KR20220022329A - Anomaly Detection Apparatus Using Loss Function Based on Latent Feature Dispersion and Method Thereof - Google Patents

Abstract

Embodiments of the present invention provide a device and method for detecting anomaly. An encoder defines a loss function so that a feature vector of latent space is distributed, and the feature vectors in the latent space are replaced with a normal image for all input data. Anomaly detection performance is improved through an image restoration model having a loss function applicable to various data sets by limiting a range of an encoder result.

Description

Anomaly Detection Apparatus Using Loss Function Based on Latent Feature Dispersion and Method Thereof}

The technical field to which the present invention pertains relates to an anomaly detection apparatus and method.

The content described in this section merely provides background information for the present embodiment and does not constitute the prior art.

Autoencoder, which is one of unsupervised learning that derives common features of a data set by learning similar data, is an encoder that abstracts and reduces the features of an image, and a decoder expands the features again and reconstructs them into an image. Since the autoencoder goes through the reduction process based on the features, the optimal feature expression can be found and the features of the original image are maintained.

A general autoencoder learns the distribution of normal data, and a model trained only on normal data does not properly restore abnormal samples. In high-dimensional space, anomaly samples are partially different from normal samples, but have an overall similar structure. Since the distance between the normal data and the abnormal data in the latent space is close, the restored image of the abnormal sample is not significantly different from the original image.

Korean Patent Publication No. 10-2020-0056183 (2020.05.22.) Korean Patent Publication No. 10-2027389 (2019.09.25.)

In the embodiments of the present invention, the encoder defines a loss function so that the feature vector in the latent space is distributed, the feature vector in the latent space is replaced with a normal image for all input data, and the range of the encoder result is limited to apply to various data sets. Its main purpose is to design an applicable loss function.

Other objects not specified in the present invention may be additionally considered within the scope that can be easily inferred from the following detailed description and effects thereof.

According to an aspect of this embodiment, in the method for detecting anomaly by a computing device, generating a restored image from an original image through an image restoration model in which an encoder and a decoder are combined, and converting the restored image into a normal image or an abnormal image It provides an anomaly detection method comprising the step of dividing and outputting anomaly results.

The loss function designed in the image reconstruction model is such that the encoder is designed such that the feature vectors of the latent space are distributed, the decoder reconstructs the feature vectors of the latent space into the reconstructed image, and the reconstructed image learns to be reconstructed into the normal image. can be

The loss function designed in the image restoration model may be learned by maximizing the anomaly score for the abnormal image while keeping the abnormality score for the normal image small.

The image restoration model may induce the generation of the normal image corresponding to the normal class to maximize the difference between the original image and the restored image with respect to the abnormal image.

The image reconstruction model applies a regularization coefficient to the loss function based on the assumption that the decoder reconstructs the image only for the trained latent features, and maximizes the variance between the latent vectors of the normal class to manage the entire latent space. It is possible to spread the vectors for the normal class in the fold space.

The decoder may generate an image having the characteristics of the normal class for all input images.

The step of outputting the abnormal result may include (i) a first situation in which the abnormal image has a relatively simpler structure than the normal image, or (ii) a structure in which the abnormal image is similar to the overall structure of the normal image within a reference range. In the second situation formed by , the abnormal image may be distinguished and output.

According to another aspect of this embodiment, in an anomaly detection device comprising one or more processors and a memory for storing one or more programs executed by the one or more processors, the processor restores an image in which an encoder and a decoder are combined from an original image Generates a reconstructed image through the model, and the processor provides an abnormality detection apparatus, characterized in that the reconstructed image is divided into a normal image or an abnormal image and an abnormality result is output.

As described above, according to the embodiments of the present invention, the encoder defines a loss function so that feature vectors in the latent space are distributed, so that all vectors in the limited space for all input data are replaced with normal images, and the encoder result It has the effect of improving the anomaly detection performance through an image reconstruction model with a loss function applicable to various data sets by limiting the range of .

Even if the effects are not explicitly mentioned herein, the effects described in the following specification expected by the technical features of the present invention and their potential effects are treated as if they were described in the specification of the present invention.

1 is a block diagram illustrating an anomaly detection apparatus according to an embodiment of the present invention.
2 is a diagram illustrating an image reconstruction model in which an encoder and a decoder are combined in an anomaly detection apparatus according to an embodiment of the present invention.
3 to 5 are diagrams illustrating distribution of feature vectors in a latent space of an anomaly detection apparatus according to an embodiment of the present invention.
6 is a flowchart illustrating an anomaly detection method according to another embodiment of the present invention.
7 to 10 are diagrams illustrating simulation results according to embodiments of the present invention.

Hereinafter, in the description of the present invention, if it is determined that the subject matter of the present invention may be unnecessarily obscure as it is obvious to those skilled in the art with respect to related known functions, the detailed description thereof will be omitted, and some embodiments of the present invention will be described. It will be described in detail with reference to exemplary drawings.

When using a general autoencoder, it is not easy to distinguish anomalies from (i) a situation in which abnormal data is formed in a relatively simple structure than normal data and (ii) in a situation in which abnormal data is formed in a structure similar to the overall structure of normal data. there is.

In order to solve this problem, the anomaly detection apparatus according to the present embodiment adds a regularization coefficient to the loss function based on the assumption that the decoder reconstructs the image well only for the trained latent features to increase the variance between the latent vectors of the normal class. By maximizing, the vector for the normal class is distributed in the entire manifold space, and the decoder generates a constant image with the characteristic of the normal class for all input images. As a result, when the input image is an abnormal image, the restoration error is significantly different.

1 is a block diagram illustrating an anomaly detection apparatus according to an embodiment of the present invention.

The anomaly detection device 110 includes at least one processor 120 , a computer-readable storage medium 130 , and a communication bus 170 .

The processor 120 may control to operate as the anomaly detection device 110 . For example, the processor 120 may execute one or more programs stored in the computer-readable storage medium 130 . The one or more programs may include one or more computer-executable instructions, which when executed by the processor 120 may be configured to cause the anomaly detection device 110 to perform operations according to the exemplary embodiment. can

Computer-readable storage medium 130 is configured to store computer-executable instructions or program code, program data, and/or other suitable form of information. The program 140 stored in the computer-readable storage medium 130 includes a set of instructions executable by the processor 120 . In one embodiment, computer-readable storage medium 130 includes memory (volatile memory, such as random access memory, non-volatile memory, or a suitable combination thereof), one or more magnetic disk storage devices, optical disk storage devices, It may be flash memory devices, other types of storage media that can be accessed by the anomaly detection apparatus 110 and store desired information, or a suitable combination thereof.

The communication bus 170 interconnects various other components of the anomaly detection device 110 including the processor 120 and the computer-readable storage medium 140 .

The anomaly detection device 110 may also include one or more input/output interfaces 150 and one or more communication interfaces 160 that provide interfaces for one or more input/output devices. The input/output interface 150 and the communication interface 160 are connected to the communication bus 170 . The input/output device (not shown) may be connected to other components of the anomaly detection device 110 through the input/output interface 150 .

The anomaly detection apparatus 110 maximizes the anomaly score for the abnormal sample while maintaining a small abnormality score for the normal sample through the newly defined loss function. will do The network of the restoration model of the anomaly detection device 110 uses an approach to always generate an image of a normal class, and with respect to an abnormal image, the difference between the original image and the restored image can be maximized. Using this method, the anomaly detection device 110 detects whether (i) abnormal data is formed in a relatively simple structure than normal data, and (ii) abnormal data is formed in a structure similar to the overall structure of normal data. can be easily distinguished.

2 is a diagram illustrating an image reconstruction model in which an encoder and a decoder are combined in an anomaly detection apparatus according to an embodiment of the present invention.

The encoder takes an image or video as an input, and uses visual characteristic information as an output. The decoder restores the visual feature information as an image or video. The encoder and decoder have multiple layers connected to the network and include a hidden layer. A layer may include parameters, and the parameters of the layer include a set of learnable filters. The parameters include weights and/or biases between nodes.

The loss function is defined as the sum of the restoration loss and the dispersion loss and is expressed as in Equation 1.

The restoration loss can be calculated as L2-Norm. x is the input image, f is the encoder, f(x) is the feature vector (latent vector) that is the output of the encoder, and g is the decoder.

The variance loss controls the variance of the latent vector. λ is the normalization coefficient, d is the dimension of the latent space, and n is the number of normal samples for a single class. The latent space is a space in which feature vectors exist, the properties of the image are mapped, and the size is set to sufficiently contain the information of the target image.

The purpose of the image reconstruction model is to minimize the loss, and the parameters of the model are trained to minimize the loss.

The image reconstruction model always induces to generate a normal image corresponding to the normal class, thereby maximizing the difference between the original image and the restored image for an abnormal image.

Based on the assumption that the decoder reconstructs images only for the latent features it has been trained on, the image reconstruction model applies regularization coefficients to the loss function to maximize the variance between the latent vectors of the stationary class, thereby maximizing the variance between the latent vectors in the latent space. Distribute the vector for the class. The manifold space means a space in which the distance or similarity between two points is Euclidean (linear distance) in the near field but Euclidean in the far distance. The decoder generates images with normal class characteristics for all input images.

3 to 5 are diagrams illustrating dispersion of feature vectors in a latent space of an anomaly detection apparatus according to an embodiment of the present invention.

3 shows a latent vector disposed in the latent space by the conventional autoencoder, FIG. 4 shows the latent vector (λ=0.01) disposed in the latent space by the anomaly detection apparatus according to the present embodiment, and FIG. A latent vector (λ=1) arranged in the latent space by the anomaly detection apparatus according to the present embodiment is shown.

As the normalization coefficient (λ) increases, the degree of dispersion in the manifold space increases. However, if the normalization coefficient is too large, since the same image is restored for all input images, it is necessary to set an optimized normalization coefficient. The normalization coefficient can be set within the range of 0.001 to 0.1, and is preferably set to 0.01.

6 is a flowchart illustrating an anomaly detection method according to another embodiment of the present invention.

In step S210, the processor generates a reconstructed image from the original image through an image reconstruction model in which an encoder and a decoder are combined.

In step S220, the processor divides the restored image into a normal image or an abnormal image and outputs an abnormal result.

In the learning process, the loss function designed in the image reconstruction model is designed so that the encoder is designed such that the feature vectors of the latent space are distributed, the decoder reconstructs the feature vectors of the latent space into the reconstructed image, and the reconstructed image is learned to be reconstructed into the normal image. do.

In the training process, the loss function designed in the image restoration model is learned by maximizing the anomaly score for the abnormal image while keeping the anomaly score small for the normal image.

The step of outputting the abnormal result may include (i) a first situation in which the abnormal image has a relatively simple structure than the normal image, or (ii) a second situation in which the abnormal image is formed in a structure similar to the overall structure of the normal image within a reference range. can be output by classifying abnormal images in .

7 to 10 are diagrams illustrating simulation results according to embodiments of the present invention.

7 is a transistor image of the MVTec data set. FIG. 7 (a) is an input abnormal image, FIGS. 7 (b) to (e) are reconstructed images, and FIG. 7 (f) is a normal image. am. An error between the input image and the reconstructed image may be calculated using Mean Squared Error (MSE).

Fig. 8 relates to a bottle image of the MVTec data set. Fig. 8 (a) is an input abnormal image, Figs. 8 (b) to (e) are reconstructed images, and Fig. 8 (f) is a normal image. am. An error between the input image and the reconstructed image may be calculated using Mean Squared Error (MSE).

Fig. 9 relates to the MNIST data set. Fig. 9 (a) is an original image, Fig. 9 (b) is a reconstructed image by an existing autoencoder, and Fig. 9 (c) is an image according to the present embodiment. A reconstructed image (λ=0.01) by the anomaly detection apparatus, and FIG. 9(d) is a reconstructed image (λ=1) by the anomaly detection apparatus according to the present embodiment. It can be seen that the reconstructed image is formed into a class of number 8 due to lossy dispersion.

Figure 10 relates to a fashion MNIST data set, (a) of Figure 10 is an original image, Figure 10 (b) is a restored image by an existing autoencoder, (c) of Figure 10 is in this embodiment It is a reconstructed image (λ=0.01) by the anomaly detection device according to the It can be seen that the reconstructed image is formed into a bag-shaped class due to the lossy dispersion.

The anomaly detection apparatus may be implemented in a logic circuit by hardware, firmware, software, or a combination thereof, or may be implemented using a general-purpose or special-purpose computer. The device may be implemented using a hardwired device, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or the like. In addition, the device may be implemented as a system on a chip (SoC) including one or more processors and controllers.

Anomaly detection apparatus may be mounted in the form of software, hardware, or a combination thereof on a computing device or server provided with hardware elements. A computing device or server is all or part of a communication device such as a communication modem for performing communication with various devices or wired/wireless communication networks, a memory for storing data for executing a program, and a microprocessor for executing operations and commands by executing the program It can mean a variety of devices, including

Although it is described that each process is sequentially executed in FIG. 6, this is merely an exemplary description, and those skilled in the art change the order described in FIG. 6 within the range that does not depart from the essential characteristics of the embodiment of the present invention Alternatively, various modifications and variations may be applied by executing one or more processes in parallel or adding other processes.

The operations according to the present embodiments may be implemented in the form of program instructions that can be performed through various computer means and recorded in a computer-readable medium. Computer-readable medium represents any medium that participates in providing instructions to a processor for execution. Computer-readable media may include program instructions, data files, data structures, or a combination thereof. For example, there may be a magnetic medium, an optical recording medium, a memory, and the like. A computer program may be distributed over a networked computer system so that computer readable code is stored and executed in a distributed manner. Functional programs, codes, and code segments for implementing the present embodiment may be easily inferred by programmers in the art to which this embodiment belongs.

The present embodiments are for explaining the technical idea of the present embodiment, and the scope of the technical idea of the present embodiment is not limited by these embodiments. The protection scope of this embodiment should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be interpreted as being included in the scope of the present embodiment.

Claims (14)

In the abnormal detection method by a computing device,
generating a reconstructed image from the original image through an image reconstruction model in which an encoder and a decoder are combined; and
and outputting an abnormality result by classifying the restored image into a normal image or an abnormal image. According to claim 1,
The loss function designed in the image reconstruction model is such that the encoder is designed such that the feature vectors of the latent space are distributed, the decoder reconstructs the feature vectors of the latent space into the reconstructed image, and the reconstructed image learns to be reconstructed into the normal image. Anomaly detection method, characterized in that. 3. The method of claim 2,
Anomaly detection method, characterized in that the loss function designed in the image restoration model is learned by maximizing the abnormality score for the abnormal image while keeping the abnormality score for the normal image small. 3. The method of claim 2,
The image restoration model always induces to generate the normal image corresponding to the normal class, and maximizes the difference between the original image and the restored image with respect to the abnormal image. 3. The method of claim 2,
The image reconstruction model applies a regularization coefficient to the loss function based on the assumption that the decoder reconstructs the image only for the trained latent features, and maximizes the variance between the latent vectors of the normal class to manage the entire latent space. Anomaly detection method, characterized in that the vector for the normal class is distributed in the fold space. 3. The method of claim 2,
The decoder generates an image having the characteristics of the normal class for all input images. According to claim 1,
The step of outputting the abnormal result is,
(i) the abnormal image in a first situation in which the abnormal image is formed to have a relatively simple structure than the normal image, or (ii) in a second situation in which the abnormal image is formed to have a similar structure to the overall structure of the normal image within a reference range; Anomaly detection method, characterized in that the output is distinguished. An anomaly detection device comprising one or more processors and a memory for storing one or more programs executed by the one or more processors,
The processor generates a restored image from the original image through an image restoration model in which an encoder and a decoder are combined,
The processor divides the restored image into a normal image or an abnormal image and outputs an abnormality result. 9. The method of claim 8,
The loss function designed in the image reconstruction model is such that the encoder is designed such that the feature vectors of the latent space are distributed, the decoder reconstructs the feature vectors of the latent space into the reconstructed image, and the reconstructed image learns to be reconstructed into the normal image. Anomaly detection device, characterized in that. 10. The method of claim 9,
Anomaly detection apparatus, characterized in that the loss function designed in the image restoration model is learned by maximizing the abnormality score for the abnormal image while keeping the abnormality score for the normal image small. 10. The method of claim 9,
The image restoration model always induces to generate the normal image corresponding to the normal class, and maximizes the difference between the original image and the restored image with respect to the abnormal image. 10. The method of claim 9,
The image reconstruction model applies a regularization coefficient to the loss function based on the assumption that the decoder reconstructs the image only for the trained latent features, and maximizes the variance between the latent vectors of the normal class to manage the entire latent space. Anomaly detection apparatus, characterized in that the vector for the normal class is distributed in the fold space. 10. The method of claim 9,
The decoder generates an image having the characteristics of the normal class with respect to all input images. 9. The method of claim 8,
The processor is configured to: (i) in a first situation in which the abnormal image is formed to have a relatively simple structure than the normal image, or (ii) in a second situation in which the abnormal image is formed to have a similar structure to the overall structure of the normal image within a reference range Anomaly detection apparatus, characterized in that the divided image is output.

Images