SAR Imaging Method and Related Device Based on Hybrid-Norm Deep Unfolding Network

Abstract

The present application discloses a SAR imaging method and related device based on a hybrid norm deep unfolding network, which relates to the field of remote sensing imaging technology. The method includes: obtaining radar operating parameters; obtaining an echo matrix according to the SAR raw echo signals collected at all sampling times; determining the observation scene of the SAR, and constructing a slant range matrix corresponding to each scene point in the observation scene at each random sampling time; determining the echo signal after range migration correction according to the echo matrix and the slant range matrix; inputting the echo signal after range migration correction into the trained hybrid norm deep unfolding network to obtain a reconstructed scene. In the present application, the optimization equation is transformed into a hybrid norm optimization equation based on the alternating direction multiplier method. The hybrid norm optimization equation includes a Tikhonov regularization optimization equation and an L 1 / 2 regularization optimization equation, and the deep unfolding network approximates the iterative process to reduce the computational complexity of the iterative process.

Description

Technical Field

[0001] The present application relates to the technical field of remote sensing imaging, and in particular to a SAR imaging method and related device based on a hybrid norm deep unfolding network. Background Art

[0002] Spaceborne Synthetic Aperture Radar (SAR) can efficiently detect hot spots and key targets through multi-level situation awareness technical means to achieve rapid discovery, detection, and identification of scene targets. With the development of Compressed Sensing (CS) theory, under the condition that the observation matrix satisfies the Restricted Isometry Property (RIP), signals can be recovered from a non-uniform sub-Nyquist sampling method that is far fewer than Nyquist sampling samples based on the sparsity of the signal. When the non-linear iterative optimization algorithm ensures high spatial resolution, the compressed sensing radar can achieve ultra-wide area observation of the scene through a flexible beam pointing design. The compressed sensing radar can increase the swath / resolution ratio to the same order of magnitude as the SIMO-SAR and MIMO-SAR systems. Moreover, the compressed sensing radar system based on sub-Nyquist sampling has a simple structure, small antenna size, and small data volume, reducing the hardware burden and the pressure of signal storage and transmission. It is a very attractive and promising research route for high-resolution wide-swath SAR systems.

[0003] The compressed sensing radar system requires prior information on the sparsity of the observation scene and has certain limitations in application. For example, in most cases, the sparsity of the observation scene is unknown, and the CS theory is applied to complex scenes. Currently, sparse imaging algorithms mainly include greedy algorithms, algorithms based on statistical distributions, and regularization algorithms. Greedy algorithms tend to fall into local optima. Algorithms based on statistical distributions have strong modeling capabilities, but they pose certain challenges for complex scenes lacking sparsity. In regularization algorithms, although there are some useful heuristic methods, there is no unified standard for how to select appropriate regularization parameters. One can only select parameters according to experience and manually optimize them through "trial and error", resulting in low execution efficiency and long running time. At the same time, due to the limitations of the algorithm itself, the compressed sensing imaging method can only obtain less image information and cannot obtain rich deep image information, which leads to poor reconstruction performance of the compressed sensing algorithm when the undersampling ratio is large. Therefore, a fast reconstruction algorithm with reliable accuracy and close to ideal theoretical performance is also a difficult problem in CS research. Summary of the Invention

[0004] The purpose of this application is to provide a SAR imaging method and related device based on a hybrid norm deep unfolding network. In the case of low signal-to-noise ratio, considering the influence of noise, clutter, etc. on the algorithm, the stability against interference during the imaging process is enhanced, and it is not affected by the scene inhomogeneity.

[0005] To achieve the above object, this application provides the following solutions:

[0006] In the first aspect, this application provides a SAR imaging method based on a hybrid norm deep unfolding network, including:

[0007] Obtain radar operating parameters; the radar operating parameters include the reference slant range at the scene center, satellite flight speed, azimuth antenna size, number of transmitted pulses, chirp rate of the transmitted signal, time vector of the radar transmitted pulses, and operating wavelength;

[0008] Obtain an echo matrix according to the SAR raw echo signals collected at all sampling times;

[0009] Determine the observation scene of the SAR, and construct a corresponding slant range matrix for each scene point in the observation scene at each random sampling time;

[0010] Determine the echo signal after range migration correction according to the echo matrix and the slant range matrix;

[0011] Input the echo signal after range migration correction into the trained hybrid norm deep unfolding network to obtain a reconstructed scene; the hybrid norm deep unfolding network includes K levels, and each level includes four sub-modules, namely the first regularization module, the sparse transformation module, the second regularization module, and the multiplier update module, and the result is output by the reconstruction module; at any level, the data input from the previous level obtains a reconstructed scene through the reconstruction module, and the first regularization module is used to solve the Tikhonov regularization equation to obtain an iterative solution result; the sparse representation of the iterative solution result is obtained through the sparse transformation module; the second regularization module is used to solve the L 1 / 2 regularization equation to obtain an iterative solution result; the multiplier update module is used to update the Lagrange multiplier.

[0012] In the second aspect, this application provides a SAR imaging device based on a hybrid norm deep unfolding network, including:

[0013] A radar operating parameter acquisition module, configured to: obtain radar operating parameters; the radar operating parameters include the reference slant range at the scene center, satellite flight speed, azimuth antenna size, number of transmitted pulses, chirp rate of the transmitted signal, time vector of the radar transmitted pulses, and operating wavelength;

[0014] An echo matrix acquisition module, configured to: obtain an echo matrix according to the SAR raw echo signals collected at all sampling moments;

[0015] A slant range matrix construction module, configured to: determine the observation scene of the SAR, and construct a corresponding slant range matrix for each scene point in the observation scene at each random sampling moment;

[0016] A determined echo signal module after range migration correction, configured to: determine the echo signal after range migration correction according to the echo matrix and the slant range matrix;

[0017] A scene reconstruction module, configured to: input the echo signal after range migration correction into a trained hybrid norm deep unfolding network to obtain a reconstructed scene; the hybrid norm deep unfolding network includes K levels, and each level includes four sub-modules, namely a first regularization module, a sparse transformation module, a second regularization module, and a multiplier update module, and outputs results from the reconstruction module; in any level, the data input from the previous level obtains a reconstructed scene through the reconstruction module, and the first regularization module is used to solve the Tikhonov regularization equation to obtain an iterative solution result; the sparse representation of the iterative solution result is obtained through the sparse transformation module; the second regularization module is used to solve the L 1 / 2 regularization equation to obtain an iterative solution result; the multiplier update module is used to implement the update of the Lagrange multiplier.

[0018] In a third aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to implement the above-mentioned SAR imaging method based on a hybrid norm deep unfolding network.

[0019] In a fourth aspect, the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the above-mentioned SAR imaging method based on a hybrid norm deep unfolding network is implemented.

[0020] According to the specific embodiments provided by the present application, the following technical effects are disclosed in the present application:

[0021] Existing compressive sensing algorithms do not take into account the effects of noise and clutter in the case of low signal-to-noise ratio. In addition, considering the non-uniformity of the scene, a single regularization parameter weighting cannot distinguish strong and weak targets well. Based on this, the present application provides a SAR imaging method and related device based on a hybrid norm deep unfolding network. Considering the non-uniformity of the scene, a single regularization parameter weighting cannot distinguish strong and weak targets well. By comprehensively considering sparsity, stability, and dynamic parameter adaptability, the optimization equation is transformed into a hybrid norm optimization equation based on the alternating direction multiplier method. The hybrid norm optimization equation includes: solving the Tikhonov regularization optimization equation through the first regularization module in the hybrid norm deep unfolding network, and the global optimum can be obtained by taking the derivative; parsing and solving the L 1 / 2 regularization optimization equation through the second regularization module in the hybrid norm deep unfolding network. The approximate iterative process of the deep unfolding network is also adopted to reduce the computational complexity of the iterative process and further optimize the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0023] Figure 1 It is an application environment diagram of a SAR imaging method based on a hybrid norm deep unfolding network in an embodiment of the present application;

[0024] Figure 2 It is a flowchart of a SAR imaging method based on a hybrid norm deep unfolding network provided by an embodiment of the present application;

[0025] Figure 3 It is a functional module diagram of a SAR imaging device based on a hybrid norm deep unfolding network provided by an embodiment of the present application;

[0026] Figure 4 It is a structural diagram of a computer device provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, rather than all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0028] The compressed sensing radar system has a simple structure, small antenna size, and small data volume, which reduces the hardware burden and the pressure of signal storage and transmission. It is a very attractive and promising research route for high-resolution wide-swath SAR systems. Due to the limitations of the algorithm itself, the compressed sensing radar can only obtain less image information and cannot obtain rich deep image information, resulting in poor reconstruction performance of the compressed sensing algorithm when the undersampling ratio is large.

[0029] To make the above objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] The SAR imaging method based on the hybrid norm deep unfolding network provided by the embodiments of the present application can be applied to, for example, Figure 1 the application environment shown in the figure. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set up separately, integrated on the server 104, or placed on the cloud or other servers. The terminal 102 can send the radar working parameters to the server 104. After receiving the radar working parameters, for the radar working parameters, the server 104 obtains the echo matrix according to the SAR raw echo signals collected at all sampling times, determines the observation scene of the SAR, and constructs the slant range matrix corresponding to each scene point in the observation scene at each random sampling time. According to the echo matrix and the slant range matrix, the echo signal after range migration correction is determined, and the echo signal after range migration correction is input into the trained hybrid norm deep unfolding network to obtain the reconstructed scene. The server 104 can feedback the obtained reconstructed scene to the terminal 102. In addition, in some embodiments, the SAR imaging method based on the hybrid norm deep unfolding network can also be implemented by the server 104 or the terminal 102 alone. For example, the terminal 102 can directly perform SAR imaging processing on the radar working parameters, or the server 104 can obtain the radar working parameters from the data storage system and perform SAR imaging processing on the radar working parameters.

[0031] Among them, the terminal 102 can be, but is not limited to, various desktop computers and laptop computers. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers, and can also be a cloud server.

[0032] In an exemplary embodiment, as Figure 2 shown in the figure, a SAR imaging method based on the hybrid norm deep unfolding network is provided. This method is executed by a computer device, and can be specifically executed by a computer device such as a terminal or a server alone, or jointly executed by a terminal and a server. In the embodiments of the present application, this method is applied toFigure 1 Taking the server 104 in as an example, the following steps 201 to 205 are included. Among them:

[0033] Step 201: Obtain radar operating parameters; the radar operating parameters include the reference slant range of the scene center, the satellite flight speed, the azimuth antenna size, the number of transmitted pulses, the chirp rate of the transmitted signal, the time vector of the radar transmitted pulses, and the operating wavelength.

[0034] Step 202: Obtain an echo matrix according to the SAR raw echo signals collected at all sampling times.

[0035] Step 203: Determine the observation scene of the SAR, and construct a corresponding slant range matrix for each scene point in the observation scene at each random sampling time.

[0036] Step 204: Determine the echo signal after range migration correction according to the echo matrix and the slant range matrix.

[0037] Step 205: Input the echo signal after range migration correction into the trained hybrid norm deep unfolding network to obtain a reconstructed scene; the hybrid norm deep unfolding network includes K levels, and each level includes four sub-modules, namely the first regularization module, the sparse transformation module, the second regularization module, and the multiplier update module, and the result is output by the reconstruction module; at any level, the data input from the previous level obtains a reconstructed scene through the reconstruction module, and the first regularization module is used to solve the Tikhonov regularization equation to obtain an iterative solution result; the sparse representation of the iterative solution result is obtained through the sparse transformation module; the second regularization module is used to solve the L 1 / 2 regularization equation to obtain an iterative solution result; the multiplier update module is used to implement the update of the Lagrange multiplier.

[0038] Implementing the above steps 201 to 205, considering the non-uniformity of the scene, a single regularization parameter weighting cannot well distinguish strong and weak targets. Considering sparsity, stability, and dynamic parameter adaptability comprehensively, the optimization equation is transformed into a hybrid norm optimization equation based on the alternating direction multiplier method. The hybrid norm optimization equation includes: solving the Tikhonov regularization optimization equation through the first regularization module in the hybrid norm deep unfolding network, and the global optimum can be obtained by taking the derivative; the L 1 / 2 regularization optimization equation is analytically solved through the second regularization module in the hybrid norm deep unfolding network. The approximate iterative process of the deep unfolding network is also adopted to reduce the computational complexity of the iterative process and further optimize the algorithm.

[0039] In step 201, the radar operating parameter is f 1 ={R 0 ,V e ,La , N a , λ, Kr, ta}, where R 0 is the reference slant range of the scene center, V e is the satellite flight speed, L a is the azimuth antenna size, N a is the number of transmitted pulses, λ is the operating wavelength, Kr is the chirp rate of the transmitted signal, and the time vector ta of the radar transmitted pulse is ta = [η 1 , η 2 ,..., η m ,... η Na T , where η m represents the non-uniform sampling time of the m-th pulse transmitted by the SAR, and [·] T represents the transpose operation.

[0040] In another exemplary embodiment of the present application, if the echo pixel points of the SAR are a two-dimensional matrix, then the echo matrix in step 202 is expressed as:

[0041]

[0042] where S echo,n,k represents the echo of the target point with the sampling time η m and the center slant range τ k ·c / 2, that is, the element in the m-th row and the k-th column of the echo matrix S echo . Additionally, for convenience of representation, S echo is also expressed as a column vector S echo = [S echo,1,col … S echo,m,cZl … S echo,Na,col T , and a row vector S echo = [S echo,row,1 … S echo,row,k … S echo,row,Nr . Where S echo,n,col = [S echo,m,1 … S echo,m,k …S echo,m,Nr represents the m-th row of S echo , and S echo,row,k = [S echo,1,k … S echo,m,k … S echo,Na,k T represents the k-th column of S echo .

[0043] ​​​The specific process of step 203 is as follows: Select a scene, establish a geometric motion model between the SAR system and the ground scene (i.e., the observation scene), and construct the slant range matrix r(η corresponding to each scene point at each random sampling moment m ).

[0044] Taking the center point of the reference scene as the center, select an area not smaller than the size of the observation scene, and divide this area into grids. Assume there are N points along the x-axis and N points along the y-axis. The spacing between grids should be less than or equal to the azimuth resolution ρ. Denote the spacing between grids as l, l ≤ ρ a , but the grid spacing l should not be taken too small, because too small a grid spacing cannot optimize its reconstruction performance, but will affect the efficiency of the algorithm. Based on the scene center coordinates (x i , y i , 0), the coordinates of each scene point can be obtained as shown in the following formula (2):

[0045]

[0046] Assume the initial satellite sampling position is [x track , 0, H]. Then the successive sampling position coordinates of the satellite are {[x track + V e η 1 , 0, H], [x track + Vη 2 , 0, H], …, [x track + Vη m , 0, H], … [x track + Vη Na , 0, H]}, where it is assumed that η 1 is the zero moment of random sampling, and H is the flight altitude of the satellite. Finally, according to the scene point coordinates and the satellite sampling position coordinates, the slant range r(η m ) of all scene points at each sampling moment is obtained.

[0047] In the present invention, any slant range r i (η m ) is calculated in the non-rotating geocentric coordinate system, and at the same time, the r(η m ) needs to consider the time difference between the pericenter passage moment and the SAR system startup moment. The calculation process of the slant range r i (η m ) is as follows:

[0048] First, use formula (3) to calculate the slant range r in the geocentric coordinate system; formula (3) is expressed in the following form:

[0049]

[0050] Among them, a is the semi-major axis of the elliptical orbit, and M is the mean anomaly.

[0051] Secondly, according to the slant range r obtained above, based on the following formula, convert the satellite sampling position coordinates to the non-rotating geocentric coordinate system:

[0052]

[0053]

[0054]

[0055]

[0056] The (x, y, z) in the last formula (7) are the satellite sampling position coordinates in the non-rotating geocentric coordinate system. In the above formula, (ξηξ) represents the coordinates of a certain point on the elliptical orbit in the reference frame ξηξ. θ represents the true anomaly, ω represents the argument of perigee, Ω represents the right ascension of the ascending node, i represents the orbital inclination, l 1 、l 2 、l 3 、m 1 、m 2 、m 3 、n 1 、n 2 、n 3 are all intermediate parameters.

[0057] Again, according to the satellite sampling position coordinates and the scene point coordinates in the non-rotating geocentric coordinate system, calculate the slant range r(η m ) of all scene points at each sampling moment.

[0058] At each random sampling moment in the azimuth direction, calculate the slant range coefficient of all scene points at the random sampling moment. The slant range coefficient is a two-dimensional matrix r, which is specifically expressed as follows:

[0059]

[0060] Among them, represents the slant range coefficient of the (n, n)th target in the scene layout when the crossing time is η m , that is, the element in the mth row and the n 2 th column of the slant range coefficient matrix r. In addition, for the sake of convenience of representation, r n is also expressed as a column vector of the element row vector and a row vector with elements as column vectors r = [r ta,1,1 … r ta,n,n … r ta,N,N . Among them represents the mth row of r, Indicates the nth 2 column of r.

[0061] In another exemplary embodiment of the present application, step 204 may be replaced by the following steps 301 to 303:

[0062] Step 301: Determine a scene delay signal matrix according to the echo matrix and the slant range matrix;

[0063] Step 302: Use projection technology to enhance the signal of the scene delay signal matrix to obtain a projected scene delay signal matrix;

[0064] Step 303: Perform summation and superposition on the projected scene delay signal matrix to determine the echo signal after range migration correction.

[0065] In another exemplary embodiment of the present application, step 301 may include the following steps 401 to 406:

[0066] Step 401: Perform range-direction Fourier transform on the echo matrix to obtain an echo signal after range-direction Fourier transform;

[0067] Step 402: Perform matched filtering for range compression on the echo signal after range-direction Fourier transform to obtain filtered echo information;

[0068] Step 403: Perform inverse range-direction Fourier transform on the filtered echo information to obtain an echo signal after range compression;

[0069] Step 404: According to the slant range matrix constructed in step 203, use the formula t = r(η m ) / 2c to determine the time delay from the transmitted signal of SAR to the received reflected signal; where r(η m ) is the slant range matrix corresponding to each scene point in the observed scene at each sampling moment η m , and c is the speed of light;

[0070] Step 405: Find the echo signals of each scene point corresponding to the time delay from the echo signal after range compression;

[0071] Step 406: Construct a first empty matrix of Na×N 2 , and put the echo signals of each scene point corresponding to the time delay into the first empty matrix of Na×N 2 to obtain a scene delay signal matrix; where N a is the number of transmitted pulses, and N is the number of scene points of the observed scene along the x-axis or y-axis.

[0072] Among them, the specific implementation processes of steps 401 to 403 are as follows:

[0073] (1) First, perform a range-direction Fourier transform on the echo matrix S echo The Fourier transform in the embodiments of the present invention is different from the fast Fourier transform (FFT). The specific operation method is as follows:

[0074] Extract the first row of S echo and swap the positions of the first N r / 2 elements of this row matrix with the last N r / 2 elements, and then perform FFT on the row matrix after the position swap. After the FFT operation, swap the first half of the elements and the second half of the elements of the row matrix after FFT again, that is, complete the range-direction Fourier transform of the first row data of the echo signal. Perform the above operations on each row of data in the echo matrix in turn to complete the operation of performing the range-direction Fourier transform on the original echo signal matrix. Denote the signal after the range-direction Fourier transform as S echo,r_fft .

[0075] (2) Then construct a matched filter for range compression. The filter expression is as follows:

[0076]

[0077] Multiply each row of the signal S echo,r_fft after the range-direction Fourier transform with H rc point by point, and denote the multiplied two-dimensional matrix as S echo,rc_fft . Then, perform a zero-padding operation on both ends of the range direction dimension of S echo,rc_fft by α times. The matrix after zero-padding is Then perform an inverse range-direction Fourier transform to complete pulse compression. Denote the signal after completing range compression as S echo,rc , and S echo,rc is the output parameter of this module. The method of inverse range-direction Fourier transform is the same as the steps of range-direction Fourier transform, except that the FFT function used in the process is replaced with the IFFT function.

[0078] In step 302, the specific process of obtaining the projected scene delay signal matrix is as follows:

[0079] First, construct a projection matrix D (n,n) as In the formula, λ is the operating wavelength, and are the slant range coefficients of the (n, n)th target in the observed scene at random sampling times η 1 , η m and η Na respectively, and [·] TDenotes the transpose operation.

[0080] Then, according to the formula S echo,projection (n,n) = D (n,n )·([D (n,n) H ·S echo,interp (n,n)), project the scene time-delay signal matrix; where S echo,projection (n,n) is the nth 2 column of the scene time-delay signal after projection, [] H denotes the conjugate transpose operation, and S echo,interp (n,n) is the nth 2 column of the scene time-delay signal in the scene time-delay signal matrix.

[0081] Finally, construct a second empty matrix of Na×N 2 , and put the projected scene time-delay signal into the second empty matrix of Na×N 2 to obtain the projected scene time-delay signal matrix.

[0082] The specific process of step 303 is as follows: construct an empty matrix of Na×N, denoted as S echo,rcm . At each sampling moment η m , sum and stack the scene time-delay signals on the same range gate and put them into S echo,rcm . For example, for the first range gate, that is, when n = 1, take out and sum the data of all the scene points arranged at n = 1 and put them into the first column of the mth row of S echo,rcm . When it is the second range gate, take out and sum the data of all the scene points arranged at n = 2 and put them into the second column of the mth row of S echo,rcm . And so on, that is, complete the range migration correction. Denote the signal after range migration correction as S echo,rcm , and S echo,rcm is the output parameter of this module.

[0083] Before inputting the echo signal after the range migration correction into the trained hybrid norm deep unfolding network, the SAR imaging method based on the hybrid norm deep unfolding network further includes: obtaining a training data set; the training data set includes a number of observation scenes and the corresponding echo for each observation scene; training the hybrid norm deep unfolding network with the training data set to obtain the trained hybrid norm deep unfolding network.

[0084] Construct a training data set based on the radar operating parameters, and the process is as follows:

[0085] Construct a reconstruction matrix according to the slant range of the defined scene, which is expressed as follows:

[0086] ​

[0087] The two-dimensional matrix D is a reconstruction matrix obtained based on the slant range of the delineated scene; where D m,n = exp{-j4πR n (η m ) / λ}·ε(((acos(R 0 / R n (η m ))-0.886*λ / L a ))) represents the reconstruction element of the target point with the penetration time of η m and the slant range of R n (η m ). ε(·) represents the step function, c is the speed of light, that is, the element in the m-th row and n-th column of the reconstruction matrix D.

[0088] The corresponding parameter pairs are the training data set of the hybrid norm depth unfolding network (compressed sensing radar super-resolution imaging network). Where P is the total amount of data in the training set. is the observation scene corresponding to the p-th set of radar operating parameters, is the corresponding echo.

[0089] Next, construct the network structure, perform network mapping, initialize network parameters, and train the network.

[0090] The hybrid norm depth unfolding network includes K levels, and each level includes four sub-modules, namely the first regularization module, the sparse transformation module, the second regularization module, and the multiplier update module, and the results are output by the reconstruction module.

[0091] Step 1, network mapping:

[0092] 1) The first regularization module z (k) Obtain the iterative solution result of the first regularization module by solving the Tikhonov regularization equation.

[0093] The Tikhonov regularization equation is expressed in the form shown in Equation (10):

[0094]

[0095] Where λ and ρ are regularization parameters, and z is the solution obtained by regularization.

[0096] The iterative solution result of the first regularization module is expressed as follows:

[0097]

[0098] Where z(k) denotes the iterative solution result of the first regularization module at the k-th iteration, where k represents the number of iterations, and ρ (k) and λ (k) denote the optimization parameters at the current k-th iteration, and σ (k-1) and y (k-1) respectively denote the iterative update results of the observation scene σ and the Lagrange multiplier y at the (k - 1)-th iteration of the previous layer.

[0099] 2) Sparse transformation module Γ(·) (k) : This sparse transformation module obtains a sparse solution through the non-linear sparse transformation domain Γ(·).

[0100] Γ(·) (k) denotes the scene sparse transformation domain of the k-th layer. For the sparse transformation domain Γ(·) (k) , in order to enhance the network representation ability and general approximation characteristics, this application constructs the sparse transformation module through a Convolutional Neural Networks (CNN), and designs a general non-linear sparse transformation function. In this application, the sparse transformation module consists of three convolutional layers (without bias term settings). Among them, the first convolutional layer and the third convolutional layer both correspond to 1 filter with a size of 3×3. The second convolutional layer corresponds to 32 filters with a size of 3×3. Mathematically, the sparse transformation is invertible, and the symmetric network Γ -1 (·) (k) is represented by a mirror symmetry framework, and Γ(·) (k) ·Γ -1 (·) (k) =Ι, where Ι is the identity matrix.

[0101] In addition, since the radar system reconstruction scene belongs to a complex operation, a real-part - imaginary-part separated model (real-complex model) is established, which is specifically expressed as follows:

[0102]

[0103]

[0104] In the formula, D mn is the reconstruction element of the target point with a crossing time of η m and a slant range of R n (η m ). (D mn ) real is the real part of the reconstruction element, and (D mn ) imag is the imaginary part of the reconstruction element; (z (k) ) real , (z (k) ) imagThey are respectively the real part and the imaginary part of the iterative solution result of the first regularization module at the k-th time.

[0105] According to the above real-complex number model, it is divided into two channels for training and learning.

[0106] 3) The second regularization module σ (k) : This second regularization module solves the L 1 / 2 norm equation to achieve good reconstruction performance, thus facilitating the recovery and reconstruction of targets with low signal-to-noise ratio.

[0107] L 1 / 2 The norm equation is expressed in the form shown in Equation (14):

[0108]

[0109] In the above formula, s represents the echo signal S after range migration correction in step 303 echo,rcm , and D is the reconstruction matrix obtained according to the slant range of the defined scene in formula (10).

[0110] The iterative solution result of the second regularization module is expressed as follows:

[0111]

[0112] Among them,

[0113]

[0114]

[0115]

[0116]

[0117] In the above formula, σ (k) represents the iterative solution result of the second regularization module at the k-th time, is an intermediate variable, u (k) represents an optimization iteration parameter in the k-th iteration process, D H represents the conjugate transpose of the reconstruction matrix D, represents the result of the backscattering coefficient of the i-th scene in the k-th iteration process, is an intermediate variable, is an intermediate variable, represents the arccosine function, usually taking values in ρ (k) is the coefficient of the quadratic penalty term.

[0118] 4) The multiplier update module y (k):The multiplier update module is used to update the Lagrange multiplier y (k) 。

[0119] y (k) =y (k-1) +τ (k) ρ (k) (z k -σ k ) (21)。

[0120] Step 2, Initialization.

[0121] Initialization of network input: Among them, is the initial input of the hybrid norm deep unfolding network.

[0122] Initialization of network parameters: τ (0) =1, ρ (0) =0.5, Among them, τ (0) is the initial iteration step size, ρ (0) and λ (0) are the coefficients of the initial quadratic penalty term.

[0123] The network parameters of each layer of the hybrid norm deep unfolding network are learned through end-to-end training, rather than being fixed.

[0124] Step 3, Train the hybrid norm deep unfolding network based on the dataset Train the hybrid norm deep unfolding network.

[0125] In this application, in the loss function, the reconstruction error and sparsity constraint are considered, and the optimization parameter κ is used to control the balance between the two constraint terms. The loss function is expressed as follows:

[0126]

[0127] Among them, Loss{} represents the loss function, τ (k) is the iteration step size, σ is the observation scene, is the reconstructed scene corresponding to the observation scene, K is the number of network layers of the hybrid norm deep unfolding network, and κ represents the regularization parameter used to balance the first 2-norm term and the second 2-norm term.

[0128] After determining the loss function, the parameters can be autonomously learned, and then the trained hybrid norm deep unfolding network can be obtained.

[0129] After obtaining the trained hybrid norm deep unfolding network, the migrated corrected echo signal S echo,rcm , is loaded into the trained hybrid norm deep unfolding network, and the required reconstructed scene can be obtained.

[0130] The present application also provides an application scenario, which applies the above SAR imaging method based on the hybrid norm deep unfolding network. Specifically: The SAR imaging method based on the hybrid norm deep unfolding network provided in this embodiment can be applied in the SAR imaging scenario. The SAR imaging scenario includes a parameter acquisition link and a SAR imaging link; the radar operating parameters enter the SAR imaging link from the parameter acquisition link to obtain the corresponding reconstructed scenario. The SAR imaging method based on the hybrid norm deep unfolding network provided in this embodiment belongs to the SAR imaging link. Specifically, in the process of the SAR imaging link for the radar operating parameters, the echo matrix can be obtained according to the SAR raw echo signals collected at all sampling times, the observation scenario of the SAR can be determined, and the slant range matrix corresponding to each scene point in the observation scenario at each random sampling time can be constructed. The echo signal after range migration correction is determined according to the echo matrix and the slant range matrix, and the echo signal after range migration correction is input into the trained hybrid norm deep unfolding network to obtain the reconstructed scenario.

[0131] Based on the same inventive concept, the embodiment of the present application also provides a SAR imaging device based on the hybrid norm deep unfolding network for implementing the above-mentioned SAR imaging method based on the hybrid norm deep unfolding network. The implementation solutions provided by this device to solve problems are similar to the implementation solutions described in the above method. Therefore, the specific limitations in one or more embodiments of the SAR imaging device based on the hybrid norm deep unfolding network provided below can refer to the limitations on the SAR imaging method based on the hybrid norm deep unfolding network in the above text, and will not be repeated here.

[0132] In an exemplary embodiment, as Figure 3 shown, a SAR imaging device based on the hybrid norm deep unfolding network is provided, including:

[0133] A radar operating parameter acquisition module T1, configured to: acquire radar operating parameters; the radar operating parameters include the reference slant range of the scene center, the satellite flight speed, the azimuth antenna size, the number of transmitted pulses, the chirp rate of the transmitted signal, the time vector of the radar transmitted pulses, and the operating wavelength;

[0134] An echo matrix acquisition module T2, configured to: acquire an echo matrix according to the SAR raw echo signals collected at all sampling times;

[0135] A slant range matrix construction module T3, configured to: determine the observation scenario of the SAR and construct the slant range matrix corresponding to each scene point in the observation scenario at each random sampling time;

[0136] An echo signal determination module T4 after range migration correction, configured to: determine the echo signal after range migration correction according to the echo matrix and the slant range matrix;

[0137] The scene reconstruction module T5 is used to: input the range migration corrected echo signal into the trained hybrid norm deep unfolding network to obtain a reconstructed scene; the hybrid norm deep unfolding network includes K levels, and each level includes four sub-modules, namely the first regularization module, the sparse transformation module, the second regularization module, and the multiplier update module, and outputs the result by the reconstruction module; in any level, the data input from the previous level obtains a reconstructed scene through the reconstruction module, and the first regularization module is used to solve the Tikhonov regularization equation to obtain an iterative solution result; the sparse representation of the iterative solution result is obtained through the sparse transformation module; the second regularization module is used to solve the L 1 / 2 regularization equation to obtain an iterative solution result; the multiplier update module is used to implement the update of the Lagrange multiplier.

[0138] In an exemplary embodiment, a computer device is provided. The computer device can be a server or a terminal, and its internal structure diagram can be as Figure 4 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store SAR imaging processing data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements a SAR imaging method based on a hybrid norm deep unfolding network.

[0139] Those skilled in the art can understand that Figure 4 the structure shown in

[0140] In an exemplary embodiment, a computer device is further provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0141] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, which, when executed by a processor, implements the steps in the above-mentioned method embodiments.

[0142] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0143] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned method embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned method embodiments. Among them, any reference to a memory, database, or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tapes, floppy disks, flash memories, optical memories, high-density embedded non-volatile memories, resistive random access memories (ReRAMs), magnetoresistive random access memories (MRAMs), ferroelectric random access memories (FRAMs), phase change memories (PCMs), graphene memories, etc. Volatile memories can include random access memory (RAM) or external cache memories, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0144] The databases involved in the embodiments provided in this application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in this application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logics, data processing logics based on quantum computing, etc., without limitation.

[0145] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0146] In this article, specific examples are used to elaborate on the principles and implementation manners of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A SAR imaging method based on a mixed norm deep unfolding network, characterized in that: The SAR imaging method based on the mixed norm deep expansion network includes: Acquire radar operating parameters; the radar operating parameters include scene center reference slant distance, satellite flight speed, azimuth antenna size, number of transmitted pulses, modulation frequency of transmitted signals, time vector of radar transmitted pulses and operating wavelength; Acquire the echo matrix according to the SAR original echo signals collected at all sampling moments; Determine the SAR observation scene and construct the corresponding slant range matrix of each scene point in the observation scene at each random sampling moment; Determine the echo signal after range migration correction according to the echo matrix and the slant range matrix; The echo signal after the range migration correction is input into the trained hybrid norm deep expansion network to obtain a reconstructed scene; the hybrid norm deep expansion network includes K levels, each level includes four submodules, namely a first regularization module, a sparse transformation module, a second regularization module and a multiplier update module, and the reconstruction module outputs the result; in any level, the data input in the previous level is used to obtain a reconstructed scene through the reconstruction module, the first regularization module is used to solve the Tikhonov regularization equation to obtain an iterative solution result; the sparse representation of the iterative solution result is obtained through the sparse transformation module; the second regularization module is used to solve L 1 / 2 Regularize the equation and obtain the iterative solution result; the multiplier update module is used to implement the update of the Lagrange multiplier.

2. The SAR imaging method based on the mixed norm deep expansion network according to claim 1 is characterized in that: The iterative solution result of the first regularization module is expressed as follows: Among them, z (k) represents the k-th iteration solution result of the first regularization module, k represents the number of iterations, ρ (k) and λ (k) represents the optimization parameter of the current k-th iteration, σ (k-1) and (k-1) They respectively represent the iterative update results of the observation scene σ and the Lagrange multiplier y of the k-1th iteration of the previous layer.

3. The SAR imaging method based on the mixed norm deep expansion network according to claim 1 is characterized in that: The sparse transformation module consists of three convolutional layers.

4. The SAR imaging method based on the mixed norm deep expansion network according to claim 1 is characterized in that: The iterative solution results of the second regularization module are expressed as follows: in, In the formula, σ (k) represents the k-th iteration solution result of the second regularization module, where k represents the number of iterations; is the intermediate variable, λ (k) represents the optimization parameter of the current k-th iteration, u (k) represents an optimized iteration parameter in the kth iteration process, σ (k-1) represents the iterative update result of the observation scene σ of the k-1th iteration of the previous layer, D H represents the conjugate transpose of the reconstruction matrix D, s represents the echo signal after range migration correction, ρ (k) is the coefficient of the quadratic penalty term, Γ(·) (k) represents the scene sparse transform domain of the kth layer, z (k) represents the k-th iteration solution of the first regularization module, y (k-1) represents the iterative update result of the Lagrange multiplier y of the k-1th iteration of the previous layer, is the intermediate variable, represents the result of the backscattering coefficient of the i-th scene during the k-th iteration, is the intermediate variable, Represents the inverse cosine function.

5. The SAR imaging method based on the mixed norm deep expansion network according to claim 1 is characterized in that: Determining the echo signal after range migration correction according to the echo matrix and the slant range matrix specifically includes: Determining a scene delay signal matrix according to the echo matrix and the slant range matrix; Using projection technology, the scene delay signal matrix is ​​enhanced to obtain a projected scene delay signal matrix; The projected scene time-delay signal matrix is ​​summed and superimposed to determine the echo signal after range migration correction.

6. The SAR imaging method based on the mixed norm deep expansion network according to claim 5 is characterized in that: Determining a scene delay signal matrix according to the echo matrix and the slant range matrix specifically includes: Performing a range-to-Fourier transform on the echo matrix to obtain an echo signal after the range-to-Fourier transform; Performing distance-compressed matched filtering on the echo signal after the distance Fourier transform to obtain filtered echo information; Performing a range-direction inverse Fourier transform on the filtered echo information to obtain a range-compressed echo signal; According to the slope distance matrix, using the formula t = r (η m ) / 2c, determine the time delay from SAR transmitting signal to receiving reflected signal; where r(η m ) is at each sampling time η m The slant distance matrix corresponding to each scene point in the observation scene, c is the speed of light; Find the echo signals of each scene point corresponding to the time delay from the echo signals after distance compression; Construct a Na×N 2 The first empty matrix, and the echo signals of each scene point corresponding to the delay are placed in Na×N 2 The first empty matrix of the scene delay signal matrix is ​​obtained; where N a is the number of transmitted pulses, and N is the number of scene points along the x-axis or y-axis of the observed scene.

7. The SAR imaging method based on a mixed norm deep expansion network according to claim 1, characterized in that: Before inputting the range migration corrected echo signal into the trained hybrid norm deep expansion network, the SAR imaging method based on the hybrid norm deep expansion network further includes: Acquire a training data set; the training data set includes a plurality of observation scenes and an echo corresponding to each observation scene; The training data set is used to train the mixed norm deep expansion network to obtain a trained mixed norm deep expansion network.

8. A SAR imaging device based on a mixed norm deep expansion network, characterized in that: The SAR imaging device based on the mixed norm deep expansion network includes: The radar working parameter acquisition module is used to: acquire radar working parameters; the radar working parameters include scene center reference slant distance, satellite flight speed, azimuth antenna size, number of transmitted pulses, modulation frequency of transmitted signal, time vector of radar transmitting pulse and working wavelength; The echo matrix acquisition module is used to: acquire the echo matrix according to the SAR original echo signals collected at all sampling moments; The slant range matrix construction module is used to: determine the SAR observation scene and construct the corresponding slant range matrix of each scene point in the observation scene at each random sampling moment; A module for determining an echo signal after range migration correction, used to: determine an echo signal after range migration correction according to the echo matrix and the slant range matrix; The scene reconstruction module is used to: input the echo signal after the range migration correction into the trained hybrid norm deep expansion network to obtain the reconstructed scene; the hybrid norm deep expansion network includes K levels, each level includes four submodules, namely the first regularization module, the sparse transformation module, the second regularization module and the multiplier update module, and the reconstruction module outputs the result; in any level, the data input in the previous level is used to obtain the reconstructed scene through the reconstruction module, the first regularization module is used to solve the Tikhonov regularization equation to obtain the iterative solution result; the sparse representation of the iterative solution result is obtained through the sparse transformation module; the second regularization module is used to solve L 1 / 2 Regularize the equation and obtain the iterative solution result; the multiplier update module is used to implement the update of the Lagrange multiplier.

9. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the SAR imaging method based on a mixed norm deep unfolding network according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the SAR imaging method based on a mixed norm deep expansion network described in any one of claims 1 to 7 is implemented.

Images