A method, system, device, and medium for classifying polarization hyperspectral camouflage targets based on polarization-guided multi-scale dynamic attention.

Abstract

A polarization-guided multi-scale dynamic attention-based method, system, device, and medium for classifying polarization-induced hyperspectral camouflage targets are disclosed. The method includes: acquiring and preprocessing polarization-induced hyperspectral data to obtain the sine and cosine components of the polarization angle AoLP, and constructing input tensors and ; constructing a parallel dual-stream network architecture of spectral and polarization flows, using and as inputs respectively, to extract spectral features and initial polarization features; designing a multi-scale dynamic AoLP attention module to convert the sine and cosine components of the polarization angle AoLP into dynamic attention weights, and physically enhancing the initial polarization features to obtain enhanced polarization features; constructing a polarization-guided space-channel attention fusion module, using the enhanced polarization features to perform dual adaptive calibration of the spectral features in both spatial and channel dimensions to obtain fused features; and then mapping these features to a category space to complete pixel-level target classification. This invention offers high classification accuracy, strong target boundary recognition capability, and good adaptability to complex backgrounds.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing imaging and image processing technology, and specifically relates to a polarization hyperspectral camouflage target classification method, system, device and medium based on polarization-guided multi-scale dynamic attention. Background Technology

[0002] Hyperspectral imaging technology, with its high resolution across a continuous narrow band, can accurately capture the spectral properties of materials. However, in complex backgrounds, camouflaged targets exhibit highly similar spectral characteristics to their environment, making accurate identification difficult by relying solely on hyperspectral information. Polarization hyperspectral imaging leverages the physical dimension of polarization imaging to complement and enhance spectral characterization. However, its information structure is complex and its modal differences are significant. Maintaining spectral discriminative power while leveraging polarization priors to achieve effective fusion has become a key challenge in polarization hyperspectral target identification.

[0003] Most existing polarization hyperspectral data fusion methods rely on linear modeling or static rules, which makes it difficult to fully explore the complex nonlinear complementary relationship between polarization and hyperspectral heterogeneous data. Moreover, polarization and hyperspectral data have significant differences in physical properties and statistical distribution. Direct fusion or shallow unified modeling can easily introduce intermodal interference, weakening the effective expression of their respective advantages. In addition, the physical geometric features contained in polarization data have not been effectively incorporated into the deep feature extraction and cross-modal fusion process, making it difficult to fully realize the potential value of polarization information in deep networks.

[0004] Patent application CN121415161A discloses an end-to-end polarization hyperspectral image classification method and system. It extracts polarization hyperspectral spatial-spectral features through 3D-CNN and uses trainable weights to perform weighted summation and fusion of the probability distributions of different spectral bands. However, due to the difficulty of its single-stream network in effectively handling the modal heterogeneity of spectral and polarization features, it leads to problems such as insufficient utilization of polarization information, large intermodal interference, and insufficient target boundary recognition ability in complex backgrounds.

[0005] Patent application CN121740763A discloses a polarization hyperspectral water quality monitoring system and method. It acquires water body image signals through a polarization hyperspectral imager and estimates water quality parameters by combining them with an inversion model. However, this method lacks a deep feature extraction network construction and a calibration and fusion mechanism between modes for automatic classification of complex land features, which results in problems such as low polarization information participation, insufficient nonlinear feature expression ability, and difficulty in being applied to target recognition tasks in complex scenes. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the present invention aims to provide a polarization-guided multi-scale dynamic attention-based method, system, device, and medium for classifying polarization hyperspectral camouflage targets. By constructing a dual-stream parallel network architecture of spectral and polarization flows, a multi-scale dynamic AoLP attention module based on the continuous sine and cosine representation of polarization angle AoLP is introduced. Furthermore, spatial attention weights and channel attention weights generated by enhanced polarization features are used to perform dual adaptive calibration of spectral features. This enables deep collaborative fusion of polarization physical structure information and hyperspectral spectral discrimination information, suppresses feature interference caused by direct fusion of heterogeneous modes, and enhances the model's ability to perceive overlapping regions of camouflage targets against spectrally similar backgrounds. The invention exhibits high classification accuracy, strong target boundary recognition capability, and good adaptability to complex backgrounds.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A polarization-guided multi-scale dynamic attention-based method for classifying polarization-induced hyperspectral camouflage targets includes the following steps: Step 1: Acquire polarization hyperspectral data, perform radiometric correction, normalization, and polarization angle AoLP decomposition on the polarization hyperspectral data to obtain the sinusoidal component of the polarization angle AoLP. Sum and cosine components And construct the spectral flow input tensor and polarization flow input tensor ; Step 2: Construct a parallel network architecture for spectral and polarization currents, respectively using... and Using this as input, spectral features and polarization features are extracted independently to obtain spectral features. With initial polarization characteristics ; Step 3: Design a multi-scale dynamic AoLP attention module to process the polarization angle AoLP sinusoidal component. With cosine component Transformed into dynamic attention weights, for the initial polarization features Physically driven enhancement is performed to obtain enhanced polarization characteristics. ; Step 4: Construct a polarization-guided spatial-channel attention fusion module, utilizing the enhanced polarization features. Regarding the spectral features Dual adaptive calibration in both spatial and channel dimensions is performed to obtain fused features. ; Step 5: Merge the features Mapping to the category space completes pixel-level target classification.

[0008] The specific method of step 1 includes: Hyperspectral images were acquired at four angles (0°, 45°, 90°, and 135°) using a polarization hyperspectral imager. Each hyperspectral data point has dimensions H×W×B, where H, W, and B represent the height, width, and number of bands, respectively. Stokes parameters were calculated from the hyperspectral data after radiometric correction. , , Linear polarization degree DoLP and polarization angle AoLP; and Normalized to relative Stokes parameters and : (2) The polarization angle AoLP is decomposed into sinusoidal components. With cosine component : (3) Construct two input tensors, one of which is the spectral flow input tensor. Includes hyperspectral reflectance data The second is the polarization flow input tensor. ,Include Five polarization channels.

[0009] The specific method for step 2 includes: Construct a parallel network architecture for spectral and polarization flow. A spatial neighborhood block of size P×P is cropped centered on the target pixel and used as input. The spectral flow is input as a tensor of the spectral flow. For the input, for the... k pixels, its input tensor is It uses a 3D convolutional network to extract spectral features, preserves the spectral correlation between bands, and mines spatial semantic information to output spectral features. ; Polarized flow input tensor For the input, for the... k pixels, its input tensor is Similarly, initial polarization features are extracted using a 3D convolutional network. .

[0010] The specific method for step 3 includes: The sinusoidal component of the polarization angle AoLP obtained in step 1 With cosine component By aggregating and splicing along the spectral dimension, the physical characteristics of AoLP are obtained. Employing multi-scale convolutional branch pairs Features at different scales are extracted and then concatenated to obtain multi-scale AoLP features. Multi-scale AoLP features Compared with the initial polarization characteristics obtained in step 2 In channel-dimensional concatenation, a mapping function consisting of 1×1 convolutions and nonlinear activations is used. Dynamically generate attention weight graph : (4) Finally, the attention weight map is used to evaluate the initial polarization features. Element-wise weighted enhancement is performed to obtain enhanced polarization features. : (5) Where, σ( ) is the Sigmoid function, and ⊙ represents element-wise multiplication.

[0011] The specific method for step 4 includes: The enhanced polarization features obtained in step 3 Max pooling and average pooling are performed along the channel dimension to obtain two 2D spatial response maps. These two 2D spatial response maps are then concatenated along the channel dimension, followed by a 7×7 convolution and a sigmoid activation function to generate spatial attention weights. ; The enhanced polarization features obtained in step 3 Global average pooling is performed to obtain channel description vectors, and channel attention weights are generated using a multilayer perceptron and a sigmoid function. ; Using spatial attention weights and channel attention weights Spectral features obtained in step 2 The final fused features are obtained by performing dual adaptive adjustment. : (5).

[0012] The specific method for step 5 includes: The fusion features obtained in step 4 Global average pooling is performed to obtain feature vectors, which are then input into a fully connected layer and mapped to the number of classes. The softmax function outputs pixel-level classification probabilities, and the classification result of the target pixel is determined based on the maximum probability value among the pixel-level classification probabilities.

[0013] This invention also provides a polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification system, comprising: The polarization hyperspectral data acquisition and preprocessing module is used to acquire polarization hyperspectral data, perform radiometric correction, normalization, and polarization angle AoLP decomposition on the polarization hyperspectral data, and obtain the sinusoidal component of the polarization angle AoLP. Sum and cosine components And construct the spectral flow input tensor and polarization flow input tensor ; The feature extraction module employs a dual-stream parallel network architecture of spectral flow and polarization flow, respectively using... and Using this as input, spectral features and polarization features are extracted independently to obtain spectral features. With initial polarization characteristics ; The physics-driven enhancement module, based on the multi-scale dynamic AoLP attention module, integrates the polarization angle AoLP sinusoidal component. With cosine component Transformed into dynamic attention weights, for the initial polarization features Physically driven enhancement is performed to obtain enhanced polarization characteristics. ; The dual adaptive calibration module, based on the polarization-guided spatial-channel attention fusion module, utilizes the enhanced polarization features. Regarding the spectral features Dual adaptive calibration in both spatial and channel dimensions is performed to obtain fused features. ; The target classification module is used to classify the fused features. Mapping to the category space completes pixel-level target classification.

[0014] This invention also provides a polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification device, comprising: Memory: A computer program that stores the above-mentioned polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method, and is a computer-readable device; Processor: Used to implement the polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method when executing the computer program.

[0015] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, enables the implementation of the aforementioned polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention reduces the mutual interference between heterogeneous modes and improves the utilization efficiency of polarization information: The present invention constructs a parallel network structure of spectral flow and polarization flow in step 2 to independently model spectral features and polarization features, avoiding the information interference problem introduced by directly splicing or uniformly processing multimodal data in the prior art, thereby improving the utilization efficiency of polarization information.

[0017] (2) This invention enhances the guiding role of polarization physics priors on deep features and improves the ability to express nonlinear features: This invention uses the multi-scale dynamic AoLP attention module constructed in step 3 to transform the surface geometric structure information represented by the polarization angle into spatial attention weights. Through multi-scale feature extraction and nonlinear mapping, it enhances the modeling ability of complex nonlinear complementary relationships between polarization and hyperspectral heterogeneous data, thereby improving the model's ability to perceive the boundary region and aliasing region of the camouflaged target.

[0018] (3) The present invention realizes deep synergistic and complementary fusion of cross-modal information: The present invention uses the polarization-guided spatial-channel attention fusion module constructed in step 4 to guide the spectral features in both spatial and channel aspects using polarization features, thereby realizing deep synergistic and complementary fusion of polarization physical structure information and hyperspectral spectral discrimination information, improving the model's ability to comprehensively utilize multi-dimensional information in complex scenarios, thereby improving classification accuracy and robustness. Attached Figure Description

[0019] Figure 1 This is a diagram illustrating the overall architecture of the polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method of this invention.

[0020] Figure 2 This is a diagram of the architecture of the multi-scale dynamic AoLP attention module of the present invention.

[0021] Figure 3 This is a diagram of the polarization-guided spatial-channel attention fusion module architecture of the present invention. Detailed Implementation

[0022] The invention will be further described in detail below with reference to the accompanying drawings and detailed model construction principles.

[0023] This embodiment uses a polarization-guided multi-scale dynamic attention method for classifying polarization-guided hyperspectral camouflage targets as an example to illustrate a method for classifying polarization-guided hyperspectral camouflage targets. The overall architecture of this method is as follows: Figure 1 As shown. Includes the following steps: Step 1: Acquire polarization hyperspectral data, perform radiometric correction, normalization, and polarization angle AoLP decomposition on the polarization hyperspectral data to obtain the sinusoidal component of the polarization angle AoLP. Sum and cosine components And construct the spectral flow input tensor and polarization flow input tensor ; First, the raw polarization hyperspectral data is preprocessed. Hyperspectral images of the same scene at 0°, 45°, 90°, and 135° polarization directions are acquired using a polarization hyperspectral imaging device. After radiometric correction and registration, polarization-related feature parameters, including Stokes parameters, degree of linear polarization (DoLP), and angle of linear polarization (AoLP), are calculated according to the method described in step 1. To eliminate the interference of intensity information on polarization characterization, the Stokes parameters are normalized to obtain relative Stokes parameters. Considering the periodicity of the polarization angle, to avoid discontinuities caused by the periodicity of the AoLP angle, the polarization angle is further decomposed into corresponding sine and cosine components. Then, two types of network input tensors are constructed: one is the spectral flow input tensor. Includes hyperspectral reflectance data The first is used to characterize the spectral reflectance information of the target. The second is the polarization flow input tensor. ,Include Five polarization channels are used to characterize the polarization features of the target.

[0024] Step 2: Construct a parallel network architecture for spectral and polarization currents, respectively using... and Using this as input, spectral features and polarization features are extracted independently to obtain spectral features. With initial polarization characteristics ; In this embodiment, a local neighborhood block centered on the target pixel is used as the network input. For the pixel to be classified, a spatial neighborhood block of size P×P is cropped centered on that pixel to simultaneously preserve pixel-level spectral information and local spatial context information. When P is 11 and the number of bands B is 209, the input tensor dimension of each sample in the spectral flow branch is (1, 209, 11, 11), and for the polarization flow branch, the input tensor dimension of each sample is (5, 209, 11, 11). Secondly, a dual-flow parallel network architecture of spectral and polarization flows is constructed to reduce feature interference caused by direct mixing of different modalities in modeling. Specifically, the spectral flow uses the spectral flow input tensor... As input, a 3D convolutional network is used to extract features from the input samples, preserving inter-band correlations while extracting local spatial semantic information, and outputting spectral features. Polarization flow is used as the input tensor. As input, polarization features are also extracted using a 3D convolutional network to obtain the initial polarization features of the polarization branch. Through this dual-stream independent modeling approach, the discriminative features of the spectral mode and polarization mode can be learned separately, providing a stable feature foundation for subsequent cross-modal fusion.

[0025] Step 3: Design a multi-scale dynamic AoLP attention module to process the sinusoidal component of the polarization angle AoLP obtained in Step 1. With cosine component Transformed into dynamic attention weights, applied to the initial polarization features obtained in step 2. Physically driven enhancement is performed to obtain enhanced polarization characteristics. ; A multi-scale dynamic AoLP attention module is constructed at the end of the polarization flow, and its architecture diagram is shown below. Figure 2 As shown, this module transforms the surface geometry information represented by the polarization angle into dynamic weights that can participate in network optimization, thereby enhancing the polarization features. First, the sine and cosine components of the polarization angle AoLP obtained in step 1 are aggregated and concatenated along the spectral dimension to obtain the AoLP physical features. Then, multi-scale convolutional branches are used to extract features from the AoLP physical features at different scales, and these features are then concatenated and fused to obtain multi-scale AoLP features that balance fine-grained edge structures with a larger spatial correlation. Next, the multi-scale AoLP features are concatenated with the initial polarization features output by the polarization flow along the channel dimension, and a dynamic attention weight map is generated using a mapping function consisting of 1×1 convolution and nonlinear activation. Finally, this dynamic attention weight map is used to perform element-wise weighted enhancement of the initial polarization features to obtain enhanced polarization features.

[0026] Step 4: Construct a polarization-guided spatial-channel attention fusion module, utilizing the enhanced polarization features obtained in Step 3. The spectral characteristics obtained in step 2 Dual adaptive calibration in both spatial and channel dimensions is performed to obtain fused features. ; The enhanced polarization features and the spectral features of the spectral stream output are jointly input into the polarization-guided spatial-channel attention fusion module, such as... Figure 3As shown, this module is used to achieve cross-modal deep collaborative complementarity. First, the enhanced polarization features are subjected to max pooling and average pooling along the channel dimension to obtain two two-dimensional spatial response maps. These two two-dimensional spatial response maps are then concatenated along the channel dimension and further processed through a 7×7 convolution and a sigmoid activation function to generate spatial attention weights. Next, global average pooling is performed on the enhanced polarization features to obtain channel description vectors, and channel attention weights are generated through a multilayer perceptron and a sigmoid function. Then, the spatial attention weights and channel attention weights are used to perform dual adaptive adjustment of the spectral features to obtain the final fused features. Compared with the direct concatenation fusion method, this invention achieves cross-modal fusion through polarization-guided multi-scale dynamic attention, which helps to reduce mutual interference between heterogeneous modes and improve the efficiency of multi-dimensional information utilization.

[0027] Step 5: Combine the fused features obtained in Step 4. Mapping to the category space completes pixel-level target classification.

[0028] Global average pooling is performed on the fused features to obtain the corresponding feature vectors. These feature vectors are then input into a fully connected layer, mapped to the number of classes, and the softmax function outputs pixel-level classification probabilities. The classification result of the target pixel is determined based on the maximum probability value among these pixel-level probabilities. After performing the above processing on all pixels to be classified, the classification result of the camouflage target corresponding to the scene can be obtained.

[0029] To verify the effectiveness of the method described in this invention, comparative experiments were conducted on three sets of self-collected polarization hyperspectral datasets, and overall accuracy (OA), average accuracy (AA), and Kappa coefficient κ were used as evaluation metrics. The training set ratio was set to 10% in the experiments. The comparison methods included the original data classification method and the classification method after using the fusion method of this invention. The original data refers to the data representation input into the classification network after directly concatenating hyperspectral and polarization information, while the fused data refers to the feature representation obtained after data fusion using the method of this invention. The experimental results are shown in Table 1.

[0030] Table 1 Comparative Experimental Results As shown in Table 1, compared with the original data processing method, the method of the present invention improves the overall accuracy, average accuracy, and Kappa coefficient to varying degrees on the three datasets. Specifically, the improvements are 1.23%, 9.25%, and 5.89% for dataset 1, 1.70%, 1.38%, and 4.28% for dataset 2, and 2.45%, 6.25%, and 6.13% for dataset 3. These results demonstrate that the method of the present invention can improve the overall classification accuracy while further enhancing the balance of identification across categories and the consistency of classification results, thereby improving the classification performance of polarization hyperspectral camouflage targets under complex background conditions.

[0031] In summary, this invention constructs a dual-stream parallel feature extraction structure, introduces a multi-scale dynamic AoLP attention mechanism, and utilizes enhanced polarization features to perform spatial and channel-guided fusion of spectral features, thereby achieving efficient collaborative expression of polarization information and hyperspectral information, thus improving the accuracy of camouflaged target classification in complex scenes.

[0032] This invention also provides a polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification system, comprising: The polarization hyperspectral data acquisition and preprocessing module is used to acquire polarization hyperspectral data in step 1, perform radiometric correction, normalization, and polarization angle AoLP decomposition on the polarization hyperspectral data, and obtain the sinusoidal component of the polarization angle AoLP. Sum and cosine components And construct the spectral flow input tensor and polarization flow input tensor ; The feature extraction module is used to implement the dual-stream parallel network architecture of spectral flow and polarization flow in step 2, respectively. and Using this as input, spectral features and polarization features are extracted independently to obtain spectral features. With initial polarization characteristics ; The physics-driven enhancement module is used to implement the multi-scale dynamic AoLP attention module in step 3, and to enhance the polarization angle AoLP sinusoidal component. With cosine component Transformed into dynamic attention weights, for the initial polarization features Physically driven enhancement is performed to obtain enhanced polarization characteristics. ; The dual adaptive calibration module is used to implement the polarization-guided spatial-channel attention fusion module in step 4, utilizing the enhanced polarization features. Regarding the spectral features Dual adaptive calibration in both spatial and channel dimensions is performed to obtain fused features. ; The target classification module is used to implement the fusion feature in step 5. Mapping to the category space completes pixel-level target classification.

[0033] This invention also provides a polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification device, comprising: Memory: A computer program that stores the above-mentioned polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method, and is a computer-readable device; Processor: Used to implement the polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method when executing the computer program.

[0034] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, enables the implementation of the aforementioned polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method.

Claims

1. A polarization-guided multi-scale dynamic attention-based method for classifying polarization-guided hyperspectral camouflage targets, characterized in that, Includes the following steps: Step 1: Acquire polarization hyperspectral data, perform radiometric correction, normalization, and polarization angle AoLP decomposition on the polarization hyperspectral data to obtain the sinusoidal component of the polarization angle AoLP. Sum and cosine components And construct the spectral flow input tensor and polarization flow input tensor ; Step 2: Construct a parallel network architecture for spectral and polarization currents, respectively using... and Using this as input, spectral features and polarization features are extracted independently to obtain spectral features. With initial polarization characteristics ; Step 3: Design a multi-scale dynamic AoLP attention module to process the polarization angle AoLP sinusoidal component. With cosine component Transformed into dynamic attention weights, for the initial polarization features Physically driven enhancement is performed to obtain enhanced polarization characteristics. ; Step 4: Construct a polarization-guided spatial-channel attention fusion module, utilizing the enhanced polarization features. Regarding the spectral features Dual adaptive calibration in both spatial and channel dimensions is performed to obtain fused features. ; Step 5: Merge the features Mapping to the category space completes pixel-level target classification.

2. The polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method according to claim 1, characterized in that, The specific method of step 1 includes: Hyperspectral images were acquired at four angles (0°, 45°, 90°, and 135°) using a polarization hyperspectral imager. Each hyperspectral data point has dimensions H×W×B, where H, W, and B represent the height, width, and number of bands, respectively. Stokes parameters were calculated from the hyperspectral data after radiometric correction. , , Linear polarization degree DoLP and polarization angle AoLP; and Normalized to relative Stokes parameters and : (2) The polarization angle AoLP is decomposed into sinusoidal components. With cosine component : (3) Construct two input tensors, one of which is the spectral flow input tensor. Includes hyperspectral reflectance data The second is the polarization flow input tensor. ,Include Five polarization channels.

3. The polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method according to claim 1, characterized in that, The specific method for step 2 includes: Construct a parallel network architecture for spectral and polarization flow. A spatial neighborhood block of size P×P is cropped centered on the target pixel and used as input. The spectral flow is input as a tensor of the spectral flow. For the input, for the... k pixels, its input tensor is It uses a 3D convolutional network to extract spectral features, preserves the spectral correlation between bands, and mines spatial semantic information to output spectral features. ; Polarized flow input tensor For the input, for the... k pixels, its input tensor is Similarly, initial polarization features are extracted using a 3D convolutional network. .

4. The polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method according to claim 1, characterized in that, The specific method for step 3 includes: The sinusoidal component of the polarization angle AoLP obtained in step 1 With cosine component By aggregating and splicing along the spectral dimension, the physical characteristics of AoLP are obtained. Employing multi-scale convolutional branch pairs Features at different scales are extracted and then concatenated to obtain multi-scale AoLP features. Multi-scale AoLP features Compared with the initial polarization characteristics obtained in step 2 In channel-dimensional concatenation, a mapping function consisting of 1×1 convolutions and nonlinear activations is used. Dynamically generate attention weight graph : (4) Finally, the attention weight map is used to evaluate the initial polarization features. Element-wise weighted enhancement is performed to obtain enhanced polarization features. : (5) Where, σ( ) is the Sigmoid function, and ⊙ represents element-wise multiplication.

5. The polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method according to claim 1, characterized in that, The specific method for step 4 includes: The enhanced polarization features obtained in step 3 Max pooling and average pooling are performed along the channel dimension to obtain two 2D spatial response maps. These two 2D spatial response maps are then concatenated along the channel dimension, followed by a 7×7 convolution and a sigmoid activation function to generate spatial attention weights. ; The enhanced polarization features obtained in step 3 Global average pooling is performed to obtain channel description vectors, and channel attention weights are generated using a multilayer perceptron and a sigmoid function. ; Using spatial attention weights and channel attention weights Spectral features obtained in step 2 The final fused features are obtained by performing dual adaptive adjustment. : (5)。 6. The polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method according to claim 1, characterized in that, The specific method for step 5 includes: The fusion features obtained in step 4 Global average pooling is performed to obtain feature vectors, which are then input into a fully connected layer and mapped to the number of classes. The softmax function outputs pixel-level classification probabilities, and the classification result of the target pixel is determined based on the maximum probability value among the pixel-level classification probabilities.

7. A polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification system, used to implement the method of claim 1, characterized in that, include: The polarization hyperspectral data acquisition and preprocessing module is used to acquire polarization hyperspectral data, perform radiometric correction, normalization, and polarization angle AoLP decomposition on the polarization hyperspectral data, and obtain the sinusoidal component of the polarization angle AoLP. Sum and cosine components And construct the spectral flow input tensor and polarization flow input tensor ; The feature extraction module employs a dual-stream parallel network architecture of spectral flow and polarization flow, respectively using... and Using this as input, spectral features and polarization features are extracted independently to obtain spectral features. With initial polarization characteristics ; The physics-driven enhancement module, based on the multi-scale dynamic AoLP attention module, integrates the polarization angle AoLP sinusoidal component. With cosine component Transformed into dynamic attention weights, for the initial polarization features Physically driven enhancement is performed to obtain enhanced polarization characteristics. ; The dual adaptive calibration module, based on the polarization-guided spatial-channel attention fusion module, utilizes the enhanced polarization features. Regarding the spectral features Dual adaptive calibration in both spatial and channel dimensions is performed to obtain fused features. ; The target classification module is used to classify the fused features. Mapping to the category space completes pixel-level target classification.

8. A polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification device, characterized in that, include: Memory: A computer program for a polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method as described in any one of claims 1-6, and is a computer-readable device; Processor: Used to implement the polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method according to any one of claims 1-6 when executing the computer program.

9. The present invention also provides a computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the polarization-guided multi-scale dynamic attention-based polarization hyperspectral camouflage target classification method according to any one of claims 1-6.

Images