JP2022158518A - Learning data selection program, learning data selection method, and information processing device - Google Patents

Abstract

To select appropriate learning data to be used in machine learning.SOLUTION: An information processing device 101 calculates a feature quantity representing a feature of a frame image included in a moving image of an abnormality detection object. The information processing device 101 calculates a period of the moving image based on a temporal change of the calculated feature quantity of the frame image. The information processing device 101 calculates, for each of a plurality of sections into which the moving image is divided according to the calculated period, a difference regarding the feature quantity of the frame image between each section and another section. The information processing device 101 identifies, based on the difference calculated for each section, a section including a frame image in an abnormal state among the plurality of sections. The information processing device 101 determines frame images corresponding to other sections, which are different from the identified section among the plurality of sections, as learning data to be used for learning of a model M that detects an abnormality of the abnormality detection object.SELECTED DRAWING: Figure 1

Description

The present invention relates to a learning data selection program, a learning data selection method, and an information processing apparatus.

In recent years, the introduction of anomaly detection technology based on machine learning using images is progressing for the purpose of monitoring the occurrence of anomalies in automatic production lines in factories. Anomaly detection using images often uses so-called semi-supervised learning (for example, AutoEncoder, etc.) in which only images in a normal state (normal images) are used as learning data.

As a prior art, for example, there is a method for determining an abnormality of a package by analyzing a still image or moving image output from a pre-inspection camera that photographs the package carried by the pre-inspection conveyor. In addition, the folded printed matter is imaged, and the autocorrelation function of each image captured at a fixed cycle timing within the imaging area is calculated while shifting the imaging position. There is a technique for using a reference image for inspection.

Japanese Patent Application Laid-Open No. 2020-46270 JP-A-11-53553

However, with the conventional technology, it is difficult to select appropriate learning data to be used for machine learning from data in which normal state images and abnormal state images are mixed.

In one aspect, an object of the present invention is to select appropriate learning data to be used for machine learning.

In one embodiment, a feature amount representing a feature of a frame image included in a moving image of an anomaly detection target is calculated, and a period of the moving image is calculated based on a temporal change in the calculated feature amount of the frame image. , for each section of a plurality of sections obtained by dividing the moving image according to the calculated period, calculating the difference regarding the feature amount of the frame image between each section and another section; Based on the difference, a section including a frame image in an abnormal state is specified among the plurality of sections, and a frame image corresponding to a section other than the specified section among the plurality of sections is detected by the abnormality detection. A learning data selection program is provided for determining learning data to be used for learning a model for detecting anomalies of a target.

According to one aspect of the present invention, it is possible to select appropriate learning data to be used for machine learning.

FIG. 1 is an explanatory diagram of an example of a learning data selection method according to an embodiment. FIG. 2 is an explanatory diagram showing a system configuration example of the information processing system 200. As shown in FIG. FIG. 3 is a block diagram showing a hardware configuration example of the learning data selection device 201. As shown in FIG. FIG. 4 is an explanatory diagram showing a specific example of an application scene for detecting an abnormality. FIG. 5 is an explanatory diagram showing an example of the contents stored in the frame image DB 220. As shown in FIG. FIG. 6 is a block diagram showing a functional configuration example of the learning data selection device 201. As shown in FIG. FIG. 7 is an explanatory diagram showing a learning example of the encoder. FIG. 8 is an explanatory diagram showing an example of the contents of the feature amount signal table 230. As shown in FIG. FIG. 9 is an explanatory diagram showing an example of updating the feature amount signal table 230. As shown in FIG. FIG. 10 is an explanatory diagram showing an example of processing for calculating the period T of the moving image Vd. FIG. 11 is an explanatory diagram illustrating an example of segment division processing. FIG. 12 is an explanatory diagram showing an example of determination of learning data. FIG. 13 is a flowchart (Part 1) showing an example of the data selection processing procedure of the learning data selection device 201 . FIG. 14 is a flowchart (part 2) showing an example of the data selection processing procedure of the learning data selection device 201 . FIG. 15 is an explanatory diagram showing an example of selection of learning data.

Exemplary embodiments of a learning data selection program, a learning data selection method, and an information processing apparatus according to the present invention will be described below in detail with reference to the drawings.

(Embodiment)
FIG. 1 is an explanatory diagram of an example of a learning data selection method according to an embodiment. In FIG. 1, an information processing apparatus 101 is a computer that determines learning data used for model M learning. The model M is a learning model for detecting an anomaly in an anomaly detection target, and is generated, for example, by machine learning using a neural network.

Anomaly detection target is, for example, an automatic production line in a factory, a robot that operates periodically, and the like. An automated production line is, for example, an assembly line assembly process designed to mass-produce the same and similar products. In an automated production line, products flow periodically. Examples of products include drinking water, electronic equipment, automobile parts, and the like.

Here, there is a case where the same position of the automatic production line is continuously photographed by a camera, and an abnormality in the product of the automatic production line is detected using the photographed image. The abnormal state is, for example, a state in which the label of the product is peeled off or the product is overturned. In anomaly detection using images, for example, a group of images in which only images in a normal state (normal images) are extracted is learned by semi-supervised learning. A method of evaluating the difference with the image is used. In this method, since the abnormal state is not reconstructed, the abnormal portion appears as a difference.

In order to accurately detect anomalies in semi-supervised learning, it is required to extract only normal images from captured images as learning data. However, there is a problem that manually extracting only normal images from data in which normal images and abnormal images are mixed requires time and effort.

On the other hand, if data in which normal state images and abnormal state images are mixed is simply used as learning data for semi-supervised learning, for example, anomalies are also reconstructed, resulting in lower anomaly detection accuracy. Therefore, it is desired to generate a learning model capable of detecting an image in an abnormal state without manually extracting only an image in a normal state.

Here, there is a technique called Isolation Forest as a conventional technique for narrowing down normal data from data in which images in a normal state and images in an abnormal state are mixed without a teacher. Isolation Forest (Prior Art 1) regards a part of the distribution in the feature space that is out of the range as abnormal.

However, prior art 1 has a problem that it is difficult to separate abnormal data when the distance between normal data and abnormal data is short in the feature space. In prior art 1, a portion that is out of existence distribution is regarded as an abnormal value candidate, and an abnormality cannot be removed unless the difference in the latent space between normal data and abnormal data is large.

For example, in the case of a moving image of an automated production line, the difference between two normal data shot at different timings and the difference between normal data and abnormal data shot at the same timing may be larger in the former. In such a case, the prior art 1 may judge abnormal data as normal and normal data as abnormal.

Further, there is a conventional technique called Robust Auto Encoder for narrowing down normal data from data in which images in a normal state and images in an abnormal state are mixed without a teacher. The Robust Auto Encoder (prior art 2) separates an image into a background part with little movement and a changed part, and learns to reduce the reconstruction error of the background part by removing the changed part.

However, according to the prior art 2, images with a large feature difference from the common part can be regarded as abnormal and removed from the learning data. data) is difficult. In other words, even with prior art 2, abnormal data whose difference from normal data is smaller than the difference between normal data cannot be removed.

As described above, in the conventional technology, an image of an abnormal state is obtained from data in which an image of a normal state and an image of an abnormal state are mixed, and the difference in features between the image of the normal state and the image of the abnormal state is not large. It is difficult to remove the images and train the model. If it is not possible to learn by removing images in an abnormal state, it is impossible to generate a learning model capable of accurately detecting images in an abnormal state.

Therefore, in the present embodiment, in a moving image with periodicity, images in a normal state and images in an abnormal state are mixed without being sorted out, and even if the difference in features between the two is small, machine learning can be performed. A learning data selection method for selecting appropriate learning data for A processing example of the information processing apparatus 101 will be described below.

(1) The information processing apparatus 101 calculates a feature quantity representing a feature of a frame image included in a moving image of an anomaly detection target. Here, the moving image of the anomaly detection target is a periodic moving image in which similar images are periodically repeated. The feature amount of the frame image is information representing the feature of the frame image, and may be, for example, a single value or a feature vector containing a plurality of components (elements). It is possible to arbitrarily set what kind of value is to be extracted as a feature amount.

In the example of FIG. 1, the object of abnormality detection is "a product on an automatic production line in a factory (for example, drinking water in a PET bottle)", and the moving image of the object of abnormality detection is "moving image 110". In the automatic production line, products flow periodically, so similar images are cyclically repeated in the moving image 110 of the automatic production line. In this case, the information processing apparatus 101 calculates, for example, feature amounts representing features of frame images (for example, frame images 111 to 113) included in the moving image 110. FIG.

(2) The information processing apparatus 101 calculates the period of the moving image based on the calculated temporal change of the feature amount of the frame image. Here, when similar images are periodically repeated in a moving image of an anomaly detection target, it can be said that temporal changes in the feature amount of frame images included in the moving image have periodicity.

For this reason, the information processing apparatus 101 determines the periodicity from the temporal change of the feature amount of the frame images included in the moving image. Specifically, for example, the information processing apparatus 101 calculates the period of the moving image based on the result obtained by frequency analysis of the feature amount signal (signal waveform) 120 indicating the temporal change of the calculated feature amount of the frame image. do.

In the example of FIG. 1, the information processing apparatus 101 calculates the period T1 of the moving image 110 from the frequency obtained by frequency analysis of the feature amount signal 120, for example. The feature amount signal 120 indicates the temporal change of the feature amount representing the features of the frame images (for example, the frame images 111 to 113) included in the moving image 110 (vertical axis: feature amount, horizontal axis: frame number). A frame number is an identifier of a frame image given in chronological order.

(3) The information processing apparatus 101 calculates, for each of a plurality of sections into which the moving image is divided according to the calculated period, the difference regarding the feature amount of the frame image between each section and another section. Other sections are, for example, sections adjacent to each section. A section is designated by, for example, a frame number.

The difference regarding the feature amount of the frame image is the difference obtained by comparing the feature amounts of the frame images of the sections. The abnormal state of the abnormality detection target often continues for a certain period of time within the section. For this reason, there tends to be a difference in the feature amount between the whole section including the frame image in the abnormal state and the other sections.

In the example of FIG. 1, the information processing apparatus 101 divides the moving image 110 according to the cycle T1 into a plurality of sections (for example, sections 110-1 to 110-3). A difference relating to the feature amount of the frame image is calculated. For example, for the section 110-2, the information processing apparatus 101 calculates the difference regarding the feature amounts of the frame images between the section 110-2 and the other sections 110-1 and 110-3.

The difference regarding the frame image feature amount between the section 110-2 and the other sections 110-1 and 110-3 is, for example, the difference between the sections 110-1 and 110-2 and the section 110-2 and 110-3. It may be the maximum value of the differences between them, or it may be the average value of the two differences. Note that the feature amount signal 130 indicates the temporal change of the feature amount representing the features of the frame images included in the moving image 110 (for example, an enlarged display of a portion of the feature amount signal 120).

(4) The information processing apparatus 101 identifies a section including a frame image in an abnormal state among the plurality of sections based on the difference calculated for each section. Specifically, for example, the information processing apparatus 101 may identify a section in which the calculated difference is equal to or greater than a threshold value as a section including a frame image in an abnormal state. The threshold can be set arbitrarily.

In the example of FIG. 1, it is assumed that section 110-2 is identified as a section containing frame images in an abnormal state.

(5) The information processing apparatus 101 determines frame images corresponding to other sections, which are different from the specified section, among the plurality of sections as learning data to be used for learning of the model M for detecting anomaly detection target. The frame images corresponding to the section may be, for example, frame images at regular intervals within the section, or may be all frame images within the section.

In the example of FIG. 1, of the frame images included in the moving image 110, the frame images corresponding to other sections (eg, sections 110-1 and 110-3) different from the section 110-2 are used for model M learning. It is determined by the learning data to be used.

As described above, the information processing apparatus 101 uses the periodicity of a moving image to compare the feature amounts of frame images for each period (section), and detects an abnormal frame image included in a section in which the periodicity is disturbed. can be removed from the training data as a possibility of Accordingly, it is possible to select appropriate learning data to be used for machine learning from a moving image of an anomaly detection target.

For example, in a moving image with periodicity, images in a normal state and images in an abnormal state are mixed without being sorted out, and even if the difference in features between the two is small, the model M can detect anomalies with high accuracy. Appropriate training data can be selected for generating Moreover, when preparing the learning data, it is not necessary to manually sort out the images in the normal state and the images in the abnormal state.

In the example of FIG. 1, the information processing apparatus 101 can remove the frame images included in the period 110-2 in which the periodicity is disturbed from the learning data as possible abnormalities. In addition, by learning the model M using the learning data 140 obtained by removing the frame images included in the period (for example, section 110-2) in which the periodicity is disturbed in the moving image 110, the abnormality of the automatic production line can be detected. A model M that can be detected with high accuracy can be generated.

(System configuration example of information processing system 200)
Next, a system configuration example of an information processing system 200 including the information processing apparatus 101 shown in FIG. 1 will be described. In the following description, a case where the information processing device 101 shown in FIG. 1 is applied to the learning data selection device 201 in the information processing system 200 will be described as an example. The information processing system 200 is applied, for example, to a computer system that detects anomalies using moving images in an automatic production line in a factory.

FIG. 2 is an explanatory diagram showing a system configuration example of the information processing system 200. As shown in FIG. In FIG. 2 , an information processing system 200 includes a learning data selection device 201 and a client device 202 . In information processing system 200 , learning data selection device 201 and client device 202 are connected via wired or wireless network 210 . The network 210 is, for example, the Internet, LAN, WAN (Wide Area Network), or the like.

Here, the learning data selection device 201 has a frame image DB (Database) 220, a feature amount signal table 230, and a learning data DB 240, and determines learning data used for model M learning. The model M is a learning model for detecting anomalies in the anomaly detection target. The learning data selection device 201 is, for example, a server.

The frame image DB 220 stores frame images included in a moving image of an anomaly detection target. The feature amount signal table 230 stores feature amounts (feature vectors) of frame images included in a moving image of an anomaly detection target. The learning data DB 240 stores learning data used for model M learning.

Note that the storage contents of the frame image DB 220 and the feature amount signal table 230 will be described later with reference to FIGS. 5 and 8. FIG.

A client device 202 is a computer used by a user of the information processing system 200 . A user is, for example, an administrator who manages an automatic production line in a factory. The client device 202 is, for example, a PC (Personal Computer), a tablet PC, or the like.

Although the learning data selection device 201 and the client device 202 are provided separately here, the present invention is not limited to this. For example, the learning data selection device 201 may be implemented by the client device 202 . The information processing system 200 may also include a plurality of client devices 202 .

(Hardware configuration example of learning data selection device 201)
FIG. 3 is a block diagram showing a hardware configuration example of the learning data selection device 201. As shown in FIG. 3, the learning data selection device 201 includes a CPU (Central Processing Unit) 301, a memory 302, a disk drive 303, a disk 304, a communication I/F (Interface) 305, and a portable recording medium I/F 306. and a portable recording medium 307 . Also, each component is connected by a bus 300 .

Here, the CPU 301 controls the learning data selection device 201 as a whole. The CPU 301 may have multiple cores. The memory 302 has, for example, a ROM (Read Only Memory), a RAM (Random Access Memory), a flash ROM, and the like. Specifically, for example, a flash ROM stores an OS (Operating System) program, a ROM stores application programs, and a RAM is used as a work area for the CPU 301 . A program stored in the memory 302 is loaded into the CPU 301 to cause the CPU 301 to execute coded processing.

The disk drive 303 controls data read/write with respect to the disk 304 under the control of the CPU 301 . The disk 304 stores data written under the control of the disk drive 303 . Examples of the disk 304 include a magnetic disk and an optical disk.

The communication I/F 305 is connected to the network 210 through a communication line, and is connected to an external computer (for example, the client device 202 shown in FIG. 2) via the network 210 . A communication I/F 305 serves as an interface between the network 210 and the inside of the apparatus, and controls input/output of data from an external computer. For the communication I/F 305, for example, a modem or a LAN adapter can be adopted.

A portable recording medium I/F 306 controls reading/writing of data from/to a portable recording medium 307 under the control of the CPU 301 . The portable recording medium 307 stores data written under control of the portable recording medium I/F 306 . Examples of the portable recording medium 307 include CD (Compact Disc)-ROM, DVD (Digital Versatile Disk), USB (Universal Serial Bus) memory, and the like.

Note that the learning data selection device 201 may have, for example, a GPU (Graphics Processing Unit), an input device, a display, etc., in addition to the components described above. Also, the client device 202 shown in FIG. 2 can be realized by a hardware configuration similar to that of the learning data selection device 201 . However, the client device 202 has, for example, an input device and a display in addition to the components described above.

(Specific examples of applicable scenes)
Next, a specific example of an application scene for detecting an anomaly will be described using a moving image of an anomaly detection target.

FIG. 4 is an explanatory diagram showing a specific example of an application scene for detecting an abnormality. In FIG. 4, a camera 400 is an imaging device having a function of shooting moving images. The camera 400 is installed in a factory, for example, and photographs the same position of the automatic production line that is the object of abnormality detection. As a result, products passing in front of the camera 400 are periodically photographed.

A moving image 410 is a moving image of a product on an automatic production line, and includes frame images 411 to 414, for example. Camera 400 may have a communication function. In this case, camera 400 may be connectable to learning data selection device 201 and client device 202 via network 210 shown in FIG.

(Stored contents of frame image DB 220)
Next, the storage contents of the frame image DB 220 of the learning data selection device 201 will be described with reference to FIG. Note that the frame image DB 220, the feature amount signal table 230, and the learning data DB 240 are realized by, for example, storage devices such as the memory 302 and disk 304 shown in FIG.

FIG. 5 is an explanatory diagram showing an example of the contents stored in the frame image DB 220. As shown in FIG. In FIG. 5, the frame image DB 220 has fields for frame numbers and frame images. ) is stored as a record.

Here, the frame number is an identifier that identifies a frame image included in the moving image. Frame numbers are assigned to frame images in chronological order. A frame image is a frame image (image data) included in a moving image. For example, the time series frame image data 500-1 indicates the frame image of frame number "1".

(Example of functional configuration of learning data selection device 201)
FIG. 6 is a block diagram showing a functional configuration example of the learning data selection device 201. As shown in FIG. In FIG. 6 , learning data selection device 201 includes acquisition section 601 , first calculation section 602 , second calculation section 603 , identification section 604 , determination section 605 , and output section 606 . Acquisition unit 601 to output unit 606 are functions of a control unit. or by the communication I/F 305, the function is realized. The processing results of each functional unit are stored in a storage device such as the memory 302 or disk 304, for example.

Acquisition unit 601 acquires video Vd in which an abnormality detection target is captured. The moving image Vd is, for example, a periodic moving image in which an abnormality detection target is photographed for a predetermined period of time. A moving image with periodicity is a moving image in which similar images are repeated periodically. A moving image 410 shown in FIG. 4 is an example of the moving image Vd.

The predetermined period is determined, for example, according to the abnormality detection target. For example, if the abnormality detection target is an automatic production line in which products are delivered at intervals of several seconds, the predetermined period is about several minutes. Specifically, for example, the acquisition unit 601 acquires the received video Vd by receiving the video Vd from the client device 202 shown in FIG.

Further, the acquiring unit 601 may acquire the designated moving image Vd from a moving image DB (not shown) by receiving the designation of the moving image Vd from the client device 202 . Further, the acquisition unit 601 may acquire the moving image Vd input by the user's operation using an input device (not shown). Further, the acquisition unit 601 may acquire the moving image Vd shot by the camera 400 from the camera 400 shown in FIG.

The first calculator 602 calculates feature amounts of frame images included in the acquired moving image Vd. The feature amount of the frame image is, for example, a feature amount vector representing the feature of the frame image. A feature vector includes a plurality of components (elements). A plurality of components is a combination of values representing features of the frame image. As the plurality of components, for example, any values (brightness, contrast, product-likeness, etc.) that can be extracted by existing feature extraction techniques may be used.

Here, the first calculation unit 602 includes an encoder learning unit 611 and a vector calculation unit 612, for example.

The encoder learning unit 611 learns an encoder ec (converter) that converts a frame image into a feature amount vector. Specifically, for example, the encoder learning unit 611 extracts frame images from the moving image Vd at regular time intervals. The fixed time can be arbitrarily set, and is set, for example, to a time such that frame images are extracted from the moving image Vd at intervals of several frames to several tens of frames.

Next, the encoder learning unit 611 assigns frame numbers to the extracted frame images in chronological order. Frame images assigned frame numbers are stored in the frame image DB 220 shown in FIG. 5, for example. Then, the encoder learning unit 611 learns the encoder ec using the extracted frame images as learning data.

Here, a learning example of the encoder ec will be described with reference to FIG. Here, a case where an encoder ec (neural network) is learned by an autoencoder will be described.

FIG. 7 is an explanatory diagram showing a learning example of the encoder. The encoder learning unit 611 uses an autoencoder to learn an encoder ec (neural network) that receives all frame images stored in the frame image DB 220 as input and outputs a reconstructed image that is close to the input.

Specifically, for example, the encoder learning unit 611 may set the encoder ec ( neural networks).

The autoencoder's latent space is dimensionally compressed while preserving the global features of the image. Therefore, by using the feature amount vector in the feature space (latent space) compressed by the encoder ec of the autoencoder, it is possible to obtain the feature amount that reflects the overall features of the image.

Although the encoder ec is learned here using the frame images included in the moving image Vd, the present invention is not limited to this. For example, the encoder learning unit 611 may acquire the learned encoder ec. Specifically, for example, the encoder learning unit 611 regards the encoder ec that has been learned using frame images included in another moving image of the anomaly detection target shot at the same position at a different timing as the learned encoder ec. You may choose to acquire it. The learned encoder ec may be acquired, for example, by a user's operation using an input device (not shown) or from another computer (eg, client device 202).

The vector calculation unit 612 uses the learned encoder ec to calculate feature amount vectors of frame images included in the moving image Vd. Specifically, for example, the vector calculation unit 612 inputs each frame image stored in the frame image DB 220 to the learned encoder ec, and outputs the feature quantity in the intermediate layer as a feature quantity vector.

The feature amount vector of each output frame image is stored in the feature amount signal table 230, for example. Here, the storage contents of the feature quantity signal table 230 will be described with reference to FIG.

FIG. 8 is an explanatory diagram showing an example of the contents of the feature amount signal table 230. As shown in FIG. 8, the feature amount signal table 230 has fields of frame number, feature amount vector, and window number. By setting information in each field, feature amount signal information (for example, feature amount signal information 800-1 800-3) are stored as records.

Here, the frame number is an identifier that identifies a frame image extracted from the moving image Vd. The feature amount vector is a feature amount vector representing the feature of the frame image calculated using the learned encoder ec. A window number is an identifier that identifies a section to which a frame image belongs. In the initial state, the window number is "-(Null)".

Returning to the description of FIG. 6, the second calculator 603 calculates the period T of the moving image Vd based on the calculated temporal change in the feature amount of the frame image. Specifically, for example, the second calculation unit 603 performs frequency analysis of a signal waveform (principal component signal waveform) that indicates the temporal change of the principal component of the calculated feature amount vector of the frame image. Based on this, the period T is calculated.

Here, the second calculator 603 includes, for example, a principal component analyzer 613 and a frequency analyzer 614 .

A principal component analysis unit 613 identifies the principal component of the calculated feature amount vector of the frame image. Specifically, for example, the principal component analysis unit 613 refers to the feature quantity signal table 230, arranges the calculated feature quantity vectors of the frame images in chronological order, performs principal component analysis, and performs principal component analysis. The component with the largest variation in is specified as the principal component.

The frequency analysis unit 614 performs frequency analysis of a feature amount signal (principal component signal waveform) that indicates the temporal change of the principal component of the specified feature amount vector. Specifically, for example, the frequency analysis unit 614 refers to the feature amount signal table 230 to identify the principal component signal waveform of the feature amount vector. Then, frequency analysis section 614 frequency-analyzes the identified main component signal waveform to identify the frequency with the highest power. In this case, the second calculator 603 sets the specified frequency as the period T of the moving image Vd.

An example of processing for calculating the period T of the moving image Vd will be described later with reference to FIG.

Note that here, the period T of the moving image Vd is calculated based on the temporal change in the feature amount of the frame images included in the moving image Vd, but the present invention is not limited to this. For example, the second calculation unit 603 may acquire a period specified in advance as the period T of the moving image Vd. Specifically, for example, the second calculation unit 603 calculates the period calculated based on the temporal change in the feature amount of the frame images included in the different moving images captured at the same position at different timings of the abnormality detection target. , may be obtained as the period T of the moving image Vd. Further, the second calculation unit 603 acquires a period designated by a user's operation using an input device (not shown) or from another computer (for example, the client device 202) as the period T of the video Vd. may

The specifying unit 604 calculates the difference df regarding the feature amount of the frame image between each section and another section, for each section of a plurality of sections obtained by dividing the moving image Vd according to the calculated cycle T. Here, the difference df is, for example, the difference between the main component signal waveforms. However, the difference df may be represented by, for example, the sum or average of the differences of the component signal waveforms of the feature vector.

Specifically, for example, the specifying unit 604 refers to the feature amount signal table 230, uses the calculated period T as the window width, and divides the frame image group (feature amount vector group) extracted from the moving image Vd. , into a plurality of intervals (windows). A window number identifying a section to which each frame image (feature vector group) belongs is set in the feature signal table 230, for example.

FIG. 9 is an explanatory diagram showing an example of updating the feature amount signal table 230. As shown in FIG. In FIG. 9, a window number identifying a section to which a frame image (feature vector) of each frame number belongs is set in the window number field of each feature signal information in the feature signal table 230 .

For example, window number "1" is set in each of the feature amount signal information 800-1 to 800-3.

An example of segment division processing will be described later with reference to FIG. 11 .

The specifying unit 604 refers to the feature amount signal table 230 and compares the feature amount vector (for example, main component signal waveform) of each section (window) with neighboring k sections to calculate the difference df. For example, let the principal component signal value of the feature amount vector of each frame image included in each section (window) be "v(t)=[x(t), x(t+1), x(t+2)...]". .

In this case, the identifying unit 604 calculates “|v(t)−v(t′)|̂2” for t′ other than t≠t′, for example. Then, the specifying unit 604 sets “P(t)=min(|v(t)−v(t′)|̂2)” as the difference df (where k=1). The difference df corresponds to a value (abnormality degree) indicating the degree to which frame images in an abnormal state are included in that section.

When k>2, the identification unit 604, for example, extracts k pieces from the calculation results of the intervals with smaller “|v(t)−v(t′)|^2”, and extracts “P(t )=avg(|v(t)−v(t1′)|̂2,|v(t)−v(t2′)|̂2 . . . )” is the difference df. A calculation example of the difference df (abnormality degree) using the k nearest neighbor method will be described later.

Then, the specifying unit 604 specifies a section including a frame image in an abnormal state among the plurality of sections based on the difference df calculated for each section. Specifically, for example, the identifying unit 604 identifies a section in which the calculated difference df is equal to or greater than the threshold Th as a section containing a frame image in an abnormal state. The threshold Th can be arbitrarily set, and is set according to, for example, an abnormality detection target.

In the following description, a section containing frame images in an abnormal state may be referred to as an "abnormal data section".

Further, the specifying unit 604 may specify, among the plurality of sections, sections excluding a predetermined number of sections in which the calculated difference df is relatively small as abnormal data sections. In other words, the specifying unit 604 may specify, for example, a predetermined number of sections in descending order of the difference df as abnormal data sections.

The predetermined number can be arbitrarily set, and is set, for example, according to the number of learning data necessary for learning the model M (required number of samples). For example, assume that one section is "60 frames" and the required number of samples is "600 frames". In this case, the identifying unit 604 identifies, among the plurality of sections, remaining sections excluding the top 10 sections with the smallest differences df as abnormal data sections.

The determination unit 605 determines frame images corresponding to other sections that are different from the identified section (abnormal data section) among the plurality of sections as learning data to be used for model M learning. That is, the determination unit 605 excludes frame images corresponding to the abnormal data section from the learning data used for learning the model M among the frame images included in the moving image Vd. The model M is a learning model that detects anomalies in the anomaly detection target.

Specifically, for example, the determination unit 605 refers to the frame image DB 220 and determines, as learning data, a frame image corresponding to a section other than the abnormal data section. Frame images determined as learning data are stored, for example, in the learning data DB 240 (see FIG. 2).

An example of determining learning data will be described later with reference to FIG. 12 .

For example, the determination unit 605 may determine all frame images other than the frame images belonging to the abnormal data section among the frame images included in the moving image Vd as learning data. That is, the determining unit 605 may determine frame images that are not stored in the frame image DB 220 as learning data.

The output unit 606 outputs the determined learning data. Specifically, for example, the output unit 606 outputs the stored frame images as learning data used for learning of the model M for detecting anomalies of the anomaly detection targets. The output format of the output unit 606 includes, for example, storage in a storage device such as the memory 302 and disk 304, transmission to another computer (for example, the client device 202) via the communication I/F 305, and the like.

Note that the learning data selection device 201 may learn the model M using the determined learning data. The model M is generated, for example, by machine learning using a neural network. As a result, it is possible to generate a model M capable of accurately detecting an abnormality in an abnormality detection target (for example, an automatic production line).

The functional units (acquisition unit 601 to output unit 606) of the learning data selection device 201 described above are realized by, for example, a plurality of computers (for example, the learning data selection device 201 and the client device 202) operating in cooperation. You can decide.

(Example of processing for calculating period T of moving image Vd)
Next, an example of processing for calculating the period T of the moving image Vd will be described with reference to FIG.

FIG. 10 is an explanatory diagram showing an example of processing for calculating the period T of the moving image Vd. In FIG. 10, a graph 1000 is a principal component signal waveform obtained by plotting the principal component of the feature amount vector of each frame image stored in the feature amount signal table 230 in frame number order (time series).

A frequency analysis unit 614 performs frequency analysis of the principal component signal waveform (graph 1000). A graph 1010 shows the result obtained by frequency analysis of the main component signal waveform (graph 1000) (vertical axis: power, horizontal axis: frequency component). Power, for example, corresponds to the amplitude of the waveform.

Here, assuming that the principal component signal of the feature amount vector is F(t), F(t) can be expressed by the following formula (1) by discrete Fourier transform. However, f(x) indicates the intensity of frequency x. N is a natural number. i is the imaginary unit. π is the circular constant.

For example, the frequency analysis unit 614 performs a discrete Fourier transform on F(t) in the form of the above formula (1) by FFT (Fast Fourier Transform) or the like, and calculates the following formula ( 2) is used to determine the power of each frequency x.

Power of frequency x=|f(x)|^2 (2)

The frequency analysis unit 614 refers to the frequency analysis result (graph 1010) to identify the _maximum power frequency xmax. The second calculator 603 sets the identified frequency x _max as the period T (window width) of the moving image Vd. As a result, the strongest frequency component in the feature quantity signal on the principal component axis can be set as the period T (window width) of the moving image Vd.

(Example of section division processing)
Next, with reference to FIG. 11, an example of segment dividing processing when dividing the moving image Vd into a plurality of segments will be described.

FIG. 11 is an explanatory diagram illustrating an example of segment division processing. In FIG. 11, a graph 1000 is a principal component signal waveform obtained by plotting the principal component of the feature amount vector of each frame image stored in the feature amount signal table 230 in frame number order (time series). The specifying unit 604 divides the main component signal waveform (graph 1000) into a plurality of sections by, for example, the calculated period T (window width).

In the example of FIG. 11, it is divided into sections S1 to S3. Each section S1 to S3 is designated by a frame number. Assuming that the section S2 is the target section, the specifying unit 604 compares the main component signal waveform of the section S2 with the main component signal waveforms of the sections S1 and S3, for example, using the sections S1 and S3 as comparison sections, thereby obtaining the difference df. (degree of anomaly) is calculated.

(Calculation example of difference df (abnormality))
Here, an example of calculating the difference df (degree of anomaly) using the k nearest neighbor method will be described. Here, a case will be described in which the data of the window width is extracted using a sliding window and the principal component signal value of the feature amount vector is calculated.

For example, if there is a moving image of 500 frames and the window width is "10", the specifying unit 604 calculates vector values v(1) to v(490) as follows. However, Xn indicates the value of the main signal component of the frame image at time n.

v(1)=(x1,x2,x3...,x10)
v(2)=(x2,x3,x4...,x11)
・・・
v(490) = (x491, x492..., x500)

Next, the identifying unit 604 calculates the difference between each v(t) and another v(t). For example, the difference between v(1) and v(2) is "(x1-x2)^2+(x2-x3)^2+...". Also, the difference between v(1) and v(3) is "(x1-x3)^2+(x2-x4)^2+...". Then, the specifying unit 604 calculates an average value of k values having the smallest difference among the calculated differences as the difference df (abnormality degree) for the section at time t.

(Example of determination of learning data)
Next, an example of learning data determination will be described with reference to FIG.

FIG. 12 is an explanatory diagram showing an example of determination of learning data. In FIG. 12, it is assumed that section S2 is identified as an abnormal data section among sections S1 to S3. In this case, the determination unit 605 refers to the frame image DB 220 and determines, as learning data, frame images corresponding to sections S1 and S3 different from the abnormal data section S2.

For example, let the section S1 be the section with the window number "1", the section S2 be the section with the window number "2", and the section S3 be the section with the window number "3". The determination unit 605 determines frame images corresponding to window numbers “1” and “3” as learning data. Then, the determining unit 605 extracts frame images corresponding to window numbers “1” and “3” from the frame image DB 220 and stores the extracted frame images in the learning data DB 240 .

As a result, the learning data used for learning the model M for detecting anomalies in the anomaly detection target (automatic production line) can be accumulated in the learning data DB 240 .

(Data selection processing procedure of learning data selection device 201)
Next, a data selection processing procedure of the learning data selection device 201 will be described.

13 and 14 are flowcharts showing an example of the data selection processing procedure of the learning data selection device 201. FIG. In the flowchart of FIG. 13 , first, the learning data selection device 201 determines whether or not a moving image Vd in which an abnormality detection target is photographed has been obtained (step S1301).

Here, the learning data selection device 201 waits to acquire the moving image Vd (step S1301: No). Then, when the learning data selection device 201 acquires the moving image Vd (step S1301: Yes), it extracts frame images from the acquired moving image Vd at regular time intervals (step S1302).

Next, the learning data selection device 201 assigns frame numbers to the extracted frame images in chronological order, and stores the frame images in the frame image DB 220 (step S1303). Then, the learning data selection device 201 refers to the frame image DB 220 and learns the encoder ec using the extracted frame images as learning data (step S1304).

Next, the learning data selection device 201 refers to the frame image DB 220, inputs each extracted frame image to the learned encoder ec, and calculates the feature amount vector of each frame image (step S1305). Then, the learning data selection device 201 arranges the calculated feature vectors of the frame images in chronological order, performs principal component analysis, and identifies the principal components of the feature vectors (step S1306).

Next, the learning data selection device 201 performs frequency analysis on the identified feature signal of the principal component (principal component signal waveform) to identify the maximum power frequency (step S1307). Then, the learning data selection device 201 sets the specified maximum power frequency as the period T of the moving image Vd (step S1308), and proceeds to step S1401 shown in FIG.

In the flowchart of FIG. 14, first, the learning data selection device 201 uses the period T of the moving image Vd as the window width, and divides the feature amount signal of the main component into a plurality of sections by the window width (step S1401). This process corresponds to the process of dividing the moving image Vd into a plurality of sections according to the period T of the moving image Vd.

Next, the learning data selection device 201 selects an unselected section from the plurality of divided sections (step S1402). Then, the learning data selection device 201 compares the feature amount signal of the selected section (here, referred to as "interested section") with neighboring k sections to calculate the difference df for the noticed section (step S1403).

Next, the learning data selection device 201 determines whether or not the calculated difference df is greater than or equal to the threshold Th (step S1404). If the difference df is greater than or equal to the threshold Th (step S1404: Yes), the learning data selection device 201 proceeds to step S1406. On the other hand, if the difference df is less than the threshold Th (step S1404: No), the learning data selection device 201 determines the frame image corresponding to the section of interest as learning data (step S1405).

Then, the learning data selection device 201 determines whether or not there is an unselected section that has not been selected from the plurality of divided sections (step S1406). Here, if there is an unselected section (step S1406: Yes), the learning data selection device 201 returns to step S1402.

On the other hand, if there is no unselected section (step S1406: No), the learning data selection device 201 outputs the frame image determined as the learning data (step S1407), and ends the series of processing according to this flowchart.

As a result, by using the periodicity of the video Vd in which the abnormality detection target is shot, the frame images included in the period in which the periodicity is disturbed in the video Vd are regarded as having the possibility of being abnormal, and the model M is learned. can be excluded from the training data used for

Although the encoder ec is learned in step S1304, the present invention is not limited to this. For example, if there is a learned encoder ec, the learning data selection device 201 may omit the process of step S1304.

Also, in step S1308, the period T of the moving image Vd is calculated, but the present invention is not limited to this. For example, if there is a cycle T specified in advance, the learning data selection device 201 may omit the processing of steps S1307 and S1308.

Also, in step S1404, learning data is determined when the difference df is less than the threshold value Th, but the present invention is not limited to this. For example, the learning data selection device 201 may select a predetermined number of sections from the smallest difference df to use as learning data. In this case, the processes of steps S1404 and S1405 are omitted, and a process of selecting the upper predetermined number of sections from the smaller difference df and determining them as learning data is added before step S1407.

As described above, according to the learning data selection device 201 according to the embodiment, the feature amount of the frame image included in the moving image Vd in which the abnormality detection target is shot is calculated, and the feature amount of the calculated frame image changes over time. , the period T of the moving image Vd can be calculated. The moving image Vd is, for example, a periodic moving image.

As a result, by utilizing the fact that similar images are periodically repeated in a moving image Vd of an anomaly detection target such as an automatic production line, the periodicity of the temporal change in the feature amount of the frame images included in the moving image Vd can be obtained. , the period T of the moving image Vd can be obtained.

Further, according to the learning data selection device 201, for each of a plurality of sections obtained by dividing the moving image Vd according to the calculated period T, the difference df regarding the feature amount of the frame image between each section and another section is calculated. Based on the difference df calculated for each section, a section (abnormal data section) containing a frame image in an abnormal state can be specified among the plurality of sections. Then, according to the learning data selection device 201, frame images corresponding to other sections different from the specified section among the plurality of sections can be determined as learning data to be used for model M learning.

As a result, by using the periodicity of the video Vd in which the abnormality detection target is shot, the frame images included in the period in which the periodicity is disturbed in the video Vd are regarded as having the possibility of being abnormal, and the model M is learned. can be excluded from the training data used for

Further, according to the learning data selection device 201, a section in which the calculated difference df is equal to or greater than the threshold Th can be identified as a section containing a frame image in an abnormal state.

As a result, it is possible to accurately identify a section in the moving image Vd in which the periodicity of the temporal change of the feature amount of the frame image is disturbed.

Further, according to the learning data selection device 201, among the plurality of sections, sections excluding a predetermined number of sections with relatively small calculated differences df can be specified as sections containing frame images in an abnormal state. .

As a result, it is possible to secure a predetermined number of learning data (required number of samples) necessary for learning the model M while removing images in an abnormal state from the learning data.

Further, according to the learning data selection device 201, the feature amount vector of the frame image included in the moving image Vd is calculated, and the frequency analysis is performed on the feature amount signal indicating the temporal change of the principal component of the calculated feature amount vector of the frame image. The period T of the moving image Vd can be calculated based on the result obtained by the above.

As a result, it is possible to set the period T of the moving image Vd to be the strongest frequency component among the feature amount signals that vary greatly along the period of the moving image.

Further, according to the learning data selection device 201, it is possible to specify, as the principal component, the component with the largest variation among the feature amount vectors of the frame images.

Thereby, it is possible to identify the feature amount component that varies greatly along the period T of the moving image Vd.

Further, according to the learning data selection device 201, frame images are extracted from the video Vd at regular time intervals, the extracted frame images are used as learning data, the encoder ec is learned, and the learned encoder ec is used to convert the video Vd. A feature amount vector of the included frame image can be calculated.

As a result, it is possible to obtain a feature amount vector in a feature space appropriate for representing the features of the frame images included in the moving image Vd. For example, by learning what kind of elements (components) should be used to accurately reproduce an image using an autoencoder, it is possible to obtain a feature vector in a feature space that is appropriate for expressing the features of a frame image. can.

Further, according to the learning data selection device 201, it is possible to output the determined learning data.

As a result, it is possible to output the learning data capable of generating the model M capable of accurately detecting the anomaly of the anomaly detection target.

For this reason, according to the learning data selection device 201, even if the normal image and the abnormal image are mixed without being sorted out, and even if the data (moving image Vd) has a small difference between the features of the two, the feature amount of the image can be determined. By detecting disturbances in the periodicity of changes, it is possible to accurately extract and learn only normal images. As a result, it is possible to generate a model M capable of accurately detecting anomalies in the anomaly detection target. In addition, when preparing the learning data, it is not necessary to manually sort out the images in the normal state and the images in the abnormal state, thereby reducing labor and man-hours.

Here, an example of selection of learning data by this learning data selection method will be described.

FIG. 15 is an explanatory diagram showing an example of selection of learning data. In FIG. 15, frame images 1501 to 1506 are chronologically arranged frame images included in a moving image Vd of an abnormality detection target (automatic production line). Here, there is a difference in how the abnormality detection target is captured due to changes in the shooting environment (illumination changes).

In this case, according to this learning data selection method, the frame images in the sections in which the appearance is different due to changes in illumination are removed from the learning data. On the other hand, if the same period is repeated after the illumination change, the frame images in that section are selected as learning data. As a result, the model M can be learned using the frame images before and after the illumination change as learning data. According to this model M, it is possible to prevent an erroneous determination of an abnormality even though the abnormality detection target is in a normal state when the illumination changes.

Note that instead of removing the frame images corresponding to the parts (sections) where the periodicity is disturbed from the training data, as in this learning data selection method, the parts where the periodicity is disturbed are directly determined as anomalies to be detected. It is also possible to consider a method to However, with this method, for example, as shown in FIG. 15, when the periodicity is disturbed due to a change in the imaging environment such as a change in lighting, there is a problem that an abnormality is erroneously determined.

The learning data selection method described in the present embodiment can be realized by executing a prepared program on a computer such as a personal computer or a workstation. This learning data selection program is recorded on a computer-readable recording medium such as a hard disk, flexible disk, CD-ROM, DVD, USB memory, etc., and is executed by being read from the recording medium by a computer. Also, the learning data selection program may be distributed via a network such as the Internet.

In addition, the learning data selection device 201 (information processing device 101) described in the present embodiment can be implemented by application-specific ICs such as standard cells and structured ASICs (Application Specific Integrated Circuits) and PLDs (Programmable Logic Devices) such as FPGAs. can also be realized.

Further, the following additional remarks are disclosed with respect to the above-described embodiment.

(Appendix 1) Calculating a feature amount representing a feature of a frame image included in a moving image of an anomaly detection target,
calculating the period of the moving image based on the time change of the calculated feature amount of the frame image;
calculating, for each section of a plurality of sections obtained by dividing the moving image according to the calculated period, calculating a difference regarding a feature amount of a frame image between each section and another section;
identifying a section including a frame image in an abnormal state among the plurality of sections based on the difference calculated for each section;
determining a frame image corresponding to another section different from the specified section among the plurality of sections as learning data to be used for learning a model for detecting anomaly of the anomaly detection target;
A learning data selection program characterized by causing a computer to execute processing.

(Appendix 2) The specifying process is
1. The learning data selection program according to appendix 1, wherein a section in which the calculated difference is equal to or greater than a threshold value is specified as a section containing a frame image in an abnormal state.

(Appendix 3) The identifying process is
3. The method according to appendix 1 or 2, wherein, among the plurality of sections, sections excluding a predetermined number of sections in which the calculated difference is relatively small are specified as sections containing frame images in an abnormal state. Learning data selection program.

(Appendix 4) The process of calculating the feature amount includes:
calculating a feature amount vector representing a feature of a frame image included in the moving image;
The process of calculating the cycle includes:
calculating the period based on a result obtained by performing a frequency analysis of a feature amount signal indicating a temporal change in the principal component of the calculated feature amount vector of the frame image;
The learning data selection program according to any one of Appendices 1 to 3, characterized by:

(Supplementary Note 5) The learning data selection program according to Supplementary Note 4, characterized in that it causes the computer to execute a process of specifying a component having the largest variation among the feature quantity vectors of the frame images as a principal component.

(Appendix 6) Extracting frame images from the moving image at regular time intervals,
causing the computer to execute a process of learning an encoder that converts a frame image into a feature vector using the extracted frame image as learning data;
The process of calculating the feature amount includes:
6. The learning data selection program according to any one of appendices 1 to 5, wherein the learned encoder is used to calculate a feature amount vector of a frame image included in the moving image.

(Appendix 7) The learning data selection program according to any one of Appendices 1 to 6, characterized by causing the computer to execute a process of outputting the determined learning data.

(Appendix 8) The learning data selection program according to any one of Appendices 1 to 7, wherein the moving image is a periodic moving image of an anomaly detection target.

(Appendix 9) The identifying process is
1. The program for selecting learning data according to appendix 1, wherein a predetermined number of sections in descending order of the calculated difference are specified as sections containing frame images in an abnormal state.

(Appendix 10) calculating a feature amount representing a feature of a frame image included in a moving image of an anomaly detection target;
calculating the period of the moving image based on the time change of the calculated feature amount of the frame image;
calculating, for each section of a plurality of sections obtained by dividing the moving image according to the calculated period, calculating a difference regarding a feature amount of a frame image between each section and another section;
identifying a section including a frame image in an abnormal state among the plurality of sections based on the difference calculated for each section;
determining a frame image corresponding to another section different from the specified section among the plurality of sections as learning data to be used for learning a model for detecting anomaly of the anomaly detection target;
A learning data selection method characterized in that processing is executed by a computer.

(Appendix 11) calculating a feature quantity representing a feature of a frame image included in a moving image of an anomaly detection target;
calculating the period of the moving image based on the time change of the calculated feature amount of the frame image;
calculating, for each section of a plurality of sections obtained by dividing the moving image according to the calculated period, calculating a difference regarding a feature amount of a frame image between each section and another section;
identifying a section including a frame image in an abnormal state among the plurality of sections based on the difference calculated for each section;
determining a frame image corresponding to another section different from the specified section among the plurality of sections as learning data to be used for learning a model for detecting anomaly of the anomaly detection target;
An information processing apparatus comprising a control unit.

101 information processing device 110, 410, Vd video 120, 130 feature signal 140 learning data 200 information processing system 201 learning data selection device 202 client device 210 network 220 frame image DB
230 Feature amount signal table 240 Learning data DB
300 Bus 301 CPU
302 memory 303 disk drive 304 disk 305 communication I/F
306 portable recording medium I/F
307 portable recording medium 400 camera 601 acquisition unit 602 first calculation unit 603 second calculation unit 604 identification unit 605 determination unit 606 output unit 611 encoder learning unit 612 vector calculation unit 613 principal component analysis unit 614 frequency analysis unit ec encoder

Claims (9)

Calculating the feature value representing the feature of the frame image included in the video of the anomaly detection target,
calculating the period of the moving image based on the time change of the calculated feature amount of the frame image;
calculating, for each section of a plurality of sections obtained by dividing the moving image according to the calculated period, calculating a difference regarding a feature amount of a frame image between each section and another section;
identifying a section including a frame image in an abnormal state among the plurality of sections based on the difference calculated for each section;
determining a frame image corresponding to another section different from the specified section among the plurality of sections as learning data to be used for learning a model for detecting anomaly of the anomaly detection target;
A learning data selection program characterized by causing a computer to execute processing. The process of specifying
2. The learning data selection program according to claim 1, wherein a section in which the calculated difference is equal to or greater than a threshold value is specified as a section containing a frame image in an abnormal state. The process of specifying
2. The learning data selection program according to claim 1, wherein a predetermined number of sections in descending order of the calculated differences are specified as sections containing frame images in an abnormal state. The process of calculating the feature amount includes:
calculating a feature amount vector representing a feature of a frame image included in the moving image;
The process of calculating the cycle includes:
calculating the period based on a result obtained by performing a frequency analysis of a feature amount signal indicating a temporal change in the principal component of the calculated feature amount vector of the frame image;
4. The learning data selection program according to any one of claims 1 to 3, characterized by: extracting frame images from the moving image at regular time intervals;
causing the computer to execute a process of learning an encoder that converts a frame image into a feature vector using the extracted frame image as learning data;
The process of calculating the feature amount includes:
5. The learning data selection program according to any one of claims 1 to 4, wherein the learned encoder is used to calculate a feature amount vector of a frame image included in the moving image. 6. The learning data selection program according to claim 1, causing the computer to execute a process of outputting the determined learning data. 7. The learning data selection program according to any one of claims 1 to 6, wherein the moving image is a periodic moving image of an anomaly detection target. Calculating the feature value representing the feature of the frame image included in the video of the anomaly detection target,
calculating the period of the moving image based on the time change of the calculated feature amount of the frame image;
calculating, for each section of a plurality of sections obtained by dividing the moving image according to the calculated period, calculating a difference regarding a feature amount of a frame image between each section and another section;
identifying a section including a frame image in an abnormal state among the plurality of sections based on the difference calculated for each section;
determining a frame image corresponding to another section different from the specified section among the plurality of sections as learning data to be used for learning a model for detecting anomaly of the anomaly detection target;
A learning data selection method characterized in that processing is executed by a computer. Calculating the feature value representing the feature of the frame image included in the video of the anomaly detection target,
calculating the period of the moving image based on the time change of the calculated feature amount of the frame image;
calculating, for each section of a plurality of sections obtained by dividing the moving image according to the calculated period, calculating a difference regarding a feature amount of a frame image between each section and another section;
identifying a section including a frame image in an abnormal state among the plurality of sections based on the difference calculated for each section;
determining a frame image corresponding to another section different from the specified section among the plurality of sections as learning data to be used for learning a model for detecting anomaly of the anomaly detection target;
An information processing apparatus comprising a control unit.

Images