Ultra-high-definition image anomaly detection method and system based on VAE and mutual K-nearest neighbor

Abstract

The invention belongs to the technical field of ultra-high-definition image processing, and provides an ultra-high-definition image anomaly detection method and system based on VAE and K nearest neighbors, and the main scheme comprises the steps: firstly, constructing and training a feature extraction model based on a variational auto-encoder according to an ultra-high-definition image set; secondly, performing feature extraction on the ultra-high-definition image set by using an encoder part of a feature extraction model based on a variational auto-encoder to obtain a feature matrix; then, according to the feature matrix, the K nearest neighbor particles of each ultra-high-definition image sample are calculated, the K nearest neighbor distance of each sample is calculated, and K is a sample neighbor number; and finally, based on the mutual K nearest neighbor distance, calculating an abnormal score of each ultra-high-definition image sample, and based on the abnormal score, judging whether the ultra-high-definition image sample is abnormal or not. According to the invention, the precision and reliability of ultra-high-definition image anomaly detection can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of ultra-high-definition image processing, and in particular to an ultra-high-definition image anomaly detection method and system based on VAE and K nearest neighbors. Background Art

[0002] As one of the important tasks of anomaly detection, ultra-high-definition image anomaly detection has been widely used in practical applications such as ultra-high-definition images, medical images, industrial images, and hyperspectral images. In medicine, ultra-high-definition image anomaly detection can assist doctors in correctly analyzing patients' conditions, and in industry, it can help technicians to promptly detect defects such as scratches and dents on industrial products, and in hyperspectral imaging technology, it can find pixels or areas that are significantly different from the surrounding environment or known samples in terms of spectral characteristics. As one of the specific tasks, ultra-high-definition image anomaly detection refers to finding abnormal ultra-high-definition images in a specified ultra-high-definition image set, and indicating the degree of abnormality of each abnormal image (generally described by anomaly scores).

[0003] The traditional UHD image anomaly detection method includes several key steps: first, UHD image acquisition and preprocessing, which uses sensors to acquire high-resolution images and perform denoising and color space conversion; second, feature extraction, which uses various filters and feature extraction algorithms to capture local and global features; then, anomaly analysis and detection, which uses anomaly detection methods to identify abnormal images; and finally, outputting the serial number and anomaly score corresponding to the abnormal image. Traditional anomaly detection methods work better when processing ordinary images, but they cannot fully utilize the advantages of the algorithm when processing large-scale data with complex features such as UHD image sets.

[0004] Traditional image anomaly detection methods are mainly divided into: statistics-based, proximity-based, clustering-based and deep learning-based methods. Statistical methods mainly use statistical models of data, such as Gaussian distribution, to calculate the degree of abnormality of each image, and identify data points that do not conform to the model by setting an abnormality score threshold; proximity-based methods use the relationship between a sample and its neighboring samples to identify samples with large distance or density differences from surrounding points as abnormal; clustering-based methods first divide the entire sample set into multiple clusters based on the similarity of feature information, and perform density estimation on each cluster, and finally use the distance from the sample to the cluster center to determine the abnormality; deep learning-based methods learn the complex feature representation of data by building a deep neural network, and perform anomaly detection based on the features. Its network structure mainly includes autoencoders (AE) and generative adversarial networks (GAN).

[0005] Therefore, the traditional image anomaly detection solution has the following problems when applied to ultra-high-definition image anomaly detection:

[0006] 1. Although statistical methods are simple and easy to understand and suitable for small samples and a small number of features, they are more difficult to process high-dimensional data and limit the ability to detect unknown anomalies;

[0007] 2. The neighbor-based method is prone to outlier masking, which reduces the detection performance of the model;

[0008] 3. When clustering-based methods are faced with high-dimensional data or a small amount of data, the complexity or sparsity of the samples will cause the distribution differences of the samples to be small, thus affecting the effect of the detection model;

[0009] 4. Deep learning-based methods usually require a large amount of labeled data to support their complex training, which is too costly. Summary of the invention

[0010] The purpose of the present invention is to provide an ultra-high-definition image anomaly detection method and system based on VAE and K-nearest neighbors, which can maintain efficient and accurate detection performance when processing complex high-dimensional ultra-high-definition image data, aiming to solve the problems of complex features and large data volume of high-dimensional ultra-high-definition image data, and can improve the accuracy and reliability of ultra-high-definition image anomaly detection.

[0011] The present invention solves the technical problem and adopts the following technical solution:

[0012] On the one hand, the present invention provides an ultra-high-definition image anomaly detection method based on VAE and mutually K-nearest neighbors, comprising the following steps:

[0013] Based on the ultra-high-definition image set, a feature extraction model based on a variational autoencoder is constructed and trained;

[0014] The encoder part of the feature extraction model based on variational autoencoder is used to extract features from the ultra-high-definition image set to obtain a feature matrix;

[0015] According to the feature matrix, the K nearest neighbor particles of each ultra-high-definition image sample are calculated, and the K nearest neighbor distance of each sample is calculated, where K is the number of sample neighbors;

[0016] Based on the K nearest neighbor distances, the anomaly score of each ultra-high-definition image sample is calculated, and whether the ultra-high-definition image sample is abnormal is determined based on the anomaly score.

[0017] As a further optimization, the feature extraction model based on the variational autoencoder is constructed and trained according to the ultra-high-definition image set, including the following steps:

[0018] Perform minimum-maximum normalization on each ultra-high-definition image sample ai in the ultra-high-definition image set A to obtain the normalized ultra-high-definition image set

[0019] Construct a feature extraction model based on variational autoencoder, VAE, using ultra-high-definition image sets The VAE-based feature extraction model is trained to obtain a trained VAE-based feature extraction model.

[0020] As a further optimization, the objective function of the variational autoencoder-based feature extraction model is defined as:

[0021] max ELBO

[0022] Among them, the objective function indicates that the model training goal is to maximize the variational lower bound ELBO, Ultra-high-definition image data is the input of the model, h is The potential representation of is learned by the encoder part of the model; the first term in the variational lower bound ELBO Represents the reconstruction term, measuring the input and the reconstruction error between the output z decoded from the latent representation h, the second term is the KL divergence, which measures the distribution of the encoder's partial output The similarity between the prior distribution p(h) and the prior distribution p(h) is replaced by the standard normal distribution N(0, 1).

[0023] As a further optimization, the encoder part of the feature extraction model based on the variational autoencoder is used to extract features from the ultra-high-definition image set to obtain a feature matrix, including the following steps:

[0024] According to the trained feature extraction model based on variational autoencoder, the encoder part of the model is used as the feature extractor f;

[0025] Ultra-high-definition image collection Each UHD image sample The feature vector d obtained by the feature extractor f i , and form a feature matrix D, the number of rows of the feature matrix D corresponds to the number of image samples n, and the number of columns corresponds to the dimension m of the feature vector.

[0026] As a further optimization, the K nearest neighbor particles of each ultra-high-definition image sample are calculated according to the feature matrix, and the K nearest neighbor distance of each sample is calculated, including the following steps:

[0027] According to the number of samples in the ultra-high-definition image set, the number of sample neighbors K is determined;

[0028] Calculate the Euclidean distance between each ultra-high-definition image sample and all other samples in the ultra-high-definition image set except the current ultra-high-definition image sample to form a distance matrix Dis of size n×n;

[0029] For each ultra-high-definition image sample calculate The K nearest neighbors of

[0030] For each ultra-high-definition image sample calculate The K nearest neighbor distances

[0031] As a further optimization, the The K nearest neighbors of Defined as:

[0032]

[0033] in, Defined as a sample The K nearest neighbors of each other represent the samples The set of samples that are closest to each other in the first K distance. For ultra-high-definition image sets, and The samples are and The K nearest neighbor set consisting of the first K samples closest to it in the distance matrix Dis;

[0034] for Any sample in if The K nearest neighbor set Also included are samples Then the sample Recognized as are each other's K nearest neighbors, All the K nearest neighbors of each other constitute the sample The K nearest neighbors of

[0035] As a further optimization, the The K nearest neighbor distances Defined as:

[0036]

[0037] in, Represents an ultra-high-definition image sample The distance between each other is K nearest neighbors. for The K nearest neighbors of The central vector formed by the mean of all samples in each feature dimension, For sample The K-th nearest neighbor of For sample The K nearest neighbor set consists of the first K samples closest to it in the distance matrix Dis, and ||·|| represents the operation of calculating the Euclidean distance.

[0038] As a further optimization, the method of calculating the anomaly score of each ultra-high-definition image sample based on the K nearest neighbor distances, and judging whether the ultra-high-definition image sample is abnormal based on the anomaly score, includes the following steps:

[0039] Calculate the mean of the K nearest neighbor distances of all ultra-high-definition image samples As the anomaly score threshold;

[0040] For each ultra-high-definition image sample If their distances to each other are K nearest neighbors If the value is greater than the threshold, the ultra-high-definition image sample is judged to be abnormal.

[0041] On the other hand, the present invention also provides an ultra-high-definition image anomaly detection system based on VAE and mutually K-nearest neighbors, which is applied to the ultra-high-definition image anomaly detection method based on VAE and mutually K-nearest neighbors, comprising:

[0042] A model training unit, used for constructing and training a feature extraction model based on a variational autoencoder according to an ultra-high-definition image set;

[0043] A feature extraction unit, used to extract features from the ultra-high-definition image set using an encoder part of a feature extraction model based on a variational autoencoder to obtain a feature matrix;

[0044] A calculation unit, used to calculate the K nearest neighbor particles of each ultra-high-definition image sample according to the feature matrix, and calculate the K nearest neighbor distance of each sample, where K is the number of sample neighbors;

[0045] The anomaly judgment unit is used to calculate the anomaly score of each ultra-high-definition image sample based on the K nearest neighbor distances, and judge whether the ultra-high-definition image sample is abnormal based on the anomaly score.

[0046] The beneficial effects of the present invention are as follows: through the above-mentioned ultra-high-definition image anomaly detection method and system based on VAE and mutual K-nearest neighbors, a variational autoencoder (VAE) is used to extract features of ultra-high-definition images, high-dimensional image data is projected into a low-dimensional space, and then mutual K-nearest neighbor particles are constructed through mutual K-nearest neighbors, and the statistical distance is used as a similarity measure between different particles and the anomaly score is calculated. Finally, according to the anomaly score of the sample, it is determined whether the image is abnormal, thereby solving the problems of complex features and large data volume of high-dimensional image data, and effectively improving the accuracy and reliability of ultra-high-definition image anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flowchart of the ultra-high-definition image anomaly detection method based on VAE and K nearest neighbors in Example 1 of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0049] Example 1

[0050] This embodiment provides an ultra-high-definition image anomaly detection method based on VAE and K nearest neighbors. Figure 1 , wherein the method comprises the following steps:

[0051] S1. Based on the ultra-high-definition image set, a feature extraction model based on variational autoencoder is constructed and trained;

[0052] S2, using the encoder part of the feature extraction model based on the variational autoencoder to extract features from the ultra-high-definition image set to obtain a feature matrix;

[0053] S3. Calculate the K nearest neighbor particles of each ultra-high-definition image sample according to the feature matrix, and calculate the K nearest neighbor distance of each sample, where K is the number of sample neighbors;

[0054] S4. Calculate the anomaly score of each ultra-high-definition image sample based on the K nearest neighbor distances, and determine whether the ultra-high-definition image sample is abnormal based on the anomaly score.

[0055] In this embodiment, a variational autoencoder can be used to extract features from ultra-high-definition images, which effectively improves the accuracy of anomaly detection in ultra-high-definition images. At the same time, by combining the advantages of deep learning and traditional machine learning, it can process complex high-dimensional ultra-high-definition image data well, and the anomaly detection model can maintain a high detection performance. In addition, this embodiment belongs to an unsupervised learning method and is not affected by parameter settings, which can greatly reduce the workload of operators.

[0056] In step S1 of the above method, constructing and training a feature extraction model based on a variational autoencoder according to an ultra-high-definition image set may include the following steps:

[0057] S101, each ultra-high-definition image sample a in the ultra-high-definition image set Ai Perform minimum-maximum normalization to obtain a normalized ultra-high-definition image set

[0058] S102, build a feature extraction model based on variational autoencoder, the variational autoencoder is identified as VAE, and use the ultra-high-definition image set The VAE-based feature extraction model is trained to obtain a trained VAE-based feature extraction model.

[0059] Here, using VAE to extract features from ultra-high-definition images can learn the latent distribution of features of the ultra-high-definition image set, expand the encoding area of ​​the Encoder, and reduce the distortion points of the latent vector space; and the trained VAE can use the Encoder part to generate a lower-dimensional latent representation h than the original data, and the latent representation h has most of the useful feature information corresponding to the original data. Therefore, using VAE for feature extraction can achieve dimensionality reduction processing of high-dimensional image data, reduce the computational complexity of subsequent anomaly detection and retain the main feature information.

[0060] It should be pointed out that during the training process, the VAE Encoder calculates the mean μ and variance δ of the potential representation h distribution of the input image set, which is composed of CNN. The part used to calculate the mean will add Gaussian noise to the output mean μ during training, so that the subsequent decoder Decoder will have a certain robustness to noise, while the part used to calculate the variance is to appropriately adjust the noise according to the training effect of the Decoder during the training process, thereby improving the generation ability and noise resistance of the Decoder; In addition, during the training process, VAE learns the potential distribution of the data by maximizing the variational lower bound ELBO, so that the model has a certain robustness to the noise and partial information missing in the input data, so the features extracted by VAE can effectively handle the incompleteness and noise of the input data. Therefore, in step S102 of the above method, the objective function of the feature extraction model based on the variational autoencoder can be defined as:

[0061] max ELBO

[0062] Among them, the objective function indicates that the model training goal is to maximize the variational lower bound ELBO, Ultra-high-definition image data is the input of the model, h is The potential representation of is learned by the encoder part of the model; the first term in the variational lower bound ELBO Represents the reconstruction term, measuring the input and the reconstruction error between the output z decoded from the latent representation h, the second term is the KL divergence, which measures the distribution of the encoder's partial output The similarity between the prior distribution p(h) and the prior distribution p(h) is replaced by the standard normal distribution N(0, 1).

[0063] It should be pointed out that since VAE training can improve the model's ability to abstractly represent ultra-high-definition image data and better capture complex features, and the feature matrix extracted using the autoencoder can effectively reduce the data dimension, which helps to improve the efficiency and performance of the subsequent anomaly detection algorithm, in step S2 of the above method, the encoder part of the feature extraction model based on the variational autoencoder is used to extract features from the ultra-high-definition image set to obtain the feature matrix, which may include the following steps:

[0064] S201, according to the trained feature extraction model based on variational autoencoder, the encoder part of the model is used as a feature extractor f;

[0065] S202, collect ultra-high-definition images Each UHD image sample The feature vector d obtained by the feature extractor f i , and form a feature matrix D, the number of rows of the feature matrix D corresponds to the number of image samples n, and the number of columns corresponds to the dimension m of the feature vector.

[0066] It should be pointed out that by calculating the Euclidean distance between each ultra-high-definition image sample and other samples, the similarities and differences between samples can be captured more accurately; and the mutual K nearest neighbor distance takes into account the relationship between the sample and the surrounding neighboring samples, which can mine more abnormal information of the sample from a local perspective. Therefore, in step S3 of the above method, the calculation of the mutual K nearest neighbor particles of each ultra-high-definition image sample according to the feature matrix and the calculation of the mutual K nearest neighbor distance of each sample can include the following steps:

[0067] S301, determining the number of sample neighbors K according to the number of samples in the ultra-high-definition image set;

[0068] S302, calculating the Euclidean distance between each ultra-high-definition image sample and all other samples in the ultra-high-definition image set except the current ultra-high-definition image sample, to form a distance matrix Dis of size n×n;

[0069] S303: for each ultra-high definition image sample calculate The K nearest neighbors of

[0070] S304: for each ultra-high definition image sample calculate The K nearest neighbor distances

[0071] Specifically, the The K nearest neighbors of It can be defined as:

[0072]

[0073] in, Defined as a sample The K nearest neighbors of each other represent the samples The set of samples that are closest to each other in the first K distance. For ultra-high-definition image sets, and The samples are and The K nearest neighbor set consisting of the first K samples closest to it in the distance matrix Dis;

[0074] for Any sample in if The K nearest neighbor set Also included are samples Then the sample Recognized as are each other's K nearest neighbors, All the K nearest neighbors of each other constitute the sample The K nearest neighbors of

[0075] At the same time, the The K nearest neighbor distances It can be defined as:

[0076]

[0077] in, Represents an ultra-high-definition image sample The distance between each other is K nearest neighbors. for The K nearest neighbors of The central vector formed by the mean of all samples in each feature dimension, For sample The K-th nearest neighbor of For sample The K nearest neighbor set consists of the first K samples closest to it in the distance matrix Dis, and ||·|| represents the operation of calculating the Euclidean distance.

[0078] It should be noted that in step S4 of the above method, the calculation of the anomaly score of each ultra-high-definition image sample based on the K nearest neighbor distances, and the determination of whether the ultra-high-definition image sample is abnormal based on the anomaly score may include the following steps:

[0079] S401, calculate the mean of the K nearest neighbor distances of all ultra-high-definition image samples As the anomaly score threshold;

[0080] S402: for each ultra-high definition image sample If their distances to each other are K nearest neighbors If the value is greater than the threshold, the ultra-high-definition image sample is judged to be abnormal.

[0081] Here, using the mean of all K-nearest neighbor distances as the threshold fully considers the collective relationship between all samples and their closest neighbors, rather than relying solely on a single distance measurement; moreover, the structure of the detection model is not complicated, and it can effectively improve the computational efficiency of anomaly detection while ensuring detection performance; at the same time, the detection model needs to determine the parameter K according to the size of the sample, which can effectively guarantee the detection effect of the model when processing different sizes of ultra-high-definition image sets.

[0082] Example 2

[0083] Based on Example 1, this embodiment provides an ultra-high-definition image anomaly detection system based on VAE and mutually K-nearest neighbors, including:

[0084] A model training unit, used for constructing and training a feature extraction model based on a variational autoencoder according to an ultra-high-definition image set;

[0085] A feature extraction unit, used to extract features from the ultra-high-definition image set using an encoder part of a feature extraction model based on a variational autoencoder to obtain a feature matrix;

[0086] A calculation unit, used to calculate the K nearest neighbor particles of each ultra-high-definition image sample according to the feature matrix, and calculate the K nearest neighbor distance of each sample, where K is the number of sample neighbors;

[0087] The anomaly judgment unit is used to calculate the anomaly score of each ultra-high-definition image sample based on the K nearest neighbor distances, and judge whether the ultra-high-definition image sample is abnormal based on the anomaly score.

[0088] According to the description of Example 1, the application scenario and implementation principle of this embodiment are consistent with those of Example 1, so they will not be described in detail.

[0089] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An ultra-high-definition image anomaly detection method based on VAE and K nearest neighbors, characterized in that: The steps include: Based on the ultra-high-definition image set, a feature extraction model based on a variational autoencoder is constructed and trained; The encoder part of the feature extraction model based on variational autoencoder is used to extract features from the ultra-high-definition image set to obtain a feature matrix; According to the feature matrix, the K nearest neighbor particles of each ultra-high-definition image sample are calculated, and the K nearest neighbor distance of each sample is calculated, where K is the number of sample neighbors; Based on the K nearest neighbor distances, the anomaly score of each ultra-high-definition image sample is calculated, and whether the ultra-high-definition image sample is abnormal is determined based on the anomaly score.

2. The method for detecting anomalies in ultra-high-definition images based on VAE and K nearest neighbors according to claim 1, characterized in that: The method of constructing and training a feature extraction model based on a variational autoencoder according to an ultra-high-definition image set comprises the following steps: Each ultra-high-definition image sample a in the ultra-high-definition image set A i Perform minimum-maximum normalization to obtain a normalized ultra-high-definition image set Construct a feature extraction model based on variational autoencoder, VAE, using ultra-high-definition image sets The VAE-based feature extraction model is trained to obtain a trained VAE-based feature extraction model.

3. The method for detecting anomalies in ultra-high-definition images based on VAE and K nearest neighbors according to claim 2, characterized in that: The objective function of the feature extraction model based on variational autoencoder is defined as: max ELBO Among them, the objective function indicates that the model training goal is to maximize the variational lower bound ELBO, Ultra-high-definition image data is the input of the model, h is The potential representation of is learned by the encoder part of the model; the first term in the variational lower bound ELBO Represents the reconstruction term, measuring the input and the reconstruction error between the output z decoded from the latent representation h, the second term is the KL divergence, which measures the distribution of the encoder's partial output The similarity between the prior distribution p(h) and the prior distribution p(h) is replaced by the standard normal distribution N(0, 1).

4. The method for detecting anomalies in ultra-high-definition images based on VAE and K nearest neighbors according to claim 1, characterized in that: The method of extracting features from the ultra-high-definition image set using the encoder part of the feature extraction model based on the variational autoencoder to obtain a feature matrix includes the following steps: According to the trained feature extraction model based on variational autoencoder, the encoder part of the model is used as the feature extractor f; Ultra-high-definition image collection Each UHD image sample The feature vector d obtained by the feature extractor f i , and form a feature matrix D, the number of rows of the feature matrix D corresponds to the number of image samples n, and the number of columns corresponds to the dimension m of the feature vector.

5. The method for detecting anomalies in ultra-high-definition images based on VAE and K nearest neighbors according to claim 1, characterized in that: The step of calculating the K nearest neighbor particles of each ultra-high-definition image sample according to the feature matrix and calculating the K nearest neighbor distance of each sample comprises the following steps: According to the number of samples in the ultra-high-definition image set, the number of sample neighbors K is determined; Calculate the Euclidean distance between each ultra-high-definition image sample and all other samples in the ultra-high-definition image set except the current ultra-high-definition image sample to form a distance matrix Dis of size n×n; For each ultra-high-definition image sample calculate The K nearest neighbors of For each ultra-high-definition image sample calculate The K nearest neighbor distances 6. The method for detecting anomalies in ultra-high-definition images based on VAE and K nearest neighbors according to claim 5, characterized in that: Said The K nearest neighbors of Defined as: in, Defined as a sample The K nearest neighbors of each other represent the samples The set of samples that are closest to each other in the first K distance. For ultra-high-definition image sets, and The samples are and The K nearest neighbor set consisting of the first K samples closest to it in the distance matrix Dis; for Any sample in if The K nearest neighbor set Also included are samples Then the sample Recognized as are each other's K nearest neighbors, All the K nearest neighbors of each other constitute the sample The K nearest neighbors of 7. The method for detecting anomalies in ultra-high-definition images based on VAE and K nearest neighbors according to claim 5, characterized in that: Said The K nearest neighbor distances Defined as: in, Represents an ultra-high-definition image sample The distance between them is K nearest neighbors. for The K nearest neighbors of The central vector formed by the mean of all samples in each feature dimension, For sample The K-th nearest neighbor of For sample The K nearest neighbor set consists of the first K samples closest to it in the distance matrix Dis, and ||·|| represents the operation of calculating the Euclidean distance.

8. The method for detecting anomalies in ultra-high-definition images based on VAE and K nearest neighbors according to claim 1, characterized in that: The method of calculating the anomaly score of each ultra-high-definition image sample based on the K nearest neighbor distances, and judging whether the ultra-high-definition image sample is abnormal based on the anomaly score, includes the following steps: Calculate the mean of the K nearest neighbor distances of all ultra-high-definition image samples As the anomaly score threshold; For each ultra-high-definition image sample If their distances to each other are K nearest neighbors If the value is greater than the threshold, the ultra-high-definition image sample is judged to be abnormal.

9. The ultra-high-definition image anomaly detection system based on VAE and mutually K-nearest neighbors applied to the ultra-high-definition image anomaly detection method based on VAE and mutually K-nearest neighbors according to any one of claims 1 to 8, characterized in that: include: A model training unit, used for constructing and training a feature extraction model based on a variational autoencoder according to an ultra-high-definition image set; A feature extraction unit, used to extract features from the ultra-high-definition image set using an encoder part of a feature extraction model based on a variational autoencoder to obtain a feature matrix; A calculation unit, used to calculate the K nearest neighbor particles of each ultra-high-definition image sample according to the feature matrix, and calculate the K nearest neighbor distance of each sample, where K is the number of sample neighbors; The anomaly judgment unit is used to calculate the anomaly score of each ultra-high-definition image sample based on the K nearest neighbor distances, and judge whether the ultra-high-definition image sample is abnormal based on the anomaly score.

Images