SAR sparse imaging method and system based on double-channel depth unfolding network

Abstract

This application relates to a SAR sparse imaging method and system based on a dual-channel depth unfolding network. The method includes updating image features using a physical observation matrix; dual-channel parallel feature mapping, thereby focusing on the sparse recovery of strong scattering points and compensating for detail loss caused by soft thresholding; fusing the outputs of the two channels to obtain the final output of the layer; and network training and imaging. This application can achieve high-quality reconstruction of three-dimensional targets under sparse observation conditions through dual-channel fusion.

Description

Technical Field

[0001] This application relates to the field of synthetic aperture radar imaging technology, and in particular to a SAR sparse imaging method and system based on a dual-channel depth unfolding network. Background Technology

[0002] Simple deep unfolding networks suffer from poor scene adaptability, at least in that single-channel unfolding networks often struggle to adapt to both "extremely sparse multi-point scenes" and "dense point / weakly sparse scenes." A singular processing logic leads to detail loss in certain complex scenarios, such as structural artifacts and information loss: relying solely on model-driven approaches can easily result in the over-thresholding and filtering of weak target information; relying solely on data-driven approaches can easily produce structural artifacts and lack physical consistency. Furthermore, parameter tuning is rigid; traditional algorithms typically use fixed step sizes and thresholds, failing to adaptively adjust based on local features of the image content. Summary of the Invention

[0003] This application provides a SAR sparse imaging method and system based on a dual-channel depth unfolding network, which can achieve high-quality reconstruction of three-dimensional targets under sparse observation conditions through dual-channel fusion.

[0004] According to one of the solutions in this application, a SAR sparse imaging method based on a dual-channel depth unfolding network is provided, including:

[0005] Construct a three-dimensional array SAR physical observation model; A dual-parallel composite network structure is constructed within the framework of iterative algorithm expansion to build a dual-channel deep expansion network architecture; Update image features using the physical observation matrix; Dual-channel parallel feature mapping focuses on sparse recovery of strong scattering points and compensates for the loss of details caused by soft thresholding; By merging the outputs of the two channels, the final output of the layer is obtained; Network training and imaging.

[0006] In some embodiments, constructing a three-dimensional array SAR physical observation model includes: Define three-dimensional geometric relationships; Construct the echo integral equation; Discretization and linear measurement model construction.

[0007] In some embodiments, constructing a dual-channel deep unfolded network architecture includes: Build includes A deep unfolded network of layers is used to find the optimal solution to the regularization problem; Based on the top-level initialization, subsequent steps are executed for each subsequent layer.

[0008] In some embodiments, updating image features using a physical observation matrix includes: Obtain the current echo residual and gradient, and update the intermediate features.

[0009] In some embodiments, the dual-channel parallel feature mapping includes: The model drives the ISTA channel to preserve the interpretability of traditional algorithms.

[0010] In some embodiments, the dual-channel parallel feature mapping further includes: Data-driven Transform channels are used to learn generalized structural priors for the scenario.

[0011] In some embodiments, the two channel outputs are fused to obtain the final output of the layer, including: Adaptive gating fusion is performed at the end of each layer. The two channel outputs are fused through a pixel-level gating mechanism to obtain the final output of the layer. This achieves adaptive adjustment, emphasizing the ISTA channel in sparse regions and the Transform channel in complex regions.

[0012] In some embodiments, network training and imaging include: Based on offline training, the network parameter set is updated using the backpropagation algorithm; Online imaging to output high-resolution three-dimensional scattering images.

[0013] According to one of the solutions in this application, a SAR sparse imaging system based on a dual-channel depth unfolding network is provided, comprising: The building module is used at least to construct a dual-parallel composite network structure based on a three-dimensional array SAR physical observation model within an iterative algorithm expansion framework. The processing module is used to update image features using the physical observation matrix; dual-channel parallel feature mapping is used to focus on sparse recovery of strong scattering points and to compensate for the loss of details caused by soft thresholding. An imaging module, used at least to obtain the final output of the layer based on the fused two-channel output; network training and imaging.

[0014] In some embodiments, the dual-channel parallel feature mapping includes: The model drives the ISTA channel to preserve the interpretability of traditional algorithms; Data-driven Transform channels are used to learn generalized structural priors for the scenario.

[0015] The SAR sparse imaging method based on a dual-channel deep unfolding network, according to various embodiments of this application, at least constructs a three-dimensional array SAR physical observation model; constructs a dual-parallel composite network structure under an iterative algorithm unfolding framework to build a dual-channel deep unfolding network architecture; updates image features using the physical observation matrix; performs dual-channel parallel feature mapping to focus on sparse recovery of strong scattering points and compensate for detail loss caused by soft thresholding; fuses the outputs of the two channels to obtain the final output of the layer; and performs network training and imaging, thereby solving the above-mentioned problems in the prior art. The dual-parallel composite network structure constructed under the iterative algorithm unfolding framework includes a model-driven ISTA channel and a data-driven Transform channel, and achieves pixel-level adaptive fusion of the two channel outputs through a gated fusion module. It adopts the idea of ​​"dual-channel complementarity": one channel introduces physical model constraints, and the other channel introduces data-driven priors. The fusion of the two achieves high-quality reconstruction of three-dimensional targets under sparse observation conditions, taking into account both physical consistency and data prior expression capabilities.

[0016] It should be understood that the foregoing general description and the following detailed description are exemplary and illustrative only, and are not intended to limit the scope of this application. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic diagram of a dual-channel overall network architecture according to an embodiment of this application is shown; Figure 2 The results of a three-dimensional threshold scan of a single point target T1 according to an embodiment of this application are shown. Detailed Implementation

[0019] Various embodiments and features of this application are described herein with reference to the accompanying drawings.

[0020] It should be understood that various modifications can be made to the embodiments described herein. Therefore, the above description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope and spirit of this application will be apparent to those skilled in the art.

[0021] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present application and, together with the general description of the present application given above and the detailed description of the embodiments given below, serve to explain the principles of the present application.

[0022] These and other features of this application will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0023] It should also be understood that although this application has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of this application.

[0024] The above and other aspects, features and advantages of this application will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0025] Specific embodiments of this application are described thereafter with reference to the accompanying drawings; however, it should be understood that the claimed embodiments are merely examples of this application, which can be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure the application. Therefore, the specific structural and functional details claimed herein are not intended to be limiting, but merely serve as the basis and representative basis for the claims to teach those skilled in the art to use this application in a variety of substantially any suitable detailed structures.

[0026] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in other embodiments,” all of which may refer to one or more of the same or different embodiments according to this application.

[0027] Synthetic Aperture Radar (SAR) can image the spatial scattering characteristics of a target scene through sparse arrays or aperture distribution optimization. In practical applications, sparse sampling methods with a sampling rate lower than the Nyquist sampling rate are often used to reduce data storage and transmission pressure.

[0028] Traditional sparse imaging methods are mainly divided into two categories: While traditional matched filtering algorithms are robust, under undersampling conditions, their imaging results are limited by sidelobe interference, leading to reduced resolution and making it difficult to balance imaging quality and efficiency. Iterative algorithms based on compressed sensing (CS) utilize the sparse prior of the target and improve resolution by solving a regularization optimization problem. However, these algorithms typically require hundreds of iterations to converge, resulting in extremely high computational complexity. Furthermore, manually designed regularization parameters are difficult to adjust, and reconstruction performance is often limited when dealing with weakly sparse targets or complex backgrounds.

[0029] In recent years, deep unfolding networks have combined model interpretability and data-driven capabilities by mapping iterative algorithms to a hierarchical neural network structure. However, existing technologies and simple deep unfolding networks suffer from poor scene adaptability: single-channel unfolding networks often struggle to adapt to both extremely sparse multi-point scenes and dense / weakly sparse scenes simultaneously. A singular processing logic leads to detail loss, structural artifacts, and information loss in certain complex scenes: relying solely on model-driven approaches can easily result in over-thresholding of weak target information; relying solely on data-driven approaches can easily produce structural artifacts and lack physical consistency. Furthermore, parameter tuning is rigid: traditional algorithms typically use fixed step sizes and thresholds, failing to adaptively adjust based on local features of the image content.

[0030] Based on the background section described above, this application provides illustrative solutions to address the deficiencies in the prior art through embodiments, but these are not intended to limit the scope of patent protection claimed in this application.

[0031] As one of the solutions, comprehensive Figure 1 and Figure 2 As shown, embodiments of this application provide a SAR sparse imaging method based on a dual-channel depth unfolding network, including: Construct a three-dimensional array SAR physical observation model; A dual-parallel composite network structure is constructed within the framework of iterative algorithm expansion to build a dual-channel deep expansion network architecture; Update image features using the physical observation matrix; Dual-channel parallel feature mapping focuses on sparse recovery of strong scattering points and compensates for the loss of details caused by soft thresholding; By merging the outputs of the two channels, the final output of the layer is obtained; Network training and imaging.

[0032] In view of the foregoing content, this application aims at least to provide a SAR sparse imaging method based on a dual-channel depth unfolding network, relating to a high-resolution SAR imaging method that combines a physical model and data-driven dual-channel depth unfolding network under sparse array or undersampling conditions. The disclosed technical solution, "dual-channel depth unfolding network," combines... Figure 1 The schematic diagram of the overall network architecture of DualPath-LSISTA-Net shown can be understood as: a dual-parallel composite network structure built under the framework of iterative algorithm expansion, including the model-driven ISTA channel and the data-driven Transform channel, and the adaptive fusion of the two channel outputs at the pixel level is achieved through the gated fusion module, thereby taking into account both physical consistency and data prior representation ability.

[0033] In the specific implementation of the embodiments of this disclosure, the SAR sparse imaging method based on dual-channel depth unfolding network can be illustrated by steps S1 to S6 in the following sections.

[0034] In some implementation schemes, the embodiments of this application may include: constructing a three-dimensional array SAR physical observation model, including: defining three-dimensional geometric relationships; constructing echo integral equations; and discretizing and constructing a linear measurement model.

[0035] Step S1: Construct a three-dimensional array SAR physical observation model.

[0036] Step S11: Define the three-dimensional geometric relationships and set the radar system to operate in a three-dimensional planar coordinate system. Define the azimuth coordinate as... The distance coordinate is The height coordinate is The radar antenna array is distributed in three-dimensional space, the first... The phase center position of each array element is denoted as The location of any scattering point within the imaging scene is denoted as... The one-way Euclidean distance between the antenna and the scattering point. Represented as: .

[0037] Step S12: Construct the echo integral equation. After the radar transmits a linear frequency modulated signal and performs pulse compression, the... Echo data received from each sampling point It can be modeled as the target scattering coefficient distribution In the three-dimensional imaging area Volume fraction on: ; in, The imaginary unit, The center wave number.

[0038] Step S13: Discretize and linearly measure the model to construct a continuous 3D scene. Discretize into The voxel mesh is used. The three-dimensional matrix of discrete scattering coefficients to be solved is flattened into column vectors, denoted as . (in ). All The radar observation echo data from each aperture sampling point is vectorized and denoted as... Construct a linear measurement equation: ; in, To determine the phase term in step S12 The constructed three-dimensional observation matrix.

[0039] In some implementations, embodiments of this application may include: constructing a dual-channel deep unfolded network architecture, including: constructing a network containing... The network is expanded in depth to find the optimal solution to the regularization problem; based on the initialization of the top layer, subsequent steps are executed for each subsequent layer.

[0040] In this embodiment, step S2 can be added on top of step S1.

[0041] Step S2: Construct a dual-channel deep unfolded network architecture. This embodiment constructs a network architecture containing... A deep unfolded network of layers is designed to solve the following Optimal solution to the regularization problem : ; The network layer 0 is initialized as follows: For each subsequent layer, perform the following steps.

[0042] In some implementations, the embodiments of this application may include: updating image features using a physical observation matrix, including: obtaining the current echo residual and gradient and updating intermediate features.

[0043] In this embodiment, step S3 can be added on top of step S2.

[0044] Step S3: Execute the gradient descent module, at the... The layer updates image features using the physical observation matrix to ensure data consistency.

[0045] Step S31 Calculate the current echo residual : .

[0046] Step S32: Calculate the gradient and update the intermediate features. : , in, This is the learnable step size parameter for this layer.

[0047] In some implementations, the embodiments of this application can be: dual-channel parallel feature mapping, including: a model-driven ISTA channel to preserve the interpretability of traditional algorithms; and a data-driven Transform channel for learning generalized structural priors of the scenario.

[0048] In this embodiment, step S4 can be added on top of step S3.

[0049] Step S4: Dual-channel parallel feature mapping.

[0050] Step S41: Channel 1: Model-driven ISTA channel. This channel retains the interpretability of traditional algorithms and focuses on the sparse recovery of strong scattering points.

[0051] , in: For feature extraction operations that include convolutional layers; This is the learnable shrinkage threshold for this layer; For soft thresholding function: ; Channel Two: Data-Driven Transform Channel. This channel is used to learn generalized structural priors for the scene (such as texture and weak targets) to compensate for the loss of detail caused by soft thresholding. A normalized gradient term is introduced. As an auxiliary input, the calculation formula is: , in, For parameters A deep convolutional neural network is used to perform nonlinear transformations and detail enhancement.

[0052] In some implementations, the embodiments of this application may be: fusing the outputs of two channels to obtain the final output of the layer, including: adaptive gating fusion at the end of each layer, fusing the outputs of two channels through a pixel-level gating mechanism to obtain the final output of the layer, thereby achieving adaptive adjustment that emphasizes the ISTA channel in sparse regions and the Transform channel in complex regions.

[0053] In this embodiment, step S5 can be added on top of step S4.

[0054] Step S5: Adaptive gated fusion. At the end of each layer, the two channel outputs are fused using a pixel-level gating mechanism to obtain the final output of that layer. .

[0055] Step S51: Calculate the fusion weight map : Utilizing lightweight convolutional networks Extract spatial features and map them to the (0,1) interval using the Sigmoid function: .

[0056] Step S52: Perform weighted fusion: ; in This indicates element-wise multiplication. This step achieves adaptive adjustment, emphasizing the ISTA channel in sparse regions and the Transform channel in complex regions.

[0057] In some implementations, embodiments of this application may include: network training and imaging, comprising: offline training to update the network parameter set using a backpropagation algorithm; and online imaging to output a high-resolution three-dimensional scattering image.

[0058] In this embodiment, step S6 can be added on top of step S5.

[0059] Step S6: Network training and imaging.

[0060] Step S61: Offline Training: Constructing images containing ground truth values and corresponding simulated echo Given a dataset, define a loss function and use the backpropagation algorithm to update the network parameter set.

[0061] Step S62: Online Imaging: Input the radar-measured 3D echo data into the trained network, and then... Layer-wise forward inference directly outputs high-resolution 3D scattering images. .

[0062] As one solution, embodiments of this application provide a SAR sparse imaging system based on a dual-channel depth unfolding network, including: The building module is used at least to construct a dual-parallel composite network structure based on a three-dimensional array SAR physical observation model within an iterative algorithm expansion framework. The processing module is used to update image features using the physical observation matrix; dual-channel parallel feature mapping is used to focus on sparse recovery of strong scattering points and to compensate for the loss of details caused by soft thresholding. An imaging module, used at least to obtain the final output of the layer based on the fused two-channel output; network training and imaging.

[0063] In some implementations, the processing module for dual-channel parallel feature mapping may include: a model-driven ISTA channel to preserve the interpretability of traditional algorithms; and a data-driven Transform channel to learn the generalized structural priors of the scenario.

[0064] Based on the content of the various technical solutions disclosed in the previous text, the technical solutions involved in the SAR sparse imaging method and system based on dual-channel depth unfolding network and the corresponding beneficial effects can be further explained in detail through specific implementation methods combined with specific experiments and results analysis.

[0065] System structure and signal parameters: The experiment employed a single-transmitter, single-receiver system for three-dimensional imaging verification. The radar's transmitted signal had a center frequency of 35 GHz, a bandwidth of 300 MHz, and 51 frequency sampling points. The wavelength, calculated based on the center frequency, was approximately 8.57 mm.

[0066] Array configuration and sparse observation settings: Array in – Planar structures form a two-dimensional array (element coordinates) The array has a side length of 0.13 m, an element spacing of 8 mm, 17 elements per side, and a total of 289 (17×17) elements in a full array. To verify the robustness of the algorithm under sparse observations, 40 elements were randomly selected from the full array to form a sparse array, thus constructing sparse observations. Under the sparse array condition, the measurement vector dimension is: , Unobserved channels are zeroed out to create sparse observation inputs.

[0067] 3D imaging region and voxel mesh: The 3D imaging area is set as follows: m, m, m, with a voxel step size of 0.2 m, therefore the 3D mesh size is The total number of prime numbers is 9261, and the target location is: .

[0068] The experimental results clearly show that the three-dimensional point spread function (PSF) index, including PSLR, ISLR and IRW, is used to reflect the sidelobe suppression capability, energy leakage degree and focusing resolution capability. The technical solutions of each embodiment of this disclosure adopt the idea of ​​"dual-channel complementarity": one channel introduces physical model constraints, and the other channel introduces data-driven priors. The fusion of the two achieves high-quality reconstruction of three-dimensional targets under sparse observation conditions.

[0069] Compared to sparse BP, the disclosed technique achieves more negative PSLR / ISLR and smaller IRW in all three directions, indicating that it can simultaneously reduce the highest sidelobe, reduce total sidelobe energy leakage, and significantly narrow the main lobe width, thereby improving three-dimensional focusing quality and resolution. The elevation ISLR improvement is the greatest, demonstrating that this technique is particularly effective in suppressing elevation energy leakage and artifacts caused by sparse arrays. See the table below for sparse BP imaging parameters:

[0070] refer to Figure 2The results of a three-dimensional threshold scan of a single point target T1 are shown. Combining the three-dimensional PSF index and the three-dimensional threshold scan results, sparse BP exhibits significant angular sidelobe lift and energy leakage under sparse observation conditions, and is highly sensitive to threshold selection, easily leading to background artifact diffusion. The embodiments of this disclosure are based on a dual-channel complementary mechanism of "physical model + data-driven," which significantly reduces the highest sidelobe and total sidelobe energy leakage while maintaining accurate geometric positioning, and narrows the main lobe width. This allows the three-dimensional reconstruction results to maintain higher background purity and a more stable target structure under different thresholds, making it more suitable for high-precision three-dimensional imaging tasks under sparse observation.

[0071] This disclosure also provides a SAR sparse imaging device based on a dual-channel depth unfolding network, including one or more processing modules configured to execute the SAR sparse imaging method based on a dual-channel depth unfolding network described above, and at least configured to execute specific implementations of steps S1 to S6.

[0072] Based on the above-mentioned inventive concept, the SAR sparse imaging method, device and system based on dual-channel depth unfolding network of various embodiments of this disclosure at least constructs a three-dimensional array SAR physical observation model; constructs a dual-parallel composite network structure under the iterative algorithm unfolding framework to construct a dual-channel depth unfolding network architecture; updates image features using the physical observation matrix; performs dual-channel parallel feature mapping, thereby focusing on the sparse recovery of strong scattering points and compensating for the loss of details caused by soft thresholding; fuses the outputs of the two channels to obtain the final output of the layer; and performs network training and imaging, thereby solving the above-mentioned problems existing in the prior art. The dual-parallel composite network structure constructed under the iterative algorithm unfolding framework includes a model-driven ISTA channel and a data-driven Transform channel, and achieves pixel-level adaptive fusion of the outputs of the two channels through a gated fusion module. It adopts the idea of ​​"dual-channel complementarity": one channel introduces physical model constraints, and the other channel introduces data-driven priors. The fusion of the two achieves high-quality reconstruction of three-dimensional targets under sparse observation conditions, taking into account both physical consistency and data prior expression capabilities.

[0073] This application also provides a computer-readable storage medium storing computer-executable instructions thereon. When executed by a processor, the computer-executable instructions mainly implement the SAR sparse imaging method based on a dual-channel depth unfolding network as described above, including: Construct a three-dimensional array SAR physical observation model; A dual-parallel composite network structure is constructed within the framework of iterative algorithm expansion to build a dual-channel deep expansion network architecture; Update image features using the physical observation matrix; Dual-channel parallel feature mapping focuses on sparse recovery of strong scattering points and compensates for the loss of details caused by soft thresholding; By merging the outputs of the two channels, the final output of the layer is obtained; Network training and imaging.

[0074] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A SAR sparse imaging method based on a dual-channel depth unfolding network, including: Construct a three-dimensional array SAR physical observation model; A dual-parallel composite network structure is constructed within the framework of iterative algorithm expansion to build a dual-channel deep expansion network architecture; Update image features using the physical observation matrix; Dual-channel parallel feature mapping focuses on sparse recovery of strong scattering points and compensates for the loss of details caused by soft thresholding; By merging the outputs of the two channels, the final output of the layer is obtained; Network training and imaging.

2. The method according to claim 1, constructing a three-dimensional array SAR physical observation model, comprising: Define three-dimensional geometric relationships; Construct the echo integral equation; Discretization and linear measurement model construction.

3. The method according to claim 2, wherein, Constructing a dual-channel deep unfolded network architecture includes: Build includes A deep unfolded network of layers is used to find the optimal solution to the regularization problem; Based on the top-level initialization, subsequent steps are executed for each subsequent layer.

4. The method according to claim 3, wherein, Updating image features using the physical observation matrix includes: Obtain the current echo residual and gradient, and update the intermediate features.

5. The method according to claim 4, wherein, Dual-channel parallel feature mapping includes: The model drives the ISTA channel to preserve the interpretability of traditional algorithms.

6. The method according to claim 4, wherein the dual-channel parallel feature mapping further comprises: Data-driven Transform channels are used to learn generalized structural priors for the scenario.

7. The method according to claim 6, fusing the outputs of two channels to obtain the final output of the layer, includes: Adaptive gating fusion is performed at the end of each layer. The two channel outputs are fused through a pixel-level gating mechanism to obtain the final output of the layer. This achieves adaptive adjustment, emphasizing the ISTA channel in sparse regions and the Transform channel in complex regions.

8. The method according to claim 7, wherein, Network training and imaging, including: Based on offline training, the network parameter set is updated using the backpropagation algorithm; Online imaging to output high-resolution three-dimensional scattering images.

9. A SAR sparse imaging system based on a dual-channel depth unfolding network, including: The building module is used at least to construct a dual-parallel composite network structure based on a three-dimensional array SAR physical observation model within an iterative algorithm expansion framework. The processing module is used to update image features using the physical observation matrix; dual-channel parallel feature mapping is used to focus on sparse recovery of strong scattering points and to compensate for the loss of details caused by soft thresholding. An imaging module, used at least to obtain the final output of the layer based on the fused two-channel output; network training and imaging.

10. The system according to claim 9, wherein, Dual-channel parallel feature mapping includes: The model drives the ISTA channel to preserve the interpretability of traditional algorithms; Data-driven Transform channels are used to learn generalized structural priors for the scenario.

Images