Data-driven SAR image generation method fusing target electromagnetic scattering mechanism

Abstract

The invention discloses a data-driven SAR image generation method fusing a target electromagnetic scattering mechanism, and belongs to the technical field of synthetic aperture radar image generation. The method comprises the following steps: firstly, jointly resolving the position and intensity distribution of a target scattering center from a noise vector and a category label through a deformable convolutional network, generating a scattering characteristic thermodynamic diagram, and embedding the scattering characteristic thermodynamic diagram into a U-Net decoding process; secondly, comparing radar sectional area statistical characteristics and scattering center space distribution differences of the generated image and the real data through a pre-training scattering feature extractor; and performing joint optimization on double physical constraints of adversarial loss, category classification loss and scattering center KL divergence-radar sectional area error to realize collaborative optimization of a scattering mechanism and data driving. The SAR image generated by the method has pixel-level fidelity and electromagnetic physical consistency at the same time, and compared with a traditional pure data driving generation method, the method has the advantages that the rationality of an SAR target scattering structure is remarkably improved, and the core problem of insufficient physical interpretability of the generated image can be solved.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular to the field of synthetic aperture radar (SAR) image generation technology. Background Art

[0002] SAR systems can provide high-resolution, all-weather, all-day ground images, and play an important role in military reconnaissance. SAR target detection, as a key link in military reconnaissance, plays an irreplaceable role in identifying potential threats and analyzing battlefield situations. The premise of accurate target recognition is a large amount of image data. However, as of now, the use of traditional methods to obtain a large amount of clear SAR image data faces many challenges.

[0003] With the development of deep learning, a new image generation technique, generative adversarial networks (GANs), has provided new insights and approaches to addressing these challenges. By training a generator (GAN) and a discriminator (Discriminator) in a competitive manner, GANs automatically learn the underlying data distribution and generate synthetic data that closely resembles the real data distribution. GANs have demonstrated significant potential in the field of SAR image generation. (For details, see "I. Goodfellow, J. Pouget-Abadie et al., "Generative adversarial nets." Advances in neural information processing systems, vol. 27, pp. 2672–2680, 2014.") The generator attempts to produce realistic SAR images to deceive the discriminator, which determines whether the input image is real or fake. As training progresses, the generator gradually learns to generate increasingly detailed SAR images that closely resemble the training data distribution. Compared to traditional methods that manually collect and annotate data, GAN-based image generation methods significantly increase data acquisition speed and overcome the sample size limitations of traditional methods. For SAR image generation, some scholars have proposed a DH-GAN generation method, which is specifically used to generate simulated SAR images to enhance the training effect of CNN. For details, see the literature "Oghim S, Kim Y, Bang H, et al. SAR image generation method using DH-GAN for automatic target recognition [J]. Sensors, 2024, 24(2): 670." The DH-GAN model can generate simulated images with high-frequency characteristics similar to real SAR images by introducing a dual discriminator structure and a high-pass filter (HPF). Specifically, the model adds a discriminator that focuses on high-frequency components on the basis of the traditional generative adversarial network. The power spectral density (PSD) analysis verifies that the generated images are highly consistent with the real SAR images in the high-frequency region.To improve the quality of generated SAR images, some scholars proposed the attention-based IAM-ACGAN (IAM-ACGAN). For details, see the paper "Li J, Wei S, Hu Y, et al. IAM-ACGAN: A High-Accuracy Approach for SAR Image Augmentation [C] / / IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2024: 9782-9785." This network structure significantly improves the generation quality and classification accuracy of SAR images by introducing the channel attention (CA) and spatial attention (SA) mechanisms. It outperforms traditional ACGAN and Re-ACGAN methods in multiple evaluation metrics (such as Fréchet Inception Distance and Inception Score). In addition to the GANs generation model, the diffusion model has also made significant progress in generating high-quality synthetic data, especially in the field of image generation. For details, see the literature "HoJ, Jain A, Abbeel P. Denoising diffusion probabilistic models [J]. Advances inneural information processing systems, 2020, 33: 6840-6851." Some scholars have used DDPM to generate SAR image data. For details, see the literature "Qosja D, Wagner S, O'Hagan D. SAR image synthesis with diffusion models [C] / / 2024 IEEE Radar Conference (RadarConf24). IEEE, 2024: 1-6." This generation model uses linear noise scheduling by training large-scale MSTAR images to accurately reconstruct the SAR target shape and background.

[0004] While advanced technologies such as generative adversarial networks (GANs) and diffusion models have achieved remarkable results in generating detailed SAR images, significantly improving the quality and efficiency of image generation and, to a certain extent, alleviating the difficulties faced by traditional methods in acquiring large-scale SAR image data, it is important to note that SAR images differ fundamentally from optical images. SAR image generation cannot simply focus on visual similarity but must also fully consider its conformance to electromagnetic scattering characteristics. Given the shortcomings of existing methods in considering electromagnetic scattering characteristics, and to further improve the accuracy and practicality of SAR image generation, this paper proposes a data-driven SAR image generation method that incorporates the target's electromagnetic scattering mechanism. Summary of the Invention

[0005] This invention belongs to the field of radar technology and discloses a SAR image generative adversarial network method based on physical scattering mechanism constraints. This method generates realistic and clear SAR images based on scattering mechanism and data-driven methods. In this method, a conditional generative adversarial network framework that integrates scattering mechanism is proposed. First, a scattering center prediction network is trained using real SAR images and extracted scattering center images, which serves as a pre-trained scattering feature extractor in the subsequent GAN network. The scattering center prediction module is used in the generator to jointly calculate the position and intensity distribution of target scattering centers from noise vectors and class labels using a deformable convolutional network. A scattering characteristic heat map is generated and embedded in the U-Net decoding process. Secondly, a dual-path discriminator is designed, adding a physical constraint channel in addition to the traditional visual discrimination channel. The pre-trained scattering feature extractor compares the radar cross-section statistical characteristics and scattering center spatial distribution differences between the generated image and real data. Finally, a hybrid loss function is constructed to jointly optimize the adversarial loss, the class classification loss, and the dual physical constraints of the scattering center KL divergence-radar cross-section error, achieving collaborative optimization of the scattering mechanism and data-driven methods.

[0006] In order to facilitate the description of the present invention, the following terms are first defined:

[0007] Definition 1. Synthetic Aperture Radar (SAR)

[0008] Synthetic Aperture Radar (SAR) is a radar imaging technology that uses the relative motion of the radar antenna and the target to create a large virtual antenna aperture, achieving high-resolution imaging in azimuth. It then transmits broadband signals and uses pulse compression technology to achieve high-resolution imaging in range, ultimately achieving two-dimensional, high-precision imaging of the observed target. For details, see the document "Principles of Synthetic Aperture Radar Imaging," edited by Pi Yiming et al. and published by the University of Electronic Science and Technology of China Press.

[0009] Definition 2. Generative Adversarial Networks (GANs)

[0010] A machine learning model based on an adversarial training framework consists of two deep neural networks: a generator and a discriminator. The generator maps random noise into synthetic data by learning the underlying probability distribution of the target dataset. The discriminator uses a binary classification task to assess the probability that the input data originates from the true distribution or the generator. Both models alternately optimize to minimize the generator's adversarial loss function (deceiving the discriminator) and maximize the discriminator's ability to discriminate between authenticity and authenticity, ultimately ensuring that the generator outputs high-quality synthetic data with statistical properties close to those of real data. For details, see "I. Goodfellow, J. Pouget-Abadie et al., "Generative adversarial nets." Advances in neural information processing systems, vol. 27, pp. 2672–2680, 2014."

[0011] Definition 3. SAR image scattering center

[0012] In synthetic aperture radar (SAR) imaging, these localized strong reflection points characterize the electromagnetic scattering characteristics of a target. Their physical origin is either geometric discontinuities (such as edges, corners, and cavities) or regions with abrupt changes in the dielectric properties of the target. Scattering centers appear as high-energy focal points in SAR images. Their spatial distribution, amplitude, and phase characteristics can be characterized using parameterized models (such as the attributed scattering center model). For details, see "Akyildiz Y, Moses RL. Scattering center model for SAR imagery [C] / / SAR image analysis, modeling, and techniques II. SPIE, 1999, 3869: 76-85."

[0013] Definition 4: KL Divergence

[0014] An asymmetric information-theoretic measure of the difference between two probability distributions, defined as the expected value of the logarithmic difference between the true probability distribution P(x) and the approximation Q(x). For details, see "Hershey JR, Olsen PA. Approximating the Kullback-Leibler divergence between Gaussian mixture models[C] / / 2007IEEE International Conference on Acoustics, Speech and SignalProcessing-ICASSP'07.IEEE, 2007, 4:IV-317-IV-320."

[0015] The technical solution of the present invention is: a data-driven SAR image generation method integrating the target electromagnetic scattering mechanism, the method comprising the following steps:

[0016] Step 1: Prepare SAR image dataset;

[0017] Collect a SAR image dataset covering a variety of target categories; all images must be unified into a single-channel grayscale format and adjusted to a fixed resolution; to generate scattering center labels, extract the key scattering center positions and intensities of the target from each SAR image and generate a corresponding scattering center heat map; let the central wavenumber at the time of imaging be k0; the wavenumber domain bandwidth at the time of imaging be Δk; and the single-station sweep angle width at the time of imaging be Δφ; and use the two-dimensional scattering center spread function h(x,y) in each calculation;

[0018] Step 2: Construct a generative adversarial network architecture that integrates the scattering mechanism;

[0019] The generative adversarial network consists of a generator and a discriminator. The generator includes a scattering center prediction module and a U-Net decoder. The input of the scattering center prediction module is defined as a noise vector With category labels Category labels are encoded using one-hot encoding, d z represents the dimension of noise, K is the number of categories;

[0020] First, concatenate z and c and map them into initial features through the fully connected layer in, represents the concatenation operation, which directly concatenates the category label vector c to the back of the noise vector z. FC(·) represents the data after the fully connected layer, and h0 represents the output data after the fully connected layer, i.e., the initial features. h, w, and C represent the length, width, and number of channels of the feature map, respectively.

[0021] Next, in the deformable convolution, the predicted offset is set to Δp, and the deformable convolution is implemented as h1 = DeformConv(h0, Δp), where DeformConv(h0, Δp) represents the convolution of the initial feature h0 with the offset Δp; then the deconvolution is gradually upsampled to the target resolution to generate the scattering center heat map. Where UpSample(·) represents the upsampling operation, ReLU(·) represents the ReLU activation function, H and W represent the height and width of the heat map respectively;

[0022] In addition, the input noise z is expanded into spatial features through the fully connected layer

[0023] Reshape(·) represents the deformation of the data size and changes its dimension; the spatial feature F z Then combine it with the heat map H to get F fused =F z ⊙Expand(H), where ⊙ is channel-by-channel multiplication. Expand(H) expands the heat map to the same size as F z Same number of channels, F fused represents the fused features, where C represents the number of channels;

[0024] Finally, the fused feature F fused Entering the U-Net decoder, the decoder first performs encoding. The encoding process includes a 4-layer structure, each layer contains 2 3*3 convolutions and 2 ReLU activation functions, and there is a 2*2 maximum pooling layer after the 4-layer structure; then decoding is performed. The decoding also has a four-layer structure, each layer contains a 2*2 upsampling convolution, 2 3*3 convolutions and 2 ReLU activation functions to generate the final SAR image Sigmoid(·) represents the sigmoid activation function, and G(z,c) represents the SAR image generated by the generator;

[0025] The discriminator consists of a visual discrimination channel and a physical constraint channel; in the visual discrimination channel, the input is a real image x or a generated image G(z,c), and multi-scale features are extracted through the convolution layer CNN(·) represents the convolutional structure; then the full connection layer outputs the discriminant result as D vis (x) = Sigmoid(FC(f vis ))∈[0,1];

[0026] The physical constraint channel first involves a scattering feature extractor. Extract the scattering center distribution features from the image and then use KL divergence The local difference between the generated and real image scattering distribution is measured and expressed as radar cross section error. Constrained global statistical properties, where D KL (·) To find the KL divergence, is to find the mathematical expectation;

[0027] Step 3: Determine the loss function of the generative adversarial network;

[0028] The loss function of the generative adversarial network consists of three parts: adversarial loss, classification loss, and physical constraint loss. It jointly optimizes the generator's generation ability and physical consistency through a weighted sum form.

[0029] Step 4: Network training;

[0030] The training process is divided into two stages: pre-training the scattering feature extractor and joint optimization of the generative adversarial network. The pre-training stage is to provide reliable scattering feature extraction capabilities for physically constrained channels.

[0031] Step 5: Use the trained generative adversarial network to generate images.

[0032] Furthermore, the specific method for extracting the key scattering center position and intensity information of the target from each SAR image in step 1 is:

[0033] Find the point with the largest amplitude from the original echo, whose amplitude is A max , the corresponding position is (x i ,y i ), where x i is the horizontal coordinate of the image, y i is the vertical coordinate of the image, i represents different iteration numbers, assuming that the number of operation iterations is N;

[0034] Using the formula The calculation moves the reference function center to (x i ,y i ) after the point scattering function value h(x i ,y i ), where j represents an imaginary number, sinc is a function, specifically sinc = sin(πx) / (πx); gx represents the difference between the horizontal coordinates of each point on the graph and the point with the maximum amplitude, and gy represents the difference between the vertical coordinates of each point on the graph and the point with the maximum amplitude;

[0035] Using the formula (Img2D) n+1 =(Img2D) n -A n ·h(xx n ,yy n), n = 1, ..., N, subtract the above calculated signal from the original data to regenerate the image residual, where (Img2D) n Represents the image with the nth iteration; A n Indicates the amplitude value of the point with the largest amplitude in the original echo in the nth iteration, (x n ,y n ) is the position of the nth iteration

[0036] Then continue to find the point with the largest amplitude in the remaining image, and repeat to get the image difference until all points are found or the signal is equal to the noise level, then stop the iteration, and use the amplitude A obtained in the iterative process max and the corresponding position information (x i ,y i ) constructs the corresponding scattering center heat map; then, the image data is normalized, the pixel values ​​are scaled to the interval [0,1], and data enhancement operations such as random rotation and horizontal flipping are applied to improve the generalization ability of the model; finally, the dataset is divided into training set and validation set in proportion, where the training set is used for model parameter optimization and the validation set is used to monitor the training process.

[0037] Furthermore, the specific method of step 3 is:

[0038] Furthermore, using the formula As an adversarial loss, it is used to measure the distribution difference between the generated image and the real image in the visual feature space. Its mathematical form is the expected difference between the generated image and the real image at the discriminator output, forcing the image output by the generator to approach the real data distribution in terms of texture, contrast and other visual aspects. vis (x) is the discrimination result output by the fully connected layer, and G(z,c) is the generated result, which is the mathematical expectation.

[0039] Using the formula As the classification loss, the classifier C predicts the category of the generated image to ensure that the generated image is consistent with the category of the input label c to avoid target confusion;

[0040] The physical constraint loss is implemented by the scattering feature extractor F, which includes two parts: local scattering distribution difference and global statistical characteristic error:

[0041] Part of the formula Calculate the relative entropy between the scattering center heat map distribution of the generated image and the real image, forcing the generator to be consistent with the real target in the position and intensity distribution of the scattering center, D KL (·) is to find the KL divergence;

[0042] The other part uses the formula As the scattering coefficient loss function, it constrains the overall scattering intensity statistical characteristics of the generated image and the real image;

[0043] Using the formula As the total loss function, λ1 and λ2 are hyperparameters that control the optimization weights of classification accuracy and physical constraints respectively, and their values ​​are usually determined by cross-validation; the generator generates physically reasonable images by minimizing the total loss, and the discriminator maximizes the difference in authenticity scores and applies a gradient penalty term Improve the stability of discrimination, is the linear interpolation sample of the real image and the generated image, λ gp is a hyperparameter, For the discriminator D The gradient of , ||·||2 is the 2-norm.

[0044] Furthermore, the specific method of step 4 is:

[0045] Furthermore, we first process the real SAR image dataset and generate the scattering center heat map corresponding to each image as the supervision label; the scattering feature extractor F adopts the U-Net structure, with the input as the SAR image and the output as the predicted scattering heat map, and uses the mean square error loss to extract the scattering center heat map. Optimize, where is the heat map label of the i-th image, F(x i ) is the scattering heatmap of the band predicted for the i-th image; the Adam optimizer is used for training, with a learning rate of l0 and a batch size of b0. The early stopping method is used to terminate the training when the validation set loss no longer decreases, ensuring that the model converges and is not overfitting;

[0046] In the joint training phase, the parameters of the pre-trained scattering feature extractor are fixed, and the generator and discriminator are initialized for adversarial training. In each round of training, the discriminator and generator are updated alternately. The specific method is as follows:

[0047] First, the discriminator is updated, the generator parameters are frozen, a batch of images x is sampled from the real data distribution, the generator generates fake images G(z,c), and the gradient penalty is calculated in is the linear interpolation sample of the real image and the generated image, and then the loss is calculated Finally, back propagation and parameter update are performed, where To freeze the generator parameters, only the discriminator's adversarial loss function is used.

[0048] Next, the generator is updated, the discriminator parameters are frozen, and the total loss is calculated Finally, back propagation and parameter update are performed, where To freeze the discriminator parameters, only the adversarial loss function of the generator is used. After a certain number of iterations, the network training is complete, and the generated SAR image of the corresponding category can be obtained by inputting the category.

[0049] This paper proposes a physically constrained and data-driven generative adversarial network for synthesizing high-fidelity, physically interpretable, and class-controllable SAR imagery. This approach ensures visual fidelity through a Wasserstein Generative Adversarial Network (WGAN-GP) gradient penalty. Consistency of electromagnetic properties is constrained by combining the KL divergence of scattering features with radar cross-section (RCS) error. Furthermore, a cross-entropy loss from a pre-trained classifier is introduced for precise semantic control, significantly improving the efficiency and reliability of SAR image generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a data-driven generation network structure diagram of the fusion target electromagnetic scattering mechanism of the present invention.

[0051] Figure 2 This is a result graph generated by the data-driven generation network of the fusion target electromagnetic scattering mechanism of the present invention. DETAILED DESCRIPTION

[0052] In order to further illustrate the above technical solution, a case study is provided below for reference. The specific implementation plan is outlined as follows:

[0053] Step 1: Prepare SAR image dataset:

[0054] A SAR image dataset was collected, with 360 images for each target type, representing three categories. All images were converted to a single-channel grayscale format and resized to a fixed resolution of 128 x 128. To generate scattering center labels, the locations and intensities of key scattering centers of the target were extracted from each SAR image, and a corresponding scattering center heat map was generated. The central beam during imaging is denoted by k0; the wavenumber domain bandwidth during imaging is Δk; and the single-station sweep angle width during imaging is denoted by Δφ. Each calculation uses the two-dimensional scattering center spread function h(x, y). The specific steps are described below.

[0055] Find the point with the largest amplitude from the original echo, whose amplitude is A max , the corresponding position is (x i ,y i ), where x i is the horizontal coordinate of the image, y i is the vertical coordinate of the image, i represents different iteration numbers, assuming the number of iterations is N.

[0056] Using the formula The calculation moves the reference function center to (x i ,y i) after the point scattering function value, where j represents an imaginary number, sinc is a function, specifically sinc = sin(πx) / (πx); gx represents the difference between the horizontal coordinates of each point on the graph and the point with the maximum amplitude, and gy represents the difference between the vertical coordinates of each point on the graph and the point with the maximum amplitude.

[0057] Using the formula (Img2D) n+1 =(Img2D) n -A n ·h(xx n ,yy n ), n=1,…,N, subtract the above calculated signal from the original data to regenerate the image residual, where (Img2D) n Represents the image with the nth iteration. n Indicates the amplitude value of the point with the largest amplitude in the original echo in the nth iteration, (x n ,y n ) is the position of the nth iteration

[0058] Then continue to find the point with the largest amplitude in the remaining image, and repeat to get the image difference until all points are found or the signal is equal to the noise level, then stop the iteration, and use the amplitude A obtained in the iterative process max and the corresponding position information (x i ,y i ) to construct the corresponding scattering center heatmap. Subsequently, the image data is normalized, pixel values ​​are scaled to the range [0, 1], and data augmentation operations such as random rotation and horizontal flipping are applied to improve model generalization. Finally, the dataset is split into a training set and a validation set in a ratio of 7:3. The training set is used for model parameter optimization, and the validation set is used to monitor the training process.

[0059] Step 2: Constructing a Generative Adversarial Network Architecture Integrating the Scattering Mechanism:

[0060] First, the generator consists of two parts: the scattering center prediction module and the U-Net decoder. The first part of the scattering center prediction module input is set as the noise vector That is, a random vector of 128 dimensions, and it conforms to Gaussian distribution sampling Since the data set is divided into 3 categories, the category labels are set Category labels are one-hot encoded.

[0061] First, z and c are concatenated and mapped to the initial feature h0 through a fully connected layer. In the deformable convolution, the predicted offset is set to Δp, and the deformable convolution is implemented as h1 = DeformConv(h0, Δp). Then, deconvolution is gradually upsampled to the target resolution to generate a scattering center heat map. Among them, UpSample(·) represents the upsampling operation, and ReLU(·) represents the ReLU activation function.

[0062] In addition, the input noise z is expanded into spatial features F through the fully connected layer z , where Reshape(·) represents the deformation of the data size and changes its dimension. Spatial feature F z Then combine it with the heat map H to get F fused =F z ⊙Expand(H), where ⊙ is channel-by-channel multiplication. Expand(H) expands the heat map to the same size as F z Same number of channels, F fused Represents the fused features.

[0063] Finally, the fused feature F fused Entering the U-Net decoder, the decoder first performs encoding, and the features pass through a 4-layer structure, each layer contains 2 3*3 convolutions and 2 ReLU activation functions, and finally there is a 2*2 maximum pooling layer. The decoding also has a four-layer structure, each layer contains a 2*2 upsampling convolution, 2 3*3 convolutions and 2 ReLU activation functions to generate the final SAR image. Sigmoid(·) represents the sigmoid activation function, and G(z,c) represents the SAR image generated by the generator.

[0064] The discriminator consists of a visual discrimination channel and a physical constraint channel. In the first part of the visual discrimination channel, the input is the real image x or the generated image G(z,c), and the multi-scale features are extracted through the convolution layer to obtain Then the judgment result is output through the fully connected layer as D vis (x) = Sigmoid(FC(f vis ))∈[0,1].

[0065] The second part is the pre-trained scattering feature extractor F(x) = U-Net phy (x). Extract the scattering center distribution features from the image through KL divergence The local difference between the generated and real image scattering distribution is measured and expressed as radar cross section error. Constrained global statistical properties, where D KL (·) To find the KL divergence, It is to find the mathematical expectation.

[0066] Step 3: Design of loss function integrating scattering mechanism

[0067] The hybrid loss function consists of three parts: adversarial loss, classification loss, and physical constraint loss. It jointly optimizes the generator's generation ability and physical consistency through a weighted sum form.

[0068] Using the formula As an adversarial loss, it measures the distribution difference between generated and real images in the visual feature space. Its mathematical form is the expected difference between the generated and real images at the discriminator output, forcing the generator's output to approximate the real data distribution in terms of texture, contrast, and other visual aspects.

[0069] Using the formula As the category classification loss, the classifier C predicts the category of the generated image to ensure that the generated image is consistent with the category of the input label c to avoid target confusion.

[0070] The physical constraint loss is implemented through the scattering feature extractor F, which includes two parts: local scattering distribution difference and global statistical characteristic error.

[0071] Using the formula The relative entropy between the scattering center heat map distributions of the generated image and the real image is calculated, forcing the generator to be consistent with the real target in the position and intensity distribution of the scattering center.

[0072] Using the formula The overall scattering intensity statistical characteristics of the constrained generated image and the real image, D KL (·) is used to find the KL divergence.

[0073] Using the formula As the total loss function, λ1 and λ2 are hyperparameters, set to 1 and 0.5 respectively, which are set to control the optimization weights of classification accuracy and physical constraints respectively. Their values ​​are usually determined by cross-validation. The generator generates physically reasonable images by minimizing the total loss, while the discriminator maximizes the difference in realism scores and applies a gradient penalty term. Improve the stability of discrimination, λ gp is a hyperparameter, For the discriminator D The gradient of ||·||2 is the 2-norm, λ gp Set to 10.

[0074] Step 4: Network training

[0075] The training process is divided into two stages: pre-training the scattering feature extractor and joint optimization of the generative adversarial network. The pre-training stage is to provide reliable scattering feature extraction capabilities for the physically constrained channel. First, the real SAR image dataset is processed to generate the scattering center heat map corresponding to each image as the supervision label. The scattering feature extractor F adopts the U-Net structure, with the input being the SAR image and the output being the predicted scattering heat map. The mean square error loss is used to calculate the scattering center heat map. Optimize, where is the heat map label of the i-th image, F(x i ) is the predicted scattering heatmap of the band for the i-th image. The Adam optimizer was used for training, with a learning rate of 0.0001 and a batch size of 16. Early stopping was used to terminate training when the KL divergence fluctuation was <1% after 100 consecutive epochs, ensuring model convergence without overfitting.

[0076] During the joint training phase, the pre-trained scattering feature extractor parameters are fixed, and the generator and discriminator are initialized for adversarial training. In each round of training, the discriminator and generator are updated alternately. The generator and discriminator use the Adam optimizer with a learning rate of 0.0001 and momentum parameters of β1 = 0.5 and β2 = 0.999, respectively. The total number of training rounds is 1000.

[0077] First, the discriminator is updated, the generator parameters are frozen, a batch of images x is sampled from the real data distribution, the generator generates fake images G(z,c), and the gradient penalty is calculated Then calculate the loss Finally, backpropagation and parameter update are performed.

[0078] Next, the generator is updated, the discriminator parameters are frozen, and the total loss is calculated Finally, backpropagation and parameter updates are performed, and training is terminated when the KL divergence fluctuation is <1% after 100 consecutive epochs. After network training is completed, the generated SAR image of the corresponding category can be obtained by inputting the category.

Claims

1. A data-driven SAR image generation method integrating target electromagnetic scattering mechanism, comprising the following steps: Step 1: Prepare SAR image dataset; Collect a SAR image dataset covering a variety of target categories; all images must be unified into a single-channel grayscale format and adjusted to a fixed resolution; to generate scattering center labels, extract the key scattering center positions and intensities of the target from each SAR image and generate a corresponding scattering center heat map; let the central wavenumber at the time of imaging be k0; the wavenumber domain bandwidth at the time of imaging be Δk; and the single-station sweep angle width at the time of imaging be Δφ; and use the two-dimensional scattering center spread function h(x,y) in each calculation; Step 2: Construct a generative adversarial network architecture that integrates the scattering mechanism; The generative adversarial network consists of a generator and a discriminator. The generator includes a scattering center prediction module and a U-Net decoder. The input of the scattering center prediction module is defined as a noise vector With category labels Category labels are encoded using one-hot encoding, d z represents the dimension of noise, K is the number of categories; First, concatenate z and c and map them into initial features through the fully connected layer in, represents the concatenation operation, which directly concatenates the category label vector c to the back of the noise vector z. FC(·) represents the data after the fully connected layer, and h0 represents the output data after the fully connected layer, i.e., the initial features. h, w, and C represent the length, width, and number of channels of the feature map, respectively. Next, in the deformable convolution, the predicted offset is set to Δp, and the deformable convolution is implemented as h1 = DeformConv(h0, Δp), where DeformConv(h0, Δp) represents the convolution of the initial feature h0 with the offset Δp; then the deconvolution is gradually upsampled to the target resolution to generate the scattering center heat map. Where UpSample(·) represents the upsampling operation, ReLU(·) represents the ReLU activation function, H and W represent the height and width of the heat map respectively; In addition, the input noise z is expanded into spatial features through the fully connected layer Reshape(·) represents the deformation of the data size and changes its dimension; the spatial feature F z Then combine it with the heat map H to get F fused =F z ⊙Expand(H), where ⊙ is channel-by-channel multiplication. Expand(H) expands the heat map to the same size as F z Same number of channels, F fused represents the fused features, where C represents the number of channels; Finally, the fused feature F fused Entering the U-Net decoder, the decoder first performs encoding. The encoding process includes a 4-layer structure, each layer contains 2 3*3 convolutions and 2 ReLU activation functions, and there is a 2*2 maximum pooling layer after the 4-layer structure; then decoding is performed. The decoding also has a four-layer structure, each layer contains a 2*2 upsampling convolution, 2 3*3 convolutions and 2 ReLU activation functions to generate the final SAR image Sigmoid(·) represents the sigmoid activation function, and G(z,c) represents the SAR image generated by the generator; The discriminator consists of a visual discrimination channel and a physical constraint channel; in the visual discrimination channel, the input is a real image x or a generated image G(z,c), and multi-scale features are extracted through the convolution layer CNN(·) represents the convolutional structure; then the full connection layer outputs the discriminant result as D vis (x) = Sigmoid(FC(f vis ))∈[0,1]; The physical constraint channel first involves a scattering feature extractor. Extract the scattering center distribution features from the image and then use KL divergence The local difference between the generated and real image scattering distribution is measured and expressed as radar cross section error. Constrained global statistical properties, where D KL (·) To find the KL divergence, is to find the mathematical expectation; Step 3: Determine the loss function of the generative adversarial network; The loss function of the generative adversarial network consists of three parts: adversarial loss, classification loss, and physical constraint loss. It jointly optimizes the generator's generation ability and physical consistency through a weighted sum form. Step 4: Network training; The training process is divided into two stages: pre-training the scattering feature extractor and joint optimization of the generative adversarial network. The pre-training stage is to provide reliable scattering feature extraction capabilities for physically constrained channels. Step 5: Use the trained generative adversarial network to generate images.

2. The data-driven SAR image generation method integrating target electromagnetic scattering mechanism according to claim 1, characterized in that: The specific method for extracting the key scattering center position and intensity information of the target from each SAR image in step 1 is: Find the point with the largest amplitude from the original echo, whose amplitude is A max , the corresponding position is (x i ,y i ), where x i is the horizontal coordinate of the image, y i is the vertical coordinate of the image, i represents different iteration numbers, assuming that the number of operation iterations is N; Using the formula The calculation moves the reference function center to (x i ,y i ) after the point scattering function value h(x i ,y i ), where j represents an imaginary number, sinc is a function, specifically sinc = sin(πx) / (πx); gx represents the difference between the horizontal coordinates of each point on the graph and the point with the maximum amplitude, and gy represents the difference between the vertical coordinates of each point on the graph and the point with the maximum amplitude; Using the formula (Img2D) n+1 =(Img2D) n -A n ·h(xx n ,yy n ), n = 1, ..., N, subtract the above calculated signal from the original data to regenerate the image residual, where (Img2D) n Represents the image with the nth iteration; A n Indicates the amplitude value of the point with the largest amplitude in the original echo in the nth iteration, (x n ,y n ) is the position of the nth iteration Then continue to find the point with the largest amplitude in the remaining image, and repeat to get the image difference until all points are found or the signal is equal to the noise level, then stop the iteration, and use the amplitude A obtained in the iterative process max and the corresponding position information (x i ,y i ) construct the corresponding scattering center heat map; Subsequently, the image data is normalized, the pixel values ​​are scaled to the range [0,1], and data augmentation operations such as random rotation and horizontal flipping are applied to improve the generalization ability of the model; finally, the dataset is divided into training and validation sets in proportion, where the training set is used for model parameter optimization and the validation set is used to monitor the training process.

3. The data-driven SAR image generation method integrating target electromagnetic scattering mechanism according to claim 1, characterized in that: The specific method of step 3 is: Using the formula as a countermeasure to loss; It is used to measure the distribution difference between the generated image and the real image in the visual feature space; its mathematical form is the expected difference between the generated image and the real image at the discriminator output, forcing the image output by the generator to approach the real data distribution in terms of texture, contrast and other visual aspects; D vis (x) is the discrimination result output by the fully connected layer, G(z,c) is the generated result, which is the mathematical expectation; Using the formula As the classification loss, the classifier C predicts the category of the generated image to ensure that the generated image is consistent with the category of the input label c to avoid target confusion; The physical constraint loss is implemented by the scattering feature extractor F, which includes two parts: local scattering distribution difference and global statistical characteristic error: Part of the formula Calculate the relative entropy between the scattering center heat map distribution of the generated image and the real image, forcing the generator to be consistent with the real target in the position and intensity distribution of the scattering center, D KL (·) is to find the KL divergence; The other part uses the formula As the scattering coefficient loss function, it constrains the overall scattering intensity statistical characteristics of the generated image and the real image; Using the formula As the total loss function, λ1 and λ2 are hyperparameters that control the optimization weights of classification accuracy and physical constraints respectively, and their values ​​are usually determined by cross-validation; the generator generates physically reasonable images by minimizing the total loss, and the discriminator maximizes the difference in authenticity scores and applies a gradient penalty term Improve the stability of discrimination, is the linear interpolation sample of the real image and the generated image, λ gp is a hyperparameter, For the discriminator D The gradient of , ||·||2 is the 2-norm.

4. The data-driven SAR image generation method integrating target electromagnetic scattering mechanism according to claim 1, characterized in that: The specific method of step 4 is: First, the real SAR image dataset is processed to generate the scattering center heat map corresponding to each image as the supervision label; the scattering feature extractor F adopts the U-Net structure, the input is the SAR image, and the output is the predicted scattering heat map, which is obtained by the mean square error loss. Optimize, where is the heat map label of the i-th image, F(x i ) is the scattering heat map of the band predicted for the i-th image; The Adam optimizer is used for training, with a learning rate of l0 and a batch size of b0. Early stopping is used to terminate training when the validation set loss stops decreasing, ensuring that the model converges without overfitting. In the joint training phase, the pre-trained scattering feature extractor parameters are fixed, and the generator and discriminator are initialized for adversarial training. In each round of training, the discriminator and generator are updated alternately. The specific method is as follows: First, the discriminator is updated, the generator parameters are frozen, a batch of images x is sampled from the real data distribution, the generator generates fake images G(z,c), and the gradient penalty is calculated in is the linear interpolation sample of the real image and the generated image, and then the loss is calculated Finally, back propagation and parameter update are performed, where To freeze the generator parameters, only the discriminator's adversarial loss function is used; Next, the generator is updated, the discriminator parameters are frozen, and the total loss is calculated Finally, back propagation and parameter update are performed, where In order to freeze the discriminator parameters, only the adversarial loss function of the generator is used; after a certain number of times, the network training is completed, and the generated SAR image of the corresponding category can be obtained by inputting the category.

Images