Remote sensing image anomaly detection model, training method and detection method

Abstract

The invention provides a remote sensing image anomaly detection model, a training method and a detection method, and the remote sensing image anomaly detection model comprises a dynamic diffusion anomaly synthesis module which is used for generating a high-simulation anomaly image according to an input sample; the feature extraction selection module comprises a feature extraction network, a feature selector and a feature selector optimization sub-module, and the adaptive fusion reconstruction module is used for fusing and reconstructing a sample feature tensor output by the trained feature selector into a normal feature tensor; and the detection module is used for calculating the abnormal score of each pixel and optimizing the reconstruction precision of the adaptive fusion reconstruction module so as to improve the abnormal detection and positioning capability. Meanwhile, the invention discloses a training method and a using method of the system. The system can effectively improve the capability of detecting and positioning abnormity. Through the mechanism, focusing on an abnormal sensitive area in a complex feature space can be realized, and fine identification and positioning optimization of an abnormal mode can be realized.

Description

Technical Field

[0001] The present invention belongs to the field of image anomaly detection, and particularly relates to a remote sensing image anomaly detection model, a training method and a detection method based on diffusion enhancement and feature selection. Background Art

[0002] Anomaly detection of remote sensing images refers to extracting discrete points from multi-spectral images to represent that pixels have abnormal values. With the rapid development of artificial intelligence, in the field of remote sensing image anomaly detection, the application of deep learning-based methods in remote sensing image anomaly detection is also increasing. The problem with existing deep learning-based anomaly detection models is that the excellent detection effect of the model requires a large amount of training data, but in anomaly detection, the number of abnormal labeled samples is often insufficient. Therefore, it is impossible to effectively suppress background information and enhance abnormal information, and a single anomaly detection model is difficult to simultaneously extract local and global discriminant features of abnormal and background information in remote sensing images, which has a certain impact on the accuracy of anomaly detection results.

[0003] For example, CN110245633A discloses a remote sensing image anomaly detection method with rotational invariance, which includes using a preselected super-resolution algorithm to perform image enhancement processing on the original remote sensing image to obtain a high-resolution remote sensing image; using a pre-trained RIFD-CNN to process the high-resolution algorithm to obtain a target remote sensing image with rotational invariance; wherein, the RIFD-CNN is generated by adding a rotational invariance regularizer and a Fisher discriminant regularizer to the network architecture of R-CNN; using a preselected target detection algorithm to perform anomaly detection on the target remote sensing image to determine abnormal targets within the target remote sensing image.

[0004] However, the above technology still has deficiencies in anomaly detection ability and positioning. Summary of the Invention

[0005] The object of the present invention is to propose a remote sensing image anomaly detection method and device based on diffusion enhancement and feature selection, which can effectively improve the detection and positioning ability of anomalies.

[0006] The present invention is realized through the following technical solutions: A remote sensing image anomaly detection model, including, A dynamic diffusion anomaly synthesis module, which is used to generate a highly realistic anomaly image according to the input sample ; ; A feature extraction and selection module, which includes a feature extraction network, a feature selector and a feature selector optimization sub-module. The feature extraction network is used to extract the input sample and the highly realistic anomaly image Sample features of n different dimensions and abnormal features , where i = 1, 2, …, n, and the feature selector is used to process each extracted sample feature and abnormal features respectively to obtain a sample feature tensor and an abnormal feature tensor . The feature selector optimization sub-module trains and optimizes the feature selector by minimizing the reconstruction loss function and the contrast loss function; An adaptive fusion and reconstruction module, which is used to fuse and reconstruct the n sample feature tensors output by the trained feature selector into n normal feature tensors ; A detection module, which is used to calculate the abnormal scores of each pixel of the sample feature tensor and the reconstructed normal feature tensor respectively, and learn and optimize the reconstruction accuracy of the adaptive fusion and reconstruction module according to the scalar loss function to improve the abnormal detection and localization capabilities.

[0007] As one of the implementation manners, the dynamic diffusion abnormal synthesis module includes A forward noise addition sub-module, whose conditional probability distribution of the noise data at time is , where , is a parameter between 0 and 1, , is a fixed variance schedule, is a latent variable, I is the Identity Matrix, whose dimension matches the vector dimension, represents the state vector at the initial moment, represents the random state vector at time t, is the Gaussian distribution function; A reverse denoising sub-module, which introduces a controllable perturbation mechanism in the reverse sampling process to direct the generation process to shift towards the low probability density region in the latent space. The probability distribution of the reverse process is . The noise in the diffusion process is modeled by using the neural network , so as to obtain , represents the random state vector at time t - 1; A loss function design sub-module, which aims to minimize the variational upper bound of the negative log-likelihood, and uses the objective function as Train a diffusion model to learn the distribution of normal images, where is the mathematical expectation, is to calculate the L2 norm, ϵ is the true noise, is the noise predicted by the model, and the conditional probability distribution , is the abnormal image obtained at time , is the introduced additional perturbation, , is the abnormal image sampled after the introduced additional perturbation, where the scalar controls the abnormal intensity.

[0008] As one of the implementation manners, the adaptive fusion reconstruction module includes a feature enhancement sub-module, which aligns and splices the sample feature tensors output by the feature selector in dimensions to obtain a feature map , divides it into m sub-channel groups, and performs multi-scale feature enhancement on the features of the channel groups to obtain enhanced features ; a feature fusion sub-module, which adaptively fuses the enhanced features to obtain a representative fused feature after fusion; a feature reconstruction sub-module, which uses a feature reconstruction network to obtain a reconstructed normal feature tensor from the fused feature , where the feature reconstruction network is designed with a structure symmetric but opposite to that of the feature extraction network.

[0009] As one of the implementation manners, the detection module includes an abnormal calculation sub-module, which obtains the abnormal score of each pixel point in the reconstructed normal image under the i-th feature dimension according to the formula , and then uses a scalar loss function to learn and optimize the reconstruction accuracy of the adaptive fusion reconstruction module to obtain the optimal abnormal score represented by ; where , , respectively represent the length and width of the normal feature tensor , is the transpose function; is the sample feature tensor the feature vector corresponding to the pixel point with coordinates (h, w) in the image under the i-th feature dimension, is the reconstructed normal feature tensor ​The feature vector corresponding to the pixel at coordinates (h, w) in the image under the i-th feature dimension, is the L1 norm used to measure the length or magnitude of a vector; Anomaly score map generation sub-module, which according to the formula obtains the final anomaly score map , where, represents the upsampling operation.

[0010] The training method of the remote sensing image anomaly detection model includes the following steps, Step S1, input samples are dynamically adjusted by the dynamic diffusion anomaly synthesis module to generate highly realistic anomaly images ; Step S2, the feature extraction network extracts the sample features of n different dimensions of the input samples and the highly realistic anomaly images respectively, where i = 1, 2,..., n. After each extracted feature, a feature selector is connected to obtain the sample feature tensor and the anomaly feature tensor . The feature selector is optimized in feature extraction and transmission by minimizing the reconstruction loss and the contrast loss; and the anomaly feature tensor . The feature selector is optimized in feature extraction and transmission by minimizing the reconstruction loss and the contrast loss; Step S3, the n sample feature tensors extracted by the trained feature selector are input into the adaptive fusion reconstruction module to obtain the fusion feature , and n normal feature tensors are reconstructed by using it. The detection module calculates the anomaly scores of each pixel of the sample feature tensor and the reconstructed normal feature tensor respectively, and generates a scalar loss function, and then uses the scalar loss function to optimize the reconstruction accuracy of the adaptive fusion reconstruction module.

[0011] The detection method for remote sensing image anomalies uses the above-mentioned remote sensing image anomaly detection model. The detection method includes the following steps, 1) The feature extraction network extracts the input features of n different dimensions of the input image , where i = 1, 2,..., n, 2) The feature selector performs feature selection to obtain the input feature tensor , 3) The n sample features extracted are input into the adaptive fusion reconstruction module to obtain the fusion feature, and n reconstructed input feature tensors are reconstructed by using it. For the reconstructed input feature tensor and the input image feature tensor Perform similarity comparison calculation to obtain the input feature tensor for each pixel in; 4) Generate and output the anomaly score map.

[0012] The present invention has the following beneficial effects: The present invention first generates a highly realistic remote sensing anomaly image through the dynamic diffusion anomaly synthesis module as the supervision signal for training the feature selector. Secondly, multi-dimensional features of the input samples are extracted, and a feature selector is integrated after each feature. Then, an anomaly contrast signal is introduced into the input samples, and by minimizing the contrast loss and feature similarity loss, the feature selector is guided to focus on the anomaly-sensitive areas and suppress the anomaly features during the feature extraction and transmission processes. Finally, the multi-scale feature enhancement and adaptive feature fusion mechanisms of the adaptive fusion reconstruction module are used to reconstruct n feature tensors, and then the anomaly detection result is obtained by calculating the similarity between the input sample feature tensor and the reconstructed feature tensor. The present invention can effectively improve the detection and localization capabilities for anomalies. Through this mechanism, it is possible to focus on the anomaly-sensitive areas within the complex feature space, achieving refined recognition and localization optimization of the anomaly patterns. Specific embodiments

[0013] An anomaly detection model for remote sensing images based on diffusion enhancement and feature selection according to the present invention includes a dynamic diffusion anomaly synthesis module, a feature extraction and selection module, an adaptive fusion reconstruction module, and a detection module, The dynamic diffusion anomaly synthesis module, which generates a highly realistic anomaly image according to the input sample ; that is, under the guidance of the input sample , a highly realistic anomaly image sample is generated through the dynamic diffusion anomaly synthesis module as the supervision signal for training the feature selector; specifically, this module uses the forward diffusion process of the denoising diffusion probability model to gradually add noise to the original data distribution ; by deeply modeling the data distribution characteristics of normal samples through the diffusion model, while maintaining the adjustability of the anomaly intensity, anomaly samples highly consistent with the real data distribution are generated. Specifically, the dynamic diffusion anomaly synthesis module uses the forward diffusion process of the denoising diffusion probability model to gradually add noise

[0014] to the original data distribution , by deeply modeling the data distribution characteristics of normal samples through the diffusion model, while maintaining the adjustability of the anomaly intensity, anomaly samples highly consistent with the real data distribution are generated, specifically including the following, The forward noise addition sub-module, at time ​When it comes to noise data The conditional probability distribution is where is a parameter between 0 and 1, is a fixed variance schedule, is a latent variable, I is the identity matrix, whose dimension matches that of the vector, represents the state vector at the initial time, represents the random state vector at time t, is the Gaussian distribution function, and the diffusion process is defined as a Markov chain, whose joint probability distribution is According to the summation rule of Gaussian random variables, at time the conditional probability distribution of is ; The reverse denoising and adding noise sub-module, by introducing a controllable perturbation mechanism during the reverse sampling process, directs the generation process to shift towards the low probability density region in the latent space. The reverse process is described as another Markov chain, where the mean and variance are determined by the parameter that is Using the neural network to model the noise in the diffusion process, thus obtaining ; The loss function design adding noise sub-module, in order to generate realistic abnormal images, during the training phase, the goal is to minimize the variational upper bound of the negative log-likelihood, using the objective function to train the diffusion model to learn the distribution of normal images. During the reverse diffusion process, the characterized by is the normal image obtained at time ; represents the random state vector at time t - 1; since the abnormal image is located in the low density region close to the normal image, an additional perturbation is introduced to sample the abnormal image, obtaining where is the introduced additional perturbation, and the scalar controls the abnormal intensity, is the abnormal image obtained at time , setting , thus the conditional probability distribution can be written as ; ​​The feature extraction and selection module includes a feature extraction network, a feature selector and a feature selector optimization submodule. The feature extraction network is used to extract input samples respectively. and high-fidelity abnormal images Sample features of n different dimensions and abnormal characteristics , where i=1,2,…,n, and the feature selector is used to select each extracted feature to obtain a sample feature tensor and the abnormal feature tensor ,The feature selector optimization submodule utilizes the minimization of reconstruction loss function to train the ,ability of the feature selector output to suppress abnormal information during the ,reasoning process, and adopts the contrast loss function to train the feature ,selector to push away abnormal information from the normal space, forcing the ,feature selector to focus on learning deeper representations of normal features.

[0015] More specifically, the feature extraction and selection module is mainly composed of a feature extraction network (such as a pre-trained WideResNet-50 network) and a feature selector. The feature extraction network extracts multi-scale and comprehensive representations of input samples; the feature selector's main function is to focus on abnormally sensitive areas and suppress abnormal features; the feature selector consists of a convolutional layer, an activation function, and a two-dimensional instance normalization unit. The convolutional layer deeply represents multi-scale features; the activation function introduces nonlinear transformations, giving the model the ability to express complex nonlinear relationships; the two-dimensional instance normalization unit stabilizes the network training process and accelerates convergence by standardizing the input, while also helping to alleviate overfitting and improve the generalization performance of the model.

[0016] As a specific embodiment, the feature extraction network is based on the formula Extract sample features of n different dimensions of the input sample ,in, represents the input sample image, represents the i-th feature extraction network, which is a pre-trained WideResNet-50 network, c i 、h i 、w i They represent the number of channels, height, and width of the tensor extracted by the i-th feature extraction network; In this embodiment, the input sample image is the Wuhan University UAV-borne Hyperspectral Image Station (WHU-Hi-Station) dataset, which is collected by the Wuhan University Remote Sensing Intelligent Data Extraction, Analysis and Application Group. This dataset is a collection of real-world remote sensing images with 270 bands and a spectral range of 400-1000 nm. The number of abnormal pixels is 1122, accounting for 0.047% of the entire image.

[0017] The feature selector corresponding to each input feature consists of a convolution , two-dimensional instance normalization and an activation function and is denoted as . According to the formula , the sample feature tensor output by the feature selector is obtained, where is the feature selector corresponding to the i-th feature; Meanwhile, the feature extraction network uses the formula to extract features of n different dimensions in the highly realistic abnormal images with pseudo-abnormalities , and uses the formula to obtain the abnormal feature tensor output by the feature selector; As a specific embodiment, the feature selector optimization sub-module trains the feature selector using the formula for minimizing the reconstruction loss function . By optimizing this learning objective, the ability of the feature selector output to suppress abnormal information during the inference process is accelerated, where is the cosine similarity function; As a specific embodiment, the feature selector optimization sub-module also includes using the formula for the contrast loss function to train the feature selector to push abnormal information away from the normal space, that is, the contrast loss is the cosine loss of the embedding margin f, and those higher than f are abnormal and will be continuously pushed away, forcing the projection layer to concentrate on learning deeper representations of normal features, where is the margin of the cosine embedding, is the maximum value function.

[0018] The adaptive fusion and reconstruction module is used to fuse and reconstruct the n sample feature tensors output by the trained feature selector into n normal feature tensors ; specifically, it includes the following, The feature enhancement sub-module aligns the dimensions of the feature tensor output after passing through the feature selector, and performs feature splicing to obtain the feature map , will be divided into m sub-channel groups. Let represent the features of the th channel group, and uses the formula for multi-scale feature enhancement, where represents global average pooling, which is used to extract channel attention information; is a convolutional layer, which is used for feature compression and restoration; Sigmoid is an activation function, which is used for normalizing to generate attention weights; is the output feature enhanced by multi-scale features; The feature fusion sub-module, after being processed by multi-scale feature enhancement, uses the formula to adaptively fuse the features of each channel group where, represents the concatenation function, represents the Gaussian error linear unit, represents the fused feature; The feature reconstruction sub-module reconstructs the features according to the formula to obtain the reconstructed normal feature tensor where, represents the i-th feature reconstruction network, and the feature reconstruction network adopts a symmetric but opposite structural design to the feature extraction network and is composed of symmetric residual decoding blocks constructed by deconvolution layers.

[0019] The detection module is used to calculate the anomaly scores of each pixel of the sample feature tensor and the reconstructed normal feature tensor respectively, and learn and optimize the reconstruction accuracy of the adaptive fusion and reconstruction module according to the scalar loss function to improve the anomaly detection and localization capabilities. It includes: The anomaly calculation sub-module, which obtains the anomaly score of each pixel in the i-th sample feature tensor according to the formula where, and ; respectively represent the length and width of, is the transpose function; is the feature vector corresponding to the pixel point with coordinates (h, w) in the image under the i-th feature dimension of the sample feature tensor , is the reconstructed normal feature tensor the feature vector corresponding to the pixel point with coordinates (h, w) in the image under the i-th feature dimension, is the L1 norm used to measure the length or magnitude of the vector. To localize the anomaly, the anomaly score is formulated as the pixel-by-pixel accumulation of the n-layer anomaly situation, and the scalar loss function is obtained. Learning is performed with this scalar loss function to obtain the optimal anomaly map represented by ; The scalar loss function is optimal when it no longer decreases, and at this time the anomaly performance index is the highest. During training, the image to be detected is a normal image, so it is optimal when the anomaly value of the anomaly map is the smallest.

[0020] The anomaly score map generation sub-module, which according to the formula Obtain the final anomaly score map , where represents the upsampling operation, that is, the anomaly score map is the final presentation after upsampling.

[0021] The model is trained by setting reasonable parameters. The final results use the area under the receiver operating characteristic curve (AUROC) and area under the per-region overlap (AUPRO), the most widely used evaluation metrics in anomaly detection, as evaluation metrics. The AUROC is 94.4% and the AUPRO is 92%.

[0022] Meanwhile, the present invention also discloses a training method for the remote sensing image anomaly detection model. This training method effectively utilizes the highly realistic anomaly images (pseudo-anomaly supervision signals) provided by the diffusion model for multi-task contrast constraints, and realizes more rapid and effective detection training for remote sensing anomalies, including the following steps: Step S1: Dynamically adjust the noise addition strategy for the input sample through the dynamic diffusion anomaly synthesis module to generate highly realistic anomaly images ; Step S2: The feature extraction network respectively extracts n different-dimensional sample features and anomaly features of the input sample and the highly realistic anomaly image , where i = 1, 2,..., n. After each extracted feature, a feature selector is connected to obtain the sample feature tensor and the anomaly feature tensor . The feature selector is optimized in feature extraction and transmission by minimizing the reconstruction loss and the contrast loss; Step S3: Input the n sample feature tensors extracted by the trained feature selector into the adaptive fusion reconstruction module to obtain the fusion feature , and use it to reconstruct n normal feature tensors . The detection module respectively calculates the anomaly scores of each pixel of the sample feature tensor and the reconstructed normal feature tensor and generates a scalar loss function, and then uses the scalar loss function to optimize the reconstruction accuracy of the adaptive fusion reconstruction module.

[0023] Among them, to achieve anomaly localization, the anomaly score is formulated as the pixel-by-pixel accumulation of the n-layer anomaly situation, and the scalar loss function is obtained.

[0024] Meanwhile, the present invention also discloses a method for remote sensing image anomaly detection. This method realizes the detection and positioning of anomalies by accurately reconstructing a normal template. It uses the aforementioned remote sensing image anomaly detection model, and its detection method includes the following steps: 1) Extract n features of different dimensions of the input image , where i = 1, 2, …, n, 2) Use the aforementioned feature selector to perform feature selection and extraction to obtain the input feature tensor , 3) Input the n sample features extracted into the adaptive fusion reconstruction module to obtain the fusion feature, and use it to reconstruct n reconstructed input feature tensors . For the reconstructed input feature tensor and the input image feature tensor perform a similarity comparison to obtain the anomaly situation of each pixel in the input feature tensor . This method significantly enhances the discrimination ability of the normal and abnormal feature boundaries and the detection sensitivity to subtle anomalies by means of diffusion enhancement and contrast constraint. Among them, in step 3, the final anomaly score map is obtained according to the formula , where represents the upsampling operation, and the anomaly position is reflected in the score map.

[0025] As mentioned above, it is only the preferred embodiment of the present invention, so the scope of implementation of the present invention cannot be limited thereby. That is, equivalent changes and modifications made according to the scope of the patent application of the present invention and the content of the specification should still fall within the scope covered by the patent of the present invention.

Claims

1. Remote sensing image anomaly detection model, characterized in that: including, A dynamic diffusion anomaly synthesis module, which is used to generate highly realistic anomaly images based on input samples ;​ a feature extraction and selection module, which includes feature extraction A network, a feature selector, and a feature selector optimization sub-module. The feature extraction network is used to extract the input samples and highly realistic abnormal images of n different-dimensional sample features and abnormal features , where i = 1, 2,..., n. The feature selector is used to perform selection processing on each extracted sample feature and abnormal feature respectively to obtain a sample feature tensor and an abnormal feature tensor . The feature selector optimization sub-module trains and optimizes the feature selector by minimizing the reconstruction loss function and the contrast loss function; An adaptive fusion and reconstruction module, which is used to fuse and reconstruct the n sample feature tensors output by the trained feature selector into n normal feature tensors ; A detection module, which is used to calculate the sample feature tensors respectively and the reconstructed normal feature tensors The anomaly score of each pixel, and learn and optimize the reconstruction accuracy of the adaptive fusion reconstruction module according to the scalar loss function to improve the anomaly detection and localization capabilities.

2. The remote sensing image anomaly detection model according to claim 1, wherein: the dynamic diffusion anomaly synthesis module includes, Forward noise addition sub-module, which at time the conditional probability distribution of the noise data is , where , is a parameter between 0 and 1, , is a fixed variance schedule, is a latent variable, I is the Identity Matrix, whose dimension matches the vector dimension, represents the state vector at the initial time, represents the random state vector at time t, is the Gaussian distribution function; The reverse denoising sub-module introduces a controllable perturbation mechanism during the reverse sampling process to directionally guide the generation process to shift towards the low-probability density region in the latent space, and the probability distribution of the reverse process is , using a neural network to model the noise in the diffusion process , thus obtaining , representing the random state vector at time t-1; The loss function design sub-module aims to minimize the variational upper bound of the negative log-likelihood and uses the objective function as to train a diffusion model to learn the distribution of normal images, where is the mathematical expectation, is to calculate the L2 norm, ϵ is the true noise, is the noise predicted by the model, and the conditional probability distribution , is the abnormal image obtained at time , is the introduced additional perturbation, , is the abnormal image sampled after the introduced additional perturbation. Among them, the scalar controls the anomaly intensity.

3. The remote sensing image anomaly detection model according to claim 2, wherein: the adaptive fusion and reconstruction module includes, A feature enhancer module that dimensionally aligns and concatenates the sample feature tensors output after passing through the feature selector to obtain a feature map , divides them into m sub-channel groups, and performs multi-scale feature enhancement on the features of the channel groups to obtain enhanced features , , ; Feature fusion sub-module, which performs adaptive feature fusion on the enhanced features to obtain the fused features representing the fused result ; The feature reconstruction sub-module, for the fused features uses a feature reconstruction network to obtain a reconstructed normal feature tensor , where the feature reconstruction network is designed with a structure symmetric but opposite to that of the feature extraction network.

4. The remote sensing image anomaly detection model according to claim 3, wherein: the detection module includes, Anomaly calculation sub-module, which according to the formula obtains the anomaly score of each pixel point in the reconstructed normal image under the i-th feature dimension , and then uses the scalar loss function to learn and optimize the reconstruction accuracy of the adaptive fusion reconstruction module to obtain the optimal anomaly score represented by ; where , , respectively represent the length and width of the normal feature tensor , is the transpose function; is the sample feature tensor the feature vector corresponding to the pixel point with coordinates (h, w) in the image under the i-th feature dimension, is the reconstructed normal feature tensor the feature vector corresponding to the pixel point with coordinates (h, w) in the image under the i-th feature dimension, is the L1 norm used to measure the length or magnitude of a vector; Anomaly score map generation sub-module, which obtains the final anomaly score map according to the formula where denotes the upsampling operation. ​ 5. The training method of the remote sensing image anomaly detection model according to any one of claims 1-4, characterized in that, including the following steps, Step S1, input the sample Generate a highly realistic abnormal image by dynamically adjusting the noise addition strategy through the dynamic diffusion anomaly synthesis module ; Step S2, the feature extraction network extracts the input samples and highly realistic abnormal images of n different-dimensional sample features and abnormal features , where i = 1, 2, …, n. After each extracted feature, a feature selector is connected to obtain a sample feature tensor and an abnormal feature tensor , and the feature selector is optimized in feature extraction and transmission by minimizing the reconstruction loss and the contrast loss; Step S3: Input the n sample feature tensors extracted by the trained feature selector into the adaptive fusion and reconstruction module to obtain the fused features , and use them to reconstruct n normal feature tensors . The detection module calculates the anomaly scores of each pixel of the sample feature tensors and the reconstructed normal feature tensors respectively to generate a scalar loss function, and then use the scalar loss function to optimize the reconstruction accuracy of the adaptive fusion and reconstruction module.

6. A method for detecting anomalies in remote sensing images, characterized in that, using the remote sensing image anomaly detection model according to any one of claims 1-4, the detection method includes the following steps, 1) The feature extraction network extracts input features of n different dimensions from the input image , where i = 1, 2, …, n 2) The feature selector performs feature selection to obtain the input feature tensor , 3) Input the extracted n sample features into the adaptive fusion and reconstruction module to obtain the fused features, and use them to reconstruct n reconstructed input feature tensors , for the reconstructed input feature tensors and the input image feature tensor perform similarity comparison calculations to obtain the anomaly situation of each pixel in the input feature tensor ; 4) generating and outputting an anomaly score map.

Images