JP2004191112A - Defect examining method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a defect examining method of high accuracy corresponding to various kinds of defects. <P>SOLUTION: This defect examining method comprises acquiring a plurality of examined images by photographing an examined object under various optical conditioners (step S1), creating a plurality of examination frequency images by performing Fourier transform on the plurality of acquired examination images (step S2), processing the plurality of created examination frequency images as input signals by a neutral network (step S3), and determining whether the examined object is good or bad on the basis of an output signal of the neutral network (step S4). The defect can be examined with high accuracy corresponding various kinds of defects by representing the difference in shapes of an examined object surface caused by the difference in optical conditions, and performing the operation in the neutral network not on the basis of the image data relating to the plurality of examined images under various optical conditions, but on the basis of the examination frequency image data transformed into the examination frequency components by Fourier transform. <P>COPYRIGHT: (C)2004,JPO&NCIPI

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a defect inspection method for performing a pass / fail judgment on a workpiece and other various inspection objects.
[0002]
[Prior art]
There are various types of defect inspection methods for determining the quality of an inspected object. As a method for detecting different types of defects with high accuracy, an inspection image obtained by imaging the inspected object under a plurality of optical conditions has been conventionally used. There is a method that uses the difference between them.
[0003]
There are also various proposals in which the type of defect is determined using a neural network (for example, see Patent Document 1).
[0004]
[Patent Document 1]
Japanese Patent No. 2758260
[0005]
[Problems to be solved by the invention]
However, in the former case, there is a problem that it is difficult to quantitatively evaluate various defects using only differences between inspection images.
[0006]
Further, in the case of Patent Document 1, the type of defect is determined by processing using a neural network. However, the processing is based on a single inspection image. In many cases, it is difficult to determine whether the surface is uneven or not, and a highly accurate defect inspection cannot always be performed.
[0007]
An object of the present invention is to provide a highly accurate defect inspection method corresponding to a plurality of types of defects.
[0008]
An object of the present invention is to provide a high-speed and high-accuracy defect inspection method.
With the goal.
[0009]
SUMMARY OF THE INVENTION An object of the present invention is to provide a defect inspection method capable of determining with high accuracy whether an abnormality in an inspection object is light or shade or unevenness.
[0010]
An object of the present invention is to provide a defect inspection method capable of determining with high accuracy whether an abnormality of an inspection object is light or dark or uneven without requiring strict alignment in difference processing between a plurality of images. is there.
[0011]
An object of the present invention is to provide a highly accurate defect inspection method capable of accurately determining an abnormality on an inspection object.
[0012]
An object of the present invention is to provide a highly accurate defect inspection method capable of accurately determining the state of unevenness on an inspection object.
[0013]
An object of the present invention is to provide a defect inspection method capable of outputting useful data for process management.
[0014]
SUMMARY OF THE INVENTION An object of the present invention is to provide a highly accurate defect inspection method capable of setting a detailed inspection level for each type of defect.
[0015]
[Means for Solving the Problems]
The invention according to claim 1 illuminates the inspection object with an illumination device, captures an image of a state of the illuminated inspection object by an imaging device, and uses the neural network based on the inspection image acquired by the imaging to execute the inspection object. A defect inspection method for determining the quality of the inspection, the inspection image acquisition step of capturing the inspection object under different optical conditions and acquiring a plurality of inspection images, and performing a Fourier transform on each of the acquired inspection images to obtain a plurality of inspection images. A conversion processing step of creating an inspection frequency image of the above, a neural network processing step of causing the neural network to perform an arithmetic processing using the created inspection frequency images as input signals, and the inspection target based on an output signal of the neural network. And a judging step of judging the quality of an object.
[0016]
Therefore, a plurality of inspection frequency images are created by performing a Fourier transform on each of a plurality of inspection images obtained by imaging the inspection object under different optical conditions, and the plurality of inspection frequency images are arithmetically processed by a neural network. It is possible to express the difference in the shape of the surface of the inspection object from the difference in the optical conditions, and to use the inspection frequency image data converted to the inspection frequency component by the Fourier transform instead of mere image data for a plurality of inspection images under such different optical conditions. Since the arithmetic processing is performed by a neural network, highly accurate defect inspection corresponding to a plurality of defect types can be performed.
[0017]
Here, as the different optical conditions, various methods such as changing the illumination angle with respect to the object to be inspected and changing the illumination method can be employed. Some examples thereof can be easily realized by the methods as described in claims 11 to 15. It is. That is, a plurality of inspection images having different optical conditions are obtained by using one imaging device and a plurality of illumination devices having different illumination conditions with respect to the object to be inspected and switching the illumination devices. Or a plurality of inspection images under different optical conditions using one illumination image device and a plurality of imaging devices having different imaging conditions for the object to be inspected, as in the invention according to claim 12. A plurality of inspection images with different optical conditions may be acquired by moving the imaging device using a pair of imaging devices and a lighting device, as in the invention according to claim 13. Or, as in the invention according to claim 14, a plurality of inspection images under different optical conditions may be obtained by moving the illumination device by using the pair of imaging devices and the illumination device. Imaging Using the location and the illumination device may acquire a plurality of test images of different optical conditions by displacing the object to be inspected.
[0018]
According to a second aspect of the present invention, in the defect inspection method according to the first aspect, in the inspection image obtaining step, the inspection image obtained by the imaging device is divided into a predetermined area, and in the conversion processing step, the inspection image is divided. A plurality of inspection frequency images are created by performing a Fourier transform on the inspection image of the created region, and in the neural network processing step, the created inspection frequency images are used as input signals of the neural network. The quality of the inspection object is determined for each of the regions.
[0019]
Therefore, the inspection image obtained by the imaging apparatus is divided into predetermined regions, a plurality of inspection frequency images are created by performing a Fourier transform on the inspection image in the divided region, and the created inspection frequency images are neurally converted. By processing on the net and determining the quality of the inspection object for each divided area, it is possible to increase the calculation speed such as frequency conversion, as compared with the case where the entire inspection image is targeted as a unit, High-speed and high-accuracy defect inspection becomes possible.
[0020]
According to a third aspect of the present invention, in the defect inspection method according to the first aspect, there is provided a unique region extracting step of extracting a unique region by processing an inspection image obtained by the imaging device with a predetermined algorithm. In the step, a plurality of inspection frequency images are created by Fourier-transforming the extracted inspection image of the singular region. In the neural network processing step, the created inspection frequency images are used as input signals of the neural network.
[0021]
Therefore, by extracting a unique region from the inspection image, Fourier transform the extracted inspection image of the unique region to create a plurality of inspection frequency images, by processing the created plurality of inspection frequency images with a neural network, Higher-speed and more accurate defect inspection can be performed than when the inspection image itself is targeted.
[0022]
According to a fourth aspect of the present invention, in the defect inspection method according to any one of the first to third aspects, the transform processing step uses an FFT (Fast Fourier Transform) for the Fourier transform.
[0023]
Therefore, by using the FFT for the Fourier transform processing, it is possible to further increase the processing speed.
[0024]
According to a fifth aspect of the present invention, in the defect inspection method according to any one of the first to fourth aspects, there is provided a difference image creating step of creating at least one difference image from a plurality of inspection images obtained by the imaging device. The conversion processing step includes a process of creating a difference frequency image by Fourier-transforming the difference image, and the neural network processing step includes the created difference frequency image as an input signal of the neural network.
[0025]
Therefore, by creating at least one difference image from a plurality of inspection images, creating a difference frequency image by performing a Fourier transform on the difference image, and including the created difference frequency image in processing in the neural network, The relationship between the images becomes clear, so that the defect inspection can determine with high accuracy whether the abnormality of the inspection object is light or dark or uneven.
[0026]
According to a sixth aspect of the present invention, in the defect inspection method according to any one of the first to fourth aspects, at least one frequency difference image is created from a plurality of inspection frequency images created by Fourier transform in the conversion processing step. The method further includes a frequency difference image creating step, wherein the neural network processing step includes the created frequency difference image in an input signal of the neural network.
[0027]
Therefore, by creating at least one frequency difference image from a plurality of inspection frequency images created by the Fourier transform and including the created frequency difference image in the processing by the neural network, the relationship between the inspection frequency images is clear. Accordingly, a defect inspection that can determine with high accuracy whether the abnormality of the inspection object is light or dark or uneven without requiring precise alignment in the difference processing between the plurality of images.
[0028]
According to a seventh aspect of the present invention, in the defect inspection method according to any one of the first to fourth aspects, a feature amount is calculated from the inspection image by processing the inspection image obtained by the imaging device using a predetermined algorithm. The neural network processing step includes the calculated feature in an input signal of the neural network.
[0029]
Therefore, by calculating the feature amount from the inspection image and including the calculated feature amount in the processing by the neural network, a highly accurate defect inspection capable of accurately determining an abnormality on the inspection object is achieved.
[0030]
According to an eighth aspect of the present invention, in the defect inspection method according to any one of the first to fourth aspects, a difference image creating step of creating at least one difference image from a plurality of inspection images obtained by the imaging device; And a feature amount calculating step of calculating a feature amount by processing the created difference image by a predetermined algorithm. In the neural network processing step, the calculated feature amount is included in an input signal of the neural network.
[0031]
Therefore, at least one difference image is created from a plurality of inspection images, a feature amount is calculated from the created difference image, and the calculated feature amount is included in the processing in the neural network, so that a This is a highly accurate defect inspection that can accurately determine the state of the unevenness.
[0032]
According to a ninth aspect of the present invention, in the defect inspection method according to any one of the first to eighth aspects, in the neural network processing step, the quality of the object to be inspected and the type of defect are included in an output signal of the neural network. .
[0033]
Therefore, by including not only the quality of the object to be inspected but also the type of defect in the output signal of the neural network, the defect inspection can output useful data for process management.
[0034]
According to a tenth aspect of the present invention, in the defect inspection method according to any one of the first to eighth aspects, in the neural network processing step, a type of the defect is output as an output signal of the neural network. The quality of the inspection object is determined by a predetermined algorithm for each type of output defect.
[0035]
Therefore, the type of defect is output as an output signal of the neural network, and the quality of the inspection object is determined for each type of output defect, so that a detailed inspection level can be set for each type of defect and high accuracy can be achieved. Defect inspection.
[0036]
BEST MODE FOR CARRYING OUT THE INVENTION
A first embodiment of the present invention will be described with reference to FIGS. In the present embodiment, for example, an inspection object such as a work is illuminated by an illumination device, the state of the illuminated inspection object is imaged by an imaging device, and the inspection object is inspected by a neural network based on the inspection image acquired by the imaging. This is applied to a defect inspection method for judging pass / fail. FIG. 1 is a schematic flowchart showing a basic flow of the method.
[0037]
First, the object to be inspected is imaged under different optical conditions, and a plurality of inspection images (image inputs 1, 2,..., N) having different characteristics on the image depending on the type of defect are acquired (step S1; inspection image acquiring step). . As the different optical conditions in step S1, for example, as shown in FIG. 2, the inspection object 1 is illuminated by a certain illumination device 2a and imaged by the imaging device 3, and then a position different from that of the illumination device 2a In this case, two inspection images having different illumination conditions (illumination angles) may be acquired by illuminating with another illuminating device 2b installed in the camera and imaging with the imaging device 3. Such two inspection images can express the difference in the shape of the surface of the inspection object 1 from the difference in the illumination angle.
[0038]
Next, a plurality of inspection frequency images are created by performing a Fourier transform (Fourier transforms 1, 2,..., N) on each of the acquired inspection images (image inputs 1, 2,..., N) separately. (Step S2; conversion processing step). That is, not only image data but also inspection frequency image data converted into inspection frequency components suitable for defect inspection according to the type of defect by Fourier transform is created for the inspection image. Here, a well-known technique may be used for Fourier transform, but it is particularly preferable to use FFT (Fast Fourier Transform).
[0039]
Next, each pixel data of the created inspection frequency image is input to the neural network as an input signal, and the neural network performs an arithmetic process (step S3; neural network processing step). In this neural network, an output signal relating to the quality of the inspection object 1 is output based on the result of the arithmetic processing, so that the quality of the inspection object 1 is determined based on the output signal (step S4; determination step).
[0040]
As described above, according to the present embodiment, a plurality of inspection frequency images are created by performing Fourier transform on a plurality of inspection images acquired by imaging the inspection object 1 under different optical conditions, and a plurality of inspection frequency images are formed. Since the frequency image is processed by the neural network, the difference in the shape of the surface of the inspection object 1 can be expressed based on the difference in the optical conditions, and the Fourier transform and the processing by the neural network are performed on such an inspection image. High-precision defect inspection corresponding to the type of defect. In particular, by using a neural network that has the generalization ability to interpolate between discrete values even when learning discrete values, it is possible to perform highly accurate defect inspection on different types of defects with less training data. Become.
[0041]
Here, well-known technology may be used for the arithmetic processing by the neural network. FIG. 3 shows an example of a schematic configuration of the neural network 4 in principle. FIG. 3 shows the neural network 4 in an extremely simplified manner. The neural network 4 includes an input layer 5, an intermediate layer 6, and an output layer 7, and pixel data 1, 2, 2,. .., N are input, and not only the quality (OK or NG) of the object to be inspected 1 but also the type of defect (shade defect or protrusion defect) is output as teacher data as teacher data by an error back propagation method or the like. It has been learned so that the error from the data is small, and the quality (OK or NG) of the inspection object 1 and the type of the defect (shade defect or dark defect) are determined according to the state of the inspection frequency images 1, 2,. (Protrusion defect). As described above, by including the type of defect as the output signal of the neural network 4, it is possible to use the data as useful data in the process management of the manufacturing process of the inspection object 1.
[0042]
In this embodiment, the quality of the test object 1 is directly determined as an output signal of the neural network 4 itself. However, as shown in FIG. And the type of the output defect is determined (step S5), and a defect inspection process (a process for a light and dark defect, a process for a concave / convex defect) is performed by a detailed method such as image measurement predetermined for each type of the defect. (Step S6), and whether the final inspection object 1 is good or bad based on the processing result (Step S7) may be made to perform the defect detection with higher accuracy. . In steps S6 and S7, a determination step of determining the quality of the object to be inspected by a predetermined algorithm is performed for each type of output defect.
[0043]
By the way, in the present embodiment, as a method of acquiring a plurality of inspection images under different optical conditions, a plurality of illumination devices 2a and 2b are switched and used, and imaging is performed by one imaging device 3. As shown in (a), two inspection images having different illumination angles are obtained by illuminating the inspection object 1 with one illumination device 2 and capturing images with a plurality of imaging devices 3a and 3b disposed at different positions. May be obtained.
[0044]
Alternatively, as illustrated in FIG. 5B, the inspection object 1 is illuminated by the illumination device 2 using the pair of illumination devices 2 and the imaging device 3, and an image is captured by the imaging device 3 positioned at the position indicated by the solid line. Later, two inspection images having different illumination angles may be acquired by moving the imaging device 3 to the position of the virtual line and capturing an image.
[0045]
Alternatively, as shown in FIG. 5C, after the inspection object 1 is illuminated by the illumination device 2 installed at the position of the solid line using the pair of illumination devices 2 and the imaging device 3, Alternatively, two inspection images having different illumination angles may be acquired by moving the illumination device 2 to the position of the virtual line and imaging the imaging device 3.
[0046]
Further, as shown in FIG. 5D, the inspection object 1 is set in a state shown by a solid line, illuminated by the illumination device 2 and imaged by the imaging device 3, and then the inspection object 1 is shown by a virtual line. Two inspection images having different illumination angles may be acquired by displacing the state and taking an image with the imaging device 3.
[0047]
Furthermore, although the description has been given of the example of the difference in the illumination angle as the different optical conditions, the present invention is not limited to this, and for example, the difference in the illumination method may be used. For example, oblique illumination, ring illumination, or coaxial epi-illumination may be used. Alternatively, the illumination may be parallel light having directivity and diffused light having no directivity. In addition, a plurality of inspection images with different optical conditions can be obtained by imaging with an imaging device having illumination of different colors and a filter corresponding thereto. It is also effective to set the type of illumination to visible light, infrared light, ultraviolet light, X-ray, or the like, and to image each with an imaging device adapted to the illumination light depending on the inspection object 1 and the defect to be detected. is there.
[0048]
Although not particularly shown in the embodiment, the unit of image processing is reduced, that is, the inspection image obtained by the imaging device 3 is divided into predetermined regions, and A plurality of inspection frequency images are created by Fourier-transforming the image data, and the created inspection frequency images are used as input signals of the neural network 4 to determine the quality of the inspection object 1 for each divided image region. You may do so. According to this, the calculation speed such as frequency conversion can be increased.
[0049]
A second embodiment of the present invention will be described with reference to FIG. The same parts as those described in the first embodiment are denoted by the same reference numerals, and description thereof is omitted (the same applies to the following embodiments).
[0050]
This embodiment is intended to further increase the speed. That is, each inspection image (image input 1, 2,..., N) obtained by the imaging device 3 in the process of step S1 is processed by a predetermined algorithm to form a unique region (singular region extraction 1, 2). ,..., N) (step S11; singular region extraction step). The predetermined algorithm for extracting the unique region may use a well-known technique in image processing or the like.
[0051]
Therefore, in the conversion processing step S13 of the present embodiment, a plurality of inspection frequency images are created by Fourier-transforming the extracted inspection image of the peculiar area. If there is no peculiar area (N in step S12), the conversion processing Do not do. That is, a plurality of inspection frequency images are created by Fourier transforming the extracted inspection image of the singular region (step S2), and the created inspection frequency images are used as input signals of the neural network to extract the extracted singular region. By performing the processing each time (step S3), it is possible to further increase the speed.
[0052]
Here, in the present embodiment, the extraction processing of the unique region is performed for all of the plurality of inspection images (step S11), but the extraction processing of the unique region is performed only for the inspection image under the specific optical condition. It may be. In addition, high-speed processing can be performed by setting the unit of image processing in the present embodiment to a power of 2 in both the main scanning and sub-scanning directions and making the Fourier transform FFT (Fast Fourier Transform).
[0053]
A third embodiment of the present invention will be described with reference to FIG. In the present embodiment, for example, in addition to the processing described in the first embodiment, a difference image is created from a plurality of inspection images (image inputs 1 and 2) obtained by the imaging device 3 (step S21; difference). Image creation step), Fourier transform of the created difference image is included as one of the conversion processes in step S2, and the obtained difference frequency image is included in the input signal of the neural network in step S3.
[0054]
According to the present embodiment, by including the difference frequency image, it is possible to determine with high accuracy whether the abnormality of the inspection object 1 is light or dark, and it is possible to detect a defect with higher accuracy. it can.
[0055]
A fourth embodiment of the present invention will be described with reference to FIG. In the present embodiment, for example, in addition to the processing described in the first embodiment, Fourier transform processing is performed on each of a plurality of inspection images (image inputs 1 and 2) (step S1) obtained by the imaging device 3. After performing (Fourier transforms 1 and 2) to create a plurality of inspection frequency images (step S2), a frequency difference image is created from the plurality of inspection frequency images (step S31; frequency difference image creation step). The obtained frequency difference image is included in the input signal of the neural network 4.
[0056]
According to the present embodiment, it is possible to determine with high accuracy whether the abnormality of the inspection object 1 is light or dark or uneven without requiring strict alignment in the difference processing between the plurality of inspection images, Defects can be detected with higher accuracy.
[0057]
A fifth embodiment of the present invention will be described with reference to FIG. In the present embodiment, for example, in addition to the processing as described in the first embodiment, the inspection images (input images 1 and 2) obtained by the imaging device are each processed by a predetermined algorithm, so that a characteristic is obtained from the inspection image. The amount is calculated (feature calculation 1, 2) (step S41; feature calculation step), and the calculated feature is included in the input signal of the neural network 4. The predetermined algorithm for calculating the characteristic amount may use a well-known technique in image processing or the like.
[0058]
According to the present embodiment, a feature amount is calculated from an inspection image, and the calculated feature amount is included in the processing in the neural network 4, so that an abnormality on the inspection object 1 can be accurately determined. This is a defect inspection.
[0059]
A sixth embodiment of the present invention will be described with reference to FIG. In the present embodiment, for example, in addition to the processing described in the first embodiment, a difference image is created from a plurality of inspection images (image inputs 1 and 2) obtained by the imaging device 3 (step S21; difference). Image forming step), a feature amount is calculated from the inspection image by processing the created difference image by a predetermined algorithm (step S51; feature amount calculating step), and the calculated feature amount is input to the neural network 4 as an input signal. It is intended to be included.
[0060]
According to the present embodiment, at least one difference image is created from a plurality of inspection images, a feature amount is calculated from the created difference image, and the calculated feature amount is included in processing in the neural network 4. As a result, a highly accurate defect inspection that can accurately determine the state of the irregularities on the inspection object 1 is achieved.
[0061]
【The invention's effect】
According to the invention described in claims 1, 11 to 15, the plurality of inspection images obtained by imaging the inspection object under different optical conditions are respectively subjected to Fourier transform to generate a plurality of inspection frequency images, and the plurality of inspection frequency images are generated. Inspection frequency images are processed by a neural network, and differences in optical conditions can be used to represent differences in the shape of the surface of the inspection object.Inspection of multiple inspection images with such different optical conditions is performed not by simple image data but by Fourier transform Since the arithmetic processing is performed by the neural network using the inspection frequency image data converted into the frequency components, a highly accurate defect inspection method corresponding to a plurality of defect types can be provided.
[0062]
According to the second aspect of the present invention, in the defect inspection method according to the first aspect, the inspection image obtained by the imaging device is divided into predetermined regions, and the inspection image of the divided region is subjected to Fourier transform. Inspection frequency images are created, the created inspection frequency images are processed by a neural network, and the quality of the inspection object is determined for each of the divided areas. Since the calculation speed of the frequency conversion and the like can be increased as compared with the case where it is performed, a high-speed and high-accuracy defect inspection method can be provided.
[0063]
According to the third aspect of the present invention, in the defect inspection method according to the first aspect, a plurality of inspection frequency images are created by extracting a unique region from the inspection image and performing a Fourier transform on the extracted inspection image of the unique region. Since the generated plurality of inspection frequency images are processed by the neural network, a faster and more accurate defect inspection method can be provided than when the inspection image itself is targeted.
[0064]
According to the fourth aspect of the present invention, in the defect inspection method according to any one of the first to third aspects, the FFT is used for the Fourier transform processing, so that further high-speed processing can be achieved. .
[0065]
According to the fifth aspect of the present invention, in the defect inspection method according to any one of the first to fourth aspects, at least one difference image is created from the plurality of inspection images, and the difference image is subjected to Fourier transform to thereby determine the difference frequency. Since the image was created and the created difference frequency image was included in the processing by the neural network, the relationship between the inspection images became clear, and therefore, it was possible to determine whether the abnormality of the inspection object was light or dark or uneven. A defect inspection method that can be determined with high accuracy can be provided.
[0066]
According to a sixth aspect of the present invention, in the defect inspection method according to any one of the first to fourth aspects, at least one frequency difference image is created from a plurality of inspection frequency images created by Fourier transform. Since the frequency difference image is included in the processing by the neural network, the relationship between the inspection frequency images becomes clear, and therefore, the inspection object can be inspected without requiring strict alignment in the difference processing between a plurality of images. It is possible to provide a defect inspection method capable of determining with high accuracy whether the abnormality of the image is light or dark or uneven.
[0067]
According to a seventh aspect of the present invention, in the defect inspection method according to any one of the first to fourth aspects, a feature amount is calculated from the inspection image, and the calculated feature amount is included in processing in the neural network. Therefore, it is possible to provide a highly accurate defect inspection method capable of accurately determining an abnormality on the inspection object.
[0068]
According to an eighth aspect of the present invention, in the defect inspection method according to any one of the first to fourth aspects, at least one difference image is created from the plurality of inspection images, and a feature amount is calculated from the created difference image. Since the calculated feature amount is included in the processing by the neural network, it is possible to provide a highly accurate defect inspection method capable of accurately determining the state of the unevenness on the inspection object.
[0069]
According to the ninth aspect of the present invention, in the defect inspection method according to any one of the first to eighth aspects, the output signal of the neural network includes not only the quality of the inspection object but also the type of the defect. Therefore, it is possible to provide a defect inspection method capable of outputting useful data for process management.
[0070]
According to a tenth aspect of the present invention, in the defect inspection method according to any one of the first to eighth aspects, the type of the defect is output as an output signal of the neural network, and the object to be inspected is output for each type of the output defect. Is determined, the detailed inspection level can be set for each type of defect, and a highly accurate defect inspection method can be provided.
[Brief description of the drawings]
FIG. 1 is a schematic flowchart showing a defect inspection method according to a first embodiment of the present invention.
FIG. 2 is a front view schematically showing an imaging state in which optical conditions are changed.
FIG. 3 is an explanatory diagram schematically showing an example of a neural network.
FIG. 4 is a schematic flowchart showing a modification.
FIG. 5 is a front view schematically showing several examples of an imaging state in which optical conditions are different.
FIG. 6 is a schematic flowchart showing a defect inspection method according to a second embodiment of the present invention.
FIG. 7 is a schematic flowchart illustrating a defect inspection method according to a third embodiment of the present invention.
FIG. 8 is a schematic flowchart showing a defect inspection method according to a fourth embodiment of the present invention.
FIG. 9 is a schematic flowchart showing a defect inspection method according to a fifth embodiment of the present invention.
FIG. 10 is a schematic flowchart illustrating a defect inspection method according to a sixth embodiment of the present invention.
[Explanation of symbols]
1 Inspection object
2 Lighting equipment
3 Imaging device
4 Neural network

Claims (15)

A defect inspection method of illuminating an inspection object with an illumination device, capturing an image of a state of the illuminated inspection object by an imaging device, and determining the acceptability of the inspection object by a neural network based on the inspection image obtained by the imaging. And
An inspection image acquisition step of acquiring the inspection object under different optical conditions and acquiring a plurality of inspection images,
A conversion processing step of Fourier transforming each of the acquired plurality of inspection images to create a plurality of inspection frequency images,
A neural network processing step of causing the neural network to perform arithmetic processing using the created plurality of test frequency images as an input signal;
A judging step of judging the quality of the inspection object based on an output signal of the neural network;
A defect inspection method, comprising: In the inspection image obtaining step, the inspection image obtained by the imaging device is divided into predetermined regions,
In the conversion processing step, a plurality of inspection frequency images are created by Fourier-transforming the inspection image of the divided region,
In the neural network processing step, the plurality of test frequency images created as input signals of the neural network,
2. The defect inspection method according to claim 1, wherein in the determining step, the quality of the inspection object is determined for each of the divided areas. A unique area extraction step of processing a test image obtained by the imaging device with a predetermined algorithm to extract a unique area,
In the conversion processing step, a plurality of inspection frequency images are created by performing a Fourier transform on the extracted inspection image of the unique region,
In the neural network processing step, the created plurality of inspection frequency images as input signals of the neural network,
The defect inspection method according to claim 1, wherein: 4. The defect inspection method according to claim 1, wherein in the conversion processing step, an FFT (Fast Fourier Transform) is used for the Fourier transform. A difference image creating step of creating at least one difference image from the plurality of inspection images obtained by the imaging device,
The conversion processing step includes a process of creating a difference frequency image by Fourier transforming the difference image,
In the neural network processing step, the created difference frequency image is included in the input signal of the neural network,
5. The defect inspection method according to claim 1, wherein: A frequency difference image creating step of creating at least one frequency difference image from the plurality of inspection frequency images created by the Fourier transform in the conversion processing step,
In the neural network processing step, the created frequency difference image is included in the input signal of the neural network,
5. The defect inspection method according to claim 1, wherein: A feature amount calculating step of calculating a feature amount from the inspection image by processing the inspection image obtained by the imaging device with a predetermined algorithm,
In the neural network processing step, the calculated feature amount is included in the input signal of the neural network,
5. The defect inspection method according to claim 1, wherein: A difference image creating step of creating at least one difference image from the plurality of inspection images obtained by the imaging device;
A feature amount calculating step of calculating a feature amount by processing the created difference image by a predetermined algorithm,
In the neural network processing step, the calculated feature amount is included in the input signal of the neural network,
5. The defect inspection method according to claim 1, wherein: In the neural network processing step, the output signal of the neural network includes the quality of the inspection object and the type of defect,
The defect inspection method according to any one of claims 1 to 8, wherein: In the neural network processing step, the type of defect is output as an output signal of the neural network,
In the determining step, the quality of the inspection object is determined by a predetermined algorithm for each type of output defect,
The defect inspection method according to any one of claims 1 to 8, wherein: In the inspection image acquiring step, a plurality of inspection images having different optical conditions are acquired by using one imaging device and a plurality of illumination devices having different illumination conditions with respect to the object to be inspected and switching the illumination device. ,
The defect inspection method according to claim 1, wherein: In the inspection image acquiring step, a plurality of inspection images under different optical conditions are acquired using one illumination image device and a plurality of imaging devices having different imaging conditions for the inspection object.
The defect inspection method according to claim 1, wherein: In the inspection image acquisition step, using a pair of imaging device and a lighting device, to obtain a plurality of inspection images of different optical conditions by moving the imaging device,
The defect inspection method according to claim 1, wherein: In the inspection image obtaining step, using a pair of imaging device and the illumination device, to obtain a plurality of inspection images of different optical conditions by moving the illumination device,
The defect inspection method according to claim 1, wherein: In the inspection image obtaining step, a plurality of inspection images of different optical conditions are obtained by displacing the object to be inspected, using a pair of imaging devices and a lighting device,
The defect inspection method according to claim 1, wherein:

Images