Method for image inspection, image inspection apparatus, and image inspection system

Abstract

To allow preparation of a dataset for machine learning of an image recognition model used for determining acceptability of an inspection target, without being restricted by actually captured images.SOLUTION: In a method for image inspection, in step S11, a plurality of annular images of an annular inspection target under different illumination conditions are acquired through multiple channels. In step S12, the annular image is converted into a rectangular image. In step S13, a plurality of cropped images, obtained by shifting cropping positions at equal intervals, are cut out from the rectangular image with an image size of a unit image, and a cropped image group is generated. In step S14, from the cropped image group, a selected image corresponding to each cropping position is selected. In step S15, using the selected image and the image recognition model, determination of acceptability of the inspection target is performed. The image recognition model is one that has been trained through machine learning using a dataset including a plurality of generated images produced with an image size corresponding to the unit image.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to an image inspection method, an image inspection device, and an image inspection system. [Background technology]

[0002] Conventionally, visual inspections are performed at industrial manufacturing facilities or manufacturing sites to check whether manufactured products have defects. Visual inspections sometimes use images of the manufactured products. Furthermore, visual inspections using product images sometimes use devices that can detect defects using machine learning.

[0003] Patent Document 1 describes an appearance inspection system including an appearance inspection device and a machine learning device. The machine learning device of the appearance inspection system of Patent Document 1 includes a defect image input unit that inputs real defect images and artificial defect images, an image extension processing unit that extends the real defect images and artificial defect images input to the defect image input unit to generate teacher image data groups, and a machine learning unit that performs machine learning based on the teacher image data groups and outputs learned data groups. The appearance inspection device of the appearance inspection system also includes an image judgment unit that judges defects that occur in the inspection workpiece. The image judgment unit makes judgments based on the learned data groups generated by the machine learning device. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2023-028564 Summary of the Invention [Problem to be solved by the invention]

[0005] In Patent Document 1, a real defect image is an image that includes a defect. On the other hand, an artificial defect image is generated by superimposing a pseudo defect image on a non-defective product image. A non-defective product image is an image that does not include a defect. The real defect image and the non-defective product image are each selected from inspection images of an inspection workpiece that are actually captured by an imaging means.

[0006] That is, in Patent Document 1, the real defect images and artificial defect images used in machine learning are based on inspection images of inspection workpieces actually captured by an imaging means. The teacher image data group and the learned data group are restricted by these real defect images and artificial defect images. Therefore, the appearance inspection system described in Patent Document 1 cannot handle inspection objects with characteristics different from the characteristics of the inspection workpieces actually captured (for example, geometric characteristics such as size, optical characteristics such as color, and other characteristics that appear in the image).

[0007] In view of the above problems, an object of the present invention is to provide an image inspection method, an image inspection device, and an image inspection system that are capable of preparing a data set for machine learning of an image recognition model used to determine the quality of an object being inspected, without being restricted by images that have actually been captured. [Means for solving the problem]

[0008] To solve the above problem, an image inspection method as an example of an embodiment of the present invention includes: acquiring annular images of a circular object under different illumination conditions by switching among a plurality of light sources arranged in a circular pattern, converting each of the annular images of the plurality of channels into a rectangular image to generate the rectangular images of the plurality of channels; extracting from each of the rectangular images of the plurality of channels a plurality of cropped images with cropping positions shifted at equal intervals at an image size corresponding to a predetermined unit image, thereby generating a group of cropped images of the plurality of channels; selecting selected images from the group of cropped images of the plurality of channels corresponding to each of the cropped positions; and performing image recognition on the selected images to determine whether the inspection object is a non-defective product using an image recognition model that identifies whether the inspection object is a non-defective product. The image recognition model is obtained by machine learning using a dataset including a plurality of generated images generated at image sizes corresponding to the unit images.

[0009] As another example of an embodiment of the present invention, an image inspection device includes an annular image acquisition unit that captures an annular object to acquire annular images for multiple channels, a shape conversion unit that converts the annular image into a rectangular image, a cropping unit that crops multiple cropped images from the rectangular image, a selection unit that selects selected images from the multiple cropped images, and a determination unit that uses the selected images to determine whether the object is good or bad. The annular image acquisition unit has multiple light sources arranged in a ring, and acquires the annular images for multiple channels, in which the portions of the object illuminated by the light sources differ for each channel, by switching the light sources to be turned on. The shape conversion unit converts each of the annular images for multiple channels into a rectangular image to generate the rectangular images for multiple channels. The cropping unit generates a group of cropped images for multiple channels by cropping multiple cropped images from each of the rectangular images for multiple channels, with the cropping positions shifted at equal intervals based on a predetermined image size of a unit image. The selection unit selects, from the group of cut-out images of multiple channels, the cut-out image of the channel in which the average brightness of pixels at each of the cut-out positions is the highest as the selected image corresponding to the cut-out position. The determination unit performs the pass / fail determination using an image recognition model that identifies whether the inspection object is a non-defective product by performing image recognition on the selected image. The image recognition model is obtained by machine learning using a dataset including a plurality of generated images generated at image sizes corresponding to the unit images.

[0010] As another example of an embodiment of the present invention, an image inspection system includes an annular image acquisition unit that captures an image of an annular object to acquire annular images for multiple channels, a shape conversion unit that converts the annular images into rectangular images, a cropping unit that crops multiple cropped images from the rectangular images, a selection unit that selects selected images from the multiple cropped images, an image generation unit that generates generated images used for training an image recognition model, a dataset storage unit that stores a dataset including the generated images generated by the image generation unit, a learning unit that performs machine learning using the dataset stored in the dataset storage unit to generate the image recognition model, and a judgment unit that judges the quality of the inspection object using the selected images and the image recognition model. The annular image acquisition unit has multiple light sources arranged in a ring, and by switching the light sources to be turned on, acquires the annular images for multiple channels, where the portions of the inspection object illuminated by the light sources differ for each channel. The shape conversion unit converts the annular images for multiple channels into rectangular images, respectively, to generate the rectangular images for multiple channels. The cropping unit generates a group of cropped images for multiple channels by cropping a plurality of cropped images from each of the rectangular images for multiple channels, the cropping positions of which are shifted at equal intervals by an image size of a predetermined unit image. The selection unit selects, from the group of cropped images for multiple channels, the cropped image of the channel in which the average pixel brightness at each cropping position is highest as the selected image corresponding to the cropping position. The image generation unit generates, as the generated images, a plurality of images with different conditions for the presence or absence of defects and the presence or absence of predetermined markings on the surface of the inspection object, with image sizes corresponding to the unit images. The learning unit performs machine learning on the image recognition model so that the image recognition model can identify the presence or absence of defects and the presence or absence of predetermined markings on the surface of the inspection object by performing image recognition on the selected images. The determination unit determines the inspection object, which is identified as having no defects and having the marking as a non-defective product as a result of image recognition on the selected images using the image recognition model. [Effects of the Invention]

[0011] According to the present invention, it is possible to prepare a data set for machine learning of an image recognition model used to determine the quality of an object to be inspected, without being restricted by images actually captured. [Brief explanation of the drawings]

[0012] [Figure 1] 1 is a diagram schematically illustrating the configuration of an image inspection device that executes an image inspection method as one example of an embodiment of the present invention, and an image inspection system that includes the image inspection device; [Figure 2] 1 is a flow chart showing the steps of an image inspection method. [Figure 3] 1 is a diagram showing a plurality of light sources arranged in a ring and portions of an inspection object illuminated by each light source. [Figure 4] FIG. 10 is a diagram illustrating how the light sources to be turned on are switched sequentially. [Figure 5] FIG. 1 is a diagram illustrating a multi-channel annular image. [Figure 6] FIG. 10 is a diagram showing an example of a rectangular image converted from a ring image. [Figure 7] 10A and 10B are diagrams for explaining how to cut out an image from a rectangular image. [Figure 8] 10A and 10B are diagrams illustrating selection of a selected image from a group of cut-out images. [Figure 9] FIG. 10 is a diagram showing a plurality of examples of selected images. [Figure 10] A diagram explaining how machine learning of an image recognition model is performed using a dataset including generated images. [Figure 11] FIG. 10 is a diagram showing an example of a dataset. [Figure 12] FIG. 10 is a diagram showing a plurality of examples of generated images. [Figure 13] FIG. 10 is a diagram illustrating a pass / fail determination using an image recognition model. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the drawings, identical or corresponding parts are designated by the same reference numerals, and descriptions thereof will not be repeated. In the following description, terms indicating positions or directions, such as front, rear, left, right, top, and bottom, may be used, but these terms are used for convenience to facilitate understanding of the embodiments. Unless otherwise specified, these terms are not limited to the strict geometric meaning of front, rear, left, right, top, bottom, etc.

[0014] An image inspection device that executes an image inspection method as an example of an embodiment of the present invention, and an image inspection system that includes the image inspection device, will be described below with reference to the drawings. Fig. 1 is a diagram that schematically shows the configuration of an image inspection device 10 that executes an image inspection method as an example of an embodiment of the present invention, and an image inspection system 1 that includes the image inspection device 10.

[0015] The image inspection system 1 includes an image inspection device 10 and a model construction device 40. The image inspection device 10 is a device that inspects the quality of an annular inspection object 12 based on an image of the inspection object 12. The model construction device 40 is a device that constructs an image recognition model 49 used by the image inspection device 10.

[0016] The annular inspection object 12 is, for example, the inner or outer ring of a bearing. The image inspection device 10 can inspect the quality of the bearing, such as whether there are any defects such as scratches on the end face of the inner or outer ring of the bearing, or whether a predetermined marking is correctly applied to the end face.

[0017] The image inspection device 10 includes an annular image acquisition unit 20, an information processing unit 30, a display unit 31, and an operation unit 32. The annular image acquisition unit 20 includes an imaging control unit 22, a camera 24, and lighting 26. The information processing unit 30 includes an image processing unit 33 and a determination unit 34. The image processing unit 33 includes a shape conversion unit 37, a cutout unit 38, and a selection unit 39. The determination unit 34 includes an image recognition model 49.

[0018] The image inspection device 10 judges the quality of the inspection object 12 in the information processing unit 30 using the annular image acquired by capturing an image of the inspection object 12 using the annular image acquisition unit 20. The result of the quality judgment is displayed on the display unit 31. The display unit 31 is a unit that displays various information. The display unit 31 is, for example, an LCD (liquid crystal display). Furthermore, operations performed by the user of the image inspection device 10 on the image inspection device 10 are accepted by the operation unit 32. The operation unit 32 is an interface unit that accepts user operations and transmits them to the image inspection device 10 as electrical signals. The operation unit 32 is, for example, a keyboard, a mouse, a joystick, or a controller designed specifically for the image inspection device 10. Furthermore, a touch panel display in which the display unit 31 and the operation unit 32 are integrated may be used.

[0019] The annular image acquisition unit 20 acquires annular images of multiple channels by capturing an image of the annular inspection object 12. The annular image acquisition unit 20 acquires annular images of multiple channels by capturing an image of the inspection object 12 with the camera 24 while changing the illumination conditions of the inspection object 12 with the illumination 26. The imaging control unit 22 is a unit that controls the camera 24 and the illumination 26. When acquiring annular images of multiple channels, the imaging control unit 22 controls the camera 24 and the illumination 26 so that the illumination conditions of the inspection object 12 differ for each channel.

[0020] The information processing unit 30 is a unit that performs various types of information processing in the image inspection device 10, and for example, a computer is used as the information processing unit 30. The information processing unit 30 includes a processor such as a CPU and storage devices such as RAM and ROM. The processor executes a program stored in the storage device, thereby performing functions such as the image processing unit 33.

[0021] The shape conversion unit 37 of the information processing unit 30 converts each of the multi-channel annular images acquired by the annular image acquisition unit 20 into a rectangular image to generate a multi-channel rectangular image. For example, the shape conversion unit 37 recognizes the center and radius of the annular image by image recognition, and calculates the distance from the center and the angle relative to the center for each coordinate of the annular image. The rectangular image after conversion by the shape conversion unit 37 is generated so that, for example, the distance from the center of the annular image corresponds to the short side, and the angle from the center corresponds to the long side.

[0022] The cropping unit 38 crops a plurality of cropped images for each channel from each of the rectangular images for the plurality of channels. More specifically, the cropping unit 38 crops a plurality of cropped images from each of the rectangular images for the plurality of channels, with the cropping positions shifted at equal intervals and having an image size of a predetermined unit image. The cropping unit 38 then assembles the plurality of cropped cropped images into a cropped image group for each channel to generate a cropped image group for the plurality of channels.

[0023] The selection unit 39 selects selected images from the plurality of cut-out images. More specifically, the selection unit 39 selects selected images corresponding to each cut-out position from the cut-out images of the plurality of channels.

[0024] The determination unit 34 determines whether the inspection object 12 is good or bad using the selected image. More specifically, the determination unit 34 performs the determination using an image recognition model 49. This image recognition model 49 is an artificial intelligence model that identifies whether the inspection object 12 is a good product or not by performing image recognition on the selected image. More specifically, the image recognition model 49 is a model that has been trained by machine learning using a data set including a plurality of generated images generated at image sizes corresponding to the unit images of the extracted image.

[0025] The model construction device 40 is a device for constructing an image recognition model 49. Because high computing power is required to construct the image recognition model 49, it is preferable that a computer with high computing power that is prepared separately from the information processing unit 30 is used as the model construction device 40, but the information processing unit 30 and the model construction device 40 may also be the same device.

[0026] The model construction device 40 includes a parameter determination unit 42, an image generation unit 44, a dataset storage unit 46, and a learning unit 48. The parameter determination unit 42 determines parameters for performing image generation. The image generation unit 44 generates generated images used for training an image recognition model 49 based on the parameters determined by the parameter determination unit 42. The dataset storage unit 46 stores a dataset including the generated images generated by the image generation unit 44. The learning unit 48 performs machine learning using the dataset stored in the dataset storage unit 46 to construct the image recognition model 49.

[0027] More specifically, the image generation unit 44 generates a plurality of images as generated images, each of which has different conditions, such as the presence or absence of defects on the surface of the object to be inspected 12 and the presence or absence of a predetermined marking, with an image size corresponding to the unit image of the extracted image.

[0028] More specifically, the learning unit 48 constructs an image recognition model 49 that uses machine learning with a dataset including the generated images to identify the presence or absence of defects and predetermined markings on the surface of the object to be inspected 12 for selected images.

[0029] The determination of parameters by the parameter determination unit 42, the generation of generated images by the image generation unit 44, the storage of the dataset by the dataset storage unit 46, and machine learning by the learning unit 48 are performed in advance before the image inspection device 10 inspects the object to be inspected 12. In other words, the construction of the image recognition model 49 by the model construction device 40 is performed independently of the inspection by the image inspection device 10. Furthermore, because the generated images generated by the image generation unit 44 are used for the dataset, the model construction device 40 can construct the image recognition model 49 without using images actually captured of the object to be inspected 12.

[0030] Next, the outline of the procedure of the image inspection method will be described with reference to Fig. 2. Fig. 2 is a flow chart showing the procedure of the image inspection method. First, in step S11, annular images of multiple channels with different illumination conditions for the annular inspection object 12 are acquired. Next, in step S12, each of the annular images of the multiple channels is converted into a rectangular image, and multiple rectangular images of the channels are generated.

[0031] Next, in step S13, a group of cut-out images for multiple channels is generated by cutting out a plurality of cut-out images from each of the rectangular images for multiple channels, the cut-out positions of which are shifted at equal intervals and have an image size of a predetermined unit image.

[0032] Next, in step S14, selected images corresponding to the respective cutout positions are selected from the group of cutout images, and in step S15, the quality of the inspection object 12 is determined using the selected images and the image recognition model 49.

[0033] For the sake of explanation, steps S11, S12, S13, S14, and S15 have been described above as being executed in order, but these steps may be executed in parallel or in a different order as necessary. For example, while conversion to a rectangular image based on an annular image of one channel is being performed, an annular image of another channel may be acquired (steps S11 and S12 may be executed in parallel). Also, for example, cropping of cropped images may be performed in parallel for rectangular images of multiple channels, and each time cropping is performed at one cropping position, a selected image corresponding to that cropping position may be selected (steps S13 and S14 may be executed alternately for each cropping position).

[0034] Next, referring to FIGS. 3 and 4, the change in illumination conditions for the inspection object 12 in step S11 of FIG. 2 will be described. FIG. 3 is a diagram illustrating a plurality of light sources 26a-26h arranged in a ring and the portions of the inspection object 12 illuminated by each light source. FIG. 4 is a diagram illustrating the sequential switching of the lit light sources 26a-26h. The light source 26 in FIG. 3 is shown from a perspective looking up at the light source 26 from diagonally below. The light source 26 in FIG. 4 is shown from a perspective looking down at the light source 26 from diagonally above, with the outline of the housing of the light source 26 and the light sources 26a-26h hidden by the housing of the light source 26 indicated by dashed lines. Although the camera 24 is not shown in FIGS. 3 and 4, it is preferable that the camera 24 be positioned at the center of the ring-shaped light sources 26a-26h. For example, it is preferable that the camera 24 be positioned so that the central axis of the arrangement of the plurality of light sources 26a-26h as a ring coincides with the optical axis of the camera 24. The camera 24 may be, for example, an image sensor using a CCD (charge coupled device) or an image sensor using a CMOS (complementary metal oxide semiconductor).

[0035] In this embodiment, as shown in FIG. 3, the illuminator 26 of the annular image acquisition unit 20 has a shape that expands downward, and multiple light sources 26a-26h are arranged in a ring shape on its underside. The light sources 26a-26h are, for example, LED (light-emitting diode) illuminators. The multiple light sources 26a-26h are arranged around the entire circumference of the annular inspection object 12 and illuminate the inspection object 12 from multiple directions. It is preferable that the light sources 26a-26h are arranged so that when all of the light sources 26a-26h are turned on, the entire circumference of the inspection object 12 is uniformly illuminated. Furthermore, each of the multiple light sources 26a-26h can be individually turned on and off. In FIGS. 3 and 4, the illuminator 26 is an eight-channel ring illuminator having eight light sources 26a, 26b, 26c, 26d, 26e, 26f, 26g, and 26h. Note that the number of light sources is not necessarily limited to eight; it may be more or less than eight.

[0036] The inspection object 12 is placed below a plurality of light sources 26a-26h included in the annular image acquisition unit 20. When one of the light sources of the illumination 26 is turned on and the other light sources are turned off, an illumination condition occurs in which only one location of the inspection object 12 is illuminated. For example, as shown in FIG. 4, when only the light source 26a is turned on and the other light sources 26b-26h are turned off, an illumination condition occurs in which only the location 14a (FIG. 3) of the inspection object 12 corresponding to the light source 26a is illuminated. When the camera 24 of the annular image acquisition unit 20 captures an image of the inspection object 12 under this illumination condition, an annular image under an illumination condition in which only the location corresponding to the light source 26a (illuminated location 14a) is illuminated is acquired.

[0037] The annular image acquisition unit 20 changes the illumination conditions for the annular inspection object 12 by switching the light sources to be turned on one by one among the multiple light sources 26a-26h arranged in a ring. For example, the annular image acquisition unit 20 switches the light sources to be turned on in a predetermined lighting order (in a counterclockwise direction ccw as viewed from above in FIG. 4), and turns off the light sources other than the one light source to be turned on. The annular image acquisition unit 20 then uses the camera 24 to acquire an annular image under illumination conditions in which only the areas corresponding to the turned-on light sources are illuminated.

[0038] The annular image acquisition unit 20 then assigns the multiple annular images captured under different illumination conditions to individual "channels." For example, an annular image captured under illumination conditions in which the first illuminated portion 14a of the inspection object 12 is illuminated by the first light source 26a is designated as the annular image of the first channel. Similarly, an annular image captured under illumination conditions in which the second illuminated portion 14b of the inspection object 12 is illuminated by the second light source 26b is designated as the annular image of the second channel. The annular image acquisition unit 20 repeats this process, acquiring annular images for multiple channels under different illumination conditions until an annular image (annular image of the eighth channel) is acquired under illumination conditions in which the last light source (the eighth light source 26h) is turned on (illumination conditions in which the eighth illuminated portion 14h is illuminated). Because the annular image acquisition unit 20 in FIG. 3 has eight light sources 26a-26h, eight corresponding annular images for the eight channels are acquired. The annular image acquisition unit 20 treats illumination conditions in which adjacent portions of the inspection object 12 are illuminated by the light sources 26a-26h as adjacent channels. As shown in FIG. 3, the illuminated portion of each channel partially overlaps with the illuminated portion of the adjacent channel (e.g., the second illuminated portion 14b and the eighth illuminated portion 14h for the first illuminated portion 14a). In this way, if the illuminated portions of adjacent channels partially overlap, there are no gaps between the illuminated portions, and all portions of the annular inspection object 12 are illuminated by one of the light sources 26a-26h at least once.

[0039] Annular images of multiple channels are shown schematically in Fig. 5. Annular images 50a-50h of eight channels are shown in Fig. 5, from annular image 50a of the first channel to annular image 50h of the eighth channel. For the sake of explanation, the horizontal direction of annular images 50a-50h of each channel will be referred to as the X direction, and the vertical direction will be referred to as the Y direction.

[0040] In the annular images 50a-50h of each channel, illuminated areas 14a-14h of the object 12 under inspection that are illuminated by one of the light sources 26a-26h appear bright, while other areas appear dark. In FIG. 5, bright (high brightness) areas are shown in white, and dark (low brightness) areas are shown in a checkered pattern. As shown in FIG. 5, the positions of the bright areas (illuminated areas 14a-14h) in the X and Y directions of the annular images 50a-50h of each channel differ from one channel to another.

[0041] Once annular image acquisition unit 20 has acquired annular images 50a-50h of multiple channels in the above manner, shape conversion unit 37 then converts these annular images 50a-50h into rectangular images (step S12 in FIG. 2).

[0042] An example of a rectangular image converted from an annular image is shown in Figure 6. The shape conversion unit 37 recognizes the center and radius of the inspection object 12 shown in the annular image by image recognition. Then, the shape conversion unit 37 converts the inspection object 12 shown in the annular image into a rectangular image 16 using the recognized center and radius.

[0043] For each pixel in the annular image, if the distance from the center of the object under inspection 12 is r and the angle with respect to the center (for example, an angle parallel to the X direction is 0°) is θ, the position in the X direction (X coordinate) and the position in the Y direction (Y coordinate) of each pixel can be expressed as X = rcosθ and Y = rsinθ.

[0044] The shape conversion unit 37 performs image processing to convert the annular inspection object 12 shown in the annular image into a rectangular shape based on the correspondence relationship of X = rcosθ and Y = rsinθ, thereby generating a rectangular image 16. The rectangular image 16 is preferably generated so that one side corresponds to the distance r from the center of the inspection object 12 and the other side corresponds to the angle θ relative to the center. FIG. 6 shows a rectangular image 16 in which the distance r from the center of the inspection object 12 corresponds to the short side and the angle θ relative to the center corresponds to the long side. Note that when generating the rectangular image 16, the distance r from the center is preferably offset by the radius of the inspection object 12 recognized by image recognition. For example, if the annular inspection object 12 is the inner or outer ring of a bearing, the offsetting causes the thickness of the inner or outer ring of the bearing to correspond to the length of the short side of the rectangular image 16.

[0045] The shape conversion unit 37 performs the above shape conversion on each of the multi-channel annular images to generate multi-channel rectangular images 16. After the multi-channel rectangular images 16 have been generated, the cropping unit 38 then crops out cropped images from each of the multi-channel rectangular images 16 (step S13 in FIG. 2).

[0046] 7, the cutting out of cut-out images from rectangular image 16 will be described. Cut-out unit 38 cuts out a plurality of cut-out images 61, 62, 63, (...omitted...), 68, 69 from rectangular image 16 of each channel, with the cut-out positions shifted at equal intervals and with an image size of a predetermined unit image.

[0047] The image size of the unit image is, for example, a square with 224 pixels on each side (224×224 pixels). When the unit image is a square, the length of one side is the same as or longer than the short side of the rectangular image 16.

[0048] 7, one end (here, the left end) of the rectangular image 16 is set as the starting point of the cut-out position, and a range equivalent to the image size of the unit image is cut out as a first cut-out image 61. Next, the cut-out position is shifted a predetermined distance from the first cut-out image 61, and a range equivalent to the image size of the unit image is cut out as a second cut-out image 62. Here, the amount of shift of the cut-out position is preferably set to a size such that the first cut-out image 61 and the second cut-out image 62 partially overlap, so that no gap is created between the first cut-out image 61 and the second cut-out image 62 (between adjacent cut-out images). In FIG. 7, about half of the range of the first cut-out image 61 overlaps with the second cut-out image 62. In Figure 7, the first cut-out image 61 and the second cut-out image 62 are shown as if they are different sizes so that they can be distinguished from each other, but in reality the first cut-out image 61 and the second cut-out image 62 are unit images of the same size.

[0049] Similarly, the cropping position is shifted at equal intervals to crop out a third cropped image 63, (...omitted...), a quasi-terminal cropped image 68, and a terminal cropped image 69. The terminal cropped image 69 is a cropped image that includes the other end (here, the right end) of the rectangular image 16, and the quasi-terminal cropped image 68 is a cropped image whose cropping position is one before the terminal cropped image 69.

[0050] The cropping unit 38 then sets a collection of multiple cropped images 61, 62, 63, (...omitted...), 68, 69 cropped from the rectangular image 16 of one channel as a group of cropped images corresponding to that channel. By performing the above cropping process on the rectangular images 16 of all channels, the cropping unit 38 generates groups of cropped images for multiple channels. The group of cropped images is stored in, for example, a storage device of the information processing unit 30. Here, each cropped image included in the group of cropped images is stored together with information on the cropping position of that cropped image.

[0051] Because the rectangular image 16 is a result of converting the shape of an annular image captured of the annular inspection object 12, the extraction position of each extracted image included in the extracted image group corresponds to the angle θ from the center of the inspection object 12. Therefore, each extracted image included in the extracted image group is stored together with information indicating the range of the angle θ from the center of the inspection object 12 that the extracted image corresponds to. The extraction unit 38 extracts extracted images from one end of the rectangular image 16 to the other end while allowing adjacent extracted images to partially overlap, so that every portion of the rectangular image 16 is included in one of the extracted images at least once. Therefore, the extracted image group includes images of the entire circumference of the annular inspection object 12.

[0052] Once the groups of cut-out images for the multiple channels have been generated, the selection unit 39 then selects selected images corresponding to the respective cut-out positions from the groups of cut-out images for the multiple channels (step S14 in FIG. 2). The selection of selected images from the groups of cut-out images will be described with reference to FIG.

[0053] The selection unit 39 processes all of the cut-out image groups of multiple channels together. FIG. 8 shows cut-out image groups 60a-60h of eight channels, from the cut-out image group 60a of the first channel to the cut-out image group 60h of the eighth channel. For the sake of explanation, each cut-out image group is shown in FIG. 8 with the cut-out images belonging to that cut-out image group connected horizontally in the figure. Therefore, although the cut-out image groups 60a-60h shown in FIG. 8 are rectangular, they are each made up of multiple connected square cut-out images.

[0054] The selection unit 39 selects the cut-out image of the channel in which the average brightness of pixels at each cut-out position is the highest as the selected image corresponding to the cut-out position. As an example, a case will be described in which selected images corresponding to three cut-out positions, cut-out position A, cut-out position B, and cut-out position C shown in Fig. 8, are selected.

[0055] When selecting a selected image corresponding to cut-out position A, the selection unit 39 calculates the average brightness of pixels included in the cut-out image groups 60a-60h cut out at cut-out position A for all channels. As described above, each cut-out image included in a cut-out image group is stored together with information about the cut-out position of that cut-out image. Therefore, the selection unit 39 only needs to calculate the average brightness of the cut-out image whose cut-out position information is cut-out position A from each of the multiple cut-out image groups 60a-60h. If there are eight channels as shown in FIG. 8, the average brightness is calculated for the eight cut-out images.

[0056] The selection unit 39 calculates the average brightness (brightness per pixel) for each of the (eight) clipped images of each channel for which the average brightness is to be calculated, based on the number of pixels contained in the clipped images and the brightness of each pixel. In Fig. 8, at clipping position A, clipped image 71e of the fifth channel (a clipped image belonging to clipped image group 60e of the fifth channel) has the highest average brightness among the eight channels. Therefore, clipped image 71e of the fifth channel is selected as the selected image corresponding to clipping position A.

[0057] Each of the cropped images 60a-60h is generated based on an annular image under illumination conditions in which any part of the annular inspection object 12 is illuminated by the illumination 26, and therefore the cropped images corresponding to the part illuminated by the illumination 26 are bright (high brightness) and the parts not illuminated are dark (low brightness). Therefore, by selecting the cropped image of the channel with the highest average brightness as the selected image, the cropped image corresponding to the part of the inspection object 12 illuminated by the illumination 26 (any of the light sources 26a-26h of the illumination 26) becomes the selected image.

[0058] Here, the selection unit 39 may select not only the channel with the highest average pixel brightness but also the cropped images of the two adjacent channels as the selected image corresponding to the crop position. Depending on the crop position, the channel with the highest average pixel brightness may have a lower average brightness than other crop positions. For example, a crop position corresponding to the edge of the illuminated area of ​​the inspection object 12 may have a lower average brightness than a crop position directly below the light sources 26a-26h. In such a case, cropped images of channels with different lighting conditions from the channel with the highest average pixel brightness but close to the channel with the highest average pixel brightness are also selected as the selected image (cropped images of three adjacent channels are selected), thereby obtaining a selected image that comprehensively better represents the surface condition of the illuminated area.

[0059] For example, at cut-out position A in Fig. 8, not only cut-out image 71e of the fifth channel but also cut-out image 71d of the fourth channel and cut-out image 71f of the sixth channel, which are adjacent to the fifth channel, may be selected as selected images corresponding to cut-out position A. In Fig. 8, cut-out image 71d of the fourth channel, cut-out image 71e of the fifth channel, and cut-out image 71f of the sixth channel are collectively shown as selected images 71 corresponding to cut-out position A.

[0060] Similarly, the selection unit 39 selects, as selected images 72 corresponding to the cut-out position B, a cut-out image 72c of the third channel, a cut-out image 72b of the second channel adjacent to the third channel, and a cut-out image 72d of the fourth channel.

[0061] Furthermore, the selection unit 39 selects a cut-out image 73h of the eighth channel, a cut-out image 73g of the seventh channel adjacent to the eighth channel, and a cut-out image 73a of the first channel as selected images 73 corresponding to the cut-out position C. As shown in FIG. 3, the eighth light source 26h and the first light source 26a are adjacent to each other, and therefore the eighth illuminated area 14h and the first illuminated area 14a illuminated by these light sources are adjacent to each other. Therefore, the eighth channel and the first channel are adjacent channels.

[0062] The selection unit 39 selects selected images corresponding to all of the cutout positions in this manner. For example, if 100 cutout images are cut out from one rectangular image, 100 selected images, or 300 selected images including cutout images from adjacent channels, are selected. The set of selected images thus selected represents the state of the surface (end face) of the entire circumference of the annular inspection object 12.

[0063] Some examples of selected images are shown in Fig. 9. Selected images 77, 78, and 79 shown in Fig. 9 are examples of selected images obtained when the inspection object 12 is the inner or outer ring of a bearing.

[0064] The selected image 77 shown in FIG. 9 includes an inscription 77s to be applied to the end face of the bearing. The inscription 77s has a predetermined shape to represent product information about the bearing (e.g., the manufacturer's name, model number, serial number, etc.). The shape of the inscription 77s may be, for example, a letter, symbol, or specific graphic (such as a manufacturer's logo). If the inscription 77s is to be applied to the end face of the bearing, then in the subsequent pass / fail judgment, if none of the selected images includes the inscription 77s, the bearing being inspected will not be judged as a pass / fail product.

[0065] 9 includes a defect 78x (a large scratch) that has occurred on the end face of the bearing due to some kind of malfunction. If any of the selected images includes defect 78x during the subsequent pass / fail determination, the bearing being inspected will not be determined to be a pass product.

[0066] 9 includes grinding marks 79z, which are minute scratches caused by grinding the end face of the bearing. In the subsequent pass / fail judgment, even if the selected image 79 includes grinding marks 79z, the bearing being inspected will be judged as a non-defective product if there are no other problems.

[0067] Once the selected images corresponding to the respective cut-out positions have been selected, the judgment unit 34 then judges the quality of the inspection object 12 using the image recognition model 49 (step S15 in FIG. 2). The image recognition model 49 is the result of machine learning using a data set including a plurality of generated images generated at image sizes corresponding to the unit images used in cutting out the cut-out images from the rectangular image.

[0068] 10, machine learning of an image recognition model 49 will be described using a dataset including generated images. The image recognition model 49 is constructed by the model construction device 40 of FIG.

[0069] First, the parameter determination unit 42 of the model construction device 40 determines generation parameters 42a to be given to the image generation unit 44. The generation parameters 42a are parameters that specify the content of the generated image generated by the image generation unit 44. The generation parameters 42a may be determined automatically in accordance with predetermined conditions as inspection conditions for the inspection object 12, or may be determined based on instructions from the user.

[0070] The generation parameters 42a may be a parameter group, with each group consisting of multiple parameters. The number of sets of the generation parameters 42a prepared corresponds to the number of generated images generated by the image generation unit 44. For example, the same number of sets of generation parameters 42a as the number of generated images may be prepared. Alternatively, a number of sets of generation parameters 42a less than the number of generated images may be prepared, with the remaining generation parameters 42a being automatically calculated based on the prepared generation parameters 42a. Alternatively, the image generation unit 44 may be capable of generating multiple generated images from the same generation parameters 42a.

[0071] The generation parameters 42a include, for example, parameters related to markings / defects / grinding marks. The parameters related to markings / defects / grinding marks specify, for example, the presence or absence of markings / defects / grinding marks in the generated image, the shape of the markings / defects / grinding marks, and the position of the markings / defects / grinding marks. Furthermore, it is preferable that a plurality of parameters related to a plurality of types of markings, each with different manufacturer names, model numbers, serial numbers, etc., are prepared as parameters related to markings.

[0072] The generation parameters 42a also include, for example, parameters related to a light source that illuminates the inspection object 12. The parameters related to the light source specify, for example, the relative position of the light source with respect to the inspection object 12, or the distance and angle of the light source. By specifying the parameters related to the light source, generated images corresponding to the inspection object 12 captured under various lighting conditions can be obtained. Note that the parameters related to the light source are preferably determined as a set of parameters related to one light source and three light sources, which are a group of two light sources adjacent to that light source (for example, parameters corresponding to the lighting conditions of the first channel, second channel, and third channel).

[0073] The generation parameters 42a also include, for example, parameters related to the texture to be superimposed on the virtual inspection object 12. The texture parameters specify, for example, the texture pattern, such as a grid pattern or a striped pattern, and the color of the texture. When a generated image is generated by superimposing the texture on the virtual inspection object 12, noise is added to the generated image. By using the generated image with noise added for machine learning, the performance of the image recognition model 49 is expected to improve.

[0074] Once the generation parameters 42a are determined, the image generation unit 44 generates an image using computer graphics technology based on the generation parameters 42a. The image generation unit 44 may generate a generated image by, for example, using random noise as an initial value and converting the random noise based on the generation parameters 42a. Alternatively, the image generation unit 44 may generate a generated image by drawing markings, defects, grinding marks, etc. based on the conditions specified by the generation parameters 42a.

[0075] The image generating unit 44 generates a plurality of images 45 that differ in the conditions of the presence or absence of defects and the presence or absence of a predetermined marking on the surface of the inspection object 12. At this time, the generated images 45 are generated with an image size corresponding to the unit image when the cut-out image is cut out from the rectangular image.

[0076] Specifically, the generated image 45 is generated to have the same size as the image size of the unit image or a size larger than the image size of the unit image. For example, if the image size of the unit image is a square of 224 pixels (224 × 224 pixels), the image size of the generated image 45 may be a square of 256 pixels (256 × 256 pixels).

[0077] A data set 46a is generated by assigning a tag 45t to each of the multiple generated images 45 generated by the image generating unit 44. The data set 46a is stored in the data set storage unit 46.

[0078] The tag 45t is information (label) that indicates how the generated image 45 should be recognized if the generated image 45 is input to the image recognition model 49. For example, the tag 45t may include an item regarding whether or not the generated image 45 includes an imprint or a defect.

[0079] Tag 45t may be assigned automatically based on generation parameters 42a, for example, or in response to a user instruction. Alternatively, model construction device 40 may recognize characteristics of generated image 45 by performing image recognition processing on generated image 45 using a means other than image recognition model 49, and tag 45t may be assigned based on the characteristics recognized by the image recognition processing.

[0080] An example of the data set 46a is shown in FIG. 11. The tag 45t includes items related to the presence or absence of markings and the presence or absence of defects. Note that the tag 45t may also include items other than these. The tag 45t assigned to each generated image 45 includes an item related to whether or not the generated image 45 includes a marking (presence or absence of markings) and an item related to whether or not the generated image 45 includes a defect (presence or absence of defects). For the sake of explanation, FIG. 11 shows multiple generated images 45 in a table format, categorized according to the combination of "presence" or "absence" of markings and "presence" or "absence" of defects.

[0081] Here, a set of generated images 45 in which the tag 45t shows "present" marking and "absent" defects is referred to as a marking image sample 81. A set of generated images 45 in which the marking is "absent" and the defect is "present" is referred to as a defect image sample 82. A set of generated images 45 in which the marking is "absent" and the defect is "absent" is referred to as a plain sample 83.

[0082] 12 shows some examples of generated images included in the data set 46a. The imprint image sample 81 includes a first imprint image 81a, a second imprint image 81b, ... (omitted) ..., an nth imprint image 81n. The defect image sample 82 includes a first defect image 82a, a second defect image 82b, ... (omitted) ..., an nth defect image 82n. The plane sample 83 includes a first plane image 83a, a second plane image 83b, ... (omitted) ..., an nth plane image 83n.

[0083] The first imprint image 81a, the second imprint image 81b, ... (omitted) ..., the nth imprint image 81n contain an imprint 81s. The first imprint image 81a, the second imprint image 81b, ... (omitted) ..., the nth imprint image 81n each have different conditions for the position, shape, etc. of the imprint 81s. By using imprint image samples 81 with various different conditions for the imprint 81s in this way for machine learning, the image recognition model 49 can correctly recognize imprints that appear under various conditions in the selected image.

[0084] First defect image 82a, second defect image 82b, ... (omitted) ..., nth defect image 82n contain defect 82x. Conditions such as the position and shape of defect 82x differ in each of first defect image 82a, second defect image 82b, ... (omitted) ..., nth defect image 82n. By using defect image samples 82 with various different conditions related to defect 82x for machine learning in this way, image recognition model 49 can correctly recognize defects that appear under various conditions in the selected images.

[0085] Furthermore, when preparing defect image sample 82 by capturing images of a product (such as a bearing) that is actually manufactured and serves as inspection target 12, as in the prior art, it may be difficult to collect a sufficient number of images containing defects. For example, an image containing a defect on the end face of the inner or outer ring of a bearing may be obtained at most once per day. However, in this embodiment, defect image sample 82 is prepared from generated images 45 generated by image generation unit 44 using computer graphics technology. This makes it easy to obtain defect image sample 82 containing a large number of images with various different conditions related to defect 82x. In other words, in this embodiment, it is possible to prepare data set 46a for machine learning without being restricted by images that have actually been captured.

[0086] In the first plain image 83a, a striped texture 83v is superimposed on an image that does not contain any markings or defects. In the second plain image 83b, a checkered texture 83w is superimposed on an image that does not contain any markings or defects. By using images with such noise-causing striped texture 83v or checkered texture 83w superimposed for machine learning, the image recognition model 49 comes to consider minute variations in color and shape appearing in the selected image to be within an acceptable range, enabling correct identification without being affected by minute variations in color and shape, which is expected to improve the performance of the image recognition model 49. In addition, the nth plain image 83n includes grinding marks 83z. By using data in which an image containing such fine scratches, such as grinding marks 83z, is tagged with tags 45t indicating "no marking" and "no defect," for machine learning, the image recognition model 49 no longer considers scratches as small as the grinding marks 83z to be defects, enabling correct identification, which is expected to improve the performance of the image recognition model 49.

[0087] The learning unit 48 uses the dataset 46a to perform machine learning on the correspondence between the generated image 45 and the tag 45t, thereby constructing an image recognition model 49. More specifically, with the generated image 45 as input and the tag 45t as output, machine learning is performed on the input-output relationship between the input and output. Once machine learning is performed with an appropriate dataset 46a, the image recognition model 49 becomes able to input any image not included in the dataset 46a and return an output on the item of the tag 45t (for example, the presence or absence of an imprint or defect).

[0088] Furthermore, if the generated image 45 is generated at a size larger than the image size of the unit image of the cropped image, the learning unit 48 crops a portion of the generated image 45 corresponding to the image size of the unit image and uses it as input for learning. For example, if the unit image of the cropped image is 224 x 224 pixels and the generated image 45 is 256 x 256 pixels, the learning unit 48 crops a 224 x 224 pixel portion of the 256 x 256 pixels of the generated image 45 while randomly changing the position, and uses it as input for learning. Even if the image size of the unit image of the cropped image is not 224 x 224 pixels, as long as it is smaller than 256 x 256 pixels, the learning unit 48 can simply crop it from 256 x 256 pixels to match the image size of the unit image of the cropped image. In this way, the same dataset 46a can be used to build multiple image recognition models 49 with different input image sizes (here, the image size of the unit image of the cropped image). In addition, by using a portion cut out from a large generated image 45 as input, the image recognition model 49 is prevented from over-fitting to a small portion of the data in the dataset 46a (over-learning), and the discrimination performance of the image recognition model 49 for any image not included in the dataset 46a is improved.

[0089] Furthermore, if the dataset 46a includes generated images 45 generated from generation parameters 42a, which are a set of parameters related to three light sources, that is, one light source and two light sources adjacent to that light source, the learning unit 48 may perform machine learning on the generated images 45 (generated images of three adjacent channels) generated from that set of generation parameters 42a as one set. When machine learning is performed on the generated images of three adjacent channels as one set, the image recognition model 49 can perform more accurate recognition when cropped images of the three adjacent channels are selected as selected images.

[0090] The quality of the inspection object 12 is determined using the image recognition model 49 constructed as described above. With reference to FIG. 13, the quality determination using the image recognition model 49 will be described. In the quality determination, selected images corresponding to all the cropping positions are input to the image recognition model 49. In FIG. 13, the selected images corresponding to all the cropping positions are collectively shown as a selected image group 70. The selected image group 70 includes selected image 71 corresponding to cropping position A, selected image 72 corresponding to cropping position B, and selected image 73 corresponding to cropping position C shown in FIG. 8, as well as all selected images corresponding to the other cropping positions.

[0091] The image recognition model 49 performs image recognition on the selected images to identify whether the inspection target 12 is a good product (good product or defective product). More specifically, for each of the selected images included in the selected image group 70, it is identified whether the selected image contains an inscription and whether the selected image contains a defect.

[0092] 13 shows in table form the classification results 91 for each of the selected images included in the selected image group 70. In the classification results 91 in Fig. 13, the selected image group 70 includes selected images with a mark "present" and a defect "absent," selected images with a mark "absent" and a defect "present," and selected images with a mark "absent" and a defect "absent."

[0093] For example, the determining unit 34 performs image recognition using the image recognition model 49, and determines that the inspection object 12 that is identified as having no defects and having an inscription is a non-defective product.

[0094] If the number of markings on one inspection object 12 is predetermined, the inspection object 12 is deemed defective if the number of markings does not match the predetermined number (for example, one). The image recognition model 49 may consider markings included in multiple selected images whose cutout positions are adjacent to each other as one marking, and markings included in multiple selected images whose cutout positions are spaced apart as multiple markings. For example, if no markings are included in the selected images at any cutout position (if there are zero markings), or if there are too many markings (for example, two), the judgment unit 34 will judge the inspection object 12 to be defective.

[0095] Furthermore, if there is even one defect, the inspection object 12 is deemed to be defective. For example, in the identification result 91 shown in Fig. 13, there is a selected image that indicates there is a defect, so the judgment unit 34 judges the inspection object 12 to be defective. However, in the case of an inspection condition in which the number of defects is not zero but is deemed to be good if it is small (for example, three or less), the allowable number of defects is predetermined, and the judgment unit 34 judges the inspection object 12 to be good if the number of defects does not exceed that allowable number.

[0096] The determination result 92 by the determination unit 34 may be displayed on the display unit 31. For example, a character string indicating the determination result 92, such as "good" or "defective," may be displayed on the display unit 31. Furthermore, if a defect is present in the inspection object 12, the location of the defect on the inspection object 12 may be displayed on the display unit 31. For example, the determination unit 34 calculates the position of a selected image including a defect on the circumferential direction of the inspection object 12 from which the selected image was extracted. Then, an image obtained by combining an annular image of the inspection object 12 with an indication indicating the location of the defect may be displayed on the display unit 31. Note that the annular image used for the combination may be, among annular images of multiple channels (e.g., eight channels), an annular image of a channel corresponding to a light source that illuminates an illuminated location closest to the location of the defect. Alternatively, the selected images included in the selected image group 70 may be connected and overlapping portions may be deleted to form an annular image for display.

[0097] The image recognition model 49 may receive the entire group of selected images 70 as input and output a pass / fail judgment regarding the inspection object 12. In this case, the judgment unit 34 outputs the image recognition model 49 as the judgment result 92.

[0098] According to this embodiment, machine learning of image recognition model 49 is performed using dataset 46a including generated images 45 generated by image generation unit 44, and therefore dataset 46a is prepared without being restricted by images that have actually been captured. For example, even if it is difficult to prepare images that include defects when preparing images by capturing images of actual products, image generation unit 44 can easily prepare generated images 45 that include defects.

[0099] Furthermore, the selected image input to the image recognition model 49 in the pass / fail judgment is selected from the cropped image, and the cropped image is obtained by transforming the annular image into a rectangular image and then cropping the rectangular image. Therefore, the generated image 45 included in the data set 46a used for training the image recognition model 49 may be an image (e.g., a square) corresponding to the cropped image. That is, the image generator 44 does not need to generate an annular image that matches the shape of the annular inspection object 12 as the image used for training the image recognition model 49. Therefore, according to this embodiment, the data set 46a and the image recognition model 49 can be prepared regardless of the actual size of the annular inspection object 12. Furthermore, the same image recognition model 49 can be used to handle inspection objects 12 of various sizes.

[0100] The learning unit 48 may be capable of re-learning the image recognition model 49 in response to changes in the inspection conditions. When re-learning is performed, the dataset 46a is updated in response to changes in the inspection conditions, and machine learning of the image recognition model 49 is performed using the updated dataset 46a. For example, if the number of types of markings applied to the inspection target object 12 increases, a generated image 45 including the new marking is added to the dataset 46a, and the learning unit 48 may re-learn the image recognition model 49 using the dataset 46a after the addition. [Explanation of symbols]

[0101] 1. Image inspection system 10. Image inspection equipment 12 Inspection object 16 Rectangular Images 20 Circular image acquisition unit 26 Lighting 26a 1st light source 34 Judgment section 37 Shape conversion section 38 Cutout part 39 Selection Department 45 Generated Images 46a Dataset 49 Image Recognition Model 50a Circular image of channel 1 61 First cutout image 70 Selected Images

Claims

1. acquiring a plurality of channel annular images of the annular inspection object under different illumination conditions by switching the light sources to be turned on among the plurality of light sources arranged in an annular shape; converting each of the circular images of the multiple channels into a rectangular image to generate the rectangular images of the multiple channels; generating a group of cut-out images for the plurality of channels by cutting out a plurality of cut-out images with a predetermined image size of a unit image and with cut-out positions shifted at equal intervals from each of the rectangular images for the plurality of channels; selecting selected images corresponding to the respective extraction positions from the group of extracted images of the plurality of channels; performing image recognition on the selected image to determine whether the inspection object is a non-defective product or not using an image recognition model; An image inspection method, wherein the image recognition model is obtained by machine learning using a dataset including a plurality of generated images generated at an image size corresponding to the unit image.

2. an annular image acquisition unit that captures an image of an annular inspection object and acquires annular images of multiple channels; a shape conversion unit that converts the annular image into a rectangular image; a cutout unit that cuts out a plurality of cutout images from the rectangular image; a selection unit that selects a selected image from the plurality of cut-out images; a determination unit that determines whether the inspection object is good or bad using the selected image; Equipped with the annular image acquisition unit has a plurality of light sources arranged in a ring, and acquires the annular images of the plurality of channels, in which the portions of the object to be inspected that are illuminated by the light sources differ for each channel, by switching the light sources to be turned on; the shape conversion unit converts the annular images of the plurality of channels into the rectangular images, respectively, to generate the rectangular images of the plurality of channels; the cropping unit generates a group of cropped images for the multiple channels by cropping a plurality of the cropped images from each of the rectangular images for the multiple channels, the cropped images being shifted at equal intervals in an image size of a predetermined unit image; the selection unit selects, from the group of cut-out images of a plurality of channels, the cut-out image of a channel in which an average luminance of pixels at each of the cut-out positions is highest, as the selected image corresponding to the cut-out position; the determination unit performs the pass / fail determination using an image recognition model that identifies whether the inspection object is a non-defective product by performing image recognition on the selected image, An image inspection device, wherein the image recognition model is one that has been machine-learned using a dataset including a plurality of generated images generated at image sizes corresponding to the unit images.

3. the annular image acquisition unit treats illumination conditions in which the portions of the inspection object illuminated by the light source are adjacent to each other as adjacent channels; 3. The image inspection device according to claim 2, wherein the selection unit also selects the cropped images of two channels adjacent to the channel having the highest average pixel brightness at each of the cropping positions as the selected images corresponding to the cropping positions.

4. the image recognition model performs image recognition on the selected image to identify the presence or absence of defects and the presence or absence of a predetermined marking on the surface of the inspection object; The image inspection device according to claim 2 , wherein the determining unit determines that the inspection object identified as having no defect and having the marking is a non-defective product.

5. an annular image acquisition unit that captures an image of an annular inspection object and acquires annular images of multiple channels; a shape conversion unit that converts the annular image into a rectangular image; a cutout unit that cuts out a plurality of cutout images from the rectangular image; a selection unit that selects a selected image from the plurality of cut-out images; an image generation unit that generates generated images used in training an image recognition model; a dataset storage unit that stores a dataset including the generated image generated by the image generation unit; a learning unit that performs machine learning using the dataset stored in the dataset storage unit to construct the image recognition model; a determination unit that determines whether the inspection object is good or bad using the selected image and the image recognition model; Equipped with the annular image acquisition unit has a plurality of light sources arranged in a ring, and acquires the annular images of the plurality of channels, in which the portions of the object to be inspected that are illuminated by the light sources differ for each channel, by switching the light sources to be turned on; the shape conversion unit converts the annular images of the plurality of channels into the rectangular images, respectively, to generate the rectangular images of the plurality of channels; the cropping unit generates a group of cropped images for the multiple channels by cropping a plurality of the cropped images from each of the rectangular images for the multiple channels, the cropped images being shifted at equal intervals in an image size of a predetermined unit image; the selection unit selects, from the group of cut-out images of a plurality of channels, the cut-out image of a channel in which an average luminance of pixels at each of the cut-out positions is highest, as the selected image corresponding to the cut-out position; the image generation unit generates, as the generated images, a plurality of images having different conditions regarding the presence or absence of defects on the surface of the inspection object and the presence or absence of a predetermined marking, with an image size corresponding to the unit image; the learning unit constructs the image recognition model that identifies the presence or absence of the defect and the presence or absence of the predetermined mark on the surface of the inspection object for the selected image by machine learning using the dataset including the generated image; The judgment unit performs image recognition on the selected image using the image recognition model, and judges the inspection object that is identified as having no defect and having the marking to be a good product.

Images