SAR (Synthetic Aperture Radar) image cross-angle generation method and system based on guidance of physical scattering model

Abstract

The invention discloses an SAR image generation method based on guidance of a physical scattering model. The SAR image generation method comprises physically guided feature modulation, attribute scattering center feature constraint and generative network training. According to the physically-guided feature modulation, pitch angle parameterized coding and multi-scale convolution are fused by designing a multi-scale attention module PMA, key scattering features are dynamically enhanced, edge attenuation caused by pitch angle changes is compensated, and geometric continuity of linear scattering is kept. According to the attribute scattering center feature constraint, an ASC loss function is provided based on an attribute scattering center model, multi-scale Gaussian filtering is carried out on frequency domain amplitude spectrums and phase spectrums of generated and real images, and the consistency of the position, the strength and the type of a scattering center is constrained through an L1 distance, so that the generated image is ensured to conform to an electromagnetic scattering physical law. And a CycleGAN generative adversarial network embedded with a PMA module is constructed by generative network training, and cross-angle high-quality SAR image generation is realized by combining end-to-end optimization of adversarial loss, cyclic consistency loss and ASC loss. The method provides a high-precision and physically interpretable solution for SAR image generation and target identification.

Description

Technical Field

[0001] The present invention relates to the technical field of synthetic aperture radar (SAR) image generation, and in particular to a method and system for generating SAR images across angles guided by a physical scattering model. Background Art

[0002] Synthetic Aperture Radar (SAR) utilizes microwave imaging, which can penetrate inclement weather such as clouds, rain, and fog, and is unaffected by day or night, enabling all-day, all-weather operation. This overcomes the limitations of optical remote sensing technology and is widely used in geological exploration, resource surveys, and environmental monitoring. Its image scattering characteristics are closely related to target geometry, material properties, and parameters such as the angle of incidence and elevation. However, its imaging mechanism is complex and susceptible to platform motion errors, signal bandwidth, and the Doppler effect, leading to geometric distortion, speckle noise, and other issues, complicating data acquisition and interpretation.

[0003] With the development of convolutional neural networks (CNNs), target recognition and detection methods in SAR image processing rely on large-scale annotated data for model training. However, SAR image acquisition is difficult (due to the high development and maintenance costs of spaceborne SAR and limited field data), and insufficient data limits model performance. Traditional data augmentation methods (such as rotation, translation, and noise injection) offer limited improvements in data diversity; the texture of images generated by electromagnetic simulation differs from real data and is computationally complex; and attribute scattering center models lack the ability to integrate end-to-end with deep learning. While methods based on generative adversarial networks (GANs) have made progress, they ignore the electromagnetic scattering characteristics of SAR and the complex imaging mechanisms, resulting in images that do not conform to physical laws and exhibit structural distortions (e.g., large errors in scattering center position, fractured line scattering, and irregular projection lengths of surface scattering).

[0004] Therefore, there is an urgent need for a method to generate cross-angle SAR images that conform to physical laws, to solve the problems of image generation that do not conform to physical laws and structural distortion in existing technologies, so as to improve the target detection and recognition performance of cross-angle SAR images. Summary of the Invention

[0005] In view of the shortcomings and deficiencies of the existing technology, the present invention aims to provide a cross-angle SAR image generation method guided by a physical scattering model. The method is designed by a generative adversarial network with a multi-scale attention mechanism and frequency domain scattering feature constraints to solve the geometric distortion and scattering degradation problems of traditional methods in cross-angle SAR image generation.

[0006] The specific technical solutions of the present invention are as follows:

[0007] The present invention provides a method for generating SAR images across angles based on a physical scattering model, comprising the following steps:

[0008] S1. Standardize and preprocess the original SAR image, including resizing to a fixed resolution and noise suppression;

[0009] S2, build a bidirectional generation architecture based on CycleGAN, which includes two generators (G A→B , G B→A ) and two discriminators (D A , D B ), supports adversarial training of incomplete paired data;

[0010] S3, each generator includes an encoder module, a residual block group and a decoder module connected in sequence, the encoder module downsamples and compresses features through three layers of convolutional layers, the decoder module upsamples and restores the resolution through three layers of deconvolutional layers, and the residual block group consists of 9 series-connected skip-connected residual blocks;

[0011] S4, embedding a physical guided multi-scale attention module PMA in the 2nd, 4th, and 6th layer residual blocks of the generator;

[0012] S5. The discriminator uses the patchGAN structure to perform authenticity judgment on a 70×70 local area of ​​the input image and output a 30×30 probability matrix;

[0013] S6. Perform fast Fourier transform on the generated image and the real image to obtain amplitude spectrum and phase spectrum of the real image and the generated image;

[0014] S7. Calculate the attribute scattering feature center of the real image and the generated image in the frequency domain, and calculate the ASC loss function based on this;

[0015] S8, perform backpropagation on the joint adversarial loss, cycle consistency loss and ASC loss to update the network weights;

[0016] S9. Repeat S3 to S8 until the network converges and save the optimal model parameters.

[0017] A preferred technical solution further includes an image generation step:

[0018] S10, resizing and normalizing the test SAR image to meet the model input requirements;

[0019] S11, input the processed test image and target pitch angle parameters into the trained generation network, and output the SAR image of the target angle;

[0020] S12. Post-process the output results, including linearly mapping the pixel values ​​from [-1, 1] to [255, 255], converting them into 8-bit grayscale images, and enhancing the visual effects through histogram equalization.

[0021] A second aspect of the present invention discloses a SAR image cross-angle generation system guided by a physical scattering model, comprising:

[0022] a preprocessing module configured to perform standardized preprocessing on the raw SAR image, including resizing to a fixed resolution and noise suppression;

[0023] The adversarial network building module is configured to build a two-way generative adversarial network based on CycleGAN, which contains two generators (G A→B , G B→A ) and two discriminators (D A , D B );

[0024] The generator module adopts an encoder-residual layer-decoder structure. The encoder compresses features through 3 layers of downsampling, and the decoder restores the resolution through 3 layers of upsampling. 9 residual blocks are embedded in the middle.

[0025] The PMA embedding module is configured to embed the physically guided multi-scale attention module in the 2nd, 4th, and 6th layer residual blocks of the generator;

[0026] The discriminator module uses the patchGAN structure to perform authenticity discrimination on a 70×70 local area of ​​the input image and output a 30×30 probability matrix;

[0027] an FFT processing module configured to perform a fast Fourier transform (FFT) on the generated image and the real image to extract an amplitude spectrum and a phase spectrum;

[0028] An ASC loss calculation module is configured to calculate the attribute scattering center (ASC) of the real image and the generated image in the frequency domain and generate an ASC loss function;

[0029] Backpropagation module, configured to perform backpropagation on the joint adversarial loss, cycle consistency loss, and ASC loss to update the network weights;

[0030] The model saving module is configured to save the optimal model parameters after the network converges.

[0031] A preferred technical solution also includes

[0032] A test preprocessing module configured to resize and normalize the test SAR images to fit the input requirements of the generative network;

[0033] A parameter input module is configured to concatenate the processed test image and target pitch angle parameters into a multi-dimensional tensor and input it into the trained generative network;

[0034] The post-processing module is configured to perform the following processing on the image output by the generated network: linearly map the pixel value from [-1, 1] to [255, 255]; convert the floating-point data into an 8-bit unsigned integer grayscale image; and enhance the image contrast through the histogram equalization algorithm.

[0035] Based on the above invention principles, the beneficial effects of the present invention are as follows:

[0036] This paper innovatively proposes a physically guided SAR image generation method. Through physically guided feature modulation and attribute scattering center constraints, it significantly improves the generation quality and physical consistency of SAR images across all elevation angles, providing a high-precision and high-robustness solution for SAR image generation and target recognition tasks, and has a wide range of application scenarios in military reconnaissance, disaster monitoring and other fields.

[0037] The proposed SAR image generation method first explicitly models the effect of pitch angle on scattering characteristics through the PMA module, such as compensating for edge attenuation and maintaining line scattering continuity, thereby improving the physical interpretability of the generated images. Secondly, the ASC loss is used to jointly optimize the frequency domain amplitude and phase spectra to avoid geometric distortion, significantly improving the generated image quality. Finally, the generated SAR image conforms to the electromagnetic scattering laws of the attributed scattering center model (ASC). This model can be directly used as training data for input into various target detection frameworks, improving cross-angle detection performance.

[0038] The method of the present invention only requires a small number of cross-angle samples for training, significantly reducing the dependence on measured data, and is suitable for scenarios with few samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is an overall flow chart of an embodiment of the SAR image cross-angle generation method guided by the physical scattering model of the present invention;

[0040] Figure 2 It is a structural diagram of the physical guided multi-scale attention module PMA in the generation method of the present invention;

[0041] Figure 3 This is a calculation flow chart of the attribute scattering center ASC loss generated by the method of the present invention;

[0042] Figure 4 It is a flow chart of an implementation method of the SAR image generation network generation method of the present invention. DETAILED DESCRIPTION

[0043] The present invention provides a SAR image generation method guided by a physical scattering model, which includes physical-guided feature modulation and attribute scattering center feature constraint. (1) Physical-guided feature modulation: A physical-guided multi-scale attention module (PMA) is designed to dynamically enhance the key scattering of SAR images by fusing pitch angle parameterized encoding with multi-scale convolution. Specifically, the pitch angle is mapped to a multi-scale modulation weight, which controls the feature response of 3*3, 5*5, and 7*7 convolution kernels respectively, compensates for the edge attenuation caused by the pitch angle change, maintains the geometric continuity of the line scattering structure, and constrains the projection length of the surface scattering. (2) Attribute scattering center feature constraint: An ASC loss function is proposed based on the attribute scattering center model, and multi-scale Gaussian filtering is performed on the frequency domain amplitude spectrum and phase spectrum of the generated image and the real image. The position, intensity, and type consistency of the scattering center are constrained by the L1 distance to ensure that the generated image conforms to the physical laws of electromagnetic scattering. (3) Generative network training: A generative adversarial network based on CycleGAN is constructed, and the PMA module is embedded in the generator. The adversarial loss, cycle consistency loss, and ASC loss are combined for end-to-end optimization to generate high-quality SAR images across angles.

[0044] Compared with the existing technology, the method of the present invention has the following beneficial effects: (1) Physically guided multi-scale attention module (PMA): Explicitly modeling the effect of pitch angle on scattering characteristics, generating interpretable modulation weights through angle encoders, and solving the problem that traditional attention mechanisms lack physical interpretability. (2) ASC loss function: Combine frequency domain amplitude features and phase features for multi-scale matching to avoid geometric distortion and scattering degradation of the generated image and improve physical consistency. (3) Universality of the generation method: The generated SAR image is compatible with different target detection frameworks, improving target detection performance across angles. The present invention provides a high-precision, physically interpretable solution for SAR image generation and target recognition tasks.

[0045] The present invention is further described below by way of examples, but should not be construed as limiting the present invention. Without departing from the spirit and substance of the present invention, modifications or substitutions made to the method, steps or conditions of the present invention are within the scope of the present invention.

[0046] The present invention provides a SAR image generation method based on physical scattering model guidance, which specifically includes two stages: SAR image generation network training based on physical guidance and image generation.

[0047] refer to Figure 1 The SAR image generation network training based on physical guidance specifically includes the following steps:

[0048] S1. Perform standardization preprocessing on the original SAR image, including resizing to a fixed resolution and noise suppression to ensure the consistency of input data distribution; for example, resize the image to 128*128.

[0049] S2, build a bidirectional generation architecture based on CycleGAN, which includes two generators (G A→B , G B→A ) and two discriminators (D A , D B ), supporting adversarial training on incompletely paired data.

[0050] The generator and the discriminator are trained against each other to generate SAR images. A→B ) and the second generator (G B→A ) are used for mapping the SAR image from source domain to target domain and from target domain to source domain respectively; the first discriminator (D A ) and the second discriminator (D B ) respectively for G A→B The generated image and the target domain real image, G B→A The generated images are compared with the real images in the source domain for authenticity discrimination.

[0051] S3. Each of the generators includes an encoder module, a residual block group and a decoder module connected in sequence. The encoder module downsamples and compresses features through three layers of convolutional layers, and the decoder module upsamples and restores the resolution through three layers of deconvolutional layers. The residual block group consists of 9 serially connected skip-connected residual blocks.

[0052] The specific method is:

[0053] 1) Input a raw SAR image of size (C × H × W) into the Res-Net generator, where C is the number of input channels, and H and W are the image spatial dimensions. After a 7x7 convolution with a stride of 1 and padding of 3, instance normalization, and ReLU activation, the output is a feature map of size (64 × H × W), which enters the subsequent encoding process.

[0054] 2) The feature of size (64×H×W) is input into a 3*3 convolution with a stride of 2 and a padding of 1 for the first downsampling, and finally outputs a feature map of size (128×H / 2×W / 2).

[0055] 3) The feature of size (128×H / 2×W / 2) is input to a 3*3 convolution with a stride of 2 and a padding of 1 for a second downsampling, and finally outputs a feature map of size (256×H / 4×W / 4).

[0056] 4) Input the features of size (256×H / 4×W / 4) into a module containing 9 residual blocks, embed the PMA module (see step 4 for details) in the 2nd, 4th, and 6th residual blocks, keep the output dimension unchanged, and finally output a feature map of size (256×H / 4×W / 4).

[0057] 5) The decoder performs upsampling. The features of size (256×H / 4×W / 4) are upsampled for the first time through a 3*3 transposed convolution kernel with a stride of 2, and the output is a feature map of size (128×H / 2×W / 2).

[0058] 6) The feature with a size of (128×H / 2×W / 2) is upsampled for the second time through a 3*3 transposed convolution kernel with a step size of 2, and the output is a feature map of size (64×H×W).

[0059] 7) The features of size (64×H×W) are input to the output layer, and through the 1*1 convolution layer and Tanh activation function, the final output size is (1×H×W).

[0060] S4. Embed a physically guided multi-scale attention module (PMA) in the 2nd, 4th, and 6th residual blocks of the generator; by explicitly encoding the pitch angle parameter θ, the pitch angle parameter θ is fused with multi-scale local features to dynamically adjust the attention distribution of the feature map.

[0061] refer to Figure 2 、Specific methods are:

[0062] 1) First, normalize the pitch angle parameter θ to [0,1].

[0063] 2) Then, the pitch angle parameter θ is mapped into three sets of 64-dimensional modulation weights W3, W5, and W7 through an MLP angle encoder consisting of two levels of fully connected layers.

[0064] 3) Capture features of different scales through parallel convolution 3*3, 5*5, 7*7 operations. The input F is (256×H / 4×W / 4) and the output C s (F) is (64×H / 4×W / 4).

[0065] 4) Multiply the convolution of each scale by the corresponding weight channel by channel C s (F)⊙W s , where W s The three groups of 64-dimensional modulation weights W3, W5, and W7 obtained in step 4 (2) have the same dimension after channel-by-channel multiplication, and the output is (64×H / 4×W / 4).

[0066] 5) After adding the weighted features of size (64×H / 4×W / 4), the sigmoid activation function is used to generate the spatial attention map Attn(F,θ) of size (64×H / 4×W / 4) at the corresponding pitch angle:

[0067]

[0068] 6) Finally, the features of the key areas are enhanced through residual connections, and the feature map of size (256×H / 4×W / 4) is output through the original features + physical modulation features.

[0069]

[0070] S5. The discriminator uses the patchGAN structure to judge the authenticity of a 70×70 local area of ​​the input image and output a 30×30 probability matrix. The authenticity of the local area of ​​the image is judged through a convolutional network.

[0071] The specific steps are as follows:

[0072] 1) Input (1×H×W) generated image.

[0073] 2) The image of size (1×H×W) is convolved with a 4*4 convolution kernel with a stride of 2, and the output is a feature of size (64×H / 2×W / 2).

[0074] 3) The feature of size (64×H / 2×W / 2) is convolved with a 4*4 convolution kernel with a stride of 2, and the output is a feature of size (128×H / 4×W / 4).

[0075] 4) The feature of size (128×H / 4×W / 4) is convolved with a 4*4 convolution kernel with a stride of 2, and the output is a feature of size (256×H / 8×W / 8).

[0076] 5) The feature of size (256×H / 8×W / 8) is convolved with a 4*4 convolution kernel with a stride of 1, and the output is a feature of size (512×H / 8×W / 8).

[0077] 6) The feature of size (512×H / 8×W / 8) is passed through a 4*4 convolution kernel with a step size of 1, and the output is a feature of size (1×H / 8×W / 8), which is then passed through a sigmoid activation function.

[0078] 7) Output a probability matrix of (H / 8×W / 8), where each element corresponds to the authenticity score of a 70*70 local area of ​​the input image.

[0079] 8) Conduct counter-loss Through the adversarial game between the generator G and the discriminator D, the generated image is forced to approach the real data distribution. The probability of the discriminator distinguishing the real image should be as close to 1 as possible. The probability of the discriminator for generating images should be as close to 0 as possible:

[0080]

[0081] S6. Perform fast Fourier transform on the generated image and the real image to obtain amplitude spectrum and phase spectrum of the real image and the generated image.

[0082] S7. Calculate the attribute scattering feature center of the real image and the generated image in the frequency domain, and calculate the ASC loss function based on this.

[0083] By performing multi-scale matching on the scattering characteristics of the generated image and the real image in the selected resonant frequency band, the band-pass filter M(f x ,f y ) retain key scattering information and use Gaussian convolution kernel G s The amplitude spectrum and phase spectrum are processed separately at three scales, and finally the scattering feature consistency between the generated image and the real image is calculated by calculating the L1 distance constraint. Figure 3 , the specific method is:

[0084] 1) Given a generated image Compared with the real image I real , through differentiable frequency domain processing to extract the physical guided scattering characteristics, first perform fast Fourier transform FFT to obtain the frequency domain representation, and then convert to the frequency domain to analyze the scattering characteristics:

[0085]

[0086] in, is the amplitude spectrum, φ(I)∈R H×W For the phase spectrum, a differentiable frequency domain processor is designed to jointly optimize the frequency domain amplitude and phase characteristics to ensure the accuracy of the scattering center position and the authenticity of the intensity distribution.

[0087] 2) Based on the geometric theory of diffraction (GTD) and physical optics (PO), the target scattering in the high-frequency region can be approximated as the superposition of multiple ASCs, and the response E of each scattering center is i (f,φ) is described by the following formula:

[0088]

[0089] Among them E i (f,φ) is the complex scattering field of the i-th scattering center at the radar operating frequency f and azimuth φ, α iis a frequency-dependent factor reflecting the geometry of the scatterer, (x i ,y i ) represents the spatial position of the scattering center and determines the geometric structure of the SAR image. i is the scattering intensity, which is related to the material properties of the target. c is the center frequency, j is the imaginary unit, and c represents the speed of light.

[0090] 3) Through the learnable frequency mask M(f x ,f y ) to focus on the target sensitive frequency band. The SAR system usually operates in the medium and high frequency band (normalized frequency of 0.1-0.4), which corresponds to the resonant scattering area of ​​typical military targets and contains the main scattering characteristics of the target. x ,f y ) retains key scattering information:

[0091]

[0092] Where (f x ,f y ) is the spatial frequency of the SAR image, The radial frequency represents the distance of the frequency component from the origin.

[0093] 4) Perform multi-scale feature extraction. The 3*3 convolution kernel size corresponds to the high-frequency local features, that is, the position of the scattering center (x i ,y i ); The convolution kernel size of 5*5 corresponds to the mid-frequency structural characteristics, that is, the length L of the scatterer; the convolution kernel size of 7*7 corresponds to the low-frequency global characteristics, that is, the scattering intensity A i . The Gaussian kernel function is:

[0094]

[0095] Among them G s (x,y) is the weight of the Gaussian convolution kernel at position (x,y), which is used to weight neighboring pixels, smooth the image, and extract features at a specific scale. τ is the standard deviation of the Gaussian function. A larger standard deviation results in a smoother convolution kernel and a lower sensitivity to small-scale features.

[0096] 5) Finally, the ASC loss function is defined as the L1 distance between the generated image and the real image feature tensor, because the L1 loss is robust to outliers and is more suitable for matching sparse scattered features:

[0097]

[0098] where λ ASC is the weight coefficient. To generate the ASC features of the image, is the ASC feature of the real image, where F ASC Defined as:

[0099]

[0100] Among them G s is a Gaussian convolution kernel of size s×s, is the amplitude spectrum of the image, and φ(I) is the phase spectrum of the image.

[0101] S8, backpropagate the joint adversarial loss, cycle consistency loss and ASC loss, update the network weights, and continue the next training. The total loss function of the generator is:

[0102]

[0103] Among them, λ is the weight coefficient, which is used to adjust different loss functions. is the cycle consistency loss, defined as follows:

[0104]

[0105] Among them I s is the source domain image, G is the generator from the source domain to the target domain, and G' is the generator from the target domain to the source domain. The cycle consistency loss constrains the reversibility of the domain conversion by forcing G' and G to be inverse transformations of each other, thereby preserving the input content structure and avoiding mode collapse. The detailed description of is in step 3, 8), which represents the adversarial loss.

[0106] S9. Repeat steps 3 to 8 until the loss approaches zero and the network finally converges, saving the model parameters with the optimal training results.

[0107] The SAR image generation network generation process based on physical guidance of the present invention is as follows: Figure 4 As shown, the specific steps include:

[0108] Step 10: Resize and normalize the SAR image for testing to ensure that it effectively fits the input requirements of the model.

[0109] Step 11: Input the test SAR image and the target pitch angle parameter θ into the trained model and output the SAR image of the target angle.

[0110] Step 12: The model output is post-processed to generate the final image. The output is linearly mapped from [-1.1] to [255, 255], converted to an 8-bit grayscale image, and histogram equalization is used to improve the visual effect.

[0111] Example 1

[0112] For example, the MSTAR dataset, collected by the Sandia National Laboratories SAR sensor platform, contains ten ground target categories: BMP2, BTR70, T72, T62, BRDM2, BTR60, ZSU23 / 4, D7, ZIL131, and 2S1, as well as background images. These images were acquired using an X-band SAR system with a resolution of 0.3 m × 0.3 m. The ground target size is approximately 128 × 128 pixels, and the background image size is approximately 1400 × 1700 pixels.

[0113] The SAR image generation network training process based on physical guidance of the present invention specifically includes the following steps:

[0114] Step 1: Read the original SAR image for training and resize the image to 128×128.

[0115] Step 2: Using a bidirectional generation architecture based on CycleGAN to support incomplete paired data, we build a physics-guided generative adversarial network consisting of two generators (G A→B , G B→A ) and two discriminators (D A , D B ). The generator and discriminator are trained adversarially to generate SAR images.

[0116] Step 3: The generator is constructed using residual blocks and adopts an encoder-residual layer-decoder structure. The encoder compresses features through 3 layers of downsampling, and the decoder restores the resolution through 3 layers of upsampling, with 9 residual blocks embedded in the middle.

[0117] The specific method is:

[0118] 1) Input a raw SAR image of size (3×128×128) into the Res-Net generator, where C is the number of input channels, and H and W are the image spatial dimensions. After a 7×7 convolution with a stride of 1 and padding of 3, instance normalization, and ReLU activation, the output is a feature map of size (64×128×128), which enters the subsequent encoding process.

[0119] 2) The feature of size (64×128×128) is input to a 3*3 convolution with a stride of 2 and a padding of 1 for the first downsampling, and finally outputs a feature map of size (128×64×64).

[0120] 3) The feature of size (128×64×64) is input into a 3*3 convolution with a stride of 2 and a padding of 1 for a second downsampling, and finally outputs a feature map of size (256×32×32).

[0121] 4) Input the features of size (256×32×32) into a module containing 9 residual blocks, embed the PMA module (see step 4 for details) in the 2nd, 4th, and 6th residual blocks, keep the output dimension unchanged, and finally output a feature map of size (256×32×32).

[0122] 5) Decoder upsampling: The features of size (256×32×32) are upsampled for the first time through a 3×3 transposed convolution kernel with a stride of 2, and the output is a feature map of size (128×64×64).

[0123] 6) The features of size (128×64×64) are upsampled for the second time through a transposed convolution kernel with a step size of 2 and 3×3, and the output is a feature map of size (64×128×128).

[0124] 7) The features of size (64×128×128) are input to the output layer, and after passing through a 1×1 convolutional layer and a Tanh activation function, the final output is a result of size (1×128×128).

[0125] Step 4: Add a physical-guided multi-scale attention mechanism (PMA) to the 2nd, 4th, and 6th residual blocks of the generator to explicitly encode the pitch angle parameter θ, fuse the pitch angle parameter θ with multi-scale local features, and dynamically adjust the attention distribution of the feature map.

[0126] The specific method is:

[0127] 1) First, normalize the pitch angle parameter θ to [0,1].

[0128] 2) Then, the pitch angle parameter θ is mapped into three sets of 64-dimensional modulation weights W3, W5, and W7 through an MLP angle encoder consisting of two levels of fully connected layers.

[0129] 3) Capture features of different scales through parallel convolution 3×3, 5×5, 7×7 operations. The input F is (256×32×32) and the output C is s (F) is (64×32×32).

[0130] 4) Multiply the convolution of each scale by the corresponding weight channel by channel C s (F)⊙W s , where W s The three groups of 64-dimensional modulation weights W3, W5, and W7 obtained in step 4 (2) have the same dimension after channel-by-channel multiplication, and the output is (64×32×32).

[0131] 5) After adding the weighted features of size (64×32×32), the sigmoid activation function is used to generate the spatial attention map Attn(F,θ) of size (64×32×32) at the corresponding pitch angle:

[0132]

[0133] 6) Finally, the features of the key areas are enhanced through residual connections, and the feature map of size (256×32×32) is output through the original features + physical modulation features.

[0134]

[0135] Step 5: The discriminator adopts the patchGAN structure and uses a convolutional network to perform authenticity judgment on the local area of ​​the image.

[0136] The specific steps are as follows:

[0137] 1) Input (1×128×128) generated image.

[0138] 2) The image of size (1×128×128) is convolved with a 4*4 convolution kernel with a stride of 2, and the output is a feature of size (64×64×64).

[0139] 3) The feature of size (64×64×64) is convolved with a 4*4 convolution kernel with a stride of 2, and the output is a feature of size (128×32×32).

[0140] 4) The feature of size (128×32×32) is convolved with a 4*4 convolution kernel with a stride of 2, and the output is a feature of size (256×16×16).

[0141] 5) The feature of size (256×16×16) is convolved with a 4*4 convolution kernel with a stride of 1, and the output is a feature of size (512×16×16).

[0142] 6) The feature of size (512×16×16) is passed through a 4*4 convolution kernel with a step size of 1, and the output is a feature of size (1×16×16), which is then passed through a sigmoid activation function.

[0143] 7) Output a (16×16) probability matrix, where each element corresponds to the authenticity score of a 70×70 local area of ​​the input image.

[0144] 8) Conduct counter-loss Through the adversarial game between the generator G and the discriminator D, the generated image is forced to approach the real data distribution. The probability of the discriminator distinguishing the real image should be as close to 1 as possible. The probability of the discriminator for generating images should be as close to 0 as possible:

[0145]

[0146] Step 6: Perform fast Fourier transform (FFT) on the generated image and the real image to obtain the amplitude spectrum and phase spectrum of the real image and the generated image.

[0147] Step 7: Calculate the difference in attribute scattering feature center ASC between the real image and the generated image, and calculate the ASC loss function.

[0148] The specific method is:

[0149] 1) Given a generated image Compared with the real image I real , through differentiable frequency domain processing to extract the physical guided scattering characteristics, first perform fast Fourier transform FFT to obtain the frequency domain representation, and then convert to the frequency domain to analyze the scattering characteristics:

[0150]

[0151] in, is the amplitude spectrum, φ(I)∈R H×W For the phase spectrum, a differentiable frequency domain processor is designed to jointly optimize the frequency domain amplitude and phase characteristics to ensure the accuracy of the scattering center position and the authenticity of the intensity distribution.

[0152] 2) Based on the geometric theory of diffraction (GTD) and physical optics (PO), the target scattering in the high-frequency region can be approximated as the superposition of multiple ASCs, and the response E of each scattering center is i (f,φ) is described by the following formula:

[0153]

[0154] Among them E i (f,φ) is the complex scattering field of the i-th scattering center at the radar operating frequency f and azimuth φ, α i is a frequency-dependent factor reflecting the geometry of the scatterer, (x i ,y i ) represents the spatial position of the scattering center and determines the geometric structure of the SAR image. i is the scattering intensity, which is related to the material properties of the target. c is the center frequency, j is the imaginary unit, and c represents the speed of light.

[0155] 3) Through the learnable frequency mask M(f x ,f y ) to focus on the target sensitive frequency band, and pass the bandpass filter M(f x,f y ) retains key scattering information:

[0156]

[0157] Where (f x ,f y ) is the spatial frequency of the SAR image, The radial frequency represents the distance of the frequency component from the origin.

[0158] 4) Perform multi-scale feature extraction. The 3×3 convolution kernel size corresponds to the high-frequency local features, that is, the location of the scattering center (x i ,y i ); The convolution kernel size of 5×5 corresponds to the mid-frequency structural characteristics, that is, the length L of the scatterer; the convolution kernel size of 7×7 corresponds to the low-frequency global characteristics, that is, the scattering intensity A i . The Gaussian kernel function is:

[0159]

[0160] Among them G s (x,y) is the weight of the Gaussian convolution kernel at position (x,y), which is used to weight neighboring pixels, smooth the image, and extract features at a specific scale. τ is the standard deviation of the Gaussian function. A larger standard deviation results in a smoother convolution kernel and a lower sensitivity to small-scale features.

[0161] 5) Finally, the ASC loss function is defined as the L1 distance between the generated image and the real image feature tensor, because the L1 loss is robust to outliers and is more suitable for matching sparse scattered features:

[0162]

[0163] Step 8: Calculate the total loss function of the generated network, use the total loss function for backpropagation, update the network weights, and continue the next training. The total loss function of the generator is:

[0164]

[0165] Among them, λ is the weight coefficient, which is used to adjust different loss functions. adv =,λ cyc =,λ ASC =.

[0166] Step 9: Repeat steps 3 to 8 until the loss approaches zero and the network finally converges, saving the model parameters with the optimal training results.

[0167] Example 2

[0168] The present invention provides a SAR image generation network generation process based on physical guidance, which further includes the following steps based on Example 1:

[0169] Step 10: Resize and normalize the SAR image for testing to ensure that it effectively fits the input requirements of the model.

[0170] Step 11: Input the SAR image with a test pitch angle of 17 degrees and the target pitch angle parameter of 30 degrees into the trained model, and output the SAR image with a pitch angle of 30 degrees.

[0171] Step 12: The model output is post-processed to generate the final image. The output is linearly mapped from [-1.1] to [255, 255], converted to an 8-bit grayscale image, and histogram equalization is used to improve the visual effect.

[0172] The present invention has multiple significant advantages: the PMA module explicitly models the changes in scattering characteristics through an angle encoder, achieving edge attenuation compensation and maintaining the continuity of line scattering; the ASC loss constrains the distribution of scattering centers in the frequency domain, so that the generated image conforms to the laws of electromagnetic scattering. In addition, the present invention can be used to improve target detection performance. A transparency channel is added to the generated target slices to produce a target detection dataset. Its effect on improving detection performance has been verified under target detection frameworks such as YOLOv8 and DETR. At the same time, the invention does not require a fully paired training set, meets the needs of a small number of samples, and exhibits wider applicability and flexibility.

[0173] It should be noted that, based on the method of the present invention, a computer-readable storage medium can also be defined, wherein the storage medium stores an executable computer program, and when the program is executed by a processor, the above-mentioned physically guided SAR image generation method can be implemented. The computer-readable storage medium includes, but is not limited to, non-transitory storage media such as floppy disks, hard disks, optical disks, solid-state memories, USB flash drives, and mobile hard disks, and the program code stored therein can be read and executed by terminal devices, servers, or other computing devices. By logically encoding the steps of the method into computer-executable instructions, the storage medium can support the generation of cross-angle SAR images that conform to the physical laws of electromagnetic scattering, and implement core functions such as physically guided feature modulation, attribute scattering center feature constraints, and generation network training, providing a reliable technical carrier for solving problems such as insufficient SAR image data and the generation of images that do not conform to the physical laws.

[0174] Furthermore, an electronic device may be defined as comprising a processor and a memory, wherein the memory stores a computer program. When the processor executes the program, the aforementioned SAR image generation method can be implemented. The processor may be a computing unit such as a central processing unit (CPU), a graphics processing unit (GPU), an application-specific integrated circuit (ASIC), or a field-programmable gate array (FPGA), and is configured to parse and execute program instructions. The memory may include a random access memory (RAM), a read-only memory (ROM), or an auxiliary storage device (such as a hard disk or solid-state drive), and is configured to store program code, training data, intermediate results, and other information. When the electronic device is running, the processor executes the program in the memory to sequentially implement a physically guided multi-scale attention module (PMA) for dynamic modulation of scattering features, attribute scattering center loss (ASC Loss) for constraining frequency domain features, and end-to-end training of a generative adversarial network based on CycleGAN, thereby generating high-quality cross-angle SAR images. The electronic device can be applied to scenarios such as server clusters, image processing workstations, or embedded terminals. It supports efficient training with few samples, significantly improves the physical consistency and target detection performance of generated images, and provides an integrated solution for SAR image data enhancement and practical engineering applications.

[0175] The various embodiments in this specification are described in a progressive manner. Similar portions between the various embodiments can be referenced to each other. Each embodiment focuses on the differences between the other embodiments. In particular, the device embodiments are described briefly because they are generally similar to the method embodiments. For relevant portions, refer to the description of the method embodiments.

[0176] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for generating SAR images across angles based on a physical scattering model, comprising the following steps: S1. Standardize and preprocess the original SAR image, including resizing to a fixed resolution and noise suppression; S2, build a bidirectional generation architecture based on CycleGAN, which includes two generators (G A→B , G B→A ) and two discriminators (D A , D B ), supports adversarial training of incomplete paired data; S3, each generator includes an encoder module, a residual block group and a decoder module connected in sequence, the encoder module downsamples and compresses features through three layers of convolutional layers, the decoder module upsamples and restores the resolution through three layers of deconvolutional layers, and the residual block group consists of 9 series-connected skip-connected residual blocks; S4, embedding a physical guided multi-scale attention module PMA in the 2nd, 4th, and 6th layer residual blocks of the generator; S5. The discriminator uses the patchGAN structure to perform authenticity judgment on a 70×70 local area of ​​the input image and output a 30×30 probability matrix; S6. Perform fast Fourier transform on the generated image and the real image to obtain amplitude spectrum and phase spectrum of the real image and the generated image; S7. Calculate the attribute scattering feature center of the real image and the generated image in the frequency domain, and calculate the ASC loss function based on this; S8, perform backpropagation on the joint adversarial loss, cycle consistency loss and ASC loss to update the network weights; S9. Repeat S3 to S8 until the network converges and save the optimal model parameters.

2. The method for generating SAR images across angles based on physical scattering model guidance according to claim 1, characterized in that: The encoder module, residual block group, and decoder module of the generator in step S3 process the input image as follows: S31. Input the original SAR image of size C×H×W into the encoder module, and pass it through 7×7 convolution with stride 1 and padding 3, instance normalization, and ReLU activation in sequence, outputting a feature map of size 64×H×W. S32, downsample the feature map of size 64×H×W twice in sequence, and each downsampling passes through a 3×3 convolutional layer with a stride of 2 and a padding of 1, outputting feature maps of sizes 128×H / 2×W / 2 and 256×H / 4×W / 4 respectively; S33, input the features of size 256×H / 4×W / 4 into the residual block group, and pass through 9 residual blocks in sequence, wherein the physical guided multi-scale attention module (PMA) is embedded in the second, fourth, and sixth residual blocks, and the output dimension is maintained at a feature map of size 256×H / 4×W / 4; S34, upsampling the feature maps output by the residual block group twice in sequence, each upsampling passes through a transposed convolution layer with a stride of 2 and a kernel of 3×3, outputting feature maps of sizes 128×H / 2×W / 2 and 64×H×W respectively; S36. Input the 64×H×W feature map into the output layer, and generate a 1×H×W SAR image through a 1×1 convolutional layer and a Tanh activation function.

3. The method for generating SAR images across angles based on physical scattering model guidance according to claim 1, characterized in that: The physically guided multi-scale attention module PMA in step S4 dynamically adjusts the feature map attention distribution by the following steps: S41, normalizing the target pitch angle parameter to the interval [0, 1]; S42, through the MLP angle encoder composed of two levels of fully connected layers, the pitch angle parameters are mapped into three sets of 64-dimensional modulation weights W3, W5, and W7; S43, perform 3×3, 5×5, and 7×7 convolutions in parallel on the input feature map F of 256×H / 4×W / 4, and output multi-scale features C s (F) is 64×H / 4×W / 4; S44, multiply the convolution of each scale by the corresponding weight channel by channel C s (F)⊙W s , the dimension remains unchanged after channel-by-channel multiplication, and the output is 64×H / 4×W / 4; where W s are the three sets of 64-dimensional modulation weights W3, W5, and W7 obtained in step S42; S45. After adding the weighted features of size 64×H / 4×W / 4 and applying the Sigmoid activation function, a spatial attention map Attn(F,θ) of size 64×H / 4×W / 4 is generated at the corresponding pitch angle: S46, add the original feature map and the physically modulated feature map element by element through the residual connection, and keep the output dimension as a feature map of 256×H / 4×W / 4 4. The method for generating SAR images across angles based on physical scattering model guidance according to claim 1, characterized in that: The processing flow of the discriminator in step S5 includes the following steps: S51, input generated image, size is 1×H×W; S52, downsample the input image through three convolutional layers with a stride of 2 and a kernel of 4×4, and the output feature map dimensions are 64×H / 2×W / 2, 128×H / 4×W / 4, and 256×H / 8×W / 8 respectively; S53, perform two convolution operations with a step size of 1 and a kernel of 4×4 on the 256×H / 8×W / 8 feature map, outputting 512×H / 8×W / 8 and 1×H / 8×W / 8 feature maps respectively; S54. Process the 1×H / 8×W / 8 feature map through a sigmoid activation function to obtain an H / 8×W / 8 probability matrix, where each element corresponds to the authenticity score of a 70×70 local region in the input image. S55, calculating the adversarial loss based on the probability matrix The adversarial loss is defined by the following formula: in, The probability of the discriminator distinguishing the real image should be as close to 1 as possible. The discriminator's probability of distinguishing the generated image should be as close to 0 as possible.

5. The method for generating SAR images across angles based on physical scattering model guidance according to claim 1, characterized in that: The attribute scattering center ASC loss in step S7 is calculated by the following steps: S71, generate image Compared with the real image I real Perform fast Fourier transform (FFT) respectively to obtain the corresponding amplitude spectrum and phase spectrum φ(I); S72. Based on the geometric diffraction theory and physical optics model, define the complex scattering field E of the target scattering center. i (f,φ) is: Among them, α i is a frequency-dependent factor reflecting the geometry of the scatterer, (x i ,y i ) represents the spatial position of the scattering center, A i is the scattering intensity, which is related to the material properties of the target. c is the center frequency, j is the imaginary unit, and c represents the speed of light; S73, through the learnable bandpass filter M(f x ,f y ) retains the key scattering frequency band of the normalized frequency interval [0.1,0.4], defined as: Where (f x , f y) is the spatial frequency of the SAR image, is the radial frequency, which represents the distance of the frequency component from the origin; S74. Perform multi-scale Gaussian convolution on the filtered amplitude spectrum and phase spectrum to extract the physically guided scattering characteristics: A 3×3 Gaussian kernel is used to extract high-frequency local features, corresponding to the scattering center position (x i ,y i ), standard deviation σ3 = 0.5; A 5 × 5 Gaussian kernel is used to extract the intermediate frequency structural features, corresponding to the scatterer length L, with a standard deviation σ5 = 5 / 6; A 7×7 Gaussian kernel is used to extract low-frequency global features, corresponding to the scattering intensity A i , standard deviation σ7=7 / 6; S75. Concatenate multi-scale features into attribute scattering feature tensors. The formula is: Among them G s is a Gaussian convolution kernel of size s×s, is the amplitude spectrum of the image, φ(I) is the phase spectrum of the image; S76. Calculate the ASC feature L1 distance loss between the generated image and the real image. The calculation formula is: where λ ASC is the weight coefficient, To generate the ASC features of the image, is the ASC feature of the real image.

6. The method for generating SAR images across angles based on physical scattering model guidance according to claim 1, characterized in that: The total loss function calculation formula of the generator in step S8 is: Among them, λ is the weight coefficient, which is used to adjust different loss functions. is the cycle consistency loss, defined as: Among them I s is the source domain image, G is the generator from the source domain to the target domain, and G' is the generator from the target domain to the source domain.

7. The method for generating SAR images across angles based on physical scattering model guidance according to any one of claims 1 to 6, characterized in that: The following steps are also included: S10, resizing and normalizing the test SAR image to meet the model input requirements; S11, input the processed test image and target pitch angle parameters into the trained generation network, and output the SAR image of the target angle; S12. Post-process the output results, including linearly mapping the pixel values ​​from [-1, 1] to [255, 255], converting them into 8-bit grayscale images, and enhancing the visual effects through histogram equalization.

8. A SAR image cross-angle generation system guided by a physical scattering model, comprising the following modules: a preprocessing module configured to perform standardized preprocessing on the raw SAR image, including resizing to a fixed resolution and noise suppression; The adversarial network building module is configured to build a two-way generative adversarial network based on CycleGAN, which contains two generators (G A→B , G B→A ) and two discriminators (D A , D B ); The generator module adopts an encoder-residual layer-decoder structure. The encoder compresses features through 3 layers of downsampling, and the decoder restores the resolution through 3 layers of upsampling. 9 residual blocks are embedded in the middle. The PMA embedding module is configured to embed the physically guided multi-scale attention module in the 2nd, 4th, and 6th layer residual blocks of the generator; The discriminator module uses the patchGAN structure to perform authenticity discrimination on a 70×70 local area of ​​the input image and output a 30×30 probability matrix; An FFT processing module configured to perform a fast Fourier transform on the generated image and the real image to extract an amplitude spectrum and a phase spectrum; An ASC loss calculation module is configured to calculate the attribute scattering center (ASC) of the real image and the generated image in the frequency domain and generate an ASC loss function; Backpropagation module, configured to perform backpropagation on the joint adversarial loss, cycle consistency loss, and ASC loss to update the network weights; The model saving module is configured to save the optimal model parameters after the network converges.

9. The SAR image cross-angle generation system based on physical scattering model guidance according to claim 8, characterized in that: Also includes A test preprocessing module configured to resize and normalize the test SAR images to fit the input requirements of the generative network; A parameter input module is configured to concatenate the processed test image and target pitch angle parameters into a multi-dimensional tensor and input it into the trained generative network; The post-processing module is configured to perform the following processing on the image output by the generated network: linearly map the pixel value from [-1, 1] to [255, 255]; convert the floating-point data into an 8-bit unsigned integer grayscale image; and enhance the image contrast through the histogram equalization algorithm.

10. A computer-readable storage medium, characterized in that A computer program for executing the method according to any one of claims 1 to 6 or claim 7 is stored.

Images