JP7327866B1 - Learning model generation method - Google Patents

Abstract

Kind Code: A1 A learning model generation method for determining the presence or absence of a defect in a transparent test object at high speed without complicating a light source device and increasing the number of captured images. A learning model generation method according to the present invention includes steps of: projecting light composed of a bright portion and a dark portion projected from a light source onto a transparent object to generate a captured image of the transparent object; a step of generating a first image obtained by inverting the luminance value of each pixel, or a first image obtained by removing a region corresponding to a bright portion or a dark portion in the captured image from the edge of the captured image; generating a second image based on absolute values of differences in luminance values for pixels having corresponding positional relationships in the image; and inputting the second image using the second image as teacher data; generating a learning model that outputs transparency defects in the images of 2. [Selection drawing] Fig. 1

Description

The present invention relates to a learning model generation method.

In visual inspection, an image processing method is known in which an image of an inspection object is captured while a black-and-white periodic pattern is irradiated onto the inspection object, and the presence or absence of defects in the inspection object is determined using the obtained image. Reference 1).

Cited Document 1 describes a method for inspecting defects such as minute unevenness of an object having specularity or light transmission properties such as glass. Specifically, the surface of an object to be inspected, such as glass, is irradiated with light using a light source having a rotating mechanism, and the object is inspected by performing image processing based on the captured image of the reflected light. The peripheral surface of the light source is covered with a black and white striped film, and the picked-up image is an image in which the phase of the black and white striped pattern produced by the light source is gradually shifted. Defects such as unevenness can be identified by obtaining the luminance difference value at the same position in the plurality of captured images.

JP-A-2000-018932

The method described in Patent Document 1 requires a rotating light source equipped with a rotary driving device as shown in FIG. 1 or FIG. 7 of Patent Document 1, which may complicate the device configuration. In addition, multiple phase-shifted images are used to discriminate defects, and if such a light source cannot be prepared, defects cannot be discriminated. Since it is necessary to change, there is a problem that inspection time is required.

In order to solve the above problems, a learning model generation method according to the present invention includes a step of projecting light consisting of a bright portion and a dark portion projected from a light source onto a transparent object, and generating a captured image of the transparent object. and generating a first image obtained by inverting the luminance value of each pixel of the captured image, or a first image obtained by deleting the area corresponding to the bright part or the dark part in the captured image from the edge of the captured image. generating a second image based on absolute values of differences in luminance values for pixels in a corresponding positional relationship in the first image and the captured image; to generate a learning model that, when input with the second image, outputs the defects of the transparency in the second image.

Further, in order to solve the above problems, another learning model generation method according to the present invention provides light composed of a bright part and a dark part projected from a light source, and the light passing through a transparent object and the light passing through the transparent object. is used as teacher data, and using the teacher data to generate a learning model that outputs the defect of the transparent object in the captured image when the captured image is input.

SUMMARY OF THE INVENTION The present invention has been made in view of the above-described problems, and can determine whether or not a transparent inspection object has defects at high speed without complicating the light source device and increasing the number of captured images.

1 is a diagram showing an image processing system according to the present invention; FIG. FIG. 4 is a diagram showing a background image according to the present invention; (A) is an inspection object having a line scratch on the surface, and (B) is a grayscale image obtained by converting a captured image of an inspection object having a line scratch along with a background image into a grayscale. 1 is a flow diagram showing an image processing method according to the present invention; FIG. (A) is a grayscale captured image, (B) is a right-shifted image according to the present invention, and (C) is a left-shifted image according to the present invention. FIG. 4 is a diagram showing an inverted image according to the present invention; (A) is a differential image according to the present invention, and (B) is a differential image with noise. (A) is a differential image with noise according to the present invention, and (B) is a diagram for explaining adaptive threshold processing. FIG. 4 is a diagram showing a binary image according to the present invention; It is a figure which shows the captured image of the Example which concerns on this invention. (A) is a captured image of an embodiment according to the present invention, and (B) to (D) are diagrams for explaining defects. It is a figure which shows the difference image of the Example which concerns on this invention. FIG. 11 shows another difference image of an embodiment according to the present invention; (A) is a binary image of an embodiment according to the present invention, and (B) is a binary image after noise processing. (A) is another binary image of an embodiment according to the present invention, and (B) is a binary image after noise processing. (A) is a defect identification image of another embodiment according to the present invention, and (B) is another defect identification image.

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the description below, the same reference numerals are given to the same components.

FIG. 1 is a diagram showing an image processing system 100 according to the present invention. The image processing system 100 determines whether or not the transparent inspection object WK has defects such as scratches. The image processing system 100 includes an imaging device 10 , an image processing device 20 and a light source device 30 . The imaging device 10 includes an optical element, a processor, an imaging element, etc., and acquires an image by imaging. It should be noted that the transparent inspection object may be any object that transmits a predetermined amount of light or more from the light source device 30 .

The image processing device 20 includes a CPU (Central Processing Unit), a memory, a storage such as a hard disk drive or SSD, an input device such as a mouse or keyboard, an output device such as a display or printer, and communication. A communication device for connecting to a network may be provided, and these may be connected via a bus.

With such a configuration, the image processing apparatus 20 can execute image processing and various types of information processing. Further, the image processing device 20 is connected to the imaging device 10, acquires an image captured by the imaging device 10, and processes the image. Note that the imaging device 10 may record the captured image in an external storage device or the like, and the image processing device 20 may read the captured image stored in the external storage device to process the image. It is also possible to cause the output device to output the image processing result, and to reflect user settings and the like in the image processing method via the input device.

The light source device 30 is a device that emits light composed of a bright portion and a dark portion, and is, for example, a display device such as a liquid crystal display device capable of displaying an image, and is capable of displaying the background image 200 shown in FIG. Wear. A pattern similar to the background image 200 is projected onto the inspection object WK. Note that the light source device 30 is not limited to a liquid crystal display device or the like. may be projected onto

The inspection object WK shown in FIG. 1 is arranged directly above the light source device 30 . The imaging device 10 images the inspection object WK together with the background image 200 displayed by the light source device 30 .

FIG. 2 shows a background image 200 according to the invention. Background image 200 is composed of white pixels and black pixels. In background image 200 , white pixels form white pixel groups such as white band 210 , and black pixels form black pixel groups such as black band 220 . As a result, the light composed of the bright portion and the dark portion is projected onto the transparent inspection object WK, which is a transparent object. The background image 200 has a plurality of white bands 210 and black bands 220. The white bands 210 and black bands 220 are of the same size and are alternately arranged.

For example, like the background image 200 shown in FIG. 2, the white belts 210 and the black belts 220 are arranged in stripes in the horizontal direction. It should be noted that the width of the white band 210 and the width of the black band 220 are preferably the same width in order to reduce the influence of noise in subsequent image processing. If the deviation is moderate, the image processing method according to the present invention can be executed, and for example, noise processing can be performed appropriately.

Next, an image processing method executed by the image processing device 20 will be described. It is assumed that the transparent inspection object WK has a line scratch 310 on its surface as shown in FIG. 3(A). In this case, the imaging device 10 acquires a captured image 300 as shown in FIG. 3(B). Then, the image processing device 20 performs image processing after converting the captured image 300 into a gray scale.

The captured image 300 includes the background image 200 transmitted through the inspection object WK, and includes white pixels reflecting bright portions and black pixels reflecting dark portions. In addition, the line flaw 310 on the inspection object WK is also represented in a neutral gray color under the influence of the defect, like the line flaw 350 in the captured image 300 .

The image processing device 20 acquires the captured image 300 from the imaging device 10 and performs image processing according to the flow of the image processing method shown in FIG.

First, in S410, the captured image 300 is converted into a grayscale captured image 500 with 256 gradations. As a result of the gray-casing, a pixel with a luminance value of "255" is a white pixel, and a pixel with a luminance value of "0" is a black pixel. It should be noted that this processing is not necessary if the acquired captured image has already been converted to grayscale. In addition, in S410, tilt correction processing is performed as necessary. The inclination correction process is performed by correcting the orientation of the black bands 501-1 to 501-6 and the white bands 502-1 to 502-6 by capturing the background image 200 included in the grayscaled captured image 500 in the horizontal direction of the image. is to correct it so that it becomes horizontal when it is tilted upward with respect to Alternatively, the tilt correction is correction so that the direction of the boundary between the black bands 501-1 to 501-6 and the white bands 502-1 to 502-6 is perpendicular to the horizontal direction of the image. For this correction, for example, it is possible to determine how many times the tilt should be corrected by specifying the boundary line based on the change in the luminance value of the pixel and obtaining the angle of the boundary line.

Next, in S420, a shift image or an inverted image is generated based on the grayscaled captured image 500. FIG.

The shift image is an area consisting of the width corresponding to the width of the black band or white band included in the grayscaled captured image and the vertical width of the grayscaled captured image. This image is deleted with the left end as the reference. An image deleted with reference to the right end is referred to as a "right-shifted image", and an image deleted with respect to the left end is referred to as a "left-shifted image". When generating a shifted image, at least one of a right shifted image, a left shifted image, or both will be generated.

FIG. 5 is a diagram for explaining a method of generating a right-shifted image 510 or a left-shifted image 520, which are shifted images, based on a grayscaled captured image 500. FIG. The grayscaled captured image 500 captures the background image 200 as black bands 501-1 to 501-6 and white bands 502-1 to 502-6. In addition, the grayscaled captured image 500 includes a line flaw 503 as a flaw that the inspection unit WK has.

A right-shifted image 510 is created by deleting, from the right end, a region having a size of the vertical width and the horizontal width of the black band 501 or the white band 502 from the grayscaled captured image 500 . An area 512 corresponds to this area. The region 512 has a vertical width H and a horizontal width W. As shown in FIG. The vertical width H is the same size as the vertical width of the grayscaled captured image 500 . The width W is obtained by, for example, scanning the left end to the right end of the grayscaled captured image, specifying the number of white bands and black bands based on the change in luminance value, and determining the number of pixels in the horizontal width of the grayscaled captured image. is divided by the specified number of lines, and the number of pixels is rounded off. The vertical width H is the vertical width of the captured image 500 converted to grayscale.

A region 512 having a width W and a height H has approximately the same size as the black band 501-1 in the captured image 500 converted to grayscale. Therefore, if the region 512 is deleted with reference to the right end of the grayscaled captured image 500, the black belt 501-1 is deleted, and as a result, the right-shifted image 510 can be generated.

The left-shifted image 520 is obtained by deleting from the left edge of the grayscaled captured image 500 an area having the size of the vertical width of the grayscaled captured image 500 and the horizontal width of the black band 501 or the white band 502 . created. A region 521 corresponds to this region. The region 521 has a vertical width H and a horizontal width W. As shown in FIG. The vertical width H is the same size as the vertical width of the grayscaled captured image 500 . The horizontal width W and the vertical width H are the same as those described regarding the generation of the right-shifted image 510 above.

A region 521 having a width W and a height H has substantially the same size as the white band 502-6 in the captured image 500 converted to grayscale. Therefore, if the area 521 is deleted with reference to the left end of the grayscaled captured image 500, the white band 502-6 is deleted, and as a result, the left-shifted image 520 can be generated.

Regarding the shifted image, an example in which a black band or a white band is deleted from the grayscaled captured image 500 has been described. is the starting point, the area to be deleted is the same size as the black or white band.

On the other hand, in the reverse image, the absolute value of the value obtained by subtracting 255 from the luminance value of the pixel of interest in the grayscaled captured image 500 is used as the value of the pixel of interest in the reverse image. For example, since a white pixel in the grayscaled captured image 500 has a luminance value of 255, subtracting 255 results in 0, and in the inverted image, it becomes a black pixel. Also, since the luminance value of black pixels in a grayscaled captured image is 0, subtracting 255 results in -255, and since the absolute value is 255, it becomes a white pixel in an inverted image.

A reverse image 600 shown in FIG. 6 is created based on the captured image 500 converted to grayscale. Comparing the grayscaled captured image 500 and the inverted image 600, the pixels at the positions of the black bands 501-1 to 501-6 in the grayscaled captured image 500 are white pixels in the reversed image 600. ~6. Also, the pixels at the positions of the white bands 502-1 to 502-6 in the grayscaled image 500 are the black bands 602-1 to 602-6 made up of black pixels in the inverted image 600. FIG. Regarding the pixel of the line flaw 503 of the grayscaled captured image 500, the luminance value is near the middle value, and if the value is 127, subtracting 255 gives -128, and the absolute value is 128, so it is inverted. The pixels in the image also exhibit similar luminance values. Therefore, like the reverse image 600, the line blemish 603 shows the intermediate value.

Next, in S430, a difference image is generated. The difference image is generated based on the absolute value of the difference between the luminance value of the pixel of interest in the grayscaled captured image and the luminance value of the pixel of interest in the shifted or inverted image.

First, an example of processing for creating a differential image based on the inverted image 600 and the grayscaled captured image 500 will be described. Since the inverted image 600 and the grayscaled captured image 500 are images of the same size, the pixels of interest have the same positional relationship. White pixels in the inverted image 600 become black pixels in the grayscaled captured image 500 , and black pixels in the inverted image 600 become white pixels in the grayscaled captured image 500 . For example, the position of the white band 601-1 in the inverted image 600 is the black band 501-1 in the grayscaled captured image 500. FIG. In this way, each pixel of interest has a relationship between a white pixel and a black pixel, and the absolute value of the difference is "255".

On the other hand, if the pixel of interest is a pixel that indicates a luminance value that indicates an intermediate value, such as the flaw 603 , the pixel that indicates a luminance value that indicates an intermediate value also exists in the grayscaled captured image 500 , like the flaw 503 . Therefore, when the difference between these luminance values is obtained, it becomes a value close to "0", and the difference image 700 is black or gray close to black.

Next, processing for creating a difference image based on the right-shifted image 510 and the grayscaled captured image 500 will be described. First, the pixel correspondence relationship between the right-shifted image 510 and the grayscaled captured image 500 is set. Specifically, the upper right corner point 513 of the right-shifted image and the upper right corner point 504 of the grayscaled captured image 500 match, and the lower right corner point 514 of the right-shifted image and the upper right corner point 506 of the grayscaled captured image 500 match. Set the positional relationship of each image so that

After setting this positional relationship, the absolute value of the difference in luminance value is obtained for the pixel of interest that has a corresponding relationship between the right-shifted image 510 and the grayscaled captured image 500, and the obtained value is used as the luminance value of the pixel in the difference image. . Since the right-shifted image 510 is smaller than the grayscaled captured image 500 by the area 512, the difference is not calculated for the region where the right-shifted image 510 and the grayscaled captured image 500 do not overlap. . This non-overlapping region is roughly the white band 502-6 of the grayscaled captured image 500. FIG. Therefore, the difference image has the same size as the right-shifted image 510 .

When obtaining the absolute value of the luminance value difference, if there is a pixel where the flaw 512 of the right-shifted image 510 and the flaw 504 of the grayscaled captured image 500 overlap, the difference in the luminance value of the overlapping pixels. has an absolute value of approximately 0. Therefore, in the difference image, the pixel is black or gray close to black. For pixels where the blemish 512 and blemish 504 do not overlap, the absolute value of the difference between the luminance value of the pixel showing the blemish and the luminance value of white or black is obtained, and the luminance value of the pixel is roughly intermediate gray.

The black band 501-1 of the grayscaled captured image 500 is at the same position as the white band 512-1 of the right-shifted image 510. FIG. Therefore, the absolute value of the difference is 255, and each pixel at the positions of the black belt 501-1 and the white belt 512-1 in the difference image is white. For example, as shown in FIG. 7A, in a differential image 700 based on the right-shifted image, each pixel corresponding to the positions of the black band 501-1 and the white band 512-1 in the area 701 is white.

The process of creating a differential image based on the left-shifted image 520 and the grayscaled captured image 500 is also the above-described process of creating a differential image based on the right-shifted image 510 and the grayscaled captured image 500. is similar to Note that the positional reference is that the upper left corner point 522 of the left-shifted image 520 and the upper left corner point 505 of the grayscaled image 500 match, and The end point 507 is set so as to match.

A difference image 700 is shown in FIG. In the difference image 700, the pixels of interest become white pixels as a result of obtaining the difference with respect to the black bands 501-1 to 501-6 and the white bands 502-1 to 502-6 included in the grayscaled captured image 500. The image does not include a band. Also, since the flaw 701 is shown in black or gray in the differential image, the flaw can be identified.

Although the differential image is obtained as described above, noise may occur in the differential image. For example, multiple lines 711 may appear as in the difference image 710 shown in FIG. 7B. This is because the width of the black band and the white band in the captured image differ, or the luminance value changes gradually near the boundary between the white band and the black band due to the influence of the imaging environment and camera performance. there is Thus, when a differential image is generated based on a captured image in which a background image is reflected, noise such as vertical lines may occur at the position of the boundary between the black band and the white band. Note that the luminance value of the line 711 can take values from 0 to around 128, which is an intermediate value.

In both the differential image 700 and the differential image 710 with noise, when compared with the grayscaled captured image 500, the black band 501 included in the grayscaled captured image 500 is removed. The line flaw 701 or the line flaw 712 can be easily visually recognized when viewed with the human eye, and the presence or absence of a defect in the inspection object WK can be easily determined.

The difference image is generated as described above, and the line flaws that were difficult to see with the human eye due to the periodic arrangement of black and white bands in the grayscaled image are removed in the difference image. This makes it easier for the human eye to visually recognize the line scratches.

Next, in S440, binarization processing is performed to generate a binary image capable of identifying defects such as scratches while removing noise contained in the difference image.

FIG. 8 is a diagram for explaining an example of binarization processing for a noise-containing difference image. Adaptive threshold processing is performed in the binarization processing. Note that the processing in the region 801 of the noise-containing difference image 710 shown in FIG. 8A will be described as an example. The Niblack method can be used for this adaptive threshold processing, but it is not limited to this, and the Sauvola method or other processing may be used.

FIG. 8B is an enlarged view of the region 801. FIG. As shown in FIG. 8B, in adaptive threshold processing, a frame 811 of a predetermined size centered on a pixel of interest 812 is set. The size of the frame 811 is, for example, 1 horizontal pixel×50 vertical pixels. Here, the horizontal width of the frame 811 is set to 1 pixel, but is set to a narrow value in order to accurately determine the change in luminance of the vertical component when pixels indicating defects such as scratches are included. Then, a threshold value is determined based on the average value and standard deviation of the luminance values of the pixels included in the frame 811, and the pixel of interest is set to "0" or "1" based on this threshold value. It should be noted that the vertical size should preferably be a pixel size that is at least about twice the size of the assumed flaw.

Then, while sliding the frame 811 rightward in the direction indicated by an arrow 813, "0" or "1" is set for the pixel of interest. If the frame 813 has moved to the right end, the pixel of interest 812 is shifted downward by one pixel and the processing proceeds in the same manner from the leftmost end.

A binary image 900 shown in FIG. 9 is obtained by the binarization processing in S440. As in the binary image 900, the flawed portion is represented by white pixels.

Next, in S450, defects such as scratches are identified from the binary image. A defect can be identified by, for example, a known labeling process or the like, and when an area of white pixels having a predetermined size exists, it can be treated as a defect such as a scratch.

<Modification 1>
In the embodiment of the invention described above, a binary image is generated in S440, and defects such as scratches are specified based on this binary image in S450. Alternatively to this process, a learned model may be used to identify defects such as scratches.

The learning model is constructed by using grayscaled captured images as teacher data. Using this teacher data, a learning model is generated in which a grayscaled captured image is input and a captured image showing a defect in a transparent object is output. It should be noted that the grayscaled captured image used as training data is preferably a captured image of an inspection object that does not include defects such as scratches. However, even if the captured image is an image of an inspection object that includes defects such as scratches, a region that does not show defects such as scratches may be extracted from the grayscaled captured image and used as teacher data.

<Modification 2>
In the embodiment of the invention described above, when generating a shift image in S420, a difference image is generated based on the shift image and the grayscaled captured image in S430. In this case, for example, in the case of a right-shifted image, the defect in the difference image may be larger than that in the captured image by the amount of the defect shifted to the right. Therefore, in S430, both a right-shifted image and a left-shifted image are created, and a first differential image is generated based on the right-shifted image and the grayscaled captured image. Also, a second difference image is generated based on the left-shifted image and the grayscaled captured image.

Next, at S440, a first binary image based on the first difference image and a second binary image based on the second difference image are generated. A third binary image is generated by logical product of corresponding pixels of the first binary image and the second binary image. identify. By performing such processing, it is possible to prevent the difference image using the shifted image from becoming a large flaw compared to the captured image. In generating the third binary image, the pixel correspondence is set so that the first binary image is aligned with the right end of the captured image, and the second binary image is aligned with the left end.

<Modification 3>
In the above-described embodiment of the invention, the captured image 300 is captured so that white bands and black bands are alternately arranged in the horizontal direction. When such a captured image is acquired, if the background image 200 is also an image in which white bands and black bands are alternately arranged in the horizontal direction, this is rotated by 90° so that white bands and black bands are alternately arranged. may be displayed in such a way that the images are arranged in the vertical direction.

As a result, the captured image becomes an image in which white stripes and black stripes are alternately arranged in the vertical direction. In this case, when generating the displaced image in S420, an upper displaced image in which the area corresponding to the white band or black band of the grayscaled image is deleted from the upper end or a lower displaced image in which the region corresponding to the gray scaled captured image is deleted from the lower end is generated. When performing local thresholding in S440, the frame is set to 50 pixels wide by 1 pixel high. This is because noise occurs in the horizontal direction.

It is also possible to combine a plurality of Modifications 1 to 3 described above and implement them as modifications of the embodiment. For example, modified example 1 may be implemented using the captured image captured in modified example 3 to acquire a captured image showing a defect in a transparent object.

<Example>
Examples are described below. A 13-inch LCD display device was used as the light source device 30 . When the background image 200 was displayed on the light source device 30, the width of the black band and the white band was 1 mm. As the imaging device 10, a camera (24 mm focal length, full HD) mounted on a mobile phone was used.

The material to be inspected WK used was polypropylene, 6.5 mm×9 mm, and 0.2 mm thick. This inspection object WK has three defects, which are line flaws with lengths of about 23 mm, about 1 mm, and about 0.8 mm, and a width of about 0.3 mm. The distance between the light source device 30 and the inspection object WK is 12 cm, and the distance between the inspection object WK and the imaging device 10 is 25 cm.

FIG. 10 shows a grayscale captured image 1000 obtained by converting the captured image to grayscale. Regarding the captured image 1000 converted to grayscale, it can be confirmed that the three flaws of the inspection object WK are reflected together with the background image. Specifically, it can be confirmed in three areas 1101, 1102, and 1103 on the grayscaled captured image 1000 shown in FIG. FIG. 11B is an enlarged view of the area 1101, and it can be confirmed that the pixels indicating the intermediate luminance value are arranged in the horizontal direction. FIG. 11C is an enlarged view of the area 1102, and FIG. 11D is an enlarged view of the area 1103. FIG. Similarly, it can be confirmed that the pixels indicating the intermediate value of the luminance value are arranged in the horizontal direction.

Next, at S420, a shifted or inverted image is generated. Then, in S430, a difference image is generated. A difference image 1200 shown in FIG. 12 is created based on a grayscaled captured image 1000 and a right-shifted image obtained by deleting a predetermined size based on the grayscaled captured image 1000 .

A differential image 1300 shown in FIG. 13 is created based on the grayscaled captured image 1000 and the inverted image created based on the grayscaled captured image 1000 .

It can be seen that in both the difference image 1200 and the difference image 1300, the line scratches are easier to visually recognize than in the captured image 1000 in which the line scratches are grayscaled.

Next, at S440, a binary image is generated. A binary image 1400 shown in FIG. 14 is a binary image generated by performing adaptive threshold processing on the difference image 1200 . For adaptive threshold processing, a frame of 1 horizontal pixel×50 vertical pixels is set.

Next, the binary image 1400 is subjected to noise removal and expansion processing to generate a binary image 1410 . In the noise removal, white pixel clusters of 1 pixel or less in width, 4 pixels or less in height, or 49 pixels or less in area are rewritten to black pixels. As for the dilation process, each block of white pixels is dilated by 2 pixels vertically and horizontally.

On the other hand, a binary image 1500 shown in FIG. 15 is generated based on the difference image 1300 in S440. A binary image 1500 is a binary image generated by performing adaptive threshold processing on the difference image 1300 . For adaptive threshold processing, a frame of 1 horizontal pixel×50 vertical pixels is set.

Next, the binary image 1500 is subjected to noise removal and dilation processing to generate a binary image 1510 . In the noise removal, white pixel clusters of 1 pixel or less in width, 4 pixels or less in height, or 49 pixels or less in area are rewritten to black pixels. As for the dilation process, each block of white pixels is dilated by 2 pixels vertically and horizontally.

Next, in S450, flaws are identified. For example, white pixels in the binary image 1410 shown in FIG. 14 or the binary image 1510 shown in FIG. 15 are identified as defects such as scratches. In order to make the user perceive it, it is also possible to enclose this white pixel group in a rectangle and display it on a binary image, or combine it on the captured image according to the position of the rectangle and display it.

When an image such as the background image 200 is not displayed, and the image processing method according to the present invention is performed by imaging the transparent inspection object WK as an imaged image, the inspection object WK is affected by halation and scratches. Could not be detected. When the image was captured with the background image 200 displayed, there was no effect of halation, so it was also found that displaying the background image 200 has the effect of suppressing halation.

Next, an example will be described in which the learning model shown in Modification 1 above is generated and defects such as scratches are specified using the learning model. An image 1600 shown in FIG. 16A is obtained by inputting a grayscale captured image of a transparent inspection object including defects such as scratches to a learning model constructed by using grayscale captured images as training data. , which outputs a picked-up image showing the defect of the inspection object.

An image 1610 shown in FIG. 16B is obtained by inputting a difference image to a learning model constructed by using a difference image based on a shift image as teacher data, and outputting a captured image showing a defect of an inspection object. . Comparing the image 1600 and the image 1610, it was found that the learning model that created the image 1600 was able to extract the defect of the inspection object better.

Although the image analysis software HALCON manufactured by MVTec (Germany) was used to construct the learning model, it is also possible to construct the learning model using other image analysis software.

REFERENCE SIGNS LIST 10 imaging device 20 image processing device 30 light source device 100 image processing system 200 background image 300 captured image 500 grayscale captured image 510 right-shifted image 520 left-shifted image 600 inverted image 700 differential image 710 differential image with noise 900 binary image 1000 Grayscaled captured image 1200 Difference image 1300 Difference image 1400 Binary image 1410 Binary image 1500 Binary image 1510 Binary image

Claims (2)

a step of projecting light consisting of a bright portion and a dark portion projected from a light source onto a transparent object to generate a captured image of the transparent object;
a step of generating a first image obtained by inverting the luminance value of each pixel of the captured image, or a first image obtained by deleting the area corresponding to the bright portion or the dark portion in the captured image from the edge of the captured image; and,
a step of generating a second image based on the absolute value of the difference in luminance values for pixels in a corresponding positional relationship in the first image and the captured image;
using the second image as teacher data to generate a learning model that outputs the defect of the transparent object in the second image when the second image is input;
A learning model generation method comprising: An image obtained by extracting a part of the second image as the training data;
The learning model generation method according to claim 1.