AI model creation method and inspection device

Abstract

To provide a generation method of an AI (artificial intelligence) model and an inspection device which corrects an image by the AI model where the accuracy of inspection by correcting noise removal and resolution enhancement may be improved by the AI model generated by learning using AI.SOLUTION: In a generation method of an AI model in an inspection device 1 which corrects an image of an inspection object using the AI model generated by machine learning, a high-quality image of the inspection object picked up by the inspection device 1 is used as a teacher image, a low-quality image of the inspection object picked up by the inspection device 1 and sharing the same visual field as the teacher image is used as a learning image, the learning image is used as an input, an image corrected by the AI model is used as an output, and the output is compared with the teacher image to generate an AI model.SELECTED DRAWING: Figure 6

Description

[Technical Field]

[0001] The present invention relates to a method for creating an AI model and an inspection device that corrects images using this AI model. [Background technology]

[0002] Inspection devices that measure the solder shape on the front and back surfaces of circuit boards perform inspections using images of the object under inspection (see, for example, Patent Documents 1 and 2). For example, in an inspection device using an automated optical inspection (AOI) system, an object under inspection is irradiated with illumination light and imaged with a camera. At this time, imaging conditions may be changed (such as widening the imaging range (field of view) or shortening the exposure time) to improve the throughput of the production line. Furthermore, tomosynthesis X-ray inspection devices are configured to capture multiple transmission images by changing the relative positions of the radiation source (radiation generator), the object under inspection, and the detector, and then generate a reconstructed image from these transmission images. Therefore, capturing multiple transmission images and generating a reconstructed image using the transmission images takes time, reducing the throughput of the production line. Furthermore, in order to increase throughput, it may be possible to only inspect a portion of the object under inspection. Furthermore, depending on the material, shape, and other configurations of the object under inspection, there is also a demand for minimizing the radiation dose (exposure) during inspection. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2012-053015 [Patent Document 2] Japanese Patent Application Laid-Open No. 2008-026334 Summary of the Invention [Problem to be solved by the invention]

[0004] When the imaging conditions are changed in the inspection device as described above, there are problems such as an increase in noise in the captured image and a decrease in resolution. For example, by reducing the intensity of radiation irradiated in the X-ray inspection device, or shortening the exposure time (radiation irradiation time), or widening the field of view (the imaging area on the object to be inspected, hereinafter referred to as "FOV (Field Of View)"), or by reducing the number of transmission images used to generate a reconstructed image, it is possible to reduce the imaging time and the amount of radiation exposure during the inspection, but as a result, noise increases in the cross-sectional images obtained from the transmission images and the reconstructed images, or the resolution decreases.

[0005] The present invention has been made in consideration of these problems, and aims to provide a method for creating an AI model that improves inspection accuracy by correcting the increased noise and reduced resolution caused by changes in imaging conditions using an AI model created through learning using AI (artificial intelligence), and an inspection device that corrects images using this AI model. [Means for solving the problem]

[0006] In order to solve the above-mentioned problems, the AI ​​model creation method according to the present invention is a method for creating an AI model in an inspection device that corrects an image of an object to be inspected using an AI model created by machine learning, the method comprising: using a high-quality image of the object to be inspected captured by the inspection device as a teacher image; The first step to A low-quality image of the object to be inspected taken by the inspection device and having the same field of view as the teacher image is used as a learning image. The second step is to , create A third step of setting an AI model, dividing each of the teacher image and the learning image into a plurality of partial regions, and selecting a partial region for learning from the partial regions. The learning image is input, and the image corrected by the AI ​​model is output, and The learning subdomain By comparing with the training images and learning, the AI ​​model update do a fourth step, a fifth step of inputting the learning image of a verification partial region among the partial regions, outputting an image corrected by the AI ​​model updated in the fourth step, and calculating a loss due to evaluation from the difference between the evaluation value of the output by evaluation and the evaluation value of the teacher image of the verification partial region, and a sixth step of repeating the fourth step and the fifth step until a predetermined condition is satisfied. . Furthermore, the fifth step of the AI ​​model creation method according to the present invention preferably includes: a step 5-1 of taking as input the training image used to update the AI ​​model in the fourth step, taking as output a training image corrected by the AI ​​model updated in the fourth step, and calculating a training loss from the difference between an evaluation value of the training output and an evaluation value of the teacher image used to update the AI ​​model in the fourth step; a step 5-2 of taking as input the training image of a verification partial region among the partial regions, taking as output an evaluation image corrected by the AI ​​model updated in the fourth step, and calculating an evaluation loss from the difference between the evaluation value of the evaluation output and an evaluation value of the teacher image of the verification partial region; and a step 5-3 of displaying on a monitor a graph of the training loss and the evaluation loss corresponding to each number of times of training from the start of learning to the present time. Furthermore, it is desirable that the sixth step of the AI ​​model creation method of the present invention accepts an operator's decision as to whether or not to terminate the creation of the AI ​​model for the graph, and terminates the creation of the AI ​​model when an input of a decision to terminate is received.

[0007] Furthermore, in the AI ​​model creation method of the present invention, when an image of the object to be inspected is acquired by dividing it into a plurality of imaging regions, it is desirable to create an AI model specialized for each of the imaging regions for at least two of the plurality of imaging regions.

[0008] In addition, the method for creating an AI model according to the present invention is a method for creating an AI model according to the present invention, wherein the image of the object to be inspected is Imaging When the imaging areas are divided into regions and acquired, it is desirable to create the AI ​​model common to at least two of the imaging areas.

[0009] Furthermore, in the method for creating an AI model according to the present invention, it is desirable to divide the image of the object to be inspected into a plurality of partial regions, and to create an AI model specialized for each of the partial regions for at least two of the plurality of partial regions.

[0010] Furthermore, in the AI ​​model creation method according to the present invention, it is desirable to create an AI model specialized for an image of a specific region of the object to be inspected, among the images of the object to be inspected.

[0011] In addition, in the AI ​​model creation method according to the present invention, it is desirable that the image quality is the amount of noise or resolution.

[0012] Further, the inspection device according to the present invention includes a light source, a holder for holding an object under inspection, a detector for detecting light emitted from the light source and reflected by or transmitted through the object under inspection to acquire an image of the object under inspection, and a control unit, wherein the control unit irradiates the object under inspection with light from the light source, acquires the teacher image and the learning image of the same field of view at different times as images of the object under inspection using the detector, and generates the AI ​​model using the teacher image and the learning image by one of the methods for creating an AI model described above. image The AI ​​model is created to correct the above.

[0013] Further, the inspection device according to the present invention teeth, a source that emits radiation Light source and , a holder for holding an object to be inspected; and Passed through the test object the light source Detecting radiation from the object to obtain an image of the object detector and , a control unit; The control unit the light source and changing the relative position of the holding unit and the detector, the light source is emitted from the radiation It is desirable to obtain the teacher image and the learning image of the same field of view at different times as transmission images of the object to be inspected by acquiring the above-mentioned with the detector, and to create the AI ​​model for correcting the transmission image using the teacher image and the learning image by one of the AI ​​model creation methods described above.

[0014] Further, the inspection device according to the present invention teeth , a source that emits radiation Light source and , a holder for holding an object to be inspected; and Passed through the test object the light source Detecting radiation from the covered Acquire an image of the object to be inspected detector and , a control unit; The control unit the light source and changing the relative position of the holding unit and the detector, the light source is emitted from the radiation by the detector, it is desirable to obtain transmission images of the object under test for the teacher image and the training image of the same field of view at different times, reconstruct the transmission image for the teacher image to generate the teacher image which is a cross-sectional image of the object under test, reconstruct the transmission image for the training image to generate the training image which is a cross-sectional image of the object under test, and create the AI ​​model for correcting the cross-sectional image using the teacher image and the training image by one of the AI ​​model creation methods described above.

[0015] In the inspection device according to the present invention, the control unit the light source and changing the relative position of the holding unit and the detector, the light sourceis emitted from the radiation It is desirable to obtain a transmission image of the object to be inspected by acquiring the above-mentioned transmission image with the detector, and to perform correction using the AI ​​model on the acquired transmission image or a cross-sectional image generated from the transmission image.

[0016] In the inspection apparatus according to the present invention, it is preferable that the correction is performed on the image of a specified region of the object under inspection.

[0017] It is also preferable that the inspection device according to the present invention has a user interface, and that the control unit accepts the learning conditions through the user interface. [Effects of the Invention]

[0018] According to the AI ​​model creation method and inspection device of the present invention, even when the imaging conditions are changed (for example, by weakening the radiation or shortening the imaging time in an X-ray inspection device), it is possible to obtain images (transmission images and cross-sectional images) with low noise and high resolution, thereby improving inspection accuracy. [Brief explanation of the drawings]

[0019] [Figure 1] FIG. 1 is an explanatory diagram for explaining a configuration of an inspection device according to an embodiment. [Figure 2] FIG. 2 is an explanatory diagram for explaining each functional block processed by a control unit of the inspection device. [Figure 3] 10 is a flowchart illustrating the flow of an examination. [Figure 4] 10 is a flowchart illustrating a flow of a process of capturing a transmission image and generating a reconstructed image. [Figure 5] 10A and 10B are explanatory diagrams for explaining the movement of the substrate holder and the detector, and the timing of X-ray emission from the radiation generator and imaging by the detector, where FIG. 10A shows a timing chart and FIG. 10B shows the timing of exposure. [Figure 6] 1 is a flowchart of the learning process of an AI model. [Figure 7] FIG. 10 is an explanatory diagram for explaining image division in learning and correction. [Figure 8] 10 is a graph showing the relationship between the number of epochs and the loss in learning. [Figure 9] FIG. 2 is an explanatory diagram for explaining the relationship between an input image (learning image), an output image, and a teacher image in learning. [Figure 10] 10 is a flowchart of a process of capturing a transmitted image and generating a reconstructed image, which is executed during an examination. DETAILED DESCRIPTION OF THE INVENTION

[0020] Preferred embodiments of the present invention will be described below with reference to the drawings. While the description will be based on an X-ray inspection apparatus, the same applies to other types of inspection apparatus, such as AOI, as will be described later. As shown in FIG. 1, the inspection apparatus 1 according to this embodiment includes a control unit 10, which is a processing device such as a personal computer (PC), a monitor 12, and an imaging unit 32. The imaging unit 32 further includes a radiation quality changing unit 14, a radiation generator driving unit 16, a substrate holder driving unit 18, a detector driving unit 20, a radiation generator 22, a substrate holder 24, and a detector 26.

[0021] The radiation generator 22 is a device (ray source) that generates radiation such as X-rays, and generates radiation by, for example, colliding accelerated electrons with a target such as tungsten or diamond. In this embodiment, the radiation will be described as X-rays, but is not limited to this. For example, the radiation may be alpha rays, beta rays, gamma rays, ultraviolet rays, visible light, or infrared rays. The radiation may also be microwaves or terahertz waves. This radiation generator 22 functions as a light source (ray source) that irradiates the object under test with light (radiation).

[0022] The substrate holding unit 24 holds a substrate, which is an object to be inspected. The substrate held by the substrate holding unit 24 is irradiated with radiation generated by the radiation generator 22, and the radiation that has passed through the substrate is captured as an image by the detector 26. Hereinafter, the radiation transmission image of the substrate captured by the detector 26 will be referred to as a "transmission image." As will be described later, in this embodiment, the substrate holding unit 24 holding the substrate and the detector 26 are moved relative to the radiation generator 22 to acquire multiple transmission images and generate a reconstructed image.

[0023] The transmission images captured by the detector 26 are sent to the control unit 10, where they are reconstructed into images including the three-dimensional shape of the solder at the joint using a known technique such as the filtered back projection (FBP) method. The reconstructed images and transmission images are then stored in a storage device within the control unit 10 or in an external storage device (not shown). Hereinafter, an image reconstructed into a three-dimensional image including the three-dimensional shape of the solder at the joint based on the transmission images will be referred to as a "reconstructed image." Furthermore, an image obtained by cutting out an arbitrary cross section from the reconstructed image will be referred to as a "cross-sectional image." Such reconstructed images and cross-sectional images are output to the monitor 12. Note that the monitor 12 displays not only the reconstructed images and cross-sectional images, but also the inspection results of the solder joint state, which will be described later. Furthermore, the reconstructed images in this embodiment are also referred to as "planar CT" because they are reconstructed from planar images captured by the detector 26, as described above.

[0024] The radiation quality changing unit 14 changes the radiation quality of the radiation generated by the radiation generator 22. The radiation quality is determined by a voltage (hereinafter referred to as "tube voltage") applied to accelerate electrons to be bombarded with a target, and a current (hereinafter referred to as "tube current") that determines the number of electrons. The radiation quality changing unit 14 is a device that controls the tube voltage and tube current. This radiation quality changing unit 14 can be realized using known technology such as a transformer or a rectifier.

[0025] Here, the quality of radiation is determined by the brightness and hardness of the radiation (the spectral distribution of the radiation). Increasing the tube current increases the number of electrons that collide with the target, thereby increasing the number of radiation photons generated. As a result, the brightness of the radiation increases. For example, some components, such as capacitors, are thicker than other components, and to capture transmission images of these components, radiation with high brightness must be irradiated. In such cases, the brightness of the radiation can be adjusted by adjusting the tube current. Furthermore, increasing the tube voltage increases the energy of the electrons that collide with the target, thereby increasing the energy (spectrum) of the generated radiation. Generally, the higher the energy of radiation, the greater its penetration power into materials and the less it is absorbed by them. Transmission images captured using such radiation have lower contrast. Therefore, the tube voltage can be used to adjust the contrast of transmission images.

[0026] The radiation generator driving unit 16 has a driving mechanism such as a motor (not shown), and can move the radiation generator 22 up and down along an axis passing through its focal point (the direction of this axis is referred to as the "Z-axis direction"). This makes it possible to change the distance between the radiation generator 22 and the object under test (substrate) held by the substrate holding unit 24, thereby changing the irradiation field and changing the magnification ratio of the transmitted image captured by the detector 26. The position of the radiation generator 22 in the Z-axis direction is detected by the generator position detection unit 23 and output to the control unit 10.

[0027] The detector driving unit 20 also has a driving mechanism such as a motor (not shown), which rotates the detector 26 along the detector rotation orbit 30. The substrate holding unit driving unit 18 also has a driving mechanism such as a motor (not shown), which translates the substrate holding unit 24 on a plane on which the substrate rotation orbit 28 is provided. The substrate holding unit 24 is configured to rotate on the substrate rotation orbit 28 in conjunction with the rotation of the detector 26. This makes it possible to capture a plurality of transmission images with different projection directions and projection angles while changing the relative positional relationship between the substrate held by the substrate holding unit 24 and the radiation generator 22.

[0028] Here, the rotation radii of the substrate rotation orbit 28 and the detector rotation orbit 30 are not fixed but can be freely changed. This makes it possible to arbitrarily change the irradiation angle of the radiation irradiated onto the components placed on the substrate. The orbital planes of the substrate rotation orbit 28 and the detector rotation orbit 30 are perpendicular to the Z-axis direction described above. If the directions perpendicular to this orbital plane are the X-axis direction and the Y-axis direction, the positions of the substrate holder 24 in the X-axis direction and the Y-axis direction are detected by the substrate position detector 29 and output to the control unit 10, and the positions of the detector 26 in the X-axis direction and the Y-axis direction are detected by the detector position detector 31 and output to the control unit 10.

[0029] The control unit 10 controls all operations of the above-mentioned inspection device 1. The functions of the control unit 10 will be explained below with reference to Fig. 2. Although not shown, input devices such as a keyboard and a mouse are connected to the control unit 10 in addition to the above-mentioned monitor 12, and these monitor, keyboard, mouse, etc. constitute the user interface of the inspection device 1.

[0030] The control unit 10 includes a memory unit 34, an imaging processing unit 35, a cross-sectional image generating unit 36, a learning processing unit 37, a substrate inspection surface detecting unit 38, a pseudo-cross-sectional image generating unit 40, and an inspection unit 42. Although not shown, the control unit 10 also includes an imaging control unit that controls the operation of the radiation quality changing unit 14, the radiation generator driving unit 16, the substrate holder driving unit 18, and the detector driving unit 20. Each of these functional blocks is realized by the cooperation of hardware, such as a CPU that performs various arithmetic processing and RAM that is used as a work area for storing data and executing programs, and software. Therefore, these functional blocks can be realized in various ways by combining hardware and software.

[0031] The storage unit 34 stores information such as imaging conditions for capturing a transmission image of the substrate and the design of the substrate to be inspected. The storage unit 34 also stores transmission images and reconstructed images (cross-sectional images, pseudo-cross-sectional images) of the substrate, as well as inspection results from the inspection unit 42 (described later). The storage unit 34 also stores an AI model created or updated by the learning processing unit 37, as will be described in detail later. The storage unit 34 also stores the speed at which the radiation generator driving unit 16 drives the radiation generator 22, the speed at which the substrate holder driving unit 18 drives the substrate holder 24, and the speed at which the detector driving unit 20 drives the detector 26.

[0032] The imaging processing unit 35 drives the radiation generator 22, the substrate holder 24, and the detector 26 using the radiation generator driving unit 16, the substrate holder driving unit 18, and the detector driving unit 20 to capture a transmission image of the object to be inspected held by the substrate holder 24, and generates a reconstructed image from the transmission image. The method of capturing a transmission image and generating a reconstructed image by this imaging processing unit 35 will be described later.

[0033] The cross-sectional image generating unit 36 ​​generates a cross-sectional image based on multiple transmission images acquired from the storage unit 34. This can be achieved using known techniques, such as the FBP method or the maximum likelihood estimation method. Different reconstruction algorithms result in different properties of the resulting reconstructed image and different times required for reconstruction. Therefore, multiple reconstruction algorithms and parameters used for the algorithms may be prepared in advance and the user may select one. This provides the user with the freedom of choice, such as prioritizing a shorter reconstruction time or higher image quality even if it takes longer. The generated cross-sectional image is output to the storage unit 34 and recorded in the storage unit 34. The imaging processing unit 35 can also correct the transmission image or the cross-sectional image using an AI model stored in the storage unit 34. A method for correcting a transmission image or a cross-sectional image will be described later.

[0034] The substrate inspection surface detection unit 38 identifies a position (cross-sectional image) that shows the surface to be inspected on the substrate (for example, the surface of the substrate) from among the multiple cross-sectional images generated by the cross-sectional image generation unit 36. Hereinafter, the cross-sectional image that shows the inspection surface of the substrate will be referred to as an "inspection surface image." A method for detecting the inspection surface image will be described in detail later.

[0035] The pseudo cross-sectional image generating unit 40 images a region of the substrate that is thicker than the cross-sectional image generated by the cross-sectional image generating unit 36 ​​by stacking a predetermined number of consecutive cross-sectional images. The number of cross-sectional images to be stacked is determined based on the thickness of the region of the substrate shown in the cross-sectional image (hereinafter referred to as the "slice thickness") and the slice thickness of the pseudo cross-sectional image. For example, if the slice thickness of the cross-sectional image is 50 μm and the height (e.g., 500 μm) of a BGA solder ball (hereinafter simply referred to as "solder") is to be used as the slice thickness for the pseudo cross-sectional image, then 500 / 50 = 10 cross-sectional images should be stacked. In this case, the inspection surface image identified by the substrate inspection surface detecting unit 38 is used to identify the position of the solder.

[0036] The inspection unit 42 inspects the solder joint state based on the cross-sectional image generated by the cross-sectional image generation unit 36, the inspection surface image identified by the board inspection surface detection unit 38, and the pseudo cross-sectional image generated by the pseudo cross-sectional image generation unit 40. Because the solder joining the board and the component is located near the board inspection surface, it is possible to determine whether the solder is properly joining the board and the component by inspecting the inspection surface image and the cross-sectional image that shows the area on the radiation generator 22 side of the inspection surface image.

[0037] Here, "solder joint condition" refers to whether the board and component are joined by solder and an appropriate conductive path is formed. Inspections of solder joint condition include bridge inspection, melting condition inspection, and void inspection. A "bridge" refers to an undesirable conductive path between conductors created by solder joining. Furthermore, a "melting condition" refers to whether the joint between the board and component is insufficient due to insufficient solder melting, or whether there is a so-called "float" condition. A "void" refers to a defect in the solder joint caused by air bubbles in the solder joint. Therefore, the inspection unit 42 includes a bridge inspection unit 44, a melting condition inspection unit 46, and a void inspection unit 48.

[0038] The operations of the bridge inspection unit 44, molten state inspection unit 46, and void inspection unit 48 will be described in detail below, but the bridge inspection unit 44 and void inspection unit 48 inspect bridges and voids, respectively, based on the pseudo cross-sectional images generated by the pseudo cross-sectional image generation unit 40, and the molten state inspection unit 46 inspects the molten state of the solder based on the inspection surface image identified by the board inspection surface detection unit 38. The inspection results of the bridge inspection unit 44, molten state inspection unit 46, and void inspection unit 48 are recorded in the memory unit 34.

[0039] Fig. 3 is a flowchart showing the flow from capturing a transmission image and generating a reconstructed image, and from identifying an inspection surface image to inspecting the solder joint state. Fig. 4 is a flowchart showing the flow of the processing portion for capturing a transmission image and generating a reconstructed image. The processing in this flowchart starts, for example, when the control unit 10 receives an instruction to start inspection from an input device (not shown).

[0040] 3, when an object to be inspected (inspected object) is carried into the inspection apparatus (step S100), the control unit 10 controls the radiation generator driving unit 16 to set the irradiation field of radiation to be emitted by the radiation generator 22, the substrate holding unit driving unit 18 to move the substrate holding unit 24, and the detector driving unit 20 to move the detector 26 to change the imaging position, while setting the radiation quality of the radiation generator 22 by the radiation quality changing unit 14, irradiating the substrate with radiation to capture transmission images, and then generating reconstructed images (cross-sectional images and pseudo-sectional images) from the multiple transmission images thus captured by the cross-sectional image generating unit 36 ​​and the pseudo-sectional image generating unit 40 (step S120). Note that the movement path of the substrate holding unit 24 by the substrate holding unit driving unit 18 and the movement path of the detector 26 by the detector driving unit 20 when capturing the transmission images are set in advance in the substrate holding unit driving unit 18 and the detector driving unit 20 by reading information stored in the memory unit 34 or by inputting it from an input device. The position of the radiation generator 22 in the Z-axis direction is also set in advance in a similar manner.

[0041] The details of the processing of step S120 will be explained using Figures 4 and 5. As shown in Figure 4, when step S120 starts, the imaging processing unit 35 of the control unit 10 turns on the activation signal output to the substrate holder drive unit 18 and the detector drive unit 20 (step S1000). This corresponds to time t0 in Figure 5(a). When this activation signal is turned on, the substrate holder drive unit 18 starts moving the substrate holder 24 (step S1002), and the detector drive unit 20 starts moving the detector 26 (step S1004). The substrate holder 24 and the detector 26 are moved along the movement path that has been set in advance as described above.

[0042] The imaging processing unit 35 determines whether it is time to capture an image (step S1006), and if it determines that it is not time to capture an image ("N" in step S1006), it repeats this step again after a predetermined time, and if it determines that it is time to capture an image ("Y" in step S1006), it transmits an imaging start signal (trigger) to the detector 26 (step S1008). For example, in the example of FIG. 5(a), the trigger for the detector 26 is turned on at time t1.

[0043] When the detector 26 detects that the trigger has been turned on by the imaging processing unit 35, it starts capturing a transmitted image and transmits a response signal indicating that imaging has started to the imaging processing unit 35 (step S1010). The detector 26 also transmits an exposure signal to the ray generator driving unit 16 (step S1012). For example, in the example of FIG. 5(a), the exposure signal output to the ray generator driving unit 16 is turned on from time t2 to time T. In this way, by configuring the detector 26 to transmit an exposure signal to the ray generator driving unit 16, it is possible to minimize the delay between the start of imaging and the start of exposure.

[0044] Upon receiving the exposure signal from the detector 26, the ray generator driver 16 generates ray from the ray generator 22 while the exposure signal is on, and the object to be inspected is irradiated with the ray (step S1014). If the detector 26 employs a rolling shutter system, information (such as intensity) of X-rays detected by the light receiving elements of the detector 26 is acquired along multiple scan lines arranged in a predetermined direction, with the start times of the scan lines being shifted. For example, as shown in FIG. 5(b), if the detector 26 is configured with n scan lines extending in the left-right direction, the information detected is acquired in the order of L1, L2, L3, . . . , Ln-1, and Ln from top to bottom, with the start times shifted. Therefore, by generating X-rays from the ray generator 22 during the time when all scan lines are acquiring data (during time T in the case of FIG. 5(b)), the information acquired from each scan line is information based on X-rays irradiated at the same time, thereby preventing distortion of the acquired transmission image.

[0045] Furthermore, upon receiving the response signal transmitted from the detector 26, the imaging processing unit 35 acquires position information of the substrate holding unit 24 from the substrate position detection unit 29 and acquires the position of the detector 26 from the detector position detection unit 31 and stores the information (step S1016). Note that the movement of the substrate holding unit 24 by the substrate holding and driving unit 18 and the movement of the detector 26 by the detector driving unit 20 are controlled along a predetermined movement path as described above. Therefore, if either the position of the substrate holding unit 24 or the position of the detector 26 is known, the position of the other can also be determined. Therefore, the positions of both the substrate holding unit 24 and the detector 26 may be stored, or only one of them may be stored. Furthermore, the positions of the substrate holding unit 24 and the detector 26 may be stored in the above-described XY Cartesian coordinate system (in the form of positions (x, y) in the X-axis and Y-axis directions), or may be stored in a polar coordinate system (in the form of positions (r, θ) specified by the distance r from the origin and the angle θ) with the centers of the orbital planes of the substrate rotation orbit 28 and the detector rotation orbit 30 as the origins.

[0046] When the capturing of the transmission image is completed in this manner, the detector 26 transmits the captured transmission image to the imaging processing unit 35 (step S1018). Then, the imaging processing unit 35, which has acquired this transmission image, associates the position information of the substrate holder 24 and the position information of the detector 26 acquired in step S1016 with the acquired transmission image and stores them in the storage unit 34 (step S1020).

[0047] The imaging processor 35 also determines whether there is a next imaging position (step S1022). If it determines that there is a next imaging position ("Y" in step S1022), it returns to step S1006 and repeats the above-described processes (steps S1006 to S1020). On the other hand, if it determines that there is no next imaging position ("N" in step S1022), it turns off the activation signal output to the substrate holder driver 18 and the detector driver 20 (step S1024). Upon detecting that the activation signal has been turned off, the substrate holder driver 18 stops the movement of the substrate holder 24 (step S1026), and the detector driver 20 stops the movement of the detector 26 (step S1028). This corresponds to time t3 in FIG. 5(a), for example.

[0048] Finally, the imaging processing unit 35 generates a reconstructed image from the transmission images stored in the storage unit 34 using the cross-sectional image generating unit 36 ​​and the pseudo cross-sectional image generating unit 40 (step S1030). The generated reconstructed image may be stored in the storage unit 34.

[0049] 3, the control unit 10 performs automatic inspection of the object to be inspected based on the preset information (step S140). Specifically, the board inspection surface detection unit 38 of the control unit 10 receives the transmission image or the reconstructed image (cross-sectional image) from the cross-sectional image generation unit 36 ​​and identifies the inspection surface image from the transmission image or the reconstructed image (cross-sectional image) (step S141). The bridge inspection unit 44 acquires a pseudo cross-sectional image having a slice thickness similar to that of the solder balls, which shows the solder balls, from the pseudo cross-sectional image generation unit 40, and inspects for the presence or absence of a bridge (step S142). If no bridge is detected ("N" in step S143), the molten state inspection unit 46 acquires the inspection surface image from the board inspection surface detection unit 38 and inspects for the presence or absence of the solder melt (step S144). If the solder is molten ("Y" in step S145), the void inspection unit 48 acquires a pseudo cross-sectional image that partially shows the solder ball from the pseudo cross-sectional image generation unit 40 and inspects whether or not voids exist (step S146). If no voids are found ("N" in step S147), the void inspection unit 48 determines that the solder joint state is normal (step S148) and outputs a message to that effect to the storage unit 34. If a bridge is detected ("Y" in step S143), if the solder is not molten ("N" in step S145), or if a void exists ("Y" in step S147), the bridge inspection unit 44, the molten state inspection unit 46, and the void inspection unit 48 each determine that the solder joint state is abnormal (step S149) and output a message to that effect to the storage unit 34. Once the solder state is output to the storage unit 34, the automatic inspection process ends.

[0050] Finally, the control unit 10 outputs the inspection results to the monitor 12 or the like (step S160), removes the object to be inspected (inspected object) (step S180), and ends the inspection according to this flowchart (or starts inspection of the next object to be inspected according to the above-mentioned processing).

[0051] According to the above method, the position at which the transmission image is captured is not information about the time when the imaging processing unit 35 sends a trigger to the detector 20, but information about the time when the detector 26 starts acquiring the image (the time when the response signal is received from the detector 26). When the relative positions of the radiation generator 22, the substrate holding unit 24, and the detector 26 are changing (when the substrate holding unit 24 and the detector 26 are continuously moving), a delay occurs between when the imaging processing unit 35 sends the trigger and when the detector 26 starts acquiring the image. Therefore, the positions of the substrate holding unit 24 and the detector 26 at the time the trigger is sent may be different from the positions at which the transmission image is actually captured. Therefore, as described above, by acquiring the positions of the substrate holding unit 24 and the detector 26 when the detector 26 starts acquiring the image and the imaging processing unit 35 receives the response signal transmitted from the detector 26 at that time, accurate position information can be acquired, thereby improving the accuracy of the reconstructed image. Furthermore, the movement path of the substrate holder 24 by the substrate holder drive unit 18 and the movement path of the detector 26 by the detector drive unit 20 may deviate from the pre-specified position due to the characteristics of the drive units, etc., but as described above, these positions are positions detected by the substrate position detection unit 29 and the detector position detection unit 31, so accurate position information can be obtained, further improving the accuracy of the reconstructed image.

[0052] In addition, the position information and transmission images of the substrate holder 24 and the detector 26 can be stored in a predetermined area of ​​the memory area (memory, hard disk, etc.) of the control unit 10 in a cyclical manner (information is stored sequentially from the beginning of the predetermined area, and when information is stored at the end of the predetermined area, it is stored again from the beginning of the predetermined area), thereby making efficient use of the memory area.

[0053] Furthermore, when the detector 26 captures a transmission image using the rolling shutter method, the image may be distorted if the transmission image is acquired while the relative position of the radiation generator 22, the substrate holder 24, and the detector 26 is being changed. However, as described above, by synchronizing the on / off of the X-rays emitted from the radiation generator 22 (on / off of the exposure signal) with the rolling shutter signal (response signal) of the detector 26, a transmission image without distortion can be acquired.

[0054] As already explained, in the inspection device 1 configured as described above, in order to shorten the inspection time or reduce the amount of radiation irradiated to the object under inspection, it is necessary to shorten the image capturing time of the object under inspection or to weaken the irradiated radiation. Possible methods for shortening the image capturing time include shortening the exposure time (radiation irradiation time), widening the FOV, or reducing the number of transmission images used to generate a reconstructed image. However, shortening the image capturing time or weakening the irradiated radiation in such a manner may increase noise in the transmission images or cross-sectional images extracted from the reconstructed images generated from the transmission images, or may reduce the resolution, which may result in a decrease in the accuracy of the inspection using these images.

[0055] The inspection device 1 according to this embodiment is configured to use an AI model to perform corrections to remove noise from transmission images or cross-sectional images and to increase the resolution of these images. That is, an image (transmission image or cross-sectional image) captured for inspection is used as an input image, and an AI model is applied to this input image to obtain an AI-corrected image as an output image. The correction method using the AI ​​model will be described below.

[0056] The correction method according to this embodiment is configured to perform machine learning such as deep learning (hereinafter referred to as "AI learning") in advance using images (transmission images or cross-sectional images) of an object to be inspected captured by the inspection device 1 as training data and training data to generate an AI model, and then correct the transmission images or cross-sectional images using the generated AI model during inspection. First, the AI ​​learning (deep learning) process will be described with reference to FIGS. 6 to 9. In the following description, a "training image" refers to an image of the object to be inspected that has little noise and is captured with sufficient exposure time and radiation quality (radiation brightness and hardness), or an image of the object to be inspected that has sufficient resolution (resolution that allows a predetermined inspection accuracy to be obtained during inspection), and a "training image" refers to an image of the object to be inspected that contains noise or has low resolution and is captured with the exposure time, radiation quality, or resolution set for the actual inspection.

[0057] In addition, in the inspection device 1 according to this embodiment, the target of correction using the AI ​​model can be either a transmission image or a cross-sectional image. Furthermore, the target of correction can be either noise reduction or resolution enhancement, but here we will explain the case of noise reduction.

[0058] Furthermore, the inspection device 1 according to this embodiment is configured to divide the entire area Ra (e.g., an area of ​​approximately 2000 × 2000 pixels) of a transmission image or cross-sectional image into multiple partial areas Rs (e.g., areas of 256 × 256 pixels), and perform learning or correction on each partial area Rs (using the partial areas Rs at the same positions in the teacher image and the learning image). Furthermore, in the AI ​​learning process, as shown in FIG. 7, 3 / 4 of the entire image area Ra is used for learning, and the remaining 1 / 4 is used for verification (performance evaluation). Note that the arrangement of the verification area and the learning area shown in FIG. 7 is merely an example, and any partial area Rs can be assigned for verification or learning. Furthermore, the ratio of the verification area to the learning area in a single image is not limited to a combination of 3 / 4 and 1 / 4, and can be set as appropriate.

[0059] As already explained, in the inspection of an object under inspection, a plurality of transmission images are taken by changing the relative positions of the radiation generator 22, the substrate holder 24 on which the object under inspection is placed, and the detector 26, so that tens to hundreds of transmission images (sets of teacher images and learning images) are obtained for one FOV, and therefore, when generating an AI model for correcting the transmission images, learning is performed using these multiple transmission images. Furthermore, reconstructed images are generated from these transmission images (reconstructed images of teacher images are generated from the transmission images of the teacher images, and reconstructed images of learning images are generated from the transmission images of the learning images), and hundreds of cross-sectional images (sets of teacher images and learning images) are obtained from these reconstructed images, so when generating an AI model for correcting the cross-sectional images, learning is performed using these multiple cross-sectional images.

[0060] Before learning begins, the object to be learned is placed on the substrate holder 24 of the inspection device 1. Furthermore, an input device or the like is used to input which part of the object to learn from, i.e., the FOV (imaging area) to be learned. Furthermore, an input device or the like is used to input the imaging conditions (radiation quality, exposure time, magnification, etc.) of the teacher image and the imaging conditions of the learning image. Furthermore, either a transmission image or a cross-sectional image is input as the target of the AI ​​model to be output.

[0061] Then, when an instruction to start the learning process is given by an input device or the like, as shown in FIG. 6, the learning processing unit 37 of the control unit 10 arranges the FOV via the imaging processing unit 35, i.e., sets the imaging area on the object to be inspected (step S200).

[0062] When the FOV is positioned in step S200, the learning processing unit 37 captures a teacher image (transmission image) via the imaging processing unit 35 using the exposure time, radiation quality, magnification, etc. that have been input in advance and stores the image in the storage unit 34 (step S202). Similarly, the learning processing unit 37 captures a training image (transmission image) using the exposure time, radiation quality, magnification, etc. that have been input in advance and stores the training image in the storage unit 34 (step S204). As described above, for one FOV, multiple transmission images are captured by changing the relative positions of the radiation generator 22, the substrate holding unit 24 on which the object to be inspected is placed, and the detector 26. Therefore, the substrate holding unit 24 and the detector 26 are moved, and steps S202 and S204 are executed at multiple positions. When AI learning is performed to correct the cross-sectional image, in step S202, a teacher reconstructed image is generated from the captured teacher image (transmission image), and a teacher cross-sectional image is further generated from the reconstructed image and stored in the storage unit 34. Similarly, in step S204, a reconstructed image for learning is generated from the captured learning image (transmission image), and a cross-sectional image for learning is further generated from the reconstructed image and stored in the storage unit .

[0063] Next, the learning processing unit 37 specifies an image set and learning parameters for AI learning (step S206). Here, the learning parameters are information related to the number of learning iterations and the network structure. These learning parameters may be stored in advance in the storage unit 34 or may be input from an input device or the like.

[0064] When the image set for AI learning is performed in step S206, the learning processing unit 37 determines whether to use an existing AI model (step S208). The AI ​​model used for correction may be an AI model specialized for each FOV, or the same AI model may be used for multiple FOVs.

[0065] Here, an AI model may be specialized for a specific object within the FOV (image), such as a BGA, chip resistor, or IC lead, or may be specialized only for solder joints. Specializing only for solder joints allows for a specialized, high-performance model. Furthermore, by identifying the inspection surface image using the above-described process and inspecting the inspection surface image and the cross-sectional image showing the area on the radiation generator 22 side of the inspection surface image, correction by the AI ​​model can be limited to only the area to be inspected, thereby shortening processing time. Note that, since the image is divided into multiple partial regions Rs as shown in FIG. 7, a specialized AI model may be created for each divided region (partial region Rs). On the other hand, by using the same AI model for multiple FOVs, a general-purpose AI model that spans FOVs can be created.

[0066] Therefore, when constructing an AI model specialized for a specific FOV, a new AI model is created without using an existing AI model. When constructing an AI model common to multiple FOVs, an existing AI model is used. Also, when an image of an object to be inspected is divided into multiple FOVs, an AI model specialized for some FOVs may be created for those FOVs, and a common AI model may be created for the remaining FOVs. Whether or not to use an existing AI model may be pre-stored in the storage unit 34 or input via an input device. When an existing AI model is not to be used ("N" in step S208), the learning processing unit 37 creates a new AI model (step S210). When an existing AI model is to be used ("Y" in step S208), the learning processing unit 37 reads the existing AI model from the storage unit 34 (step S212). When storing the AI ​​model in the storage unit 34, information on the FOV corresponding to each AI model is also managed.

[0067] Furthermore, the learning processing unit 37 performs the above-described AI learning using the teacher image and the learning image (step S214). As described with reference to FIG. 7, AI learning is performed on each of the learning partial regions Rs among the divided partial regions Rs. In AI learning, a learning image (which may be a transparent image or a cross-sectional image) is input using a technique such as deep learning, and an AI model is trained to become an image similar to the teacher image (which may be a transparent image or a cross-sectional image, as with the learning image). Then, the learning processing unit 37 updates the AI ​​model based on the results of the AI ​​learning (step S216). The learning processing unit 37 may also be configured to display the progress of the learning on the monitor 12, and if the progress of the learning is being displayed, the display is updated (step S218). The progress includes a performance evaluation of the AI ​​model.

[0068] The performance of an AI model is evaluated by calculating loss. An AI model trained through AI learning is applied to both training and validation images (input images) to generate a corrected image (high-quality output image). The loss is calculated by taking the difference in evaluation value between this generated output image and the training image (for example, the mean squared error of brightness values). Here, loss is the quality that is lost or impaired (or cannot be restored) between the training image and the image corrected by the AI ​​model. Therefore, loss can be expressed by simply comparing brightness values ​​as described above, or it can also be expressed by emphasizing edges or comparing frequency components. Therefore, loss is calculated based on a loss function determined according to the above guidelines.

[0069] Figure 8 is a graph showing the image loss (output image, shown as "training" in Figure 8) when an AI model is applied to training images and the loss in performance evaluation (shown as "evaluation" in Figure 8) versus the number of epochs (number of training iterations). When the loss in the output image is small but the loss in the performance evaluation is large (dashed line in Figure 8), overlearning (a state in which an AI model is overly specialized for the training data and lacks versatility) is considered to be occurring. On the other hand, when the loss in both the output image and the performance evaluation becomes small, learning is considered to be progressing normally. In Figure 9, (a) is the input image (training image), and (b) is the output image corrected by applying an AI model to this input image. If training is proceeding normally, the output image will be similar to the training image (c). However, if overlearning occurs, the output image will be abnormal, as shown in (d). Note that an abnormal state refers to a state in which there are areas with significantly different brightness values ​​or the shape of the captured object is significantly different from its original shape.

[0070] As described above, when the progress of learning is displayed on the monitor 12 or the like in step S218, the learning status can be visually confirmed by displaying the loss graph shown in FIG. 8 or the input image, output image, and teacher image shown in FIG. 9 side by side. As described above, when the loss graph shows that the loss in performance evaluation is large relative to the loss in the output image, or when the output image and the teacher image are significantly different, it can be determined that the learning is abnormal (overlearning), and abnormal learning (wasted time) can be reduced by interrupting the learning. Furthermore, by terminating the learning when the loss becomes small or the output image becomes close to the teacher image, the learning time can be reduced. Shorten The display of the learning progress may include the time that has elapsed since the learning started and the estimated time remaining until the learning is completed.

[0071] Returning to FIG. 6 , the learning processing unit 37 determines whether to perform the next learning (step S220). If there are transmission images or cross-sectional images that have not yet been learned, or if the predetermined number of learning iterations has not been completed ("Y" in step S220), the learning processing unit 37 returns to step S216 and repeats the above-described process. On the other hand, if there are no transmission images or cross-sectional images that have not yet been learned, or if an instruction to end learning is issued from an input device or the like ("N" in step S220), the learning processing unit 37 outputs the AI ​​model to the storage unit 34 ("Y" in step S220) and ends the AI ​​learning process. Note that, as described above, if the AI ​​learning process is interrupted due to a determination that overlearning has occurred ("Aborted" in step S220), the learning processing unit 37 ends the AI ​​learning process without outputting the AI ​​model.

[0072] The above explanation has been about learning an AI model for correction that removes noise, but when learning an AI model for correction that increases resolution, it is necessary to acquire teacher images and learning images with different resolutions. For example, in step S202 of FIG. 6, when acquiring a teacher image (transparent image), the FOV during actual inspection is divided into four, and the images are captured at a high magnification (enlargement ratio), and the four images are combined into one image to be used as a teacher image. The learning image (transparent image) acquired in step S204 is captured with the FOV during actual inspection. This makes it possible to acquire high-resolution teacher images and low-resolution learning images.

[0073] Furthermore, reducing the number of transmission images used to generate a reconstructed image is also effective in shortening imaging time and reducing radiation exposure. Reducing the number of transmission images reduces the resolution of the reconstructed image (cross-sectional image), which may affect the accuracy of the examination. Therefore, by generating pseudo transmission images from the captured transmission images and increasing the number of transmission images, the resolution of the reconstructed image (cross-sectional image) can be increased. Here, the pseudo transmission image can be obtained by generating a transmission image between two adjacent transmission images from the captured transmission images. This type of interpolation is called sinogram interpolation.

[0074] The AI ​​learning process described above can also be applied to sinogram interpolation. When generating an AI model for sinogram interpolation, training images (transmission images) are acquired in step S202 of FIG. 6. At this time, the number of transmission images for one FOV is set to a number that allows the resolution of a reconstructed image (cross-sectional image) generated from those transmission images to achieve a predetermined inspection accuracy. Then, in step S204, instead of capturing training images (transmission images), a predetermined number of transmission images are thinned out from the training images acquired in step S202 and used as training images. For example, if 150 transmission images are acquired as training images in step S202, 75 of the training images (transmission images) are thinned out in step S204, and the remaining 75 are used as training images (transmission images).

[0075] Furthermore, in the AI ​​learning process of steps S214 to S220, an AI model is constructed by learning to generate a transparent image between any two adjacent transparent images of the learning images. For example, when interpolating the Nth transparent image, the N-1th transparent image and the N+1th transparent image are used as input images, and learning is performed to apply the AI ​​model to these input images to obtain the Nth transparent image (pseudo transparent image) as an output image.

[0076] In the above description, the learning conditions, such as whether the correction target of the AI ​​model is a transmission image or a cross-sectional image, whether to remove noise from or increase the resolution of a transmission image or a cross-sectional image, whether to perform sinogram interpolation, whether to learn an image of a specific region of the object to be inspected (a substrate surface or a specific component), whether to create an AI model for each FOV or an AI model common to all FOVs, and the image quality of the teacher image and learning image, and the learning termination conditions, can be set in the control unit 10 via a user interface consisting of the monitor 12 and a keyboard and mouse. In addition, the learning status can be checked and the learning process can be terminated via this user interface.

[0077] When the AI ​​model obtained as described above is applied to an inspection, in step S120 shown in Fig. 3, the AI ​​model described above is applied to the acquired transmission image or reconstructed image (cross-sectional image) to perform correction (noise correction, high resolution, sinogram interpolation, etc.). Below, the process of correcting the transmission image or reconstructed image (cross-sectional image) using the AI ​​model will be described with reference to Fig. 10.

[0078] 10(a) shows a process for applying correction using an AI model to a transmission image. The control unit 10 captures a transmission image of an object to be inspected under imaging conditions for inspection (step S121a), and performs correction on the captured transmission image using the AI ​​model (step S122a). Then, the control unit 10 generates a reconstructed image (cross-sectional image) from the corrected transmission image (step S123a), and ends the transmission image capturing / reconstructed image generation process S120.

[0079] 10(b) shows a process in which correction by the AI ​​model is applied to a reconstructed image (cross-sectional image). The control unit 10 captures a transmission image of the object under inspection under imaging conditions for inspection (step S121b), and generates a reconstructed image (cross-sectional image) from the captured transmission image (step S122b). The control unit 10 then performs correction using the AI ​​model on the generated reconstructed image (cross-sectional image) (step S123b), and ends the transmission image capturing / reconstructed image generation process S120.

[0080] 10(c) shows a process for applying an AI model to only the reconstructed image (cross-sectional image) of a specified region among the reconstructed images (cross-sectional images). The control unit 10 captures a transmission image of the object to be inspected under the imaging conditions for inspection (step S121c), and generates a reconstructed image (cross-sectional image) from the captured transmission image (step S122c). The control unit 10 then identifies the substrate surface (inspection surface image) from the generated reconstructed image (cross-sectional image) using the method described in step S141 or the like (step S123c), and performs correction using the AI ​​model on the identified inspection surface image and the cross-sectional image showing the region on the ray generator 22 side of the inspection surface image (step S124c), thereby completing the transmission image capturing / reconstructed image generation process S120.

[0081] In this case, if an AI model specialized for an FOV has been created, the AI ​​model corresponding to that FOV is applied for each FOV to perform correction. As described above, among multiple FOVs, an AI model specialized for that FOV may be applied to some FOVs, and a common AI model may be applied to some FOVs.

[0082] As described above, by performing corrections using an AI model on the transmitted image or reconstructed image (cross-sectional image) during the examination, even if the irradiated radiation is weakened or the imaging time is shortened, the examination can be performed using the reconstructed image (cross-sectional image / pseudo-cross-sectional image) from which noise has been removed or whose resolution has been increased in the subsequent processing in Figure 3, thereby improving the examination accuracy.

[0083] As described above, in the inspection device according to this embodiment, noise is removed, resolution is increased, or sinograms are interpolated by correction using an AI model, so that high-quality images (transmission images or tomographic images) of the object to be inspected can be acquired even if the radiation dose is reduced or the imaging time is shortened, thereby improving inspection accuracy. Furthermore, because the time required for inspection can be shortened, the overall throughput does not decrease even when the entire area of ​​the object to be inspected is inspected.

[0084] Here, the AI ​​model is trained using images captured by an inspection device equipped with the AI ​​model, which serve as teacher images and training images of the object to be inspected. For example, when the object to be inspected is an electronic circuit board, various types of IC packages (e.g., SOJ (Small Outline J-leaded) and BGA (Ball Grid Array)) are mounted on the board. Therefore, according to the above-described method, AI models specialized for each object can be constructed by training using images captured of the object itself as teacher images and training images, which can improve the quality of images (transmission images and cross-sectional images) during inspection and can be expected to improve inspection accuracy. The effectiveness of learning can be further improved by capturing such training images (teacher images and training images) using an inspection device that applies an AI model and corrects the images during inspection using this AI model.

[0085] In addition, the teacher images and learning images may be acquired by the inspection device 1, and the acquired images may be used to perform learning on a computer other than the control unit 10, and the results obtained from the learning (AI model) may be implemented in the control unit 10 for inspection, or the learning may be performed in the control unit 10 as described above, and the results (AI model) may be used for inspection.

[0086] Furthermore, the above-described AI model learning method and image correction using the generated AI model can be applied not only to X-ray inspection devices, but also to AOI-type inspection devices, specifically, inspection devices that place a light source and a camera serving as a detector above the object under inspection, irradiate the object with light from the light source (such as illumination light with a stripe pattern), and acquire images using the camera. In an inspection device with such a configuration, images with different resolutions can be acquired as training images and training images by changing the magnification of the camera's zoom lens or using pixel shifting. Furthermore, by averaging multiple acquired images to generate a single image, images with different amounts of noise can be used as training images and training images. [Explanation of symbols]

[0087] 1. Inspection equipment 10 Control Unit 22 Radiation generator (light source, radiation source) 24 Board holding part (holding part) 26 detector

Claims

1. A method for creating an AI model in an inspection device that corrects an image of an object to be inspected using an AI model created by machine learning, comprising: a first step of using a high-quality image of the object to be inspected, captured by the inspection device, as a teacher image; a second step of using, as a learning image, an image of the object to be inspected, which has a low image quality and is captured at a timing different from that of the teacher image by the inspection device, and which has the same field of view as that of the teacher image; A third step of configuring the AI ​​model to be created; a fourth step of updating the AI ​​model by dividing each of the teacher image and the learning image into a plurality of partial regions, inputting the learning image of a learning partial region among the partial regions, outputting an image corrected by the AI ​​model, and comparing the output with the teacher image of the learning partial region for learning; a fifth step of inputting the learning image of a verification partial region among the partial regions, outputting an image corrected by the AI ​​model updated in the fourth step by evaluation, and calculating a loss due to evaluation from the difference between the evaluation value of the output by evaluation and the evaluation value of the teacher image of the verification partial region; and a sixth step of repeating the fourth step and the fifth step until a predetermined condition is satisfied. How to create an AI model.

2. The fifth step comprises: A fifth step (5-1) of inputting the learning image used to update the AI ​​model in the fourth step, outputting an image corrected by the AI ​​model updated in the fourth step, and calculating a learning loss from the difference between the evaluation value of the learning output and the evaluation value of the teacher image used to update the AI ​​model in the fourth step; A fifth-2 step in which the learning image of a verification partial region among the partial regions is used as an input, an image corrected by the AI ​​model updated in the fourth step is used as an evaluation output, and a loss due to evaluation is calculated from the difference between the evaluation value of the output by the evaluation and the evaluation value of the teacher image of the verification partial region; and a fifth-third step of displaying on a monitor a graph of the loss due to the learning and the loss due to the evaluation corresponding to each number of times of learning from the start of learning to the present time. The method for creating an AI model according to claim 1.

3. The sixth step receives an operator's decision as to whether or not to terminate the creation of the AI ​​model for the graph, and terminates the creation of the AI ​​model when an input of a decision to terminate is received. The method for creating an AI model according to claim 2.

4. When the image of the object to be inspected is divided into a plurality of imaging regions and acquired, the AI ​​model specialized for each of the imaging regions is created for at least two of the imaging regions. The method for creating an AI model according to any one of claims 1 to 3.

5. When the image of the object to be inspected is divided into a plurality of imaging regions and acquired, the AI ​​model common to the imaging regions is created in at least two imaging regions among the plurality of imaging regions. A method for creating an AI model according to any one of claims 1 to 4.

6. Dividing the image of the object to be inspected into a plurality of partial regions, and creating the AI ​​model specialized for each of the partial regions for at least two of the plurality of partial regions. A method for creating an AI model according to any one of claims 1 to 5.

7. An AI model specialized for an image of a specific region of the object to be inspected is created. A method for creating an AI model according to any one of claims 1 to 6.

8. The image quality is the amount of noise or the resolution A method for creating an AI model according to any one of claims 1 to 7.

9. A light source and a holder for holding the object to be inspected; a detector that detects light emitted from the light source and reflected by or transmitted through the object under inspection to acquire an image of the object under inspection; a control unit; The control unit irradiating the object under inspection with light from the light source, and acquiring the teacher image and the learning image of the same field of view at different times as images of the object under inspection by the detector; The AI ​​model for correcting the image is created by the AI ​​model creation method according to any one of claims 1 to 8 using the teacher image and the learning image. Inspection equipment.

10. a light source that emits radiation; a holder for holding the object to be inspected; a detector that detects radiation emitted from the light source and transmitted through the object under inspection to obtain an image of the object under inspection; a control unit; The control unit changing a relative position of the light source, the holding unit, and the detector, and acquiring, with the detector, radiation emitted from the light source and transmitted through the object under inspection, the teacher image and the learning image of the same field of view as transmitted images of the object under inspection at different times; The AI ​​model for correcting the transmission image is created by the AI ​​model creation method according to any one of claims 1 to 8 using the teacher image and the learning image. Inspection equipment.

11. a light source that emits radiation; a holder for holding the object to be inspected; a detector that detects radiation emitted from the light source and transmitted through the object under inspection to obtain an image of the object under inspection; a control unit; The control unit changing a relative position of the light source, the holder, and the detector, and acquiring, with the detector, radiation emitted from the light source and transmitted through the object under inspection, thereby acquiring transmitted images of the object under inspection for the teacher image and the learning image of the same field of view at different timings; reconstructing the transmission image for the teacher image to generate the teacher image which is a cross-sectional image of the object to be inspected, and reconstructing the transmission image for the learning image to generate the learning image which is a cross-sectional image of the object to be inspected; The AI ​​model for correcting the cross-sectional image is created by the AI ​​model creation method according to any one of claims 1 to 8 using the teacher image and the learning image. Inspection equipment.

12. The control unit changing a relative position of the light source, the holder, and the detector, and acquiring, with the detector, radiation emitted from the light source and transmitted through the object under inspection, thereby acquiring a transmission image of the object under inspection; Correction is performed using the AI ​​model on the acquired transmission image or a cross-sectional image generated from the transmission image. The inspection device according to claim 10 or 11.

13. The correction is performed on the image of the specified region of the object. The inspection device according to any one of claims 9 to 12.

14. having a user interface, The control unit receives the learning conditions through the user interface. The inspection device according to any one of claims 9 to 13.

Images