JP7735801B2 - Anomaly detection device, method, and program - Google Patents

Description

The present invention relates to an anomaly detection device, method, and program for detecting anomalies in an object by visually inspecting an image of the object.

Traditionally, abnormalities in objects are detected by visual inspection of images of the object. Methods used for this type of abnormality detection include autoencoders and one-class classification.

An autoencoder extracts features from an input image, reconstructs an image that recreates the input image from the features, and detects anomalies based on the error between the input image and the reconstructed image (see, for example, Patent Document 1).

In one-class classification, features are extracted from the input image, and anomalies are detected based on the distance between the feature and the center point of the feature extracted from a normal input image during training (see, for example, Non-Patent Document 1).

Patent No. 6599294 specification

Ruff, L., Vandermeulen, R., Goernitz, N., Deecke, L., Siddiqui, S.A., Binder, A., Muller, E., Kloft, M., “Deep One-Class Classification”, Proceedings of the 35th International Conference on Machine Learning (2018)

Conventional anomaly detection methods such as one-class classification and autoencoders run the risk of missing anomalies, failing to detect them even when they exist in an object.

The present invention was made in consideration of the above-mentioned problems, and its purpose is to provide an anomaly detection device, method, and program that can prevent abnormalities from being overlooked during anomaly detection.

A first embodiment of the present invention is an anomaly detection device comprising: a feature extraction unit that extracts features from an input image; an image reconstruction unit that constructs a reconstructed image from the features; a feature anomaly score calculation unit that calculates a feature anomaly score indicating the distance between the feature and a predetermined center point; a reconstruction anomaly score calculation unit that calculates a reconstruction anomaly score indicating the error between the reconstructed image and the input image; and an anomaly determination unit that determines an anomaly using both the feature anomaly score and the reconstruction anomaly score.

In this embodiment, the feature anomaly score and the reconstructed anomaly score have complementary properties with respect to anomalies that may be overlooked. Since anomalies are determined using both the feature anomaly score and the reconstructed anomaly score, it is possible to prevent anomalies from being overlooked during anomaly detection.

A second embodiment of the present invention is a learning device that performs training on the anomaly detection device of the first embodiment, and includes a feature extraction update unit that updates the feature extraction unit using the feature anomaly score as a loss function, an image reconstruction update unit that updates the feature extraction unit and the image reconstruction unit using the reconstruction anomaly score as a loss function, and a gradient regularization unit that updates the image reconstruction unit so as to regularize the gradient of the image reconstruction unit.

In this embodiment, the anomaly detection device is trained to regularize the gradient of the image reconstruction unit. This promotes complementarity between the feature anomaly score and the reconstructed anomaly score in anomaly detection in the trained anomaly detection device, making it possible to more reliably prevent anomalies from being overlooked in anomaly detection.

A third embodiment of the present invention is an anomaly detection method comprising: a feature extraction step of extracting features from an input image; an image reconstruction step of constructing a reconstructed image from the features; a feature anomaly score calculation step of calculating a feature anomaly score indicating a distance between the feature and a predetermined center point; a reconstructed anomaly score calculation step of calculating a reconstructed anomaly score indicating an error between the reconstructed image and the input image; and an anomaly determination step of determining an anomaly using both the feature anomaly score and the reconstructed anomaly score.
This embodiment provides the same effects as the first embodiment.

A fourth embodiment of the present invention is a learning method for the anomaly detection device of the first embodiment, comprising: a feature extraction updating step of updating the feature extraction unit using the feature anomaly score as a loss function; an image reconstruction updating step of updating the feature extraction unit and the image reconstruction unit using the reconstruction anomaly score as a loss function; and a gradient regularization step of updating the image reconstruction unit so as to regularize the gradient of the image reconstruction unit.
This embodiment provides the same effects as the second embodiment.

A fifth embodiment of the present invention is an anomaly detection program that causes a computer to realize a feature extraction function that extracts features from an input image, an image reconstruction function that constructs a reconstructed image from the features, a feature anomaly score calculation function that calculates a feature anomaly score that indicates the distance between the feature and a predetermined center point, a reconstruction anomaly score calculation function that calculates a reconstruction anomaly score that indicates the error between the reconstructed image and the input image, and an anomaly determination function that determines an anomaly using both the feature anomaly score and the reconstruction anomaly score.
This embodiment provides the same effects as the first embodiment.

A sixth embodiment of the present invention is a learning program for training the anomaly detection device of the first embodiment, which causes a computer to realize a feature extraction update function for updating the feature extraction unit using the feature anomaly score as a loss function, an image reconstruction update function for updating the feature extraction unit and the image reconstruction unit using the reconstruction anomaly score as a loss function, and a gradient regularization function for updating the image reconstruction unit so as to regularize the gradient of the image reconstruction unit.
This embodiment provides the same effects as the second embodiment.

This invention makes it possible to prevent abnormalities from being overlooked during anomaly detection.

1 is a block diagram showing an anomaly detection device according to an embodiment of the present invention; FIG. 1 is a block diagram showing a learning device according to an embodiment of the present invention. FIG. 2 is a flow diagram illustrating an anomaly detection method according to an embodiment of the present invention. 1 is a flow diagram illustrating a training method according to an embodiment of the present invention. FIG. 2 is a schematic diagram showing an input image according to an embodiment of the present invention.

An outline of one embodiment of the present invention will be described below.
A wide variety of anomaly detection techniques have been used to detect anomalies in an object by visually inspecting an image of the object. The inventors of the present application were the first to discover that among these various anomaly detection techniques, autoencoders and one-class classification have complementary properties with regard to anomalies that may be overlooked, leading to the invention of the present embodiment. The anomaly detection of the present embodiment combines autoencoders and one-class classification and utilizes their complementary properties to prevent overlooking of anomalies. Furthermore, in addition to combining autoencoders and one-class classification, the anomaly detection of the present embodiment is trained to regularize the gradient of the image reconstruction portion of the autoencoder. This promotes the complementary properties during anomaly detection in a trained anomaly detection device, thereby enabling more reliable prevention of overlooking of anomalies.

An embodiment of the present invention will be described with reference to FIGS.
The abnormality detection device of this embodiment will be described with reference to FIG.

In the anomaly detection device, the feature extraction unit 12 extracts a feature vector Z as a feature from the input image X, as shown in the following equation (1): where φ represents a parameter of the feature extraction unit 12.
The feature extraction unit 12 is formed by a convolutional neural network, a multilayer perceptron, etc. As the feature extraction unit 12, an initial network that has been pre-trained using a predetermined training data set may be used.

The image reconstruction unit 14 reconstructs a reconstructed image X̂ that reproduces the input image X from the feature vector Z extracted from the input image X by the feature extraction unit 12, as shown in the following equation (2), where ψ represents a parameter of the image reconstruction unit 14.
The image reconstruction unit 14 is formed by a convolutional neural network, a multilayer perceptron, or the like.

The feature anomaly score calculation unit 16 calculates the feature anomaly score _L1 . The feature anomaly score _L1 indicates the distance in the feature vector space between the feature vector Z extracted from the input image X by the feature extraction unit 12 and the central vector C as a predetermined central point. As the feature anomaly score _L1 , an arbitrary norm of the feature vector Z and the central vector C is used, for example, the L2 norm, as shown in the following equation (3).
Furthermore, the predetermined central vector C may be (a) the average value of the feature vectors extracted from the input image when training the initial network, or (b) the exponential moving average value of the feature vectors extracted from the input image when training the anomaly detection device.

During training of the anomaly detection device, the feature anomaly score _L1 is used as a loss function, and the parameter φ of the feature extraction unit 12 is updated so that the feature vector Z extracted from a normal input image X approaches the central vector C. For this reason, when an abnormal input image X is input during anomaly detection by the anomaly detection device, the feature anomaly score _L1 is expected to increase.

The reconstructed anomaly score calculation unit 18 calculates the reconstructed anomaly score _L2 . The reconstructed anomaly score _L2 indicates the error in image space between the reconstructed image X^ reconstructed by the image reconstruction unit 14 from the feature vector Z and the original input image X. As the reconstructed anomaly score _L2 , an arbitrary error function dist between the reconstructed image X^ and the input image X is used, as shown in the following equation (4).
The error function dist may be an L2 norm or an index specific to an image, such as SSIM (structural similarity).

During training of the anomaly detection device, the reconstructed anomaly score _L2 is used as a loss function, and the parameters φ and ψ of the feature extraction unit 12 and the image reconstruction unit 14 are updated so that the reconstructed image X̂ constructed from a normal input image X approximates the original input image X. For this reason, when an abnormal input image X is input during anomaly detection by the anomaly detection device, the reconstructed anomaly score _L2 is expected to increase.

The anomaly score integrating unit 20 calculates an integrated anomaly score L. The integrated anomaly score L is obtained by integrating the feature amount anomaly score _L1 calculated by the feature amount anomaly score calculating unit 16 and the reconstructed anomaly score _L2 calculated by the reconstructed anomaly score calculating unit 18.

Here, the feature anomaly score _L1 and the reconstructed anomaly score _L2 have complementary properties with respect to anomalies that can be eliminated. That is, (i) if the feature anomaly score _L1 is small despite the input of an abnormal input image X, then in the feature vector space, the feature vector Z is close to the center vector C. Therefore, in the image space, the reconstructed image X^ reconstructed from the feature vector Z should be similar to the normal input image. Since the error between the original abnormal input image X and the reconstructed image X^ is large, the reconstructed anomaly score _L2 , which indicates the error between the reconstructed image X^ and the input image X, is large. On the other hand, (ii) if the reconstructed anomaly score L2 is small despite the input of an abnormal input image _X , then in the image space, the error between the reconstructed image X^ and the original abnormal input image X is small. Therefore, in the feature vector space, the feature vector Z reconstructed from the reconstructed image X^ that approximates the abnormal input image X should be far from the center vector C. Therefore, the feature anomaly score _L1 , which indicates the distance between the feature vector Z and the center vector C, is large.

The anomaly score integrating unit 20 integrates the feature amount anomaly score _L1 and the reconstructed anomaly score _L2 while maintaining their complementarity to calculate an integrated anomaly score L. The integrated anomaly score L may be, for example, a weighted sum of the feature amount anomaly score _L1 and the reconstructed anomaly score _L2 , as shown in the following equation (5):
Here, _λ1 and _λ2 represent the weights of the feature anomaly score _L1 and the reconstructed anomaly score _L2 , respectively. The weights _λ1 and _λ2 can be set arbitrarily, and (a) the reciprocal of the average value of the feature anomaly score _L1 and the reconstructed anomaly score L2 of the normal input image X used when training the anomaly detection device, or ( _b ) if a verification image is available, the reciprocal of the average value of the feature anomaly score _L1 and the reconstructed anomaly score _L2 of the normal verification image, etc., can be used.

The anomaly determination unit 22 determines an anomaly in the input image X using both the feature amount anomaly score _L1 and the reconstructed anomaly score _L2 , which are complementary to each other. In this embodiment, the anomaly determination unit 22 determines an anomaly in the input image X using the integrated anomaly score L calculated by the anomaly score integrating unit 20. For example, the anomaly determination unit 22 determines that the input image X has an anomaly when the integrated anomaly score L is equal to or greater than a predetermined threshold.

The learning device of this embodiment will be described with reference to FIG.
The learning device of this embodiment is used to train the anomaly detection device of this embodiment, and performs the training using a normal input image X.

In the learning device, the feature extraction update unit 32 updates the parameter φ of the feature extraction unit 12 using the feature anomaly score _L1 calculated by the feature anomaly score calculation unit 16 as a loss function. Here, the feature anomaly score _L1 indicates the distance between the feature vector Z and the central vector C in the feature vector space, and the parameter φ of the feature extraction unit 12 is updated so that the feature vector Z extracted from a normal input image X approaches the central vector C. For this reason, if an abnormal input image X is input during anomaly detection by the anomaly detection device, the feature anomaly score _L1 is expected to increase.

The image reconstruction update unit 34 updates the parameters φ and ψ of the feature extraction unit 12 and the image reconstruction unit 14 using the reconstructed anomaly score _L2 calculated by the reconstructed anomaly score calculation unit 18 as a loss function. Here, the reconstructed anomaly score _L2 indicates the error between the reconstructed image X^ and the original input image X in image space, and the parameters φ and ψ of the feature extraction unit 12 and the image reconstruction unit 14 are updated so that the reconstructed image X^ constructed from the normal input image X approximates the original input image X. For this reason, if an abnormal input image X is input when the anomaly detection device detects an anomaly, the reconstructed anomaly score _L2 is expected to increase.

The gradient regularizer 36 updates the parameter ψ of the image reconstructor 14 so as to regularize the gradient of the image reconstructor 14 .
As described above, the feature anomaly score _L1 and the reconstructed anomaly score _L2 are complementary with respect to anomalies that can be eliminated. For the complementarity between the feature anomaly score _L1 and the reconstructed anomaly score _L2 to be realized, the gradient of the image reconstructor 14 must be regularized to a certain degree, and the image reconstructor 14 must be smooth as a function to a certain degree. The more the gradient of the image reconstructor 14 is regularized and the smoother the function of the image reconstructor 14, the more the complementarity between the feature anomaly score _L1 and the reconstructed anomaly score _L2 is enhanced. In other words, if the reconstructed image X^ reconstructed from the feature vector Z changes independently or extremely significantly in image space with respect to changes in the feature vector Z in the feature vector space, the complementarity (i) and (ii) between the feature anomaly score _L1 and the reconstructed anomaly score _L2 described above will no longer be valid. On the other hand, the more closely the change in the image space of the reconstructed image X^ reconstructed from the feature vector Z corresponds to the change in the feature vector Z in the feature vector space, the more the complementarity (i) and (ii) between the feature anomaly score _L1 and the reconstructed anomaly score _L2 described above is promoted.

In this embodiment, the gradient regularizer 36 regularizes the gradient of the image reconstructor 14 by updating the parameter ψ of the image reconstructor 14 using, as a loss function, the gradient regularization term _L3 that promotes regularization of the gradient of the image reconstructor 14. As the gradient regularization term _L3 , the Frobenius norm of the Jacobian of the center vector C of the image reconstructor 14, or the like, is used, as shown in the following equation (6).
The Frobenius norm of the Jacobian for the central vector C of the image reconstruction unit 14 may be calculated directly using the above equation (6) when the size of the input image is relatively small, or may be calculated using a random projection approximation as shown in the following equation (7) when the size of the input image is relatively large and direct calculation is difficult.

The gradient regularizer 36 of one modified example of this embodiment uses spectral normalization to regularize the gradients of the image reconstructor 14. That is, as shown in the following equation (8), the parameter ω of each layer of the image reconstructor 14 is normalized by the maximum singular value σ(ω) calculated approximately, thereby regularizing the gradients of each layer of the image reconstructor 14.
Spectral normalization is disclosed in the following documents, etc.
Miyato, T., and Kataoka, T. and Koyama, M. and Yoshida, Y., “Spectral Normalization for Generative Adversarial Networks”, Proceedings of the 6th International Conference on Learning Representations (2018)

In addition to the gradient regularization unit 36, the model of the image reconstruction unit 14 itself may be simplified to regularize the gradient of the image reconstruction unit 14 and smooth the image reconstruction unit 14 as a function. The model of the image reconstruction unit 14 itself may be simplified by, for example, reducing the number of stacked layers that make up the model of the image reconstruction unit 14.

The abnormality detection method of this embodiment will be described with reference to FIG.
As shown in FIG. 3, the anomaly detection method includes the following steps.

Feature vector extraction step S12
In the feature vector extraction step S12, the feature amount extraction unit 12 extracts a feature vector Z from the input image X.

Image reconstruction step S14
In the image reconstruction step S14, the image reconstruction unit 14 reconstructs a reconstructed image X^ that reproduces the input image X from the feature vector Z extracted from the input image X in the feature extraction step S16.

Feature abnormality score calculation step S16
In the feature anomaly score calculation step S16, the feature anomaly score calculation unit 16 calculates a feature anomaly score _L1 . The feature anomaly score _L1 indicates the distance between the feature vector Z extracted from the input image X in the feature vector extraction step S12 and a predetermined center vector C in the feature vector space.

Reconstruction abnormality score calculation step S18
In the reconstruction anomaly score calculation step S18, the reconstruction anomaly score _L2 is calculated by the reconstruction anomaly score calculation unit 18. The reconstruction anomaly score _L2 indicates the error in the image space between the reconstructed image Ẑ reconstructed in the image reconstruction step S14 and the original input image X.

Anomaly score integration step S20
In the anomaly score integration step S20, the anomaly score integration unit 20 integrates the feature amount anomaly score _L1 calculated in the feature amount anomaly score calculation step S16 and the reconstructed anomaly score _L2 calculated in the reconstructed anomaly score calculation step S18 while maintaining complementarity, to calculate an integrated anomaly score L.

Abnormality determination step S22
In the anomaly determination step S22, the anomaly determination unit 22 uses both the feature amount anomaly score _L1 and the reconstructed anomaly score _L2 , which are complementary to each other, to determine an anomaly in the input image X. In this embodiment, the integrated anomaly score L calculated in the anomaly score integration step S20 is used to determine an anomaly in the input image X.

The learning method of this embodiment will be described with reference to FIG.
The learning method of this embodiment is a learning method for the anomaly detection device of this embodiment, and learning is performed using normal input images.
As shown in FIG. 4, the learning method of this embodiment includes the following steps.

Feature vector extraction step S32, image reconstruction step S34, feature amount abnormality score calculation step S36, reconstruction abnormality score calculation step S38
These steps are the same as the feature vector extraction step S12, image reconstruction step S14, feature amount anomaly score calculation step S16, and reconstruction anomaly score calculation step S18 of the anomaly detection method.

Gradient regularization term calculation step S40
In the gradient regularization term calculation step S40, the gradient regularization unit 36 calculates the gradient regularization term _L3 that promotes regularization of the gradient in the image reconstruction unit 14.

Feature extraction and update step S42
In the feature extraction and update step S42, the feature extraction and update unit 32 updates the parameter φ of the feature extraction unit 12 using the feature anomaly score _L1 calculated in the feature anomaly score calculation step S36 as a loss function.

Image reconstruction update step S44
In the image reconstruction update step S44, the image reconstruction update unit 34 updates the parameters φ and ψ of the feature extraction unit 12 and the image reconstruction unit 14 using the reconstruction abnormality score _L2 calculated in the reconstruction abnormality score calculation step S38 as a loss function.

Gradient regularization step S46
In the gradient regularization step S46, the gradient regularizer 36 updates the image reconstruction unit 14 so as to regularize the gradient of the image reconstruction unit 14. In this embodiment, the parameter ψ of the image reconstruction unit 14 is updated using the gradient regularization term _L3 calculated in the gradient regularization term calculation step S40 as a loss function.

Parameter optimization step S48
In the parameter optimization step S48, the above steps are repeated until the parameters φ and ψ of the feature extraction unit 12 and the image reconstruction unit 14 are optimized.

In one variation of this embodiment, spectral normalization is used to regularize the gradients of the image reconstruction unit 14. That is, instead of the gradient regularization term calculation step, a maximum singular value calculation step is executed to approximately calculate the maximum singular value σ(ω) of the parameters of each layer of the image reconstruction unit 14. Then, in the gradient regularization step, the parameter ω of each layer of the image reconstruction unit 14 is normalized by the maximum singular value σ(ω) calculated in the maximum singular value calculation step, thereby regularizing the gradients of each layer of the image reconstruction unit 14.

The complementarity in anomaly detection in this embodiment will be described with reference to FIG.
FIG. 5( a ) shows a normal input image, FIG. 5( b ) shows an input image with a local anomaly, and FIG. 5( c ) shows an input image with a global anomaly. For an input image with a local anomaly such as that shown in FIG. 5( b ), it is difficult to detect an anomaly using a feature anomaly score based on a feature vector obtained by compressing the input image, but it is easy to detect an anomaly using a reconstructed anomaly score. On the other hand, for an input image with a global anomaly such as that shown in FIG. 5( c ), it is difficult to detect an anomaly using the reconstructed anomaly score because the image is locally normal, but it is easy to detect an anomaly using the feature anomaly score. In this embodiment, anomaly detection is performed using both the feature anomaly score and the reconstructed anomaly score, so that anomalies can be detected without missing any of the input images with local anomalies and those with global anomalies.

The anomaly detection device and method, and learning device and method of this embodiment have the following advantages:

In the anomaly detection device and method of this embodiment, the feature anomaly score _L1 , which indicates the distance in feature vector space between a feature vector Z extracted from an input image X and a predetermined center point C, and the reconstructed anomaly score _L2 , which indicates the error in image space between a reconstructed image X^ reconstructed from the feature vector Z and the original input image X, have complementary properties with respect to anomalies that may be overlooked. Since anomalies are determined using an integrated anomaly score L, which is obtained by integrating the feature anomaly score _L1 and the reconstructed anomaly score _L2 while maintaining their complementarity, it is possible to prevent anomalies from being overlooked during anomaly detection.

In the learning device and method of this embodiment, the anomaly detection device is trained to regularize the gradient of the image reconstruction unit 14. This makes it possible to promote complementarity between the feature anomaly score _L1 and the reconstructed anomaly score _L2 in anomaly detection in the trained anomaly detection device, and more reliably prevent anomalies from being overlooked in anomaly detection.

In the above embodiments, we have described an anomaly detection device and method, and a learning device and method. However, the scope of the present invention also includes programs that cause a computer to realize the functions of the device, or programs that cause a computer to execute the steps of the method. Furthermore, the scope of the present invention also includes an anomaly detection device that has been trained using the learning device, method, and program.

12: Feature extraction unit 14: Image reconstruction unit 16: Feature abnormality score calculation unit 18: Reconstruction abnormality score calculation unit 20: Abnormality score integration unit 22: Abnormality determination unit 32: Feature extraction update unit 34: Image reconstruction update unit 36: Gradient regularization unit

Claims (9)

a feature extraction unit that extracts features from an input image;
an image reconstruction unit that constructs a reconstructed image from the feature amount;
a feature anomaly score calculation unit that calculates a feature anomaly score indicating a distance between the feature and a predetermined center point;
a reconstruction anomaly score calculation unit that calculates a reconstruction anomaly score indicating an error between the reconstructed image and the input image;
an abnormality determination unit that determines an abnormality using both the feature amount abnormality score and the reconstructed abnormality score;
Equipped with
an anomaly score integrating unit that integrates the feature amount anomaly score and the reconstructed anomaly score while maintaining complementarity to calculate an integrated anomaly score;
the abnormality determination unit determines an abnormality using the integrated abnormality score.
Anomaly detection device. A learning device that performs learning on an anomaly detection device,
The abnormality detection device
a feature extraction unit that extracts features from an input image;
an image reconstruction unit that constructs a reconstructed image from the feature amount;
a feature anomaly score calculation unit that calculates a feature anomaly score indicating a distance between the feature and a predetermined center point;
a reconstruction anomaly score calculation unit that calculates a reconstruction anomaly score indicating an error between the reconstructed image and the input image;
an abnormality determination unit that determines an abnormality using both the feature amount abnormality score and the reconstructed abnormality score;
Equipped with
The learning device
a feature extraction update unit that updates the feature extraction unit using the feature anomaly score as a loss function;
an image reconstruction update unit that updates the feature amount extraction unit and the image reconstruction unit using the reconstruction anomaly score as a loss function;
a gradient regularizer that updates the image reconstruction unit to regularize the gradient of the image reconstruction unit;
A learning device comprising: the gradient regularization unit calculates a Frobenius norm of a Jacobian at the center point of the image reconstruction unit as a gradient regularization term, and updates the image reconstruction unit using the gradient regularization term as a loss function;
The learning device according to claim 2 . the gradient regularization unit regularizes the gradient of each layer of the image reconstruction unit by normalizing the parameter of each layer of the image reconstruction unit using the maximum singular value of the parameter;
The learning device according to claim 2 . 5. The anomaly detection device according to claim 1, wherein learning is performed by the learning device according to any one of claims 2 to 4. a feature extraction step of extracting features from an input image;
an image reconstruction step of constructing a reconstructed image from the feature amount;
a feature amount anomaly score calculation step of calculating a feature amount anomaly score indicating a distance between the feature amount and a predetermined center point;
a reconstruction anomaly score calculation step of calculating a reconstruction anomaly score indicating an error between the reconstructed image and the input image;
an anomaly determination step of determining an anomaly using both the feature amount anomaly score and the reconstructed anomaly score;
Equipped with
an anomaly score integration step of integrating the feature anomaly score and the reconstructed anomaly score while maintaining complementarity to calculate an integrated anomaly score,
In the abnormality determination step, an abnormality is determined using the integrated abnormality score.
Anomaly detection methods. A learning method for an anomaly detection device, comprising:
The abnormality detection device
a feature extraction unit that extracts features from an input image;
an image reconstruction unit that constructs a reconstructed image from the feature amount;
a feature anomaly score calculation unit that calculates a feature anomaly score indicating a distance between the feature and a predetermined center point;
a reconstruction anomaly score calculation unit that calculates a reconstruction anomaly score indicating an error between the reconstructed image and the input image;
an abnormality determination unit that determines an abnormality using both the feature amount abnormality score and the reconstructed abnormality score;
Equipped with
The learning method includes:
a feature extraction update step of updating the feature extraction unit using the feature anomaly score as a loss function;
an image reconstruction updating step of updating the feature amount extraction unit and the image reconstruction unit using the reconstruction anomaly score as a loss function;
a gradient regularization step of updating the image reconstructor to regularize the gradient of the image reconstructor;
A learning method that includes the following. On the computer,
A feature extraction function that extracts features from input images,
an image reconstruction function for constructing a reconstructed image from the feature amount;
a feature anomaly score calculation function for calculating a feature anomaly score indicating a distance between the feature and a predetermined center point;
a reconstruction anomaly score calculation function for calculating a reconstruction anomaly score indicating an error between the reconstructed image and the input image;
an anomaly determination function that determines an anomaly using both the feature amount anomaly score and the reconstructed anomaly score;
To achieve this,
further realizing an anomaly score integration function that calculates an integrated anomaly score by integrating the feature anomaly score and the reconstructed anomaly score while maintaining complementarity;
The abnormality determination function determines an abnormality using the integrated abnormality score.
Anomaly detection programs. A learning program for learning an anomaly detection device,
The abnormality detection device
a feature extraction unit that extracts features from an input image;
an image reconstruction unit that constructs a reconstructed image from the feature amount;
a feature anomaly score calculation unit that calculates a feature anomaly score indicating a distance between the feature and a predetermined center point;
a reconstruction anomaly score calculation unit that calculates a reconstruction anomaly score indicating an error between the reconstructed image and the input image;
an abnormality determination unit that determines an abnormality using both the feature amount abnormality score and the reconstructed abnormality score;
Equipped with
The learning program
On the computer,
a feature extraction update function that updates the feature extraction unit using the feature anomaly score as a loss function;
an image reconstruction update function that updates the feature extraction unit and the image reconstruction unit using the reconstruction anomaly score as a loss function;
a gradient regularization function that updates the image reconstruction unit to regularize the gradient of the image reconstruction unit;
A learning program that makes this possible.