Hyperspectral remote sensing image abnormal target detection method based on double-decoder double-Transform background enhancement reconstruction network

Abstract

The invention discloses a hyperspectral remote sensing image abnormal target detection method based on a double-decoder double-Transform background enhancement reconstruction network. According to the method, a novel background feature purification decoding task is introduced and designed, and the influence of abnormal pixels in the reconstruction process is reduced by focusing on background features. And in consideration of the characteristic that the exception is different from the background, an exception stripping shielding auxiliary branch module is designed to weaken the exception feature, and richer non-exception background information is supplemented for the decoding stage. Besides, in order to further guide reconstruction of a pure background, a novel cross guidance Transform module is further designed and is used for globally integrating multi-scale background attribute features. And finally, by using a collaborative optimization loss function, carrying out double constraint learning on potential anomalies in a spatial spectrum and a frequency domain, and completing accurate detection of an abnormal target. Experimental results on a hyperspectral data set show that the accuracy and robustness of anomaly detection can be remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of abnormal target detection in remote sensing images, and in particular, relates to a problem and a solution for abnormal target detection in hyperspectral remote sensing images based on a double-decoding double-Transformer background enhancement reconstruction network. Background Art

[0002] Hyperspectral remote sensing is an earth observation technology that can accurately analyze and identify surface features by acquiring observation data of reflection or radiation of surface objects in multiple spectral bands. Anomaly detection in hyperspectral remote sensing images refers to the identification of abnormal targets with spectral characteristics significantly different from the background without prior information about the target. Hyperspectral images (HSI) have the ability to identify and locate the fine spectral characteristics of ground objects. They are of great significance in remote sensing observations and are widely used in earth science, remote sensing, agriculture, military and environmental monitoring. In this task, the lack of prior knowledge of the target and the background makes it a key challenge to accurately detect abnormal targets from the background. To address this problem, researchers have proposed various methods to model the background and use different strategies to detect abnormal targets that deviate from the background pattern.

[0003] Existing anomaly target detection methods can be divided into three categories: statistical theory-based methods, representation learning-based methods, and deep learning-based methods. Statistical theory-based anomaly detection methods usually model the background by using the statistical characteristics of HSI. This method assumes that the background follows a multivariate Gaussian distribution and detects anomalies through the Mahalanobis distance. However, due to the complexity of the HSI background and the interference of various factors, it is very challenging to build a model consistent with the real scene. In contrast, representation-based methods reconstruct background pixels through dictionaries, and abnormal pixels cannot be effectively reconstructed. The reconstruction error is used to detect anomalies. Such methods include sparse representation (SR) and collaborative representation (CR). The SR method assumes that normal samples can be represented by a small number of atoms, while abnormal samples cannot be reconstructed in this way; the CR method emphasizes the collaborative relationship between dictionary atoms and evaluates whether pixels can be represented by linear combinations of adjacent pixels. Aiming at this feature, an abnormal target detection (CRD) method based on collaborative representation learning is proposed.

[0004] The HAD method based on deep learning enhances the recognition ability of background and abnormal targets by extracting deep features. Common methods are divided into models based on similarity measurement and models based on reconstruction. Similarity measurement models usually use convolutional neural networks (CNN) as the core framework and select a large number of pixel pairs from labeled hyperspectral images during training. Unsupervised deep learning technology based on reconstruction has good effectiveness and adaptability in anomaly detection tasks. Common network structures include autoencoders (AE) and generative adversarial networks (GAN). For example, Jiang et al. proposed a weakly supervised GAN based on spectral constraints to enhance the recognition ability of abnormal targets. The AE method reconstructs the background through the encoding and decoding process and uses the reconstruction error to detect anomalies. In order to improve robustness, a fully autointegrated autoencoder (Auto-AD) was developed, which combined with adaptive weighted loss to suppress anomalies. The guided autoencoder (GAED) proposed by Xiang et al. enhances the background representation ability. Self-supervised anomaly detection technology has also received attention. PDBSNet and BS3LNet use blind spot networks to reduce the feature representation of abnormal targets. The self-supervised multi-scale network (MSNet) proposed by Liu et al. combines multi-scale feature extraction to handle anomalies of different scales.

[0005] In recent years, deep learning has made significant progress in the fields of computer vision and remote sensing. However, existing methods cannot fully extract effective anomaly information from hyperspectral remote sensing images for separation, and it is difficult to reconstruct high-quality background information. Therefore, we propose a dual-decoder dual-Transformer background enhancement reconstruction network to capture more useful information, and further improve the accuracy of abnormal target detection results by using the spatial dependency of the background learned by each stage of the network encoder-decoder structure. Summary of the invention

[0006] The purpose of this invention is to solve the reconstruction mapping problem caused by anomalies in the feature encoding-decoding mechanism of the hyperspectral remote sensing image reconstruction process, and specifically propose a background enhancement reconstruction network based on dual decoders and dual Transformers, which dynamically adjusts its attention to background and anomalies, reduces the spatial dependence related to abnormal pixels, and makes full use of the hyperspectral information at each stage. Its core lies in the following four parts:

[0007] 1. Dual decoder

[0008] Conventional U-shaped networks usually include an encoder, a decoder, and regular skip connections. Among them, although the skip connection can partially alleviate the problem of feature loss, it may still cause the loss of important information when processing complex hyperspectral data. In order to enhance the extraction of information, a simple way is to increase the depth of the network by stacking convolution blocks, which can obtain powerful semantic information extraction capabilities, but this operation will make it difficult to avoid the problem of equivalent reconstruction of abnormal targets in the process of background generation, and the background image cannot be accurately estimated, resulting in a decrease in the ability to detect abnormal targets. In addition, the single path in the decoding stage encounters challenges in generating robust feature maps. Therefore, we designed a double decoding structure to obtain more accurate background information without abnormalities while maintaining feature sharing between the upper and lower layers.

[0009] The dual decoding network backbone of the present invention is designed based on a U-shaped network and consists of two layers of parallel decoding subnetworks. First, the hyperspectral remote sensing image HSI is represented as Where H, W, and C represent the height, width, and number of spectral bands, respectively. The encoder at the upper layer accepts the original hyperspectral image as input and changes the number of channels from C to N through a 3×3 convolution layer. Then, the feature map passes through two Swin Transformer modules, a 1×1 convolution layer, and a downsampling layer for shallow feature extraction. The downsampling layer is a 2×2 stride convolution layer, which halves the spatial resolution of the feature map and doubles the channels. The sizes of feature maps at different nodes become At the same time, the bottom layer of the model is a single SwinTransformer module, which further extracts features and keeps the number of channels and spatial resolution of the feature map unchanged as Then, symmetrically with the encoder, the upper decoder uses two 2×2 stride deconvolution layers, a 1×1 convolution layer and Swi n Transform r The module is used to learn the deep feature map representation. After decoding, the size of the feature map returns to H×W×C.

[0010] The lower decoder uses the mapping input uniform noise, and obtains the feature map through the background feature fusion module and the upsampling layer. Then, the feature maps from the upper encoding and decoding stages are passed to the lower decoder through skip connections for feature interaction, so that the lower decoder can refer to more context information during the reconstruction process. After this stage, the feature map finally generates a real hyperspectral background space.

[0011] 2. Abnormal stripping shielding auxiliary branch module

[0012] In hyperspectral anomaly detection, traditional deep learning methods may lack sufficient ability to effectively distinguish anomalies from background, thereby reconstructing anomaly targets during background reconstruction. In order to enhance the difference between background and anomaly features and make the model network avoid focusing on anomalies as much as possible, the present invention proposes an anomaly stripping shielding auxiliary branch module. This branch aims to strip the anomaly part from the reconstruction process through constraints, thereby promoting the model to better focus on the learning and reconstruction of background features. This branch module mainly includes the following:

[0013] 1) Background anomaly separation: The previous stage of the entire branch uses a convolutional autoencoder to capture features to complete background reconstruction, providing the model with an information flow of global features of the background. The encoding block consists of a 3×3 convolution (stride=2) and a 1×1 convolution (BN and LeakReLU). The input of each encoding block can be expressed as:

[0014] E i =Conv 3×3 ((Conv 1×1 (E i-1 ))

[0015] Among them, 0<i<5. The decoding block consists of an upsampling layer and a 1×1 convolution (BN and LeakReLU), combined with a skip connection to retain shallow features. The input of the decoding process can be expressed as:

[0016] D i =Concat[E i-1 , Conv 1×1 (Upsample(D i-1 ))]

[0017] The independent path of the convolutional autoencoder provides a new reconstruction perspective for the model and globally separates the background and anomalies. With the iteration of training, the anomalies are progressively stripped away to obtain a sparse matrix of anomaly targets from coarse to fine. The calculation method of the entire residual is as follows:

[0018]

[0019] Where (·) i，j，： is the spectrum vector at position (i, j).

[0020] 2) Abnormal feature shielding

[0021] After obtaining the separated feature maps, we further introduced an abnormal shielding mechanism to eliminate the feature space mapping of abnormal target information. In the abnormal shielding stage, we divided it into three parts to achieve hierarchical suppression of abnormalities. First, we used a parallel structure of two sets of large convolution kernels, a single 5×5 convolution kernel and a cascade of a 5×5 convolution kernel and a 7×7 convolution kernel. The large convolution kernel provides a wider receptive field and focuses on a larger background area, thereby weakening the abnormal information and dispersing it into the feature map S3.

[0022] S3=Concat[DwConv 5×5 (S),DwConv 7×7 (DwConv 5×5 (S))]

[0023] Then, the obtained feature map is passed to a combination of average pooling (MEP) and maximum pooling (MAP) for processing. In this process, average pooling helps the model focus on the globally consistent background, while maximum pooling selects local background information, which has the effect of guiding the abnormal target to no longer be the focus of global information. Then, a 1×1 convolution and activation function are performed to obtain the background selection W1 and W2.

[0024] In order to further optimize the suppression effect of abnormal features, we introduced a channel attention mechanism to dynamically adjust the weights of each channel, especially to impose lower weights on channels containing anomalies. Finally, the channel attention output is element-wise multiplied with the feature information of the parallel convolution structure, thereby achieving suppression and optimization of abnormal features in multiple dimensions. In this way, the abnormal information is effectively weakened at the channel, space, and scale levels, and the output feature map is more focused on background information, while supplementing more abundant background information for abnormal suppression in the subsequent decoding process.

[0025] 3. Background content cross-guides the Transformer module

[0026] The Transformer model has been applied to hyperspectral anomaly detection because of its excellent global feature modeling capabilities. Although it performs well in acquiring overall data information, it still has limitations such as insensitivity to capturing local details and insufficient robustness to complex backgrounds. In order to address the impact of these problems and enhance the model's ability to distinguish between background and abnormal targets during training, a novel Transformer module, namely the background content cross-guided Transformer module, is proposed.

[0027] First, the module applies the conditional position embedding to the lower decoder input through a 3×3 deep convolutional layer. Then, the entire module is divided into a decoding multi-head self-attention branch and an encoding multi-head self-attention branch. The input lower layer decoded information X1 and encoded information X2 data are linearly transformed to obtain the query, key, and value respectively:

[0028]

[0029] in, Represents the projection weights of query, key and value respectively

[0030] After the mapping is completed, the decoding multi-head attention branch performs background smoothing recovery by capturing the deep features of the spectral dimension of the decoded information. The obtained Q1, K1, and V1 are used with non-overlapping windows (using a size of M×M) to obtain Subsequently, they are divided into multiple heads along the channel dimension.

[0031] In the encoding multi-head attention branch, we focus on fine-grained background reconstruction of the encoded information. Before non-overlapping window division, K2 and V2 are respectively subjected to 3×3 deep convolution layers, which can be obtained.

[0032] The attention output from the decoding branch is used as the query Q and the key k after window partitioning to perform cross-guided attention calculation with the value of the encoding branch attention, and then permutation and linear layers are performed to integrate multiple heads. The final attention is as follows:

[0033]

[0034] Finally, Att Out Passed through a feedforward network, CGA jointly processes the inputs of the encoder and decoder, utilizing the information interaction between the encoder and decoder to enhance background confidence and weaken abnormal local spatial dependencies.

[0035] 4. Collaborative Optimization Loss Module

[0036] During the training process, the goal of the overall network learning is to make the background pixels in the generated image consistent with the background pixels in the original hyperspectral image as much as possible, while effectively suppressing the reconstruction of abnormal areas. To this end, this method adopts a background priority mapping method, giving background pixels a higher confidence and reducing the weight of abnormal target pixels, thereby reducing the impact of abnormal areas on the loss function and achieving the effect of suppressing abnormal targets. In order to evaluate the reconstruction effect, MSE loss and SmoothL1 loss are combined. MSE loss mainly calculates the pixel difference between the original hyperspectral image and the reconstructed image, focusing on subtle pixel-level errors, while SmoothL1 loss introduces structural smoothness constraints to ensure accurate reconstruction of the background area.

[0037] In actual image reconstruction tasks, key structural information and abnormal features are usually concentrated in high-frequency parts. Traditional pixel-level losses (such as MSE and SmoothL1) ignore these important information in the frequency domain, making it difficult for the model to distinguish between background and abnormal areas. Therefore, a new frequency domain loss function is introduced, which converts the image to the frequency domain and compares the frequency distribution difference between the reconstructed image and the original image, thereby enhancing the model's sensitivity to high-frequency abnormal information. The final frequency domain loss function calculates the difference in the frequency domain and assigns different weights according to the frequency, thereby reducing the impact of abnormal areas and better focusing on the reconstruction of background areas. The overall loss function is shown below:

[0038]

[0039] In short, our strategy is to help the model pay more attention to the background area by introducing the frequency domain loss function, while suppressing the interference of abnormal areas on the quality of reconstructed background. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Schematic diagram of the invented dual decoder dual Transformer background enhancement reconstruction network hyperspectral anomaly detection model;

[0041] Figure 2 A schematic diagram of a dual decoder in the invented model;

[0042] Figure 3 It is a schematic diagram of the abnormal stripping shielding auxiliary branch of the invention;

[0043] Figure 4 A schematic diagram of a cross-guided Transformer module for the background content of the invention;

[0044] Figure 5 Schematic diagram of background feature fusion;

[0045] Figure 6 The data sets used for the invention and the corresponding real surface locations;

[0046] Figure 7 Visualization results of anomaly detection of the method of the present invention and the comparative method in various typical scenarios of the data set;

[0047] Figure 8 The ROC curves of the method of the present invention and the comparative method in various typical scenarios of the data set are shown; DETAILED DESCRIPTION

[0048] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0049] Step 1: Input the preprocessed data and the generated uniform noise into the encoder for feature extraction;

[0050] Step 2: Pass the background features processed by the abnormal stripping and masking module into the lower layer decoder;

[0051] Step 3: Input the encoder and lower-layer background decoder features into the background content guided Transformer module;

[0052] Step 4: Constrain the generated error using the collaborative optimization loss function;

[0053] Step 5: The information processed by the upper and lower decoders is improved using the background feature fusion module to obtain the final prediction map;

[0054] Further, the step 1 specifically comprises the following steps:

[0055] Step 1-1: Normalize the hyperspectral remote sensing image dataset used to facilitate the model to process the data and improve the computational efficiency. At the same time, construct uniform noise to prepare for background feature modeling.

[0056] Step 1-2: If Figure 1 The figure shows a schematic diagram of the invented dual decoder dual Transformer background enhancement reconstruction network hyperspectral remote sensing image anomaly detection model. The processed image data is input into the encoder, and the number of channels is changed from C to N through a 3×3 convolution layer. Then, the feature map is subjected to two Swin Transformer modules, a 1×1 convolution layer, and a downsampling layer for shallow feature extraction. The downsampling layer is a 2×2 stride convolution layer, which halves the spatial resolution of the feature map and doubles the channels. The sizes of feature maps of different nodes are changed to Then use the skip connection to feed the encoded features into the upper decoder to avoid information loss. The double decoding structure is as follows Figure 2 At the same time, we input the noise into the decoder of the lower layer for feature extraction.

[0057] Further, the step 2 specifically includes the following steps:

[0058] Step 2-1: If Figure 3 As shown in the figure, the original hyperspectral data is passed as input to the entire anomaly stripping and shielding auxiliary branch. The previous stage of the entire branch uses a convolutional autoencoder to capture features to complete background reconstruction, providing the model with an information flow of global features of the background. The encoding block consists of a 3×3 convolution (stride=2) and a 1×1 convolution, layer normalization, and activation function. The corresponding decoding block consists of a 3×3 deconvolution layer and a 1×1 convolution, layer normalization, and activation function. The generated anomaly residual can be expressed as:

[0059]

[0060] Step 2-2: After processing the feature map and obtaining its separated representation, the receptive field can be significantly expanded through a large convolution kernel to eliminate the adverse effects of abnormal target information in the feature space and obtain a preliminary abnormal masked feature map.

[0061] S3=Concat[DwConv 5×5 (S),DwConv 7×7 (DwConv 5×5 (S))]

[0062] Specifically, in the anomaly shielding stage, we gradually suppress the abnormal targets in a hierarchical manner, thereby ensuring that the model can focus on learning and reasoning about normal patterns. Next, we input the obtained feature map into a processing module that combines average pooling (MEP) and maximum pooling (MAP). Through this combined pooling operation, the abnormal information no longer dominates the global features, but is effectively suppressed, making the background a more prominent focus.

[0063] [W1, W2] = Sigmoid(Conv 1×1 [MAE(S3),MAX(S3)])

[0064] Subsequently, we use channel attention to select background W1 and W2 in order to fine-tune and select background information in subsequent steps. Finally, the channel attention output is element-wise multiplied with the feature information of the parallel convolution structure, thereby achieving the suppression and optimization of abnormal features in multiple dimensions and obtaining supplementary information. This operation helps to further optimize the feature map, ensuring that the model can more accurately identify and process background areas, thereby improving the overall reconstruction and recognition effects.

[0065] Further, the step 3 specifically includes the following steps:

[0066] Step 3-1: If Figure 4 As shown, we transform the feature map X of the lower decoder of the image i and additional information Add them together, and then use the background content to cross-guide the Transformer to adaptively learn the background content. The decoding multi-head attention branch recovers the background smoothness by capturing the deep features of the spectral dimension of the decoded information. The obtained Q1, K1, and V1 are obtained using a non-overlapping window (using a size of M×M) Then, they are divided into multiple heads along the channel dimension as follows:

[0067]

[0068] Where h represents the number of heads. In this process, different attention heads can focus on the correlation between different bands, thereby performing a detailed background representation in the spectral dimension. Then, a multi-head self-attention mechanism is calculated for each head, ensuring that the model can more accurately describe the local details in the background.

[0069]

[0070] Step 3-2: Calculate using the same method In order to give greater weight to the decoupled background information, the attention from the decoding branch is then cross-guided for attention calculation, and the multi-head attention is integrated by rearrangement to finally obtain the synthetic attention output.

[0071]

[0072] The final output attention result will focus on the background features, weakening the impact of anomalies and preparing for the subsequent background feature fusion.

[0073] Further, the step 4 specifically comprises the following steps:

[0074] Step 4-1: For the generated background image, we introduce the frequency domain loss function during the training process to model the frequency domain information of the image. Because abnormal targets often have more obvious high-frequency components. By applying a smaller weight suppression to the high-frequency components, the model can focus more on the reconstruction of the background area, as shown below.

[0075]

[0076] in ω is the weight coefficient, which is used to impose constraints in the frequency domain.

[0077] Step 4-2: In order to further achieve effective suppression of abnormal targets and ensure accurate reconstruction of the background area, we combined three loss functions. Ultimately, the optimization goal is to minimize the sum of these loss functions so that the generated image is closer to the real hyperspectral image and maintains the consistency of the background area.

[0078]

[0079] in represents the mean square error function, represents the smooth loss function, represents the frequency domain loss function.

[0080] Further, the step 5 specifically includes the following steps:

[0081] Step 5-1: If Figure 5 As shown in Figure 1, in this background feature fusion module, the output from the lower decoder and the output from the upper decoder are first connected along the channel dimension. The fused feature map is then processed by the spatial attention mechanism to enhance the information interaction between different spaces and highlight the key information. The output processed by the deep attention is then added to the spatial attention output result, and after deep attention refinement, it is element-wise multiplied with the feature map of the spatial attention to enhance important background information while suppressing abnormal interference, generating the final clean background.

[0082] Step 5-2: Use the obtained background image and the original observation image to calculate the reconstruction error between the two to obtain the final detection result:

[0083]

[0084] Finally, the obtained detection map is verified with the real surface value.

[0085] The Pavia dataset used in this experiment comes from an airborne sensor called Reflection Optical System Imaging Spectrometer (ROSIS) with a spatial resolution of 1.3m and a wavelength range of 430-860nm. The image is 100×100 in size, has 102 spectral bands, and contains 71 anomalies. The San Diego dataset was captured by the Airborne Visible / Infrared Imaging Spectrometer (AVIRIS) sensor in San Diego, California, USA, with a spatial resolution of 3.5m and a wavelength range of 370-2510nm. The Texas Coast datasets were captured by the AVIRIS sensor on the coast of Texas, USA in 2010. Both scenes have a spatial resolution of 17.2m and 100×100 pixels. The first scene contains 67 anomalous pixels and 204 spectral bands with a wavelength range of 450-1350nm. Buildings of different sizes are considered anomalies. The second scene has 207 bands, and all bands were used in each experiment. The two HSIs are called TC-1 and TC-2. The Gainesville dataset represents an urban scene captured by the AVIRIS airborne sensor in Gainesville, Florida, USA. By removing the water absorption band and the signal-to-noise ratio band from the original data, 191 bands were retained for the experiment. The Salinas dataset is a background sub-image selected from the Salinas scene. The data and the ground truth locations of the anomaly targets are shown in Figure 6 .

[0086] Among the data-driven methods, there are ten representative anomaly detection methods for comparison, including GRX, CRD, FrFE, RCRD, PCA-TLRSR, LEBSR, GAED, Auto-AD, PDBSNet and MSNet. GRX is a probability distribution-based method; CRD, RCRD and PCA-TLRSR are representation-based methods; FrFE uses frequency domain transformation to detect anomalies; LEBSR detects anomalies by learning background end members; GAED, Auto-AD, PDBSNet and MSNet reconstruct the background in a self-supervised manner. Ten comparison methods are used to compare with the present invention, and the visualization results and detection results are shown in Figure 2. Figure 7 and Figure 8 As shown, the method of the present invention uses the area under the curve (AUC) to quantitatively evaluate and analyze the method of the present invention and several other anomaly detection methods. The AUC value ranges from 0 to 1. The closer the value is to 1, the better the performance of the detection method.

[0087] Table 1 Comparison of AUC values ​​for anomaly detection in hyperspectral remote sensing images

[0088]

[0089] The results are shown in Table 1. From the experimental results, the detection performance of this method is obviously better than that of the comparison algorithm, which fully verifies the superiority of this method.

[0090] The visualization results of the six data sets in typical scenarios are shown in the figure below. Figure 7 and Figure 8 As shown. Although the CRD and RCRD methods present a relatively complete background, this makes it difficult to distinguish anomalies from the background. FrFE and PCA-TLRSR also produce many erroneous anomaly results during the detection process. Although LEBSR and Auto-AD are able to detect abnormal targets on all datasets, the background is still quite obvious. In contrast, GAED, PDBSNet and MSNet methods can detect anomalies better. Compared with other methods, the present invention not only successfully detects abnormal targets, but also effectively separates anomalies from the background, significantly improves the detection accuracy, and achieves good background suppression.

[0091] In order to further qualitatively analyze the detection performance of different methods, the ROC curves of different methods are drawn, such as Figure 8 As shown. For better visualization, they are displayed in logarithmic scale. In general, the present invention outperforms other methods in ROC curve performance, which significantly demonstrates the performance advantage of the present invention method.

Claims

1. A method for abnormal target detection in hyperspectral remote sensing images based on a dual-decoder dual-Transformer background enhancement reconstruction network, characterized in that: The following steps are involved: S1. Construct a dual decoder for feature refinement: Select the observed hyperspectral remote sensing image as the original data, and generate uniform noise image data that strictly obeys the uniform distribution to ensure that the probability of each pixel value is equal, and perform data normalization on the original number. Use the encoder to extract features from the image. The image encoder and the upper-layer decoding feature extractor use the Swin Transformer module as a feature extractor to extract high-level semantic information. The lower-layer decoder uses the background content cross-guided Transformer module as a feature extractor to obtain a reconstructed pure background; S2. Construct anomaly stripping and shielding branch: An additional auxiliary branch for anomaly separation is designed to extract anomaly features and use them as supplementary input for the fused feature maps at each decoding stage, thereby improving the anomaly suppression ability of the model; S3, constructing a background content cross-guided Transformer module: Its feature is that the background content cross-guided Transformer module is embedded in the lower decoder, which guides attention to the background semantic information and enhances the model's ability to capture and learn complex abnormal patterns; S4. Construct a collaborative optimization loss function: Its characteristic is to combine frequency domain loss, MSE loss and SmoothL1 loss, punish abnormal targets from multiple dimensions, and strengthen the ability to suppress abnormal targets. The loss function network model is used for optimization and adjustment throughout the training process to improve the robustness of the model and training stability. S5. The features output by the encoder and decoder are fused using the background feature fusion module to restore the reconstructed background image to the original spatial resolution and number of channels, thereby obtaining an accurate background image. The original hyperspectral remote sensing image is then combined to calculate the reconstruction error to obtain the final detection image.

2. The method for abnormal target detection in hyperspectral remote sensing images based on dual decoder and dual Transformer background enhancement reconstruction network according to claim 1 is characterized in that: The step S1 comprises the following steps: S1.1 normalizes the hyperspectral remote sensing image dataset used to improve computational efficiency and constructs uniform noise for background image modeling. S1.2 inputs the processed image data into the encoder and adjusts the number of channels from C to N through a 3×3 convolution layer. Then, the feature map passes through two Swin Transformer modules, a 1×1 convolution layer, and a downsampling layer for shallow feature extraction. The downsampling layer is a 2×2 stride convolution layer, which halves the spatial resolution of the feature map and doubles the channels. The sizes of feature maps of different nodes are adjusted to E i，j Here, i and j represent the number of encoders and the number of internal feature extractor layers, respectively. Then, the encoded features are fed into the upper-layer decoder using skip connections to avoid information loss. The double decoding structure is shown in Figure 2. At the same time, the noise is input into the lower-layer decoder for feature extraction.

3. The method for abnormal target detection in hyperspectral remote sensing images based on dual decoder and dual Transformer background enhancement reconstruction network according to claim 1 is characterized in that: The step S2 comprises the following steps: S2.1: Take the original hyperspectral remote sensing image data as input and pass it into the entire anomaly stripping and shielding auxiliary branch network. The previous stage of the entire branch uses a convolutional autoencoder to capture features to complete background reconstruction, providing the model with an information flow of global features of the background. The encoding block consists of a 3×3 convolution (stride=2) and a 1×1 convolution, layer normalization, and activation function. The corresponding decoding block consists of a 3×3 deconvolution layer and a 1×1 convolution, layer normalization, and activation function. The generated anomaly residual can be expressed as: After processing the feature map and obtaining its separated representation, S2.2 uses a large convolution kernel to expand the receptive field, eliminate the adverse effects of abnormal target information in the feature space, and obtain a preliminary feature map where the abnormal target is shielded. S3=Concat[DwConv 5×5 (S),DwConv 7×7 (DwConv 5×5 (S))] Specifically, in the anomaly shielding stage, the abnormal targets are gradually suppressed in a hierarchical manner to ensure that the model can focus on the learning and reasoning of the background. Next, the obtained feature map is input into a processing module that combines average pooling (MEP) and maximum pooling (MAP). Through this combined pooling operation, the abnormal target information can be effectively suppressed, promoting the effective reconstruction of the background image. [W1,W2]=Sigmoid(Conv 1×1 [MAE(S3),MAX(S3)]) S2.3 Then, the background selection features W1 and W2 are finely regulated using channel attention. Finally, the channel attention output is element-wise multiplied with the feature information of the parallel convolution structure, thereby achieving the suppression and optimization of abnormal features in multiple dimensions and obtaining supplementary information.

4. The method for abnormal target detection in hyperspectral remote sensing images based on dual decoder and dual Transformer background enhancement reconstruction network according to claim 1 is characterized in that: The step S3 comprises the following steps: S3.1 The feature map X of the lower layer decoder of the image i and additional information Add them together, and then use the background content to cross-guide the Transformer to adaptively learn the background content. The decoding multi-head attention branch recovers the background smoothness by capturing the deep features of the spectral dimension of the decoded information. The obtained Q1, K1, and V1 are obtained using a non-overlapping window (using a size of M×M) Then, they are divided into multiple heads along the channel dimension as follows: Where h represents the number of heads. In this process, different attention heads can focus on the correlation between different bands, thereby performing a detailed background representation in the spectral dimension. Then, a multi-head self-attention mechanism is calculated for each head, ensuring that the model can more accurately describe the local details in the background. S3.2 Calculate the encoded multi-head self-attention using the same method In order to give greater weight to the decoupled background information, the attention from the decoding branch is then cross-guided for attention calculation, and the multi-head attention is integrated by rearrangement to finally obtain the synthetic attention output. The final output attention Att Out Focusing on background features will weaken the impact of abnormal targets and enhance the accuracy of subsequent background feature fusion.

5. The method for abnormal target detection in hyperspectral remote sensing images based on dual decoder and dual Transformer background enhancement reconstruction network according to claim 1 is characterized in that: The step S4 comprises the following steps: S4.1 For the generated background image, the frequency domain information of the image is modeled by introducing the frequency domain loss function during the training process. Because abnormal targets often have more obvious high-frequency components. By applying a smaller weight suppression to the high-frequency components, the model can focus more on the reconstruction of the background area, as shown below. Where ω is the weight coefficient, which is used to impose anomaly constraints in the frequency domain. S4.2 To further achieve effective suppression of abnormal targets while ensuring the reconstruction accuracy of the background area, this method combines three loss functions. Frequency domain loss helps the model capture abnormal features in the frequency domain and achieves abnormal suppression from the frequency domain; MSE loss enhances the pixel-level reconstruction quality of the background area; SmoothL1 loss smoothes and optimizes small errors, reduces the impact of errors, and improves the robustness of the model. Ultimately, the optimization goal is to minimize the sum of these loss functions so that the generated image is closer to the real hyperspectral image and maintains the consistency of the background area. in represents the mean square error function, represents the smooth loss function, represents the frequency domain loss function.

6. The method for abnormal target detection in hyperspectral remote sensing images based on dual decoder and dual Transformer background enhancement reconstruction network according to claim 1 is characterized in that: The step S5 comprises the following steps: The output from the lower decoder and the output from the upper decoder are passed to the background feature fusion module. First, they are concatenated along the channel dimension, and then the fused feature map is processed by the spatial attention mechanism to enhance the information interaction between different spaces. Then, the output after deep attention processing is added to the spatial attention output result, and element-wise multiplied with the feature map of spatial attention to restore it to the original dimension to generate the final clean background. The obtained background is used to calculate the reconstruction error to obtain the final detection result: Finally, the obtained detection map is verified with the real surface value.

Images