Hyperspectral anomaly detection method based on detail and saliency difference convolutional network

Abstract

The application discloses a hyperspectral anomaly detection method based on a difference and saliency difference convolutional network. First, assuming that an abnormal pixel is difficult to be reconstructed by the features of surrounding pixels, a difference convolutional network is introduced to extract local recurrent neighborhood features of a hyperspectral image (HSI), i.e., difference features obtained by taking the difference between outer window features and inner window features. The difference convolution includes two convolution kernels with different receptive fields, and the outer window features and the inner window features are obtained through 5*5 and 3*3 convolution kernels respectively. Second, an attention map of the outer window is extracted through a feature fusion module based on a detail-based attention (LDA) and a saliency-based attention (LTA) to guide the 3*3 features of the inner window, and the feature extraction capability of the 3*3 features of the inner window is enhanced. Then, the obtained local recurrent neighborhood features are used to reconstruct a pure background image B of the hyperspectral image. Finally, an abnormal target detection result is extracted according to a difference image E between the input hyperspectral image Y and the reconstructed hyperspectral image background B. The method enhances the local feature extraction capability of the network and improves the anomaly detection performance.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a hyperspectral anomaly detection method based on detail and saliency difference convolutional networks. Background Technology

[0002] As an important branch of image processing, hyperspectral anomaly detection is a highly valuable research topic. Hyperspectral images have a wide field of view, targets occupy only a few pixels, are very small, and have low spatial resolution, making hyperspectral anomaly detection challenging. Hyperspectral anomaly detection involves unsupervised detection of targets whose spatial and spectral features differ from the surrounding background without any prior information. Hyperspectral anomaly detection methods are broadly classified into two categories: traditional methods and deep learning-based methods.

[0003] Traditional methods can be further divided into statistical methods and data representation-based methods. Statistical methods aim to establish a background distribution model of hyperspectral images (HSI) and determine target locations through hypothesis testing. However, in real hyperspectral images, when background information is complex, the mean and covariance used by statistical methods are easily affected by anomalies and noise, which impacts detection performance. Data representation-based methods assume that background pixels can be reconstructed from some similar background pixels. Data representation-based methods face challenges in establishing accurate background dictionaries, reducing the accuracy of the obtained low-rank and sparse matrices.

[0004] Deep learning-based methods can extract latent deep features from hyperspectral images (HSI) through their outstanding nonlinear fitting capabilities, achieving accurate mapping of complex scenes. Deep learning-based methods can be categorized into supervised and unsupervised methods based on label availability. Supervised methods train classifiers with a large number of labels to facilitate anomaly detection. However, due to the lack of ground truth labels, these methods use other methods to generate pseudo-labels. Unsupervised methods extract deep features from hyperspectral images (HSI) without labels. However, they neglect local features surrounding anomalies, which are crucial for hyperspectral anomaly detection. Training the network solely by extracting features globally makes it difficult to achieve a clean hyperspectral image background, limiting the effectiveness of these methods in anomaly detection. Summary of the Invention

[0005] In view of this, the main objective of the present invention is to provide a hyperspectral anomaly detection method based on detail and saliency difference convolutional networks.

[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows: This invention discloses a hyperspectral anomaly detection method based on detail and saliency difference convolutional networks. The method is as follows: Step 1: Read in the t-th frame hyperspectral image Y. The input HSI is represented as Y = (y1, y2, ..., y3). M·N )∈R M×N×B M and N are the spatial dimensions of HSI, and B is the spectral dimension; Step 2, Build 5 5's outer window and 3 The inner window of 3 forms a local loop-shaped neighborhood S; Step 3, extract 5 5. External window features 3 3. Inner window features ; Step 4: Extract local loop-shaped neighborhood S features from the input image Y, and extract the differential features F; Step 5: Based on the learnable weighting parameter λ introduced in the difference operation, we obtain... ; Step 6: Based on the detail-based attention LDA module... Extract detailed features and add features ,get ;Will replace The input is fed into a difference network to obtain local circular neighborhood features. F 1; Step six specifically includes: Given the features of the inner window (3×3) of the input and input outer window 5×5 features ,and , , ; Based on the 5×5 features of the outer window of input Y In the learning process, the basic convolutional kernel W={Wi|i=1,2,...,nbase}∈R nbase×C×Cgroup×k×k The fusion coefficient D; where W is a group of nbase convolutional kernels, each kernel having a size of C×C. group ×k×k,C group It is the number of channels in the basic convolutional group W, where each kernel performs grouped convolutions, and the number of groups is C / Cgroup. group k is the spatial dimension of the convolution kernel; Features Local detail information is extracted sequentially through 3×3 convolution, followed by simple gate activation to enhance feature propagation. Finally, 1×1 convolution maps the features to the number of base convolution kernels, nbase, to obtain the fusion coefficient D. ; Based on the fusion coefficient D, the basic convolutional kernels W are linearly multiplied to obtain the fused convolutional kernel K. i , j )∈R C ×Cgroup×k×k , In the formula, D represents the t-th convolution kernel in the base convolution kernel W. i , j , t ) represents the pixel in D ( i , j The t-th fusion coefficient at position t; According to the fusion convolution kernel K( i , j )and Perform grouped convolutions to compute detail-focused features. D ( i , j ) Location characteristics , ; Based on detailed attention to features F DA With input inner window features F in Weighted summation yields the output inner window features. F inda This brings more attention to detail to the 3x3 inner window feature. Where γF represents the learnable weight parameters; replace As input to the next level of differential convolution, we obtain F1; Step 7: Based on the saliency-based attention LTA module... Perform salient feature extraction and add features , to obtain features Then, through 1×1 convolution, we obtain... ;Will replace The input is fed into a difference network to obtain local circular neighborhood features. F 4; Step 8: Concatenate the four sets of local loop-shaped neighborhood features along the channel dimension to obtain the combined difference features. F combine ; Step 9: Use a 1×1 convolutional layer and LeakyReLU activation layer Fusion combined difference features F combine The predicted hyperspectral image background B is obtained; Step 10: Calculate the difference between the input hyperspectral image Y and the predicted hyperspectral image background B to obtain the anomaly detection result E; Step 11: During training, reduce the loss function by 5. 5. The reconstructed image Y1 from the outer window feature extraction branch is compared with the input hyperspectral image Y, and the image is reduced to 3. The difference between the reconstructed image Y2 and the input Y is reduced by the inner window feature extraction branch, which narrows the gap between the hyperspectral background image B and the input hyperspectral image Y. The difference network parameters are optimized, including convolution weights w and weighting parameters λ.

[0007] Optionally, step three specifically includes: Features are extracted from the input hyperspectral image Y using a 5×5 convolution kernel. Features are extracted from the input hyperspectral image Y using a 3×3 convolution kernel. ; Through 5 5. Outer window feature extraction branch and 3 3. The feature extraction branch of the inner window is subtracted to functionally extract a local loop-shaped neighborhood S; In a local circular neighborhood window, the objective function is represented as follows: In the formula, Xs( The set of all pixels yi in a local circular neighborhood s, represented by ) is the network parameter to be optimized, f(·) is the mapping function from network input to output, and a is the regularization parameter.

[0008] Optionally, step four specifically includes: extracting local loop-shaped neighborhood S features from the t-th frame hyperspectral image Y, and extracting the difference features F, denoted as... Where w is the convolution weight defined on the set Xs, and b is the bias defined on the set Xs. This represents the convolution operation, and F represents the convolution output feature. The shape representation Xs of the local loop-shaped neighborhood features is based on this. ; where X out and X in In pixels y i Centered pixel sets of the outer and inner windows in a localized area; This represents the set difference operation; it is worth noting that the convolution weights w and bias b act on the local loop-shaped neighborhood S. The difference feature F is represented as

[0009]

[0010]

[0011] and They represent X out and X in The convolution output features.

[0012] Optionally, step five specifically includes: The difference convolution is represented as

[0013] Four sets of differential features obtained using differential convolutional networks , ; where, obtained directly from the above formula F 2， F 3。

[0014] Optionally, step seven specifically includes: Given input inner window characteristics and input outer window features Through the Performing a 1×1 convolution yields the query map. ; For the query mapping graph Q, after passing through an r×r convolutional layer, a Leaky ReLU activation layer, and a 1×1 convolutional layer, the coefficient mapping is obtained. The r×r convolutional layer is used to capture nonlocal receptive fields and map feature channels to r. 2 The LeakyReLU function activation layer enhances the nonlinear representation capability, and the 1×1 convolutional layer merges channel features to establish the correlation of attention coefficients within the receptive field. Extract bits ( i , j correlation coefficient F qv ( i , j ), and reshape it into an r×r coefficient matrix T; Calculate pixel position ( i , j ) based on saliency attention features , In the formula, Indicates in Location( i , j A nonlocal region of size r; Attention features based on salience With features F in The composite features are obtained by splicing. F tac, ; Feature fusion is performed using a 1×1 convolutional layer to obtain the output inner window features. ; Used to replace As input to the next level of differential convolution.

[0015] Optionally, step eight specifically includes: according to Four sets of difference features are concatenated along the channel dimension to obtain the combined difference feature. F combine .

[0016] Optionally, step nine specifically includes: Using 1×1 convolutional layers and LeakyReLU activation layer By fusing all the differential features, the predicted hyperspectral image background B is obtained. .

[0017] Optionally, step ten specifically includes: Anomaly detection result E is represented as , In the formula, For input pixels y i The j One element, To predict background pixels b i The j One element, E i For position i The value of E in the anomaly detection result, where C represents the band number.

[0018] Optionally, step eleven specifically includes: The total loss function is .

[0019] Compared with existing technologies, this invention enhances the ability to extract local network features and improves anomaly detection performance. Attached Figure Description

[0020] Figure 1 This is a flowchart of the present invention.

[0021] Figure 2 This is a schematic diagram of the local loop-shaped neighborhood S constructed in this invention.

[0022] Figure 3 In this embodiment of the invention, the input hyperspectral image WHU-Hi-River and 3 3. Window features Fin and 5 5. Window feature Fout.

[0023] Figure 4 The difference features of the input hyperspectral image WHU-Hi-River are shown in this embodiment of the invention.

[0024] Figure 5 In an embodiment of the present invention, λ 1, λ 2, λ 3, λ Screenshot of parameters 4.

[0025] Figure 6 In this embodiment of the invention, the fusion coefficient D is used.

[0026] Figure 7 In this embodiment of the invention, the learning parameter γF is used.

[0027] Figure 8 This is the simulation result of anomaly detection of the input hyperspectral image Y in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0029] This invention provides a hyperspectral anomaly detection method based on detail and saliency difference convolutional networks, such as... Figure 1 As shown, the method is as follows: Assume that background pixels of the input HSI can be represented by other pixels within their local loop-shaped neighborhood, while outlier pixels cannot.

[0030] Step 1: Read in the t-th frame hyperspectral image Y. The input HSI is represented as Y = (y1, y2, ..., y3). M·N )∈R M×N×B M and N are the spatial dimensions of HSI, and B is the spectral dimension.

[0031] In this invention, the input hyperspectral image Y is the dataset WHU-Hi-River, and the resolution of the hyperspectral data is 105. 168, spectral dimension B=135. Figure 3 (a) is the input image Y in this embodiment of the invention.

[0032] Step 2, Build 5 5's outer window and 3 The inner window of 3 forms a local loop-shaped neighborhood S. Figure 2The gray area S, which is a local loop-shaped neighborhood constructed in this invention.

[0033] Step 3, extract 5 5. External window features 3 3. Inner window features . Figure 3 (b) is the inner window feature in the embodiment of the present invention. , Figure 3 (c) is the outer window feature in the embodiment of the present invention. .

[0034] Specifically, a 5×5 convolution kernel is used to extract features from the input hyperspectral image Y. Features are extracted from the input hyperspectral image Y using a 3×3 convolution kernel. .

[0035] Through 5 5. Outer window feature extraction branch and 3 The inner window feature extraction branch is subtracted to functionally extract local loop-shaped neighborhood S features.

[0036] In a local circular neighborhood window, the objective function can be represented as

[0037] In the formula Xs( The local circular neighborhood S pixels y represents the local circular neighborhood S pixels y. i All sets, is the network parameter to be optimized, f(·) is the mapping function from network input to output, and a is the regularization parameter.

[0038] Local circular neighborhood features (difference features) of input Y are extracted using a four-layer difference network. F 1 , F 2 , F 3 , F 4 .

[0039] Step 4: Extract local loop-shaped neighborhood S features from the input image Y, and extract the differential features F.

[0040] Specifically, the general form of the differential convolution F can be expressed as:

[0041] Where w is the convolution weight value defined on the set Xs, and b is the bias defined on the set Xs. denoted as convolution operation, and F represents the output feature of convolution.

[0042] Based on the shape of the circular neighborhood S, Xs is represented as...

[0043] In the formula, X out and X in In pixels y i 5 centrally located exterior windows 5 and partial inner window 3 3-pixel set; This represents the set difference operation. It's worth noting that the convolution weights w and bias b act on the local loop-shaped neighborhood S. Therefore, their values ​​are local within the inner window X. in The value above is 0.

[0044] F It can also be expressed as

[0045]

[0046]

[0047] F 'and They represent X respectively out and X in The convolution output features.

[0048] Step 5: Based on the learnable weighting parameter λ introduced in the difference operation, we obtain... .

[0049] Specifically: The difference convolution can be expressed as

[0050]

[0051] Four sets of differential features obtained using a differential convolutional network (four convolutional layers) .

[0052] The difference features are obtained directly through the above formula. F 2， Difference features F 3 In this embodiment of the invention, Figure 4 (b) The present invention inputs the F2 of the differential features of the hyperspectral image WHU-Hi-River. Figure 4 (c) The F3 of the differential features of the input hyperspectral image WHU-Hi-River in this invention. Parameter vector in the embodiments of this invention. The values ​​are as follows Figure 5Embodiments of the present invention λ 1, λ 2, λ 3, λ Screenshot of parameters 4.

[0053] Step 6: Based on the detail-based attention (LDA) module... Extract detailed features and add features ,get ;Will replace The input is fed into the differential network to obtain... F 1.

[0054] Specifically: Given the features of the inner window (3×3) of the input and input outer window 5×5 features ,and , , .

[0055] Based on the 5×5 features of the outer window of input Y In the learning process, the basic convolutional kernel W={Wi|i=1,2,...,nbase}∈R nbase×C×Cgroup×k×k The fusion coefficient D.

[0056] Where W is a group of nbase convolutional kernels, each kernel having a size of C×C. group ×k×k,C group It is the number of channels in the basic convolutional group W, where each kernel performs grouped convolutions, and the number of groups is C / Cgroup. group , where k is the spatial dimension of the convolution kernel. In this invention example, C group =4 ， k=3, C=64.

[0057] 5x5 feature in outer window Perform fusion attention operation on the features, specifically: Local detail information is extracted sequentially through 3×3 convolution, followed by simple gate activation to enhance feature propagation. Finally, 1×1 convolution maps the features to the number of base convolution kernels, nbase, to obtain the fusion coefficient D. In this invention, the fusion coefficient D is a feature map, such as... Figure 6 The fusion coefficient D in this embodiment of the invention.

[0058] Based on the fusion coefficient D, the basic convolutional kernels W are linearly multiplied to obtain the fused convolutional kernel K. i , j )∈R C×Cgroup×k×k Specifically:

[0059] In the formula, D represents the t-th convolution kernel in the base convolution kernel W. i , j , t ) represents the pixel in D ( i , j The t-th fusion coefficient at position ).

[0060] According to the fusion convolution kernel K( i , j )and Perform grouped convolutions to compute detail-focused features. D ( i , j ) Location characteristics

[0061]

[0062] Based on detailed attention to features F DA With input inner window features F in Weighted summation yields the output inner window features. F inda This brings more attention to detail to the 3x3 inner window feature.

[0063]

[0064] Where γF represents the learnable weight parameters. In this embodiment of the invention, γF is a vector. Figure 7 Screenshot of the learning parameter γF in an embodiment of the present invention.

[0065] replace As input to the next level of differential convolution, F1 is obtained. In this embodiment of the invention, Figure 4 (a) F1 of the differential features of the input hyperspectral image WHU-Hi-River in this invention.

[0066] Step 7: Based on the saliency-based attention (LTA) module... Perform salient feature extraction and add features , to obtain features Then through Convolution yields .Will replace Input to the differential network F 4.

[0067] Specifically: given the input inner window characteristics and input outer window features Through the Performing a 1×1 convolution yields the query map. .

[0068] Two convolutional layers are used to approximate the matrix multiplication in the attention mechanism (Transformer) to perform Q-convolution on the query mapping graph, resulting in a relevant mapping graph. .

[0069] Specifically, the query mapping graph Q is processed through an r×r convolutional layer, a Leaky ReLU activation layer, and a 1×1 convolutional layer to obtain the coefficient mapping. In this example of the invention, r=7.

[0070] The r×r convolutional layer is used to capture nonlocal receptive fields and map feature channels to r. 2 The LeakyReLU activation layer enhances the nonlinear representation capability, and the 1×1 convolutional layer merges channel features, establishing the correlation of attention coefficients within the receptive field. In this invention example, r 2 =49. Extract pixel position ( i , j correlation coefficient F qv ( i , j ), and reshape it into an r×r coefficient matrix T. Calculate pixel position ( i , j ) based on saliency attention features

[0071]

[0072] In the formula, Indicates in Location( i , j A nonlocal region of size r.

[0073] Attention features based on salience With features F in The composite features are obtained by splicing. F tac .

[0074] F tac It can be represented as

[0075] Feature fusion is performed using a 1×1 convolutional layer to obtain the output inner window features. . Used to replace As input to the next level of differential convolution, F4 is obtained.

[0076] Step 8: Concatenate the four sets of difference features along the channel dimension to obtain the combined difference features. F combine

[0077] Step 9: Fuse all differential features using a 1×1 convolutional layer and a LeakyReLU activation layer. The predicted hyperspectral image background B is obtained.

[0078] Specifically, it involves using a 1×1 convolutional layer. and LeakyReLU activation layer By fusing all the differential features, the predicted hyperspectral image background B is obtained. In this invention example, Figure 8 (b) Background B of the hyperspectral image of an embodiment of the present invention.

[0079] B can be described as

[0080] Step 10: Calculate the difference between the input hyperspectral image Y and the predicted hyperspectral image background B to obtain the anomaly detection result E.

[0081] The anomaly detection result E can be represented as ,

[0082] In the formula, For input pixels y i The j One element, To predict background pixels b i The j One element, E i For position i The value of the anomaly detection result E is given by the formula, where C represents the number of bands. In this embodiment of the invention, the number of bands C = 135. In this embodiment of the invention, Figure 8 (c) Hyperspectral image anomaly detection result E of the present invention embodiment.

[0083] Step 11: During training, reduce the loss function by 5. 5. The reconstructed image Y1 from the outer window feature extraction branch is compared with the input hyperspectral image Y, and the image is reduced to 3. The difference between the reconstructed image Y2 and the input Y is calculated by the inner window feature extraction branch, which reduces the difference between the hyperspectral background image B and the input hyperspectral image Y. The parameters of the difference network are optimized, such as the convolution weight w and the weighting parameter λ.

[0084] The loss function during training is as follows:

[0085] Where SmoothL1(·) is the smooth L1 function.

[0086] For the differential network predicting background B, since anomalies account for a small proportion of HSI, and considering computational efficiency, input Y is chosen as the optimization objective. The loss function for the background is expressed as follows:

[0087] Therefore, the total loss function 。

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. A method for detecting hyperspectral anomalies based on detail and saliency difference convolutional networks, characterized in that, The method is as follows: Step 1: Read in the t-th frame hyperspectral image Y. The input HSI is represented as Y = (y1, y2, ..., y3). M·N )∈R M×N×B M and N are the spatial dimensions of HSI, and B is the spectral dimension; Step 2, Build 5 5's outer window and 3 The inner window of 3 forms a local loop-shaped neighborhood S; Step 3, extract 5 5. External window features 3 3. Inner window features ; Step 4: Extract local loop-shaped neighborhood S features from the input image Y, and extract the differential features F; Step 5: Based on the learnable weighting parameter λ introduced in the difference operation, we obtain... ; Step 6: Based on the detail-based attention LDA module... Extract detailed features and add features ,get ;Will replace The input is fed into a difference network to obtain local circular neighborhood features. F 1; Step six specifically includes: Given the features of the inner window (3×3) of the input and input outer window 5×5 features ,and , , ; Based on the 5×5 features of the outer window of input Y In the learning process, the basic convolutional kernel W={Wi|i=1,2,...,nbase}∈R nbase ×C×Cgroup×k×k The fusion coefficient D; where W is a group of nbase convolutional kernels, each kernel having a size of C×C. group ×k×k,C group It is the number of channels in the basic convolutional group W, where each kernel performs grouped convolutions, and the number of groups is C / Cgroup. group k is the spatial dimension of the convolution kernel; Features Local detail information is extracted sequentially through 3×3 convolution, followed by simple gate activation to enhance feature propagation. Finally, 1×1 convolution maps the features to the number of base convolution kernels, nbase, to obtain the fusion coefficient D. ; Based on the fusion coefficient D, the basic convolutional kernels W are linearly multiplied to obtain the fused convolutional kernel K. i , j )∈R C×Cgroup×k×k , In the formula, D represents the t-th convolution kernel in the base convolution kernel W. i , j , t ) represents the pixel in D ( i , j The t-th fusion coefficient at position t; According to the fusion convolution kernel K( i , j )and Perform grouped convolutions to compute detail-focused features. D ( i , j ) Location characteristics , ; Based on detailed attention to features F DA With input inner window features F in Weighted summation yields the output inner window features. F inda This brings more attention to detail to the 3x3 inner window feature. , where γF represents the learnable weight parameters; replace As input to the next level of differential convolution, we obtain F1; Step 7: Based on the saliency-based attention LTA module... Perform salient feature extraction and add features , to obtain features Then, through 1×1 convolution, we obtain... ;Will replace The input is fed into a difference network to obtain local circular neighborhood features. F 4; Step 8: Concatenate the four sets of local loop-shaped neighborhood features along the channel dimension to obtain the combined difference features. F combine ; Step 9: Use a 1×1 convolutional layer and LeakyReLU activation layer Fusion combined difference features F combine The predicted hyperspectral image background B is obtained; Step 10: Calculate the difference between the input hyperspectral image Y and the predicted hyperspectral image background B to obtain the anomaly detection result E; Step 11: During training, reduce the loss function by 5.

5. The reconstructed image Y1 from the outer window feature extraction branch is compared with the input hyperspectral image Y, and the image is reduced to 3. The difference between the reconstructed image Y2 and the input Y is reduced by the inner window feature extraction branch, which narrows the gap between the hyperspectral background image B and the input hyperspectral image Y. The difference network parameters are optimized, including convolution weights w and weighting parameters λ.

2. The hyperspectral anomaly detection method based on detail and saliency difference convolutional networks according to claim 1, characterized in that, Step three specifically involves: Features are extracted from the input hyperspectral image Y using a 5×5 convolution kernel. Features are extracted from the input hyperspectral image Y using a 3×3 convolution kernel. ; Through 5 5. Outer window feature extraction branch and 3 3. The feature extraction branch of the inner window is subtracted to functionally extract a local loop-shaped neighborhood S; In a local circular neighborhood window, the objective function is represented as follows: In the formula, Xs( The set of all pixels yi in a local circular neighborhood s, represented by ) is the network parameter to be optimized, f(·) is the mapping function from network input to output, and a is the regularization parameter.

3. The hyperspectral anomaly detection method based on detail and saliency difference convolutional networks according to claim 2, characterized in that, Step four specifically includes: extracting local loop-shaped neighborhood S features from the t-th frame hyperspectral image Y, and extracting the difference features F, denoted as... Where w is the convolution weight defined on the set Xs, and b is the bias defined on the set Xs. This represents the convolution operation, and F represents the convolution output feature. The shape representation Xs of the local loop-shaped neighborhood features is based on this. ; where X out and X in In pixels y i Centered pixel sets of the outer and inner windows in a localized area; This represents the set difference operation; it is worth noting that the convolution weights w and bias b act on the local loop-shaped neighborhood S. The difference feature F is represented as and They represent X out and X in The convolution output features.

4. The hyperspectral anomaly detection method based on detail and saliency difference convolutional networks according to claim 3, characterized in that, Step five specifically includes: The difference convolution is represented as Four sets of differential features obtained using differential convolutional networks , ; where, obtained directly from the above formula F 2， F 3。 5. The hyperspectral anomaly detection method based on detail and saliency difference convolutional networks according to claim 1, characterized in that, Step seven specifically includes: Given input inner window characteristics and input outer window features Through the Performing a 1×1 convolution yields the query mapping graph. ; For the query mapping graph Q, after passing through an r×r convolutional layer, a Leaky ReLU activation layer, and a 1×1 convolutional layer, the coefficient mapping is obtained. The r×r convolutional layer is used to capture nonlocal receptive fields and map feature channels to r. 2 The LeakyReLU function activation layer enhances the nonlinear representation capability, and the 1×1 convolutional layer merges channel features to establish the correlation of attention coefficients within the receptive field. Extract bits ( i , j correlation coefficient F qv ( i , j ), and reshape it into an r×r coefficient matrix T; Calculate pixel position ( i , j ) based on saliency attention features , In the formula, Indicates in Location( i , j A nonlocal region of size r; Attention features based on salience With features F in The composite features are obtained by splicing. F tac , ; Feature fusion is performed using a 1×1 convolutional layer to obtain the output inner window features. ; Used to replace As input to the next level of differential convolution.

6. The hyperspectral anomaly detection method based on detail and saliency difference convolutional networks according to claim 5, characterized in that, Step eight specifically includes: according to Four sets of difference features are spliced ​​along the channel dimension to obtain the combined difference feature. F combine .

7. The hyperspectral anomaly detection method based on detail and saliency difference convolutional networks according to claim 6, characterized in that, Step nine specifically includes: Using 1×1 convolutional layers and LeakyReLU activation layer By fusing all the differential features, the predicted hyperspectral image background B is obtained. .

8. The hyperspectral anomaly detection method based on detail and saliency difference convolutional networks according to claim 7, characterized in that, Step ten specifically includes: Anomaly detection result E is represented as , In the formula, For input pixels y i The j One element, To predict background pixels b i The j One element, E i For position i The value of E in the anomaly detection result, where C represents the band number.

9. The hyperspectral anomaly detection method based on detail and saliency difference convolutional networks according to claim 8, characterized in that, Step eleven specifically includes: The total loss function is .

Images