JP2013167596A - Defect inspection device, defect inspection method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a defect inspection device, a defect inspection method, and a program capable of properly determining the presence/absence of defects in an inspection object.SOLUTION: A defect inspection device includes: an area information obtaining portion 103 for performing binarization processing on image information of an inspection object, and obtaining area information of a defect candidate; a position calculating portion 105a for calculating a position of the defect candidate; an area calculating portion 104 for calculating an area of the defect candidate; a feature vector extracting area deciding portion 105b for deciding an area for extracting a feature vector of the defect candidate; a feature vector extracting portion 107 for extracting a feature vector of the detect candidate; a type discriminating portion 108 for discriminating a type of the defect candidate by comparing each feature vector extracted by the feature vector extracting portion 107 with a boundary condition given in advance; and a defect determining portion 110 for determining the presence/absence of a defect in the inspection object by comparing the type, position, and area of the defect candidate respectively with discrimination references given in advance.

Description

  The present invention relates to a defect inspection apparatus, a defect inspection method, and a program for inspecting whether there is a defect in an inspection object based on image information obtained from the inspection object.

A technique for inspecting whether or not a surface of an electrode is defective in a manufacturing process of a fuel cell, a lithium ion battery, or the like is known.
For example, Patent Document 1 describes a technique for acquiring a transmission image by irradiating an electrode substrate, which is an inspection object, with X-rays, and detecting a portion where metal impurity particles (defects) exist based on the transmission image. Has been.

  Patent Document 2 uses the first image data obtained by irradiating light from the upstream side of the electrode material being conveyed, and the second image data obtained by irradiating light from the downstream side, A technique for discriminating the shape of a foreign object (defect) is described.

JP 2006-179424 A Japanese Unexamined Patent Publication No. 2010-25566

  However, although the technique described in Patent Document 1 can detect the presence or absence of metal impurity particles, there is a problem that the shape of the metal impurity particles cannot be automatically identified. Moreover, even if a missing portion (a hollow portion) exists on the surface of the electrode material, there is a possibility that the missing portion cannot be detected in the X-ray transmission image.

  In the technique described in Patent Document 2, the total number of pixels in a range (outline) in which the luminance value is equal to or greater than a predetermined value is used in the total image data obtained by combining the first image data and the second image data. The shape of the foreign matter (defect) is discriminated. Therefore, it is only possible to estimate the rough shape of a foreign object, such as “circular” or “horse-shoe”.

  Generally, when detecting the presence or absence of a defect from image information of an electrode material that is an inspection object, it is necessary to appropriately determine the type of the defect. In addition to the above types, it is desirable to comprehensively determine the presence / absence of a defect in consideration of the position and area of the defect. In order to make such a detailed determination, conventionally, it has been necessary to finally determine the presence or absence of a defect by visually observing it, and therefore, a great deal of time and cost are required.

  Then, this invention makes it a subject to provide the defect inspection apparatus, the defect inspection method, and program which can determine appropriately the presence or absence of the defect of a test target object.

  In order to solve the above-described problem, a defect inspection apparatus according to the present invention is a defect inspection apparatus that inspects whether there is a defect in an inspection object based on image information of the inspection object, and is input from the outside. By performing binarization processing on the image information of the inspection object, the coordinate value of the pixel included in the defect candidate is extracted, and the labeling value, which is identification information for each defect candidate, is associated with the coordinate value. Area information acquisition means for acquiring area information of the defect candidates, and position calculation means for calculating the position of each defect candidate as a barycentric coordinate value based on the area information of the defect candidates acquired by the area information acquisition means; , An area calculation unit that calculates an area for each defect candidate based on the region information of the defect candidate acquired by the region information acquisition unit, and before the acquisition by the region information acquisition unit In the region determined by the feature vector extraction region determination unit, a feature vector extraction region determination unit that determines a region for extracting a feature vector for each defect candidate based on the region information of the defect candidate, and the binarization process A feature vector extracting unit that performs high-order local autocorrelation processing on the previous image information and extracts a feature vector for each defect candidate, and each feature vector extracted by the feature vector extracting unit is given in advance By comparing the boundary condition, a type determining unit that determines the type of the defect candidate, a type of defect candidate determined by the type determining unit, a position of the defect candidate calculated by the position calculating unit, Each of the defect candidate areas calculated by the area calculating means is compared with a predetermined criterion. Characterized in that it comprises a defect determination unit determines the presence or absence of a defect of the inspection object by.

According to such a configuration, the area information acquisition unit extracts the coordinate value of the pixel included in the defect candidate by performing binarization processing on the image information of the inspection object input from the outside. Then, area information of defect candidates is obtained by associating a labeling value, which is identification information for each defect candidate, with the coordinate value (area information acquisition step). Then, the feature vector extraction region determination unit determines a region for extracting a feature vector for each defect candidate based on the defect candidate region information acquired by the region information acquisition unit (a feature vector extraction region determination step).
Further, the feature vector extraction unit performs high-order local autocorrelation processing on the image information before the binarization processing in the region determined by the feature vector extraction region determination unit, and obtains a feature vector for each defect candidate. Extract (feature vector extraction step).

Therefore, the area information acquisition means performs binarization processing and associates the labeling value and the coordinate value for each defect candidate, so that the area information of the defect candidate can be acquired easily and accurately.
In addition, since the area from which the feature vector extraction unit extracts the feature vector of the defect candidate is not the entire image of the inspection object but a predetermined region determined by the feature vector extraction region determination unit, the amount of calculation in the feature vector extraction unit Can be greatly reduced.

Further, the type discriminating unit discriminates the type of the defect candidate by comparing each feature vector extracted by the feature vector extracting unit with a boundary condition given in advance (type discriminating step).
The defect determination means is calculated by the defect candidate type determined by the type determination means, the position of the defect candidate calculated by the position calculation means (position calculation step), and the area calculation means (area calculation step). The presence / absence of a defect in the inspection object is determined by comparing each of the areas of the defect candidates with a predetermined determination criterion (defect determination step).

  That is, the defect determination means comprehensively determines whether or not a defect exists in the inspection target based on the type, position, and area of the defect candidate. Therefore, according to the defect inspection apparatus according to the present invention, it is possible to appropriately determine the presence or absence of a defect in the inspection object.

  In the defect inspection apparatus according to the present invention, the defect inspection apparatus further includes a filter processing unit that applies a bilateral filter to image information of the inspection object input from the outside, and the feature vector extraction unit is provided by the filter processing unit. It is preferable that high-order local autocorrelation processing is performed on the region determined by the feature vector extraction region determination unit in the image information subjected to the filter processing to extract the feature vector of the defect candidate.

According to such a configuration, the filter processing unit applies a bilateral filter to the image information of the inspection object, and the feature vector extraction unit outputs the feature vector of the defect candidate to the image information subjected to the bilateral filter. Extract (filtering step).
Here, the bilateral filter performs smoothing processing on the entire image while retaining edges. As a result, the feature vector extracting means extracts the feature vector of the defect candidate from the image information subjected to the bilateral filter, and thus can appropriately extract the feature vector for each defect candidate.

  In the defect inspection apparatus according to the present invention, in the image information filtered by the filter processing unit, the luminance value is not less than a predetermined value in the region input from the feature vector extraction region determining unit. A degree value calculating unit that calculates a ratio of pixels as a degree value of the defect candidate, and the defect determining unit includes a defect candidate type determined by the type determining unit and a defect calculated by the position calculating unit. Each of the candidate position, the area of the defect candidate calculated by the area calculating unit, and the degree value of the defect candidate calculated by the degree value calculating unit are compared with the predetermined determination criterion. It is preferable to determine the presence or absence of a defect in the inspection object.

According to such a configuration, in the image information input from the filter processing unit, a ratio of pixels having a luminance value equal to or greater than a predetermined value in the region input from the feature vector extraction region determination unit is determined as the defect candidate. Is calculated as a degree value (degree value calculation step).
Then, the defect determination means reflects the degree value of the defect candidate in addition to the type, position, and area of the defect candidate when determining the presence or absence of the defect. Therefore, the defect determination means can appropriately determine the presence or absence of a defect in the inspection object.
It should be noted that the relationship between the ratio and the degree value occupied by pixels having a luminance value equal to or greater than a predetermined value can be set as appropriate for each type of defect.

  Further, in the defect inspection apparatus according to the present invention, the type determination unit uses the hyperplane determined by a determination function derived in advance from teacher image information in which the type of defect candidate is known as the boundary condition. It is preferable to determine the type of candidate.

  According to such a configuration, the type determination unit appropriately determines the type of defect candidate by determining the type of defect candidate using a hyperplane determined by a discriminant function derived in advance from the teacher image information as a boundary condition. Can be determined.

Moreover, this invention is a program for making the said defect inspection apparatus which is a computer perform the said defect inspection method.
According to the program according to the present invention, it is possible to cause a defect inspection apparatus to execute a defect inspection method that can appropriately determine the presence or absence of a defect in an inspection object.

  According to the present invention, it is possible to provide a defect inspection apparatus, a defect inspection method, and a program that can appropriately determine the presence or absence of a defect in an inspection object.

1 is a configuration diagram of a defect inspection apparatus according to a first embodiment of the present invention. It is the figure which calculated and plotted the Mahalanobis general distance from the missing gravity center and the Mahalanobis general distance from the foreign material gravity center for a plurality of teacher images. (A) is explanatory drawing of the image of a test object, (b) is explanatory drawing of the image after the binarization process by a binarization process part, (c) is after the filter process by a filter process part It is explanatory drawing of an image. (A) is a brightness | luminance histogram in the image information of the test object input from the outside, (b) is a brightness | luminance histogram when distribution is changed by the filter process part. It is explanatory drawing of the high-order local autocorrelation process which a feature vector extraction part performs. It is explanatory drawing which shows the reference | standard at the time of discriminating whether it corresponds to a defect based on the kind, position, and area of a defect candidate. It is a flowchart which shows the flow of a process of a defect determination part. It is explanatory drawing which shows the determination result at the time of performing various filter processing as pre-processing which extracts a feature vector, (a) is a case where an edge emphasis filter is used as a comparative example, (b) is this implementation This is a case where the bilateral filter according to the embodiment is used. It is a block diagram of the defect inspection apparatus which concerns on 2nd Embodiment of this invention. It is a block diagram of the defect inspection apparatus which concerns on 3rd Embodiment of this invention.

  Hereinafter, modes for carrying out the present invention (hereinafter referred to as embodiments) will be described in detail with reference to the drawings as appropriate. In addition, the same code | symbol is attached | subjected to the common part in each figure, and the overlapping description is abbreviate | omitted.

<< First Embodiment >>
<Configuration of defect inspection device>
FIG. 1 is a configuration diagram of a defect inspection apparatus according to the first embodiment of the present invention. The defect inspection apparatus 100 inspects the presence / absence of a defect in the inspection object based on image information of the inspection object. In the following description, as an example, a case will be described in which the defect inspection apparatus 100 inspects whether or not there is a defect on the surface of a fuel cell electrode and a solid polymer membrane (Proton Exchange Membrane: PEM).
In the following description, objects to be inspected by the defect inspection apparatus 100 such as the electrodes and the solid polymer film are collectively referred to as “inspection objects”.

Further, the “defect” is a defective portion existing in the inspection target, and includes, for example, a portion where a missing portion (dent) or a foreign object exists. An inspection object having a defect becomes a target for disposal or reuse.
Further, the “defect candidate” indicates a portion that may correspond to the above-described defect in the inspection object.

As illustrated in FIG. 1, the defect inspection apparatus 100 includes a pre-learning unit 101, a first storage unit 102, a region information acquisition unit 103, an area calculation unit 104, a position information acquisition unit 105, and a filter processing unit 106. A feature vector extraction unit 107, a type determination unit 108, a second storage unit 109, and a defect determination unit 110.
The defect inspection apparatus 100 includes an electronic circuit (not shown) such as a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), and various interfaces, and a program stored in the ROM. Are loaded into the RAM, and the CPU executes various processes.

The defect inspection apparatus 100 according to the present embodiment performs a pre-learning process based on teacher image information input from the outside before starting an inspection process for checking the presence / absence of a defect.
Therefore, after first describing the configuration related to the pre-learning process, the configuration related to the defect inspection process using the learning result will be described.

(1. Configuration related to pre-learning processing)
The pre-learning unit 101 performs pre-learning using teacher image information input from the outside. The teacher image information is image information of defect candidates prepared in advance by a person who sets the defect inspection apparatus 100 (hereinafter referred to as a setter).
Incidentally, there are a plurality of types of “defect candidates” (for example, foreign matter P1, missing Q1, etc .: see FIG. 3A), and “defects” are predetermined criteria (type, position, and area) of defect candidate images. To meet.
In the present embodiment, as an example, a case where there are 10 types of defect candidates in total will be described.

First, a plurality of (for example, six) samples are prepared in advance by the setter for each defect candidate type, and input to the pre-learning unit 101 as teacher image information. In addition, the sample type (type 1,..., Type 10) of the teacher image information is input to the pre-learning unit 101 in association with the teacher image information.
The pre-learning unit 101 associates a predetermined identification symbol with the input teacher image information and type, and stores them in the first storage unit 102.

Next, the pre-learning unit 101 sequentially reads the teacher image information and the identification symbol from the first storage unit 102 and outputs them to the filter processing unit 106. Then, for each input teacher image information, the filter processing unit 106 performs a filter process, and the feature vector extraction unit 107 extracts a feature vector of the teacher image information. Further, the feature vector extraction unit 107 stores information on the extracted feature vector in the first storage unit 102 in association with the above-described identification symbol.
Details of processing by the filter processing unit 106 and the feature vector extraction unit 107 will be described later.

When feature vectors are extracted for all teacher image information by the feature vector extraction unit 107, the pre-learning unit 101 reads the feature vector of the teacher image information from the first storage unit 102, and performs discriminant analysis processing using the feature vector. Run. That is, the pre-learning unit 101, based on the feature vector of each teaching image data, a correlation ratio eta ² represented by the ratio of intergroup variation S _B to the total variation S _T derives a discriminant function to maximize, Store in the first storage unit 102.

Next, the pre-learning unit 101 outputs a command signal that causes the type determination unit 108 to execute the type determination process. When the signal is input, the type discriminating unit 108 sequentially reads the feature vector of the teacher image information from the first storage unit 102, and uses the discriminant function to determine the type (type 1,..., Type 10). Is determined). Then, the type determination unit 108 stores the determined type in the first storage unit 102 in association with the identification symbol of the teacher image information.
Further, the pre-learning unit 101 calculates the barycentric coordinates of the feature vectors of the plurality of teacher image information of the same type, and stores them in the first storage unit 102 in association with the types.

Therefore, in the first storage unit 102, the type of the teacher image information input in advance by the setter and the type of the teacher image information determined by the type determination unit 108 using the discrimination function are associated with the identification symbols. Will be stored.
Further, the first storage unit 102 stores information regarding the barycentric coordinate value of the teacher image information for each type.

  Next, the pre-learning unit 101 determines the suitability of learning based on the correct answer rate of the type determining unit 108. The correct answer rate is represented by a ratio between the number of samples in which the type determined by the type determination unit 108 is correct (matches the type input by the setter) and the number of samples of all teacher image information. When the correct answer rate is equal to or higher than a predetermined value set in advance, the pre-learning unit 101 determines that learning is appropriate. On the other hand, when the correct answer rate is less than the predetermined value, the pre-learning unit 101 determines that the learning is inappropriate.

Moreover, even when the correct answer rate is equal to or higher than the predetermined value, there is a possibility that the teacher image information itself is not appropriate. That is, there may be a feature vector of the teacher image data selected by the setter that is separated from the barycentric coordinates of that type.
Accordingly, the pre-learning unit 101 reads out the barycentric coordinates for each type (for each type determined by the type determining unit 108) from the first storage unit 102, and calculates the Mahalanobis general distance between each feature vector and the barycentric coordinates. The data is output to the display device 200. Incidentally, the Mahalanobis generalized distance is a distance obtained by standardizing Euclidean distance by dispersion.

Thus, the setter can know the discrimination result of the type discrimination unit 108 and the Mahalanobis general distance from the barycentric coordinates for each input teacher image information.
FIG. 2 is a diagram in which the Mahalanobis general distance from the missing center of gravity and the Mahalanobis general distance from the foreign object center of gravity are calculated and plotted for a plurality of teacher images. In the example shown in FIG. 2, feature vectors are extracted from images of two types of defect samples (missing samples: 6, foreign matter samples: 6) as teacher image information, and Mahalanobis general distances from the centroid coordinates of each type are extracted. Calculated and plotted.

In the example shown in FIG. 2, the data included in the region p1 corresponds to the teacher image information of the missing sample, and the data included in the regions q1 and q2 corresponds to the teacher image information of the foreign material sample. Therefore, it can be seen that the type discriminating unit 108 appropriately discriminates the type of defect based on the discriminant function derived by the prior learning unit 101.
In FIG. 2, the distance q from the missing center of gravity is relatively small in the data of the region q2, but the shape of the foreign material sample corresponding to the data is slightly different from the other foreign material samples.

Therefore, the setter knows that the teacher image information input to the pre-learning unit 101 is appropriate in the example shown in FIG. Incidentally, when it is determined that the pre-learning is inappropriate, the pre-learning unit 101 performs the pre-learning again based on the teacher image information changed by the setter to derive a discriminant function.
In this way, the pre-learning unit 101 derives an appropriate discriminant function and stores it in the first storage unit 102 shown in FIG.

Incidentally, the number of dimensions of the explanatory variable of the teacher image information is the number of dimensions of the feature vector (33 dimensions) extracted by the feature vector extraction unit 107 described above. Therefore, the pre-learning unit 101 specifies a plurality (for example, five) explanatory variables contributing to the identification of each type among the 33 dimensions representing the teacher image information using a variable increase / decrease method, It is preferable to represent the discriminant function using the explanatory variable.
As a result, a discriminant function that can appropriately classify the type of teacher image information can be derived, and the processing load of the type discriminating unit 108 to be described later can be reduced.

  Further, as shown in FIG. 1, discrimination reference information (position information and area information) serving as a reference when the defect determination unit 110 determines the presence / absence of a defect is input to the first storage unit 102 in advance by a setter. The Details of the discrimination criterion information will be described later.

(2. Configuration related to defect inspection processing)
In the defect inspection process, after the pre-learning process described above is completed, the image information of the inspection object is input to the defect inspection apparatus 100, and the user performs a predetermined operation via an input unit (such as a mouse: not shown). Is started.
The image information of the inspection object is obtained by, for example, imaging the inspection object irradiated with light by an illumination device (not shown) with an imaging device (not shown) such as a CCD (Charge Coupled Device) camera. Can be obtained by:

When the inspection object is an electrode of a fuel cell and a solid polymer membrane, image information (hereinafter simply referred to as image information) input from an external imaging device is as shown in FIG.
As shown in FIG. 3A, the inspection object has an electrode E bonded on a solid polymer film M. Further, there is a missing P1 on the upper right side of the electrode E, and a foreign substance Q1 on the lower left side of the electrode E.
Incidentally, in FIG. 3A, for the purpose of explanation, defect candidates (missing P1 and foreign matter Q1) are shown larger than actual dimensions (the same applies to FIGS. 3B and 3C described later). .

The region information acquisition unit 103 performs a binarization process on the image information of the inspection object input from the outside, thereby extracting the coordinate value of the pixel included in the defect candidate, and using the identification information for each defect candidate. The defect candidate area information in which a certain labeling value is associated with the coordinate value is acquired.
Here, the “defect candidate area information” is information for distinguishing the area occupied by the defect candidate from other areas and identifying each defect candidate. For example, the area information acquisition unit 103 sets the pixel value of the area occupied by the defect candidate to 1 and sets the pixel value of the area that is not the defect candidate to 0.

  The area information acquisition unit 103 includes a luminance information generation unit 103a and a binarization processing unit 103b. The luminance information generation unit 103a generates luminance information from image information input from an external imaging device. That is, all the pixels of the input image information are converted from RGB data to YUV data, and the Y value (luminance value) of each pixel is output to the binarization processing unit 103b as luminance information.

  The binarization processing unit 103b performs binarization processing based on the luminance information input from the luminance information generation unit 103a. For example, the binarization processing unit 103b calculates a binarization threshold value at which the two groups are most separated from the luminance value histogram. Then, the binarization processing unit 103b associates 1 with a pixel whose luminance value is greater than or equal to the binarization threshold, and associates 0 with a pixel whose luminance value is less than the binarization threshold.

Further, the binarization processing unit 103b identifies adjacent pixels (or a Euclidean distance equal to or smaller than a predetermined value) among pixels having a pixel value of 1 as the same region, and labels each region.
For example, the binarization processing unit 103b performs binarization processing on the image illustrated in FIG. 3A, and sets the pixel values of the region P2 corresponding to the missing P1 and the pixel value of the region Q2 corresponding to the foreign matter Q1 to 1. (See FIG. 3B). In addition, the binarization processing unit 103b sets all pixel values in regions other than the regions P2 and Q2 to 0.

Further, the binarization processing unit 103b identifies each of the areas P2 and Q2 shown in FIG. 3B as one area, and performs labeling on the areas P2 and Q2. That is, the binarization processing unit 103b associates a labeling value, which is identification information for each defect candidate, with a coordinate value of a pixel (a group of pixels in each defect candidate) included in the defect candidate. Get region information.
In this way, the binarization processing unit 103b acquires defect candidate region information and outputs the defect candidate region information to the area calculation unit 104 and the position information acquisition unit 105.

  The area calculation unit 104 calculates the area for each defect candidate based on the defect candidate region information acquired by the region information acquisition unit 103. In other words, the area calculation unit 104 calculates the number of pixels in the region having the same labeling value (that is, the area of the defect candidate). Then, the area calculation unit 104 stores the area value for each defect candidate in the second storage unit 109 in association with the labeling value described above.

The position information acquisition unit 105 acquires information regarding the position of the defect candidate. The position information acquisition unit 105 includes a position calculation unit 105a and a feature vector extraction region determination unit 105b.
The position calculation unit 105 a calculates the position of each defect candidate as the barycentric coordinate value based on the defect candidate region information acquired by the region information acquisition unit 103. That is, the position calculation unit 105a calculates a centroid coordinate value for each of the defect candidates labeled by the binarization processing unit 103b, and stores it in the second storage unit 109 in association with the labeling value.

The feature vector extraction region determination unit 105b determines a region for extracting a feature vector for each defect candidate based on the defect candidate region information acquired by the region information acquisition unit 103.
The feature vector extraction region determination unit 105b determines a region from which a feature vector is extracted (hereinafter referred to as “feature vector extraction region”), for example, as follows. That is, the feature vector extraction region determination unit 105b specifies a rectangular region including a defect candidate region for each defect candidate input from the binarization processing unit 103b, and associates the coordinate value with the labeling value described above. To the feature vector extraction unit 107.

In the example shown in FIG. 3B, the feature vector extraction region determination unit 105b associates the coordinate values specifying the rectangular region TP2 including the region P2 and the rectangular region TQ2 including the region Q2 with the labeling values. The result is output to the feature vector extraction unit 107.
The coordinate value only needs to be able to specify a rectangular area. For example, the coordinate value of the pixel located in the uppermost right of each rectangular area and the coordinate value of the pixel located in the lowermost left of the rectangular area may be used.

The filter processing unit 106 applies a bilateral filter to the image information (that is, all the pixels of the input image information) of the inspection object input from an external imaging device (not shown). . Incidentally, the binarization processing is not performed on the image information input to the filter unit 106.
The bilateral filter is an edge-preserving smoothing filter, and emphasizes edges (contours) by canceling light and shade changes and fine noise in the entire image by performing smoothing processing on the entire image. Thereby, the feature vector extraction unit 107 described later can more appropriately extract the feature vector of the defect.

Note that the image processing by the filter processing unit 106 and the image processing by the region information acquisition unit 103 are executed in parallel. That is, the area information acquisition unit 103 performs a binarization process on the image information in parallel with the filter processing unit 106 performing a bilateral filter on the image information input from the outside. Region information.
Thereby, the process of the defect inspection apparatus 100 can be performed more promptly.

In the present embodiment, the filter processing unit 106 applies a data-dependent bilateral filter. The data-dependent bilateral filter calculates a parameter for each local region of an image by, for example, fuzzy inference. And the filter process part 106 performs a filter process by applying the bilateral filter based on the parameter which changes for every said local area | region.
As a result, it is possible to avoid recognizing and emphasizing a pixel with a lot of noise as an edge.

As preprocessing for applying the data-dependent bilateral filter, it is preferable to change the luminance histogram distribution to obtain a distribution suitable for detecting defects.
4A is a luminance histogram in the image information of the inspection target inputted from the outside, and FIG. 4B is a luminance histogram in the case where the distribution is changed by the filter processing unit 106.

  For example, as shown in FIG. 3A, consider a case in which the inspection object is an electrode E and a solid polymer film M used in a fuel cell. Incidentally, the electrode E is composed of a C / C composite in which graphite is reinforced with carbon fibers. Then, since the electrode E is black, the luminance value of the image obtained by imaging the electrode E by the imaging device (not shown) is concentrated in the range A in which the value is small (FIG. 4A). reference). Incidentally, the range B where the luminance value is relatively large corresponds to the solid polymer film M (see FIG. 3A).

  Since most of the defects exist in the electrodes, the filter processing unit 106 extracts pixels that exist in the range A corresponding to the electrodes. That is, the filter processing unit 106 has a luminance value histogram having a normal distribution and a median value of the normal distribution having a predetermined value set in advance in pixels existing within the range A shown in FIG. The luminance value distribution is changed.

For example, for the pixels whose luminance values are within the range A shown in FIG. 4A, the filter processing unit 106 has a normal distribution of luminance histograms for those pixels, and the median value of the normal distribution is a predetermined value. Change the brightness value so that
Thus, by changing the histogram distribution of luminance values, the feature vector extraction unit 107 described later can more appropriately extract feature vectors.

  Then, the filter processing unit 106 applies a data-dependent bilateral filter to the image information whose luminance value has been changed, and outputs the result to the feature vector extraction unit 107. The information output from the filter processing unit 106 to the feature vector extraction unit 107 is, for example, image information corresponding to a smoothed image as a whole image while retaining edges as shown in FIG. It becomes.

The feature vector extraction unit 107 performs high-order local autocorrelation processing on the image information before binarization processing in the region determined by the feature vector extraction region determination unit 105b, and extracts feature vectors for each defect candidate. .
For example, when inspecting whether or not the electrode surface of the fuel cell is defective one by one for each electrode, the electrode having the defect is a part of the whole. Also, the area occupied by the defects is very small relative to the area of the electrode. In such a case, if high-order autocorrelation processing is performed on the entire surface of the electrode, an enormous amount of calculation processing is performed. On the other hand, most of the electrode is not a defective region, and thus waste occurs.

  Therefore, in the present embodiment, the feature vector extraction unit 107 performs high-order local autocorrelation processing only for the rectangular region (one or a plurality of rectangular regions including defect candidates) input from the feature vector extraction region determination unit 105b. . As a result, feature vectors are extracted only for rectangular areas where defects may exist, so that the amount of calculation when performing defect inspection can be greatly reduced, and inspection processing can be performed quickly. Become.

Next, high-order local autocorrelation processing performed by the feature vector extraction unit 107 will be described. Higher-Order Local Auto-Correlation (HLAC) calculates features by counting the correlation patterns in a two-dimensional local region as a whole, and has superior properties such as additiveness. Have
The Nth-order autocorrelation function _XN is expressed by the following (formula 1). In (Expression 1), f (r) is a value of image information, and (a ₁ , a ₂ ,..., A _N ) represents a displacement direction viewed from the reference pixel r.

  In the present embodiment, N = 2 (two-dimensional image information) in (Expression 1). The feature vector extraction unit 107, for example, the pixel value of the reference pixel in the feature vector extraction region (rectangular region TP3 shown in FIG. 5) and the pixel value of the pixel masked by the predetermined mask pattern K (see FIG. 5). The feature value (component value of the feature vector) corresponding to the mask pattern K is calculated by integrating the value obtained by multiplying all the pixels in the rectangular area.

Incidentally, the above-described mask pattern is a predetermined pattern set for pixels located in the vicinity of the Moore of reference pixels (that is, eight pixels in the vicinity surrounding the reference pixel). Further, if masks that have the same value when the mask is moved in parallel are excluded, there are 33 mask patterns.
The feature vector extraction unit 107 calculates a feature amount (component value of the feature vector) for each of these 33 mask patterns, extracts a 33-dimensional feature vector representing the feature of the rectangular area including the defect candidate, and The data is output to the type determination unit 108 in association with the labeling value.

The type discriminating unit 108 discriminates the type of defect candidate by comparing each feature vector extracted by the feature vector extracting unit 107 with a boundary condition given in advance. More specifically, the type discriminating unit 108 discriminates the type of defect candidate using the hyperplane determined by the discriminant function derived in advance from the teacher image information whose type of defect candidate is known as the boundary condition. The discriminant function derived in advance from the teacher image information is acquired in advance by the pre-learning process by the pre-learning unit 101 and stored in the first storage unit 102.

  When the information corresponding to the feature vector of the defect candidate is input from the feature vector extraction unit 107, the type determination unit 108 reads out the determination function previously derived by the pre-learning unit 101 from the first storage unit 102. Then, the type discriminating unit 108 discriminates the type of the defect candidate depending on which space the feature vector of the defect candidate exists in the space partitioned by the hyperplane determined by the read discriminant function.

Further, the type determination unit 108 stores the determined type of defect candidate in the second storage unit 109 in association with the labeling value of the defect candidate.
Therefore, the second storage unit 109 stores the position, area, and type of the defect candidate in association with the labeling value determined for each defect candidate.

The defect determination unit 110 determines each of the defect candidate type determined by the type determination unit 108, the position of the defect candidate calculated by the position calculation unit 105a, and the area of the defect candidate calculated by the area calculation unit 104. The presence / absence of a defect in the inspection object is determined by comparing with a predetermined criterion.
Here, the above-mentioned “discrimination criteria” are positional information (for example, corresponding to the regions X, Y, and Z shown in FIG. 6) and area information (for example, the areas α, β) and type information (for example, types 1, 2,..., 10 shown in FIG. 6), and information indicating a reference value when determining the presence or absence of a defect in the inspection object.
Further, the position, area, and type of the defect candidate described above are acquired by the defect determination unit 110 reading each information from the second storage unit 109 in association with the labeling value described above.

  FIG. 6 is an explanatory diagram showing a criterion for determining whether or not a defect corresponds to the defect based on the type, position and area of the defect candidate. In FIG. 6, a circle indicates a case that does not correspond to a defect, and a cross indicates a case that corresponds to a defect.

Further, the “position” shown in FIG. 6 is the position of the defect candidate (coordinate value of the center of gravity) calculated by the position calculation unit 105a, and the “area S” is the defect candidate calculated by the area calculation unit 104. Area.
6 is, for example, a predetermined rectangular region in the electrode E (see FIG. 3A), and the “region Y” is a solid polymer membrane M (see FIG. 3B). The “region Z” is a predetermined region in the vicinity of the outer edge of the solid polymer film M.
As described above, the regions X, Y, and Z (position information) and the area threshold values α and β (area information) shown in FIG. 6 are discrimination reference information given in advance by the setter, The value is stored in advance in the first storage unit 102 (see FIG. 1).

The position and area criteria used when the defect determination unit 110 determines the presence or absence of a defect may differ depending on the type of defect.
For example, when the type of the missing P1 (see FIG. 3A) corresponds to the type 1 shown in FIG. 6 and the missing P1 is located in the region X, the defect determining unit 110 detects the missing P1 that is a defect candidate. The following determination is made for.
That is, when the value of the area S of the missing P1 is less than β, the defect determining unit 110 determines that the missing P1 does not correspond to a defect. On the other hand, when the value of the area S of the missing P1 is equal to or larger than β, the defect determining unit 110 determines that the missing P1 corresponds to a defect.

Further, when the type of the foreign matter Q1 (see FIG. 3A) is the type 10 shown in FIG. 6 and the foreign matter Q1 is located in the region X, the defect determination unit 110 determines the foreign matter Q1 that is a defect candidate. The following determination is made.
That is, when the value of the area S of the foreign matter Q1 is α or less, the defect determination unit 110 determines that the foreign matter Q1 does not correspond to a defect. On the other hand, when the value of the area S of the foreign matter Q1 is larger than α, the defect determination unit 110 determines that the foreign matter Q1 corresponds to a defect.
In this way, the defect determination unit 110 sequentially determines whether or not the defect corresponds to each defect candidate, and outputs information corresponding to the final determination result to the external display device 200. The display device 200 displays the information received from the defect determination unit 110.

  FIG. 7 is a flowchart showing a flow of processing of the defect determination unit. In step S101, the defect determination unit 110 sets the value of the labeling value n of the defect candidate to be inspected as 1. That is, the defect determination unit 110 starts an inspection for the first defect candidate among one or more defect candidates.

In step S102, the defect determination unit 110 determines whether the type of defect candidate selected in step S101 is a type that may be a defect. Here, “there is a possibility of a defect” indicates that the defect candidate may correspond to a defect depending on the position and area conditions of the defect described later. Incidentally, in the example shown in FIG. 6, type 1, type 2, and type 10 may be a defect.
Note that the defect determination unit 110 may determine that the defect is a defect only by corresponding to a predetermined type regardless of the position and area of the defect.

When the type of defect candidate is a type that may be a defect (S102 → Yes), the process of the defect determination unit 110 proceeds to step S103. On the other hand, if the type of defect candidate is not a type that is likely to be a defect (S102 → No), the process of the defect determination unit 110 proceeds to step S106.
In step S <b> 103, the defect determination unit 110 determines whether or not the position of the defect candidate is a position that may be a defect. That is, the defect determination unit 110 determines whether or not the position of the defect candidate (centroid coordinate value) calculated by the position calculation unit 105a is a position that may be a defect. For example, in FIG. 6, when the type of defect candidate is “type 1” and the position is in the region Y, the defect determination unit 110 determines that there is a possibility of a defect.

When the position of the defect candidate is a position where there is a possibility of being a defect (S103 → Yes), the processing of the defect determination unit 110 proceeds to step S104. On the other hand, when the position of the defect candidate is not a position where there is a possibility of being a defect (S103 → No), the process of the defect determination unit 110 proceeds to step S106.
In step S104, the defect determination unit 110 determines whether the area corresponds to a defect. That is, the defect determination unit 110 determines whether the area (number of pixels) of the defect candidate calculated by the area calculation unit 104 is within a predetermined range set in advance.
For example, in FIG. 6, when the defect candidate type is “type 1”, the position is the region Y, and the area S value is α <S <β, the defect determination unit 110 It is determined that the defect candidate corresponds to a defect (see the shaded portion in FIG. 6).

When the area of the defect candidate is within the predetermined range (S104 → Yes), the process of the defect determination unit 110 proceeds to step S105. On the other hand, when the area of the defect candidate is not within the predetermined range (S104 → No), the process of the defect determination unit 110 proceeds to step S106.
In step S <b> 105, the defect determination unit 110 determines that the inspection target has a defect, and outputs a signal corresponding to the determination result to the display device 200. When the signal is received from the defect determination unit 110, the display device 200 displays an image corresponding to “defect”.

In step S106, the defect determination unit 110 determines whether or not all defect candidates have been checked for the presence or absence of defects. That is, when the total number of defect candidates is N, it is determined whether or not the value n of the defect candidates to be inspected is equal to the total number N of defect candidates.
When the presence / absence of defects is examined for all defect candidates (that is, when n = N: S106 → Yes), the process of the defect determination unit 110 proceeds to step S107. On the other hand, when there is at least one defect candidate whose presence or absence of a defect has not been examined (that is, when n <N: S106 → No), the process of the defect determination unit 110 proceeds to step S108.

In step S <b> 107, the defect determination unit 110 determines that the inspection target is not defective, and outputs a signal corresponding to the determination result to the display device 200. When the signal is received from the defect determination unit 110, the display device 200 displays an image corresponding to “no defect”.
In step S108, the defect determination unit 110 increments the labeling value n of the defect candidate and returns to the process of step S102. In this case, in steps S102 to S104, the defect determination unit 110 inspects whether or not the defect candidate corresponding to the incremented labeling value n corresponds to a defect.

<Effect 1>
According to the defect inspection apparatus 100 according to the present embodiment, the feature vector extraction unit 107 performs feature vector extraction processing using higher-order local autocorrelation processing. In high-order local autocorrelation processing, feature vectors are calculated by counting correlation patterns in a two-dimensional local region. Thereby, the features of the defect candidate can be extracted as a 33-dimensional feature vector including not only the shape (contour) of the defect candidate but also the texture of the surface.
Therefore, the type determination unit 108 can appropriately determine the type of defect candidate as in the case where a human makes a visual determination by determining the type of defect candidate based on the above-described feature vector.

The feature vector extraction unit 107 performs high-order local autocorrelation processing only for the feature vector extraction region (rectangular region) input from the feature vector extraction region determination unit 105b, and performs high-order local autocorrelation for the other regions. Do not process. Therefore, the amount of calculation required for the feature vector extraction process can be reduced, and the time required for defect inspection can be greatly shortened.
Usually, the higher-order local autocorrelation function is applied to the entire image of the inspection object. When image processing was performed by applying a higher-order local autocorrelation function to the entire image of the inspection object corresponding to the image shown in FIG. 3A, 85.3 hours were required for each image.

On the other hand, when the high-order local autocorrelation process is performed only for the feature vector extraction region (rectangular region) using the defect inspection apparatus 100 according to the present embodiment, the processing may be completed in 64 seconds per image. did it. That is, the processing speed is about 5000 times that of the case where the high-order local autocorrelation processing is performed on the entire image information.
As described above, by performing the feature vector extraction process only for the rectangular region including the defect candidate, the time required for the feature vector extraction process can be significantly shortened.

In addition, before the feature vector extraction processing unit 108 performs high-order local autocorrelation processing, the filter processing unit 106 applies a bilateral filter. As described above, the bilateral filter performs a smoothing process (blurring process) on the entire image while maintaining an edge.
Normally, when an inspection object is imaged by an imaging device (not shown), the peripheral portion of the image is slightly darker than the central portion. In the present embodiment, edge enhancement processing can be performed while canceling out the luminance difference and the luminance difference due to noise by applying a bilateral filter.
By performing such preprocessing, the feature vector extraction accuracy by the feature vector extraction unit 107 can be improved.

FIG. 8 is an explanatory diagram illustrating determination results when various filter processes are performed as pre-processing for extracting feature vectors. FIG. 8A illustrates a case where an edge enhancement filter is used as a comparative example. ) Is a case where the bilateral filter according to the present embodiment is used.
It should be noted that as a defect sample, a foreign matter sample (6 pieces) having a known defect type and a missing sample (6 pieces) having a known defect type are used, and the defect type (that is, whether a foreign matter or a missing piece is applicable). It was made to distinguish. Incidentally, in both cases of FIGS. 8A and 8B, the processing contents in the feature vector extraction unit 107 and the type determination unit 108 are the same as those described above.

In the case of using the edge enhancement processing shown in FIG. 8A (comparative example), three of the six foreign matter samples could be judged as “foreign matter”, but the remaining three were judged as “missing”. . In addition, two of the six missing samples could be determined as “missing”, but the remaining four were determined as “foreign matter”. As a result, the correct answer rate for the foreign sample was 50%, and the correct answer rate for the missing sample was 33%.
This is considered to be because there was a case where a defect candidate had a missing edge, and one defect candidate was erroneously recognized as a plurality of defect candidates.

On the other hand, when the bilateral filter shown in FIG. 8B is used (this embodiment), all six foreign samples are determined as “foreign” and all six missing samples are “missed”. It was possible to determine. That is, the correct answer rates for the foreign material sample and the missing sample were both 100%.
Therefore, in this embodiment, by using a bilateral filter as the preprocessing, the feature vector extraction unit 107 accurately extracts the feature vector of the defect candidate, and the type determination unit 108 appropriately selects the type of defect candidate based on the feature vector. Can be determined.

  In addition, since the defect inspection apparatus 100 according to the present embodiment determines whether or not the defect candidate corresponds to the defect by combining the types, positions, and areas of the defect candidates, it can more appropriately determine the presence or absence of the defect. . For example, when inspecting whether or not an electrode (see FIG. 3A) has a defect, even if the defect candidate type is the same and the area is almost the same, the defect is determined depending on its position. Whether or not to determine may be different.

  If high-order local autocorrelation processing with shift invariance is performed on the entire image, if the type and area are almost the same, the same conclusion (applicable to defect or not applicable) is issued regardless of position. End up. On the other hand, in the present embodiment, as in the case where a human visually inspects, determination processing is performed in consideration of not only the type of defect candidate but also the position and area, and therefore it is appropriately determined whether or not the defect is applicable. Can do.

Further, although the filter processing unit 106 performs processing using a bilateral filter, the feature vector extraction accuracy is improved, but it is difficult to recognize defect candidates as a group for each region. This is because a luminance difference is generated even in the edge of one defect candidate region depending on how the light from the illumination device (not shown) hits.
On the other hand, in the defect inspection apparatus 100 according to the present embodiment, the region information acquisition unit 103 acquires defect candidate region information based on the binarized image information, and the feature vector extraction region is based on the region information. The determination unit 105b determines a feature vector extraction region.
Therefore, the feature vector extraction unit 107 only needs to perform feature vector extraction as a defect candidate exists in each of the determined feature vector extraction regions (rectangular regions), and thus solves the above-described problem. Can do.

<< Second Embodiment >>
FIG. 9 is a configuration diagram of a defect inspection apparatus according to the second embodiment of the present invention.
The defect inspection apparatus 100A according to the present embodiment is different from the defect inspection apparatus 100 according to the first embodiment in that a degree value calculation unit 111 is added, but the others are the same as in the first embodiment. is there.
Therefore, the different parts will be described, and the description of the other parts will be omitted.

The degree value calculation unit 111 has a defective ratio of pixels whose luminance value is equal to or greater than a predetermined value in the region input from the feature vector extraction region determination unit 105b in the image information filtered by the filter processing unit 106. Calculated as a candidate degree value.
Note that the “degree value” indicates, for example, the degree of the depth of a defect with respect to the surface of the inspection object, the degree of protrusion of foreign matter, and the like, and the greater the degree value, the higher the possibility of being a defect. In addition, the relationship between the ratio of the pixels whose luminance value is equal to or greater than a predetermined value and the degree value can be set as appropriate for each type of defect.

The degree value calculation unit 111 is an image in an area (rectangular area) input from the feature vector extraction area determination unit 105b among the image information (image information subjected to the bilateral filter) input from the filter processing unit 106. For example, the following processing is performed on the information.
That is, the degree value calculation unit 111 calculates, as the degree value, the ratio of pixels having a luminance equal to or higher than a predetermined value to the pixels in the rectangular area, and stores the degree value in the second storage unit 109.

  Then, the defect determination unit 110 includes the type of defect candidate determined by the type determination unit 108, the position of the defect candidate calculated by the position calculation unit 105a, the area of the defect candidate calculated by the area calculation unit 104, The presence / absence of a defect in the inspection object is determined by comparing each of the defect candidate degree values calculated by the degree value calculation unit 111 with a predetermined criterion.

Here, the above-mentioned “discrimination criterion” includes degree value information in addition to the discrimination criterion (position information, area information, and type information) described in the first embodiment, and whether there is a defect in the inspection object. It is the information which shows the reference value at the time of determining.
The degree value information (information about the threshold value of the degree value) is a value given in advance by the setter and stored in the first storage unit 102.

Then, the defect determination unit 110 determines the discrimination reference value (here, the threshold value of the degree value) set in correspondence with the type, position, and area of the detected defect candidate, and the degree value of the defect candidate. Compare
If the degree value is equal to or greater than the threshold value, the defect determination unit 110 determines that the defect candidate corresponds to a defect. On the other hand, when the degree value is less than the threshold value, the defect determination unit 110 determines that the defect candidate does not correspond to a defect.

For example, the defect determination unit 110 has a positive correlation between the probability of being a defect and a luminance value in a teacher image as a defect candidate sample (that is, the higher the luminance value, the more likely it is a defect) For the types that are high, the following determination process is performed.
That is, the defect determination unit 110 compares the threshold corresponding to the type of defect candidate with the degree value of the defect candidate, and determines that the defect corresponds to a defect when the degree value is equal to or greater than the threshold. To do.
On the other hand, among teacher images as defect candidate samples, there is a negative correlation between the probability of being a defect and the luminance value (that is, the smaller the luminance value, the more likely it is a defect) The defect determination unit 110 determines that the defect corresponds to a defect when the degree value is equal to or less than a threshold value corresponding to the type of defect candidate.

<Effect 2>
According to the defect inspection apparatus 100A according to the present embodiment, when the defect determination unit 110 determines the presence or absence of a defect, the defect candidate calculated by the degree value calculation unit 111 in addition to the type, position, and area of the defect candidate. The degree value of is also reflected. That is, the discrimination reference value (here, the threshold value of the degree value) set corresponding to the type, position, and area of the detected defect candidate is compared with the degree value of the defect candidate, and the defect The presence or absence of is determined.
Therefore, the defect determination unit 110 can appropriately determine the presence / absence of a defect in the inspection object.

«Third embodiment»
FIG. 10 is a configuration diagram of a defect inspection apparatus according to the third embodiment of the present invention.
Compared to the defect inspection apparatus 100 according to the first embodiment, the defect information inspection apparatus 100B according to the present embodiment includes a position information acquisition unit 105B in addition to the position calculation unit 105a and the feature vector extraction region determination unit 105b. The other points are the same as those in the first embodiment, except that the sorting processing unit 105c is provided.
Therefore, the different parts will be described, and the description of the other parts will be omitted.

The selection processing unit 105c illustrated in FIG. 10 performs feature vector extraction by the feature vector extraction unit 107 based on a selection criterion given in advance from one or a plurality of regions input from the feature vector extraction region determination unit 105b. The area to be processed is selected (that is, narrowed down).
Here, the selection criterion given in advance is information including position information and area information input in advance by the setter and indicating a reference value when narrowing down the feature vector extraction region.

For example, when the defect candidate of the area S1 is near the center of the electrode E shown in FIG. 3A, if the area S1 is equal to or less than a predetermined value, the type does not need to be identified and does not correspond to a defect. It may be good.
On the other hand, when the defect candidate of the area S2 is in the vicinity of the edge of the solid polymer film M shown in FIG. 3A, it should be determined that it corresponds to the defect even if the area S2 is smaller than the area S1. There is also.
It should be noted that the reference value (that is, the selection criterion information) is input in advance by the setter and stored in the first storage unit 102 as described above.

  Then, the sorting processing unit 105 c reads position information and area information of defect candidates corresponding to a predetermined labeling value from the second storage unit 103, and further reads sorting reference information from the first storage unit 102. Then, it is determined whether or not the feature vector extraction process should be performed by comparing the position and area of the defect candidate with the selection criterion.

That is, when compared with the selection criteria, if at least one of the position and area does not satisfy the selection condition (that is, if it is clear that the defect does not correspond to the type determination process), the selection is performed. The processing unit 105c excludes the defect candidate.
On the other hand, when the position and area satisfy the selection condition when compared with the selection criterion (that is, when the defect determination process needs to be performed after performing the type determination process), the selection processing unit 105c. Performs the following process.

That is, the sorting processing unit 105c selects the defect selected by the above processing from the information acquired by the feature vector extraction region determining unit 105b (for example, information for determining the rectangular regions TP2 and TQ2 shown in FIG. 3B). Only the information regarding the candidate is output to the feature vector extraction unit 107.
Then, the feature vector extraction unit 107 performs feature vector extraction processing only on the feature vector extraction region (rectangular region) corresponding to the defect candidate selected by the selection processing unit 105c.

<Effect 3>
According to the defect inspection apparatus 100B according to the present embodiment, the selection processing unit 105c selects defect candidates to be subjected to the feature vector extraction process by comparing the position and area of the defect candidates with the selection reference. Then, the feature vector extraction unit 107 may perform feature vector extraction processing only for the rectangular area corresponding to the defect candidate.
Therefore, the amount of calculation in the feature vector extraction unit 107 can be greatly reduced, and the cost and time required for a series of defect inspection processes can be greatly reduced.

≪Modification≫
The defect inspection apparatus according to the present invention has been described above with reference to the embodiments. However, the embodiments of the present invention are not limited to these descriptions, and various modifications can be made.
For example, in each of the above embodiments, it is determined whether or not the electrode E and the solid polymer film M used in the fuel cell are defective, but the present invention is not limited to this. That is, the presence or absence of defects can be inspected for various inspection objects such as lithium ion batteries and semiconductor materials.

In each of the above embodiments, image information is input from an external imaging device, but the present invention is not limited to this. The image information may be stored in a recording medium or received via a network.
In the embodiment, the binarization processing unit 103b calculates a binarization threshold value from which the two groups are most separated from the luminance value histogram, and binarizes the image information based on the binarization threshold value. However, this is not a limitation. For example, the threshold value may be a predetermined value set in advance.

In each of the above embodiments, the case where the position calculation unit 105a calculates the barycentric coordinates of the defect candidate and outputs it as position information has been described. However, the present invention is not limited to this. For example, the position of the defect candidate may be the center position (position where diagonal lines intersect) of the rectangular area determined by the feature vector extraction area determination unit 105b.
In each of the above embodiments, the case has been described in which the feature vector extraction region determination unit 105b sets a rectangular region including a defect candidate as a feature vector extraction region, but the shape is not limited thereto. For example, the feature vector extraction region may be a circular region determined by the center coordinates and radius.

In each of the above embodiments, the filter processing unit 106 performs the filter processing using the bilateral filter, but the present invention is not limited to this. For example, the filter processing unit 106 may perform filter processing using an ABF (Adaptive Bilateral Filter), a MAX-MIN filter, or a Sobel filter.
Further, depending on the inspection object and the imaging apparatus, an image suitable for feature vector extraction may be obtained without using the filter processing unit 106. In such a case, the filter processing unit 106 may be omitted.

In the above embodiments, the pre-learning unit 101 determines the suitability of the teacher image information using the Mahalanobis general distance, but the present invention is not limited to this. For example, Kullback-Leibler (K-L) information criterion, Akaike information criteria (Akaike's Information Criterion: AIC) , or Marozu using C _p (Mallows C _p) of, to determine the appropriateness of teacher image information It is good.

In each of the above embodiments, the number of defect candidates is ten, but the number of defect candidates may be nine or less or eleven or more.
In the embodiment, the pre-learning unit 101 performs a discriminant analysis process to derive a discriminant function, and the type discriminating unit 108 discriminates the type of defect candidate based on the discriminant function. Absent.

For example, the pre-learning unit 101 may learn using a support vector machine and derive a hyperplane that maximizes the distance from each teacher image information. In this case, the type determination unit 108 determines the type of defect candidate based on the hyperplane.
In each of the above embodiments, the defect inspection apparatus 100 (100A) includes the first storage unit 102 and the second storage unit 109, but these may be integrated. The first storage unit 102 and the second storage unit 109 may be provided outside the defect inspection apparatus 100 (100A).

In each of the above embodiments, after the pre-learning unit 101 performs pre-learning based on teacher image information (image information whose type is known) input in advance by the setter, the defect inspection process is started. However, it is not limited to this.
That is, the defect inspection apparatus may include a learning unit (not shown) that performs unsupervised learning instead of the pre-learning unit 101. The learning unit performs, for example, clustering processing using the feature vector extracted by the feature vector extraction unit 107 on sample image information (image information whose type is unknown) input to the defect inspection apparatus.

As a clustering method, for example, hierarchical clustering such as the shortest distance method or group average method, or non-hierarchical clustering such as the k-means method can be performed.
The learning unit may perform a learning process using a neural network.

In each of the above embodiments, the defect inspection apparatus 100 (100A) includes the pre-learning unit 101. However, the pre-learning unit 101 may be omitted. For example, a device including the pre-learning unit 101 is provided separately, and pre-learning and testing are performed in advance using the device, and the device stores the derived discriminant function in the first storage unit 102 of the defect inspection device. It is good to do.
In this case, the apparatus needs to have the same configuration as the filter processing unit 106 and the feature vector extraction unit 107.

In addition, when the defect inspection apparatus does not include the pre-learning unit 101, information on the discriminant function may be input from the storage medium to the defect inspection apparatus, or acquired from the outside via the network and input to the defect inspection apparatus May be.
The processing according to each embodiment can be realized by a computer and a software program, and can be provided by recording the program on a computer-readable recording medium and via a network. It is also possible to provide.

100, 100A, 100B Defect inspection apparatus 101 Pre-learning unit 102 First storage unit 103 Region information acquisition unit (region information acquisition means)
103a Luminance information generation unit 103b Binarization processing unit 104 Area calculation unit (area calculation means)
105 position information acquisition unit 105a position calculation unit (position calculation means)
105b Feature vector extraction region determination unit (feature vector extraction region determination means)
106 Filter processing section (filter processing means)
107 feature vector extraction unit (feature vector extraction means)
108 type discriminator (type discriminating means)
109 Second storage unit 110 Defect determination unit (defect determination means)
111 degree value calculation unit (degree value calculation means)
200 Display device

Claims (9)

A defect inspection apparatus that inspects the presence or absence of a defect of an inspection object based on image information of the inspection object,
By performing a binarization process on the image information of the inspection object input from the outside, the coordinate value of the pixel included in the defect candidate is extracted, and the labeling value and the coordinates as identification information for each defect candidate Area information acquisition means for acquiring area information of defect candidates associated with values;
Position calculation means for calculating the position of each defect candidate as a barycentric coordinate value based on the area information of the defect candidate acquired by the area information acquisition means;
Area calculating means for calculating an area for each defect candidate based on the area information of the defect candidate acquired by the area information acquiring means;
A feature vector extraction region determination unit that determines a region for extracting a feature vector for each defect candidate based on the region information of the defect candidate acquired by the region information acquisition unit;
Feature vector extraction means for performing high-order local autocorrelation processing on the image information before the binarization processing in the region determined by the feature vector extraction region determination means and extracting a feature vector for each defect candidate When,
A type discriminating unit that discriminates the type of the defect candidate by comparing each feature vector extracted by the feature vector extracting unit with a predetermined boundary condition;
Discrimination criteria given in advance for the types of defect candidates discriminated by the type discriminating means, the positions of defect candidates calculated by the position calculating means, and the areas of defect candidates calculated by the area calculating means, respectively. And a defect determination means for determining the presence or absence of a defect of the inspection object by comparing with the defect inspection apparatus. Filter processing means for applying a bilateral filter to the image information of the inspection object input from the outside,
The feature vector extraction unit performs high-order local autocorrelation processing on the region determined by the feature vector extraction region determination unit in the image information filtered by the filter processing unit, and the feature of the defect candidate The defect inspection apparatus according to claim 1, wherein a vector is extracted. Of the image information filtered by the filter processing unit, a ratio of pixels having a luminance value equal to or greater than a predetermined value in the region input from the feature vector extraction region determination unit is set as the defect candidate degree value. A degree value calculating means for calculating,
The defect determination means includes a defect candidate type determined by the type determination means, a position of the defect candidate calculated by the position calculation means, an area of the defect candidate calculated by the area calculation means, and the degree 3. The presence / absence of a defect in the inspection object is determined by comparing each of the defect candidate degree values calculated by the value calculating means with the determination criterion given in advance. Defect inspection apparatus as described. The type discriminating unit discriminates the type of the defect candidate by using, as the boundary condition, a hyperplane determined by a discriminant function derived in advance from teacher image information whose type of defect candidate is known. The defect inspection apparatus according to any one of claims 1 to 3. A defect inspection method performed by a defect inspection apparatus that inspects the presence or absence of a defect of an inspection object based on image information of the inspection object,
The defect inspection apparatus is
By performing a binarization process on the image information of the inspection object input from the outside, the coordinate value of the pixel included in the defect candidate is extracted, and the labeling value and the coordinates as identification information for each defect candidate An area information acquisition step of acquiring area information of defect candidates associated with values;
A position calculating step for calculating the position of each defect candidate as a barycentric coordinate value based on the area information of the defect candidate acquired in the area information acquiring step;
An area calculating step for calculating an area for each defect candidate based on the area information of the defect candidate acquired in the area information acquiring step;
A feature vector extraction region determination step for determining a region for extracting a feature vector for each defect candidate based on the region information of the defect candidate acquired in the region information acquisition step;
A feature vector extraction step of performing a high-order local autocorrelation process on the image information before the binarization process and extracting a feature vector for each defect candidate in the region determined in the feature vector extraction region determination step When,
A type determination step of determining the type of the defect candidate by comparing each feature vector extracted in the feature vector extraction step with a boundary condition given in advance;
The defect candidate type determined in the type determination step, the position of the defect candidate calculated in the position calculation step, and the area of the defect candidate calculated in the area calculation step are given in advance. A defect determination step of performing a defect determination step of determining the presence or absence of a defect of the inspection object by comparing with a determination criterion. Including a filtering step of applying a bilateral filter to the image information of the inspection object input from the outside,
In the feature vector extraction step, high-order local autocorrelation processing is performed in the region determined in the feature vector extraction region determination step in the image information filtered in the filter processing step, and the feature of the defect candidate The defect inspection method according to claim 5, wherein a vector is extracted. Of the image information subjected to the filtering process in the filtering process step, a ratio of pixels having a luminance value equal to or larger than a predetermined value in the area determined in the feature vector extraction area determining step is set as the defect candidate degree value. A degree value calculating step for calculating,
In the defect determination step, the type of defect candidate determined in the type determination step, the position of the defect candidate calculated in the position calculation step, the area of the defect candidate calculated in the area calculation step, and the degree The presence / absence of a defect in the inspection object is determined by comparing each of the defect candidate degree values calculated in the value calculation step with the predetermined determination criterion. Described defect inspection method. The type of defect determination is characterized in that the type of defect candidate is determined using the hyperplane determined by a discriminant function derived in advance from teacher image information whose type of defect candidate is known as the boundary condition. Item 8. The defect inspection method according to any one of Items 5 to 7.   The program for making the said defect inspection apparatus which is a computer perform the defect inspection method as described in any one of Claims 5-8.

Images