Learning device, simulated data generation device, and learning method

Abstract

To improve accuracy of an AI model through machine learning using sufficient data.SOLUTION: A learning device is provided, comprising a detection unit for detecting anomalies in image data, a generation unit configured to process the image data subjected to the anomaly detection by the detection unit to generate simulated data similar to the image data, and a learning control unit for controlling learning of a model using the simulated data generated by the generation unit.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to a learning device, a simulation data generating device, and a learning method. [Background technology]

[0002] Generally, for example, in image inspection of products, anomaly detection is sometimes performed using an AI (Artificial Intelligence) model. The AI ​​model used for such anomaly detection is generated by, for example, having the AI ​​learn machine learning of images of an inspection target that have no anomalies and images of an inspection target that have anomalies.

[0003] In addition, when an abnormality is detected in an inspection object, it is considered to classify the detected abnormality using an AI model. In other words, the AI ​​model classifies the detected abnormality as, for example, a surface scratch or a dent. The AI ​​model used for such anomaly classification is generated by having the AI ​​learn images of inspection objects with various abnormalities through machine learning. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2021-157303 A [Patent Document 2] Patent Publication No. 2022-24541 [Patent Document 3] JP 2022-189247 A Summary of the Invention [Problem to be solved by the invention]

[0005] However, if the AI ​​does not learn a sufficient amount of data, the accuracy of the AI ​​model does not improve, and there is a problem that it is difficult to perform anomaly detection and anomaly classification with high accuracy. Specifically, for example, when operating anomaly detection using an AI model, if an anomaly is falsely detected in an inspection subject that does not have an anomaly, the false positive data may be re-trained to the AI. However, since the occurrence frequency of such false positive data is low and it is difficult to collect sufficient false positive data, even if the false positive data is re-trained to the AI, the improvement in the accuracy of the AI ​​model is limited, and it is difficult to prevent similar false positives.

[0006] In addition, since only a small number of inspection objects typically have abnormalities, it is difficult to obtain the large amounts of data required to train an AI to classify abnormalities, and it is not easy to improve the accuracy of an AI model for classifying anomalies.

[0007] The present disclosure has been made in consideration of the above, and aims to provide a learning device, a simulated data generation device, and a learning method that can improve the accuracy of an AI model through machine learning using sufficient data. [Means for solving the problem]

[0008] According to one aspect of the present disclosure, a learning device includes a detection unit that detects anomalies in image data, a generation unit that processes image data in which anomalies have been detected by the detection unit to generate simulated data similar to the image data, and a learning control unit that controls learning of a model using the simulated data generated by the generation unit.

[0009] According to another aspect of the present disclosure, the simulated data generation device has a detection unit that detects abnormalities in image data, and a generation unit that processes image data in which an abnormality has been detected by the detection unit, and generates simulated data similar to the image data.

[0010] According to yet another aspect of the present disclosure, the learning method detects anomalies in image data, processes the image data in which the abnormality is detected to generate simulated data similar to the image data, and performs model learning using the generated simulated data. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 is a block diagram showing a configuration of a learning device according to the first embodiment. [Diagram 2] FIG. 2 is a flow diagram showing the learning method according to the first embodiment. [Diagram 3] FIG. 3 is a diagram showing a specific example of the simulation data according to the first embodiment. [Figure 4] FIG. 4 is a block diagram showing a configuration of a learning device according to the second embodiment. [Diagram 5] FIG. 5 is a flow diagram showing a learning method according to the second embodiment. [Figure 6] FIG. 6 is a diagram showing a specific example of the simulation data according to the second embodiment. [Figure 7] FIG. 7 is a block diagram showing a configuration of a learning device according to the third embodiment. [Figure 8] FIG. 8 is a flow diagram showing a learning method according to the third embodiment. [Figure 9] FIG. 9 is a block diagram showing a specific example of the hardware configuration of the learning device. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. The embodiments described below are merely examples, and should not be construed as being limited by the description.

[0013] (Embodiment 1) Fig. 1 is a block diagram showing the configuration of a learning device 100 according to embodiment 1. The learning device 100 shown in Fig. 1 includes an image acquisition unit 110, an abnormality detection unit 120, a normal data storage unit 130, a learning control unit 140, a result confirmation unit 150, an abnormal data storage unit 160, an erroneous detection data storage unit 170, and a data generation unit 180.

[0014] The image acquisition unit 110 has an imaging device such as a camera, and acquires image data of a product or the like to be inspected.

[0015] The anomaly detection unit 120 generates an AI model (hereinafter referred to as "anomaly detection model") that detects anomalies from image data by machine learning. That is, the anomaly detection unit 120 executes learning of the anomaly detection model using normal data provided from the learning control unit 140. The anomaly detection model used for learning by the anomaly detection unit 120 may be a model that learns anomalies end-to-end, such as Patch Core or CS (Cross Scale)-Flow, or may be a model that uses a reconstruction error, such as an autoencoder or VAE (Variational AutoEncoder).

[0016] During anomaly detection operation, the anomaly detection unit 120 uses an anomaly detection model obtained by learning to detect anomalies in the image data input from the image acquisition unit 110. That is, the anomaly detection unit 120 applies the anomaly detection model to the image data to obtain an anomaly score value and an anomaly location, and determines that an anomaly exists in the anomaly location when the anomaly score value is equal to or greater than a predetermined threshold value.

[0017] The normal data storage unit 130 stores normal data including image data of inspection targets having no abnormalities.

[0018] The learning control unit 140 acquires normal data stored by the normal data storage unit 130, and causes the anomaly detection unit 120 to learn the anomaly detection model. Furthermore, when erroneous detection data and simulation data, which will be described later, are newly stored by the normal data storage unit 130, the learning control unit 140 acquires normal data including the erroneous detection data and simulation data, and causes the anomaly detection unit 120 to re-learn the anomaly detection model.

[0019] The result confirmation unit 150 displays image data in which an anomaly has been detected by the anomaly detection unit 120, clearly indicating the location of the anomaly, and accepts input of a confirmation result as to whether or not the image data in which an anomaly has been detected is image data of an inspection target that truly contains an anomaly. That is, the result confirmation unit 150 presents image data in which an anomaly has been detected by the anomaly detection model to an administrator or the like, and accepts input of a confirmation result as to whether or not this image data is an image that truly contains an anomaly.

[0020] Then, when a confirmation result indicating that the image truly contains an abnormality is input, the result confirmation unit 150 stores the image data and abnormality data including the abnormality score value and information on the abnormal location in the abnormality data storage unit 160, and when a confirmation result indicating that the image does not truly contain an abnormality is input, the result confirmation unit 150 stores the image data and false detection data including the abnormality score value and information on the abnormal location in the false detection data storage unit 170. In other words, the result confirmation unit 150 stores, among the image data in which an abnormality has been detected by the anomaly detection model, abnormality data relating to image data confirmed to contain an abnormality in the abnormality data storage unit 160, and stores, in the false detection data storage unit 170, false detection data relating to image data confirmed to contain no abnormality.

[0021] The abnormal data storage unit 160 stores abnormal data including image data confirmed to have an abnormality among image data in which an abnormality has been detected by the anomaly detection unit 120. That is, the abnormal data storage unit 160 stores image data in which an abnormality has been correctly detected by the anomaly detection model, as well as the abnormality score value and abnormal location of this image data.

[0022] The false positive data storage unit 170 stores false positive data including image data confirmed to be free of anomalies among image data in which anomalies have been detected by the anomaly detection unit 120. That is, the false positive data storage unit 170 stores image data in which an anomaly has been erroneously detected by the anomaly detection model, as well as the anomaly score value and abnormal location of this image data.

[0023] The data generating unit 180 generates simulation data similar to the false detection data based on the false detection data stored in the false detection data storage unit 170. That is, the data generating unit 180 generates simulation data in which an abnormality may be falsely detected, using the normal data stored in the normal data storage unit 130 and the false detection data stored in the false detection data storage unit 170. Specifically, the data generating unit 180 extracts a false detection point where an abnormality is falsely detected from the false detection data, processes the image of the false detection point, and adds it to the normal data to generate the simulation data. When processing the image of the false detection point, the data generating unit 180 executes processing such as enlargement / reduction, affine transformation, translation, rotation, or projective transformation. Then, the data generating unit 180 stores the false detection data used to generate the simulation data in the normal data storage unit 130, and stores the generated simulation data in the normal data storage unit 130.

[0024] Next, a learning method by the learning device 100 configured as described above will be described with reference to the flow chart shown in Fig. 2. In the following, re-learning of the anomaly detection model will be described in the case where anomaly detection is performed by an anomaly detection model that has been learned using normal data previously stored in the normal data storage unit 130.

[0025] When anomaly detection is being implemented, image data of an inspection target is acquired by the image acquisition unit 110 (step S101). The acquired image data is output to the anomaly detection unit 120, and anomalies in the image data of the inspection target are detected by the anomaly detection unit 120 using an anomaly detection model (step S102). Specifically, the anomaly detection model is applied to the image data to obtain an anomaly score value and an abnormality location, and if the anomaly score value is equal to or greater than a predetermined threshold, it is determined that an anomaly exists in the abnormality location.

[0026] If no abnormality is detected in the image data of the inspection target as a result of the abnormality detection, an abnormality detection result indicating that there is no abnormality may be notified to an administrator or the like, and the image data in which no abnormality is detected may be stored in the normal data storage unit 130. Here, the explanation will be continued assuming that an abnormality is detected in the image data of the inspection target as a result of the abnormality detection.

[0027] When the anomaly detection unit 120 detects an anomaly in the image data of the inspection target, the result confirmation unit 150 displays the image data in which the anomaly was detected, and requests the administrator or the like to confirm whether or not there is an anomaly. That is, the image data in which the anomaly was detected is displayed with the abnormality clearly indicated, and the administrator or the like is prompted to confirm whether or not the abnormality is truly anomaly. Then, the result confirmation unit 150 accepts input of a confirmation result as to whether or not the anomaly detection is a false positive (step S103), and when a confirmation result indicating that the image is truly anomaly and not a false positive is input (step S103 No), the abnormal data including the image data is stored in the abnormal data storage unit 160 (step S109).

[0028] On the other hand, if a confirmation result indicating that the image is not truly abnormal but is a false positive (Yes in step S103), false positive data including the image data is stored in the false positive data storage unit 170 (step S104). Then, the data generation unit 180 acquires an image of the false positive location where the abnormality was falsely detected from the false positive data stored in the false positive data storage unit 170 (step S105).

[0029] The image of the false detection location is processed by the data generating unit 180 and added to the image of the normal data to generate simulation data (step S106). The image of the false detection location is processed by, for example, scaling, affine transformation, translation, rotation, or projective transformation. The processing does not have to be of only one type, and a combination of multiple types of processing may be applied to the image of the false detection location. Furthermore, multiple processed images are generated by performing different processing on the image of the false detection location, and each processed image is added to the image of the normal data to generate multiple simulation data.

[0030] Here, the generation of the simulation data will be described using a specific example with reference to FIG.

[0031] 3, the simulation data is generated using an image of normal data (hereinafter referred to as a "normal image") 210 and an image of false positive data (hereinafter referred to as a "false positive image") 220. Since the false positive image 220 includes a false positive portion identified as an abnormal portion by the anomaly detection model, an image of the false positive portion (hereinafter referred to as a "false positive portion image") 230 is obtained from the false positive image 220. The false positive portion image 230 includes a false positive portion 231.

[0032] The false detection portion image 230 is subjected to processing such as enlargement / reduction, affine transformation, translation, rotation, or projective transformation of the false detection portion 231 to generate a processed image 240. At this time, a plurality of processed images 240 each having a different processing may be generated. The processed image 240 includes a false detection portion 241 that has been processed.

[0033] Then, the processed image 240 is added to the normal image 210 to generate a simulated data image 250. When a plurality of processed images 240 are generated, the normal image 210 is added to each of the processed images 240 to generate a plurality of simulated data images 250. The simulated data image 250 generated in this manner is an image different from the erroneous detection image 220, but is a normal image in which an abnormality may be erroneously detected.

[0034] 2, when the simulation data is generated, the generated simulation data is stored in the normal data storage unit 130 together with the false positive data used to generate the simulation data (step S107). That is, both the false positive data and the simulation data are image data that are falsely detected as an anomaly by the current anomaly detection model, but are actually normal image data that do not contain an anomaly. For this reason, the false positive data and the simulation data are stored in the normal data storage unit 130.

[0035] When the false positive data and the simulation data are stored in the normal data storage unit 130, the newly stored normal data including the false positive data and the simulation data is acquired by the learning control unit 140, and the anomaly detection model of the anomaly detection unit 120 is re-learned using the acquired normal data (step S108). In re-learning the anomaly detection model, learning is performed using not only the false positive data but also the simulation data, so that the accuracy of the anomaly detection model can be improved by machine learning using sufficient data.

[0036] As described above, according to the present embodiment, false positive data in which an anomaly is falsely detected by the anomaly detection model is processed to generate mock data, which is normal data that may falsely detect an anomaly, and the anomaly detection model is retrained using the normal data including the mock data. As a result, training is performed using a wide range of sufficient normal data, and the accuracy of the anomaly detection model can be improved.

[0037] (Embodiment 2) Fig. 4 is a block diagram showing the configuration of a learning device 100 according to embodiment 2. In Fig. 4, the same components as those in Fig. 1 are given the same reference numerals, and their description will be omitted. The learning device 100 shown in Fig. 4 has a learning control unit 320 and a data generation unit 330 instead of the learning control unit 140 and the data generation unit 180 of the learning device 100 shown in Fig. 1, and adopts a configuration in which a region division unit 310 is added.

[0038] The region division unit 310 generates an AI model (hereinafter referred to as "region division model") that divides regions in image data by machine learning. That is, the region division unit 310 executes learning of the region division model using normal data provided from the learning control unit 320. The region division model used for learning by the region division unit 310 may be, for example, a segmentation model such as U-net or Deeplab.

[0039] During anomaly detection, the area division unit 310 divides the area of ​​the image data input from the image acquisition unit 110 using an area division model obtained by learning. That is, the area division unit 310 applies the area division model to the image data to detect an area that can be divided in the image data. Therefore, the area division unit 310 divides the area of ​​the image data of, for example, a product to be inspected, and detects areas corresponding to parts that constitute the product.

[0040] The learning control unit 320 acquires normal data stored by the normal data storage unit 130, and causes the anomaly detection unit 120 to learn the anomaly detection model. When erroneous detection data and simulation data are newly stored by the normal data storage unit 130, the learning control unit 320 acquires normal data including the erroneous detection data and simulation data, and causes the anomaly detection unit 120 to re-learn the anomaly detection model. Furthermore, the learning control unit 320 acquires normal data stored by the normal data storage unit 130, and causes the region division unit 310 to learn the region division model. When erroneous detection data and simulation data are newly stored by the normal data storage unit 130, the learning control unit 320 does not need to re-learn the region division model.

[0041] The data generating unit 330 generates simulation data similar to the false detection data based on the false detection data stored in the false detection data storage unit 170. That is, the data generating unit 330 generates simulation data in which an abnormality may be falsely detected, using the normal data stored in the normal data storage unit 130 and the false detection data stored in the false detection data storage unit 170. Specifically, the data generating unit 330 extracts a false detection point where an abnormality is falsely detected from the false detection data, processes the image of the false detection point, and adds it to the normal data to generate the simulation data. When processing the image of the false detection point, the data generating unit 330 uses the region division result by the region dividing unit 310 and processes the image within the range of the region where the false detection point is located. In this processing, the data generating unit 330 executes processing such as enlargement / reduction, affine transformation, translation, rotation, or projective transformation. Then, the data generating section 330 stores the erroneous detection data used to generate the simulation data in the normal data storing section 130, and also stores the generated simulation data in the normal data storing section 130.

[0042] Next, a learning method by the learning device 100 configured as described above will be described with reference to the flow chart shown in Fig. 5. In Fig. 5, the same parts as in Fig. 2 are given the same reference numerals, and detailed description thereof will be omitted. In the following, re-learning of the anomaly detection model will be described in the case where anomaly detection is performed by an anomaly detection model that has been learned using normal data previously stored in the normal data storage unit 130.

[0043] When anomaly detection is being implemented, image data of an inspection target is acquired by the image acquisition unit 110 (step S101). The acquired image data is output to the anomaly detection unit 120 and the area division unit 310, and the area division unit 310 uses an area division model to perform area division of the image data of the inspection target (step S201). In addition, the anomaly detection unit 120 uses the anomaly detection model to detect an anomaly in the image data of the inspection target (step S102).

[0044] If no abnormality is detected in the image data of the inspection target as a result of the abnormality detection, an abnormality detection result indicating that there is no abnormality may be notified to an administrator or the like, and the image data in which no abnormality is detected may be stored in the normal data storage unit 130. Here, the explanation will be continued assuming that an abnormality is detected in the image data of the inspection target as a result of the abnormality detection.

[0045] When the anomaly detection unit 120 detects an anomaly in the image data of the inspection target, the result confirmation unit 150 displays the image data in which the anomaly was detected, and requests the administrator or the like to confirm the presence or absence of the anomaly. Then, the result confirmation unit 150 accepts input of a confirmation result as to whether or not the anomaly detection is a false positive (step S103), and when a confirmation result indicating that the anomaly detection is not a false positive is input (step S103 No), the abnormal data including the image data is stored in the abnormal data storage unit 160 (step S109).

[0046] On the other hand, if a confirmation result indicating a false positive is input (Yes in step S103), false positive data including the image data is stored in false positive data storage unit 170 (step S104). Then, data generation unit 330 acquires an image of the false positive location where an abnormality was falsely detected from the false positive data stored in false positive data storage unit 170 (step S105).

[0047] The image of the false detection location is processed by the data generating unit 330 using the region division result, and is added to the image of the normal data to generate simulation data (step S202). The image of the false detection location is processed by, for example, enlarging / reducing, affine transformation, translation, rotation, or projective transformation. At this time, various processes are performed within the range of the area where the false detection location is located by using the region division result. In other words, areas other than the area including the false detection location are not processed. In addition, it is not necessary to perform only one type of processing during processing, and a combination of multiple types of processing may be performed on the image of the false detection location. Furthermore, multiple processed images are generated by performing different processes on the image of the false detection location, and each processed image is added to the image of the normal data to generate multiple simulation data.

[0048] Here, the generation of the simulation data will be described using a specific example with reference to FIG.

[0049] 6, the simulation data is generated using a normal image 410 and a false positive image 420. The normal image 410 and the false positive image 420 each include regions 411, 412, and 413 that are divided by a region division model. The false positive portion is included in region 413 of the false positive image 420. From this false positive image 420, a false positive portion image 430 including a false positive portion 431 is obtained.

[0050] The false detection part image 430 is subjected to processing such as enlargement / reduction, affine transformation, translation, rotation, or projective transformation of the false detection part 431 to generate a processed image 440. At this time, the processing of the false detection part image 430 is performed within the range of the region 413 in which the false detection part 431 is located. Therefore, in the processed image 440, the processed false detection part 441 is located within the region 413. Therefore, it is possible to generate a processed image 440 including the false detection part 441 that may actually occur.

[0051] Then, the processed image 440 is added to the normal image 410 to generate the simulated data image 450. When multiple processed images 440 are generated, the normal image 410 is added to each processed image 440 to generate multiple simulated data images 450. The simulated data image 450 generated in this manner is an image different from the false detection image 420, but is a normal image in which an abnormality may be falsely detected. In addition, since the processed image 440 includes a false detection portion 441 that may actually occur, the simulated data image 450 also includes a false detection portion that may actually occur.

[0052] 5, when the simulated data is generated, the generated simulated data is stored in the normal data storage unit 130 together with the false positive data used to generate the simulated data (step S107). Then, the newly stored false positive data and normal data including the simulated data are acquired by the learning control unit 320, and the anomaly detection model of the anomaly detection unit 120 is re-learned using the acquired normal data (step S108). In re-learning the anomaly detection model, learning is performed using not only the false positive data but also the simulated data, so that the accuracy of the anomaly detection model can be improved by machine learning using sufficient data.

[0053] As described above, according to this embodiment, false positive data in which an anomaly is falsely detected by the anomaly detection model is processed while utilizing the region segmentation result by the region segmentation model to generate simulated data that is normal data that may falsely detect an anomaly, and the anomaly detection model is re-trained using the normal data including the simulated data. Therefore, training is performed using a wide range of sufficient normal data, and the accuracy of the anomaly detection model can be improved. In addition, since false positive data is processed while utilizing the region segmentation result, simulated data including false positive locations that may actually occur can be generated.

[0054] (Embodiment 3) Fig. 7 is a block diagram showing the configuration of a learning device 100 according to embodiment 3. In Fig. 7, the same components as those in Fig. 1 are given the same reference numerals, and their description will be omitted. The learning device 100 shown in Fig. 7 has a learning control unit 520 and a data generation unit 530 instead of the learning control unit 140 and the data generation unit 180 of the learning device 100 shown in Fig. 1, and adopts a configuration in which an anomaly classification unit 510 is added.

[0055] The anomaly classification unit 510 generates an AI model (hereinafter referred to as the "anomaly classification model") that classifies detected anomalies through machine learning. That is, the anomaly classification unit 510 executes learning of the anomaly classification model using anomaly data provided from the learning control unit 520. The anomaly classification model used for learning by the anomaly classification unit 510 may be, for example, a model used for image classification such as RESNET or ViT.

[0056] During anomaly detection operation, anomaly classification unit 510 classifies anomalies detected by anomaly detection unit 120 using an anomaly classification model obtained by learning. That is, anomaly classification unit 510 applies the anomaly classification model to image data in which an anomaly has been detected by anomaly detection unit 120, and classifies the anomaly detected from the image data. Therefore, anomaly classification unit 510 classifies, for example, an anomaly detected from image data of an inspection target as being a scratch on the surface of the inspection target, a dent in the inspection target, or the like.

[0057] The learning control unit 520 acquires normal data stored by the normal data storage unit 130, and causes the anomaly detection unit 120 to learn the anomaly detection model. The learning control unit 520 also acquires abnormal data stored by the abnormal data storage unit 160, and causes the anomaly classification unit 510 to learn the anomaly classification model. When simulated abnormal data, which will be described later, is newly stored by the abnormal data storage unit 160, the learning control unit 520 acquires abnormal data including the simulated abnormal data, and causes the anomaly classification unit 510 to re-learn the anomaly classification model.

[0058] The data generating unit 530 generates simulated abnormal data similar to the abnormal data based on the abnormal data stored in the abnormal data storage unit 160. That is, the data generating unit 530 generates simulated abnormal data, which is a variation of the abnormal data, using the normal data stored in the normal data storage unit 130 and the abnormal data stored in the abnormal data storage unit 160. Specifically, the data generating unit 530 extracts an abnormal portion where an abnormality is detected from the abnormal data, processes an image of the abnormal portion, and adds it to the normal data to generate the simulated abnormal data. When processing the image of the abnormal portion, the data generating unit 530 executes processes such as enlargement / reduction, affine transformation, translation, rotation, or projective transformation. Then, the data generating unit 530 stores the generated simulated abnormal data in the abnormal data storage unit 160.

[0059] Next, a learning method by the learning device 100 configured as described above will be described with reference to the flow chart shown in Fig. 8. In Fig. 8, the same parts as in Fig. 2 are given the same reference numerals, and detailed description thereof will be omitted. In the following, re-learning of an anomaly classification model will be described in the case where anomaly detection and anomaly classification are performed using an anomaly detection model and an anomaly classification model that have been trained in advance.

[0060] When anomaly detection and anomaly classification are implemented, image data of an inspection target is acquired by the image acquisition unit 110 (step S101). The acquired image data is output to the anomaly detection unit 120, and anomalies in the image data of the inspection target are detected by the anomaly detection unit 120 using an anomaly detection model (step S102).

[0061] If no abnormality is detected in the image data of the inspection target as a result of the abnormality detection, an abnormality detection result indicating that there is no abnormality may be notified to an administrator or the like, and the image data in which no abnormality is detected may be stored in the normal data storage unit 130. Here, the explanation will be continued assuming that an abnormality is detected in the image data of the inspection target as a result of the abnormality detection.

[0062] When the anomaly detection unit 120 detects an anomaly in the image data of the inspection target, the anomaly classification unit 510 uses the anomaly classification model to classify the detected anomaly (step S301). Specifically, the anomaly classification model is applied to the image data in which the anomaly is detected, and the anomaly location is classified as to which category of anomaly it belongs to. When the anomaly classification is performed, the result confirmation unit 150 displays the image data in which the anomaly is detected, and requests the administrator or the like to confirm the presence or absence of an anomaly. Then, the result confirmation unit 150 accepts input of the confirmation result of the presence or absence of an anomaly, and when a confirmation result indicating the presence of an anomaly is input, the anomaly data including the image data and the classification result of the anomaly is stored in the anomaly data storage unit 160 (step S302). Then, the data generation unit 530 acquires an image of the anomaly location in which the anomaly is detected from the anomaly data stored in the anomaly data storage unit 160 (step S303).

[0063] The image of the abnormal area is processed by the data generating unit 530 and added to the image of the normal data to generate simulated abnormal data (step S304). The image of the abnormal area is processed by, for example, scaling, affine transformation, translation, rotation, or projective transformation. The processing does not have to be of only one type, and a combination of multiple types of processing may be applied to the image of the abnormal area. Furthermore, multiple processed images may be generated by performing different processing on the image of the abnormal area, and each processed image may be added to the image of the normal data to generate multiple simulated abnormal data.

[0064] When the simulated abnormality data is generated, the generated simulated abnormality data is stored in the abnormality data storage unit 160 (step S305). That is, the simulated abnormality data is stored in the abnormality data storage unit 160 because it is variable abnormality data obtained by processing image data including an abnormality detected by the anomaly detection model.

[0065] When the simulated abnormal data is stored in the abnormal data storage unit 160, abnormal data including the newly stored simulated abnormal data is acquired by the learning control unit 520, and the acquired abnormal data is used to re-learn the anomaly classification model of the anomaly classification unit 510 (step S306). In re-learning the anomaly classification model, learning is performed using not only abnormal data in which an abnormality has been detected, but also simulated abnormal data that is a variation of the abnormal data, so that the accuracy of the anomaly classification model can be improved by machine learning using sufficient data.

[0066] As described above, according to this embodiment, abnormal data in which an abnormality is detected by the anomaly detection model is processed to generate simulated abnormal data that is a variation of the abnormal data, and the abnormal data including the simulated abnormal data is used to re-learn the anomaly classification model. Therefore, learning is performed using a wide range of sufficient abnormal data, and the accuracy of the anomaly classification model can be improved.

[0067] The above-mentioned embodiment 3 can be implemented in combination with the above-mentioned embodiments 1 and 2. That is, in addition to re-learning the anomaly classification model using the simulated anomaly data, re-learning of the anomaly detection model using the simulated data may be performed.

[0068] The learning device 100 according to the above-mentioned embodiments 1 to 3 can be configured using a processor and a memory. Fig. 9 is a block diagram showing an example of a hardware configuration of the learning device 100 according to the embodiments 1 to 3. As shown in Fig. 9, the learning device 100 has a processor 101, a memory 102, an input / output unit 103, and a storage 104.

[0069] The processor 101 includes, for example, a central processing unit (CPU), a field programmable gate array (FPGA), or a digital signal processor (DSP), and controls the entire learning device 100 and executes various types of arithmetic processing.

[0070] The memory 102 includes, for example, a random access memory (RAM) or a read only memory (ROM), and stores information used in the arithmetic processing executed by the processor 101.

[0071] The input / output unit 103 is an interface through which the user inputs information and outputs information to the user, and may include, for example, a keyboard, a display, a touch panel, a microphone, or a speaker.

[0072] The storage 104 includes, for example, a hard disk drive (HDD) or a solid state drive (SSD), and stores various types of data.

[0073] The processing by the learning device 100 described in the above first to third embodiments can also be written as a computer-executable program. In this case, the program can be stored in a computer-readable and non-transitory recording medium and introduced into the computer. Examples of such recording media include portable recording media such as CD-ROMs, DVD disks, and USB memories, and semiconductor memories such as flash memories. [Explanation of symbols]

[0074] 101 Processor 102 Memory 103 Input / output section 104 Storage 110 Image acquisition unit 120 Abnormality detection unit 130 Normal data storage unit 140, 320, 520 Learning control unit 150 Result Confirmation Department 160 Abnormal data storage unit 170 False positive data storage unit 180, 330, 530 Data Generation Section 310 Area division part 510 Abnormal Classification Section

Claims

1. A detection unit that detects an abnormality from the image data; a generating unit that processes the image data in which an abnormality has been detected by the detecting unit and generates simulation data similar to the image data; a learning control unit that controls learning of a model using the simulation data generated by the generation unit; A learning device having the above configuration.

2. The detection unit is Detect anomalies from image data using an anomaly detection model, The generation unit is generating simulation data in which an anomaly may be erroneously detected by processing false detection data in which an anomaly has been erroneously detected among the image data in which an anomaly has been detected by the detection unit; The learning control unit is Controlling learning of the anomaly detection model using the simulated data generated by the generation unit The learning device according to claim 1 .

3. A dividing unit that divides an area in the image data, The generation unit is The false positive data is processed based on the regions divided by the dividing unit. The learning device according to claim 2 .

4. A classification unit that classifies an anomaly detected by the detection unit using an anomaly classification model, The generation unit is processing abnormal data that contains a true abnormality among the image data in which an abnormality has been detected by the detection unit, to generate simulated abnormal data that is a variation of the abnormal data; The learning control unit is Controlling learning of the anomaly classification model using the simulated anomaly data generated by the generation unit The learning device according to claim 1 .

5. A detection unit that detects an abnormality from the image data; a generating unit that processes the image data in which an abnormality has been detected by the detecting unit and generates simulation data similar to the image data; A simulated data generating device having the following:

6. Detect abnormalities from image data, Processing the image data in which an anomaly was detected to generate simulated data similar to the image data; Train the model using the generated simulated data A learning method with processes.

Images