A meta-learning based cross-system parameter adaptive sar image reconstruction method

Abstract

The application discloses a cross-system parameter adaptive SAR image reconstruction method based on meta learning. An imaging framework combining model-independent meta learning (MAML) and a deep complex convolutional fully convolutional network (Complex-CVFCN) is proposed. First, an 8-layer deep FCNN with an instance normalization layer and a global residual connection is constructed to utilize coupled convolution to deeply mine the phase features of echo real and imaginary parts. Second, a double-layer optimization mechanism of MAML is utilized to perform meta training on multiple bandwidth tasks to learn meta-initialization weights with high gradient sensitivity. Finally, when facing unknown bandwidth or new scene tasks, only a small amount of support set samples are required for 1-5 step gradient updates to realize rapid adaptation of network parameters, effectively correct the point spread function mismatch caused by system parameter drift, suppress sidelobe artifacts, and realize SAR image reconstruction under non-ideal and cross-system observation conditions.

Description

Technical Field

[0001] This invention relates to the intersection of radar imaging and deep learning technology, specifically to a high-quality and fast imaging method for synthetic aperture radar (SAR) sparse sampled data across scenes and system parameters using model-independent meta-learning (MAML) and deep complex fully convolutional neural networks (Complex-CVFCN). Background Technology

[0002] Synthetic Aperture Radar (SAR) possesses all-weather, all-day imaging capabilities and has wide-ranging applications in remote sensing mapping, disaster monitoring, and national defense reconnaissance. With the development of imaging technology, SAR images acquired under different platforms (spaceborne, airborne, and UAV-borne), different frequency bands (L, S, C, X, Ku, etc.), and different system parameters exhibit significant differences in resolution, bandwidth, and imaging algorithms. Effective bandwidth directly determines the range resolution of SAR; changes in bandwidth lead to significant alterations in the range-direction point spread function (PSF), thereby affecting image quality.

[0003] While traditional compressed sensing imaging methods possess some sparse reconstruction capabilities, they suffer from complex iterative computations, poor real-time performance, and high parameter sensitivity, making them unsuitable for real-time processing requirements in engineering. In recent years, deep learning has achieved significant success in solving inverse problems and has been rapidly applied to sparse SAR imaging, giving rise to physically inspired deep unfolding networks such as ISTA-Net and CISTA-Net. However, existing end-to-end deep neural network imaging operators face a fundamental challenge in generalization during practical industrial deployment and applications. Traditional deep learning imaging models (including those based on multilayer perceptrons (MLPs), ordinary single-layer convolutions, or lightweight networks) are typically task-specific, with their network parameters trained only under fixed radar system parameters (especially effective bandwidth) and a fixed sampling rate. When applied to entirely new imaging tasks or different radar configurations, the physical matching of network operators is disrupted, leading to decreased imaging resolution, sidelobe contamination, or aliasing artifacts, severely limiting the engineering application capabilities of deep learning SAR imaging models.

[0004] Meta-learning, as a new paradigm of "learning how to learn," enables models to extract transferable knowledge across multiple tasks and quickly adapt to new tasks with minimal gradient updates. MAML (Model-Independent Meta-Learning), one of the most representative meta-learning frameworks, has demonstrated good generalization capabilities in areas such as image classification and regression prediction, and has been initially applied to ISAR imaging.

[0005] However, existing meta-learning-based radar imaging methods (such as the ISAR imaging method disclosed in CN120610266A) employ L1-norm-based optimized unfolded networks, aiming to improve robustness to signal-to-noise ratio and loss rate, rather than addressing the PSF mismatch problem caused by changes in system parameters (especially effective bandwidth). Introducing MAML into end-to-end complex domain SAR imaging requires a network architecture adapted to the processing requirements of complex signals, and the drastic fluctuations in signal amplitude under different system parameters can interfere with the gradient updates of MAML—a problem that has not yet been effectively resolved.

[0006] Therefore, there is an urgent need to propose a SAR image reconstruction method that combines cross-system parameter generalization capability with rapid adaptation capability, which can effectively cope with the impact of radar system parameter changes (especially effective bandwidth changes) on imaging quality. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing deep learning SAR imaging networks, such as insufficient generalization ability across system parameters and inability to quickly adapt to new system parameters, and to provide a meta-learning-based cross-system parameter adaptive SAR image reconstruction method and apparatus. This method innovatively combines the MAML meta-learning framework with the Complex-CVFCN complex fully convolutional network, endowing the model with the ability to quickly adapt to changes in system parameters (such as effective bandwidth).

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A meta-learning-based cross-system parameter adaptive SAR image reconstruction method includes the following steps:

[0010] Step 1: Acquire raw SAR echo signals and construct a multi-task dataset. Obtain a full-bandwidth reference signal through simulation or field measurement. Crop the signal into rectangular windows with different effective bandwidths in the frequency domain to generate various imaging tasks with different range-direction point spread functions. Each task corresponds to different system parameter settings.

[0011] Step 2: Construct the Complex-CVFCN fully convolutional network. This network takes the original complex-domain echo signal as input and performs joint modeling of the real and imaginary parts through complex convolution operations, preserving complete amplitude and phase coupling information. The network uses the complex ReLU activation function, introduces an instance normalization layer after each convolutional layer, and constructs global residual connections to achieve end-to-end reconstruction from complex-domain echo to SAR image.

[0012] Step 3: Employ the MAML framework for two-layer optimization meta-training. In the inner loop, a small amount of gradient updates are performed on the support set for each imaging task to obtain task-specific parameters. In the outer loop, the meta-initialization parameters are updated using the query set for each task. This two-layer optimization mechanism enables the network to learn highly gradient-sensitive meta-initialization weights during multi-task training, providing cross-system parameter generalization capabilities.

[0013] Step 4: Perform fast adaptive inference under unknown system parameters. Load the meta-initialized weights obtained from meta-training, and perform 1-5 gradient updates using a very small number of support set samples to obtain an adapted network for the current system parameters. Use the adapted network to perform full-data inference and output high-quality SAR reconstructed images.

[0014] Furthermore, the frequency domain window clipping in step one specifically includes: transforming the time-domain range-compressed signal to the range frequency domain, constructing a rectangular window function in the frequency domain to clip the signal spectrum according to the target effective bandwidth, and inversely transforming the clipped spectrum back to the time domain to obtain the range-compressed signal of the target task.

[0015] The effective bandwidths corresponding to the different imaging tasks generated are different, resulting in different range resolutions for different imaging tasks.

[0016] Furthermore, the complex convolution operation in step two follows the complex multiplication rule, for the first... Complex convolution operation of layers, with input feature map as The convolution kernel is Output the real part of the feature map and the virtual part Calculate according to the following formulas respectively:

[0017]

[0018]

[0019] Where * represents convolution operation, and j is the imaginary unit.

[0020] Furthermore, the Complex-CVFCN complex fully convolutional network described in step two employs the complex ReLU activation function, which is defined as follows: Nonlinear transformations are performed on the real and imaginary parts respectively to preserve the phase coherence.

[0021] The instance normalization layer in step two is applied after each convolutional layer to eliminate the interference of signal amplitude fluctuations under different system parameter conditions on weight updates, providing a stable training foundation for rapid cross-task iteration under the MAML framework. Global residual connections enable the network to focus on compensating for defocusing errors, avoiding the gradient vanishing problem in deep network training.

[0022] Furthermore, the training of the complex fully convolutional network described in step two adopts a method based on... Norm-based reconstruction loss function:

[0023]

[0024] in, For the first Sampling echoes under each task For the corresponding reference image, These are the network parameters to be optimized for Complex-CVFCN.

[0025] Furthermore, the MAML two-layer optimization training described in step three includes:

[0026] During the inner parameter update phase, in the support set Calculate loss And perform one or more gradient descent iterations to obtain task-specific parameters. :

[0027]

[0028] Where θ represents the initial parameters of the network. The learning rate for the inner loop;

[0029] In the outer optimization phase, in the query set Calculate the adapted parameters loss Summing or averaging yields the total aggregation loss, i.e.:

[0030] = ;

[0031] Calculate the gradient with respect to the original meta-parameters θ, and update the meta-parameters θ. Then calculate the gradient of the total loss with respect to the original meta-parameters θ, and update the meta-parameters as well.

[0032] ;

[0033] Where β is the outer loop learning rate.

[0034] Furthermore, the number of small support set samples mentioned in step four is 1 to 5, and the gradient update of the predetermined number of steps is a gradient descent update of no more than 5 times.

[0035] The present invention also provides a meta-learning-based cross-system parameter adaptive SAR image reconstruction apparatus for performing the above method, the apparatus comprising:

[0036] The data acquisition module is used to acquire raw SAR echo signals and generate various multi-task data with different system parameters;

[0037] A network building module is provided for constructing the Complex-CVFCN complex fully convolutional network according to any one of claims 2 to 6;

[0038] The training module is used to obtain meta-initialized weights by employing the MAML two-layer optimization training as described in claim 8;

[0039] The fast adaptation module is used to perform gradient updates on the meta-initialized weights for a predetermined number of steps under tasks with unknown system parameters, using a small number of support set samples, to obtain the adapted network parameters.

[0040] The imaging module is used for SAR image reconstruction using the adapted network.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] First, it exhibits excellent generalization ability across system parameters. This invention innovatively applies the MAML meta-learning framework to SAR end-to-end complex domain imaging, enabling meta-training on various bandwidth tasks and learning highly gradient-sensitive meta-initialization weights. In unknown bandwidth tasks deviating from the training center frequency, the network can effectively suppress defocusing caused by bandwidth mismatch after a small number of gradient updates, significantly improving imaging quality under cross-system parameter conditions.

[0043] Second, it has a highly efficient and rapid deployment capability. When faced with new tasks with different system parameters, there is no need to collect large-scale data for retraining or fine-tuning for a long time. The network adaptation can be completed in just 1-5 steps of gradient updates, and the convergence speed is far superior to that of traditional networks (traditional networks require thousands of iterations). This significantly reduces the data acquisition and training costs when upgrading or migrating SAR systems.

[0044] Third, complete phase information is preserved. Through the complex convolution mechanism of the Complex-CVFCN complex fully convolutional network, the amplitude and phase coupling information of the SAR echo is completely preserved, avoiding the distortion of electromagnetic scattering characteristics caused by the decomposition of complex components in traditional methods, and achieving more accurate defocus correction and sidelobe suppression.

[0045] Fourth, excellent robustness and generalization performance. The invention combines instance normalization layers to eliminate the impact of signal amplitude fluctuations on training, and global residual connections to ensure the training stability of deep networks. Experimental results show that the PSNR of this invention is significantly better than existing methods under all test bandwidth conditions. Attached Figure Description

[0046] Figure 1This is a flowchart illustrating the overall processing flow of the method of the present invention;

[0047] Figure 2 A schematic diagram of the MAML two-layer optimization training framework;

[0048] Figure 3 Comparison of SAR image focusing results under different bandwidths and cross-scenario tasks;

[0049] Figure 4 A comparison chart of imaging performance (PSNR curves) of different methods under different radar bandwidths. Detailed Implementation

[0050] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this is not intended to limit the scope of the invention.

[0051] Example 1

[0052] This embodiment provides a cross-system parameter adaptive SAR image reconstruction method based on meta-learning, such as... Figure 1 As shown, it includes the following steps:

[0053] Step 1: Acquire raw SAR echo signals and construct a multi-task dataset.

[0054] Assuming the airborne radar platform is at an altitude At a constant speed Moving along the azimuth direction, a target exists in the imaging scene. Its three-dimensional spatial coordinates are As the radar platform moves, the instantaneous slant range between the target and the phase center of the radar antenna... Slower time with direction The signal undergoes continuous change. The two-dimensional baseband echo signal received by the radar can be modeled as follows:

[0055] in, Indicates distance in terms of time. Indicates direction in slow time. Indicates the moment when the radar beam center crosses. This represents the scattering coefficient of the target. At the speed of light, Represents carrier frequency, For range-directed frequency modulation, and These are the window functions for the range and azimuth directions, respectively.

[0056] The range resolution of SAR is mainly determined by the effective bandwidth of the signal. The decision is made using the following formula: Where c is the speed of light.

[0057] A full-bandwidth reference signal is obtained through simulation. To simulate the performance of different radar systems or different imaging modes, a portion of the signal's spectrum is truncated in the frequency domain using digital processing. Since the signal is a linear frequency modulated (LFM) signal in the time domain, and the time-domain sampling points have a linear mapping relationship with the frequency domain frequencies, center-clipping in the frequency domain is equivalent to limiting the effective frequency modulation range of the transmitted pulse. Assume the signal after full-bandwidth range compression is... ,in For direction and time, For distance to time.

[0058] To simulate different radar system parameter configurations, the reference signal is center-cropped in the frequency domain. Specific steps include:

[0059] (1) Distance to Fourier Transform:

[0060] The time-domain range-compressed signal is transformed to the range-frequency domain using a Fast Fourier Transform (FFT):

[0061] ;

[0062] in Represents the distance to the frequency axis;

[0063] (2) Frequency domain window clipping:

[0064] Set the target effective bandwidth for the current task to The full bandwidth is A rectangular window function needs to be constructed. Its bandwidth corresponds to the number of sampling points of the target bandwidth:

[0065] ;

[0066] Clipped frequency domain signal Represented as:

[0067] ;

[0068] (3) Inverse Fourier Transform (IFFT):

[0069] Transforming the cropped spectrum back into the time domain yields the distance-compressed signal required for the task:

[0070] .

[0071] In this way, multiple imaging tasks with different range-direction point spread functions (PSF) can be generated based on the same reference echo (as shown in Table 1), thereby training the MAML model to capture the feature shift caused by bandwidth variations. The signals for each task are fully focused using the traditional RDA algorithm to generate high-quality images as training labels (Ground Truth), and the signals for each task are undersampled in the azimuth direction to generate echoes containing aliasing artifacts as network input.

[0072] Step 2: Construct the Complex-CVFCN complex fully convolutional network.

[0073] Construct an 8-layer Complex-CVFCN network with the following layer structure:

[0074] Layer 1 (Input Layer): Directly input the original complex domain echo signal, preserving complete amplitude and phase information.

[0075] Layers 2-3 (complex convolutional layers): Eight 3×3 complex convolutional kernels with a stride of 1×1 are used to extract the initial complex domain topological features. Each complex convolutional layer is followed by an instance normalization layer and a complex ReLU activation layer.

[0076] Layer 4 (Pooling Layer): Uses a 2×2 pooling kernel with a stride of 1×1 to perform local feature filtering.

[0077] Layers 5-6 (complex convolutional layers): 16 output channels, 3×3 kernel size, 1×1 stride, for feature level expansion and non-linear modeling. Each complex convolutional layer is followed by an instance normalization layer and a complex ReLU activation layer.

[0078] Layer 7 (Feature Flattening Layer): High-dimensional semantic feature vectorization.

[0079] Layer 8 (fully connected regression layer): Image domain grayscale mapping and imaging output.

[0080] Global residual connections: These connect directly from the input layer to the output layer, allowing the network to focus on compensating for defocusing errors while preventing gradient vanishing.

[0081] The network as a whole can be represented as:

[0082] ;

[0083] In the FCNN architecture, the real and imaginary parts are used as multi-channel inputs.

[0084] Complex convolution operation rules: For the first... Complex convolution operation of layers, assuming the input feature map is... The convolution kernel is Output the real part of the feature map and the virtual part Calculate according to the following formulas respectively:

[0085]

[0086]

[0087] In this way, FCNN can learn the amplitude and phase characteristics of the echo signal simultaneously, thereby achieving more accurate defocus correction in the complex domain.

[0088] To maintain phase coherence, the network employs the Complex-ReLU activation function, performing nonlinear transformations on the real and imaginary parts separately:

[0089] ;

[0090] In addition, instance normalization is introduced after each convolutional layer to eliminate the impact of drastic fluctuations in signal amplitude between different tasks (such as different bandwidths) on weight updates, which is crucial for rapid iteration under the subsequent MAML framework.

[0091] To enable network output Approximate the standard reference image generated by the RD algorithm as closely as possible. The network's loss function adopts a method based on... Norm-based reconstruction loss function:

[0092]

[0093] in, For the first Sampling echoes under each task For the corresponding RDA reference image These are the network parameters to be optimized for Complex-CVFCN. This loss function not only constrains pixel-level brightness consistency but also implicitly requires the network to achieve optimal matched filtering by learning the convolutional kernel parameters of FCNN.

[0094] Step 3: Use the MAML framework to perform two-layer optimization meta-training.

[0095] The training data includes imaging tasks with five different bandwidth parameters (30MHz, 50MHz, 75MHz, 100MHz, and 150MHz), sourced from various target scenarios such as ships, vessels, and shore-based systems. The dataset contains 2000 samples, including 1400 samples in the meta-training set, 300 samples in the meta-validation set, and 300 samples in the test set with unknown parameters.

[0096] like Figure 2 As shown, the MAML two-layer optimization training framework specifically includes the following elements:

[0097] (1) Task sampling: A batch of imaging tasks are randomly sampled from the multi-task dataset (containing imaging tasks with different bandwidths and different scenarios). Each imaging task Divided into support sets and query set , The support set is used to simulate rapid adaptation to new tasks with small samples, and the query set is used to evaluate the generalization performance after adaptation and update the meta-parameters.

[0098] (2) Inner loop (Fast Adaptation): For each task Starting with the current meta-parameter θ (initialized to random or pre-trained values), in the support set... Calculate loss And perform one or more gradient descent iterations (arrows in the diagram indicate gradient backpropagation and parameter update directions) to obtain task-specific parameters. :

[0099]

[0100] Here, α is the learning rate of the inner loop, and the number of steps in the inner loop is usually set to 1-5 steps to simulate the rapid adaptation process in real deployment scenarios where only a small number of samples can be obtained.

[0101] (3) Outer loop (Meta Optimization): After all tasks have completed the inner loop, the query set corresponding to each task is used. Calculate the adapted parameters loss Summing or averaging yields the total aggregation loss, i.e.

[0102] = ;

[0103] Calculate the gradient with respect to the original meta-parameter θ, and update the meta-parameter θ (as shown by the arrows in the diagram). Then, calculate the gradient of the total loss with respect to the original meta-parameter θ (the arrows in the diagram indicate the gradient backpropagation path), and update the meta-parameters.

[0104] ;

[0105] Where β is the outer loop learning rate.

[0106] Note that the gradient calculation of the outer loop needs to go through the gradient update process of the inner loop (i.e., the second gradient), but in actual implementation, a first-order approximation can also be used to reduce the amount of computation.

[0107] (4) Iterative convergence: Repeat the above task sampling, inner loop, and outer loop process until the meta-parameter θ converges or reaches the preset maximum number of iterations. The meta-parameter θ obtained after training... ∗ This refers to the meta-initialized weights that have the ability to generalize across system parameters.

[0108] Figure 2 The "meta-parameter θ" and "task-specific parameters" The bidirectional arrows between the inner loop and the outer loop's meta-parameter updates visually demonstrate the dependency between the inner loop's rapid adaptation and the outer loop's meta-parameter updates.

[0109] exist Figure 2 In the framework shown, the support set and query set for each task come from the same system parameter distribution (e.g., the same effective bandwidth), but the bandwidth parameters differ between different tasks. Through this design, the meta-learning model can extract common patterns regarding the changes in the point spread function (PSF) from bandwidth variations. Thus, when faced with unseen bandwidth parameters, it can quickly adjust the matched filtering characteristics of the convolution kernel with only a small number of support set samples, effectively correcting imaging defocus caused by system parameter drift.

[0110] Training parameter settings: total number of iterations (epochs) = 2000, batch size = 16, Adam optimizer used, learning rate set to 0.002. The meta-initialized weights θ* obtained after meta-training have generalization ability across system parameters (bandwidth) and are highly gradient sensitive to tasks with unknown bandwidth.

[0111] Step 4: Perform fast adaptation inference under unknown system parameter tasks.

[0112] Performing Fast Few-shot Inference on an Unknown Bandwidth Task: Setting up a bandwidth task that has never appeared in the training set (e.g., jumping from the training 75MHz to an unknown 40MHz or 120MHz).

[0113] Before inference begins, the Complex-CVFCN network loads the globally optimal initial weights obtained during the meta-learning phase. This weight is not the optimal solution for a specific task, but rather it is located at a "characteristic singularity" with high gradient sensitivity, capable of sensing the common topology of echo signals under different bandwidths.

[0114] The fast inference process requires only a small amount of support set data. That will complete the process. The specific steps are as follows:

[0115] Forward propagation: sparsely sampled echoes of unknown bandwidth Input Complex-CVFCN to calculate the current weights. The reconstructed image below.

[0116] Loss calculation: Calculate the inference loss using a very small number of reference images (e.g., 1-5 samples). Single-step gradient update: executed once or very rarely ( Gradient descent update to obtain the needle

[0117] Specific weights for the current bandwidth :

[0118]

[0119] in is the learning rate. Since the FCNN architecture has already learned the general physical mapping of the complex domain during the meta-learning stage, a single-step update allows the network to quickly capture the sidelobe characteristics and resolution scale under the current bandwidth.

[0120] The input is unknown bandwidth task data. Meta weight Loading FCNN meta weights Calculate the gradient ,renew ,use Perform full data inference and output images. The final output is a high-quality, adapted SAR image. .

[0121] Effect verification:

[0122] To verify the effectiveness and generalization ability of the proposed deep FCNN imaging method based on the MAML framework across system parameters, the implementation example uses five different bandwidth system parameters listed in Table 1 to perform inverse imaging from SAR complex images of several different target scenarios (ships, vessels, shore-based, etc.) to generate original echoes. Frequency domain cropping is performed according to the effective bandwidth of the task, followed by sparsity processing (sparseness of 0.5). Each meta-learning task corresponds to one effective bandwidth, generating multi-task data pairs. The dataset contains a total of 2000 samples, including a 1400-sample meta-training set, a 300-sample meta-validation set, and a 300-sample unseen-task test set.

[0123] Table 1. Five different adaptive parameter settings

[0124] .

[0125] In the experiment, the constructed 8-layer deep Complex-CVFCN imaging network was first trained using MAML using a meta-training set to learn a general optimal starting point for initial weights. Subsequently, in the meta-testing phase, to simulate extreme non-ideal conditions of physical mismatch in system parameters, the radar configuration was artificially modified, for example, by adjusting the effective bandwidth (…). Set it to a value different from the one used during training.

[0126] During the testing phase, the network loaded with meta-general initial weights was rapidly adapted in a very small number of steps (e.g., 5 steps) on small sample data of unknown bandwidth tasks to obtain a task-specific focused network. The experimental results are as follows: Figure 3 As shown, Figure 3 The left column shows the model input (aliased echo imaging), the middle column shows the model output, and the right column shows the original imaging scene. It can be seen that even at previously unseen bandwidths such as 90MHz and 120MHz, the images output by the method of this invention can still effectively suppress sidelobes and aliasing artifacts, closely resembling the original scene.

[0127] like Figure 4 As shown, the impact of three algorithms on the peak signal-to-noise ratio (PSNR) of reconstructed images under different radar transmit bandwidths is compared using the method of this invention. The results show that the method of this invention significantly outperforms the comparative methods FCNN(Static) and CISTA-Net under all bandwidth conditions, achieving a peak PSNR of 39.6 dB at a bandwidth of 75 MHz. It also maintains excellent imaging quality on unseen bandwidth tasks such as 40 MHz and 120 MHz, verifying that the invention has stronger robustness and imaging performance in cross-system parameter generalization.

[0128] Experimental results demonstrate that this invention, by introducing a deep fully convolutional network within the Meta-Learning Framework (MAML) and combining the InstanceNorm stabilization mechanism with a global residual learning strategy, can achieve high-quality SAR imaging under non-ideal cross-system parameter observation conditions. This method combines the rapid cross-parameter adaptive capability of MAML, the powerful modeling ability of deep FCNN for complex surface target features, and excellent generalization performance when the system's physical configuration changes.

Claims

1. A cross-system parameter adaptive SAR image reconstruction method based on meta-learning, characterized in that, Includes the following steps: Step S1: Acquire the raw SAR echo signal, perform frequency domain window cropping according to different effective bandwidths, and generate various imaging tasks with different range-direction point spread functions (PSFs). Step S2: Construct a Complex-CVFCN network. The network uses complex convolution to jointly model the real and imaginary parts of the SAR echo signal, retaining amplitude and phase coupling information, and realizing end-to-end reconstruction from complex domain echo to SAR image. Step S3: The complex fully convolutional network is trained using a two-layer optimization framework of Model Independent Meta-Learning (MAML). The inner loop is used to quickly adapt to multiple imaging tasks on the support set, and the outer loop is used to update the meta-initialization parameters on the query set, thereby obtaining meta-weights with cross-system parameter generalization ability. Step S4: Load the meta-weights. For tasks with unknown system parameters, perform gradient updates for a predetermined number of steps using a small number of support set samples to quickly adapt the network parameters, and then use the adapted network to reconstruct SAR images.

2. The cross-system parameter adaptive SAR image reconstruction method based on meta-learning according to claim 1, characterized in that, The operation method of complex convolution in step S2 is as follows: For the Complex convolution operation of layers, with input feature map as The convolution kernel is Output the real part of the feature map and the virtual part Calculate according to the following formulas respectively: Where * represents convolution operation, and j is the imaginary unit.

3. The cross-system parameter adaptive SAR image reconstruction method based on meta-learning according to claim 1, characterized in that, The Complex-CVFCN network described in step S2 uses the complex ReLU activation function, which is defined as follows: , where ReLU is the corrected linear unit; The Complex-CVFCN introduces an instance normalization layer after each convolutional layer to eliminate the impact of signal amplitude fluctuations under different system parameter conditions on weight updates.

4. The cross-system parameter adaptive SAR image reconstruction method based on meta-learning according to claim 1, characterized in that, The Complex-CVFCN network mentioned in step S2 is an 8-layer network structure. Its input layer directly processes the original echo signal in the complex domain and includes at least one convolutional layer with a 3×3 convolutional kernel, a pooling layer with a 2×2 pooling kernel, and a feature flattening layer. The network output is the grayscale mapping result in the SAR image domain. The Complex-CVFCN network introduces global residual connections, enabling the network to focus on compensating for defocusing errors.

5. The cross-system parameter adaptive SAR image reconstruction method based on meta-learning according to claim 1, characterized in that, The training of the complex fully convolutional network in step S2 is based on... Norm-based reconstruction loss function: in, For the first Sampling echoes under each task For the corresponding reference image, These are the network parameters to be optimized for Complex-CVFCN.

6. The cross-system parameter adaptive SAR image reconstruction method based on meta-learning according to claim 1, characterized in that, The frequency domain window clipping in step S1 specifically includes: transforming the time-domain range-compressed signal to the range frequency domain, constructing a rectangular window function in the frequency domain to clip the signal spectrum according to the target effective bandwidth, and inversely transforming the clipped spectrum back to the time domain to obtain the range-compressed signal of the target task; the effective bandwidths corresponding to the generated different imaging tasks are different, so that different imaging tasks have different range resolutions.

7. The cross-system parameter adaptive SAR image reconstruction method based on meta-learning according to claim 1, characterized in that, The MAML two-layer optimization training described in step S3 includes: During the inner parameter update phase, in the support set Calculate loss And perform one or more gradient descent iterations to obtain task-specific parameters. : Where θ represents the initial parameters of the network. The learning rate for the inner loop; In the outer optimization phase, in the query set Calculate the adapted parameters loss Summing or averaging yields the total aggregation loss, i.e.: = ； Calculate the gradient with respect to the original meta-parameters θ, update the meta-parameters θ, then calculate the gradient of the total loss with respect to the original meta-parameters θ, and update the meta-parameters: ； Where β is the outer loop learning rate.

8. The cross-system parameter adaptive SAR image reconstruction method based on meta-learning according to claim 7, characterized in that, The meta-parameters for the MAML two-layer optimization training are set as follows: epochs=2000, batch size=16, Adam optimizer is used, and learning rate is set to 0.

002.

9. The cross-system parameter adaptive SAR image reconstruction method based on meta-learning according to claim 1, characterized in that, The number of small support set samples mentioned in step S4 is 1 to 5, and the gradient update of the predetermined number of steps is a gradient descent update of no more than 5 times.

10. A cross-system parameter adaptive SAR image reconstruction device based on meta-learning, characterized in that, include: The data acquisition module is used to acquire raw SAR echo signals and generate various multi-task data with different system parameters; A network building module is provided for constructing the Complex-CVFCN complex fully convolutional network according to any one of claims 2 to 6; The training module is used to obtain meta-initialized weights by employing the MAML two-layer optimization training as described in claim 8; The fast adaptation module is used to perform gradient updates on the meta-initialized weights for a predetermined number of steps under tasks with unknown system parameters, using a small number of support set samples, to obtain the adapted network parameters. The imaging module is used for SAR image reconstruction using the adapted network.

Images