JP7793285B2 - Abnormality judgment model generation method, abnormality judgment model generation device, and inspection device - Google Patents

Description

The present invention relates to an anomaly detection model generation method, an anomaly detection model generation device, and an inspection device.

In recent years, AI (Artificial Intelligence) technologies such as deep learning have advanced, and the use of AI in various inspections is being considered. Common methods of automating inspections using AI technology include anomaly detection using supervised machine learning and unsupervised machine learning.

When using supervised machine learning to inspect for external abnormalities, it is necessary to prepare a large amount of training image data in advance and label them as normal or abnormal, or by level of abnormality, before conducting the machine learning. If there is insufficient training image data, unknown abnormalities cannot be detected and effective learning cannot be carried out. For this reason, new training image data is generated by processing training image data of actual abnormalities, thereby increasing the amount of training image data.

For example, Patent Document 1 proposes technology for generating high-quality teacher image data when there is an insufficient amount of teacher image data to use in machine learning. Specifically, Patent Document 1 discloses a machine learning system that acquires teacher image data classified into multiple patterns, identifies a missing pattern that has a small number of teacher image data items based on predetermined criteria, and generates new teacher image data that belongs to the missing pattern by spatially inverting certain teacher image data or changing the color tone.

Japanese Patent Application Laid-Open No. 2018-169672

The machine learning system disclosed in Patent Document 1 increases the number of pieces of teacher image data by converting the color tones of the actual teacher image data. It is also possible to increase the number of pieces of teacher image data not only by converting color tones, but also by inverting or linearly converting the actual teacher image data. However, when increasing the number of pieces of teacher image data by converting color tones or linearly converting them, there are limitations to the content of the teacher image data because the newly added teacher image data is created based on the characteristics of the original teacher image data, and there is a risk that the accuracy of abnormality detection using machine learning will decrease.

The present invention was made in consideration of the above-mentioned problems, and aims to provide an anomaly detection model generation method, an anomaly detection model generation device, and an inspection device that can improve the accuracy of anomaly detection.

In order to solve the above-mentioned problems, one aspect of the present invention provides an anomaly detection model generation method that generates an anomaly detection model for detecting abnormalities in an inspection object using inspection image data obtained by measuring the inspection object as input data, the anomaly detection model generation method including the steps of generating a training dataset for generating the anomaly detection model based on pseudo data created by simulating characteristic patterns of abnormalities, and generating the anomaly detection model using the training dataset.

In order to solve the above-mentioned problems, according to another aspect of the present invention, there is provided an anomaly determination model generation device that generates an anomaly determination model for determining abnormalities in an inspection object using inspection image data obtained by measuring the inspection object as input data, and that includes a learning unit that generates the anomaly determination model using a learning dataset generated based on pseudo data created by simulating characteristic patterns of abnormalities.

Furthermore, in order to solve the above-mentioned problems, according to yet another aspect of the present invention, there is provided an inspection device that inspects an inspection object based on inspection image data obtained by measuring the inspection object, the inspection device including: a storage unit that stores an abnormality determination model generated using a learning dataset that is generated based on pseudo data created by simulating characteristic patterns of abnormalities; and a determination unit that inputs the inspection image data obtained by measuring the inspection object into the abnormality determination model and determines whether or not the inspection object has an abnormality.

As described above, according to the present invention, even when there is little teacher image data of abnormalities, it is possible to add new teacher image data of various types, thereby improving the reliability of inspections of abnormalities that appear externally.

1 is a schematic diagram illustrating a configuration example of an inspection system including an inspection device according to an embodiment of the present invention. FIG. 2 is a block diagram showing a functional configuration of the inspection device according to the embodiment. 4 is a flowchart illustrating an example of an abnormality determination model generation method according to the embodiment. FIG. 1 is an explanatory diagram showing a method for creating a learning dataset. 10 is a flowchart illustrating an example of an abnormality determination method using the inspection device according to the embodiment. FIG. 10 is a schematic diagram showing an application example of an inspection system including the inspection device according to the embodiment. FIG. 10 is a schematic diagram showing another application example of the inspection system including the inspection device according to the embodiment.

Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. Note that in this specification and drawings, components having substantially the same functional configuration will be designated by the same reference numerals, and redundant explanations will be omitted.

<1. Overview of the inspection device>
First, an overview of an inspection device according to an embodiment of the present invention will be described.

Figure 1 is a schematic diagram showing an example configuration of an inspection system 100 including an inspection device 20 according to this embodiment. The inspection system 100 comprises the inspection device 20, a detector 11, an input unit 15, and a display unit 13. The inspection system 100 is constructed as a system that automatically performs an abnormality determination for an inspection target using an abnormality determination model based on image data of the inspection target.

Detector 11 generates image data by measuring the inspection object and outputs the image data to inspection device 20. Examples of detector 11 include an imaging camera or LiDAR (Light Detection and Ranging or Light Imaging Detection and Ranging). An imaging camera is equipped with an imaging element such as a CCD (Charge Coupled Device) or CMOS (Complementary Metal Oxide Semiconductor) and generates image data of the object being photographed. LiDAR generates image data that reflects the external appearance of the object by scanning a laser beam onto the object and observing the scattered and reflected light of the laser beam.

The detector 11 may be any device capable of generating image data that reflects the external appearance of the object being measured, and is not limited to an imaging camera or LiDAR. Furthermore, the number of detectors 11 provided in the inspection system 100 is not limited to one, but may be two or more. The detector 11 may include multiple types of devices, such as an imaging camera and LiDAR.

The input unit 15 accepts operation inputs from the user and transmits operation signals to the inspection device 20. For example, the input unit 15 includes at least one input means such as an operation button or operation switch that accepts operation inputs from the user, and is configured, for example, by a keyboard, touch panel, tablet computer, smartphone, etc.

The display unit 13 is driven based on a drive signal output from the inspection device 20, and displays predetermined information such as inspection results. For example, the display unit 13 is configured as a display device equipped with a liquid crystal panel or the like. The input unit 15 and the display unit 13 may be integrated and configured as, for example, a touch panel.

The inspection device 20 includes a control unit 30, a memory 21, and a storage device 23. Part or all of the control unit 30 is configured with a processor such as a CPU (Central Processing Unit). In particular, the control unit 30 may include a GPU (Graphic Processing Unit) in addition to the CPU. In addition, the inspection device 20 includes interfaces for sending and receiving signals to and from the detector 11, input unit 15, and display unit 13.

Memory 21 is composed of at least one storage element, such as RAM (Random Access Memory) or ROM (Read Only Memory). Storage device 23 is composed of at least one storage medium, such as an HDD (Hard Disk Drive), CD (Compact Disc), DVD (Digital Versatile Disc), SSD (Solid State Drive), USB (Universal Serial Bus) flash, or storage device. Memory 21 or storage device 23 functions as a storage unit, storing programs for arithmetic processing executed by the processor, various calculation parameters used in arithmetic processing, acquired data, calculation results, etc. Memory 21 or storage device 23 also stores an anomaly determination model used in testing.

2. General configuration of inspection device
So far, we have explained the overall configuration of the inspection system 100. Next, we will specifically explain an example of the schematic configuration of the inspection device 20.

Figure 2 is a block diagram showing the functional configuration of the inspection device 20. The control unit 30 of the inspection device 20 includes an image processing unit 31, a calculation unit 33, and a display control unit 39. The calculation unit 33 includes a learning unit 35 and a determination unit 37. Specifically, each function of the control unit 30 may be a software function realized by a processor executing a program, or some functions may be realized by hardware functions.

(2-1. Image Processing Section)
The image processing unit 31 processes the image data output from the detector 11 to generate inspection image data to be used for inspection. For example, the image processing unit 31 may perform a process of adjusting the contrast of the image data output from the detector 11 using at least one of a dilation filter, a contraction filter, an averaging filter, and a median filter, or may perform a process of adjusting the contrast or emphasizing the edges using an edge extraction filter or an edge enhancement filter. The image processing unit 31 may also perform a trimming process to extract image data of a predetermined size or range from the image data output from the detector 11. The function of the image processing unit 31 may be provided in the detector 11.

(2-2. Arithmetic section)
The calculation unit 33 includes a learning unit 35 that generates an abnormality determination model used to determine whether an object to be inspected is abnormal, and a determination unit 37 that uses the generated abnormality determination model to determine whether an object to be inspected is abnormal.

The learning unit 35 generates an anomaly determination model to be used for anomaly determination. Specifically, the learning unit 35 performs machine learning of the anomaly determination model using image data (learning data) representing the characteristic patterns of anomalies in the inspection target as input data. The generated anomaly determination model is stored in the memory 21 or the storage device 23 and is used in the anomaly determination process by the determination unit 37. In this embodiment, the inspection device 20 including the learning unit 35 also functions as an anomaly determination model generation device.

In this embodiment, the training data used is not actual image data obtained by actually measuring an inspection object in which an abnormality has occurred, but pseudo data created by simulating characteristic patterns of abnormalities that may occur in the inspection object. The pseudo data is created, for example, by artificially drawing characteristic patterns of abnormalities that may occur, using as reference the characteristic patterns of scratches that actually occur, or based on the user's past knowledge or empirical rules. The training data may also include pseudo data generated by randomly performing image conversion processing based on pseudo data created by artificially drawing characteristic patterns of abnormalities. This makes it possible to generate an abnormality determination model before an abnormality actually occurs when building a new inspection system 100 or when a new inspection object is installed. For example, a random image generation algorithm, such as GAN (Generative Adversarial Network), can be used as this image conversion processing. However, the image generation algorithm is not limited to GAN.

Furthermore, when determining anomalies, if it is desired to determine not only the presence or absence of anomalies but also the type of anomaly, the pseudo data can be classified by type of anomaly when artificially drawing the characteristic patterns of the anomaly to create pseudo data. For example, if the characteristic patterns that appear differ depending on the type of anomaly, such as when scratches, breakage, or cracks occur on the surface of the object being inspected, the pseudo data can be classified by type of anomaly.

The learning unit 35 uses the pseudo-data as learning data and generates an anomaly detection model using a machine learning model. The machine learning model may be any existing machine learning model that can be used for machine learning using image data as training data. For example, the machine learning model may be a computational model that uses linear regression, a filter such as a Kalman filter, a support vector machine, a random forest, a nearest neighbor method, a neural network such as deep learning, or a Bayesian network.

Furthermore, the learning unit 35 may update the abnormality determination model based on the inspection image data generated by the image processing unit 31 and information on the abnormality determination results obtained by the determination unit 37, which will be described later. This allows the abnormality determination model, which was generated using pseudo data created by artificially drawing abnormality characteristic patterns, to be automatically corrected based on the normality/abnormality determination results of actually obtained inspection image data. Therefore, the accuracy of the inspection results using the abnormality determination model can be improved.

The judgment unit 37 uses the anomaly judgment model generated by the learning unit 35 and stored in the memory 21 or storage device 23 to judge whether the inspection object has an abnormality based on the inspection image data generated by the image processing unit 31. Specifically, the judgment unit 37 inputs the inspection image data into the anomaly judgment model and determines whether the inspection object has an abnormality based on the output from the anomaly judgment model. For example, the anomaly judgment model calculates the probability that the inspection image data contains a feature pattern corresponding to an abnormal feature pattern contained in the pseudo data input as learning data. The judgment unit 37 judges that an abnormality exists if the probability output from the anomaly judgment model is equal to or greater than a predetermined threshold.

Note that the method of determining an abnormality using the anomaly determination model is not limited to the above example. Furthermore, if the anomaly determination model requires not only the probability that an abnormality has occurred but also a classification of the type of abnormality, the determination unit 37 not only determines whether or not an abnormality exists in the test object, but also determines the type of abnormality.

(2-3. Display control unit)
The display control unit 39 generates an image signal to be displayed on the display unit 13 and controls the display on the display unit 13. The display control unit 39 causes at least the processing results of the learning unit 35 and the determination unit 37 to be displayed on the display unit 13. For example, when machine learning is performed by the learning unit 35, the display control unit 39 causes the display unit 13 to display a setting screen for machine learning conditions and a progress status. Furthermore, when an abnormality determination process is performed by the determination unit 37, the display control unit 39 causes the display unit 13 to display the determination result. However, the display content on the display unit 13 is not particularly limited.

<3. Method for generating an abnormality determination model>
A method for generating an anomaly determination model will be described.

FIG. 3 is a flowchart showing an example of a method for generating an abnormality determination model.
First, the user creates pseudo data by artificially drawing characteristic patterns of abnormalities that may occur in the inspection object (step S11). For example, the user generates pseudo data by drawing characteristic patterns of abnormalities that the user can imagine as abnormalities that may occur in the inspection object. The characteristic patterns of possible abnormalities may be drawn, for example, based on empirical rules. Alternatively, depending on the type of abnormality, they may be randomly drawn characteristic patterns. The target of abnormality detection by the inspection device 20 according to this embodiment is an abnormality that appears in the appearance of the inspection object. Specifically, abnormalities that are the target of abnormality detection include at least one of, for example, scratches, abrasions, breaks, deformations, and adhesion of contaminants on the surface of the inspection object. The user generates pseudo data by drawing patterns of possible abnormality shapes, taking into account the usage environment and conditions of the inspection object, the shape patterns of abnormalities that have occurred in the past, and so on.

For example, a user may use graphics drawing software to draw a characteristic pattern of an abnormality and perform data conversion processing to generate pseudo data, or a user may generate pseudo data by using data creation software to read a handwritten characteristic pattern of an abnormality. There are no particular limitations on the method of artificially drawing a characteristic pattern of an abnormality to create pseudo data. Furthermore, if the characteristic pattern of the generated pseudo data is drawn too specifically, overfitting may occur during machine learning. For this reason, it is preferable that the drawing pattern of the generated pseudo data is a simplified one, rather than a detailed depiction of the shape of the abnormality.

Next, the user randomly performs image transformation processing based on the pseudo data created in step S11 to add new pseudo data and generate a dataset for machine learning (step S13). For example, the user uses a device capable of executing a random image generation algorithm such as GAN to perform color transformation, linear transformation, etc. on the pseudo data created in step S11 to create further pseudo data. If the number of pseudo data is too small, overfitting may occur, making it impossible to respond to characteristic patterns of unknown abnormalities. For this reason, it is preferable to compile an appropriate amount of learning data while verifying the effectiveness of the anomaly detection model by using image data obtained by actually measuring an inspection target in an abnormal state as verification image data.

FIG. 4 is an explanatory diagram showing a method for creating a data set for machine learning.
For example, when generating an abnormality determination model for inspecting scratches that occur on the surface of an object to be inspected, when a scratch that occurs on the surface of the object to be inspected is actually measured, image data Img_act having a feature pattern such as that shown in the upper part of Figure 4 is obtained.

In the example shown in Figure 4, the user creates pseudo data Img_fak, which depicts characteristic patterns of scratches that may occur, using as reference characteristic patterns of scratches that actually occur, or based on the user's past knowledge or experience. Furthermore, based on the created pseudo data Img_fak, the user randomly performs image conversion processing to add pseudo data, thereby generating a learning dataset Img_tch consisting of a predetermined number of pseudo data.

Note that if the number of pseudo data created in step S11 is sufficient as a dataset for machine learning, step S13, in which pseudo data is added through image conversion processing, may be omitted. The generated learning dataset Img_tch may also include image data obtained by actually measuring an inspection object exhibiting an abnormality.

The user then inputs the generated machine learning dataset into the inspection device 20 and causes the learning unit 35 to perform machine learning to generate an anomaly determination model (step S15). As described above, the learning unit 35 performs machine learning of the anomaly determination model using an existing machine learning model that can be used for machine learning using image data as training data.

Next, the user verifies the validity of the generated anomaly detection model using verification image data obtained by actually measuring an abnormality that has occurred in the test object (step S17). The verification image data used to verify validity is actual image data that has not been used in the machine learning of the anomaly detection model. The verification image data includes, for example, image data Img_act in which an abnormality has occurred, as shown in Figure 4, and image data in a normal state where no abnormality has occurred. The number of verification image data may be greater than the number of pseudo data used in the machine learning of the anomaly detection model. The user inputs the verification image data into the anomaly detection model and obtains data on the accuracy rate of the judgment results output.

The user then determines whether the obtained accuracy rate meets the desired requirements (step S19). If the accuracy rate does not meet the desired requirements (S19/No), the user returns to step S11, creates new pseudo data Img_fak, and re-runs the generation of the anomaly detection model. Instead of returning to step S11, the user may return to step S13 and add pseudo data through image conversion processing. This allows for adaptation to improve the accuracy of the anomaly detection results using the anomaly detection model.

On the other hand, if the accuracy rate meets the desired requirements (S19/Yes), the user ends the task of generating the anomaly determination model. The generated anomaly determination model is stored in memory 21 or storage device 23.

The anomaly determination model may not only determine the presence or absence of an anomaly, but may also be a model that can classify and output the type of anomaly. For example, machine learning is performed by associating the pseudo data Img_fak of the learning dataset Img_tch used in machine learning of the anomaly determination model with the type of anomaly, to generate an anomaly determination model. This makes it possible to estimate the anomaly occurring in the inspection object from the determination results obtained by inputting inspection image data into the anomaly determination model, and to select how to deal with the anomaly and its level of urgency.

4. Operation of the Inspection Device
The operation of the abnormality determination process performed by the determination unit 37 of the inspection device 20 will be described.

Figure 5 is a flowchart showing the operation of the abnormality determination process for the test object, which is mainly performed by the determination unit 37.

First, the image processing unit 31 acquires image data obtained by measuring the object to be inspected and output from the detector 11, and generates inspection image data by processing the image data (step S21).

The determination unit 37 then inputs the generated inspection image data into the anomaly determination model to determine whether or not an abnormality exists in the inspection target (step S23). For example, the determination unit 37 obtains data on the probability of an abnormality occurring as the determination result output from the anomaly determination model, and determines that an abnormality has occurred if the probability of an abnormality occurring is equal to or greater than a predetermined threshold.

Next, the display control unit 39 generates a drive signal for the display unit 13 and causes the display unit 13 to display information about the judgment result (step S25). In addition to displaying the result on the display unit 13, the display control unit 39 may also save a log of the judgment result in the memory 21 or the storage device 23. Furthermore, if an abnormality has occurred in the test object, the display control unit 39 may notify the user by generating an alarm or the like.

In addition, if the abnormality determination model is a model that can classify not only the presence or absence of an abnormality but also the type of abnormality that has occurred, the display control unit 39 may cause the display unit 13 to display information on the type of abnormality based on the determination result data output from the abnormality determination model. At that time, the display control unit 39 may also display information such as the countermeasure and urgency level according to the type of abnormality.

The inspection device 20 may, for example, continuously acquire inspection image data obtained by measuring the inspection object using a detector 11 fixed at a fixed point, and repeat steps S21 to S25 to monitor the inspection object for abnormalities. Alternatively, the user may measure the inspection object using the detector 11 at any time, and the inspection device 20 may execute steps S21 to S25 based on the obtained inspection image data to determine whether or not the inspection object has an abnormality.

Alternatively, the inspection object may be configured to be movable relative to the detector 11 and positioned so as to pass sequentially through the measurement range of the detector 11, and the inspection device 20 may continuously acquire inspection image data obtained by measuring the inspection object with the detector 11, and monitor the inspection object for abnormalities by repeating the processes of steps S21 to S25. This makes it possible to determine the presence or absence of abnormalities using an anomaly determination model, even for long inspection objects, without having to measure the entire object at the same time.

<5. Application Examples>
An application example of the inspection system 100 using the inspection device 20 according to this embodiment will be described.

(5-1. First example)
6 is an explanatory diagram showing an example of an inspection system 100A applied to a product manufacturing line. In the example shown in Fig. 6, two detectors 11a and 11b are installed so as to be able to measure a manufacturing line in which products are produced by assembling parts using multiple arm robots 51a to 51c and 53 to equipment moving on a belt conveyor 55. In such a manufacturing line, for example, it is assumed that there is a risk of scratches occurring on the arm robots 51a to 51c or the leg 55a of the belt conveyor 55 due to contact or sliding with other objects.

In this case, the user artificially draws possible characteristic patterns of flaws and then performs image conversion processing to generate a learning dataset Img_tch containing multiple pseudo data Img_fak. The user also performs machine learning in advance using these learning datasets Img_tch as input data to generate an anomaly determination model. The inspection device 20 then inputs inspection image data generated from the image data output from detectors 11a and 11b into the anomaly determination model, and repeatedly performs processing to determine the presence or absence of an anomaly based on the output from the anomaly determination model.

This makes it possible to monitor for scratches and other abnormalities on the production line, even when the inspection target, such as arm robots 51a-51c and 55, includes moving parts. Furthermore, because the system determines the presence or absence of an abnormality using an abnormality determination model trained using pseudo-data created by artificially drawing characteristic patterns of scratches and other abnormalities as training data, it is possible to detect anticipated abnormalities even when there is no image data of an actual abnormality. The application example shown in Figure 6 can be used as an inspection system for the appearance of various inspection targets, such as boxes on containers moving within a warehouse, or floors or walls within a factory or warehouse.

(5-2. Second example)
7 shows an example of an inspection system 100B applied to the inspection of inspection target items 63a to 63e placed and transported on a belt conveyor 61. In the example shown in Fig. 7, two detectors 11c and 11d are installed facing specific measurement ranges Ra and Rb, respectively, of the belt conveyor 61. It is assumed that the inspection target items 63a to 63e may be subject to scratches, abrasions, dirt, damage, etc.

In this case, the user artificially draws characteristic patterns of possible external anomalies or characteristic patterns of shapes that do not exist in the appearance of normal inspection target products 63a-63e, and then performs image conversion processing to generate a training data set Img_tch containing multiple pseudo data Img_fak. The user also performs machine learning in advance using these training data sets Img_tch as input data to generate an anomaly determination model. Then, as inspection target products 63a-63e sequentially pass through each measurement range Ra, Rb, inspection device 20 inputs inspection image data generated from image data output from detectors 11c, 11d into the anomaly determination model, and repeatedly performs processing to determine whether or not there is an abnormality in inspection target products 63a-63e based on the output from the anomaly determination model.

This allows for automatic and continuous inspection of inspection targets 63a-63e as they pass sequentially through the measurement ranges Ra and Rb of detectors 11c and 11d. Furthermore, because the presence or absence of an abnormality is determined using an abnormality determination model trained using pseudo-data created by artificially drawing abnormality characteristic patterns as training data, it is possible to detect anticipated or unknown abnormalities even when image data of an actual abnormality is not available.

In the application example shown in Figure 7, the measurement ranges Ra and Rb of detectors 11c and 11d are fixed, and abnormality determination is performed using image data obtained by measuring inspection targets 63a-63b passing through these measurement ranges Ra and Rb. However, an inspection system can also be constructed so that the inspection target does not move, but the detector's measurement range moves relative to the inspection target. Also, instead of the inspection target being placed on a conveyor belt and transported, the inspection target itself may move relative to the detector's measurement range. The application example shown in Figure 7 can also be used, for example, as an inspection system for the exterior of vehicles such as trains and automobiles.

<6. Effects>
As described above, according to the anomaly determination model generation method and anomaly determination model generation device of the present embodiment, pseudo data (e.g., data artificially created by a user) created by simulating characteristic patterns of an anomaly is used as training data for generating an anomaly determination model used for anomaly determination, rather than image data obtained by actually measuring an inspection target exhibiting an anomaly. Therefore, because training data including characteristic patterns of an anomaly can be freely generated, an anomaly determination model can be generated that can detect anomalies that are predicted in advance based on knowledge or empirical rules, as well as unexpected, unknown anomalies, even before an actual anomaly occurs or when the number of samples showing a specific anomaly is small. Furthermore, because training data that is recognized as an anomaly in advance can be freely generated, the burden of classifying actually measured image data as normal or abnormal, or by type of anomaly, can be reduced, thereby reducing costs.

Furthermore, according to this embodiment, the effectiveness of an anomaly determination model generated by machine learning using pseudo data as learning data is verified using verification image data obtained by actually measuring the inspection target. Therefore, an anomaly determination model with high determination accuracy can be generated while creating learning data so that the accuracy rate of the generated anomaly determination model is the desired accuracy rate. Therefore, even when using pseudo data created by a user artificially drawing characteristic patterns of an anomaly, an anomaly determination model with high determination accuracy can be generated.

Furthermore, according to this embodiment, pseudo data created by a user artificially drawing characteristic patterns of abnormalities can be used as training data. This makes it possible to create training data with characteristic patterns of various shapes, and to generate an abnormality determination model capable of detecting various abnormalities. Furthermore, because pseudo data created by a user artificially drawing characteristic patterns of abnormalities can be used as training data, it is possible to use training data that includes patterns of simplified shapes, reducing the risk of overfitting during machine learning.

Furthermore, with the inspection device according to this embodiment, the appearance of the inspection object can be inspected using the anomaly judgment model generated by the above-described anomaly judgment model generation method and anomaly judgment model generation device. Therefore, even before an abnormality actually occurs or when the number of samples in which a specific abnormality occurs is small, an abnormality that has occurred in the appearance of the inspection object can be more easily detected, thereby improving the accuracy of the judgment. Furthermore, with the inspection device according to this embodiment, an anomaly judgment is performed using an anomaly judgment model in which overfitting is suppressed, which is generated using learning data including simplified depictions of abnormality feature patterns, thereby improving the accuracy of the judgment of an abnormality that has occurred in the appearance of the inspection object.

The above describes in detail preferred embodiments of the present invention with reference to the accompanying drawings, but the present invention is not limited to such examples. It is clear that a person with ordinary skill in the technical field to which the present invention pertains can conceive of various modifications or alterations within the scope of the technical ideas set forth in the claims, and it is understood that these also naturally fall within the technical scope of the present invention.

For example, the present invention can be applied to an inspection object to which the anomaly determination model generation method, anomaly determination model generation device, and an inspection device according to the present invention are applied, as long as a feature pattern can be extracted from the image data of the inspection object.

Furthermore, the learning dataset Img_tch does not have to be composed of pseudo data only, but may also include at least one of normal data and abnormal data obtained by actually measuring the test subject.

In addition, in the above embodiment, an example was described in which the inspection device 20 was equipped with the learning unit 35 and the judgment unit 37, but the present invention is not limited to such an example. For example, the learning unit 35 may be provided in a separate device and configured as an independent anomaly judgment model generation device. In this case, an anomaly judgment model is generated by performing machine learning in advance using the anomaly judgment model generation device including the learning unit 35, and the anomaly judgment model is stored in the memory 21 or storage device 23 of the inspection device 20 including the judgment unit 37.

Furthermore, when abnormality determination is performed offline rather than in real time using the inspection device 20, the inspection device 20 may be configured without including the image processing unit 31. For example, inspection image data extracted from image data output from the detector 11 may be stored in a storage medium, and the inspection image data stored in the storage medium may be read into the determination unit 37 of the inspection device 20. Even in this configuration, the inspection device 20 can perform abnormality determination of the inspection object based on the abnormality determination model.

1... inspection device, 3... detector, 7... display unit, 10... control unit, 11... image processing unit, 13... calculation unit, 15... learning unit, 17... determination unit, 19... display control unit, 21... memory, 23... storage device

Claims (5)

1. A method for generating an abnormality determination model for determining an abnormality in an inspection object using inspection image data obtained by measuring the inspection object as input data, the method comprising:
generating a learning data set (Img_tch) for generating the anomaly determination model based on pseudo data (Img_fak) that is not actual image data obtained by measuring the inspection object;
generating the anomaly determination model using the learning dataset (Img_tch);
a step of verifying the validity of the anomaly determination model by using image data obtained by actually measuring an anomaly occurring in the inspection object as verification image data;
Including,
The method for generating an anomaly determination model is characterized in that the pseudo data (Img_fak) is data created by a user drawing a characteristic pattern of an abnormality that may occur in the inspection object, and is also data generated by using data creation software to read the characteristic pattern of an abnormality that the user created by hand . The anomaly determination model generation method described in claim 1, characterized in that the abnormality characteristic pattern of the pseudo data (Img_fak) is a pattern of the shape of an abnormality formed on the surface of the inspection object. An abnormality determination model generation device (20) that generates an abnormality determination model for determining an abnormality in an inspection object using inspection image data obtained by measuring the inspection object as input data,
a learning unit (35) that generates the anomaly determination model using a learning dataset (Img_tch) that is generated based on pseudo data (Img_fak) that is not actual image data obtained by measuring the inspection object, and verifies the validity of the anomaly determination model by using image data obtained by actually measuring an anomaly that has occurred in the inspection object as verification image data ,
The anomaly determination model generating device is characterized in that the pseudo data (Img_fak) is data created by a user drawing a characteristic pattern of an abnormality that may occur in the inspection object, and is also data generated by using data creation software to read the characteristic pattern of an abnormality that the user hand-drawn. In an inspection device (20) that inspects an object to be inspected based on inspection image data obtained by measuring the object to be inspected,
a storage unit (21, 23) for storing an anomaly determination model generated using a learning dataset (Img_tch) generated based on pseudo data (Img_fak) that is not actual image data obtained by measuring the inspection object;
a determination unit (37) that inputs the inspection image data obtained by measuring the inspection object into the abnormality determination model and determines whether or not the inspection object has an abnormality;
Including,
the pseudo data (Img_fak) is data created by a user drawing a characteristic pattern of an abnormality that may occur in the inspection object , and is also data generated by reading the characteristic pattern of an abnormality handwritten by the user using data creation software;
The inspection apparatus is characterized in that the validity of the abnormality determination model is verified by using image data obtained by actually measuring an abnormality occurring in the inspection object as verification image data . 5. The inspection device according to claim 4, wherein the object to be inspected is configured to be movable relative to the measurement range (Ra, Rb) of the detector (11) of the inspection image data, and is arranged so as to pass sequentially through the measurement range (Ra, Rb) of the detector (11).