Inspection method and inspection device
Patent Information
- Application Number
- JP2023204223
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-12-01
- Publication Date
- 2026-09-09
AI Technical Summary
The accuracy of existing inspection methods for determining the normality and abnormality of products relies heavily on the variability of training data, making it challenging to handle diverse abnormal modes without requiring extensive and costly data generation.
An inspection method utilizing an image generator constructed by machine learning to complement missing areas in images, allowing for the determination of normality or abnormality by comparing differences between original and complemented images with a predetermined reference.
This approach reduces labor and cost by enabling efficient determination of normality and abnormality without the need for extensive machine learning on diverse abnormal modes, relying instead on a simple comparison of image differences.
Smart Images
Figure 00000000_0000_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an inspection technique for determining normality and abnormality based on an image of an object to be inspected.
Background Art
[0002] At the time of shipping products manufactured in a factory, an inspection at the time of shipping is performed to check whether there are any abnormalities in the products. In such inspections, for example, an inspector visually checks an image of the product to determine whether there are any defective parts or foreign substances in the product, and determines the normality and abnormality of the product (good or bad determination). However, when an inspector visually inspects images of a large number of products manufactured in a factory, there are problems such as overlooking defective products, high labor costs, and limitations in inspection speed.
[0003] Therefore, attempts have been made to determine the quality of a product to be inspected from an image of the product by using, as an identifier, a learned model constructed by performing machine learning using a large number of images of products for which normality and abnormality have already been determined as teacher data.
Prior Art Documents
Non-Patent Documents
[0004]
Non-Patent Document 1
Non-Patent Document 2
Non-Patent Document 3
Summary of the Invention
Problems to be Solved by the Invention
[0005] Since the abnormal modes that appear in products are not uniform, the accuracy of the discriminator using the trained model constructed by machine learning depends on the variations of the training data used in that machine learning, specifically, the variations of the abnormal parts to be discriminated. In order to handle many abnormalities, machine learning must be performed using images of products with more diverse abnormal modes including them as training data. Generally, most of the products manufactured in factories are normal, and it is difficult to obtain images of products with various abnormal modes. It is possible to manufacture products with intentionally provided various abnormal modes or to generate images of products with abnormalities using dedicated software as described in Non-Patent Document 1, but there is a problem that such work is time-consuming and costly.
[0006] Here, the case of inspecting products manufactured in a factory is taken as an example for explanation, but the same problems as above also exist when determining the normality and abnormality of inspection objects other than industrial products.
[0007] The problem to be solved by the present invention is to reduce the labor and cost of determining whether an inspection object is normal or abnormal.
Means for Solving the Problem
[0008] An inspection method according to one aspect of the present invention made to solve the above problem includes: an image generator preparation step of preparing an image generator that generates an image in which a missing area is complemented from an image having a missing area in part by machine learning; an inspection image preparation step of preparing an inspection target image; a missing image generation step of specifying a part of the inspection target image with a window having a predetermined shape and size, and generating a missing image in which the specified internal area is missing; a complemented image generation step of generating a complemented image in which the internal area of the window is complemented by inputting the missing image into the image generator; a difference acquisition step of obtaining a difference between the inspection target image and the complemented image; a determination step of determining whether the internal area of the window is normal or abnormal by comparing the difference with a predetermined reference and is provided with.
[0009] An inspection apparatus according to another aspect of the present invention made to solve the above problem includes: an image generator storage unit that stores an image generator constructed by machine learning and that generates an image in which a missing area is complemented from an image having a missing area in part; an image storage unit that stores an inspection target image; a missing image generation unit that generates a missing image in which the internal area of a window having a predetermined shape and size is missing from the inspection target image; a complemented image generation unit that generates a complemented image in which the internal area of the window is complemented by inputting the missing image into the image generator; a difference acquisition unit that obtains a difference between the inspection target image and the complemented image; a determination unit that determines whether the internal area of the window is normal or abnormal by comparing the difference with a predetermined reference comprises.
Advantages of the Invention
[0010] In the present invention, an image generator that generates an image in which a missing region is complemented from an image having a missing region in part by machine learning is prepared in advance. The image generator used in the present invention is called an image generation AI and has a function of complementing a missing region by estimating a normal image that should be included in the missing region from the image around the missing region. This image generator is constructed by machine learning using an image of a normal object (preferably an object to be inspected), and complements the missing region of the inspection target image with an image of a normal object (object to be inspected).
[0011] In the present invention, a window having a predetermined shape and size is applied to an image of an object to be inspected (inspection target image) to make the internal region missing, and the image of the missing region is complemented by an image generator. As described above, since the image generator complements the missing region with an image of a normal object to be inspected, when the inspection target image is that of a normal object to be inspected, there is no difference between the inspection target image and the complemented image. On the other hand, when the region to which the window is applied in the inspection target image contains an abnormality, a significant difference appears when the difference (for example, the difference for each pixel value) between the inspection target image and the complemented image is obtained. By extracting such a significant difference that appears according to the predetermined criterion, it is determined that the region to which the window is applied contains an abnormality.
[0012] In the conventional method, since an identifier was used to identify whether or not the inspection target image contains an abnormality, in order to accurately identify the normality and abnormality of the inspection target object, it was necessary to perform machine learning using images of a large number of inspection target objects including various types of abnormalities as teacher data. On the other hand, in the present invention, an image generator that complements the defective area inside the window applied to the inspection target image with an image of a normal inspection target object is used. Whether or not the inspection target image contains an abnormality is determined only by a simple process of comparing the difference between the pixel values of the inspection target image and the pixel values of the complementary image with a predetermined criterion, and an identifier is not required in this process. In the present invention, an image generator constructed by machine learning using an image of a normal object may be used. Therefore, compared with the case of using an identifier that discriminates between the normality and abnormality of the inspection target object from the image of the inspection target object as in the prior art, the image generator can be constructed with less labor and cost, and the normality and abnormality of the inspection target object can be determined.
Brief Description of the Drawings
[0013]
Figure 1
Figure 2
Figure 3
Figure 4
Figure 5
Figure 6
Figure 7
Figure 8
Figure 9
Embodiments for Carrying Out the Invention
[0014] Embodiments of the inspection method and inspection apparatus according to the present invention will be described below with reference to the drawings.
[0015] <Configuration of Inspection System 1> FIG. 1 shows a main configuration of an inspection system 1 including the inspection apparatus of the present embodiment. The inspection system 1 of the present embodiment includes an X-ray CT apparatus 10 and a control / processing apparatus 20. The control / processing apparatus 20 corresponds to an embodiment of the inspection apparatus according to the present invention.
[0016] The X-ray CT apparatus 10 irradiates an object to be inspected with X-rays and captures a tomographic image of the inside thereof. Further, the X-ray CT apparatus 10 can perform dual-energy imaging in which X-rays of a plurality of different energies are irradiated to capture tomographic images at each energy.
[0017] The control / processing apparatus 20 includes a storage unit 30. In the storage unit 30, there are provided an image generator storage unit 31 in which an image generator constructed by a learned model created by machine learning is stored, and an image storage unit 32 in which tomographic images obtained by photographing an object to be inspected with the X-ray CT apparatus 10 are stored. Further, in the storage unit 30, information serving as a criterion for determining whether the object to be inspected is normal or abnormal (in this embodiment, a threshold value of the standard deviation of the pixel values of each pixel of the difference image 65 (see FIG. 5; described later)) is stored. Such a criterion may be determined based on values obtained by using images of one or more normal objects to be inspected as inspection target images and executing a series of processes described later.
[0018] The control and processing device 20 also includes, as functional blocks, a shooting control unit 41, an image preprocessing unit 42, a window setting unit 43, a missing image generation unit 44, a complementary image generation unit 45, a difference image acquisition unit 46, a determination unit 47, and an inspection result display processing unit 48. The control and processing device 20 can be configured by, for example, a general personal computer, and these functional blocks are realized by executing a program of dedicated inspection software 40 installed in advance by a processor. Further, an input unit 51 composed of a keyboard, a mouse, etc. and a display unit 52 composed of a liquid crystal display, etc. are connected to the control and processing device 20.
[0019] In this embodiment, before determining normal or abnormal based on an image of an object to be inspected, an image generator is created and stored in the image generator storage unit 31. The image generator of this embodiment is constructed by a generative adversarial network (GAN), which is a method of unsupervised machine learning that learns features without giving correct data. Since the generative adversarial network itself is a conventionally known method, it will be briefly described here.
[0020] Fig. 2 shows the structure of the generative adversarial network. The generative adversarial network has two neural networks. One is a generator, which is given a vector of random values (latent input) as input and generates data having the same structure as the learning data. The other is a discriminator, which is given a batch of data including observation values of both the learning data and the data generated by the generator, and attempts to classify whether the observation value is real or generated.
[0021] When training with a GAN, usually, the performance of both networks is maximized by training them simultaneously. Specifically, for the generator, by inputting noise corresponding to the type of features of the generated data, machine learning is performed to map this noise to approximate the desired data, generating data such that the discriminator determines it to be real. For the discriminator, machine learning is performed to distinguish the images generated by the generator from real images. In this machine learning, it is preferable to have the generator generate images of the same type as the object to be inspected as the generated data. This enables the creation of a generator that has learned more features of the object to be inspected. In a GAN, since it is only necessary to construct an image generator that generates images of normal objects to be inspected, there is no need to perform machine learning on images of abnormal objects to be inspected as was the case when creating a conventional discriminator. In this embodiment, only the generator generated by the GAN needs to be used as the image generator, and there is no need to perform machine learning on the discriminator. Therefore, in this embodiment, a discriminator with a certain level of discrimination ability may be left as it is (without further machine learning), and only the generator may be machine-learned by the GAN.
[0022] <Inspection method using inspection system 1> Next, the procedure for determining whether an object to be inspected is normal or abnormal using the inspection apparatus of this embodiment will be described with reference to the flowchart of FIG. 3.
[0023] The user first prepares a tomographic image (inspection target image) of the object to be inspected (step 1). If the tomographic image has already been acquired, the tomographic image of the object to be inspected is read from the image storage unit 32. If the tomographic image has not been acquired, the object to be inspected is set at a predetermined position of the X-ray CT apparatus 10 and imaging is instructed. When imaging of the tomographic image is instructed, the imaging control unit 41 irradiates X-rays of a plurality of different energies individually and captures each tomographic image. The captured tomographic images are stored in the image storage unit 32 in association with information (such as an ID number) for specifying the object to be inspected.
[0024] Next, the image preprocessing unit 42 performs predetermined preprocessing on the tomographic image of the object to be inspected (step 2). The preprocessing of the image includes, for example, a process of normalizing the size of the tomographic image, a process of removing artifacts in the image, and the like. The normalization of the size of the tomographic image is a process of making the number of pixels in the vertical and horizontal directions of the captured image match a predetermined number of pixels. It is preferable that this predetermined number of pixels is approximately (preferably the same as) the number of pixels of the image generated by the generator in the machine learning for constructing the image generator used. Thereby, when generating a complementary image (described later), the features learned by the image generator are likely to be reflected. Also, the removal of artifacts in the image is, for example, a process of extracting only the part corresponding to the object to be inspected from the tomographic images obtained by irradiating X-rays of different energies, and removing the artifacts peculiar to the images taken using the X-ray CT apparatus 10. For example, when determining the normality and abnormality of the electrodes of a storage battery, it is a process of distinguishing the electrodes from other members, specifying the electrode part, or removing unnecessary parts from the plurality of tomographic images. Note that the preprocessing of the image is not an essential step, but by performing such preprocessing, a highly accurate generated image can be generated by the generator, and false determination due to artifacts in the image can be prevented.
[0025] Subsequently, the window setting unit 43 sets a window 62 of a predetermined size at the initial position (for example, the upper left corner) of the preprocessed image (preprocessed image 61) (step 3. Refer to the upper left of FIG. 4). The window 62 used in this embodiment is rectangular, and in this embodiment, the lengths of two sides of the window 62 are each 10% of the lengths of two sides of the preprocessed image.
[0026] When the window 62 is set, the missing image generation unit 44 generates a missing image 63 in which the pixel data of the part specified by the window 62 is erased from the preprocessed image 61 (step 4. Refer to FIG. 5).
[0027] When the defective image 63 is generated, the complementary image generation unit 45 uses the image generator stored in the image generator storage unit 31 to generate a complementary image 64 obtained by complementing the image of the defective area of the defective image 63 (step 5; see FIG. 5).
[0028] When the complementary image 64 is generated, the difference image acquisition unit 46 creates a difference image 65 that is the difference between the pre-processed image 61 and the complementary image 64 (step 6; see FIG. 5).
[0029] When the difference image 65 is created, the determination unit 47 calculates the standard deviation of the difference image 65 (step 7). Specifically, for the difference image 65, the difference in pixel values of each pixel within the area where the window is set is obtained, and the standard deviation is calculated. The determination unit 47 compares the calculated standard deviation value with the threshold value stored in the storage unit 30 (step 8). Then, when the standard deviation value is less than or equal to the threshold value (NO in step 8), it is determined that the area to which the window 62 is applied is normal (step 9), and when the standard deviation value exceeds the threshold value (YES in step 8), it is determined that the area to which the window 62 is applied contains an abnormality (step 10).
[0030] After finishing the determination of whether the area to which the window 62 is applied is normal or abnormal, the window setting unit 43 determines whether the window 62 is set at all positions of the pre-processed image 61 (step 11). Here, since the window 62 is only set at the initial position, it is determined that there are remaining positions where the window 62 is not set.
[0031] If it is determined that there is a remaining position where the window 62 is not set, the window setting unit 43 re-sets the window 62 at a position shifted by a predetermined direction and number of pixels (for example, 1 to several pixels in the right direction) from the position where the window 62 was previously set (here, the initial position) (step 12). As shown in FIG. 4, the window setting unit 43 sequentially repeats the setting of the window 62 on the pre-processed image 61 using a method called a sliding window, which is used in, for example, the field of object detection. Note that this is an example when the abnormal location is unknown. When the abnormal location is predicted, instead of moving the window, the user may specify the predicted area and only process the specified area.
[0032] When the window 62 is re-set, the process returns to step 4, and the processes from step 4 to step 11 are executed again in the same manner as above. Then, in step 11, when it is determined that the window 62 is set at all positions, the inspection result display processing unit 48 displays, as an inspection result, an image in which the positions of the windows 62 determined to include abnormalities are superimposed on the pre-processed image 61 based on the determination results for each of the windows 62 set so far on the display unit 52 (step 13).
[0033] <Example 1> Here, as a specific example, the results of inspecting the electrodes of a storage battery will be described. In this example, an image after a test of repeatedly charging and discharging a cylindrical lithium-ion battery (a sheet-like material in which a positive electrode, a separator, and a negative electrode are stacked and wound inside a cylindrical container; 18650-type LIB) was used to inspect the presence or absence of abnormalities in the electrodes. Here too, in the same manner as in the above embodiment, windows were sequentially set on the pre-processed image 61 using a sliding window, and the presence or absence of abnormalities at the positions of each window was determined by the processes of steps 4 to 10 above.
[0034] Figure 6 shows the pre-processed image 61. In this example, a pre-processed image 61 of 700×700 pixels was used, and a square window of 70×70 pixels was applied. The left side of Figure 7 is a partial enlarged view of the tomographic image before the test (Image A), and the right side of Figure 7 is a partial enlarged view of the tomographic image after the test (Image B). An abnormality (a wrinkled state) of the electrode has occurred at the position indicated by the circle in the tomographic image B. Note that in this example, the abnormality of the electrode appears in a wrinkled shape, but the abnormality of the electrode can appear in various forms.
[0035] Figure 8 shows the pre-processed image 61, the missing image 63, the complemented image 64, and the difference image 65 for one of the windows where an abnormality was detected. As described above, since the image generator of the present embodiment is constructed by machine learning that generates an image in which a missing area is complemented from an image having a partially missing area, an area including an abnormal part (the area where the window is set) is made missing, and when the complemented image 64 in which the missing area is complemented by the image generator is generated, an image of a normal electrode is complemented in the missing area. Therefore, when the area to which the window is applied is normal, there is no significant difference between the pre-processed image 61 and the complemented image 64. On the other hand, when the area to which the window is applied includes an abnormality, as shown in the partial enlarged view in Figure 8, a significant difference occurs between the pre-processed image 61 and the complemented image 64, and pixel value variations occur in each pixel of their difference image. Therefore, by calculating the standard deviation representing the pixel value variation of each pixel of the difference image and comparing it with a predetermined threshold value, it is possible to determine whether the area includes an abnormality.
[0036] <Example 2> FIG. 9 is an example of the result of inspecting the presence or absence of electrode abnormalities before and after a test of repeatedly charging and discharging a cylindrical lithium-ion battery, similar to Example 1. Example 1 is the inspection result using a tomographic image of a horizontal cross-section (a cross-section perpendicular to the direction in which the electrodes extend) of the lithium-ion battery, and Example 2 is the inspection result using a tomographic image of a vertical cross-section (a cross-section parallel to the direction in which the electrodes extend and including the central axis of the wound electrodes) of the lithium-ion battery. Also in this Example 2, in the difference image, the locations where electrode abnormalities have occurred are clearly shown, and it can be seen that there is variation (the standard deviation is large) in the pixel values of each pixel.
[0037] As described above, in this embodiment and the examples, a complementary image is generated by an image generator that performs machine learning to generate an image of a normal object (preferably the object to be inspected), and the subsequent determination of normal and abnormal is performed by a simple process of obtaining the standard deviation of the difference image between the pre-processed image (original image) and the complementary image. Therefore, it is not necessary to learn a large number of images of the object to be inspected including abnormalities that appear in various forms, as in the conventional method of constructing a discriminator for discriminating between normal and abnormal of the object to be inspected by machine learning, and the labor and cost of determining whether the object to be inspected is normal or abnormal can be reduced compared to the conventional method.
[0038] The above embodiment and examples are merely examples, and can be appropriately modified in accordance with the gist of the present invention.
[0039] In the above embodiment, a tomographic image of the object to be inspected (such as a storage battery) was acquired using the X-ray CT apparatus 10, and the normality and abnormality of the electrodes of the storage battery were determined based on the tomographic image. However, the type of the image to be inspected is arbitrary, and an appropriate one according to the purpose of the inspection may be used. For example, when performing an appearance inspection, an image obtained by photographing the appearance of the object to be inspected may be used. Also, the object to be inspected is not limited to only objects such as storage batteries. For example, various images such as an image of a human tissue (organ) photographed using the X-ray CT apparatus 10 and an image showing the distribution of a target compound inside a sample obtained by an imaging mass spectrometer can be used as the image to be inspected.
[0040] In the above-described embodiment, an example was described in which dual-energy imaging is performed to remove artifacts from an image and a plurality of images are acquired. However, alternatively, an image may be acquired by an imaging method according to the characteristics of the object to be inspected, such as photon-counting imaging or phase imaging, and the composition of the object to be inspected may be discriminated and the part to be inspected may be specified.
[0041] In the above-described embodiment and examples, the normalcy and abnormality of the region where the window is set are determined by comparing the standard deviation calculated for the difference image of the region where the window is set with a threshold value. However, the determination of the normalcy and abnormality of the region may be performed based on a quantity (feature quantity) representing the features of the difference image of the region, and something other than the standard deviation may be used. For example, the average value, variance value, eigenvalue, representative value of the norm, etc. of the pixel values of each pixel in the difference image of the region where the window is set can be used. Further, in the above-described embodiment and examples, the normalcy and abnormality of the object to be inspected are determined based on the monochrome image of the object to be inspected, but a color image may also be used. When using a color image, RGB values can also be used as pixel values.
[0042] In the above-described embodiments and examples, the case where the image generator generates a difference image using only one complementary image for the missing image has been described. However, in an image generator (so-called generative AI) constructed by machine learning, in many cases, a plurality of images are generated together with their accuracies, and the one with the highest accuracy among them is presented as the generated image. In the above-described embodiment, the difference image is generated using only one complementary image presented by the image generator. However, the accuracy determination by the image generator is not always correct. For example, when generating the complementary image 64 from the missing image 63 in which the peripheral portion of the post-preprocessing image 61 or the image of the inspection object with low periodicity is missing, the accuracy of the complementary image 64 generated by the image generator is low, and thus the accuracy may also be low. In such a case, a difference image 65 may be created for each of all the complementary images 64 (or a predetermined number of complementary images 64 in descending order of accuracy) generated by the image generator, and normality and abnormality may be determined based on the feature amounts (such as the standard deviation in the above-described embodiments and examples) calculated from these plurality of difference images 65. Thereby, even when it is difficult for the image generator to generate a complementary image with high accuracy, abnormalities can be detected more reliably.
[0043] In the above-described embodiment, only one window is used, but a plurality of windows with different sizes may be used. For example, when the abnormal portion included in the inspection object is small, if a large window is used for the abnormal portion, pixel values representing abnormalities appear only in a very small part of the pixels within that window, and no significant difference occurs in the other pixels, and the value of the standard deviation may not exceed the threshold. In such a case, by using a plurality of windows with different sizes, an abnormal portion of any size can be detected by a window with a size adapted to the size of the abnormal portion.
[0044] In the above-described embodiment, the windows are set at a plurality of different positions such that a part of the windows overlaps. However, when the position where an abnormality appears in the inspection object is known in advance, the window may be set only at that location.
[0045] In the above-described embodiment, an image generator created by an adversarial generation network is used. However, the image generator can also be created by other methods, and an image generator created by such a method may be used. For example, an image generator created by a diffusion model, a VAE (Variational AutoEncoder), a flow-based model, an autoregressive model, etc., as described in Non-Patent Document 3, can also be used.
[0046] [Aspect] It is obvious to those skilled in the art that the above-described exemplary embodiments are specific examples of the following aspects.
[0047] (Item 1) An inspection method according to an aspect of the present invention includes an image generator preparation step of preparing, by machine learning, an image generator that generates an image in which a missing region is complemented from an image having a missing region in part; an inspection image preparation step of preparing an inspection target image; a missing image generation step of designating a part of the inspection target image with a window having a predetermined shape and size from the inspection target image, and generating a missing image in which the designated internal region is made missing; a complemented image generation step of generating a complemented image in which the internal region of the window is complemented by inputting the missing image into the image generator; a difference acquisition step of obtaining a difference between the inspection target image and the complemented image; a determination step of determining normality or abnormality of the internal region of the window by comparing the difference with a predetermined reference and comprising.
[0048] (Item 2) An inspection apparatus according to Item 2 includes an image generator storage unit in which an image generator constructed by machine learning and generating an image in which a missing region is complemented from an image having a missing region in part is stored; an image storage unit in which an inspection target image is stored; A defect image generation unit that generates a defect image in which an internal region of a window having a predetermined shape and size is deleted from the inspection target image; A complementary image generation unit that generates a complementary image in which the internal region of the window is complemented by inputting the defect image into the image generator; A difference acquisition unit that obtains a difference between the inspection target image and the complementary image; A determination unit that determines whether the internal region of the window is normal or abnormal by comparing the difference with a predetermined reference; and comprises.
[0049] (Item 3) The inspection apparatus according to Item 3 is the inspection apparatus according to Item 2, wherein the image generator is constituted by a generator used together with a discriminator in adversarial learning.
[0050] In the inspection method according to Item 1 and the inspection apparatus according to Item 2, an image generator that generates an image in which a defective region is complemented from an image having a defective region in part by machine learning is prepared in advance. As such an image generator, for example, one constituted by a generator used together with a discriminator in adversarial learning as used in the inspection apparatus according to Item 3 can be used. The image generator used in the inspection method according to Item 1 and the inspection apparatus according to Item 2 is called an image generation AI, and has a function of complementing a defective region by estimating a normal image that should be included in the defective region included in the image from the image around the defective region. This image generator is constructed by machine learning using images of normal objects (preferably inspection targets), and complements the defective regions of the inspection target image with images of normal objects (inspection targets).
[0051] In the inspection method according to the first aspect and the inspection apparatus according to the second aspect, a window having a predetermined shape and size is applied to an image of an object to be inspected (inspection target image) to delete its internal area, and an image of the deleted area (defect area) is complemented by an image generator. As described above, since the image generator complements the defect area with an image of a normal object to be inspected, when the inspection target image is that of a normal object to be inspected, no difference occurs between the inspection target image and the complemented image. On the other hand, when an abnormality is included in the area of the inspection target image to which the window is applied, when the difference between the inspection target image and the complemented image (for example, the difference for each pixel value) is obtained, a significant difference appears there. By extracting such a significant difference according to the predetermined criterion, it is determined that an abnormality is included in the area to which the window is applied.
[0052] In the conventional method, since an identifier is used to identify whether or not an abnormality is included in the inspection target image, in order to accurately identify the normality and abnormality of the object to be inspected, it is necessary to perform machine learning using a large number of images of the object to be inspected including various types of abnormalities as teacher data. On the other hand, in the inspection method according to the first aspect and the inspection apparatus according to the second aspect, an image generator that complements the defect area inside the window applied to the inspection target image with an image of a normal object to be inspected is used. Whether or not an abnormality is included in the inspection target image is determined only by a simple process of comparing the difference between the pixel values of the inspection target image and the complemented image with a predetermined criterion, and an identifier is not required in this process. In the inspection method according to the first aspect and the inspection apparatus according to the second aspect, an image generator constructed by machine learning using an image of a normal object to be inspected may be used. Therefore, compared with the case of using an identifier that discriminates between the normality and abnormality of the object to be inspected from the image of the object to be inspected as in the prior art, in the inspection method according to the first aspect and the inspection apparatus according to the second aspect, an image generator constructed by machine learning using an image of a normal object can be used, and therefore, compared with the case of using an identifier that discriminates between the normality and abnormality of the object to be inspected from the image of the object to be inspected as in the prior art, the trouble and cost can be reduced to construct the image generator, and the normality and abnormality of the object to be inspected can be determined.
[0053] (Item 4) The inspection apparatus according to claim 4 is the inspection apparatus according to either claim 2 or claim 3, wherein the window is set such that a part of the window overlaps each of a plurality of different positions of the image to be inspected.
[0054] In the inspection apparatus according to claim 4, even if an abnormality is included in any part of the image of the object to be inspected, the abnormality can be detected based on the difference obtained when the window is applied to that position.
[0055] (Claim 5) The inspection apparatus according to claim 5 is the inspection apparatus according to any one of claims 2 to 4, wherein the complementary image generation unit generates a plurality of complementary images for one of the defective images, the difference acquisition unit obtains the difference between the image to be inspected and each of the plurality of complementary images.
[0056] For example, when generating a complementary image from a defective image in which the peripheral portion of the image to be inspected or an image of an inspection object with low periodicity is missing, the accuracy of the image generated by the image generator may be low. In the inspection apparatus according to claim 5, since the difference is obtained for each of the plurality of complementary images generated by the image generator, even when it is difficult for the image generator to generate a highly accurate complementary image, the abnormality can be detected more reliably.
Explanation of Signs
[0057] 1... Inspection system 10... X-ray CT apparatus 20... Control / processing apparatus 30... Storage unit 31... Image generator storage unit 32... Image storage unit 40... Inspection software 41... Imaging control unit 42... Image preprocessing unit 43... Window setting unit 44... Defective image generation unit 45... Complementary image generation unit 46... Difference image acquisition unit 47…Determination unit 48…Inspection result display processing unit 51…Input unit 52…Display unit 61…Pre-processed image 62…Window 63…Defective image 64…Complemented image 65…Difference image
Claims
1. The process includes an image generator preparation step, which involves preparing an image generator that uses machine learning to generate an image in which a missing region has been filled in from an image in which a missing region has been partially filled in, and The process involves preparing the images to be examined, A defect image generation step involves specifying a window of a predetermined shape and size, and generating a defect image by removing the region to which the window is applied from the image to be inspected. A complementary image generation step involves inputting the missing image into the image generator to generate a complementary image in which the region has been filled in, A difference acquisition step to obtain the difference between the image to be inspected and the complementary image, A determination step in which the difference is compared with a predetermined standard to determine whether the region is normal or abnormal. An inspection method that includes the following features.
2. An image generator storage unit stores an image generator constructed using machine learning that generates an image with missing regions from an image that has missing regions in part, An image storage unit where the images to be examined are stored, A defect image generation unit generates a defect image from the aforementioned image to be inspected by removing a region to which a window of a predetermined shape and size has been applied. A complementary image generation unit generates a complementary image by inputting the missing image to the image generator, which then fills in the missing region. A difference acquisition unit that calculates the difference between the image to be inspected and the complementary image, A determination unit that determines whether the region is normal or abnormal by comparing the difference with a predetermined standard. An inspection device equipped with the following features.
3. The inspection apparatus according to claim 2, wherein the image generator is composed of a generator used together with a discriminator in adversarial learning.
4. The inspection apparatus according to claim 2, wherein the window is set so that a portion of the window overlaps each of a plurality of different positions in the image to be inspected.
5. The aforementioned interpolation image generation unit generates multiple interpolation images for one of the missing images, The difference acquisition unit calculates the difference between the image to be inspected and the plurality of complementary images. The inspection apparatus according to claim 2.