Two-dimensional intelligent imaging network construction and training method under radar sparse sampling
By constructing a two-dimensional intelligent imaging network based on the ADMM optimization algorithm and combining sparse prior and equal-variable constraint loss, the problem of low accuracy in sparse ISAR imaging is solved, and efficient, high-quality imaging is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-13
- Publication Date
- 2026-04-03
AI Technical Summary
Existing sparse ISAR imaging methods have low imaging accuracy under sparse conditions. Traditional methods rely on hyperparameter adjustment and are computationally complex. End-to-end network training requires paired datasets that are difficult to obtain.
A two-dimensional intelligent imaging network based on the ADMM optimization algorithm is constructed. By combining sparse prior modules and equivariant constraint loss, the network is trained through gradient descent and self-supervised learning to reduce the need for paired data.
It improves imaging quality under sparse sampling conditions, reduces the need for paired data, and achieves efficient, high-quality imaging.
Smart Images
Figure CN121784732A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar imaging technology, and in particular relates to a method for constructing and training a two-dimensional intelligent imaging network under radar sparse sampling. Background Technology
[0002] Inverse Synthetic Aperture Radar (ISAR) plays a crucial role in space situational awareness and target monitoring, possessing all-weather, all-time imaging capabilities. ISAR achieves high range resolution by transmitting broadband signals and high azimuth resolution by utilizing the relative motion (especially the rotation component) between the radar and the target. However, in practical ISAR imaging scenarios, due to interference and radar system resource scheduling, it is difficult to acquire complete echo data in both azimuth and range dimensions according to the Nyquist sampling theorem, leading to the sparse ISAR imaging problem. The traditional Range-Doppler (RD) method is unsuitable under sparse conditions, as it produces grating lobes that cause image blurring.
[0003] Existing sparse ISAR imaging methods are mainly based on compressed sensing (CS) and deep learning (DL) techniques: CS-based methods can be divided into greedy algorithms, sparse Bayesian learning (SBL), and convex optimization algorithms based on sparse constraints. Greedy algorithms achieve reconstruction by iteratively selecting locally optimal matching elements, but the imaging accuracy decreases significantly when the sparsity of the target is unknown. The SBL algorithm assumes that the reconstructed signal elements follow a predefined sparse prior distribution and obtains the optimal estimate by maximizing the posterior probability through Bayes' theorem. It has strong modeling ability, but the computational complexity is high, the hyperparameters are sensitive, and it is easily affected by sparse model mismatch. Convex optimization methods use iterative solvers (such as Fastiterative shrinkage threshold algorithm, FISTA and alternating direction method of multiplier, ADMM) to handle sparse ISAR imaging problems [4]. Among them, ADMM has attracted much attention because of its ability to handle multivariate optimization problems. However, the selection of related hyperparameters in these methods depends on empirical adjustment, which can easily affect the imaging accuracy, and the regularized sparse prior for the continuous structural features of the target may lead to reconstruction mismatch.
[0004] Deep learning (DL) methods include purely data-driven approaches and model-guided approaches. Purely data-driven methods train end-to-end deep neural networks directly using radar echoes or preprocessed images as input and corresponding ISAR images as output. While achieving good results, they lack interpretability and generalization performance. Model-guided methods, on the other hand, form interpretable iterative neural networks through optimization algorithms (such as ADMM), achieving computational efficiency and high imaging accuracy, and can integrate reinforcement learning to adaptively learn iterative hyperparameters. Nevertheless, training these end-to-end networks typically requires paired radar echo and image data, and obtaining sufficiently large training datasets remains a significant challenge in the ISAR field. Summary of the Invention
[0005] To address the aforementioned issues, this invention provides a method for constructing and training a two-dimensional intelligent imaging network under sparse radar sampling, which can significantly improve imaging quality under sparse sampling conditions while reducing the need for paired data.
[0006] A method for constructing and training a two-dimensional intelligent imaging network under radar sparse sampling, wherein the two-dimensional intelligent imaging network includes a measurement reconstruction module, a sparse prior module, and a dual ascent module; If the set number of iterations has not been reached, the measurement and reconstruction module uses gradient descent to reconstruct the regularized constraint image obtained from the previous iteration. residual Reconstructing sparsely sampled echoes The reconstructed image of this iteration is obtained. After reaching the set number of iterations, the reconstructed image obtained in the last iteration is used as the final output of the two-dimensional intelligent imaging network. The sparse prior module is based on sparse prior regularization filtering, using the residual obtained from the previous iteration. The reconstructed image obtained in this iteration Apply regularization constraints to obtain the regularization constraint image for this iteration. ; The dual ascent module is based on a dual ascent method, using the reconstructed image obtained in this iteration. Regularized constrained images Obtain the residual value of this iteration. .
[0007] Furthermore, the measurement and reconstruction module acquires the reconstructed image. The method is as follows:
[0008] in, For the first The first iteration in the process of the iteration The reconstructed image obtained by subgradient descent For the first The first iteration in the process of the iteration The reconstructed image obtained by subgradient descent Let be the total number of gradient descent iterations, then the th... Subgradient descent yields the final reconstructed image. ; These are the gradient descent weights. Let be the learning rate for gradient descent, and and All are learnable variables; The range sampling matrix, This is the azimuth sampling matrix. This is the conjugate transpose of the matrix.
[0009] Furthermore, the sparse prior module acquires the regularized constraint image. The method is as follows:
[0010] in, It is a soft threshold function. This indicates the adaptive threshold module.
[0011] Furthermore, the adaptive threshold module The calculation method is as follows:
[0012] in, and These represent two convolutional layers. This represents the activation function layer.
[0013] Furthermore, the dual ascent module obtains the residual amount. The method is as follows:
[0014] in, No. Learnable variables during the iteration process.
[0015] Furthermore, the loss function used when training the two-dimensional intelligent imaging network is as follows:
[0016] in, and These represent measurement constraints and isotropic constraints, respectively. To constrain the weights, For the first in the equivariant constraint There are rotation operators, where |G| represents the total number of rotation operators in the isovariant constraint. For the forward sampling operator, For two-dimensional intelligent imaging networks, For the first training sample set A sparsely sampled echo, The total number of sparsely sampled echoes in the training sample set. This represents the square of the Frobenius norm of the matrix.
[0017] Beneficial effects: This invention provides a method for constructing and training a two-dimensional intelligent imaging network under sparse radar sampling. First, a sparsely sampled two-dimensional imaging network is constructed based on the ADMM optimization algorithm, ensuring network resolvability and algorithm robustness. In addition, the effect of sparse prior filtering is improved by designing a globally learnable threshold. By exploring the rotational equivariance characteristics in radar imaging, a self-supervised training strategy based on equivariance constraints is designed, which can achieve high-quality network training using only incomplete radar echo data, reducing the need for paired data, and thus significantly improving the imaging quality under sparse sampling conditions. Attached Figure Description
[0018] Figure 1 A flowchart illustrating a method for constructing and training a two-dimensional intelligent imaging network under radar sparse sampling. Figure 2 A schematic diagram of a sparse sampling two-dimensional imaging network with adaptive thresholding; Figure 3 This is a schematic diagram of rotational equalization constraints in radar imaging. Figure 4 This is a schematic diagram of electromagnetic simulation data observation in the example; Figure 5 The images show sparse sampling imaging results for different methods used in the implementation examples. Detailed Implementation
[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.
[0020] This invention provides a method for constructing and training a two-dimensional intelligent imaging network under sparse radar sampling. The method first constructs an analytically sparsely sampled two-dimensional imaging network based on the ADMM optimization algorithm, where a globally learnable threshold is designed to effectively suppress grating lobe interference. Furthermore, the rotational isovariance characteristics in radar imaging are explored, and an isovariance constraint loss is constructed to train the radar imaging network in a self-supervised mode. The effectiveness of the proposed algorithm is verified through simulation. The two-dimensional intelligent imaging network includes a measurement reconstruction module, a sparse prior module, and a dual ascent module. It should be noted that the radar 2D imaging problem under sparse sampling is an ill-posed problem. High-quality imaging cannot be achieved directly using traditional matched filtering methods, requiring the introduction of additional prior information. Sparse priors can effectively characterize the scattering characteristics of the imaging target, constructing a joint constraint problem of measurement constraints and sparse priors. This problem is generally solved using optimization algorithms. However, traditional optimization methods require manual setting of hyperparameters, which affects imaging results and efficiency. Therefore, in this invention, the optimization method is expanded into an iterative neural network architecture. The network is trained to adaptively learn the hyperparameters of the iterative process. The processing flowchart is as follows: Figure 1 As shown, the specific steps are as follows: If the set number of iterations has not been reached, the measurement and reconstruction module uses gradient descent to reconstruct the regularized constraint image obtained from the previous iteration. residual Reconstructing sparsely sampled echoes The reconstructed image of this iteration is obtained. After reaching the set number of iterations, the reconstructed image obtained in the last iteration is used as the final output of the two-dimensional intelligent imaging network. It should be noted that, in order to ensure that the reconstruction results are consistent with the sampled data, the following measurement constraints and prior constraints are established:
[0021] in, , and These are represented as sparse sampled echo, range image sampling matrix, azimuth sampling matrix, and the image to be reconstructed, respectively. For sparse prior constraints As an auxiliary variable, Let be the weight coefficients of the prior terms. First, solve the measurement constraint iterations. Based on the above global optimization objective, the following objective function can be obtained:
[0022] in As dual variables, the above problem can be solved using gradient descent, updating the reconstructed image acquired by the measurement and reconstruction module. The expression is as follows:
[0023] in, For the first The first iteration in the process of the iteration The reconstructed image obtained by subgradient descent For the first The first iteration in the process of the iteration The reconstructed image obtained by subgradient descent Let be the total number of gradient descent iterations, then the th... Subgradient descent yields the final reconstructed image. ; These are the gradient descent weights. Let be the learning rate for gradient descent, and and All are learnable variables; For the distance image sampling matrix, This is the azimuth sampling matrix. This is the conjugate transpose of the matrix.
[0024] The sparse prior module is based on sparse prior regularization filtering, using the residual obtained from the previous iteration. The reconstructed image obtained in this iteration Apply regularization constraints to obtain the regularization constraint image for this iteration. ; It should be noted that the sparse prior module, by adding prior regularization, can find a solution space that meets the constraints from the underdetermined problem. In sparse sampling ISAR imaging, sparse prior constraints can express target characteristics and can be used as an effective regularization constraint. Based on the above equation... The following prior regular iteration objective can be obtained:
[0025] Its theoretical analytical solution is:
[0026] in, This is a soft thresholding function. The above analytical solution only considers the sparse statistical properties of the target and fails to consider its structural properties. Generally, targets have structural properties, meaning that a point will have structural relationships with surrounding points. Therefore, in this invention, to perform prior regularization filtering more efficiently, we replace the fixed threshold with a learnable convolutional layer. That is, using the following regularization constraint image Forms of expression:
[0027] in, It is a soft threshold function. This represents the adaptive threshold module. The calculation method is as follows:
[0028] in, and These represent two convolutional layers. This represents the activation function layer.
[0029] The dual ascent module is based on a dual ascent method, using the reconstructed image obtained in this iteration. Regularized constrained images Obtain the residual value of this iteration. .
[0030] The dual ascent module obtains the residual amount. The method is as follows:
[0031] in, No. Learnable variables during the iteration process.
[0032] The adaptive threshold sparse 2D imaging network is now complete. The overall process is as follows: Figure 2 As shown.
[0033] Next, we can use the rotational equivariance properties in radar imaging to construct an equivariant constraint loss to train the radar imaging network.
[0034] It should be noted that in sparse sampling imaging, a property can be observed: for radar sampling with specific sparse measurements, when the target rotates around its centroid, the image of the target rotates accordingly after imaging, while the grating lobes do not rotate with it, but are determined solely by the sparse sampling pattern. This property is called the isovariant constraint. This property provides the mathematical basis for suppressing grating lobes using a self-supervised learning strategy, and it can be expressed as:
[0035] in and These are the rotation operator and the forward sampling operator, respectively. This is a radar imaging network. Based on the ISAR imaging operator, the following expression can be obtained:
[0036] in Forward operator The null space, in order to satisfy the equivariant constraint, i.e.: Therefore, we can conclude that:
[0037] This demonstrates that rotationally equivariant constraints can reduce the null space acquisition and understanding, as illustrated in the overall diagram below. Figure 3 As shown. Based on the above inferences, the loss function used when constructing and training the two-dimensional intelligent imaging network is as follows:
[0038] in, and These represent measurement constraints and isotropic constraints, respectively. To constrain the weights, For the first in the equivariant constraint There are rotation operators, where |G| represents the total number of rotation operators in the isovariant constraint. For the forward sampling operator, For intelligent imaging networks, For the first training sample set A sparsely sampled echo, The total number of sparsely sampled echoes in the training sample set. This represents the square of the Frobenius norm of the matrix.
[0039] Thus, the proposed method for constructing and training two-dimensional intelligent imaging networks under radar sparse sampling was completed.
[0040] In this invention, an electromagnetic simulation experiment based on the publicly available dataset Gotcha was conducted, and the data acquisition and observation process is as follows: Figure 4 As shown, the target radar echo was acquired through circular trajectory mode with an echo signal-to-noise ratio (SNR) of 30dB. Sparse sampling reconstruction was performed on an electromagnetic simulation vehicle target. Specifically, the original data was collected using downsampling rates of 50% and 70%.
[0041] To verify the effectiveness of the algorithm, we compared it with traditional RD algorithms and supervised class unfolding methods. The experimental results are as follows: Figure 5 As shown, it can be seen from Figure 5 As can be seen, the proposed self-supervised radar imaging network can effectively suppress grating lobes, with results similar to those of supervised methods. In contrast, traditional RD methods suffer from severe grating lobes due to sparse sampling, which affects image interpretation. This demonstrates that the proposed method has a certain two-dimensional imaging capability under sparse sampling conditions.
[0042] In summary, this invention provides a method for constructing and training a two-dimensional intelligent imaging network under sparse radar sampling. First, a sparse sampling two-dimensional imaging network is constructed based on the ADMM optimization algorithm, and the hyperparameters in the optimization algorithm are set as learnable variables. Second, the rotational isovariance property in radar imaging is used to construct an isovariance constraint loss to train the sparse sampling two-dimensional imaging network.
[0043] Therefore, this invention aims to provide a highly intelligent and efficient solution for radar sparse sampling two-dimensional imaging. First, it constructs an analytically sparse sampling two-dimensional imaging network based on the ADMM optimization algorithm. The threshold filtering uses learnable convolutional layers to learn the filtering threshold instead of the traditional fixed threshold, improving the filtering effect. Second, it uses rotational constraints to achieve self-supervised learning, i.e., high-quality network training is achieved using only incomplete radar echo data. This invention can be used in fields such as high-quality imaging under radar sparse sampling conditions, significantly improving imaging quality under sparse sampling conditions while reducing the need for paired data.
[0044] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.
Claims
1. A method for constructing and training a two-dimensional intelligent imaging network under radar sparse sampling, characterized in that, The two-dimensional intelligent imaging network includes a measurement and reconstruction module, a sparse prior module, and a dual ascent module. If the set number of iterations has not been reached, the measurement and reconstruction module uses gradient descent to reconstruct the regularized constraint image obtained from the previous iteration. residual Reconstructing sparsely sampled echoes The reconstructed image of this iteration is obtained. ; After reaching the set number of iterations, the reconstructed image obtained in the last iteration is used as the final output of the two-dimensional intelligent imaging network. The sparse prior module is based on sparse prior regularization filtering, using the residual obtained from the previous iteration. The reconstructed image obtained in this iteration Apply regularization constraints to obtain the regularization constraint image for this iteration. ; The dual ascent module is based on a dual ascent method, using the reconstructed image obtained in this iteration. Regularized constrained images Obtain the residual value of this iteration. .
2. The method for constructing and training a two-dimensional intelligent imaging network under sparse radar sampling as described in claim 1, characterized in that, The measurement and reconstruction module acquires the reconstructed image. The method is as follows: in, For the first The first iteration in the process of the second iteration The reconstructed image obtained by subgradient descent For the first The first iteration in the process of the second iteration The reconstructed image obtained by subgradient descent Let be the total number of gradient descent iterations, then the th... Subgradient descent yields the final reconstructed image. ; These are the gradient descent weights. Let be the learning rate for gradient descent, and and All are learnable variables; The range sampling matrix, This is the azimuth sampling matrix. This is the conjugate transpose of the matrix.
3. The method for constructing and training a two-dimensional intelligent imaging network under sparse radar sampling as described in claim 1, characterized in that, The sparse prior module obtains the regularization constraint image. The method is as follows: in, It is a soft thresholding function. This indicates the adaptive threshold module.
4. The method for constructing and training a two-dimensional intelligent imaging network under sparse radar sampling as described in claim 3, characterized in that, Adaptive threshold module The calculation method is as follows: in, and These represent two convolutional layers. This represents the activation function layer.
5. The method for constructing and training a two-dimensional intelligent imaging network under sparse radar sampling as described in claim 1, characterized in that, The dual ascent module obtains the residual. The method is as follows: in, No. Learnable variables during the iteration process.
6. The method for constructing and training a two-dimensional intelligent imaging network under sparse radar sampling as described in claim 1, characterized in that, The loss function used when training the two-dimensional intelligent imaging network is as follows: in, and These represent measurement constraints and isotropic constraints, respectively. To constrain the weights, For the first in the equivariant constraint There are rotation operators, where |G| represents the total number of rotation operators in the isovariant constraint. For the forward sampling operator, For two-dimensional intelligent imaging networks, For the first training sample set A sparsely sampled echo, The total number of sparsely sampled echoes in the training sample set. This represents the square of the Frobenius norm of the matrix.