Inspection device and determination model generation method

The inspection device improves defect detection accuracy by generating pseudo-defect images based on the state and position of normal and pseudo-defect elements, using machine learning to create a determination model that effectively addresses the limitations of existing technologies.

WO2024247380A1PCT designated stage expired Publication Date: 2024-12-05ASTEMO LTD
View PDF 13 Cites 0 Cited by

Patent Information

Application Number
PCT/JP2024/004462
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-05-30
Filing Date
2024-02-09
Publication Date
2024-12-05

AI Technical Summary

Technical Problem

The accuracy of determining the quality of inspected objects in existing inspection devices deteriorates due to randomly selected or user-input-based generation parameters, which do not consider the state or position of non-defective and defective images.

Method used

An inspection device with a pseudo-defect target image generation unit that combines pseudo-defect element images with normal target images based on the position and state of both, using machine learning to create a determination model for improved defect detection accuracy.

Benefits of technology

The solution generates pseudo-defect target images that closely resemble actual defects, enhancing the accuracy of inspection judgments by considering the state and position parameters of both images.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2024004462_05122024_PF_FP_ABST
    Figure JP2024004462_05122024_PF_FP_ABST
Patent Text Reader

Abstract

An inspection device according to the present invention comprises: a pseudo defect subject image generation unit that generates a pseudo defect subject image by combining a normal subject image with a pseudo defect element image in which a defective part is simulated, that, in accordance with a combining position of the normal subject image at which the pseudo defect element image is combined, a state at the combining position of the normal subject image, and a state of the pseudo defect element image, changes at least one of the state of the pseudo defect element image, the combining position of the normal subject image, and the state at the combining position of the normal subject image, and that combines the pseudo defect element image at the combining position of the normal subject image; a training processing unit that prepares a determination model by machine learning using the pseudo defect subject image; and a determination processing unit that uses the determination model to detect a defect in a subject image obtained by imaging a subject to be inspected.
Need to check novelty before this filing date? Find Prior Art

Description

Inspection device and decision model generation method

[0001] The present invention relates to an inspection device and a decision model generation method.

[0002] Patent Document 1 discloses an inspection device equipped with a learning data generation unit that generates learning data, which is image data of a learning image in which a good product image is combined with a defective product image, based on defective part data, good product data, and generation parameters to determine whether an object to be inspected is good or bad.

[0003] Japanese Patent Application Laid-Open No. 2020-27424

[0004] However, in the inspection device of Patent Document 1, the generation parameters are generated by random selection or generation according to a predetermined rule, or by user input, and therefore the state of the image of a non-defective product or an image of a defective part is not taken into consideration, which poses a problem that the accuracy of the pass / fail judgment of the inspected object may be reduced. One of the objects of the present invention is to provide an inspection device and a judgment model generation method that improves the accuracy of inspection judgment.

[0005] An inspection device in one embodiment of the present invention has a pseudo-defect object image generation unit that generates a pseudo-defect object image by combining a pseudo-defect element image that simulates a defect portion with a normal object image, and the pseudo-defect object image generation unit changes at least one of the state of the pseudo-defect element image, the combination position of the normal object image, and the state of the combination position of the normal object image depending on the combination position of the normal object image at which the pseudo-defect element image is combined, the state of the combination position of the normal object image, and the state of the pseudo-defect element image, thereby combining the pseudo-defect element image with the combination position of the normal object image; a learning processing unit that uses machine learning to train the pseudo-defect object image to create a judgment model; and a judgment processing unit that uses the judgment model to detect defects in an object image photographed from an object to be inspected.

[0006] According to one embodiment of the present invention, it is possible to improve the accuracy of test determination.

[0007] 1 is an overall view of an inspection apparatus according to a first embodiment. FIG. 2 is a block diagram illustrating the inside of a computer according to the first embodiment. FIG. 3 is a flowchart illustrating the operation of a pseudo defect object image generation unit according to the first embodiment. FIG. 4 is a diagram illustrating state parameters of a synthesis position of a normal object image according to the first embodiment. FIG. 5 is a diagram illustrating state parameters of a pseudo defect element image according to the first embodiment. FIG. 6 is a graph illustrating a relationship between the brightness of a synthesis position of a normal object image and the optimum brightness of a pseudo defect element image according to the first embodiment. FIG. 7 is a graph illustrating a brightness difference between the brightness of a synthesis position of a normal object image and the optimum brightness of a pseudo defect element image according to the first embodiment. FIG. 8 is a diagram illustrating the operation of a state parameter processing unit according to a state parameter history of a state parameter history storage unit according to the first embodiment. FIG. 9 is a diagram illustrating the operation of a state parameter processing unit using a feature point according to the first embodiment. FIG. 10 is a diagram illustrating the operation of a state parameter processing unit using a boundary specification mask image to be set according to the first embodiment. FIG. 11 is a diagram illustrating state parameters of a synthesis position of a normal object image according to the second embodiment. FIG. 12 is a diagram illustrating a change in angle at a synthesis position at a region boundary of a pseudo defect element image according to the second embodiment. FIG. 13 is a diagram illustrating a filter process of a synthesis position of a normal object image according to the third embodiment.

[0008] First Embodiment FIG. 1 is an overall view of an inspection device according to a first embodiment.

[0009] The inspection device 1 of embodiment 1 includes a camera 2, a robot 3, a computer 4, and a judgment result display unit 6. The camera 2 captures an image of the crown surface 5a of a piston (inspection object) 5. The crown surface 5a of the piston 5 has a machined surface (surface type and condition) with cutting marks and a cast surface (surface type and condition) whose surface properties are significantly different. The robot 3 changes the relative orientation of the piston 5 with respect to the camera 2. The computer 4 is, for example, a personal computer. The judgment result display unit 6 displays the judgment result obtained by determining defects in the target image of the crown surface 5a of the piston 5 captured by the camera 2.

[0010] FIG. 2 is a block diagram showing the inside of the computer according to the first embodiment.

[0011] (Configuration of Computer) The computer 4 includes a normal object image storage unit 401 that stores a plurality of normal object images captured by the camera 2, a normal object image synthesis position state parameter detection processing unit 402, a state parameter processing unit 403, a state parameter history storage unit 404, a pseudo defect element image generation processing unit 405, a pseudo defect element image storage unit 406, a pseudo defect element image state parameter detection processing unit 407, a boundary designation mask image setting and storage unit 408, a synthesis processing unit 409, a pseudo defect object image storage unit 410, an annotation data storage unit 411, a learning processing unit 412, a judgment model storage unit 413, and a judgment processing unit 414 that performs a pass / fail judgment to detect defects in the object image of the crown surface 5 a of the piston 5 captured by the camera 2 (judgment step) and transmits the judgment result to the judgment result display unit 6. The pseudo defect object image generating unit 4a includes a normal object image synthesis position state parameter detection processing unit 402, a state parameter processing unit 403, a state parameter history storage unit 404, a pseudo defect element image generation processing unit 405, a pseudo defect element image storage unit 406, a pseudo defect element image state parameter detection processing unit 407, a boundary designation mask image setting and storage unit 408, a synthesis processing unit 409, a pseudo defect object image storage unit 410, and an annotation data storage unit 411. A learning processing unit 412 performs machine learning using the coordinate data of the pseudo defect object image stored in the annotation data storage unit 411 and the pseudo defect object image stored in the pseudo defect object image storage unit 410, and stores the machine learning results in a judgment model storage unit 413 (judgment model creation step). This machine learning is, for example, learning using a neural network, and in the first embodiment, deep learning, which is a multilayered version of a neural network, is adopted.

[0012] (Operation of Pseudo Defect Image Generating Unit) FIG. 3 is a flowchart for explaining the operation (combining step) of the pseudo defect object image generating unit in the first embodiment.

[0013] In step S1, a normal target image captured by the camera 2 is acquired and stored in the normal target image storage unit 401. In step S2, a boundary-specified mask image is set and stored in the boundary-specified mask image setting / storage unit 408. In step S3, the state parameter history stored in the state parameter history storage unit 404 is processed. In step S4, it is confirmed whether there are multiple normal target images stored in the normal target image storage unit 401. If there are not multiple normal target images, the process proceeds to step S5. If there are multiple normal target images, the process proceeds to step S12. In step S5, a pseudo-defect element image simulating a defect portion generated by the pseudo-defect element image generation processing unit 405 and stored in the pseudo-defect element image storage unit 406 is acquired, and the pseudo-defect element image state parameter detection processing unit 407 detects the state parameters (states) of the pseudo-defect element image and transmits them to the state parameter processing unit 403. In step S6, the normal target image stored in the normal target image storage unit 401 is acquired, and it is confirmed whether there are any candidate synthesis positions. If there is a candidate for the combining position, the process proceeds to step S7; if there is no candidate for the combining position, the process returns to step S1. In step S7, the normal target image combining position state parameter detection processing unit 402 detects the state parameters (state) of the combining position of the normal target image and transmits them to the state parameter processing unit 403. In step S8, it is confirmed whether combining is possible based on the state parameters of the combining position of the normal target image. If combining is possible, the process proceeds to step S9; if combining is not possible, the process returns to step S1. In step S9, the state parameter processing unit 403 changes the state parameters of the combining position of the normal target image or the state parameters of the pseudo defect element image according to the state parameters of the combining position of the normal target image and the boundary position designation mask image stored in the boundary designation mask image setting and storage unit 408. The state parameter processing unit 403 transmits the combining position of the normal target image, the state parameters of the combining position of the normal target image, and the changed state parameters of the pseudo defect element image to the state parameter history storage unit 404 for storage. In step S10, the synthesis processing unit 409 synthesizes the pseudo-defect element image at the synthesis position of the normal object image to generate a pseudo-defect object image.In step 11, the generated pseudo defect object image is stored in the pseudo defect object image storage unit 410, and the coordinate data of the generated pseudo defect object image is stored in the annotation data storage unit 411.

[0014] In step S12, pseudo defect element images corresponding to multiple normal target images simulating defect portions, which have been generated by the pseudo defect element image generation processing unit 405 and stored in the pseudo defect element image storage unit 406, are obtained. The pseudo defect element image state parameter detection processing unit 407 detects state parameters of the pseudo defect element images and transmits them to the state parameter processing unit 403. In step S13, multiple normal target images stored in the normal target image storage unit 401 are obtained and checked for the presence or absence of candidate synthesis positions. If candidate synthesis positions are present, the process proceeds to step S14. If candidate synthesis positions are not present, the process returns to step S1. In step S14, locations and shapes that are common feature points of the multiple normal target images are extracted. In step S15, the overall positions of the multiple normal target images are corrected based on the locations of the feature points and shapes. In step S16, the normal target image synthesis position state parameter detection processing unit 402 detects state parameters of the synthesis positions of the multiple normal target images and transmits them to the state parameter processing unit 403. In step S17, it is confirmed whether or not compositing is possible based on the state parameters of the compositing positions of the normal target images. If compositing is possible, the process proceeds to step S18; if compositing is not possible, the process returns to step S1. In step S18, the state parameter processing unit 403 changes the state parameters of the compositing positions of each normal target image or the state parameters of the pseudo defect element images according to the state parameters of the compositing positions of each normal target image and the boundary position designation mask image stored in the boundary designation mask image setting and storage unit 408. In step S19, the compositing processing unit 409 performs a compositing process on the pseudo defect element images at the compositing positions of the normal target images to generate pseudo defect object images. In step S20, the generated pseudo defect object images are stored in the pseudo defect object image storage unit 410, and the coordinate data of the generated pseudo defect object images are stored in the annotation data storage unit 411.In this way, the state parameters of the synthesis position of the normal target image are detected, and the state parameters of the synthesis position of the normal target image or the state parameters of the pseudo-defect element image are changed according to the state parameters of the synthesis position of the normal target image, and the pseudo-defect element image is synthesized at the synthesis position of the normal target image to generate a pseudo-defect object image.Therefore, it is possible to generate a pseudo-defect object image that is close to an actual defect, and the accuracy of inspection judgment can be improved.

[0015] FIG. 4 is a diagram illustrating state parameters of the synthesis position of the normal target image in the first embodiment.

[0016] In the first embodiment, the luminance (background luminance) ka of the compositing position G of the normal target image 7 is used as a state parameter (state) of the compositing position of the normal target image by the normal target image compositing position state detection processing unit 402. The detected compositing position G and the luminance ka of the compositing position G are transmitted to the state parameter processing unit 403. Note that, instead of the luminance ka, a luminance variation Δk may be used.

[0017] FIG. 5 is a diagram illustrating state parameters of a pseudo defect element image in the first embodiment.

[0018] The pseudo defect element image state parameter detection processing unit 407 can use one or a combination of the following as the state parameter (state) of the pseudo defect element image 8: type (hole, scratch), major axis a (size), minor axis b (size), aspect ratio (= major axis a: minor axis b), angle θ1 of the major axis axis α, and brightness kb. In the first embodiment, brightness kb is used as the state parameter of the pseudo defect element image 8.

[0019] FIG. 6 is a graph showing the relationship between the luminance of the synthesis position of the normal target image and the optimum luminance of the pseudo defect element image in the first embodiment.

[0020] A predetermined relationship is established by a regression equation between the brightness ka of the combining position G of the normal target image 7 and the optimal brightness (brightness after change) kc of the pseudo defect element image 8. Therefore, the brightness ka of the combining position G of the normal target image 7 is detected, and the optimal brightness kc of the pseudo defect element image 8 is calculated by the regression equation from the brightness ka of the detected combining position G of the normal target image 7. Then, if there is a brightness difference between the brightness kb of the detected actual pseudo defect element image 8 and the optimal brightness kc of the pseudo defect element image 8, the brightness kb of the actual pseudo defect element image 8 is changed to the optimal brightness kc of the pseudo defect element image 8, and the pseudo defect element image 8 with the changed brightness is combined at the combining position of the normal target image to generate a pseudo defect object image. This makes it possible to efficiently generate an image that is closer to an actual defect.

[0021] FIG. 7 is a graph showing the difference in brightness between the brightness at the synthesis position of the normal target image and the optimum brightness of the pseudo defect element image in the first embodiment.

[0022] The greater the luminance ka of the combining position G of the normal target image 7, the greater the optimal luminance kc of the pseudo defect element image 8. That is, the greater the luminance ka of the combining position G of the normal target image 7, the greater the luminance difference kd between the luminance ka of the combining position G of the normal target image 7 and the optimal luminance kc of the pseudo defect element image 8. As a result, the greater the luminance ka of the combining position G of the normal target image 7, the clearer the contrast with the luminance kb of the actual pseudo defect element image 8. However, by increasing the luminance kb of the actual pseudo defect element image 8 to the optimal luminance kc of the pseudo defect element image 8, the pseudo defect element image 8 does not become too clear, and an image closer to an actual defect can be efficiently generated.

[0023] Furthermore, when the brightness ka of the combining position G of the normal target image 7 is equal to or greater than the threshold value s, the pseudo defect element image 8 is combined, and when the brightness ka of the combining position G of the normal target image 7 is less than the threshold value s, the pseudo defect element image 8 is not combined. As a result, when the brightness ka of the combining position G of the normal target image 7 is less than the threshold value s, the brightness difference kd with the brightness kb of the pseudo defect element image 8 is small, making it impossible to perform an accurate inspection determination. This can be prevented, thereby improving the accuracy of the inspection determination.

[0024] FIG. 8 is a diagram illustrating the operation of the state parameter determination processing unit based on the state parameter history of the state parameter history storage unit in the first embodiment.

[0025] The state parameter processing unit 403 checks the past synthesis position of the machined surface 10 at the region boundary R1 between the machined surface 10 having the cutting marks and the casting surface 11 in the history of the synthesis position G of the normal target image 7 stored in the state parameter history storage unit 404 in the history target image 9, and identifies a priority synthesis area that has not been synthesized (step S3 in FIG. 3). This changes the synthesis position G of the normal target image 7 to within the priority synthesis area, and generates a pseudo-defect target image, thereby improving the comprehensiveness of the pseudo-defect element images 8. The priority synthesis area is not limited to the machined surface 10 within the region boundary R1, but can also be set to the casting surface 11 or the region boundary R1. Instead of the synthesis position G of the normal target image 7, the state parameters of the synthesis position G of the normal target image 7 or the history of the state parameters of the pseudo-defect element images 8 may be used.

[0026] FIG. 9 is a diagram illustrating the operation of the state parameter processing unit using feature points in the first embodiment.

[0027] When there are multiple normal target images, the state parameter processing unit 403 extracts a ring-shaped depression T (feature point) caused by the mold and corrects the overall positions of the multiple normal target images using the ring-shaped depression T (feature point) caused by the mold as a reference so that the multiple normal target images are in the same overall position (steps S14 and S15 in FIG. 3). This makes it possible to correct a state in which the overall positions of the multiple normal target images are slightly misaligned due to misalignment caused by photography, and therefore it is possible to reuse the boundary position designation mask image (boundary position designation mask) M stored in the boundary designation mask image setting and storage unit 408 (described later) for the multiple normal target images, making it possible to efficiently generate images that resemble actual defects.

[0028] FIG. 10 is a diagram illustrating the operation of the state parameter processing unit using the boundary specification mask image set in the first embodiment.

[0029] The state parameter processing unit 403 uses a boundary position designation mask image (boundary position designation mask) M having an area boundary R1 between the machined surface 10 and the casting surface 11 and an area boundary R2 indicating the outer periphery of the crown surface 5a of the piston 5, which are stored in the boundary designation mask image setting and storage unit 408, to composite the pseudo defect element image 8 at the composite position G of the normal target image 7 so that a part of the pseudo defect element image 8 overlaps the area boundary R1 (steps S9 and S18 in FIG. 3). In this way, the pseudo defect element image 8 can be arranged at the area boundary R1 where the judgment is difficult, so that the judgment accuracy of the inspection can be more efficiently improved.

[0030] Next, a description will be given of the effects of embodiment 1. Embodiment 1 provides the following effects.

[0031] (1) The brightness as a state parameter of the compositing position of the normal target image is detected, the brightness as a state parameter of the pseudo defect element image is changed according to the brightness as a state parameter of the compositing position of the normal target image, and the pseudo defect element image with the brightness as a state parameter changed is composited at the compositing position of the normal target image to generate a pseudo defect object image. Thus, a pseudo defect object image close to an actual defect can be generated, and the judgment accuracy of the inspection can be improved.

[0032] (2) When there is a difference in brightness between the brightness kb of the detected actual pseudo defect element image 8 and the optimal brightness kc of the pseudo defect element image 8 calculated by the regression equation, the brightness kb of the actual pseudo defect element image 8 is changed to the optimal brightness kc of the pseudo defect element image 8, and the pseudo defect element image 8 with the changed brightness is synthesized at the synthesis position of the normal target image to generate a pseudo defect object image. Thus, a pseudo defect object image closer to an actual defect can be efficiently generated.

[0033] (3) The greater the luminance ka of the combining position G of the normal target image 7, the greater the luminance difference kd between the luminance ka of the combining position G of the normal target image 7 and the optimum luminance kc of the pseudo defect element image 8. Therefore, the greater the luminance ka of the combining position G of the normal target image 7, the clearer the contrast with the luminance kb of the actual pseudo defect element image 8. However, by increasing the luminance kb of the actual pseudo defect element image 8 to the optimum luminance kc of the pseudo defect element image 8, the pseudo defect element image 8 does not become too clear, and a pseudo defect object image that is closer to an actual defect can be efficiently generated.

[0034] (4) When the brightness ka of the combining position G of the normal target image 7 is equal to or greater than the threshold value s, the pseudo defect element image 8 is combined, and when the brightness ka of the combining position G of the normal target image 7 is less than the threshold value s, the pseudo defect element image 8 is not combined. Therefore, when the brightness ka of the combining position G of the normal target image 7 is less than the threshold value s, the brightness difference kd with the brightness kb of the pseudo defect element image 8 is small, making it impossible to perform an accurate inspection determination. This can be prevented to improve the accuracy of the inspection determination.

[0035] (5) The state parameter processing unit 403 checks the past compositing position of the machined surface 10 at the region boundary R1 between the machined surface 10 and the casting surface 11 in the history object image 9 based on the history of the compositing position G of the normal object image 7 stored in the state parameter history storage unit 404, identifies the priority compositing area that has not been composited, changes the compositing position of the pseudo-defect element image 8 to within the priority compositing area, and generates a pseudo-defect object image. This improves the comprehensiveness of the pseudo-defect object images.

[0036] (6) When there are multiple normal target images, the state parameter processing unit 403 extracts a ring-shaped depression T (feature point) caused by the mold, and corrects the overall positions of the multiple normal target images so that the multiple normal target images are in the same overall position using the ring-shaped depression T (feature point) caused by the mold as a reference. Therefore, since it is possible to correct a state in which the overall positions of the multiple normal target images are slightly shifted due to shifts caused by shooting, it is possible to reuse the boundary position specification mask image stored in the boundary specification mask image setting and storage unit 408 for the multiple normal target images, and it is possible to efficiently generate pseudo defect target images that are similar to actual defects.

[0037] (7) The state parameter processing unit 403 generates a pseudo-defect object image by combining the pseudo-defect element image 8 at the combining position G of the normal object image 7 so that a part of the pseudo-defect element image 8 overlaps the area boundary R1, using the boundary position specifying mask image M having the area boundary R1 between the machined surface 10 and the casting surface 11 and the area boundary R2 indicating the outer periphery of the crown surface 5a of the piston 5, which are stored in the boundary specifying mask image setting and storage unit 408. Therefore, the pseudo-defect element image 8 can be placed at the area boundary R1 where judgment is difficult, and the judgment accuracy of the inspection can be improved more efficiently.

[0038] Second Embodiment FIG. 11 is a diagram illustrating state parameters of a synthesis position of a normal target image in a second embodiment.

[0039] In the first embodiment, the brightness ka of the compositing position G of the normal target image 7 was used as the state parameter (state) of the compositing position of the normal target image by the normal target image compositing position state detection processing unit 402, but in the second embodiment, the distance c from the region boundary R1 between the machined surface 10 having the cutting mark 10a at the compositing position G of the normal target image 7 and the casting surface 11, and the angle θ2 between the region boundary R1 are used as the state parameter (state) of the compositing position of the normal target image. Since the other configurations are the same as those in the first embodiment, the same reference numerals are used for the same configurations and the description thereof will be omitted.

[0040] FIG. 12 is a diagram illustrating a change in angle at a synthesis position at a boundary between regions of pseudo defect element images in the second embodiment.

[0041] If the major axis α of the pseudo defect element image 8 is directly above the region boundary R1 between the machined surface 10 having the cutting marks 10a and the casting surface 11, it becomes difficult to distinguish between the region boundary R1 and the pseudo defect element image 8, resulting in a deterioration in the accuracy of the inspection judgment. Therefore, the pseudo defect element image 8 is synthesized at a synthesis position of a normal object image obtained by rotating the major axis α of the pseudo defect element image 8 relative to the region boundary R1 by an angle θ3 larger than a predetermined angle θa, thereby generating a pseudo defect object image. For example, the predetermined angle θa is 45°. Therefore, in addition to the effects of the first embodiment, the second embodiment achieves the effect of being able to generate pseudo defect object images that are more similar to actual defects even near the region boundary R1.

[0042] Third Embodiment FIG. 13 is a diagram illustrating a filtering process for a synthesis position of a normal target image in a third embodiment.

[0043] In the first embodiment, the pseudo defect element image 8, whose brightness has been changed, is composited at the composite position G of the normal object image 7 of the machined surface 10 having the cutting marks 10a on the crown surface 5a of the piston 5. However, in the third embodiment, a smoothing filter process (filter process) is performed on the normal object image 7 of the machined surface (surface type) 10 having the cutting marks 10a at the composite position G to eliminate the cutting marks, change the surface state, and then composite the pseudo defect element image 8 to generate a pseudo defect object image. Since the other configurations are the same as those of the first embodiment, the same components are denoted by the same reference numerals and description thereof will be omitted. As a result, the pseudo defect object image is not affected by the cutting marks 10a on the machined surface 10, and therefore, a pseudo defect object image that is closer to an actual defect can be generated. Therefore, in addition to the effects of the first embodiment, the third embodiment achieves the effect of generating a pseudo defect object image that is not affected by the cutting marks on the machined surface 10, and thus, an image that is closer to an actual defect can be generated.

[0044] [Other Embodiments] While the embodiments for carrying out the present invention have been described above, the specific configuration of the present invention is not limited to the configuration of the embodiments, and design changes within the scope of the invention are also included in the present invention. For example, the object to be inspected is not limited to a piston, and the machine learning is not limited to deep learning, which is a multilayered neural network. Furthermore, although the posture control unit changes the posture of the piston using a robot, this may also be done by changing the position of a camera.

[0045] The present invention is not limited to the above-described embodiments and includes various modifications. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those including all of the described configurations. Furthermore, it is possible to replace part of the configuration of one embodiment with the configuration of another embodiment, or to add the configuration of another embodiment to the configuration of one embodiment. Furthermore, it is possible to add, delete, or replace part of the configuration of each embodiment with other configurations.

[0046] This application claims priority to Japanese Patent Application No. 2023-088315, filed May 30, 2023. The entire disclosure of Japanese Patent Application No. 2023-088315, filed May 30, 2023, including the specification, claims, drawings, and abstract, is incorporated herein by reference in its entirety.

[0047] 1 Inspection device, 4a Pseudo defect object image generation unit, 412 Learning processing unit, 414 Judgment processing unit, 5 Piston (inspected object), 7 Normal object image, 8 Pseudo defect element image, 10 Machined surface (surface type, condition), 11 Cast surface (surface type, condition), a Major axis (size, condition), b Minor axis (size, condition), c Distance from area boundary of synthesis position of normal object image (condition), α Major axis direction, θ1 Angle of major axis direction (condition), θ2 Angle with area boundary of synthesis position of normal object image (condition), θa Predetermined angle of major axis direction, G Synthesis position of normal object image, ka Brightness of synthesis position of normal object image (background brightness, condition), kb Brightness of actual pseudo defect element image (condition), kc Optimal brightness of pseudo defect element image (brightness of pseudo defect element image after change), kd Brightness difference between brightness of synthesis position of normal object image and optimal brightness of pseudo defect element image, R1 Area boundary, s Brightness threshold of the normal target image synthesis position

Claims

1. An inspection device for detecting defects in an object to be inspected using an object image photographed of the object, comprising: a pseudo-defect object image generation unit that generates a pseudo-defect object image by combining a pseudo-defect element image that simulates a defective portion with a normal object image, the pseudo-defect object image generation unit changing at least one of a state of the pseudo-defect element image, a combination position of the normal object image, and a combination position of the normal object image depending on a combination position of the normal object image at which the pseudo-defect element image is to be combined, a state of the combination position of the normal object image, and a state of the pseudo-defect element image, to combine the pseudo-defect element image at the combination position of the normal object image; a learning processing unit that machine-learns the pseudo-defect object image to create a judgment model; and a judgment processing unit that uses the judgment model to detect defects in the object image photographed of the object to be inspected.

2. An inspection device according to claim 1, wherein the state of the synthesis position of the normal object image is at least one of brightness, brightness variation, distance from the region boundary, angle from the region boundary, and type of surface.

3. An inspection apparatus according to claim 1, wherein the state of the pseudo defect element image is at least one of type, size, aspect ratio, angle, and brightness.

4. An inspection device as described in claim 2, wherein the state of the synthesis position of the normal target image is brightness, and the pseudo-defect object image generation unit changes the brightness of the pseudo-defect element image according to background brightness, which is the brightness of the synthesis position of the normal target image, and synthesizes the pseudo-defect element image at the synthesis position of the normal target image.

5. An inspection device as described in claim 4, wherein the pseudo-defect object image generation unit increases the luminance difference between the background luminance and the luminance after the modification of the pseudo-defect element image when the background luminance of the synthesis position of the normal object image is high compared to the luminance difference between the background luminance and the luminance after the modification of the pseudo-defect element image when the background luminance of the synthesis position of the normal object image is low, and synthesizes the pseudo-defect element image at the synthesis position of the normal object image.

6. An inspection device as described in claim 2, wherein the state of the synthesis position of the normal target image is brightness, and the pseudo-defect object image generation unit synthesizes the pseudo-defect element image at the synthesis position of the normal target image when background brightness, which is the brightness of the synthesis position of the normal target image, is equal to or greater than a threshold value.

7. An inspection device as described in claim 2, wherein the state of the synthesis position of the normal object image is a surface type, and the pseudo-defect object image generation unit performs filter processing on the synthesis position of the normal object image according to the surface type, and then synthesizes the pseudo-defect element image at the synthesis position of the normal object image.

8. An inspection device according to claim 7, wherein the filter used in the filtering process is a smoothing filter that smoothes the synthesis position of the normal target image.

9. An inspection device as described in claim 2, wherein the state of the synthesis position of the normal object image is a distance from the area boundary, and when a portion of the pseudo defect element image is at a distance that overlaps with the area boundary and the pseudo defect element image has an elliptical shape having a major axis and a minor axis, the pseudo defect element image is synthesized at the synthesis position of the normal object image so that the axis in the major axis direction has an angle with the area boundary that is greater than a predetermined angle.

10. An inspection device as described in claim 1, wherein the pseudo-defect object image generation unit saves the synthesis position of the normal object image, the state of the synthesis position of the normal object image, and the state of the pseudo-defect element image as history, and changes at least one of the state of the pseudo-defect element image, the synthesis position of the normal object image, and the state of the synthesis position of the normal object image based on the history, and synthesizes the pseudo-defect element image at the synthesis position of the normal object image.

11. An inspection device as described in claim 1, wherein, when there are multiple normal target images, the pseudo-defect object image generation unit extracts the locations and shapes of feature points in each of the multiple normal target images, corrects the overall positions of the multiple normal target images based on the positions of the feature points of each of the multiple normal target images, and determines a position for synthesizing the pseudo-defect element image.

12. An inspection device as described in claim 1, wherein the pseudo-defect object image generation unit sets a boundary position designation mask near an area boundary in the normal object image, and synthesizes the pseudo-defect element image at a synthesis position of the normal object image so that a portion of the pseudo-defect element image overlaps the area boundary of the boundary position designation mask.

13. An inspection method for detecting defects in an object to be inspected using an object image photographed of the object to be inspected, comprising: a synthesis step of generating a pseudo-defect object image by changing at least one of the state of the pseudo-defect element image, the synthesis position of the normal object image, and the synthesis position of the normal object image depending on the synthesis position of a normal object image into which a pseudo-defect element image is synthesized, the state of the synthesis position of the normal object image, and the state of the pseudo-defect element image; a judgment model creation step of generating a judgment model by machine learning the pseudo-defect object image; and a judgment step of detecting defects in the object image photographed of the object to be inspected using the judgment model.

Citation Information

Patent Citations

  • Image inspection device, image inspection method, and control program of image inspection device

    JP2014074709A

  • Image inspection device, image inspection system, and image inspection method

    JP2014074710A

  • Image generation device and image generation method of generating an inspection-purpose image for making performance adjustment of image inspection system

    JP2018205123A

  • Defect inspection device for display device

    JP2020003219A

  • Image selection method, image selection program, and image selection apparatus

    JP2020201713A