Polarimetric sar image generation method for automatic target recognition

Abstract

The application discloses a polarimetric SAR image generation method for automatic target recognition, and relates to the technical field of polarimetric synthetic aperture radar. The method acquires a measured polarimetric SAR image of two different azimuth angles and a polarimetric SAR image of an intermediate azimuth angle; the measured polarimetric SAR image of the azimuth angle is input into a low-dimensional manifold sparse sensing feature encoder to obtain a GAS feature of the azimuth angle; the GAS feature of the azimuth angle and the measured polarimetric SAR image of the two different azimuth angles are input into a residual topology generator based on a generative adversarial network to obtain a generated polarimetric SAR image of the azimuth angle; and in the training process of the residual topology generator, the measured polarimetric SAR image of the intermediate azimuth angle is taken as a training reference image, and a multi-level constraint loss function is used for control training. The application can realize high-resolution and physically consistent polarimetric SAR image generation.

Description

Technical Field

[0001] This invention relates to the field of polarimetric synthetic aperture radar (SAR) technology, and more particularly to a method for generating polarimetric SAR images for automatic target identification. Background Technology

[0002] High-resolution polarimetric synthetic aperture radar (PolSAR) is widely used in automatic target recognition (ATR) tasks. The performance and generalization ability of a recognition model largely depend on the scale and quality of the training data. However, high-resolution polarimetric SAR data is sensitive to azimuth changes and primarily relies on costly flight experiments for acquisition. This results in usable data typically limited to a few azimuth angles, making it difficult to scale up. When the azimuth changes, the polarimetric scattering distribution and geometry change significantly. This "azimuth sensitivity" leads to inconsistent characteristics of measured data across multiple viewing angles, causing a decrease in the generalization ability of the recognition model at unseen angles.

[0003] Currently, polarimetric SAR image generation methods are mainly divided into two categories: traditional methods and deep learning-based methods. Traditional methods typically employ sample-level augmentation or electromagnetic simulation to augment the data. Sample-level augmentation increases the number of samples at the visual level through methods such as rotation, translation, and noise addition, thereby improving the robustness of the model. Electromagnetic simulation methods utilize a 3D model of the target and radar scattering equations to simulate the polarimetric scattering process and generate synthetic SAR data. Deep learning-based methods utilize generative adversarial networks or diffusion models to learn the mapping relationship from input conditions to SAR images, thereby generating high-fidelity data samples.

[0004] Traditional polarimetric SAR image generation methods have several drawbacks: 1) While image enhancement methods (such as rotation and noise addition) can generate visually diverse samples, they cannot introduce new target scattering information into polarimetric SAR and often violate physical imaging mechanisms, introducing false features that make it difficult for the model to learn the true scattering characteristics, thus reducing recognition performance. 2) Although electromagnetic simulation can simulate the polarimetric scattering process based on the target's 3D model, its generated results are affected by errors in target structure modeling, inaccurate estimation of material parameters, and computational complexity, resulting in discrepancies with measured polarimetric characteristics. Furthermore, electromagnetic simulation has low generation efficiency, making it difficult to meet the large-scale data requirements of multi-target, multi-angle scenarios, thus limiting its practical application in automatic target recognition tasks. In addition, there are some limitations in deep learning-based generation methods, such as: 1) Current research on deep generation models mainly focuses on single-polarization SAR images and has not yet been extended to fully polarized target scenes; 2) Most models still use pixel-level loss functions for natural images, which cannot effectively model the polarization scattering mechanism and the physical correlation between multiple channels, resulting in insufficient polarization feature constraints; 3) Some models use random noise as input and lack explicit control over the azimuth angle, so that although the generated images are visually realistic, they have obvious polarization characteristics deviations and poor physical consistency of scattering characteristics.

[0005] In summary, existing methods cannot simultaneously achieve polarization fidelity, azimuth controllability, and structural consistency in the multi-angle generation of polarimetric SAR vehicle targets. The lack of an effective scheme to generate multi-angle, fully polarimetric images under physical constraints makes it difficult to provide high-quality, continuous-azimuth training data support for downstream automatic target recognition. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a polarimetric SAR image generation method for automatic target recognition. This method overcomes the issue that the polarimetric scattering characteristics of vehicle targets are highly sensitive to the relative azimuth angle between the radar line of sight and the target, leading to polarimetric feature confusion in data across multiple azimuth angles. This confusion restricts the accuracy and generalization of the recognition model. Furthermore, it overcomes the limitations of existing target data generation methods in terms of physical consistency of scattering characteristics and azimuth control accuracy. This method can accurately characterize the polarimetric scattering characteristics of vehicle targets under multi-azimuth conditions, achieving high-resolution, physically consistent polarimetric SAR image generation. Simultaneously, it reduces the learning complexity of high-dimensional space during polarimetric feature modeling, simplifying model training and enabling multi-angle polarimetric SAR data expansion even under azimuth-constrained conditions, effectively improving the accuracy and generalization of the recognition model.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] A polarimetric SAR image generation method for automatic target recognition includes the following steps:

[0009] Obtain two different measured azimuth angles , Polarimetric SAR images and intermediate azimuth angles Polarimetric SAR images;

[0010] Two different azimuth angles were measured. , The polarimetric SAR image is input into a low-dimensional manifold sparse sensing feature encoder to obtain the azimuth angle. The GAS features; the azimuth angle GAS characteristics, two measured different azimuth angles , The polarimetric SAR image is simultaneously input into a residual topology generator based on a generative adversarial network to obtain the generated azimuth angle. The polarimetric SAR image; during the training process of the residual topology generator, the measured intermediate azimuth angle is used. The polarimetric SAR image is used as the training reference image, and a multi-level constraint loss function with coordinated constraints in the polarimetric domain, image domain, and physical domain is used for control training.

[0011] Polarimetric SAR images of target angles are generated based on a trained residual topology generator.

[0012] Preferably, the processing procedure of the low-dimensional manifold sparse perceptual feature encoder includes the following steps:

[0013] Dipole scattering of targets in polarimetric SAR images Dihedral scattering and Bragg scattering The coherent superposition yields the boundary scattering model. :

[0014] (4)

[0015] in, , and The scattering intensity coefficient is related to the edge length and radar wavelength. Let be a rotation matrix. For the equivalent Bragg incident angle, σ h The standard deviation of surface roughness. The angle between the edge direction and the radar incident direction. To determine the phase difference caused by path delay The resulting complex exponential phase factor;

[0016] The offset Δu of the geometric edge structure in a polarimetric SAR image can be obtained from the target attitude change Δα:

[0017] ( )

[0018] in, H For the radar platform height, φ The radar incident angle, H , φ With target attitude α Together, they determine the position of the geometric edge structure in the image;

[0019] By combining formulas (4) and (6), the correspondence between the change in scattering intensity and the change in the displacement of the scattering center is established, and the geometric orientation coupled scattering boundary GAS is obtained, as follows:

[0020] ( )

[0021] in, This represents the target polarization domain scattering model. This represents the target image domain projection model.

[0022] Preferably, the multi-level constraint loss function As shown below:

[0023] ( )

[0024] in, For polarization domain constraint loss, For image domain constraint loss, This represents the loss due to physical domain constraints.

[0025] Preferably, the polarization domain constraint loss The weighted combination of reciprocity loss and channel difference loss is shown below:

[0026] ( )

[0027] Reciprocity loss As shown below:

[0028] ( )

[0029] Among them, S HV (x,y) and S VH (x,y) represent the complex values ​​of the cross-polarization channel at the pixel (x,y) position, and N is the total number of pixels in the image;

[0030] Overall consistency across polarization channels introduces channel difference loss in four-channel complex polarization images. As shown below:

[0031] ( )

[0032] in, and These represent the complex polarization values ​​of the generated image and the measured image on channel c, respectively; and Operators for real and imaginary parts respectively; It represents the square of the L2 norm, which is the sum of the squares of the differences between the real and imaginary parts of all pixels in the image, reflecting the overall channel error.

[0033] Preferably, the image domain constraint loss The weighted combination of reconstruction loss, multi-scale structural similarity loss, perceptual feature consistency loss, and generator loss is as follows:

[0034] ( )

[0035] in, , , and For loss weighting coefficients, The reconstruction loss is used to minimize the absolute difference between the generated image and the ground truth image at the pixel level. This is a multi-scale structural similarity loss, used to measure the brightness, contrast, and structural similarity of an image at different scales.

[0036] Perceptual feature consistency loss By measuring the difference between the generated image and the measured image in the deep feature space, high-level semantic and structural information is captured, low-frequency artifacts are weakened, and the loss is expressed as:

[0037] ( )

[0038] in, Indicates the first l Layered convolutional features;

[0039] The generator and discriminator adopt the WGAN-GP framework, and the generator loss is... With discriminator loss As shown below:

[0040] ( )

[0041] in, , Azimuth , Measured polarimetric SAR images, This is a measured polarimetric SAR image at the target azimuth angle. The sparse structure features are guided by the angle corresponding to the target azimuth. λ is the random interpolation sample used to calculate the gradient penalty, and λ is the gradient penalty weight.

[0042] Preferably, the physical domain constraint loss As shown below:

[0043] ( )

[0044] in, and The amplitude values ​​of the generated image and the corresponding measured image under polarization channel c are given. c Here, N represents the sparse prior under polarization channel c, and N is the total number of pixels in the image.

[0045] Based on the above technical solution, the beneficial effects of the present invention are:

[0046] 1) The low-dimensional manifold sparse feature encoding mechanism proposed in this invention can effectively extract the geometric-orientation coupling features of vehicle targets under multi-directional conditions, realize the compact expression of polarization scattering information, and improve the physical interpretability and angle discrimination of feature representation.

[0047] 2) This invention achieves continuous and controllable generation of multi-angle polarimetric SAR images through a physical prior-guided azimuth interpolation generation strategy, ensuring the consistency of the generation results across polarimetric channels and the structural coherence under different azimuths.

[0048] 3) The multi-domain information constraints of this invention, under the joint optimization of the polarization domain, image domain and physical domain, effectively suppress problems such as polarization distortion, structural ambiguity and scattering drift, simplify the model learning difficulty, and improve the physical consistency and visual quality of the scattering characteristics of the generated image. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of a polarimetric SAR image generation method for automatic target identification in one embodiment;

[0050] Figure 2 This is a schematic diagram of a low-dimensional manifold sparse perceptual feature encoder in one embodiment;

[0051] Figure 3 This is a schematic diagram of a physical prior-guided orientation interpolation generation framework in one embodiment.

[0052] Figure 4This is a schematic diagram of a residual topology generator in one embodiment. Detailed Implementation

[0053] The technical solution of this embodiment will be clearly and completely described below with reference to the accompanying drawings.

[0054] like Figures 1 to 4 As shown, this embodiment provides a polarimetric SAR image generation method for automatic target recognition, including the following steps:

[0055] Obtain two different measured azimuth angles , Polarimetric SAR images and intermediate azimuth angles Polarimetric SAR images;

[0056] Two different azimuth angles were measured. , The polarimetric SAR image is input into a low-dimensional manifold sparse sensing feature encoder to obtain the azimuth angle. The GAS features; the azimuth angle GAS characteristics, two measured different azimuth angles , The polarimetric SAR image is simultaneously input into a residual topology generator based on a generative adversarial network to obtain the generated azimuth angle. The polarimetric SAR image; during the training process of the residual topology generator, the measured intermediate azimuth angle is used. The polarimetric SAR image is used as the training reference image, and a multi-level constraint loss function with coordinated constraints in the polarimetric domain, image domain, and physical domain is used for control training.

[0057] Polarimetric SAR images of target angles are generated based on a trained residual topology generator.

[0058] In this embodiment, firstly, the low-dimensional manifold sparse feature encoding module, based on sparse representation theory and manifold learning, guides the gradual learning of the target's low-dimensional manifold structure under continuous azimuth changes through specific structural design and input format, thereby achieving accurate feature representation of images at any specific azimuth. Simultaneously, this embodiment defines a geometry-azimuth coupled polarimetric SAR feature sparse representation, establishing a closed-loop relationship between geometric edges, polarimetric features, and imaging parameters based on the target's position and attitude. Secondly, the physics-prior-guided azimuth interpolation generation module, based on the structural mapping between polarimetric SAR target images at equal azimuth angles, explicitly embeds radar physical imaging laws and target azimuth evolution characteristics, achieving high-quality continuous azimuth generation while maintaining image physical consistency. Furthermore, this embodiment achieves unified modeling of physical guidance and azimuth control through a progressive optimization strategy, significantly simplifying the network's learning difficulty and effectively solving the inherent bias problem in polarimetric SAR generation tasks that makes it difficult to balance physical laws and generation quality. Finally, the multi-domain information constraint module forms a closed-loop optimization between the polarization domain, image domain, and physical domain through a dynamic weight adjustment mechanism, which respectively ensures physical consistency, visual fidelity, and scattering prior constraints, effectively suppressing artifacts such as polarization channel mismatch and scattering center drift, improving the structural fidelity and polarization reliability of the generated image, and providing important support for the high-quality and controllable generation of polarimetric SAR images.

[0059] In the low-dimensional manifold sparse feature encoding module, this embodiment designs a geometry-azimuth coupled polarimetric SAR feature sparse representation method. First, for high-resolution polarimetric SAR vehicle targets, their geometric contour edges will generate strong dipole scattering, dihedral scattering, and Bragg scattering. Specifically, when the radar wave incident direction is perpendicular to the target edge (such as the edge of a vehicle roof or metal corner), the edge structure will induce linear polarimetric resonance, and its scattering matrix can be expressed as:

[0060] (1)

[0061] in, Let be a rotation matrix. This is the angle between the edge direction and the radar incident direction, typically characterized by a bright red line in Pauli images. . This represents the scattering response of a radar when it transmits with horizontal polarization (H) and receives with horizontal polarization (H). This represents the scattering response of a radar when it transmits and receives signals with vertical polarization (V).

[0062] When two planes form a right angle (such as a vehicle and the ground), a corner reflector effect occurs, and its scattering matrix is:

[0063] ( )

[0064] in, This is the phase difference caused by path delay, which is typically characterized by a strong blue tint in Pauli images. Quantity.

[0065] When the target surface has micro-roughness (such as weld seams on a vehicle surface), coherent superposition of surface scattering occurs:

[0066] ( )

[0067] in, and For horizontal / vertical polarization reflectance, The equivalent Bragg angle of incidence is typically characterized by green dots in the Pauli image. and Quantity. This represents the scattering response of a radar when it transmits with horizontal polarization (H) and receives with vertical polarization (V). This represents the scattering response of a radar when it transmits with vertical polarization (V) and receives with horizontal polarization (H). For man-made targets such as vehicles, the scattering response of their geometric edge structures (such as the vehicle's geometric corners, the junctions between the roof and body, and between the body and wheels) can be simplified into a coherent superposition of three typical scattering mechanisms:

[0068] ( )

[0069] in, , and σ is the scattering intensity coefficient related to the edge length and radar wavelength. h This represents the standard deviation of surface roughness. The angle between the edge direction and the radar incident direction. To determine the phase difference caused by path delay The resulting complex exponential phase factor. This model shows that the azimuth angle θ of the geometric edge structure directly determines the rotational characteristics of the polarization scattering matrix, while the physical length of the edge affects the scattering intensity k.

[0070] For position projection in polarimetric SAR imaging geometry, the projection of the target's 3D coordinates (x, y, z) onto the image plane (u, v) can be expressed as:

[0071] ( )

[0072] Where H is the radar platform height, The radar incident angle, and Characterizes the target attitude α (pitch / yaw angle). The offset Δu of the geometric edge structure position can be obtained from the target attitude change Δα:

[0073] ( )

[0074] This indicates that: H For the radar platform height, φ The radar incident angle, both ( H,φ In this paper, these are collectively referred to as platform parameters, used to describe the spatial location of the radar platform and its corresponding imaging geometry. Platform parameters ( H,φ ) and target attitude ( α These factors together determine the position of the geometric edge structure in the image.

[0075] Meanwhile, due to the structural rigidity of vehicle targets, the scattering characteristics of their geometric edge structures are usually highly stable, maintaining a consistent shape under different radar incident directions, reflecting the physical structure of the target. Equation (4) describes the scattering mechanism and polarization characteristics of the target boundary as the azimuth angle changes, and Equation (6) describes the positional change of the boundary in the imaging coordinates. Therefore, this paper establishes a correspondence between the change in scattering intensity and the change in the displacement of the scattering center by combining equations (4) and (6), thus forming a closed-loop relationship of geometric edge-polarization characteristics-imaging parameters, which is called the Geometric-Azimuth Coupled Scattering Boundary (GAS):

[0076] ( )

[0077] in, This represents the target polarization domain scattering model. The target image domain projection model is represented. Through polarimetric scattering modeling and imaging geometry analysis, the GAS can be rigorously described as a joint representation of the target's geometric contour in the polarimetric and image domains.

[0078] To establish a unified description between target scattering mechanism and imaging geometry, this embodiment of the invention completes the entire modeling process from polarization scattering model to low-dimensional sparse feature representation through formulas (1) to (7). First, formulas (1)–(3) construct the basic polarization scattering expression of the target boundary using three typical scattering mechanisms; formula (4) further combines these scattering atoms according to weights and introduces azimuth-related rotation operators and phase terms to obtain a boundary scattering model that can reflect the changes in scattering intensity and polarization mode of the target under different observation angles. Subsequently, geometric projection relationships are established through formulas (5)–(6) to describe the displacement and deformation law of the target boundary on the radar imaging plane, so that the azimuth change of scattering intensity and the geometric position change of scattering center are uniformly characterized under the same framework. Based on this, formula (7) represents the scattering of the target polarization domain. Projection representation of the target image domain Coupling is achieved to form a geometrically oriented coupled scattering boundary (GAS), thus realizing a closed-loop correlation between "scattering mechanism—geometric structure—observation azimuth" at the physical level. Low-dimensional representations are obtained through sparse mapping and manifold compression, enabling azimuth-sensitive scattering modes, boundary displacements, and structural perturbations to form continuous and separable feature trajectories in the encoding space. For example... Figure 2 As shown, specifically, the low-dimensional manifold sparse sensing feature encoder takes two pairs of polarimetric SAR images from different azimuths as input. Through adversarial training, it maps these images to the manifold position corresponding to the intermediate azimuth angle. The discriminator jointly evaluates the manifold fit and sparsity compliance of the GAS to enhance physical plausibility. On the generator side, cross-layer sparsity selection and channel partitioning are adopted, retaining only channels strongly correlated with azimuth to reduce computational complexity and improve representation consistency. In the feature fusion stage, polarimetric channel attention (PCA) based on the Convolutional Block Attention Module (CBAM) is introduced. This adaptively enhances the response related to the dominant scattering mechanism and suppresses redundant noise according to the channel contribution. Through attention weighting and residual skip connection constraints, an interpretable projection of high-dimensional polarimetric information onto the low-dimensional GAS space is achieved. This overall design achieves sparse representation with a small number of physically driven basis functions, thereby improving the generalization ability for unknown azimuth angles and ensuring the physical consistency and structural fidelity of the generated results.

[0079] This embodiment proposes a physically prior-guided orientation interpolation generation framework, such as... Figure 3 As shown, this framework explicitly embeds the physical laws of radar imaging and the evolution characteristics of target azimuth, achieving high-quality generation that combines physical consistency and azimuth continuity, forming a closed-loop optimization chain of "physical prior extraction - fine generation - multi-dimensional evaluation".

[0080] The generator employs a cascaded residual topology, combining target geometry and azimuth coupling for multi-stage feature interaction. GAS features, serving as low-dimensional sparse physical priors and multi-polarization deep semantic features, are fused under channel attention PCA and residual jump-connection constraints. This dynamically adjusts polarization channel contributions while maintaining spatial consistency of the scattering center, outputting the generated result at a specified intermediate azimuth angle. The input receives measured images of adjacent azimuth angles and their sparse priors; three-branch joint encoding provides global and local physical constraints. The discriminator uses a dual-path progressive evaluation, jointly considering manifold fit, sparsity compliance, and structural coherence. From data-driven to physical modeling, multi-level consistency optimization is achieved, suppressing artifacts such as polarization channel mismatch and scattering center drift. This framework can systematically complete multi-azimuth samples of similar targets, enhancing intra-class pattern learning under attitude changes. Simultaneously, it generates different categories of samples at each azimuth angle, improving feature discriminative power and resistance to azimuth interference, providing high-quality, controllable, and physically reliable data support for polarimetric SAR vehicle ATR.

[0081] This embodiment proposes a polarimetric SAR image generation method for automatic target recognition, and specifically discloses a multi-level constraint loss function. This multi-level constraint loss function achieves coordinated optimization through constraints at the following three levels (polarimetric domain constraint loss, image domain constraint loss, and physical domain constraint loss):

[0082] 1) Polarization domain constraint loss To ensure the physical plausibility of the generated image in terms of polarization characteristics, this embodiment applies two constraints. First, the cross-polarization channels are constrained to satisfy the polarization reciprocity principle; second, the overall coordination between channels is enhanced by constraining the amplitude and phase consistency between the generated data and the measured data in each polarization channel. The reciprocity loss is defined as follows:

[0083] ( )

[0084] Among them, S HV (x,y) and S VH (x, y) represent the complex values ​​of the cross-polarization channels at pixel (x, y), where N is the total number of pixels. The overall consistency between polarization channels introduces a channel difference loss for the four-channel complex polarization image:

[0085] ( )

[0086] in, and These represent the complex polarization values ​​of the generated image and the measured image on channel c, respectively. and Operators for real and imaginary parts respectively. It represents the square of the L2 norm, which is the sum of the squares of the differences between the real and imaginary parts of all pixels in the image, reflecting the overall channel error.

[0087] Ultimately, the polarization domain loss is a weighted combination of the two terms mentioned above:

[0088] ( )

[0089] in, , The loss weighting coefficient is used to balance the effects of reciprocity constraints and channel difference constraints.

[0090] 2) Image domain constraint loss Image domain constraints strike a balance between pixel-level fidelity and deep feature consistency. To improve the quality of generated images in terms of structural reconstruction and texture fidelity, this paper designs a multi-scale loss function that fuses pixel-level and semantic-level loss in the image domain. Specifically, the L1 reconstruction loss is used to minimize the absolute difference between the generated image and the real image at the pixel level, and is defined as follows:

[0091] ( )

[0092] in, and These represent the intensity maps of the generated image and the measured image, respectively. Where is the pixel position, and N is the total number of pixels.

[0093] Multi-scale structural similarity loss measures the brightness, contrast, and structural similarity of an image at different scales, providing a more comprehensive reflection of the visual consistency of an image. It is defined as follows:

[0094] ( )

[0095] in, Here are the measured and generated images of the j-th pyramid level, where M is the pyramid level. , , These are the contrast similarity term, the structural similarity term, and the brightness similarity term, respectively. and This is the weighted index for each layer.

[0096] Perceptual feature consistency loss measures the difference between the generated image and the measured image in the deep feature space, captures high-level semantic and structural information, and weakens low-frequency artifacts. The loss is expressed as:

[0097] ( )

[0098] in, Indicates the first l Layered convolution features.

[0099] The generator and discriminator adopt the Wasserstein GAN with Gradient Penalty (WGAN-GP) framework, whose gradient stability is more suitable for this type of polarimetric SAR multi-channel physical data, and are defined as follows:

[0100] ( )

[0101] in, , Azimuth , Measured polarimetric SAR images, This is a measured polarimetric SAR image at the target azimuth angle. The sparse structure features are guided by the angle corresponding to the target azimuth. λ is the random interpolation sample used to calculate the gradient penalty, and λ is the gradient penalty weight.

[0102] Ultimately, the image domain loss is a weighted combination of the above terms:

[0103] ( )

[0104] in, , , and This is the loss weighting coefficient.

[0105] 3) Physical domain constraint loss Physical domain constraints embed prior knowledge of electromagnetic scattering into the optimization process, imposing global constraints on the overall directional consistency and scattering stability of the generated image. This ensures that the generated image is physically consistent with polarimetric SAR imaging in terms of scattering response and structural stability. Its mathematical definition is as follows:

[0106] ( )

[0107] in, and The amplitude values ​​of the generated image and the corresponding measured image under polarization channel c are given. c Here, N represents the sparse prior under polarization channel c, and N is the total number of pixels in the image.

[0108] Finally, the overall optimization objective of the multi-level constraint loss function is defined as follows:

[0109] ( )

[0110] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0111] The above are merely preferred embodiments of the present application and are not intended to limit the embodiments of the present application. For those skilled in the art, the embodiments of the present application can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of the present application should be included within the protection scope of the embodiments of the present application.

Claims

1. A polarimetric SAR image generation method for automatic target recognition, characterized in that, The method comprises the following steps: Obtain two different measured azimuth angles , Polarimetric SAR images and intermediate azimuth angles Polarimetric SAR images; the measured polarization SAR images of two different azimuth angles , are input into a low-dimensional manifold sparse perception feature encoder to obtain the GAS features of the azimuth angle ; the GAS features of the azimuth angle , the measured polarization SAR images of two different azimuth angles , are simultaneously input into a residual topology generator based on a generative adversarial network to obtain the generated polarization SAR image of the azimuth angle ; in the training process of the residual topology generator, the polarization SAR image of the intermediate azimuth angle is used as a training reference image, and a multi-level constraint loss function with polarization domain-image domain-physical domain collaborative constraint is used for control training, wherein the processing process of the low-dimensional manifold sparse perception feature encoder includes the following steps: Dipole scattering of targets in polarimetric SAR images , dihedral scattering , and bragg scattering coherent superposition of which yields the boundary scattering model : (4) wherein, , and is a scattering intensity coefficient related to the edge length and the radar wavelength, is a rotation matrix, is an equivalent Bragg incidence angle, σ h is a standard deviation of surface roughness, is an angle between the edge direction and the radar incidence direction, is a phase difference caused by path delay a complex exponential phase factor formed; The polarization SAR image geometric edge structure position offset amount Δu can be obtained from the target attitude change Δα: ( ) where H is the radar platform height, φ is the radar incidence angle, H, The target attitude α together determines the position of the geometric edge structure in the image. The corresponding relationship between the scattering intensity change and the scattering center displacement change is established by combining formula (4) and formula (6), and the geometric azimuth coupling scattering boundary (GAS) is obtained, and the formula is as follows: ( ) wherein, denotes a target polarized domain scattering model, denotes a target image domain projection model; Generate the polarized SAR image of the target angle based on the trained residual topology generator.

2. The polarimetric SAR image generation method for automatic target recognition according to claim 1, characterized in that, The multi-level constraint loss function As follows: ( ) wherein, is a polar domain constraint loss, is an image domain constraint loss, is a physical domain constraint loss.

3. The polarimetric SAR image generation method for automatic target recognition according to claim 2, characterized in that, the polar domain constraint loss is a weighted combination of the reciprocity loss and the channel difference loss, as follows: ( ) Reciprocity loss As follows: ( ) where S HV (x,y) and S VH (x,y) represent the cross-polar channel complex value at pixel (x,y) position, and N is the total number of pixels in the image. Polarization channel-wise overall consistency introduces channel difference loss for four-channel complex polarized images As shown below: ( ) wherein, and and and are real and imaginary operators, respectively; denotes the square of the L2 norm, i.e. the sum of the squared differences of the real or imaginary parts of all pixels in the image, reflecting the channel error over the whole image.

4. The polarimetric SAR image generation method for automatic target recognition according to claim 2, characterized in that, the image domain constraint loss is a weighted combination of the reconstruction loss, the multi-scale structural similarity loss, the perceptual feature consistency loss, and the generator loss, as follows: ( ) wherein, , , and is a loss weight coefficient, is a reconstruction loss for minimizing the absolute difference between the generated image and the real image at the pixel level, is a multi-scale structural similarity loss for measuring the brightness, contrast, and structural similarity of the image at different scales. Perceptual feature consistency loss By measuring the difference between the generated image and the measured image in the deep feature space, the high-level semantic and structural information is captured, and the low-frequency artifacts are weakened. The loss is expressed as: ( ) wherein, represents the l-th layer convolutional feature; The generator and the discriminator adopt the WGAN-GP framework, and the generator loss and the discriminator loss are as follows: ( ) wherein, , is an azimuth angle , of the measured polarimetric SAR image, is the measured polarimetric SAR image at the target azimuth angle, is the angle-guided sparse structure feature corresponding to the target azimuth angle, is a gradient-penalized random interpolation sample for computing, and λ is a gradient-penalized weight.

5. The polarimetric SAR image generation method for automatic target recognition according to claim 2, characterized in that, the physical domain constraint loss as follows: ( ) wherein, and are the amplitude values of the generated image and the corresponding measured image under polarization channel c, c is the sparse prior under polarization channel c, and N is the total number of pixels of the image.

Images