JP5359377B2 - Defect detection apparatus, defect detection method, and defect detection program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a learning-type defect detection device, method and program, capable of maintaining the accuracy of defect detection regardless of the size of a defect generated in an inspection object. <P>SOLUTION: A plurality of partial domains BLK<SP>(1)</SP>, BLK<SP>(2)</SP>, etc, BLK<SP>(m)</SP>having each different size are extracted from an inspection image IMG. Each image vector is calculated from each image information of the extracted partial domains BLK<SP>(1)</SP>, BLK<SP>(2)</SP>, etc, BLK<SP>(m)</SP>, and each distance Out<SB>1</SB>, Out<SB>2</SB>, etc, Out<SB>m</SB>between each image vector and a corresponding eigenspace is calculated. Further, a response vector using each distance as elements is calculated, and the response vector is mapped on a characteristic space. Then, it is determined whether the response vector is a normal value or a deviation value based on a discrimination function determined beforehand. <P>COPYRIGHT: (C)2010,JPO&amp;INPIT

Description

  The present invention relates to a defect detection apparatus, a defect detection method, and a defect detection program for detecting a defect from an inspection image, and particularly to a technique for detecting a defect based on a learning result acquired from a learning image.

  Conventionally, in the FA (Factory Automation) field and the like, a defect inspection technique for detecting defects such as scratches and dust appearing on an inspection object (target work) has been developed. As a typical method of such defect inspection technology, defect inspection using so-called image measurement, in which defects are detected by performing image processing on an inspection image obtained by photographing an inspection object, has been put into practical use. ing.

  Among such defect inspections using image measurement, a method for detecting defects by acquiring a model corresponding to a defect in advance and comparing each area of the inspection image with this model is generally employed. I came.

  On the other hand, there has been proposed a method for detecting a portion different from the whole in the image due to the properties that hold in the whole of the image. For example, Non-Patent Document 1 discloses a defect detection method that focuses on the correlation between local regions of an image obtained by fractal nature and periodicity of the image.

  According to this defect detection method, a plurality of samples are cut out from the inspection image, and an image vector is calculated for each sample. Then, an eigenspace is generated by calculating eigenvector sequences for these image vector sequences, and a defect location is detected using the distance between the eigenspace and a sample (image vector). Such a detection method is called an Eigen Space Defect Detector (ESDD), and in this specification, a defect detection method based on a distance between an eigenspace and an image vector is also collectively referred to as “ESDD”. To do.

  Thus, according to ESDD, since it is not necessary to acquire a model corresponding to a defect in advance, even if it is difficult to model a defect that may occur in the inspection object in advance, the defect is detected. There is an advantage that you can.

Toshiyuki Amano, “Defect Detection Based on Image Correlation”, Image Recognition and Understanding Symposium (MIRU2005), July 2005

  However, when the defect generated in the inspection object is extremely large or extremely small as compared with the inspection sample, the defect detection accuracy may be lowered. For this reason, when there are variations in the size of the defect generated in the inspection object, there may be a problem that only a defect having a specific size is detected as a detection target, while other defects are not detected.

  Accordingly, the present invention has been made to solve such a problem, and the object thereof is a learning type defect detection apparatus capable of maintaining the detection accuracy of defects regardless of the size of defects occurring in an inspection object. A defect detection method and a defect detection program are provided.

  According to an aspect of the present invention, there is provided a defect detection device that detects a defect in an inspection image based on a learning result acquired from a learning image that does not include a defect. The defect detection apparatus includes means for extracting image information of a plurality of partial areas each including a first evaluation area set in an inspection image and having different sizes, and each of image information extracted from the inspection image A means for generating an inspection image vector for each size of the partial region, a means for calculating a distance between the corresponding eigenspace for each of the inspection image vectors for each size of the partial region, Means for referring to the discriminant function and determining the presence or absence of a defect in the first evaluation region based on a first response vector whose element is a distance for each size of the partial region.

  According to the present invention, a plurality of partial areas having different sizes are set for the first evaluation area, and between each partial area, an image vector generated from the image information and a corresponding eigenspace are provided. The distance is calculated. Furthermore, a first response vector having a distance for each size of each partial region as an element is generated, and the first response vector is evaluated with reference to a predetermined discriminant function, whereby the first response vector is evaluated. The presence or absence of defects in the evaluation area is determined. As described above, since the first evaluation area is determined based on the evaluation results in the units of the plurality of partial areas having different sizes, the presence / absence of the defect is determined. Therefore, the size of the defect that may occur in the inspection image. It is possible to suppress the influence due to the variation of the defect and to further improve the detection accuracy of the defect.

  Preferably, the defect detection apparatus includes image information of a plurality of partial areas each including a second evaluation area sequentially set in the learning image and having the same size as each of the plurality of partial areas extracted from the inspection image. From each of the means for extracting, the image information extracted from the learning image, the means for sequentially generating the evaluation area image vector for each size of the partial area, and for each of the evaluation area image vector for each size of the partial area, Based on the means for sequentially calculating the distance between the corresponding eigenspaces and the set of second response vectors each having a distance for each size of the subregion, the boundary between the inside and outside the set is determined. And a means for determining a discriminant function to be defined.

  More preferably, the means for determining the discriminant function determines the discriminant function according to the outlier detection method.

  Preferably, the defect detection apparatus sequentially extracts image information of a third partial area having the same size as one partial area from among the plurality of partial areas, and a third part that is sequentially extracted. Means for generating a learning image vector from image information of a region, means for calculating an eigenvector from a set of vectors including the learning image vector, and using a predetermined number of eigenvectors in descending order of eigenvalues among the calculated eigenvectors And a means for determining an eigenspace in the size of the third partial region.

  If another situation of this invention is followed, the defect detection apparatus which detects the defect in an inspection image based on the learning result acquired from the learning image which does not contain a defect is provided. The defect detection apparatus includes means for extracting image information of a plurality of partial areas each including a first evaluation area that is sequentially set in a learning image, and each of image information extracted from the learning image. Means for sequentially generating an evaluation area image vector for each size of the partial area, and means for sequentially calculating a distance from the corresponding eigenspace for each of the evaluation area image vectors for each size of the partial area; , A means for determining a discriminant function that defines the boundary between the inside and outside of the set based on a set of response vectors each having a distance for each size of the partial region as an element, Image information of a plurality of partial areas including two evaluation areas and having the same size as each of the plurality of partial areas extracted from the learning image is extracted and generated from the extracted image information And 査 image vector, based on a determination function, and means for determining the presence or absence of defects in the second evaluation region.

  Preferably, the means for determining the discriminant function determines the discriminant function according to the outlier detection method.

  According to still another aspect of the present invention, there is provided a defect detection method for detecting a defect in an inspection image based on a learning result acquired from a learning image using a computer including an arithmetic device and a storage device. In the defect detection method, the arithmetic device includes a first evaluation region set in the inspection image stored in the storage device and extracts image information of each of the plurality of partial regions having different sizes from each other; A step of generating an inspection image vector for each size of the partial area from each piece of image information extracted from the inspection image, and a distance between the corresponding eigenspace for each of the inspection image vectors for each size of the partial area And the presence / absence of a defect in the first evaluation region based on a first response vector whose element is a distance for each size of the partial region with reference to a discriminant function stored in advance in a storage device Determining.

  According to still another aspect of the present invention, there is provided a defect detection method for detecting a defect in an inspection image based on a learning result acquired from a learning image using a computer including an arithmetic device and a storage device. In the defect detection device, the arithmetic device includes a first evaluation region that is sequentially set in the learning image stored in the storage device, and extracts image information of each of the plurality of partial regions having different sizes from each other; A step of sequentially generating an evaluation area image vector for each size of the partial area from each piece of image information extracted from the learning image, and a corresponding eigenspace for each of the evaluation area image vectors for each size of the partial area A step of sequentially calculating a distance between them, a step of determining a discriminant function that defines a boundary between the inside and outside of the set based on a set of response vectors each having a distance for each size of a partial region as an element; A plurality of partial areas including a second evaluation area sequentially set in the examination image stored in the storage device and having the same size as each of the plurality of partial areas extracted from the learning image Extracted image information, respectively, and the inspection image vector generated from the image information to be the extraction, based on the discriminant function, and determining the presence or absence of a defect in the second evaluation region.

  According to still another aspect of the present invention, there is provided a defect detection program for detecting a defect in an inspection image based on a learning result acquired from a learning image using a computer. The defect detection program includes a computer that includes a first evaluation region set in an inspection image and that extracts image information of each of a plurality of partial regions having different sizes, and an image extracted from the inspection image Means for generating an inspection image vector for each size of the partial area from each of the information; means for calculating a distance between the corresponding eigenspace for each of the inspection image vectors for each size of the partial area; With reference to a predetermined discriminant function, it is made to function as means for determining the presence or absence of a defect in the first evaluation region based on a first response vector whose element is a distance for each size of the partial region.

  According to still another aspect of the present invention, there is provided a defect detection program for detecting a defect in an inspection image based on a learning result acquired from a learning image using a computer. The defect detection program includes a computer that includes a first evaluation region that is sequentially set in a learning image, and that extracts image information of a plurality of partial regions having different sizes, and an image extracted from the learning image From each piece of information, a means for sequentially generating an evaluation area image vector for each size of the partial area and a distance between the corresponding eigenspace for each evaluation area image vector for each size of the partial area are sequentially calculated. Means for determining a discriminant function that defines the boundary between the inside and outside of the set based on a set of response vectors each having a distance for each size of the partial area as an element, and sequentially setting the inspection image Image information of a plurality of partial areas including the second evaluation area to be extracted and having the same size as each of the plurality of partial areas extracted from the learning image is extracted. A test image vector generated from the image information, based on the discriminant function, to function as means for determining the presence or absence of a defect in the second evaluation region.

  According to the present invention, it is possible to maintain the detection accuracy of a defect regardless of the size of the defect generated in the inspection object.

It is the schematic which shows the whole structure of the defect detection system containing the defect detection apparatus according to Embodiment 1 of this invention. It is a schematic block diagram which shows the hardware constitutions of a computer. It is a figure for demonstrating the process in the case of performing defect detection using learning type | mold ESDD. It is a figure which shows an example of the defect which generate | occur | produces in a test | inspection image. It is a figure for demonstrating notionally the magnitude | size of the partial area | region set to the defect detection apparatus according to Embodiment 1 of this invention, and its evaluation. It is a figure for demonstrating notionally the learning process and defect detection process in the defect detection apparatus according to Embodiment 1 of this invention. It is a functional block diagram which shows the control structure of the defect detection apparatus according to Embodiment 1 of this invention. It is a figure for demonstrating notionally the production | generation process of the image vector in the defect detection apparatus according to Embodiment 1 of this invention. It is a figure for demonstrating notionally the image information of the partial area | region in the defect detection apparatus according to Embodiment 1 of this invention. It is a figure which shows an example of the setting method of several partial area | regions. It is a figure which shows an example of the setting method of the partial area | region set corresponding to an evaluation area | region. It is a figure which shows the upper limit vector size and lower limit vector size of the partial area which can be included in a partial area set. It is a figure for demonstrating the process when a part of partial area set protrudes from a learning image. It is a figure for demonstrating the distance of an eigenspace and an image vector. It is a figure for demonstrating the distance of a one-dimensional eigenspace and an image vector. It is a figure for demonstrating the calculation method of the distance with respect to the eigenspace of the area | region where a partial area | region is overlapped and set. It is a flowchart which shows the process sequence for detecting the defect which may generate | occur | produce in the object workpiece | work in the defect detection system according to Embodiment 1 of this invention. It is a flowchart which shows the process sequence of the learning process according to Embodiment 1 of this invention. It is a flowchart which shows the process sequence of the defect detection process according to Embodiment 1 of this invention. It is a functional block diagram which shows the control structure of the defect detection apparatus according to Embodiment 2 of this invention. It is a figure which shows an example of the object workpiece | work which has a periodic texture. It is a figure which shows an example of the partial area | region extracted when the position of the periodic shading pattern which appears in a test | inspection image varies. It is an example of the figure which showed visually the distance with respect to eigenspace about the typical partial area | region BLK corresponding to the evaluation area | region set to the test | inspection image IMG. It is a figure for demonstrating the search process of the reference position in a test | inspection image. It is a figure for demonstrating the determination method of the search area | region for searching the reference | standard position in a test | inspection image. It is a figure for demonstrating the period of a shading pattern. It is a figure which shows an example of the distance calculated in order in a distance calculation part. It is a figure which shows an example of the candidate point extracted with respect to the test | inspection image. It is a figure for demonstrating the partial area | region set with respect to a test | inspection image by the partial area | region setting part. It is a figure for demonstrating the process in case a part of set partial area | region protrudes from a test | inspection image. It is a figure which shows an example of the test | inspection image with the periodic shading pattern made into object. It is a figure which shows the result at the time of applying the method according to the conventional method and this Embodiment with respect to the test | inspection image shown in FIG. It is a flowchart which shows the process sequence for detecting the defect which may generate | occur | produce in the object workpiece | work in the defect detection system according to Embodiment 2 of this invention. It is a flowchart which shows the process sequence of the reference | standard position search process according to Embodiment 2 of this invention. It is a figure which shows an example of the image SMP obtained by imaging the object workpiece | work according to the modification of Embodiment 2 of this invention. It is a flowchart which shows the process sequence for producing | generating the image vector about the target workpiece | work which has periodicity in the defect detection system according to the modification of Embodiment 2 of this invention. It is the schematic which shows the defect detection system containing the defect detection apparatus according to Embodiment 3 of this invention.

  Embodiments of the present invention will be described in detail with reference to the drawings. Note that the same or corresponding parts in the drawings are denoted by the same reference numerals and description thereof will not be repeated.

[Embodiment 1]
(Overall equipment configuration)
FIG. 1 is a schematic diagram showing an overall configuration of a defect detection system 1 including a defect detection apparatus according to Embodiment 1 of the present invention.

  Referring to FIG. 1, a defect detection system 1 is typically incorporated in a production line or the like, and detects the presence / absence of a defect and a position where a defect occurs in a workpiece to be subjected to defect inspection (hereinafter also referred to as a target workpiece). . Defects typically mean scratches, dust, etc. that occur in the target workpiece.

  In the defect detection system 1 according to the present embodiment, the target work 2 is transported by a transport mechanism 6 such as a belt conveyor, and the transported target work 2 is sequentially photographed by the imaging unit 8. An image (hereinafter, also referred to as “inspection image”) captured by the imaging unit 8 is transmitted to a computer 100 that is a representative example that implements the defect detection apparatus according to the present embodiment.

  The imaging unit 8 includes, for example, an imaging element such as a CCD (Coupled Charged Device) or a CMOS (Complementary Metal Oxide Semiconductor) sensor and a lens, and images the target workpiece 2 or a reference workpiece 4 described later. Note that the imaging unit 8 performs imaging according to an imaging command generated by the computer 100 in accordance with the conveyance operation of the conveyance mechanism 6.

  An image captured by the imaging unit 8 typically has image information in units of pixels. Alternatively, image information may be provided in units of blocks in which a plurality of pixels are collected. As such “image information”, if the imaging unit 8 is a three-band color camera, the “red”, “green”, and “blue” gradation values (R component value, G component value, Information specified by (B component value) is used. Alternatively, information specified by gradation values (C component value, M component value, Y component value) of “cyan”, “magenta”, and “yellow”, which are complementary colors of the three primary colors of light, may be used as the image information. Good. If the imaging unit 8 is a color camera having a larger number of bands, information specified by a number of gradation values corresponding to the number of bands is used.

  Alternatively, if the imaging unit 8 is a monochrome camera, information specified by gradation values (gray scale) indicating light and shade may be used. When this gradation value is divided into two stages, it represents a binarized image of “white” or “black”.

  Furthermore, instead of the gradation value as described above, a color attribute including parameters of “Hue”, “Lightness (Value)”, and “Saturation (Chroma)” may be used as the image information.

  In particular, the defect detection apparatus according to the present embodiment images a reference work 4 to be used as a reference for defect inspection in advance, and acquires a learning result from an image obtained by this imaging (hereinafter also referred to as “learning image”). Based on this learning result, a defect in the inspection image is detected (hereinafter, defect detection by the defect detection apparatus according to the present invention is also referred to as “learning ESDD”). Therefore, it is necessary to select the same type of product as the target workpiece 2 as the reference workpiece 4. Note that it is desirable that the reference workpiece 4 (that is, the learning image) does not include a defect. Here, “does not include a defect” means that a portion determined to be a defect under a predetermined determination criterion is not included.

  Further, in this specification, as a representative example of the learning result, a discriminant function for evaluating a response vector having as an element the distance between a plurality of image vectors having different vector sizes and the corresponding eigenspace is used. The discriminant function and response vector will be described later.

  On the other hand, as the target workpiece 2, the same type of products that are continuously produced, such as electronic parts and plastic parts, are suitable. In addition, the magnitude | size, the kind, etc. of the object workpiece | work which the defect detection apparatus based on this invention makes the object of defect detection are not restrict | limited. Furthermore, the present invention is not limited to those manufactured on the production line of a factory, and scratches on agricultural products such as apples and scratches on movie films can also be detected.

  The defect detection system 1 further includes an illumination unit 10 for illuminating the vicinity of the imaging range of the imaging unit 8 and a light source unit 12 that drives the illumination unit 10. The light source unit 12 irradiates a predetermined amount of light from the illumination unit 10 in accordance with an illumination command from the computer 100. It is desirable that the light emitted from the illumination unit 10 has a relatively wide wavelength spectrum.

  The computer 100 also includes a computer main body 101 having an FD (Flexible Disk) driving device 111 and a CD-ROM (Compact Disk-Read Only Memory) driving device 113, a monitor 102, a keyboard 103, and a mouse 104. . Computer 100 implements a defect detection apparatus according to the present embodiment by executing a program stored in advance.

(Hardware configuration)
FIG. 2 is a schematic configuration diagram illustrating a hardware configuration of the computer 100.

  2, in addition to the FD drive device 111 and the CD-ROM drive device 113 shown in FIG. 1, a computer main body 101 includes a CPU (Central Processing Unit) 105 that is an arithmetic device connected to each other via a bus. A memory 106, a fixed disk 107 as a storage device, and an interface 109.

  An FD 112 is mounted on the FD driving device 111, and a CD-ROM 114 is mounted on the CD-ROM driving device 113. As described above, the defect detection apparatus according to the present embodiment is realized by CPU 105 executing a program using computer hardware such as memory 106. In general, such a program is stored in a recording medium such as the FD 112 or the CD-ROM 114, or distributed via a network or the like. Such a program is read from a recording medium by the FD driving device 111 or the CD-ROM driving device 113 or received by the interface 109 and stored in the fixed disk 107. Further, it is read from the fixed disk 107 to the memory 106 and executed by the CPU 105.

  The monitor 102 is a display unit for displaying information output by the CPU 105, and includes, for example, an LCD (Liquid Crystal Display) or a CRT (Cathode Ray Tube). The mouse 104 receives a command from the user according to an operation such as click or slide. The keyboard 103 receives a command from the user corresponding to the input key. The CPU 105 is an arithmetic processing unit that performs various calculations by sequentially executing programmed instructions. The memory 106 stores various types of information in accordance with program execution by the CPU 105. The interface 109 is a device for establishing communication between the computer 100 and the imaging unit 8 (FIG. 1) and the illumination unit 10 (FIG. 1). The information output from the CPU 105 is converted into, for example, an electrical signal, and the like. Are sent to these devices, and electrical signals are received from these devices and converted into information that can be used by the CPU 105. In particular, image data captured by the imaging unit 8 is stored in the memory 106 or the fixed disk 107 via the interface 109. The fixed disk 107 is a non-volatile storage device that stores programs executed by the CPU 105, image data, and the like. The computer 100 may be connected to another output device such as a printer as necessary.

(Defect detection processing by learning type ESDD)
FIG. 3 is a diagram for explaining processing when defect detection is performed using learning-type ESDD. 3A shows an example of a learning image LNIMG that does not include a defect, FIG. 3B shows an example of an inspection image IMG that includes a defect DEF, and FIG. 3C does not include a defect DEF. An example of the partial region BLK1 is shown, and FIG. 3D shows an example of the partial region BLK2 including the defect DEF.

Referring to FIG. 3A, first, according to learning-type ESDD, partial regions BLK ^LN of a predetermined size are sequentially extracted from a learning image LNIMG that does not include a defect obtained by photographing the reference workpiece 4. The Then, properties (features) appearing as a whole of the learning image LNIMG are acquired in advance (learning). 3A shows an example in which the partial regions BLK ^LN are sequentially extracted so as not to overlap each other in order to simplify the drawing, the partial regions BLK ^LN may be set so as to overlap each other.

With reference to FIG. 3B, subsequently, according to learning-type ESDD, partial areas BLK1, BLK2, and the like having a predetermined size are extracted from the inspection image IMG obtained by photographing the target workpiece 2. Note that the partial areas BLK1 and BLK2 extracted from the inspection image IMG have the same size as the partial areas BLK ^LN extracted from the learning image LNIMG.

  As shown in FIG. 3C, when the partial region BLK1 does not include the defect DEF, the partial region BLK1 has substantially the same property as that of the entire learning image LNIMG. On the other hand, as shown in FIG. 3D, when the partial region BLK2 includes a defect DEF, the partial region BLK2 has a characteristic property that is different from the property that appears as a whole of the learning image LNIMG. .

  As described above, according to the learning type ESDD, the property that appears as the entire learning image LNIMG obtained by photographing the reference workpiece 4 is acquired (learned) in advance, and the property that appears as the entire learning image LNIMG acquired in advance. And the property appearing in the corresponding partial region BLK of the inspection image IMG, thereby detecting the defect DEF in the inspection image IMG.

FIG. 4 is a diagram illustrating an example of the defects DEF1 and DEF2 generated in the inspection image IMG.
Referring to FIG. 4, depending on the target workpiece, the size of the generated defect may vary. Therefore, in the inspection image IMG, there may be a defect DEF1 that fits in the partial area BLK, or a defect DEF2 that does not fit in one partial area BLK.

On the other hand, it is often difficult to predict the size of the generated defect in advance, and it is difficult to optimize the sizes of the partial regions BLK ^LN and BLK in advance. Further, it is difficult to optimize the sizes of the partial regions BLK ^LN and BLK even when the defect sizes vary.

Therefore, unlike the ESDD disclosed in Non-Patent Document 1, the defect detection apparatus according to the present invention does not fix the sizes of the partial regions BLK ^LN and BLK, and a plurality of partial regions BLK having different sizes from each other. Images are evaluated using ^LN and BLK to detect defects. Hereinafter, the sizes of the partial areas BLK ^LN and BLK are also referred to as “vector sizes”.

(Learning process and defect detection process)
FIG. 5 is a diagram for conceptually explaining the size of the partial area set in the defect detection apparatus according to the first embodiment of the present invention and its evaluation.

  FIG. 6 is a diagram for conceptually explaining learning processing and defect detection processing in the defect detection device according to the first embodiment of the present invention. FIG. 6A conceptually shows the learning process, and FIG. 6B conceptually shows the defect detection process.

Referring to FIG. 5, the defect detection apparatus according to the present embodiment includes a plurality of partial areas BLK ^{EV having} different sizes from learning image LNIMG or inspection image IMG (hereinafter sometimes simply referred to as “image”). ^{(1), BLK EV (2} ), ···, BLK EV (m) or ^{BLK (1), BLK (2} ), extracts · · ·, ^{BLK (m)} is, respectively. At this ^{time, BLK EV (1), BLK} EV (2), ···, BLK EV (m) or partial regions ^{BLK (1), BLK (2} ), ···, of ^{BLK (m),} the smallest (Narrow) partial areas are the respective evaluation areas. The other partial areas are set so as to include the smallest area corresponding to the evaluation area (partial area BLK ^{EV (1)} or BLK ^{(1) in} FIG. 5). Hereinafter, the partial area corresponding to a certain evaluation area ^{BLK EV (1), BLK EV} (2), ···, BLK EV (m) or ^{BLK (1), BLK (2} ), ···, BLK (m) Is also referred to as “partial region set SET”.

Then, each image vector is calculated from each image information of the extracted partial areas BLK ^{EV (1)} , BLK ^{EV (2)} ,..., BLK ^{EV (m)} , and further, a unique corresponding to each image vector is calculated. The distances Out ₁ , Out ₂ ,..., Out _m between the spaces are calculated. Through such a process, the response vector RV = [Out ₁ , Out ₂ ,..., Out _m ] for the partial region set SET corresponding to each evaluation region can be calculated.

As described above, the partial areas BLK ^{EV (1)} , BLK ^{EV (2)} ,..., BLK ^{EV (m)} set in the learning image LNIMG are used for calculating the response vector RV. This is used to calculate a discriminant function for evaluating each evaluation region set in the inspection image IMG. On the other hand, a partial region BLK ^LN described later is mainly used for calculating an eigenspace (eigenvector) in the learning image LNIMG. Actually, it is necessary to calculate eigenspaces (eigenvectors) for each size (vector size) of the partial areas BLK ^{EV (1)} , BLK ^{EV (2)} ,..., BLK ^{EV (m).} Although it may be a region, it is described with a different symbol for easy understanding.

  Referring to FIG. 6A, the defect detection apparatus according to the present embodiment maps response vector RV calculated corresponding to each evaluation region of learning image LNIMG to the feature space in the learning process. This feature space is a superspace having a dimension corresponding to each component constituting the response vector. That is, in this feature space, the distance between the image vector and the eigenspace at each vector size is set as a base (coordinate).

  Then, the defect detection apparatus according to the present embodiment distinguishes a normal value from an outlier value with respect to a set of response vectors mapped to the feature space. Such a distinction is realized by using an outlier detection method in the data mining field. More specifically, the defect detection apparatus according to the present embodiment determines a discriminant function for distinguishing between a normal value and a deviation value. That is, the discriminant function is a function that defines the boundary between the inside and outside of the set of response vectors.

  The procedure for determining such a discriminant function is as follows. First, a plurality of response vectors RV calculated from the learning image LNIMG are clustered. For this clustered result, one or more classes (corresponding to non-defect portions) are defined based on the distance and density distribution. That is, the standard for the difference of each class is determined. A function indicating a criterion for this class difference corresponds to a discriminant function. Then, based on this discriminant function, if the response vector RV calculated from any inspection image IMG does not belong to any defined class, it is classified as a “reject class”. To be judged.

  In the defect detection apparatus according to the present embodiment, since it is only necessary to determine whether the response vector RV is “normal value” or “outlier value”, a so-called “one-class problem” in which only one class is defined. It can be determined using the method. As typical techniques for such a one-class problem, 1-class SVM (One-Class Support Vector Machine) and SVDD (Support Vector Data Description) are known. For 1-class SVM, refer to, for example, literature (Siwei Lyu and Hany Farid, Steganalysis Using Color Wavelet Statistics and One-Class Support Vector Machines, Department of Science, Dartmouth College, Hanover, NH 03755, USA). .

  Alternatively, K-NN method, K-Means method, MCD (Minimum Covariance Determinant) method, GMM (Gaussian Mixture Model) method, GP (Gaussian Process) method or the like is used as a method for determining such discriminant function. You can also. The outlier detection method described above is described in the literature (Sanjay Chawla and Pei Sun, "Clustering and Outlier Detection in High Dimensional Data", School of Information Technologies, University of Sydney NSW, Australia, http: //www.dbs .ifi.lmu.de / CODHDD09 / [Internet]).

  As described above, the defect detection device according to the present embodiment determines a discriminant function (learning result) on the feature space in advance in the learning process.

  Referring to FIG. 6B, in the defect detection process, the defect detection apparatus according to the present embodiment is similar to the above-described process in that the response vector for partial area set SET corresponding to each evaluation area of inspection image IMG. RV ′ is sequentially calculated and mapped to the feature space. Then, the defect detection apparatus according to the present embodiment determines whether or not this mapped response vector RV ′ is a normal value or a deviation value based on a predetermined discriminant function. That is, it is determined whether or not the mapped response vector RV ′ is present in the set of response vectors calculated from the learning image. If the response vector is an outlier, the defect detection device according to the present embodiment detects that a defect exists in the corresponding evaluation region.

  As described above, the defect detection apparatus according to the present embodiment determines the presence / absence of a defect in the target evaluation region by integrating the evaluation results for each of the plurality of partial regions having different sizes including the evaluation region. To do. Therefore, even when there is a relative variation between the size of the defect and the size of the partial area, the influence of the defect detection due to this variation can be suppressed, and the detection accuracy of the defect can be further increased.

(Control structure)
Next, a control structure for realizing the above processing will be described.

  FIG. 7 is a functional block diagram showing a control structure of the defect detection apparatus according to the first embodiment of the present invention.

  Referring to FIG. 7, the defect detection apparatus according to the present embodiment functions as image buffer units 202 and 222, extraction units 204, 214 and 224, vector generation units 206, 216 and 226, and eigenvector calculation. Unit 208, eigenspace determination unit 210, distance calculation units 218 and 228, classification unit 220, determination unit 230, defect detection unit 232, and learning result storage unit 240.

  The image buffer unit 202 is formed in a predetermined area of the memory 106 (FIG. 1), and temporarily stores a learning image input from the outside of the apparatus. In this embodiment, a configuration in which a learning image is input from the imaging unit 8 to the image buffer unit 202 is illustrated. However, a learning image captured by a device other than the imaging unit 8 is transmitted via a recording medium, a network, or the like. Thus, it may be stored in advance in the fixed disk 107 (FIG. 2).

The extraction unit 204, the vector generation unit 206, the eigenvector calculation unit 208, and the eigenspace determination unit 210 determine an eigenspace that serves as a reference for calculating the distance between each image vector. The eigenspace is determined for each size (vector size) of the partial area BLK ^EV included in the partial area set SET. That is, when each partial region set SET includes m types of partial regions BLK ^EV (or BLK) having different vector sizes, m eigenspaces are determined.

The extraction unit 204 extracts, for each of the plurality of partial regions BLK ^EV included in the partial region set SET, partial regions BLK ^LN having a corresponding vector size from the learning image temporarily stored in the image buffer unit 202. Then, the image information of the learning image corresponding to each extracted partial region BLK ^LN is output to the vector generation unit 206. In other words, the extraction unit 204 sequentially extracts learning samples (local regions) for each vector size from the learning image temporarily stored in the image buffer unit 202.

The vector generation unit 206 uses the learning image vectors x ₁ ^{LN (j)} , x ₂ ^{LN (j)} , ^... For each vector size j from the image information of the partial regions BLK ^LN of the learning images sequentially extracted by the extraction unit 204. .., X _{R (j)} ^{LN (j)} is generated sequentially (where R (j) is the total number of learning samples at each vector size j).

  FIG. 8 conceptually illustrates image vector generation processing in the defect detection apparatus according to the first embodiment of the present invention.

  FIG. 9 is a diagram for conceptually explaining image information of a partial region in the defect detection device according to the first embodiment of the present invention.

Referring to FIG. 8, extraction unit 204 sequentially sets partial areas for each of a plurality of different vector sizes for learning image LNIMG stored in image buffer unit 202. FIG. 8A shows a state in which a partial area BLK ^{LN (1) of} vector size 1 is set for the learning image LNIMG. Similarly, FIG. 8B and FIG. A state where a partial area BLK ^{LN (2)} of vector size ² and a partial area BLK ^{LN (m) of} vector size m are set for each of the learning images LNIMG.

Referring to FIG. 9, if a partial area ^{BLK LN} is w pixels × h magnitude of the pixel (vector size), extraction unit 204, for each of the partial regions ^{BLK LN (j)} is set, N (= The w × h) pieces of image information c _i are output to the vector generation unit 206. Then, the vector generation unit 206 generates the following learning image vector for each of the set partial areas BLK ^LN .

Learning image vector x _i ^{LN (j)} = [c ₁ , c ₂ ,..., C _{N (j)} ] ^T
However, N (j) = w × h, 1 ≦ j ≦ m
Each pixel is the R component value, G component value, if made of the B component value, the image information c _i is the 3-dimensional array containing these component values. Therefore, the learning image vector x _i ^{LN (j)} is as follows.

x _i ^{LN (j)} = [R ₁ , G ₁ , B ₁ , R ₂ , G ₂ , B ₂ ,..., R _{N (j)} , G _{N (j)} , B _{N (j)} ] ^T
The learning image vector x _i ^{LN (j)} for each partial region BLK ^{LN (j)} generated by the vector generation unit 206 can be defined as a point on the N-dimensional Euclidean space.

Referring again to FIG. 7, eigenvector calculation section 208 calculates eigenvectors for each vector size from a set of learning image vectors x _i ^{LN (j)} sequentially generated by vector generation section 206. More specifically, the eigenvector calculation unit 208 sets the following set vector X ^{LN (j)} for each vector size as a set (set) of learning image vectors x _i ^{LN (j)} generated by the vector generation unit 206. Is calculated.

Set vector X ^{LN (j)} = [x ₁ ^{LN (j)} , x ₂ ^{LN (j)} ,..., X _{R (j)} ^{LN (j)} ]
However, R (j) is the total number (the total number of learning samples) of the partial areas BLK ^LN set in the learning image LNIMG in the corresponding vector size j. The eigenvector calculation unit 208 uses the set vector X ^{LN (j)} The eigenvector V ^(j) for diagonalizing the variance matrix Q ^(j) = X ^{LN (j)} · X ^{LN (j) T} is calculated. That is, the eigenvector calculation unit 208 calculates the eigenvector V ^(j) and the eigenvalue matrix Δ ^(j) so that the following expression is established.

Dispersion matrix Q ^(j) = X ^{LN (j)} · X ^{LN (j) T} = V ^(j) · Δ ^(j) · V ^{(j) T}
Eigenvector V ^(j) = [e ₁ ^(j) , e ₂ ^(j) ,..., E _R ^(j) ]
Here, the basis vectors e ₁ ^(j) , e ₂ ^(j) ,..., E _R ^(j) satisfy the following conditions.

The eigenvalue matrix Δ ^(j) is represented as a diagonal matrix of eigenvalues λ ₁ ^(j) , λ ₂ ^(j) ,..., Λ _R ^(j) as follows.

  Note that the eigenvector and eigenvalue calculation processes as described above can be realized using a known algorithm such as KL (Karhunen-Loeve) expansion.

Here, each of the basis vectors e ₁ ^(j) , e ₂ ^(j) ,..., E _R ^(j) constituting the eigenvector V ^(j) is obtained by sampling the learning image LNIMG with a corresponding vector size. In addition, it represents the property that appears as a whole of the learning image LNIMG, and the greater the eigenvalue, the greater the degree of representation. Here, since it is the eigenvalues ^{_{λ 1 (j) ≧ λ 2}} (j) ≧ ··· ≧ λ R (j), basis vectors ^{_{e 1 (j), e 2}} (j), ···, e R ( ^This means that the degree of representing the property appearing as a whole of the learning image LNIMG in the order of ^j) is large.

Therefore, in the defect detection device according to the present embodiment, the eigenvalue is considered in consideration of the physical meaning of such basis vectors e ₁ ^(j) , e ₂ ^(j) ,..., E _R ^(j) . The eigenspace in each vector size is determined using a predetermined number of basis vectors in order from the largest. That is, the eigenspace is a subspace calculated using a part of the basis vectors e ₁ ^(j) , e ₂ ^(j) ,..., E _R ^(j) generated by the eigenvector calculation unit 208. . The eigenspace calculated in this way reflects the properties appearing as a whole of the learning image LNIMG to the extent corresponding to the dimension (order) of the base vector. Here, if the fractal nature and periodicity in the learning image LNIMG are dominant, the characteristics of most of the learning image LNIMG are approximately approximated even in the eigenspace defined by the relatively small dimensional basis vectors. obtain.

The eigenspace determination unit 210 has a large eigenvalue among eigenvectors V ^(j) = [e ₁ ^(j) , e ₂ ^(j) ,..., E _R ^(j) ] calculated by the eigenvector calculation unit 208. Using a predetermined number of basis vectors e ₁ ^(j) , e ₂ ^(j) ,..., E _{dim (j)} ^(j) (dim (j) ≦ R (j)) The eigenspace E ^(j) is determined. This eigenspace E ^(j) is a hyperplane of dimension dim (j). Note that the dimension dim (j) can be arbitrarily determined according to the property appearing in the learning image LNIMG, but the calculation amount increases to a projection vector, which will be described later, as the dimension dim (j) increases. Therefore, it is desirable to determine an appropriate dimension based on the cumulative contribution as disclosed in Non-Patent Document 1. Then, the eigenspace determination unit 210 outputs the determined eigenspace E ^(j) to the distance calculation unit 218 and stores it in the learning result storage unit 240.

  Next, the extraction unit 214, the vector generation unit 216, the distance calculation unit 218, and the classification unit 220 calculate a response vector for the partial region set SET corresponding to each evaluation region set in the learning image LNIMG. At the same time, a discriminant function for discriminating between a normal value and an outlier value is determined for the set of response vectors.

The extraction unit 214 sequentially sets evaluation regions in the learning image LNIMG, and also associates with each evaluation region, and includes partial regions BLK ^{EV (1)} , BLK ^{EV (2)} ,. BLK ^{EV (m)} is sequentially extracted. The sizes of the partial areas BLK ^{EV (1)} , BLK ^{EV (2)} ,..., BLK ^{EV (m)} are equal to the eigenspace E ^(j) generated by the vector generation unit 206 (where 1 ≦ j ≦ m) is equal to the size of the partial area corresponding to each.

Then, the extraction unit 214 stores image information of learning images corresponding to the partial areas BLK ^{EV (1)} , BLK ^{EV (2)} ,..., BLK ^{EV (m)} (partial area set) to be sequentially extracted. The information is read from the unit 202 and the read image information is output to the vector generation unit 216. As described above, the partial areas BLK ^{EV (1)} , BLK ^{EV (2)} ,..., BLK ^{EV (m)} are all set to include the corresponding evaluation areas.

Here, the partial areas BLK ^{EV (1)} , BLK ^{EV (2)} ,..., BLK ^{EV (m)} are set in any manner as long as they include the evaluation area (corresponding to the smallest partial area BLK). May be. Therefore, as shown in FIG. 5, the partial areas may be set so as to sequentially expand in the lower right direction on the paper with respect to the evaluation area (corresponding to the partial area BLK ^{EV (1))} , or the evaluation area is centered. The partial area may be set so as to sequentially expand in the peripheral direction.

FIG. 10 is a diagram illustrating an example of a method of setting a plurality of partial areas BLK ^{EV (1)} , BLK ^{EV (2)} ,..., BLK ^{EV (m)} . In FIG. 10, the partial areas BLK ^{EV (2) to} BLK ^{EV (m )} are enlarged so as to expand in the peripheral direction with the evaluation area (corresponding to the partial area BLK ^{EV (1))} set in the learning image LNIMG as the center. ⁾ Is set.

The evaluation area can be set at an arbitrary location in the learning image LNIMG.
FIG. 11 is a diagram illustrating an example of a setting method of the partial region set SET corresponding to the evaluation region. As illustrated in FIG. 11A, the extraction unit 214 (FIG. 7) sequentially sets the partial region set SET by shifting the partial region set SET by a predetermined interval with respect to the learning image LNIMG as an example. At this time, the partial area set SET1 and a part of the adjacent partial area set SET2 may overlap each other. Furthermore, as shown in FIG. 11B, a partial region set SET1 and a partial region set SET2 that is separated from the partial region set SET1 by a predetermined interval may be set. Further, the interval between adjacent partial region sets may not be constant, such as between the partial region set SET1 and the partial region set SET2, and between the partial region set SET2 and the partial region set SET3.

  The size of the partial area included in the partial area set set by the extraction unit 204 and the extraction unit 214 (FIG. 7) can be determined according to the minimum / maximum vector size setting from the user or the like. That is, the size of the minimum partial region included in the partial region set SET is set to the minimum vector size, and the size of the maximum partial region is also set to the maximum vector size. Further, a necessary number of partial areas are set between the minimum vector size and the maximum vector size.

  FIG. 12 is a diagram illustrating an upper limit vector size and a lower limit vector size of a partial area that can be included in the partial area set. Referring to FIG. 12, the lower limit value of the vector size that can be set as a partial region is one pixel that is the minimum unit constituting learning image LNIMG. On the other hand, the upper limit of the vector size that can be set as the partial area is the entire learning image LNIMG, that is, the image size of the learning image LNIMG.

  In addition, when a part of the partial region set protrudes from the learning image LNIMG, a specific process may be performed.

  FIG. 13 is a diagram for explaining processing when a part of the partial region set SET protrudes from the learning image LNIMG.

As an example of processing when a part of the partial region set SET protrudes from the learning image LNIMG, as shown in FIG. 13A, the output of the image information of the protruding part may be omitted. In the example shown in FIG. 13A, since a partial area of the partial area BLK ^{EV (m)} is a protruding part, image information about this part is not output. In such a case, the size of a corresponding image vector generated by the vector generation unit 216 described later is reduced, but the presence or absence of a defect is evaluated based on the distance between the image vector and the corresponding eigenspace. Therefore, there is no influence due to the variation in the size of the image vector.

  Alternatively, as shown in FIG. 13B, interpolation processing may be performed on a portion that protrudes from the learning image LNIMG based on image information of a portion adjacent to the portion. By performing such interpolation processing, even if a part of the partial region set includes a partial region that protrudes from the learning image LNIMG, the same image vector as the partial region set in which all the partial regions fit within the learning image LNIMG. Can be generated. Since this interpolation process treats the protruding part as having the same property as the adjacent part, in the defect detection method based on ESDD, the decrease in detection accuracy due to this interpolation process can be almost ignored.

Referring to FIG. 7 again, the vector generation unit 216 is configured to extract the partial region sets (partial regions BLK ^{EV (1)} , BLK ^{EV (2)} ,..., BLK ^{EV (m)} sequentially extracted by the extraction unit 214. ), The evaluation area image vectors x _k ^{EV (1)} , x _k ^{EV (2)} ,..., X _k ^{EV (m)} are sequentially generated. Note that k is a label number indicating a partial region set (evaluation region) set in the learning image LNIMG.

  Since the generation process of the evaluation area image vector in the vector generation unit 216 is the same as the generation process of the learning image vector in the vector generation unit 206 described above, detailed description will not be repeated. Then, the vector generation unit 216 outputs the generated evaluation area image vector to the distance calculation unit 218.

The distance calculation unit 218 outputs a vector size corresponding to each of the evaluation region image vectors x _k ^{EV (1)} , x _k ^{EV (2)} ,..., X _k ^{EV (m)} output from the vector generation unit 216. The distances Out ₁ (k), Out ₂ (k),..., Out _m (k) to the eigenspace E ^(j ) are sequentially calculated.

  Here, a method for calculating the distance between the eigenspace E and the image vector x will be generalized and described.

FIG. 14 is a diagram for explaining the distance Out between the eigenspace E and the image vector x.
Referring to FIG. 14, the distance Out between the eigenspace E and the image vector x is defined as the Euclidean distance, and this Euclidean distance is expressed as follows using the projection point x ′ of the image vector x on the eigenspace E: It can be expressed by the following formula.

| Out | ² = | xx ′ | ² = (xx ′) ^T (xx ′)
Further, the projection point x ′ of the image vector x on the eigenspace E can be expressed by the following equation.

Next, the distance between the eigenspace E ⁽¹⁾ and the image vector x will be described with reference to FIG.

FIG. 15 is a diagram for explaining the distance Out between the one-dimensional eigenspace E ⁽¹⁾ and the image vector x.

With reference to FIG. 15, consider the distance P from the origin to the projection point x ′. Since the one-dimensional eigenspace E ⁽¹⁾ is defined by the basis vector e ₁ , the vector set P can be expressed by the following equation using the image vector x and the basis vector e ₁ .

P = x · e ₁
When this is expanded to the dim dimension, a set of distances P from the origin to the projection point x ′ can be expressed by the following equation.

P = (P ₁ , P ₂ ,..., P _dim ) ^T
Therefore, the distance from the origin to the projection point x ′ in an arbitrary component is P _m = x · e _m (1 ≦ m ≦ dim).

As shown in FIG. 15, if the eigenspace ⁽¹⁾ is one-dimensional, the projection point x ′ on the eigenspace ⁽¹⁾ (base vector e ₁ plane) is considered to be P times in the direction of the base vector e _1. Therefore, x ′ = Pe ₁ is satisfied. However, the basis vector e ₁ is a unit vector.

  When this is expanded to the dim dimension, the projection point x ′ can be expressed by the following equation.

  Consider again the distance Out between the eigenspace E and the image vector x. As mentioned above,

So
| Out | ² = | xx ′ | ² = (xx ′) ^T (xx ′)
= (X-EP) ^T (x-EP) (where P = E ^T x)
^{= X T x-x T EP} -P T E T x + P T E T EP
Here, the following equation holds, so

^{| Out | 2 = x T x} -P T P-P T P + P T P
= ^X T ^x-P T ^P
As described above, the distance Out between the eigenspace E and the image vector x can be expressed by the following equation using the image vector x and the vector set P.

^{| Out | 2 = x T x} -P T P
Here, vector set P corresponds to the projection vector for projecting an image vector x to the projection point x ', P T ^P is equivalent to the inner product value of the projection vector.

Referring to FIG. 7 again, the distance calculation unit 218 outputs the evaluation area image vectors x _k ^{EV (1)} , x _k ^{EV (2)} ,..., X _k ^{EV (m )} output from the vector generation unit 216. ^{) Between} the eigenspaces E ^(j) output from the eigenspace determining unit 210 and eigenspaces having the same vector size Out ₁ (k), Out ₂ (k),.・, Out _m (k) is calculated sequentially.

| Out _j (k) | ² = x _k ^{EV (j) T} · x _k ^{EV (j)} −P _k ^{(j) T} · P _k ^{(j)
_{However, P k (j) = E}} (j) T · x k EV (j)
Then, the distance calculation unit 218 outputs the distances Out ₁ (k), Out ₂ (k),..., Out _m (k) sequentially calculated in this way to the classification unit 220.

  Note that the distance to the eigenspace of the area where the partial areas are set to overlap each other may be calculated as the sum of the distances for the partial areas set to overlap.

  FIG. 16 is a diagram for explaining a method for calculating the distance to the eigenspace of an area set with overlapping partial areas. FIG. 16A shows a case where two partial regions BLKa and BLKb overlap, and FIG. 16B shows a case where four partial regions BLKa, BLKb, BLKc, and BLKd overlap.

  Referring to FIG. 16 (a), the distance to the eigenspace for the overlapping portion of partial area BLKa and partial area BLKb is the sum of the distance to eigenspace of partial area BLKa and the distance to eigenspace of partial area BLKb. Or an average value thereof.

  Referring to FIG. 16B, the distance to the eigenspace for the overlapping portion of the partial areas BLKa, BLKb, BLKc, and BLKd is the sum of the distances to the eigenspace of the partial areas BLKa, BLKb, BLKc, and BLKd. Or an average value thereof.

Referring to FIG. 7 again, the classification unit 220 first determines the distances Out ₁ (k), Out ₂ (k),..., Out _m (k) in each evaluation region calculated by the distance calculation unit 218. Are sequentially generated as response vector RV (k) = [Out ₁ (k), Out ₂ (k),..., Out _m (k)]. Then, the classification unit 220 maps the response vectors RV (k) that are sequentially generated into the m-dimensional feature space. Further, the classification unit 220 determines a discriminant function for distinguishing a normal value from an outlier value for a set of k response vectors RV (k) mapped on the m-dimensional feature space. (See FIG. 6 (a)).

  As described above, the classification unit 220 according to the present embodiment determines the discriminant function using the outlier detection method (typically 1-class SVM). Then, the classification unit 220 stores the determined discriminant function in the learning result storage unit 240.

  As described above, the discriminant function determined by the classification unit 220 is stored in the learning result storage unit 240 as a learning result. Thereby, the learning process is completed.

  Next, the image buffer unit 222, the extraction unit 224, the vector generation unit 226, the distance calculation unit 228, the determination unit 230, and the defect detection unit 232 perform a defect detection process on the inspection image.

  The image buffer unit 222 is formed in a predetermined area of the memory 106 (FIG. 1), and temporarily stores an inspection image input from the outside of the apparatus. In the present embodiment, the configuration in which the inspection image is input from the imaging unit 8 to the image buffer unit 222 is illustrated. However, the inspection image captured by a device other than the imaging unit 8 is transmitted via a recording medium or a network. Thus, it may be stored in advance in the fixed disk 107 (FIG. 2).

The extraction unit 224 sequentially sets evaluation regions in the inspection image IMG temporarily stored in the image buffer unit 222 and associates with each evaluation region with partial regions BLK ⁽¹⁾ , 2 having different vector sizes. BLK ⁽²⁾ ,..., BLK ^(m) are sequentially extracted. The portion area ^{BLK (1), BLK (2} ), ···, BLK size of ^(m) is the extraction unit 214 by the learning image portion area ^BLK (1) to be sequentially extracted from LNIMG, ^{BLK (2)} ,..., Matches the size of BLK ^(m) . In addition, the evaluation area set by the extraction unit 224 in the inspection image IMG corresponds to one of the evaluation areas set by the extraction unit 214 in the learning image LNIMG.

  Further, the extraction unit 224 outputs to the determination unit 230 the position (label information) of the evaluation region that is sequentially extracted from the inspection image IMG.

The vector generation unit 226 uses the inspection image vector x _k ⁽ from the image information of the partial region sets (partial regions BLK ⁽¹⁾ , BLK ⁽²⁾ ,..., BLK ^(m) ) sequentially extracted by the extraction unit 224. ¹⁾ , x _k ⁽²⁾ ,..., X _k ^(m) are generated. Note that k is a label number indicating which of the partial region set (evaluation region) set in the learning image LNIMG corresponds to the evaluation region set in the inspection image vector. Note that the inspection image vector generation processing in vector generation unit 226 is similar to the learning image vector generation processing in vector generation unit 206 described above, and therefore detailed description thereof will not be repeated. Then, the vector generation unit 226 outputs the generated inspection image vector to the distance calculation unit 228.

The distance calculation unit 228, for each of the inspection image vectors x _k ⁽¹⁾ , x _k ⁽²⁾ ,..., X _k ^(m) output from the vector generation unit 226, corresponds to the eigenspace E of the corresponding vector size. ^{(1), E (2)} , ···, E (m) distance _{Out '1 (k), Out} ' between the ₂ (k), ···, and sequentially calculates the Out _'m (k) . More specifically, the distance calculation unit 228 includes the inspection image vector x _k ^(j) output from the vector generation unit 226 and the corresponding vector size eigenspace E ^(D) stored in the learning result storage unit 240. On the basis of the,
| Out ′ _j (k) | ² = x _k ^{(j) T} · x _k ^(j) −P _k ^{(j) T} · P _k ^(j)
(However, _Pk ^(j) = E ^(D) _Txk ^(j) )
The distance Out ′ _j (k) between the inspection image vector x _k ^(j) and the corresponding eigenspace E ^(D) is calculated according to the following relational expression. In this calculation, the distance calculation unit 228 reads the corresponding eigenspace E ^(D) from the learning result storage unit 240.

As described above, the distance calculation unit 228 uses the eigenspace E ^(D) stored in the learning result storage unit 240 to use the distance Out ′ _j between the eigenspace E ^(D) and the inspection image vector x _k ^(j). since calculating the (k), after the eigenspace E based on the learning image LNIMG ^(D) is calculated once again calculates the eigenspace E ^(D) for each test image IMG is input to the subsequent do not have to. Therefore, the processing time can be shortened as compared with the conventional case. Moreover, since the eigen space E ^{(D) serving as} a reference does not include information on a defective portion included in the image, the accuracy of defect detection can be increased.

The determination unit 230 collects the distances Out ′ ₁ (k), Out ′ ₂ (k),..., Out ′ _m (k) in the target evaluation area calculated by the distance calculation unit 228 to collect the response vector RV ′. (K) = [Out ′ ₁ (k), Out ′ ₂ (k),..., Out ′ _m (k)] is generated. Then, the determination unit 230 maps the generated response vector RV ′ (k) to the m-dimensional feature space, reads a discriminant function calculated in advance from the learning result storage unit 240, and the mapped response vector RV ′ It is determined in which region the normal value or the outlier exists (see FIG. 6B). If the response vector RV ′ is present in the outlier region, the determination unit 230 determines that a defect exists in the target evaluation region. On the other hand, when the response vector RV ′ is present in the normal value region, the determination unit 230 determines that there is no defect in the target evaluation region. Further, when determining that a defect exists, the determination unit 230 notifies the defect detection unit 232 of information for specifying the defect position.

  The defect detection unit 232 refers to the inspection image IMG stored in the image buffer unit 222 based on the defect position specifying information from the determination unit 230, and detailed information about the defect such as the type and size of the defect. Is detected. Specifically, the defect detection unit 232 performs edge detection and the like on the inspection image IMG. And the defect detection part 232 outputs the detailed information etc. of the detected defect as a defect detection result.

  With the functional configuration as described above, the defect detection device according to the present embodiment detects a defect in inspection image IMG based on learning image LNIMG that does not include a defect.

(Processing procedure)
FIG. 17 is a flowchart showing a processing procedure for detecting a defect that may occur in the target work 2 in the defect detection system 1 according to the first embodiment of the present invention.

  Referring to FIGS. 1 and 17, the user places reference work 4 that is known to have no defect in advance in the imaging range of imaging unit 8 on transport mechanism 6 (step S2). Then, the user gives a predetermined command to the defect detection device to acquire a learning image obtained by photographing the reference workpiece 4 (step S4). At this time, the illumination unit 10 irradiates a predetermined amount of light toward the reference workpiece. As described above, a learning image captured in advance by an apparatus other than the imaging unit 8 may be given to the defect detection apparatus.

  After acquiring the learning image, the defect detection apparatus executes a learning process (step S6) and stores the learning result (discriminant function or eigenspace) (step S8). The learning process will be described later.

  When the learning process and the storage of the learning result are finished, the user operates the transport mechanism 6 and the like so that the target work 2 is sequentially transported to the photographing range of the imaging unit 8 on the transport mechanism 6 (step S10). When the target workpiece 2 is transported to the imaging range of the imaging unit 8, the defect detection apparatus acquires an inspection image by imaging the target workpiece 2 by the imaging unit 8 (step S12). The defect detection device detects the arrival of the target workpiece 2 by a position sensor (not shown) arranged on the transport mechanism 6 or the like.

  When the inspection image is acquired, the defect detection apparatus executes defect detection processing (step S14) and outputs the defect detection result (step S16). The defect detection result is stored in the defect detection apparatus or output to the monitor 102 or the like.

  Then, the defect detection device determines whether or not an end instruction has been given (step S18). This end instruction is given by the user or the like, or in conjunction with the detection of the absence of the target workpiece 2 as the defect detection target. If an end instruction has not been given (NO in step S18), the defect detection device executes the processes in and after step S12 again. If an end instruction is given (YES in step S18), the process ends.

  FIG. 18 is a flowchart showing a processing procedure of learning processing according to the first embodiment of the present invention. FIG. 19 is a flowchart showing a processing procedure for defect detection processing according to the first embodiment of the present invention. The flowcharts shown in FIGS. 18 and 19 are executed by realizing the functions shown in FIG. 7 by the CPU 105 reading a program stored in advance in the fixed disk 107 or the like into the memory 106 and executing it. .

(Learning process)
Referring to FIGS. 7 and 18, CPU 105 functioning as extraction unit 204 accepts the setting of the minimum / maximum vector size of the partial area and the number of partial areas included in the partial area set from the user (step S <b> 100). Then, the CPU 105 functioning as the extraction unit 204 determines a partial region set according to the setting information from the user in step S100 (step S102).

The CPU 105 functioning as the extracting unit 204 extracts a partial region from the learning image LNIMG for one vector size included in the partial region set (step S104). Then, the CPU 105 functioning as the vector generation unit 206 generates a learning image vector x _i ^{LN (j)} from the extracted partial region image information (step S106). Further, the CPU 105 determines whether or not the partial region BLK ^LN to be extracted from the learning image LNIMG still remains (step S108). When partial region BLK ^LN to be extracted from learning image LNIMG remains (YES in step S108), CPU 105 repeats the processing in step S104 and subsequent steps.

On the other hand, when the partial region BLK to be extracted from the learning image LNIMG does not remain (NO in step S108), the CPU 105 functioning as the eigenvector calculation unit 208 determines the learning image vector x _i ^{LN (j).} An eigenvector V ^(j) is calculated from a set of vectors consisting of (step S110). Then, the CPU 105 functioning as the eigenspace determining unit 210 determines the eigenspace E ^(j) in the vector size using a predetermined number of eigenvectors in descending order from the calculated eigenvector V ^(j). (Step S112). Then, the CPU 105 stores the determined eigenspace E ^(j) in the learning result storage unit 240 (step S114).

Further, the CPU 105 determines whether or not the determination of the eigenspace E ^(j) is completed for all the partial areas of the vector size included in the partial area set (step S116). When the calculation of the eigenspace E ^(j) is not completed for all the partial areas of the vector size included in the partial area set (NO in step S116), the CPU 105 functioning as the extracting unit 204 For the remaining vector sizes included in the set, a partial region is extracted from the learning image LNIMG (step S118). And the process after step S106 is performed again.

When extraction has been completed for all the partial regions of the vector size included in the partial region set (YES in step S116), CPU 105 functioning as extraction unit 214 sets an evaluation region in learning image LNIMG ( Then, partial areas BLK ^{EV (1)} , BLK ^{EV (2)} ,..., BLK ^{EV (m)} having a plurality of different vector sizes are sequentially extracted from the evaluation area (step S122). Then, the CPU 105 functioning as the vector generation unit 216 determines the evaluation area image vector x _k ^EV from the image information of the extracted partial areas BLK ^{EV (1)} , BLK ^{EV (2)} ,..., BLK ^{EV (m). (1)} , x _k ^{EV (2)} ,..., X _k ^{EV (m)} are generated (step S124).

Subsequently, the CPU 105 functioning as the distance calculation unit 218 has a vector size corresponding to each of the evaluation area image vectors x _k ^{EV (1)} , x _k ^{EV (2)} ,..., X _k ^{EV (m)} . eigenspace ^{E (1), E (1} ), ···, E (m) distance _Out 1 _(k) between, Out 2 (k), ··· , calculates the _Out m (k) ( Step S126). Then, the CPU 105 functioning as the distance calculation unit 218 collects the calculated distances Out ₁ (k), Out ₂ (k),..., Out _m (k) and sets a response vector RV (k) = [Out ₁ (K), Out ₂ (k),..., Out _m (k)] are generated (step S128).

  Further, the CPU 105 determines whether or not there is still a region where an evaluation region should be set in the learning image LNIMG (step S130). If there is an area for which an evaluation area is to be set in learning image LNIMG (YES in step S130), CPU 105 repeats the processes in and after step S120.

  On the other hand, when there is no region for which the evaluation region should be set in the learning image LNIMG (NO in step S130), the CPU 105 functioning as the classification unit 220 generates the response vector RV ( For the set k), a discriminant function for discriminating between a normal value and an outlier value is determined (step S132), and the determined discriminant function is stored in the learning result storage unit 240 (step S134). Then, the process returns to step S8.

(Defect inspection processing procedure)
Referring to FIGS. 7 and 19, CPU 105 functioning as extraction unit 224 sets an evaluation area in inspection image IMG (step S200), and partial areas BLK ⁽¹ having different vector sizes for the evaluation area) ^{), BLK (2), ···} , sequentially extracted ^{BLK (m)} (step S202). Then, the CPU 105 functioning as the vector generation unit 226 determines the inspection image vector x _k ⁽¹⁾ , x from the image information of the extracted partial areas BLK ⁽¹⁾ , BLK ⁽²⁾ ,..., BLK ^(m). _k ⁽²⁾ ,..., x _k ^(m) are generated (step S204).

Subsequently, the CPU 105 functioning as the distance calculation unit 228 receives eigenspaces corresponding to the inspection image vectors x _k ⁽¹⁾ , x _k ⁽²⁾ ,..., X _k ^(m) from the learning result storage unit 240, respectively. E ^(D) is read (step S206). The CPU 105 functioning as the distance calculation unit 228 and the inspection image vectors x _k ⁽¹⁾ , x _k ⁽²⁾ ,..., X _k ^(m) and the read inner product values P _k ^{(j) T} Based on P _k ^(j) , the distances Out ′ ₁ (k), Out ′ ₂ (k),..., Out ′ _m (k) with the corresponding eigenspace E ^(j) of the vector size Calculate (step S208).

Subsequently, the CPU 105 functioning as the determination unit 230 collects the calculated distances Out ′ ₁ (k), Out ′ ₂ (k),..., Out ′ _m (k) to collect the response vector RV ′ (k). = [Out ' ₁ (k), Out' ₂ (k), ..., Out ' _m (k)] is generated (step S210). Then, the CPU 105 functioning as the determination unit 230 reads the discriminant function from the learning result storage unit 240 (step S212), and the calculated response vector RV ′ (k) based on the read discriminant function has a normal value and an outlier value. It is determined in which area it exists (step S214). Then, the CPU 105 functioning as the determination unit 230 maps the defect presence / absence state in the inspection image IMG in association with the label information of the target evaluation area (step S216). That is, the CPU 105 identifies the position of the defect in the inspection image IMG.

  Thereafter, the CPU 105 determines whether or not there is still an area for setting the evaluation area in the inspection image IMG (step S218). If there remains an area for which an evaluation area should be set in inspection image IMG (YES in step S218), CPU 105 repeats the processes in and after step S200.

  On the other hand, when there is no region where the evaluation region should be set in the inspection image IMG (NO in step S218), the CPU 105 functioning as the determination unit 230 determines the position of the defect with respect to the inspection image IMG. Edge detection or the like is performed on the part identified as “” and detailed information about the defect such as the type and size of the defect is detected (step S220). Then, the process returns to step S16.

(Operational effects of the present embodiment)
According to the first embodiment of the present invention, a plurality of partial areas having different sizes are set for the evaluation area of the inspection image, and each partial area is associated with an image vector generated from the image information. Calculate the distance to the eigenspace. Further, a response vector having a distance for each size as a component for each partial region is generated, and this response vector is evaluated with reference to a discriminant function determined in advance by a learning process. Judgment is made. In this way, since the evaluation area is determined based on the evaluation results in the units of a plurality of partial areas having different sizes, the presence / absence of a defect is determined. The influence can be suppressed and the detection accuracy of defects can be further increased.

  Further, according to the first embodiment of the present invention, the eigenspace serving as a reference for calculating the distance to the image vector is determined based on a learning image not including a defect in advance. For this reason, the eigenspace in which the influence of the defective portion is eliminated can be determined, so that the defect detection accuracy for the inspection image can be further increased.

  According to the first embodiment of the present invention, the distance between the image vector and the eigenspace during the defect detection process is calculated using the inner product value of the projection vectors calculated during the learning process. As a result, the amount of calculation associated with the distance calculation in the defect detection process can be reduced, so that a faster defect detection process can be realized.

[Embodiment 2]
According to the defect detection process according to the above-described first embodiment, it is possible to suppress the influence due to the variation in the size of the defect that may occur in the inspection image, and to further improve the defect detection accuracy. However, an error tends to occur when the inspection image has a periodic texture.

  Specifically, even if the position of the sample cut out from the inspection image is exactly the same as that of the learning image, the periodic texture appearing there is not always exactly the same. Therefore, a spatial phase difference occurs in the periodic gray pattern appearing in the cut out evaluation area (partial area set SET), and as a result, the distance to the eigenspace can be increased even if there is no defect. Therefore, in the present embodiment, a configuration and processing capable of accurately detecting a defect even for an inspection image including a periodic shading pattern will be exemplified.

  More specifically, the reference position in each inspection image IMG is searched in order to appropriately determine the start position of the arrangement pattern of the evaluation area with respect to the inspection image IMG. In particular, when the learning image LNIMG and the inspection image IMG have a periodic shading pattern, the defect detection accuracy is greatly influenced by which position the arrangement pattern is arranged. Note that an image having a periodic shading pattern means an image in which the same pattern (texture) appears continuously at a predetermined period (constant period). Typically, it means an image including a geometric pattern or the like as shown in FIGS.

Here, the position (spatial phase) of the periodic shading pattern generated in the learning image LNIMG and the inspection image IMG is not always the same, and the imaging unit 8 also depends on the transport state of the production line and the like. Deviation occurs in the periodic shading pattern appearing in the captured image. In such a case, if the partial areas BLK ^LN and BLK corresponding to the evaluation area existing at a certain position of the arrangement pattern are viewed, the feature amounts appearing in the image vectors respectively generated from the learning image LNIMG and the inspection image IMG Are different from each other. Therefore, before performing the defect detection with respect to the inspection image IMG, along a predetermined search direction of the learning image LNIMG, continue to sequentially set the partial area BLK ^S having a predetermined size, the partial area BLK ^S at each position The distance to the eigenspace of the image vector generated from the image information is sequentially calculated. And based on the change about the search direction of the distance with respect to this eigenspace, the corresponding reference position in the inspection image IMG is determined. Then, according to the arrangement pattern based on the determined reference position, an evaluation region (that is, a corresponding BLK ⁽¹⁾ , BLK ⁽²⁾ ,..., BLK ^(m) ) is set for the inspection image IMG, and a defect is detected. The presence or absence of is detected.

(Control structure)
FIG. 20 is a functional block diagram showing a control structure of the defect detection apparatus according to the second embodiment of the present invention.

  Referring to FIG. 20, the defect detection device according to the present embodiment is the same as the defect detection device according to the first embodiment shown in FIG. 7, but includes region setting unit 244, vector generation unit 246, distance calculation unit 248, reference This corresponds to the addition of the position determination unit 250 and the search area reception unit 256. Since other parts have been described with reference to FIG. 7, detailed description will not be repeated.

First, let us consider a case where the target work 2 has a periodic texture and a periodic shading pattern appears in the learning image LNIMG and the inspection image IMG. In such a case, the evaluation area (that is, the corresponding BLK ^{EV (1)} , BLK ^{EV (2)} ,..., BLK ^{EV (m)} or BLK ⁽¹⁾ , BLK set in the inspection image IMG. The arrangement pattern of ⁽²⁾ ,..., BLK ^(m) ) needs to be appropriately set according to the periodic shading pattern appearing in the inspection image IMG.

  Hereinafter, the reason why an error occurs when the defect detection method according to the present embodiment is applied to the learning image LNIMG and the inspection image IMG in which such a periodic shading pattern appears will be described.

  FIG. 21 is a diagram illustrating an example of the target workpiece 2 having a periodic texture. For example, when the target workpiece 2 shown in FIG. 21A and the target workpiece 2 shown in FIG. 21B are compared, the relative start positions of the periodic shading patterns in each target workpiece 2 are different. I understand. Specifically, it can be seen that the shading pattern is entirely displaced in the vertical direction of the paper. As described above, even if the workpieces are manufactured on the same manufacturing line, the texture appearing on the surface does not always appear at the same position.

  FIG. 22 is a diagram illustrating an example of the partial region BLK that is extracted when the positions of the periodic shading patterns that appear in the inspection image IMG vary. FIG. 22A shows an example of the entire inspection image IMG, and FIGS. 22B and 22C show examples of the partial areas BLK that are extracted.

In the inspection image IMG having an overall periodic shading pattern as shown in FIG. 22A, if the position of the periodic shading pattern varies, the evaluation is performed based on the corresponding coordinate position in the inspection image IMG. Even if the region is set, the features included in the extracted partial regions BLK ⁽¹⁾ , BLK ⁽²⁾ ,..., BLK ^(m) are different. For example, even if the evaluation region is set with reference to the upper left vertex of the inspection image IMG, if the position of the periodic shading pattern of the inspection image IMG varies, the states shown in FIGS. 22 (b) and 22 (c) are obtained. As shown, the features between the partial areas BLK ^{EV (m)} and BLK ^(m) of the same size corresponding to the evaluation area are different from each other.

  FIG. 23 is an example of a diagram visually showing the distance to the eigenspace for a representative partial region BLK corresponding to the evaluation region set in the inspection image IMG. FIG. 23A shows a case where the reference position in the inspection image IMG is shifted, and FIG. 23B shows a case where the reference position in the inspection image IMG corresponds to the reference position of the learning image LNIMG. In FIG. 23, the higher the density (black), the smaller the distance to the eigenspace, and the brighter (white), the greater the distance to the eigenspace.

Referring to FIG. 23A, when the reference position of the inspection image IMG is deviated, it can be seen that a cyclic error occurs due to this deviation. On the other hand, as shown in FIG. 23B, when the reference position of the inspection image IMG corresponds to the reference position of the learning image LNIMG, each part except for the partial area BLK including the defect DEF. It can be seen that the distance to the eigenspace for the region BLK is small and uniform.

  Therefore, the defect image device according to the present embodiment searches the inspection image IMG for a reference position corresponding to the evaluation region set in the learning image LNIMG, and performs defect detection on the basis of the searched reference position. Set the target evaluation area. That is, instead of sharing the relative position in the learning image LNIMG and the inspection image IMG, the reference point in the inspection image IMG is associated with the periodic shade pattern portion in the reference point of the learning image LNIMG. Is to search. Thus, accurate defect detection can be realized even if there is a deviation between the periodic shading pattern appearing in the learning image LNIMG and the shading pattern appearing in the inspection image IMG.

  Referring again to FIG. 20, such a reference position search process includes an area setting unit 244, a vector generation unit 246, a distance calculation unit 248, a reference position determination unit 250, and a search region reception unit 256. It is realized by.

FIG. 24 is a diagram for explaining reference position search processing in the inspection image IMG. FIG. 24A shows an evaluation region (represented as a partial region BLK ^{EV (1)} having a minimum size for simplification) set in the learning image LNIMG, and FIG. 24B shows an inspection image IMG. A partial region BLK ^S for searching for a reference position at is shown.

Referring to FIG. 24A, for example, in the learning image LNIMG, a plurality of evaluation regions are regularly set according to a predetermined arrangement pattern based on a predetermined reference position (for example, the upper left vertex). The eigenspace is calculated based on all or part of the partial areas BLK ⁽¹⁾ , BLK ⁽²⁾ ,..., BLK ^(m) corresponding to the plurality of set evaluation areas.

On the other hand, referring to FIG. 24 (b), the along the partial region BLK ^S set for the test image IMG in a predetermined search direction, as well as progressive scan, image vectors corresponding to the respective partial regions BLK ^S ( Hereinafter, the distance between the “search image vector” and the eigenspace is sequentially calculated. For convenience, the horizontal direction of the test image IMG is referred to as the X direction, and the vertical direction of the paper is also referred to as the Y direction. Furthermore, a reference position in the inspection image IMG is determined based on a change in distance with respect to the eigenspace with respect to the search direction.

In order to search for the reference position in the inspection image IMG, it is necessary to sequentially set the partial areas BLK ^S at intervals shorter than the length of the partial area BLK ^S along the search direction, and the pixels (pixels) constituting the inspection image IMG It is more preferable to set the partial areas BLK ^S at unit intervals. In this manner, the reference position in the inspection image IMG can be determined more accurately by moving the partial region BLK ^S in units of one pixel.

Further, the reference position in the inspection image IMG can be searched by scanning the partial region BLK ^S over the entire inspection image IMG. However, there may be a case where excessive processing time is required. Therefore, it is preferable to scan the partial area BLK ^S within the predetermined search area SRC. Incidentally, X-direction alone or the Y direction alone, to be subjected to scanning the partial area BLK ^S by any of the methods of the X · Y-direction simultaneously.

FIG. 25 is a diagram for explaining a method of determining a search area SRC for searching for a reference position in the inspection image IMG. FIG. 25A shows a case where the period of the shade pattern is shorter than the size of the partial region BLK ^S , and FIG. 25B shows a case where the cycle of the shade pattern is longer than the size of the partial region BLK ^S. FIG. 26 is a diagram for explaining the period of the shading pattern.

Referring to FIG. 25A, when the period of the light and shade pattern is shorter than the size of the partial region BLK ^S ,
In other words, if there is a light and shade pattern of one period or more in one partial area BLK ^S , the search area SRC may be at most 2 × 2 in size of the partial area BLK ^S. This is because the reference position can be determined by scanning the partial region BLK ^S over a distance corresponding to one period of the light and shade pattern.

On the other hand, referring to FIG. 25 (b), the if the period of the shading pattern is longer than the size of the partial region BLK ^S, i.e. one partial area BLK ^S does not include a shading pattern of one period is the gray The search area SRC needs to be determined according to the pattern period.

  With reference to FIG. 26, in this specification, the period of a light and shade pattern means the distance between the same adjacent patterns (textures). Basically, the period of the shading pattern can be defined for each of the X direction and the Y direction. It is preferable to set a search region SRC having a length of one cycle or more in the X direction and one cycle or more in the Y direction.

  Such a search area SRC may be determined in advance after the user confirms a periodic shading pattern appearing in the learning image LNIMG when calculating the eigenspace based on the learning image LNIMG not including a defect. Good. In such a case, the search area receiving unit 256 (FIG. 20) receives the size of the search area SRC from the user and stores the value in the learning result storage unit 240. Alternatively, the search area SRC may be calculated after performing a known pattern extraction process or the like on the learning image LNIMG and calculating the period of such a shading pattern. In such a case, the search area receiving unit 256 (FIG. 20) may guide the user of areas that should be the search area SRC.

Referring to FIG. 20 again, region setting unit 244 sequentially sets partial regions BLK ^S according to a predetermined search direction of inspection image IMG for inspection image IMG temporarily stored in image buffer unit 222. To do. The region setting unit 244 sequentially sets the partial regions BLK ^S according to the search region SRC stored in the learning result storage unit 240 as shown in FIG.

Then, the region setting unit 244 outputs the image information of the inspection image IMG corresponding to each set partial region BLK ^S to the vector generation unit 246. In other words, the region setting unit 244 sequentially extracts search samples (local regions) having a predetermined size from the inspection image temporarily stored in the image buffer unit 222.

The vector generation unit 246 sequentially generates the search image vector x ^S _i from the image information of the partial region BLK ^S of the inspection image IMG sequentially extracted by the region setting unit 244. The search image vector x ^S _i generation processing in the vector generation unit 246 includes the learning image vector x ^LN _i generation processing in the vector generation units 206 and 216 and the inspection image vector x _i generation processing in the vector generation unit 226. Since they are similar, detailed description will not be repeated.

Distance calculating unit 248, a search image vector ^x _{S i} output from the vector generator 246, based on the eigenspace ^{E (D)} which is previously stored in the learning result storage unit 240, an eigenspace ^{E (D)} A distance Δx ^S _i with the search image vector x ^S _i is calculated. The distance calculation processing distance [Delta] x ^S _i of the eigenspace ^E and ^(D) and the search image vector ^x _{S i} is the distance calculation unit 218 described above, the inspection image vector _{x i} with eigenspace ^{E (D)} because the [Delta] x _i is the same as the calculation processing, the detailed description thereof will not be repeated.

The reference position determination unit 250, based on a change in the search direction of the distance [Delta] x ^S _i that are sequentially calculated in the distance calculation unit 248, the inspection image IMG, determines the reference position corresponding to the reference position of the learning images LNIMG.

FIG. 27 is a diagram illustrating an example of the distance Δx ^S _i sequentially calculated by the distance calculation unit 248. FIG. 27A shows an example of a change when the inspection image IMG is searched in the X direction, and FIG. 27B shows an example of a change when the inspection image IMG is searched in the Y direction.

Referring to FIGS. 27 (a) and 27 (b), the reference position determination unit 250, with respect to the distance [Delta] x ^S _i which varies according to the position of the partial area BLK ^S set in the inspection image IMG, minimum A point position is extracted as a reference position candidate position. At the time of extracting a candidate position of the reference position, it is preferred that the distance [Delta] x ^S _i is on condition that is below a predetermined threshold value Th. The threshold Th can be determined in accordance with the overall average value of the distance [Delta] x ^S _i.

As a method of searching the reference position, independently, fix one of the X or Y direction of the test image IMG, and a method of sequentially setting a partial area BLK ^S along the other axis, the X and Y directions It is possible to adopt a method in which the partial regions BLK ^S are sequentially set by changing to.

FIG. 28 is a diagram illustrating an example of candidate points extracted for the inspection image IMG.
Referring to FIG. 28, according to the method as described above, candidate positions in the X direction and the Y direction can be acquired. Therefore, based on these candidate positions, a plurality of candidate points with respect to the inspection image IMG can be obtained. Can be calculated. Here, a candidate point is a coordinate uniquely determined by a candidate position in the X direction and a candidate position in the Y direction.

The reference position determination unit 250 determines one candidate point as a reference position in the inspection image IMG among the plurality of candidate points. Typically, the reference position determination unit 250 determines the candidate point having the smallest distance Dis from the origin corresponding to the reference position of the learning image LNIMG as the reference position. That is, if the coordinates of the candidate point are (x, y), the candidate point with the smallest distance from the origin Dis = (x ² + y ² ) ^1/2 is determined as the reference position.

  This is because, in general, since a periodic shading pattern shift appearing in the target work 2 appears in units of cycles, by selecting a position where the shift with respect to the periodic shading pattern appearing in the learning image LNIMG is minimized, This is because an appropriate reference position can be determined.

  Thus, the reference position determined by the reference position determination unit 250 is output to the extraction unit 224, and defect detection processing for the inspection image IMG is performed.

FIG. 29 is a diagram for explaining an evaluation region (expressed as a partial region BLK ⁽¹⁾ of the minimum size for simplification) set for the inspection image IMG by the extraction unit 224.

  Referring to FIG. 29, based on the reference position determined by reference position determination unit 250, extraction unit 224 sequentially sets evaluation regions for inspection image IMG according to an arrangement pattern with reference to the reference position. .

  Comparing FIG. 24A and FIG. 29, it can be seen that the arrangement pattern for setting the evaluation region is arranged so as to protrude from the entire frame of the inspection image IMG. However, the relative positional relationship of this arrangement pattern with respect to the shading pattern appearing in the inspection image IMG substantially coincides with the relative positional relationship of the arrangement pattern for setting the evaluation region with respect to the shading pattern appearing in the learning image LNIMG. I understand that. That is, for example, it is understood that the shading pattern surrounded by the evaluation region adjacent to the reference position of the learning image LNIMG is the same as the shading pattern surrounded by the evaluation region adjacent to the reference position of the inspection image IMG. In this way, by accurately matching the shade patterns appearing in the partial areas corresponding to the corresponding pair of evaluation areas between the learning image LNIMG and the inspection image IMG, it is possible to perform highly accurate error detection. .

(Variation 1)
In the above description, (preferably, 1 pixel) for a predetermined search area SRC, a predetermined interval partial area BLK ^S configuration has been described to sequentially set at, sets a partial region BLK ^S Interval May be divided into a plurality of stages. That is, after the reference position is roughly searched at a relatively large interval, the reference position may be searched more finely at a relatively small interval.

Specifically, the inspection image IMG, a relatively large interval partial area BLK ^S (e.g., 8 pixels) are sequentially set in, looking for candidate points by the same method as described above. Then, after a limited search area around the candidate point, a relatively small distance with respect to the search image (e.g., 1 pixel) are sequentially set the partial area BLK ^S in, the same manner as described above To search for candidate points more precisely.

According to this method, it is possible to streamline the distance calculation with respect to setting and eigenspace subregion BLK ^S, the search process of the reference position can be speeded.

  Note that when the method as described above is employed, the search area SRC may not be set in advance.

(Variation 2)
In the above description, the configuration in which the partial areas BLK ^S are sequentially set for the periodic shading pattern in the inspection image IMG has been described. However, the partial areas BLK ^S may be set so as to protrude from the inspection image IMG. .

FIG. 30 is a diagram for explaining processing when a part of the set partial region BLK ^S protrudes from the inspection image IMG.

  Referring to FIG. 30, interpolation processing may be performed on a portion that protrudes from inspection image IMG based on image information of an adjacent portion. Such interpolation processing is described in the literature (Toshiyuki Amano, Yukio Sato, “Image interpolation by high-dimensional nonlinear projection using kBPLP method”, IEICE Transactions D-II Vol. J86-D-II No. 4). pp. 525-534, April 2003).

By performing such an interpolation process, the reference position can be searched even when there is a partial region BLK ^S that partially protrudes from the inspection image IMG.

(Comparative example)
Conventionally, as a method of detecting a defect from an inspection image, a method of performing processing based on a spatial frequency represented by a luminance value of a pixel constituting the image is known. However, with such a method, it is difficult to detect a defect from an inspection image having a periodic shading pattern that is a subject of the present invention.

FIG. 31 is a diagram illustrating an example of an inspection image having a periodic shading pattern as a target.
FIG. 32 is a diagram showing a result when the conventional method and the method according to the present embodiment are applied to the inspection image shown in FIG. 31.

  Referring to FIG. 31, consider a case where the conventional method and the method according to the present embodiment are applied to an inspection image having a periodic gray pattern along the horizontal direction of the paper surface and the vertical direction of the paper surface.

  FIG. 32A shows a graph in which the R component value, the G component value, and the B component value in the sample region of FIG. 31 are plotted in association with the positions. As described above, when the evaluation is performed with the spatial frequency composed of the components of each pixel, even if the inspection image has a periodic shading pattern, the difference value or the positive / negative value in each component value varies. Therefore, it is very difficult to determine a color component to be used as a reference and to determine the threshold value.

  On the other hand, as shown in FIG. 32 (b), the eigenspace is calculated in advance based on the learning image LNIMG, and then evaluated using the distance from the eigenspace in each partial region. It is possible to perform dimension detection and to perform defect detection after removing noise. For this reason, the method according to the present invention can be applied universally without being affected by the color components of the periodic shading pattern appearing in the inspection image.

(Processing procedure)
FIG. 33 is a flowchart showing a processing procedure for detecting a defect that may occur in the target work 2 in the defect detection system 1 according to the second embodiment of the present invention.

  The flowchart shown in FIG. 33 corresponds to the flowchart shown in FIG. 7 with steps S7 and S13 added. The same steps are given the same step numbers as in FIG. 7, and the processing of each step has been described above, and thus detailed description will not be repeated.

  Referring to FIGS. 20 and 33, after the execution of the learning process (step S6), the defect detection device accepts search area SRC from the user (step S7). The user operates the keyboard 103 and the mouse 104 while checking the captured learning image to set the search area SRC.

  After acquiring the inspection image by photographing the target workpiece 2 by the imaging unit 8 (step S12), the defect detection device first executes a reference position search process (step S13). In addition, the defect detection device executes a defect detection process (step S14). Then, the defect detection device outputs the defect detection result (step S16).

  FIG. 34 is a flowchart showing a procedure of the reference position search process according to the second embodiment of the present invention. The flowchart shown in FIG. 34 is executed by realizing the functions shown in FIG. 20 by the CPU 105 reading a program stored in advance in the fixed disk 107 or the like into the memory 106 and executing it.

Referring to FIGS. 20 and 34, CPU 105 functioning as region setting unit 244 reads search region SRC from learned value storage unit 212 (step S300). Then, the CPU 105 functioning as the area setting unit 244 sequentially sets the partial areas BLK ^S in the search area SRC according to a predetermined search direction for the inspection image IMG (step S302). Then, the CPU 105 functioning as the vector generation unit 246 generates a search image vector x ^S _i from the image information of each partial area BLK ^S (step S304). The CPU 105 functioning as the distance calculation unit 248 reads the eigenspace E ^(D) from the learning value storage unit 212 (step S306). Then, the CPU 105 calculates a distance Δx ^S _i from the eigenspace E ^(D) to the search image vector x ^S _i based on the retrieved image vector x ^S _i and the read eigenspace E ^(D) (step) S308). Further, the CPU 105 determines whether or not the partial area BLK ^S to be extracted remains in the search area SRC (step S310). When partial area BLK ^S to be extracted remains in search area SRC (YES in step S310), CPU 105 repeats the processes in and after step S302.

When there is no partial region BLK ^S to be extracted in the search region SRC (NO in step S310), the CPU 105 functioning as the reference position determination unit 250 minimizes each of the search directions of the distance Δx ^S _i . A position that becomes a point is extracted as a candidate position (step S312), and among the extracted minimum points, those that exceed a predetermined threshold Th are excluded (step S314). Further, the CPU 105 calculates candidate points in the inspection image IMG based on the candidate positions in the respective search directions (step S316), and among these candidate points, the distance from the origin corresponding to the reference position of the learning image LNIMG. The candidate point with the smallest is determined as the reference position (step S318).

(Operational effects of the present embodiment)
According to the second embodiment of the present invention, when an evaluation area is set in an inspection image, a position where the distance to the eigenspace has a minimum value is searched along a predetermined search direction. Based on the searched position, an arrangement pattern for setting the evaluation area is positioned. As a result, even when the learning image and the inspection image include a periodic shading pattern, the relative start positions of the arrangement pattern with respect to the periodic shading pattern can be matched with each other in the learning image and the inspection image. . Therefore, it is possible to improve the accuracy of defect detection even for an inspection image including a periodic shading pattern.

[Modification of Embodiment 2]
In the above-described second embodiment, the configuration for optimizing the arrangement pattern of the evaluation region is illustrated when the learning image LNIMG and the inspection image IMG have a periodic shading pattern. By the way, in an actual production line or the like, a process for manufacturing a plurality of products from one base material, for example, a manufacturing process for a liquid crystal panel, a semiconductor substrate, or the like is known in order to increase production efficiency. In such a process, since each product has the same size, the products before separation are periodically arranged on one substrate. In such a case, it is desirable to determine a partial region set corresponding to the size of one product and calculate an image vector and a response vector. Therefore, in this modification, a configuration that can be easily applied to such an application will be described below.

  FIG. 35 is a diagram showing an example of an image SMP obtained by imaging a target workpiece according to the modification of the second embodiment of the present invention. FIG. 35 shows a state where a plurality of circuit boards are formed on one base material as a typical example.

  The defect detection apparatus according to the present modification can automatically determine the size and position of the partial area for generating the image vector for the target workpiece as shown in FIG. More specifically, a pattern (template pattern) MDL representing at least a part of one circuit board included in the target work is registered in advance, and a plurality of positions matching the registered pattern MDL on the image SMP is registered. Is identified. The number of positions specified based on the pattern MDL matches the number of products included in the inspection object, and the position period thereof matches the arrangement period of products included in the inspection object. Therefore, as shown in FIG. 35 (b), by measuring the interval between positions that match the registered pattern, the size of the partial area BLK (X-direction lengths Lx and Y) to be set for each product is measured. The direction length Ly) can be determined. Hereinafter, a more detailed processing procedure will be described with reference to the flowchart shown in FIG.

  FIG. 36 is a flowchart showing a processing procedure for generating an image vector for a target workpiece having periodicity in the defect detection system according to the modification of the second embodiment of the present invention. The flowchart shown in FIG. 36 is realized by the CPU 105 reading a program stored in advance in the fixed disk 107 or the like into the memory 106 and executing it.

  Referring to FIG. 36, CPU 105 accepts a template pattern designated by the user (step S400). That is, the user registers a part of an image such as one product that periodically appears in the target work as a pattern MDL indicating a model image. This pattern MDL may be cut out from image data obtained by photographing a workpiece to be used as a reference, or image data such as a catalog may be externally input. Furthermore, the CPU 105 accepts a periodic size (values in the X direction and the Y direction) that appears periodically in the target work specified by the user (step S402). That is, the user inputs an approximate size such as one product that periodically appears in the target workpiece. This size is for shortening the time required for the search processing described later.

  When these inputs are completed, the CPU 105 executes a rough search process for searching for a position that matches the previously registered pattern MDL on the image data obtained by imaging the target workpiece (step S404). . In this rough search process, a feature amount between the pattern MDL and the target region, typically, a position having a relatively high correlation value is searched (normalized correlation matching process). As a result, the pattern MDL is positioned at a position where the correlation value is relatively high.

  Subsequently, the CPU 105 sets a search area SRC based on each positioned part (step S406). The search area SRC is set in a range corresponding to the period size input in step S402. Subsequently, the CPU 105 executes a fine search process for searching for a position that matches the pattern MDL in more detail in each search area SRC (step S408).

  When the position that matches the pattern MDL is specified, the CPU 105 determines the interval between the positions that match the specified pattern MDL (the X-direction length Lx and the Y-direction length Ly shown in FIG. 35A). Is measured (step S410). Subsequently, the CPU 105 sets the partial area BLK having the measured X-direction length Lx and Y-direction length Ly in association with the position that matches the pattern MDL specified in step S408 (step S412). Further, the CPU 105 extracts image information for each of the partial areas BLK set on the target work, and generates an image vector based on the extracted image information (step S414). A response vector and the like are calculated based on the generated image vector.

  The processing procedure as described above is incorporated as part of the procedure for generating the image vector and the response vector for each of the learning image LNIMG and the inspection image IMG.

(Operational effects of the modification of the present embodiment)
According to the modification of the second embodiment of the present invention, an image vector and a response vector can be generated more quickly and reliably for a structure having periodicity. Therefore, defect detection can be reliably performed even in a process of manufacturing a plurality of products from a single substrate.

[Embodiment 3]
In FIG. 1 described above, a configuration in which a relatively small object such as an electronic component or a plastic part is an inspection target is illustrated as the target workpiece 2, but the defect detection apparatus according to the present invention is a relatively large one. It can also be an inspection target.

  FIG. 37 is a schematic diagram showing a defect detection system 1 # including a defect detection device according to the third embodiment of the present invention.

  Referring to FIG. 37, defect detection system 1 # shows, as an example, a configuration in which a defect is an inspection target for a relatively large plate-shaped target workpiece 2 such as a liquid crystal panel. The defect detection system 1 # is movable in a direction orthogonal to the conveyance direction of the target workpiece 2, the conveyance frame 20, the conveyance table 22 that can slide on the conveyance frame 20 in the conveyance direction, the imaging unit 14, and the imaging unit 14. And an image pickup unit driving mechanism 24. The target workpiece 2 is arranged on the transport table 22 and moves along the transport direction. The imaging unit 14 shoots the target workpiece 2 moving on the conveyance platform 22 in the conveyance direction at predetermined intervals, and is moved by the imaging unit driving mechanism 24 so that the target workpiece 2 is also orthogonal to the conveyance direction. Shoot at every predetermined interval (however, different from the predetermined interval in the transport direction).

  An image photographed by the imaging unit 14 is transmitted to the control device 100 # that realizes the defect detection device via the interface 16. The imaging unit 14 may be integrated with an illumination unit. In this case, a power source for driving the illumination unit is supplied from the interface 16.

  Control device 100 # further includes a main body 101 #, a monitor 102 #, a keyboard 103 #, and the like. Main body 101 #, monitor 102 #, and keyboard 103 # implement the same functions as main body 101, monitor 102, and keyboard 103 of computer 100 shown in FIG.

  In particular, in defect detection system 1 # according to the present embodiment, since all of target workpiece 2 cannot be accommodated within the imaging range of imaging unit 14, the position of target workpiece 2 in the conveyance direction and the conveyance direction of imaging unit 14 It is necessary to specify the defect position by processing associated with the position in the orthogonal direction.

  In addition, it is not necessary to use a reference workpiece having the same size as the target workpiece 2, and it is sufficient to use a reference workpiece having a size that can be entirely accommodated in the imaging range of the imaging unit 14.

  Since other configurations and processes are the same as those of the defect detection apparatus according to the above-described embodiment, detailed description thereof will not be repeated.

[Other forms]
The program according to the present invention may be a program module that is provided as a part of a computer operating system (OS) and that calls necessary modules in a predetermined arrangement at a predetermined timing to execute processing. . In that case, the program itself does not include the module, and the process is executed in cooperation with the OS. A program that does not include such a module can also be included in the program according to the present invention.

  The program according to the present invention may be provided by being incorporated in a part of another program. Even in this case, the program itself does not include the module included in the other program, and the process is executed in cooperation with the other program. Such a program incorporated in another program can also be included in the program according to the present invention.

  The provided program product is installed in a program storage unit such as a hard disk and executed. Note that the program product includes the program itself and a recording medium in which the program is stored.

  Furthermore, part or all of the functions realized by the program according to the present invention may be configured by dedicated hardware.

  The embodiment disclosed this time should be considered as illustrative in all points and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

  1, 1 # Defect detection system, 2 target workpiece, 4 reference workpiece, 6 transport mechanism, 8 imaging unit, 10 illumination unit, 12 light source unit, 14 imaging unit, 16, 109 interface, 20 transport platform, 22 transport platform, 24 Imaging unit drive mechanism, 100 computer, 100 # control device, 101 computer main body, 101 # main body, 102, 102 # monitor, 103, 103 # keyboard, 104, 104 # mouse, 106 memory, 107 fixed disk, 111 FD drive device 113 CD-ROM drive, 202, 222 image buffer unit, 204, 214, 224 extraction unit, 206, 216, 226 vector generation unit, 208 eigenvector calculation unit, 210 eigenspace determination unit, 218, 228 distance calculation unit, 220 classification unit, 230 determination unit, 2 2 defect detecting unit, 240 learning result storage unit, 244 area setting unit, 246 vector generation unit, 248 distance calculating section, 250 reference position determination unit, 256 search area accepting unit.

Claims (10)

A defect detection device that detects a defect in an inspection image based on a learning result acquired from a learning image that does not include a defect,
Means for extracting image information of each of a plurality of partial areas including the first evaluation area set in the inspection image and having different sizes from each other;
Means for generating an inspection image vector for each size of the partial area from each of the image information extracted from the inspection image;
Means for calculating a distance between the corresponding eigenspace for each of the inspection image vectors for each size of the partial region;
Means for referring to a predetermined discriminant function and determining the presence or absence of a defect in the first evaluation region based on a first response vector having the distance for each size of the partial region as an element; A defect detection device. Means for extracting image information of a plurality of partial areas each including a second evaluation area sequentially set in the learning image and having the same size as each of the plurality of partial areas extracted from the inspection image;
Means for sequentially generating an evaluation area image vector for each size of the partial area from each of the image information extracted from the learning image;
Means for sequentially calculating a distance from a corresponding eigenspace for each of the evaluation area image vectors for each size of the partial area;
Means for determining the discriminant function defining a boundary between the inside of the set and the outside of the set based on a set of second response vectors each having the distance for each size of the partial region as an element; The defect detection apparatus of Claim 1 provided.   The defect detection apparatus according to claim 2, wherein the means for determining the discriminant function determines the discriminant function according to an outlier detection method. Means for sequentially extracting image information of a third partial area having the same size as one of the plurality of partial areas from the learning image;
Means for generating a learning image vector from the image information of the third partial region extracted sequentially;
Means for calculating an eigenvector from a set of vectors including the learning image vector;
4. The apparatus according to claim 1, further comprising means for determining the eigenspace in the size of the third partial region using a predetermined number of eigenvectors in descending order of eigenvalues among the calculated eigenvectors. The defect detection apparatus of any one of Claims. A defect detection device that detects a defect in an inspection image based on a learning result acquired from a learning image that does not include a defect,
Means for extracting image information of each of a plurality of partial areas including first evaluation areas sequentially set in the learning image and having different sizes;
Means for sequentially generating an evaluation area image vector for each size of the partial area from each of the image information extracted from the learning image;
Means for sequentially calculating a distance from a corresponding eigenspace for each of the evaluation area image vectors for each size of the partial area;
It means for each based on a set of response vector to the distance elements for each size of the partial region, to determine the determine specific functions that define the boundary between the set outer and the inner set,
Extract image information of a plurality of partial areas each including a second evaluation area sequentially set in the inspection image and having the same size as the plurality of partial areas extracted from the learning image. A defect detection apparatus comprising: means for determining the presence or absence of a defect in the second evaluation region based on the inspection image vector generated from the image information to be performed and the discriminant function.   The defect detection apparatus according to claim 5, wherein the means for determining the discriminant function determines the discriminant function according to an outlier detection method. A defect detection method for detecting defects in an inspection image based on a learning result acquired from a learning image using a computer including an arithmetic device and a storage device,
The arithmetic device including a first evaluation region set in the inspection image stored in the storage device and extracting image information of each of a plurality of partial regions having different sizes;
The arithmetic unit generates an inspection image vector for each size of the partial region from each of the image information extracted from the inspection image;
The arithmetic unit calculating a distance between the corresponding eigenspace for each inspection image vector for each size of the partial region;
The arithmetic device refers to a discriminant function stored in advance in the storage device, and based on a first response vector whose element is the distance for each size of the partial region, in the first evaluation region Determining a presence or absence of a defect. A defect detection method for detecting defects in an inspection image based on a learning result acquired from a learning image using a computer including an arithmetic device and a storage device,
The arithmetic device includes a first evaluation area sequentially set in the learning image stored in the storage device, and extracts image information of each of a plurality of partial areas having different sizes;
The arithmetic device sequentially generates an evaluation area image vector for each size of the partial area from each of the image information extracted from the learning image;
The arithmetic device sequentially calculating a distance from the corresponding eigenspace for each of the evaluation region image vectors for each size of the partial region;
Step the arithmetic unit, based on the set of response vectors, each of which with the distance elements for each size of the partial region, to determine the determine specific functions that define the boundary between the set outer and the inner set When,
The arithmetic device includes a plurality of second evaluation regions sequentially set in the inspection image stored in the storage device, and a plurality of partial regions extracted from the learning image, each having the same size Extracting image information of each partial area, and determining whether or not there is a defect in the second evaluation area based on the inspection image vector generated from the extracted image information and the discriminant function; Including a defect detection method. A defect detection program for detecting defects in an inspection image based on a learning result obtained from a learning image using a computer, the computer comprising:
Means for extracting image information of each of a plurality of partial areas including the first evaluation area set in the inspection image and having different sizes from each other;
Means for generating an inspection image vector for each size of the partial area from each of the image information extracted from the inspection image;
Means for calculating a distance between the corresponding eigenspace for each of the inspection image vectors for each size of the partial region;
Function as means for judging presence / absence of a defect in the first evaluation region based on a first response vector having the distance for each size of the partial region as an element with reference to a predetermined discriminant function Defect detection program. A defect detection program for detecting defects in an inspection image based on a learning result obtained from a learning image using a computer, the computer comprising:
Means for extracting image information of each of a plurality of partial areas including first evaluation areas sequentially set in the learning image and having different sizes;
Means for sequentially generating an evaluation area image vector for each size of the partial area from each of the image information extracted from the learning image;
Means for sequentially calculating a distance from a corresponding eigenspace for each of the evaluation area image vectors for each size of the partial area;
It means for each based on a set of response vector to the distance elements for each size of the partial region, to determine the determine specific functions that define the boundary between the set outer and the inner set,
Extract image information of a plurality of partial areas each including a second evaluation area sequentially set in the inspection image and having the same size as the plurality of partial areas extracted from the learning image. A defect detection program that functions as means for determining the presence / absence of a defect in the second evaluation region based on the inspection image vector generated from the image information and the discriminant function.