) <0, it is determined to be black.

[0048]

[Image recognition processing] When the feature amount of the defect is obtained in this way, the shape of the defect is then obtained in the image recognition processing means 64, the grade of the defect is discriminated, and finally a comprehensive judgment is made for the defect.

[0049]

[Shape Recognition Processing] The shape of the defect is processed by the shape recognition processing 66. This processing 66 is performed in two steps such that the defect shape is roughly classified by the neural network recognition method, and further the defect shape is finely classified by the decision tree recognition method.

[0050]

[Neural Network] FIG. 16 is a diagram showing distribution of shape categories (points, lines, surfaces, etc.) in the feature space. Here, "area" and "width / length" are used as the feature quantities, and a two-dimensional feature space having these axes is determined.

On this feature space, categories of shapes to be identified from these are arranged. Then, an identification plane F _P that serves as a boundary between each category of points, planes, and lines is assumed. In this example, there are three identification planes F _P1 , F _P2 and F _P3 . Since these feature quantities are within the range of [0, 1.0], they are normalized by the normalization constant.

The neural network inputs two feature quantities, "area" and "width / length", and classifies three categories of points, planes, and lines by three identification planes F _P in the two-dimensional feature space. I do. Therefore, the neural network in this case has a three-layer structure of an input layer, an intermediate layer, and an output layer, as shown in FIG. 17, and sets the number of units in the input layer to the feature quantity, that is, the dimension number 2 of the feature space. Is the number of identification planes, and the number of units in the output layer is the number of categories of dots, lines, and planes.

The units of each layer have the input / output relationship shown in FIG. The inputs to this unit are (x ₁ , x ₂ ..., x
_{If n} ), the input / output operation follows the following equation. u = Σω _i x _i −θ (Σ is the sum of i = 1 to n) z = f (u) where ω _i is a weighting coefficient, θ is a threshold value, and f is a response function. Further, u is an internal potential (membrane potential).

The response function f (u) which is the output z is a binary model which allows two values of 0 and 1 (linear threshold element model).
However, in this embodiment, a continuous value model using the sigmoid (s-shaped) function shown in FIG. 19 is used. A typical sigmoid function is f (u) = 1 / {1 + exp (−2u / u ₀ )}. Where u ₀ is a constant.

The feature values of "area" and "width / length" are input to the two units of the input layer, respectively. These inputs operate according to the input / output relationship (FIG. 18) of the unit described above, and are processed toward the output layer through the intermediate layer. Since the three units in the output layer respectively correspond to the categories of dots, lines, and planes, the unit with the maximum output is selected and the category corresponding to this unit is used as the classification result. The value of the weighting coefficient ω _i between units is assumed to be obtained by learning in advance, and is set to the network in advance before this processing.

[0056]

[Decision tree] As described above, shape categories (dots, lines,
When the face) is identified, a decision tree is then used to subdivide each defect shape. This decision tree is, for example, the decision tree shown in (A), (B), and (C) of FIG. The feature amounts and classification categories used for fine classification for each defect shape are shown in the following table.

[0057]

[Table 1]

The threshold value preset in each node of the tree is compared with the feature quantity, and the course is selected according to the result. When the end is reached, the category associated therewith is used as the result of subclassification. Although the result does not relate to the shape category, it is desirable to add information on the polarity of the defect when the feature amount measuring unit 62 obtains the polarity of the defect.

[0059]

[Grade Recognition Processing] The grade of the defect is performed by the grade recognition processing 68. This processing 68 uses two feature quantities, for example, "area" and "density average value" of the defect for each detected defect shape, and uses a neural network to classify the defect into three levels: light, medium, and heavy. Can be divided into

FIG. 21 shows the categories of the grades in the feature space used for the neural network used here (light / medium /
FIG. 22 is a structural diagram of the neural network. This feature space is a two-dimensional space having "area" and "density average value" as axes.

The categories of the grade to be discriminated are arranged on this feature space, and the discrimination plane F _{G which} is the boundary of each category is arranged.
Suppose In this example, there are two identification planes F _G1 and F _G2 . Note that the feature amount used here is normalized by a normalization constant so as to be within the range of [0, 1.0].

The structure of the neural network for deriving such an identification plane F _G can be a three-layer structure as shown in FIG. Here, the number of units in the input layer is 2 corresponding to the dimension of the feature space, and the number of units in the intermediate layer is the identification plane F _G.
The number of units in the output layer is set to 3 in accordance with the category (light / medium / heavy).

Six types of neural networks are prepared. That is, combinations of shapes (points / lines / planes) classified by the shape recognition neural network (FIG. 17) in the shape recognition processing 66 and polarities (positive / negative) are prepared for each of six types. Therefore, based on the result of the large classification of shapes (FIG. 17) and the polarity, the neural network of the grade to be used is selected.

When the "area" and "density average value" are input to the selected neural network as the normalized feature amount, this input is subjected to predetermined processing by each unit of each layer and output. Each unit in the output layer is heavy / medium /
Since it corresponds to the three light categories, the category of the unit with the maximum output is selected as the grade determination result. The input / output operation of each unit is the same as that described in the neural network used for determining the shape, and only the weighting coefficient ω _i , the threshold θ, and the like are different, and thus the description thereof will not be repeated.

[0065]

[Learning of weighting factor ω _i ] Here, the shape recognition neural network used for large classification of shapes (FIG. 17) and the grade recognition neural network used for class determination (FIG. 2)
In 2), learning for obtaining a desired output with respect to an input will be described.

The learning data used here is formed, for example, in the case of shape classification, by a set of the characteristic amount obtained by the characteristic amount measuring process and the teacher signal of the classification category. Here, the teacher signal is a desired output that the neural network should output for a certain input, and is given from the outside.

The learning data for shape recognition can be in the following format, for example. Here, the feature quantity is normalized. {Input data} = {Area, width / length} {Teacher signal} = {0.98, 0.02, 0.02} ... In case of surface defect {Teacher signal} = {0.02, 0.98, 0.02} ... In case of line defect {Teacher signal} = {0.02, 0.02, 0.98} ... In case of point defect

The learning data for grade recognition can be determined as follows, for example. {Input data} = {Area, density average value} {Teacher signal} = {0.98, 0.02, 0.02} ... In case of heavy defect {Teacher signal} = {0.02, 0.98, 0 .02} ... Medium defect {Teacher signal} = {0.02, 0.02, 0.98} ... Light defect The values of 0.98 and 0.02 used for the teacher signal are
It is simply one of values close to 1 or close to 0, and is not limited to this value. As a learning method, the general error back propagation method (Error Back Propagation
Algorithm) is used.

When learning the neural network for shape recognition, input data and a teacher signal are created for an appropriate shape (pattern), and the output obtained from the output layer when this input data is input is used. The difference from the teacher signal is obtained for all units in the output layer. Then, the following error function E is defined by the sum of squared differences, and this is used for evaluation of learning. E = (1/2) Σ _i (T _i −O _i ) ²

Here, T _i represents the teacher signal for the i-th unit in the output layer, and O _i represents the output of each unit in the output layer. Since the output O _i is determined by the weight ω _i of the coupling between the units at that time, the error function E is also a function implicitly defined with respect to the weight ω _i . Therefore, if we consider a space that can use the value of each weight ω _i as an axis, and further consider the value defined by this error function E as the height, the error function E becomes an “error curved surface” as a hypercurved surface on this weight space.

In order to reach the minimum value of this error curved surface, it is sufficient to advance on this error curved surface in the steepest inclination direction (steepest descent method, gradient decent method). That is, this error function E
The gradient ω _i may be obtained and the weight ω _i may be corrected in that direction.

Grad E = (∂E / ∂ω ₁ , ∂E / ∂ω
₂ , ... ∂E / ∂ω _n ), each weight ω _i can be changed by Δω _i = −ε · ∂E / ∂ω _i (ε> 0). Thus ω _i
Is changed and the calculation is repeated until the value of the error function E becomes sufficiently small.

[0073]

[Addition of certainty factor by membership function] In the above description, the process of recognizing a large classification of defect shapes or a defect grade by a neural network was shown. The teacher signal T _i used for learning may be replaced with a certainty factor value obtained from the membership function. This membership function is a function that assigns one value between 0 and 1 to each unit in the output layer according to the degree of belonging to the category to which the unit corresponds.

FIG. 23 is a diagram showing an example of this membership function. For example, when considering a "surface-like" shape having a certain input data, the unit corresponding to the "surface" of the output layer has a certainty factor of "0.75" as the teacher signal T _i for the determination "it looks like a surface". Can be given to each of the other units, and a confidence level of “0.25” can be given to the other units. In this case, when the input data of the tally image is input, the output of each unit of the output layer indicates the certainty factor of each of these units.
Therefore, by writing the outputs of all the units in the output layer side by side as the output result, the certainty factor can be added to the recognition result.

[0075]

[Comprehensive Judgment Processing] When the shape and grade of the defect are judged in this way, then a comprehensive judgment processing 70 makes a comprehensive judgment. At this time, if the shapes of a plurality of defects are all in the same category, information to that effect is added. Further, when there are a plurality of defects having the same shape, the grade may be set higher than the grade for each single defect, so that the influence of the number of defects may be taken into consideration.

[0076]

As described above, according to the first aspect of the present invention, a neural network having a feature amount including at least a length, a width and an area of a defect as an input is formed to identify a defect shape category. Since the learning is performed, it is possible to save the trouble of constructing the discrimination logic by trial and error as in the conventional device, and it is possible to perform highly accurate identification in a short time.

If a neural network for determining the defect grade (heavy, medium, light, etc.) is input with the feature amount including at least the area of the defect and the density information, the defect grade can be determined with high accuracy. It is possible (Claim 4).

The neural network used for these has a three-layer structure having an input layer, an intermediate layer and an output layer,
The number of units in the input layer is the number of dimensions of the feature vector that has the input feature as a component, the number of units in the intermediate layer is the number of identification planes in the feature space, and the number of units in the output layer is the number of categories to be identified. Match each (claims 2 and 5).

Although it is necessary to learn the weighting coefficient ω _i in these neural networks, it is possible to give the certainty factor by the membership function as the teacher signal T _i used for this learning. That is, the certainty factor between 0 and 1 is applied to all units in the output layer. In this case, the certainty factor for the defect category to be recognized is output to each unit in the output layer. Therefore, by outputting the output of each of these units at the same time, the certainty factor can be added to the identification result (claims 3 and 6).

As the image information of the defect used for recognizing the shape of the defect, it is desirable to use a binarized image in which only a region including the defect is filled with a value indicating the defect (claim 7). That is, first, only the presence or absence of a defect is determined from the differential image to obtain its address, and only the area containing the address of the defect is applied to the defect value in the image obtained by binarizing the differential image with a predetermined threshold value. Crush. If the feature amount is obtained by using the image obtained by extracting the defect area in this way, the accuracy of recognizing the shape of the defect is improved.

[Brief description of drawings]

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

FIG. 2 is a detailed block diagram of a part thereof.

FIG. 3 is an explanatory diagram of an example of a method of determining a threshold TH1.

FIG. 4 is an explanatory diagram of an example of a method of determining a threshold TH2.

FIG. 5 is an explanatory diagram of a threshold TH3.

FIG. 6 is an explanatory diagram of a vertical contraction algorithm.

FIG. 7 is a conceptual explanatory diagram of vertical contraction.

FIG. 8 is an explanatory diagram of a region extraction algorithm.

FIG. 9 is an explanatory diagram of region extraction processing

FIG. 10 is an explanatory diagram of an expansion processing algorithm.

FIG. 11 is an explanatory diagram of a filling process.

FIG. 12 is an algorithm explanatory diagram of contraction processing.

FIG. 13 is a diagram showing a process of correcting the shape of a defect image.

FIG. 14 is a diagram showing a concept of a feature amount.

FIG. 15 is a diagram showing a method for measuring an average concentration value.

FIG. 16 is a diagram showing distribution of shape categories.

FIG. 17 is a structural diagram of a shape recognition neural network.

FIG. 18 is a diagram showing input / output relationships of units.

FIG. 19 is a diagram showing a sigmoid function as an example of a response function.

FIG. 20 is a diagram showing the structure of a decision tree.

FIG. 21 is a diagram showing distribution of grade categories.

FIG. 22 is a structural diagram of a grade recognition neural network.

FIG. 23 is a diagram showing an example of a membership function.

FIG. 24 is a diagram showing that a teacher signal is given with confidence.

[Explanation of symbols]

10 inspection object 38 differential filter 40 defect address detection means 42 binarization processing means 48 defect area extraction means 62 feature amount measurement means 66 shape recognition processing means 68 grade recognition processing means A ₀ differential image A ₁ binary image A ₂ defect image

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification code Internal reference number FI technical display location G06F 15/70 330 N 9071-5L

Claims (7)

[Claims] 1. A defect inspection for identifying a defect shape category by extracting a region including a defect of an inspection target by using a differential image obtained by performing spatial filtering processing including a differential filter on an image signal obtained by scanning the inspection target. The apparatus is provided with a feature quantity measurement processing means for obtaining a feature quantity including at least a length, a width and an area of the defect, and a neural network for identifying a category of a defect shape by using the feature quantity as an input. Defect inspection device. 2. The neural network has a three-layer structure including an input layer, an intermediate layer, and an output layer, and the unit number of the input layer is set to the dimension number of a feature vector having the feature quantity as a component. The number of layer units is made to correspond to the number of identification planes required for classifying the shape categories in the feature space formed by the feature vector, and the number of unit units of the output layer is made to correspond to the number of categories. The defect inspection device according to claim 1. 3. The teaching signal according to claim 2, wherein a teacher signal given at the time of learning of the neural network is given by a membership function, and the output value of each unit in the output layer is added to the classification result of the shape as a certainty factor. Defect inspection device. 4. A defect inspection apparatus for detecting a defect of an inspection object by using a differential image obtained by spatially filtering an image signal obtained by scanning the inspection object and including a differential filter, and identifying a category of the defect grade. A defect inspection comprising: a feature amount measurement processing means for obtaining a feature amount including at least the area of the defect and density information; and a neural network for identifying the category of the defect grade with the feature amount as an input. apparatus. 5. The neural network has a three-layer structure including an input layer, an intermediate layer, and an output layer, and the unit number of the input layer is set to the dimension number of a feature vector having the feature amount as a component. The number of layer units was made to correspond to the number of identification planes required to classify the category of the grade on the feature space formed by the feature vector, and the number of unit of the output layer was made to correspond to the number of categories. The defect inspection apparatus according to claim 4, wherein 6. The method according to claim 5, wherein a teacher signal given at the time of learning of the neural network is given by a membership function, and the output value of each unit in the output layer is added to the classification category discrimination result as a certainty factor. Defect inspection device. 7. The defect address detecting means for obtaining the address of a defect based on a differential image, the binarizing processing means for binarizing the differential image, and the binary value according to claim 1. A defect area extracting unit that fills all pixels in the area including the defect address in the image processed to a value indicating a defect, and identifies the category of the defect shape using the output of the defect area extracting unit. A defect inspection apparatus characterized by performing.