Apparatus and method for updating learning model for image inspection

Abstract

An apparatus and method for updating a learning model for image inspection to ensure the accuracy of decisions made using a learning model for image inspection. The apparatus includes: a light source information acquisition section that acquires first light source information representing one or both of the position and orientation of a light source in an actual space captured using a captured image; a light source information comparison section that determines whether second light source information representing one or both of the position and orientation of a light source of a first learning image used for learning of a learning model is different from the first light source information; a learning image generation section that, in the case where it is determined that the first light source information is different from the second light source information, configures a three-dimensional model simulating an object of inspection and a virtual light source simulating a light source in accordance with the first light source information in a virtual space, changes one or both of the position and orientation of the three-dimensional model, and generates a plurality of second learning images representing the three-dimensional model; and a learning model generation section that updates the learning model using the plurality of second learning images.

Description

Technical Field

[0001] This disclosure relates to apparatus and methods for updating learning models used for image inspection. Background Technology

[0002] Patent Document 1 discloses a technique for changing the position of a light source in an image obtained by photographing an object using a camera or the like.

[0003] Existing technical documents

[0004] Patent documents

[0005] Patent Document 1: Japanese Patent Application Publication No. 2013-235537 Summary of the Invention

[0006] The problem that the invention aims to solve

[0007] Learning models generated through machine learning are sometimes used for image inspection. However, the accuracy of image inspection can sometimes decrease when the position or orientation of the light source illuminating the inspected object changes, for example, due to changes in factory layout. Therefore, as mentioned in the literature above, it is possible to consider changing the position of the light source in the image obtained by taking a picture of the inspected object, but this alone cannot guarantee the accuracy of image inspection.

[0008] Technical solutions for solving the problem

[0009] This disclosure can be implemented in the following manner.

[0010] (1) According to a first aspect of this disclosure, an apparatus for updating a learning model for image inspection is provided. The apparatus comprises: a light source information acquisition unit that acquires first light source information, the first light source information representing information on one or both of the position and orientation of a light source in the actual space obtained by capturing an image of the actual space of an object to be inspected for image inspection; a light source information comparison unit that determines whether second light source information is different from the first light source information, the second light source information representing information on one or both of the position and orientation of the light source in a first learning image used for learning the learning model, which is represented by the first light source information; a learning image generation unit that, when the light source information comparison unit determines that the first light source information is different from the second light source information, arranges a three-dimensional model simulated for the object to be inspected and a virtual light source simulated according to the first light source information in a virtual space, while making one or both of the position and orientation of the three-dimensional model in the virtual space different, and generates a plurality of second learning images representing the three-dimensional model; and a learning model generation unit that updates the learning model using the plurality of second learning images generated by the learning image generation unit.

[0011] According to this apparatus, when one or both of the position and orientation of the light source in the actual space change, multiple second learning images with different positions and orientations of the 3D model can be generated using a 3D model illuminated by a virtual light source positioned at a position and orientation corresponding to the position and orientation of the light source in the actual space. These multiple second learning images are then used to update the learning model. Therefore, the accuracy of image inspection of the inspected object can be ensured.

[0012] (2) In the apparatus described above, the apparatus may also include an image inspection unit that performs an image inspection of the subject being inspected using an image obtained by taking a picture of the subject being inspected after the learning model is updated by the learning model generation unit and the learning model updated by the learning model generation unit.

[0013] According to this method, the device can use an updated learning model to perform image inspection of the subject.

[0014] (3) In the apparatus described above, the light source information comparison unit may determine whether the first light source information and the second light source information are different if the misjudgment rate of the image inspection performed by the image inspection unit exceeds a predetermined value.

[0015] According to this method, when the false positive rate of image inspection increases due to changes in the position or orientation of the light source, the learning model can be automatically updated to reduce the false positive rate of image inspection.

[0016] (4) In the apparatus described above, the apparatus may also include a storage unit for storing the first learning image, and the light source information acquisition unit may acquire the second light source information by parsing the first learning image stored in the storage unit.

[0017] According to this method, the device can use a first learning image stored in the storage unit to obtain information about a second light source.

[0018] (5) According to a second aspect of this disclosure, a method for updating a learning model for image inspection is provided. The method includes: a light source information acquisition step, acquiring first light source information, the first light source information representing information on one or both of the position and orientation of a light source in the actual space obtained by taking a photographic image of the actual space of the object to be inspected for image inspection; a light source information comparison step, determining whether second light source information is different from the first light source information, the second light source information representing information on one or both of the position and orientation of the light source in the first learning image used for learning the learning model, which is represented by the first light source information; a learning image generation step, in which, if it is determined in the light source information comparison step that the first light source information is different from the second light source information, a three-dimensional model simulated for the object to be inspected and a virtual light source simulated according to the first light source information are configured in a virtual space, while making one or both of the position and orientation of the three-dimensional model in the virtual space different, and generating a plurality of second learning images representing the three-dimensional model; and a learning model generation step, using the plurality of second learning images generated in the learning image generation step to update the learning model.

[0019] According to this method, when one or both of the position and orientation of the light source in the actual space change, multiple second learning images with different positions and orientations of the 3D model can be generated using a 3D model illuminated by a virtual light source positioned and oriented in a position and orientation corresponding to the position and orientation of the light source in the actual space. These multiple second learning images are then used to update the learning model. Therefore, the accuracy of image inspection of the inspected object can be ensured.

[0020] This disclosure can also be implemented by various means other than an apparatus for updating a learning model for image inspection and an apparatus for updating a learning model for image inspection. For example, it can be implemented by an image inspection system, an image inspection apparatus, an image inspection method, etc. Attached Figure Description

[0021] Figure 1 This is an illustrative diagram that schematically represents the general structure of an image inspection system.

[0022] Figure 2 This is a flowchart representing the content of the learning model update process.

[0023] Figure 3 This is a flowchart representing the content generated and processed using images during learning.

[0024] Figure 4 This is an explanatory diagram representing a three-dimensional model configured in virtual space. Detailed Implementation

[0025] A. First implementation method:

[0026] Figure 1 This is an explanatory diagram schematically showing the general structure of the image inspection system 10 in the first embodiment. In this embodiment, the image inspection system 10 includes a stage 20, a light source 30, a camera 40, a camera controller 45, an information processing device 50, and a display device 60. The information processing device 50 is sometimes referred to as a device for updating the learning model GM used for image inspection, or an image inspection device.

[0027] The stage 20 is disposed, for example, indoors in a factory or similar facility. The object to be inspected, RM, which is the subject of image inspection, is placed on the stage 20. In this embodiment, the object to be inspected, RM, is a car bumper. Even for bumpers of the same car model, various components such as the emblem or grille mounted on the bumper may differ depending on the model grade or sales region. The image inspection system 10 inspects the bumper RM to check the appropriate installation of various components according to the model grade or sales region. The object to be inspected, for example, is transported by a robotic arm (not shown) and placed on the stage 20.

[0028] The light source 30 illuminates the object RM to be inspected, which is placed on the stage 20. The light source 30 may be an LED (Light Emitting Diode). The light source 30 is fixed by a light source fixing clamp (not shown).

[0029] Camera 40 takes an image of the subject RM placed on stage 20. Camera 40 takes an image of the subject RM illuminated by light source 30. Camera 40 may be, for example, a camera with a CCD (Charge-Coupled Device) image sensor or a CMOS (Complementary Metal Oxide Semiconductor) image sensor. Camera 40 is secured by a camera clamp (not shown). Camera 40 is connected to camera controller 45 via wired or wireless communication, and output signals from camera 40 are sent to camera controller 45.

[0030] The camera controller 45 generates an image RG by processing the output signal from the camera 40. Furthermore, in this embodiment, the camera controller 45 estimates the position and orientation of the light source 30 by analyzing the image RG, and generates light source information LD representing the estimated position and orientation of the light source 30. The camera controller 45 estimates the position and orientation of the light source 30, for example, using a successive relaxation method, in which it searches for reflection parameters that simultaneously satisfy the conditions that the diffuse reflection component of the reflected light represented by the image RG follows the Lambertian model, the specular reflection component of the reflected light represented by the image RG follows the Torrance-Sparrow reflection model, the position of the light source 30, and the orientation of the light source 30. The camera controller 45 can estimate the position and orientation of the light source 30 based on an image RG using the above method. The camera controller 45 is connected to the information processing device 50 via wired or wireless communication, and the image RG and light source information LD generated by the camera controller 45 are sent to the information processing device 50.

[0031] The information processing device 50 includes a CPU 101, a memory 102, and an input / output interface 103. In this embodiment, the information processing device 50 includes an image acquisition unit 110, a light source information acquisition unit 120, a light source information comparison unit 130, a learning image generation unit 140, a learning model generation unit 150, and an image inspection unit 160. The image acquisition unit 110, the light source information acquisition unit 120, the light source information comparison unit 130, the learning image generation unit 140, the learning model generation unit 150, and the image inspection unit 160 are implemented in software by the CPU 101 executing a computer program stored in the memory 102. Alternatively, at least one of the image acquisition unit 110, the light source information acquisition unit 120, the light source information comparison unit 130, the learning image generation unit 140, the learning model generation unit 150, and the image inspection unit 160 can also be implemented using hardware circuitry. The memory 102 is sometimes referred to as a storage unit.

[0032] The image acquisition unit 110 acquires the captured image RG. In this embodiment, the image acquisition unit 110 acquires the captured image RG sent from the camera controller 45. The captured image RG acquired by the image acquisition unit 110 is stored in the memory 102.

[0033] The light source information acquisition unit 120 acquires light source information LD, which indicates the position and orientation of the light source 30. In this embodiment, the light source information acquisition unit 120 acquires the light source information LD sent from the camera controller 45. The light source information LD acquired by the light source information acquisition unit 120 is stored in the memory 102.

[0034] The light source information comparison unit 130 determines whether at least one of the positions and orientations of the light source 30 represented by the two light source information LDs is different by comparing the two light source information LDs.

[0035] The learning image generation unit 140 generates a learning image VG for machine learning performed by the learning model generation unit 150. In this embodiment, the learning image generation unit 140 is configured as a 3D simulator. The learning image VG generated by the learning image generation unit 140 is stored in the memory 102.

[0036] The learning model generation unit 150 generates a learning model GM by performing machine learning using the learning image VG. The learning model GM generated by the learning model generation unit 150 is stored in the memory 102.

[0037] The image inspection unit 160 performs an image inspection of the object RM using the captured image RG obtained by the camera 40 and the learning model GM generated by the learning model generation unit 150. The inspection result RD of the image inspection performed by the image inspection unit 160 is stored in the memory 102.

[0038] The display device 60 is, for example, a liquid crystal display (LCD) or an organic EL display. The display device 60 displays the inspection result RD of the image inspection performed by the image inspection unit 160. For example, if the image inspection unit 160 determines that the bumper, which is the inspected object RM, is properly fitted with various components corresponding to its grade or sales region, the word "OK" is displayed on the display device 60; if the various components are not properly fitted, the word "NG" is displayed on the display device 60. The display device 60 is, for example, located near the workbench 20 in the factory. By checking the inspection result RD displayed on the display device 60, the inspector can determine whether the various components corresponding to the grade or sales region are properly fitted to the bumper. If the word "NG" is displayed on the display device 60, the inspector performs a visual inspection. If the visual inspection result indicates that the various components corresponding to the grade or sales region are properly fitted to the bumper, the inspector inputs a false judgment from the image inspection unit 160 into the information processing device 50. The inspection result RD contains information about the false judgment rate.

[0039] In the image inspection system 10 of this embodiment, since the image inspection unit 160 performs image inspection using the learning model GM, the workload of the inspector is reduced compared to the method of visual inspection of the inspected object RM by an inspector. Furthermore, in image inspection using the learning model GM as in this embodiment, the inspection can usually be completed in a shorter time compared to visual inspection by an inspector. However, for example, when the position or orientation of the light source 30 changes due to changes in the factory layout, the judgment accuracy of the image inspection performed by the image inspection unit 160 may sometimes decrease. Decreased judgment accuracy means an increased false positive rate. This decrease in judgment accuracy can occur, for example, due to changes in the color tone of the inspected object RM represented by the image RG taken before and after a change in the position or orientation of the light source 30. In this embodiment, even if the position or orientation of the light source 30 changes, the false positive rate of image inspection can be reduced by updating the learning model GM used for image inspection by performing learning model update processing by the information processing device 50.

[0040] Figure 2 This is a flowchart representing the content of the learning model update process. Figure 3 This is a flowchart representing the image generation process performed during the learning model update process. Figure 4 This is an explanatory diagram showing the 3D model VM configured in the virtual space (VS) used in learning image generation processing.

[0041] like Figure 2As shown, in this embodiment, if the false judgment rate of the image inspection represented by the inspection result RD stored in the memory 102 exceeds a predetermined value, the CPU 101 of the information processing device 50 starts the learning model update process. After starting the learning model update process, firstly, in step S110, the image acquisition unit 110 controls the camera 40 via the camera controller 45, thereby using the camera 40 to capture an image of the inspected object RM, and acquires the captured image RG of the inspected object RM from the camera controller 45.

[0042] Next, in step S120, the light source information acquisition unit 120 acquires light source information LD, which represents the position and orientation of the light source 30. In this embodiment, the camera controller 45 estimates the position and orientation of the light source 30 by analyzing the captured image RG, thereby generating the light source information LD, and the light source information acquisition unit 120 acquires the light source information LD generated by the camera controller 45. The light source information LD acquired by the light source information acquisition unit 120 is stored in the memory 102. In addition, step S120 is sometimes referred to as the light source information acquisition step of the method for updating the learning model GM for image inspection, or the light source information acquisition step of the image inspection method.

[0043] In step S130, the light source information comparison unit 130 compares the light source information LD (first light source information LD1) representing the current position and orientation of the light source 30, obtained in step S120, with the light source information LD (second light source information LD2) representing the past position and orientation of the light source 30 stored in the memory 102, thereby determining whether at least one of the position and orientation of the light source 30 has changed. In this embodiment, if at least one of the position and orientation of the light source 30 represented by the first light source information LD1 deviates from at least one of the position and orientation of the light source 30 represented by the second light source information LD2, the light source information comparison unit 130 determines that at least one of the position and orientation of the light source 30 has changed. The statement that at least one of the position and orientation of the light source 30 represented by the first light source information LD1 is deviated from at least one of the position and orientation of the light source 30 represented by the second light source information LD2 means that at least one of the following states exists: the position of the light source 30 represented by the first light source information LD1 is separated from the position of the light source 30 represented by the second light source information LD2 by a predetermined distance or more; or the orientation of the light source 30 represented by the first light source information LD1 is separated from the orientation of the light source 30 represented by the second light source information LD2 by a predetermined angle or more. Furthermore, step S130 is sometimes referred to as a light source information comparison step in a method for updating the learning model GM for image inspection, or a light source information comparison step in an image inspection method. The light source information comparison unit 130 uses light source information LD as the second light source information LD2, which represents the position and orientation of the light source 30 in the learning image VG used for learning the learning model GM for image inspection before the learning model update process begins. Sometimes the learning image VG used for learning the learning model GM for image inspection before the learning model update process begins is referred to as the first learning image VG1.

[0044] If, in step S130, it is not determined that at least one of the position and orientation of the light source 30 has changed, the information processing device 50 terminates the processing. If, in step S130, it is determined that at least one of the position and orientation of the light source 30 has changed, in step S140, the learning image generation unit 140 begins learning image generation processing.

[0045] like Figure 3 As shown, after starting the learning process for image generation, firstly, in step S141, the learning image generation unit 140... Figure 4The virtual space VS shown is configured to simulate a three-dimensional model VM of the object under inspection RM. In this embodiment, the learning image generation unit 140 reads the three-dimensional model VM pre-stored in the memory 102 and configures the three-dimensional model VM in the virtual space VS. The three-dimensional model VM simulates not only the three-dimensional shape of the object under inspection RM, but also the material of the object under inspection RM. In addition, the real space in which the object under inspection RM is configured is sometimes referred to as the actual space.

[0046] Next, in step S142, the learning image generation unit 140 modifies the position and orientation of the virtual light source VL disposed in the virtual space VS according to the position and orientation of the light source 30 represented by the first light source information LD1 obtained in step S120 of the learning model update process. Alternatively, the order of steps S141 and S142 can be reversed.

[0047] In step S143, the learning image generation unit 140 captures a 3D model VM positioned in the virtual space VS using a virtual camera VC that simulates the viewpoint of the camera 40, thereby generating a learning image VG representing the 3D model VM illuminated by a virtual light source VL. In the learning image VG, the 3D model VM is represented from the same viewpoint as when the camera 40 captures the object under inspection RM. The learning image VG generated by the learning image generation unit 140 is stored in the memory 102.

[0048] In step S144, the learning image generation unit 140 determines whether a predetermined number of learning images VG have been generated. For example, the learning image generation unit 140 determines whether 1000 learning images VG have been generated.

[0049] If it is not determined in step S144 that a predetermined number of learning images VG have been generated, the learning image generation unit 140, in step S145, changes at least one of the position and orientation of the 3D model VM in the virtual space VS without changing the position and orientation of the virtual light source VL, and then returns to step S143 to generate the learning images VG. The learning image generation unit 140 repeatedly performs the process from step S143 to step S145 until it is determined in step S144 that a predetermined number of learning images VG have been obtained, thereby generating a predetermined number of learning images VG where at least one of the position and orientation of the 3D model VM is different.

[0050] If, in step S144, it is determined that a predetermined number of learning images VG have been obtained, the learning image generation unit 140 determines in step S146 whether learning images VG have been generated for all 3D model VMs. If, in step S146, it is not determined that learning images VG have been generated for all 3D model VMs, the learning image generation unit 140 switches the 3D model VMs configured in the virtual space VS in step S147. In this embodiment, multiple 3D model VMs are pre-stored in the memory 102. Each 3D model VM simulates multiple inspection objects RMs, i.e., multiple bumpers with different grades or sales regions. After switching the 3D model VMs configured in the virtual space VS by reading in different types of 3D model VMs from those that have completed the generation of a predetermined number of learning images VG, the learning image generation unit 140 returns to step S143 to generate learning images VG.

[0051] The learning image generation unit 140 repeatedly performs the processes from step S143 to step S146 until it is determined in step S146 that learning images VG have been generated for all 3D models VM, thereby generating a predetermined number of learning images VG for each 3D model VM. If it is determined in step S146 that learning images VG have been generated for all 3D models VM, the learning image generation unit 140 ends the learning image generation process. Furthermore, the learning image generation process is sometimes referred to as the learning image generation step of a method for updating the learning model GM for image inspection, or the learning image generation step of an image inspection method. Sometimes the learning images VG generated by the learning image generation process are referred to as the second learning image VG2.

[0052] like Figure 3 As shown, after the image generation process for learning is completed, in step S150 of the learning model update process, the learning model generation unit 150 assigns labels representing grades or sales regions to the learning image VG, generating a learning dataset. In step S160, the learning model generation unit 150 generates a learning model GM by reading the learning dataset and performing machine learning, thereby updating the learning model GM stored in the memory 102. In this embodiment, the learning model generation unit 150 generates the learning model GM by performing machine learning using a convolutional neural network. Afterward, the information processing apparatus 50 ends the learning model update process. Additionally, step S160 is sometimes referred to as the learning model generation step of a method for updating the learning model GM for image inspection, or the learning model generation step of an image inspection method.

[0053] According to the image inspection system 10 of this embodiment described above, when at least one of the position and orientation of the light source 30 in the actual space changes, the information processing device 50 uses a three-dimensional model VM illuminated by a virtual light source VL positioned at a position and orientation corresponding to the position and orientation of the light source 30 in the actual space to generate multiple learning images VG, each with a different position and orientation, and uses these multiple learning images VG to update the learning model GM. Therefore, even when at least one of the position and orientation of the light source 30 in the actual space changes, the accuracy of the image inspection of the inspected object RM can be ensured. In particular, in this embodiment, since multiple learning images VG, each with a different position and orientation, are used to update the learning model GM, the accuracy of the image inspection can be ensured even when the position or orientation of the inspected object RM placed on the stage 20 deviates from the reference position or orientation. Furthermore, in this embodiment, since the image generated using the three-dimensional model VM is used as the learning image VG, the learning image VG can be collected in a shorter time compared to the method of using the captured image RG obtained by using the camera 40 to capture the subject RM as the learning image VG.

[0054] Furthermore, in this embodiment, the information processing apparatus 50 includes an image inspection unit 160, which performs an image inspection of the object RM using an image RG obtained by capturing the object RM after the learning model GM is updated by the learning model generation unit 150, and the updated learning model GM. Therefore, the information processing apparatus 50 can perform an image inspection of the object RM using the updated learning model GM.

[0055] Furthermore, in this embodiment, the information processing device 50 initiates learning model update processing when the false positive rate of the image inspection performed by the image inspection unit 160 exceeds a predetermined value. Therefore, when the false positive rate of image inspection increases, the learning model GM can be automatically updated to reduce the false positive rate. In particular, in this embodiment, even if the inspector is unaware that the position or orientation of the light source 30 has changed and the false positive rate of image inspection has increased, the learning model GM can still be automatically updated to reduce the false positive rate.

[0056] B. Other implementation methods:

[0057] (B1) In the image inspection system 10 of the above embodiment, the camera controller 45 has the function of receiving the output signal from the camera 40 to generate an image RG, and the function of estimating the position and orientation of the light source 30 by analyzing the image RG to generate light source information LD. In contrast, the information processing device 50 may also have at least some of the functions of the camera controller 45. For example, the camera controller 45 may not be provided in the image inspection system 10, and the information processing device 50 may have all the above-mentioned functions of the camera controller 45. Alternatively, in the image inspection system 10, the camera controller 45 receives the output signal from the camera 40 to generate an image RG, and the light source information acquisition unit 120 of the information processing device 50 estimates the position and orientation of the light source 30 by analyzing the image RG sent from the camera controller 45 to generate light source information LD.

[0058] (B2) In the image inspection system 10 of the above embodiment, the second light source information LD2 is stored in the memory 102 of the information processing device 50. Conversely, the second light source information LD2 may not be stored in the memory 102 of the information processing device 50. In this case, for example, the light source information acquisition unit 120 may be configured to have the function of estimating the position and orientation of the light source 30 by analyzing the captured image RG, and generating the light source information LD. The light source information acquisition unit 120 generates the second light source information LD2 by reading from the memory 102 and analyzing the captured image RG taken before the position or orientation of the light source 30 changes.

[0059] (B3) In the image inspection system 10 of the above embodiments, an image inspection unit 160 is provided in the information processing device 50. In contrast, the image inspection unit 160 may not be provided in the information processing device 50, but may be provided in the camera controller 45, for example. In this case, the image inspection of the object to be inspected RM can be performed using the captured image RG obtained by the camera 40 and the learning model GM obtained from the information processing device 50.

[0060] (B4) In the image inspection system 10 of the above embodiments, if the false positive rate represented by the inspection result RD of the image inspection stored in the memory 102 of the information processing device 50 exceeds a predetermined value, the information processing device 50 automatically starts the learning model update process. Conversely, the information processing device 50 may not automatically start the learning model update process. For example, during the periodic inspection of the image inspection system 10, if the operator performing the periodic inspection performs a predetermined start operation on the information processing device 50, the information processing device 50 may start the learning model update process.

[0061] (B5) In the image inspection system 10 of the above embodiments, the learning image generation unit 140 uses multiple three-dimensional models (VMs) to generate a learning image (VG) during the learning image generation process. In contrast, the learning image generation unit 140 may also use only one three-dimensional model (VM) to generate a learning image (VG) during the learning image generation process.

[0062] (B6) In the image inspection system 10 of the above embodiments, the first light source information LD1 and the second light source information LD2 represent the position and orientation of the light source 30. Conversely, the first light source information LD1 and the second light source information LD2 may not represent the orientation of the light source 30. In this case, if the position of the light source 30 represented by the first light source information LD1 deviates from the position of the light source 30 represented by the second light source information LD2, the learning image generation unit 140 can change the position of the virtual light source VL according to the first light source information LD1 to generate a learning image VG representing the three-dimensional model VM illuminated by the virtual light source VL with the changed position. Alternatively, the first light source information LD1 and the second light source information LD2 may not represent the position of the light source 30. In this case, if the orientation of the light source 30 represented by the first light source information LD1 deviates from the orientation of the light source 30 represented by the second light source information LD2, the learning image generation unit 140 can change the orientation of the virtual light source VL according to the first light source information LD1 to generate a learning image VG representing the three-dimensional model VM illuminated by the virtual light source VL with the changed orientation.

[0063] This disclosure is not limited to the embodiments described above, and can be implemented in various structures without departing from the spirit of this disclosure. For example, in order to solve some or all of the above-described problems, or to achieve some or all of the above-described effects, the technical features in the embodiments corresponding to the technical features in the various methods described in the Summary of the Invention section can be appropriately replaced or combined. Furthermore, if a technical feature is not described as a necessary technical feature in this specification, it can be appropriately deleted.

[0064] Label Explanation

[0065] 10…Image inspection system, 20…Stage, 30…Light source, 40…Camera, 45…Camera controller, 50…Information processing device, 60…Display device, 101…CPU, 102…Memory, 103…Input / output interface, 110…Image acquisition unit, 120…Light source information acquisition unit, 130…Light source information comparison unit, 140…Learning image generation unit, 150…Learning model generation unit, 160…Image inspection unit.

Claims

1. An apparatus for updating a learning model used for image inspection, comprising: The light source information acquisition unit acquires first light source information, which represents information on one or both of the position and orientation of a light source in the actual space estimated by taking a picture of the actual space of the inspected object being inspected using a captured image. The light source information comparison unit determines whether at least one of the position and orientation of the light source represented by the second light source information and the first light source information deviates from each other. The second light source information represents one or both of the position and orientation of the light source in the first learning image used by the learning model. The learning image generation unit, when the light source information comparison unit determines that the first light source information and the second light source information are different, arranges a three-dimensional model simulated for the inspected object and a virtual light source simulated for the light source according to the first light source information in a virtual space, and generates a plurality of second learning images representing the three-dimensional model while making one or both of the position and orientation of the three-dimensional model in the virtual space different; and The learning model generation unit updates the learning model using the plurality of second learning images generated by the learning image generation unit. The apparatus includes an image inspection unit that performs image inspection on the subject by using an image captured after the learning model is updated by the learning model generation unit and the updated learning model. If the misjudgment rate of the image inspection performed by the image inspection unit exceeds a predetermined value, the light source information comparison unit determines whether the first light source information and the second light source information are different.

2. The apparatus according to claim 1, wherein, The device includes a storage unit for storing the first learning image. The light source information acquisition unit obtains the second light source information by parsing the first learning image stored in the storage unit.

3. A method for updating a learning model used for image inspection, comprising: The light source information acquisition process acquires first light source information, which represents information on one or both of the position and orientation of the light source in the actual space estimated by taking a picture of the actual space of the object being inspected using the image of the object being inspected. The light source information comparison process determines whether at least one of the position and orientation of the light source represented by the second light source information and the first light source information deviates from each other. The second light source information represents one or both of the position and orientation of the light source in the first learning image used by the learning model. In the learning image generation process, if it is determined in the light source information comparison process that the first light source information and the second light source information are different, a three-dimensional model simulated by the inspected object and a virtual light source simulated by the light source according to the first light source information are configured in a virtual space, while making one or both of the position and orientation of the three-dimensional model in the virtual space different, and generating multiple second learning images representing the three-dimensional model. and The learning model generation step uses the plurality of second learning images generated in the learning image generation step to update the learning model. The method further includes an image inspection step, in which an image of the subject being inspected, obtained by taking a picture of the subject after updating the learning model in the learning model generation step, and the updated learning model in the learning model generation step are used to perform an image inspection of the subject being inspected. In the light source information comparison process, if the false judgment rate of the image inspection performed in the image inspection process exceeds a predetermined value, it is determined whether the first light source information and the second light source information are different.

Images