JP7499337B2 - Inspection system, learning method, inspection method, and storage medium - Google Patents

Description

The present invention relates to an inspection system, a learning method, an inspection method, and a storage medium.

In recent years, deep learning has been used for visual inspection in the manufacturing industry. Specifically, a trained model based on deep learning is used to perform visual inspection of the inspection target (manufactured parts, etc.).

In this case, deep learning involves unsupervised learning using only images of normal inspection objects. The reason for this is that in manufacturing sites, etc., the number of abnormal inspection objects (inspection objects with external defects such as scratches) is far less than the number of normal inspection objects, so it is not possible to collect many images of abnormal inspection objects. Also, the abnormalities (external defects, etc.) that may occur in the inspection objects are unknown.

An autoencoder (AE: AutoEncorder) is known as a representative example of a trained model that has undergone unsupervised learning. In an autoencoder used in visual inspection, learning is performed using only images of normal inspection objects during learning, so only the features of the images of normal inspection objects are extracted. Therefore, when an image of a normal inspection object is input, the image of the normal inspection object is output, and when an image of an abnormal inspection object is input, an image in which the abnormal part is not restored (which can also be said to be an image of a normal inspection object) is output. In visual inspection using an autoencoder, for example, the difference between the image of the inspection object input to the autoencoder and the image output from the autoencoder is taken, and the presence or absence of an abnormality in the inspection object is determined based on the difference. For example, if the difference is large, it is determined that the inspection object has an abnormality (is an abnormal inspection object).

Other known techniques for visual inspection using autoencoders include the visual inspection device described in Patent Document 1. This visual inspection device learns using pseudo-defect images that combine images of non-defective products with images that show defects.

International Publication No. 2018/225745

In the above-mentioned autoencoder that has been trained using only images of normal test objects, when an image of an abnormal test object is input, an image in which the abnormal part is also restored may be output, even though an image in which the abnormal part is not restored should be output. For example, when the abnormal part is a scratch or a low-contrast part. In this case, the difference between the image input to the autoencoder and the image output from the autoencoder is almost zero, so that what should be judged as an abnormal test object (abnormality present) is erroneously judged as a normal test object (no abnormality), and an accurate abnormality judgment cannot be made for the test object.

On the other hand, in the visual inspection device described in the above-mentioned Patent Document 1, the defects synthesized into the image of a good product often deviate from defects that may actually occur, so it is difficult to say that the device can accurately determine whether the object being inspected is abnormal.

In view of the above-mentioned circumstances, the present invention aims to provide an inspection system, a learning method, an inspection method, and a storage medium capable of accurately determining abnormalities in an object to be inspected.

One aspect of the present invention is an inspection system comprising: a memory for storing a trained model; a processor; an input interface for inputting a processing target image, which is an image of an inspection target, to the processor; and a user interface for outputting a signal in response to a user's input operation, wherein the trained model is a multi-layer neural network having an encoder structure for extracting features of the processing target image by encoding the processing target image, and a decoder structure for generating an image having the same number of pixels and pixel size as the processing target image using the features, and during learning, the trained model is deep-trained to output a normal image when a normal image is input, and to output a second abnormal image when a first abnormal image is input, wherein the normal image is an image of a normal inspection target, the first abnormal image is an image of an inspection target having an abnormal portion, or an image based on the image, and the second abnormal image is an image of the abnormal portion in the first abnormal image, and a processor for detecting an abnormality in the image to be processed according to the detected abnormality, the processor inputs the image to be processed into the trained model during inspection, acquires features extracted by the encoder structure, and determines whether the image to be processed contains an abnormality based on the features; and/or acquires an output image output in response to the input of the image to be processed, calculates a difference in pixel values between the image to be processed and the output image, and detects an abnormality in the image to be processed according to the difference . After detecting the abnormality in the image to be processed, the processor receives as input a signal output by the user interface in response to a user's input operation for a region in the image to be processed that contains at least an undetected abnormality, and performs mask processing in response to the signal to identify the abnormal region in the image to be processed, and generates from the image to be processed a second abnormal image to be used during additional training or re-training of the trained model .

Another aspect of the present invention is an inspection method, the learning method for deep learning a learning model includes a step of deep learning the learning model so as to output a normal image when a normal image is input, and a step of deep learning the learning model so as to output a second abnormal image when a first abnormal image is input, the normal image being an image of a normal inspection target, the first abnormal image being an image of an inspection target having an abnormal portion, or an image based on said image, the second abnormal image being an image in which pixel values of an abnormal region in the first abnormal image, the abnormal region being a region including at least a region corresponding to the abnormal portion, are changed to different pixel values, and a trained model obtained by deep learning the learning model by the learning method is a multi-layer neural network, and an encoder structure for extracting feature quantities of a processing target image, which is an image of an inspection target, by encoding the processing target image. and a decoder structure that uses the feature quantities to generate an image having the same number of pixels and pixel size as the image to be processed, the trained model comprising the steps of: inputting the image to be processed to the trained model; acquiring the feature quantities extracted by the encoder structure and determining whether or not the image to be processed contains an anomaly based on the feature quantities, and/or acquiring an output image in response to the input of the image to be processed, calculating a difference in pixel values between the image to be processed and the output image , and detecting an anomaly in the image to be processed according to the difference; and after detecting the anomaly in the image to be processed, performing a mask process in response to a user's input operation on an area in the image to be processed that contains at least an undetected anomaly, thereby identifying an anomaly area in the image to be processed, and generating from the image to be processed a second anomaly image to be used during additional training or re-training of the trained model .

The present invention makes it possible to accurately determine abnormalities in the test subject.

FIG. 1 is a diagram illustrating a configuration of an inspection system according to an embodiment. FIG. 2 is a diagram illustrating an example of a hardware configuration of an inspection device. 11 is a flowchart illustrating the flow of a learning process performed by a processor. FIG. 13 is a diagram illustrating an example of generation of a second abnormality image performed by the processor. 11 is a flowchart illustrating a flow of an inspection process performed by a processor. FIG. 1 is a schematic diagram (part 1) illustrating an example of an image to be processed and an output image of the trained model when the image to be processed is input. FIG. 2 is a schematic diagram (part 2) illustrating an example of an image to be processed and an output image of the trained model when the image to be processed is input. FIG. 3 is a schematic diagram (part 3) illustrating an example of an image to be processed and an output image of the trained model when the image to be processed is input.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
FIG. 1 is a diagram illustrating a configuration of an inspection system according to an embodiment.
The inspection system 1 illustrated in Figure 1 is a system used for visual inspection of manufactured parts and the like produced at a manufacturing site, and includes an industrial microscope device 10, an inspection device 20 connected to the microscope device 10, and a keyboard 30, a mouse 40, and a display device 50 connected to the inspection device 20.

The microscope device 10 includes a stage 11 on which an inspection object (e.g., manufactured part S) is placed, an objective lens 12 that collects light from the inspection object, a revolver 13 that switches between the objective lenses arranged on the optical path, and an imaging device 14. The imaging device 14 is, for example, a digital camera that includes an image sensor that converts incident observation light into an electrical signal. The imaging device 14 captures an image of the inspection object to generate an inspection object image. The inspection object image is output from the imaging device 14 to the inspection device 20, for example, as an image to be processed.

The inspection device 20 performs an inspection process to inspect the appearance of the inspection object based on the processing target image input from the microscope device 10 (imaging device 14). In this inspection process, a trained model is used to determine whether the inspection object has an abnormality or not, detect an abnormality in the inspection object, etc. In addition, the inspection device 20 also performs a learning process to deep-learn a learning model (pre-learning model) in order to generate a trained model to be used in the inspection process.

The keyboard 30 and mouse 40 input signals corresponding to user input operations to the inspection device 20. The display device 50 is, for example, an LCD (Liquid Crystal Display) and displays the judgment results and detection results of the inspection device 20.

The inspection system 1 may further include a touch panel connected to the inspection device 20. In this case, the touch panel may be provided on the display screen of the display device 50. The inspection system 1 may further include a speaker connected to the inspection device 20. In this case, the speaker may output the judgment results, detection results, etc. of the inspection device 20 by voice.

Figure 2 is a diagram illustrating an example of the hardware configuration of an inspection device. Note that the hardware configuration illustrated in Figure 2 also illustrates the hardware configuration of a computer, and thus the inspection device 20 may be realized by a computer.

The inspection device 20 illustrated in Figure 2 comprises a processor 21, a memory 22, a storage device 23, an input/output interface 24, a portable storage medium drive device 25, and a communication interface 26, each of which is connected to a bus 27 and capable of sending and receiving signals (data) to and from each other.

The processor 21 includes a CPU (Central Processing Unit) and executes an OS (Operating System) program and application programs (including programs for performing inspection and learning processes) to control various processes (including inspection and learning processes) performed by the inspection device 20. The processor 21 may further include a GPU (Graphics Processing Unit) or the like.

The memory 22 includes a RAM (Random Access Memory) and a ROM (Read Only Memory). The RAM temporarily stores parts of the programs executed by the processor 21 and is used as a working memory area for the processor 21. The ROM stores the programs executed by the processor 21, various data required for executing the programs, trained models, etc. The ROM may be a mask ROM, an EPROM (Erasable Programmable ROM), an EEPROM (Electrically EPROM), a flash memory, etc.

The storage device 23 is a hard disk drive (HDD) or a solid state drive (SSD), etc. The storage device 23 may store the program executed by the processor 21, various data required for executing the program, a trained model, etc.

The input/output interface 24 is an interface between an input device and an output device. The input devices are the microscope device 10, the keyboard 30, the mouse 40, a touch panel, etc. The keyboard 30, the mouse 40, the touch panel, etc. are examples of a user interface that outputs a signal in response to a user's input operation. The input/output interface 24, for example, inputs an image to be processed output by the microscope device 10 (imaging device 14) to the processor 21, and inputs a signal output by the user interface to the processor 21. The output devices are the display device 50, a speaker, etc.

The portable storage medium drive device 25 drives the portable storage medium 25a and accesses the stored contents. The portable storage medium 25a may store the program executed by the processor 21, various data required for executing the program, and trained models. The portable storage medium 25a is a memory device, a flexible disk, an optical disk, a magneto-optical disk, etc., and also includes a CD-ROM (Compact Disc Read Only Memory), a DVD (Digital Versatile Disc), a Blu-ray disk, a USB (Universal Serial Bus) memory, etc.

Note that the memory 22, the storage device 23, and the portable storage medium 25a are examples of non-transitory computer-readable storage media. The storage device 23 and the portable storage medium 25a may also be referred to as memories.

The communication interface 26 is connected to a network (not shown) and is an interface for communicating with an external device (such as a server device) connected to the network.

The inspection device 20 is not limited to the hardware configuration illustrated in FIG. 2, but may be configured to include one or more of the components illustrated in FIG. 2, or may be configured to exclude some of the components. For example, the inspection device 20 is not limited to one processor 21, but may include multiple processors. Furthermore, the processing may be performed by one processor, or the processing may be performed by one or more processors simultaneously, sequentially, or in other manners.

In addition, the inspection device 20 may be configured to include hardware such as a microprocessor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a programmable logic device (PLD), a field programmable gate array (FPGA), etc. For example, the processor 21 may be implemented using at least one of these pieces of hardware.

Next, a trained model used in the inspection process performed by the inspection device 20 will be described.
The trained model used in the inspection process is a multi-layer neural network and has an encoder structure and a decoder structure. The encoder structure extracts features of a target image input to the trained model by encoding the target image. The decoder structure uses the features extracted by the encoder structure to generate an image with the same number of pixels and pixel size as the input target image. In other words, the trained model has a structure similar to that of an autoencoder.

In addition, the trained model used in the inspection process is generated, for example, by a learning process performed by the processor 21 of the inspection device 20.
3 is a flowchart illustrating the flow of a learning process performed by a processor. This flowchart also illustrates a learning method.

In the learning process illustrated in Fig. 3, the processor 21 first deep-learns the learning model so that when a normal image is input, the model outputs the input normal image (step S11). Here, the normal image is an image of a normal inspection object, for example, an image of a normal inspection object captured by the imaging device 14 of the microscope device 10. In step S11, deep learning is performed using multiple such normal images.

Next, the processor 21 further performs deep learning on the learning model so that the second abnormal image is output when the first abnormal image is input (step S12). Here, the first abnormal image is an image of an inspection object having an abnormal part, for example, an image of an inspection object having an abnormal part captured by the imaging device 14 of the microscope device 10. Or, the first abnormal image is an image based on an image of an inspection object having an abnormal part, for example, an image obtained by rotating, reducing, enlarging, or the like an image of an inspection object having an abnormal part. An abnormal part is, for example, a scratch part and/or a low contrast part, which is an abnormal part that is difficult to detect by a conventional appearance inspection using an autoencoder. A low contrast part is a part with low contrast relative to the surroundings, or a part with low contrast relative to the corresponding part of a normal inspection object. The second abnormal image is an image in which the pixel values of an abnormal area, which is an area that includes at least an area corresponding to the abnormal part of the inspection object in the first abnormal image, are changed to different pixel values. Changing the pixel values of the abnormal region to different pixel values means, for example, changing the pixel values of the abnormal region to pixel values whose difference from the pixel values is maximum or equal to or greater than a predetermined value, or inverting the luminance of the pixel values of the abnormal region. In step S12, deep learning is performed using multiple pairs of such first and second abnormal images.

By the learning process illustrated in Fig. 3, a trained model to be used in the inspection process is generated. The trained model thus generated is then stored, for example, in the ROM of the memory 22.

The second abnormal image used in step S12 may be generated by the processor 21 in response to, for example, a user's input operation using a user interface (keyboard 30, mouse 40, touch panel, etc.). In this case, the processor 21 may receive a signal corresponding to the user's input operation using the user interface, perform a mask process in response to the signal, identify an abnormal area in the first abnormal image, and generate the second abnormal image by changing the pixel values of the abnormal area to different pixel values.

FIG. 4 is a diagram illustrating an example of generation of a second abnormality image performed by the processor.
4 is an example in which an abnormal area (at least an area 61a corresponding to the scratched area) of a first abnormal image 61, which is an image of an inspection object having a scratched area, is identified by a mask process (user annotation) in response to a user's input operation using a user interface, and the pixel values of the abnormal area are inverted in luminance to generate a second abnormal image 62. Due to this luminance inversion, the second abnormal image 62 becomes an image in which the abnormal area of the first abnormal image 61 is emphasized.

Next, the inspection process performed by the processor 21 of the inspection device 20 will be described.
5 is a flowchart illustrating the flow of the inspection process performed by the processor. Note that this inspection process also illustrates an inspection method.

In the inspection process illustrated in Figure 5, the processor 21 first acquires an inspection object image, which is an image of the inspection object captured by the imaging device 14, from the microscope device 10 as an image to be processed (step S21).

Next, the processor 21 reads out a trained model stored, for example, in the ROM of the memory 22, and inputs the image to be processed obtained in step S21 into the trained model (step S22).

Next, the processor 21 performs an anomaly determination process and/or anomaly detection process (step S23). In the anomaly determination process, the processor 21 acquires the feature of the processing target image extracted by the encoder structure of the trained model, and determines whether the processing target image contains an anomaly based on the feature. In this determination, for example, the difference between the acquired feature and the reference feature may be calculated, and whether the processing target image contains an anomaly may be determined depending on whether the difference is equal to or greater than a predetermined threshold. The reference feature is a feature that can be extracted by the encoder structure when an image of a normal inspection target is input to the trained model, and is also a feature that can be used to determine that the processing target image does not contain an anomaly.

In addition, in step S23, the anomaly detection process obtains an output image output from the trained model in response to the input of the image to be processed in step S22, calculates the difference in pixel values between the image to be processed and the output image, and detects an anomaly (abnormal area) in the image to be processed according to the difference. In this detection, an area where the difference in pixel values is equal to or greater than a predetermined threshold may be detected as an anomaly. The difference in pixel values may also be calculated by obtaining a difference image between the image to be processed and the output image.

Next, the processor 21 displays the results of the abnormality determination process and/or abnormality detection process in step S23 on the display device 50 (step S24). Specifically, the display of the results of the abnormality determination process displays the result of the determination as to whether or not the image to be processed contains an abnormality (i.e., the result of the determination as to whether or not the inspection target contains an abnormality when the inspection target image set as the image to be processed is captured). The display of the results of the abnormality detection process displays the result of the detection of an abnormality in the image to be processed (i.e., the result of the detection of an abnormality in the inspection target when the inspection target image set as the image to be processed is captured). In this display, if no abnormal area is detected in the image to be processed, a message to that effect is displayed, and if an abnormal area is detected, the abnormal area may be displayed so as to be distinguishable from other areas.

6, 7, and 8 are diagrams illustrating schematic examples of a processing target image and an output image of the trained model when the processing target image is input.
6 illustrates a processing target image 71, which is an inspection target image obtained by capturing an image of a normal inspection target, and an output image 73 of a trained model 72 when the processing target image 71 is input. In this case, the output image 73 is the same image as the processing target image 71, and therefore there is no region in which the difference in pixel values with respect to the processing target image 71 is equal to or greater than a predetermined threshold, and no abnormality in the processing target image 71 is detected.

On the other hand, FIG. 7 illustrates a processing target image 74, which is an image of an inspection target (abnormal inspection target) having a scratch, and an output image 75 of the trained model 72 when the processing target image 74 is input. In this case, the output image 75 is an image in which an area (abnormal area) including at least an area 74a corresponding to the scratch in the processing target image 74 is emphasized, so that an area exists in which the difference in pixel value with respect to the processing target image 74 is equal to or greater than a predetermined threshold, and an abnormality in the processing target image 74 is detected. Note that the difference in pixel value in this case is calculated, for example, by obtaining a difference image 76 between the processing target image 74 and the output image 75.

8 illustrates an example of a processing target image 77, which is an image of an inspection target having a low contrast area (an abnormal inspection target), and an output image 78 of the trained model 72 when the processing target image 77 is input. In this case, the output image 78 is an image in which an area (an abnormal area) in the processing target image 77 that includes at least an area 77a corresponding to the low contrast area is emphasized, and therefore there is an area in which the difference in pixel value with respect to the processing target image 77 is equal to or greater than a predetermined threshold, and an abnormality in the processing target image 77 is detected.

As described above, according to the inspection system 1, for example, by performing visual inspection of the inspection object by the inspection process illustrated in Figure 5 using a trained model that has been deep-learned by the learning process illustrated in Figure 3, it becomes possible to detect abnormalities (e.g., scratches and low-contrast areas) that are difficult to detect by visual inspection using conventional autoencoders, thereby enabling accurate detection of abnormalities in the inspection object.

The inspection system 1 may be configured as follows.
For example, in step S23 of the inspection process illustrated in FIG. 5, the processor 21 may generate a second abnormal image to be used when additionally learning or re-learning the trained model from the processing target image (corresponding to the first abnormal image) having the undetected abnormal part when the user checks the processing result displayed in S24 for the processing target image on which the abnormality detection process has been performed and an undetected abnormal part is found. In this case, the processor 21 may identify an area including at least an undetected abnormal part in the processing target image as an abnormal area by masking in response to an input operation by the user using a user interface, and may generate a second abnormal image by changing the pixel value of the abnormal area to a different pixel value. In this case, the pixel value may be changed to a pixel value whose difference from the pixel value is maximum or equal to or greater than a predetermined value, as described in step S12 of FIG. 3, or the pixel value of the abnormal area may be inverted in luminance. The inspection performance is improved by the additional learning or re-learning.

In addition, the trained model used in the inspection process performed by the processor 21 may be generated by an external device (such as a server device). In this case, the trained model generated by an external device that executes a process similar to the learning process illustrated in FIG. 3 may be stored in the ROM of the memory 22 via a network and communication interface 26 (not shown). In this case, the second abnormal image as illustrated in FIG. 4 may also be generated by the external device.

In addition, the inspection system 1 is not limited to use in the manufacturing field, and may be used in fields other than the manufacturing field, such as biology and medicine. For example, when used in the biology field, the inspection system 1 may include a biological microscope device equipped with an imaging device instead of the industrial microscope device 10.

Furthermore, the inspection system 1 is not limited to the configuration illustrated in FIG. 1, and as long as it includes at least the inspection device 20, it may be configured without some of the devices, or may be configured to further include other devices.

As mentioned above, the present invention is not limited to the above-described embodiment, and in the implementation stage, the components can be modified and embodied without departing from the gist of the invention. Furthermore, various inventions can be formed by appropriate combinations of multiple components disclosed in the above-described embodiment. For example, some of the components shown in the embodiment may be deleted. Furthermore, components from different embodiments may be appropriately combined.

1 Inspection system 10 Microscope device 11 Stage 12 Objective lens 13 Revolver 14 Imaging device 20 Inspection device 21 Processor 22 Memory 23 Storage device 24 Input/output interface 25 Portable storage medium drive device 25a Portable storage medium 26 Communication interface 27 Bus 30 Keyboard 40 Mouse 50 Display device 61 First abnormal image 61a Area 62 Second abnormal image 71 Processing target image 72 Learned model 73 Output image 74 Processing target image 74a Area 75 Output image 76 Difference image 77 Processing target image 77a Area 78 Output image

Claims (9)

A memory for storing a trained model;
A processor;
an input interface for inputting a processing target image, which is an image of an inspection target, to the processor;
a user interface that outputs a signal in response to an input operation by a user;
Equipped with
The trained model is
1. A multi-layer neural network, comprising:
an encoder structure for extracting features of the processing target image by encoding the processing target image;
a decoder structure for generating an image having the same number of pixels and the same pixel size as the processing target image by using the feature amount;
When studying,
Deep learning is performed so that, when a normal image is input, the normal image is output;
A trained model that is deep-trained to output a second abnormal image when a first abnormal image is input,
The normal image is an image of a normal test subject,
The first abnormality image is an image of an inspection object having an abnormality or an image based on the image,
The second abnormal image is an image in which pixel values of an abnormal region in the first abnormal image, which is a region including at least a region corresponding to the abnormal portion, are changed to different pixel values,
The processor,
During the inspection,
The processing target image is input to the trained model;
acquiring a feature amount extracted by the encoder structure, and determining whether or not the target image includes an abnormality based on the feature amount; and/or acquiring an output image output in response to an input of the target image, calculating a difference in pixel values between the target image and the output image, and detecting an abnormality in the target image based on the difference amount ;
After detecting an anomaly in the processing target image, the processor receives a signal output by the user interface in response to a user's input operation on an area in the processing target image that includes at least an undetected anomaly, and performs mask processing in response to the signal to identify an abnormal area in the processing target image, and generates a second abnormal image from the processing target image to be used during additional learning or re-learning of the trained model.
1. An inspection system comprising: The inspection system according to claim 1, characterized in that the processor detects an area where the degree of difference is equal to or greater than a predetermined threshold as an abnormality. The inspection system of claim 1, characterized in that the processor calculates the degree of difference by obtaining a difference image between the image to be processed and the output image. The inspection system according to claim 1, characterized in that the second abnormal image is an image in which the pixel values of the abnormal region in the first abnormal image are changed to pixel values whose difference from the pixel values is maximum or equal to or greater than a predetermined value. The inspection system according to claim 1, characterized in that the second abnormal image is an image in which the pixel values of the abnormal area in the first abnormal image are inverted in brightness. The inspection system according to claim 1, characterized in that the abnormality is a scratch. The inspection system according to claim 1, characterized in that the abnormal area is a low-contrast area. The microscope device further includes an imaging device.
2. The inspection system according to claim 1, wherein the normal image, the first abnormal image which is an image of an inspection object having the abnormal area, and the processing object image are images acquired by the microscope device. The learning method for deep learning the learning model is
A step of deep learning the learning model so as to output a normal image when a normal image is input;
A step of deep learning the learning model so as to output a second abnormal image when a first abnormal image is input;
Including,
The normal image is an image of a normal test subject,
The first abnormality image is an image of an inspection object having an abnormality or an image based on the image,
The second abnormal image is an image in which pixel values of an abnormal region in the first abnormal image, which is a region including at least a region corresponding to the abnormal portion, are changed to different pixel values,
A trained model obtained by deep learning the learning model by the learning method,
1. A multi-layer neural network, comprising:
an encoder structure for extracting features of a processing target image, the processing target image being an image of an inspection target, by encoding the processing target image;
a decoder structure for generating an image having the same number of pixels and pixel size as the target image using the feature amount;
having
Inputting the target image to the trained model;
acquiring features extracted by the encoder structure, and determining whether or not the image to be processed includes an abnormality based on the features; and/or acquiring an output image in response to an input of the image to be processed, calculating a degree of difference in pixel values between the image to be processed and the output image, and detecting an abnormality in the image to be processed based on the degree of difference ;
After detecting an anomaly in the processing target image, a mask process is performed in response to a user's input operation on an area in the processing target image that includes at least an undetected anomaly, thereby identifying an anomaly area in the processing target image, and generating a second anomaly image from the processing target image to be used during additional learning or re-learning of the trained model;
An inspection method comprising:

Images