MIMO radar high-resolution self-focusing learning imaging method

Abstract

The invention provides an MIMO radar high-resolution self-focusing learning imaging method. The method comprises the following steps: constructing an MIMO radar learning imaging training data set; an MIMO radar learning imaging network LI-Net is constructed; performing channel processing on an output image XLI of the LI-Net, namely splitting a single-channel complex value image XLI into a real part and an imaginary part, and converting the single-channel complex value image XLI into a dual-channel real value image XIN; an MIMO radar learning imaging network IF-Net is constructed; and training the MIMO radar self-focusing learning imaging network. According to the method, the LI-Net is designed based on a two-dimensional analysis-learning near-end operator method, a gate loop unit is introduced into each layer of the LI-Net, and the sparse recovery performance better than that of a two-dimensional AePOM method is obtained; designing an IF-Net based on a stacked convolution auto-encoder SCAE, processing a pre-image obtained by an LI-Net, and converting a partially focused low-resolution image into a fully focused high-resolution image; a jump connection operation is embedded in the IF-Net so as to fuse different features in the whole processing process to improve the performance of image transformation. The method is used for realizing MIMO radar target high-resolution imaging under the condition of imaging model mismatching.

Description

Technical Field

[0001] The present invention relates to the field of radar technology, and in particular to a high-resolution self-focusing learning imaging method for a multiple-input and multiple-output (MIMO) radar. Background Art

[0002] The MIMO radar target imaging method based on compressed sensing (CS) has the advantages of high resolution and low sampling cost. Traditionally, the CS imaging method uses a one-dimensional sparse recovery (SR) algorithm to solve the sparse vector recovery problem established by the one-dimensional imaging model of the MIMO radar to obtain a two-dimensional high-resolution image of the target. However, since the two-dimensional signal matrix and the target image need to be vectorized, the imaging process of the one-dimensional method is very time-consuming, especially in high-resolution and large imaging scenarios. To solve this problem, after approximating the radar echo signal, the CS imaging method uses a two-dimensional SR algorithm to solve the two-dimensional sparse matrix recovery problem established by the two-dimensional approximate imaging model of the MIMO radar, thereby reducing the computational complexity and improving the imaging efficiency.

[0003] In practice, the CS imaging method has a more thorny problem, that is, the parameters of the SR algorithm need to be manually adjusted. If the parameters are not appropriate, the SR algorithm may require multiple iterations to obtain satisfactory results, thereby increasing the operation time. To overcome this problem, deep learning (DL) technology is applied to the CS imaging method. By unfolding the SR algorithm into a deep learning network, the deep unfolding network (DUN) can automatically learn its optimal parameters from the training data, thereby optimizing the iteration process and improving the imaging quality and efficiency. However, when the established signal model does not match the actual situation, the performance of radar imaging methods based on SR, DUN, etc. is significantly reduced. For example, for MIMO radar imaging, when the target is in the far field and its azimuth angle is small, a two-dimensional sparse matrix recovery model can be established, and two-dimensional SR, DUN and other methods can be used to obtain a well-focused high-resolution target image. However, when the target is in the near field or the azimuth angle is large, the model error caused by the signal approximation is large, and the sparse matrix recovery model is no longer valid. In this case, the model mismatch will cause the target image obtained by two-dimensional SR, DUN and other methods to be defocused. Therefore, studying the high-resolution self-focusing learning imaging method under model mismatch conditions is of great significance for improving the target imaging and recognition performance of MIMO radar. Summary of the invention

[0004] In view of the problems existing in the prior art, the present invention provides a MIMO radar high-resolution self-focusing learning imaging method, which specifically includes the following steps:

[0005] Step 1: Construct a MIMO radar learning imaging training data set;

[0006] The specific steps include:

[0007] Step 1: Determine the MIMO radar system parameters;

[0008] Establish a plane rectangular coordinate system, with the positive direction of the X axis pointing to the east and the positive direction of the Y axis pointing to the north. m 、x n are the x-axis coordinates of the mth transmitting array element and the nth receiving array element, respectively. m=1,2,…,M,n=1,2,…,N,M and N represent the number of transmitting array elements and receiving array elements of the MIMO radar, respectively. p is the x-axis coordinate of the pth virtual array element, p = 1, 2, ..., MN, r k ,θ k are the distance and angle of the kth target scattering point relative to the radar center, k = 1, 2, ..., K, K represents the number of target scattering points; according to actual needs, determine the number of transmitting array elements M, the number of receiving array elements N, and the position of the transmitting array element (x m ,0), receiving array element position (x n ,0), and the starting frequency f1 and center frequency f of the step-frequency continuous wave signal c , frequency step Δf and frequency number Q;

[0009] Step 2: Determine the MIMO radar imaging space;

[0010] According to the possible location of the target, the range of the MIMO radar imaging [r min ,r max ] and angular imaging range [θ min ,θ max ], where r min and r max are the minimum imaging distance and the maximum imaging distance, θ min and θ max are the minimum imaging angle and the maximum imaging angle respectively;

[0011] Step 3: Divide the imaging space and construct the measurement matrix;

[0012] The MIMO radar imaging space is divided into I×J grids, where I and J are the number of grids divided in distance and angle respectively, and the distance full sampling measurement matrix Ψ1 and the angle full sampling measurement matrix Ψ2 in the pseudo-polar coordinate system are constructed;

[0013]

[0014]

[0015] Among them, P = MN is the number of virtual array elements, α i =2r i Δf / c is the i-th distance coordinate in the pseudo-polar coordinate system, β j = -2dsinθ j / λ c is the jth angle coordinate in the pseudo-polar coordinate system, c is the speed of light, λ c =c / f c is the center frequency f c The corresponding wavelength, i = 1, 2, ..., I, j = 1, 2, ..., J, d is the radar virtual array element spacing, r i =r min +(i-1)(r max -r min ) / (I-1) and θ j =θ min +(j-1)(θ max -θ min ) / (J-1) are the i-th discrete distance and j-th discrete angle respectively;

[0016] Randomly select frequency points and virtual array elements, extract row vectors corresponding to the selected frequency points and virtual array elements from the range and angle full sampling measurement matrices Ψ1 and Ψ2 respectively, and construct the range measurement matrix Φ1 and the angle measurement matrix Φ2;

[0017] Step 4: Calculate the optimal matrix based on the measurement matrix;

[0018] Given the distance and angle measurement matrices Φ1 and Φ2, the gradient projection method is used to solve equation (3) to calculate the distance and angle optimal matrices B1 and B2;

[0019]

[0020] Among them, ||·|| F is the Frobenius norm of the matrix, (·) 2 represents the square, diag(·) is the vector of matrix diagonal elements, 1 I and 1 J are all-1 vectors of size I and J respectively, (·) H represents conjugate transpose;

[0021] Step 5: Randomly generate an ideal target image, calculate the MIMO radar echo signal matrix, and generate a training data set;

[0022] The number of targets is randomly selected within a certain numerical range, and the amplitude of each target is randomly determined based on a Gaussian distribution with a certain mean and standard deviation. min ,r max ] and [θ min ,θ max ] to randomly set the position of each target and generate O ideal target images Calculate MIMO radar echo signal matrix Where O is the size of the training dataset, is the oth target ideal image, is the oth MIMO radar signal matrix, o=1,2,...,O;

[0023] Step 2: Construct the MIMO radar learning imaging network LI-Net;

[0024] Construct an L-layer deep neural network LI-Net with the same structure in each layer, whose input is the MIMO radar echo signal matrix S input , distance measurement matrix Φ1, angle measurement matrix Φ2, distance optimal matrix B1, angle optimal matrix B2, target image initialization result X (0) =0, the hidden state initialization result of the gate recurrent unit h0=0, S input Indicates any one The output is the target image X LI , specifically including the following steps:

[0025] Step 1: Calculate the approximate value 1r of the imaging error of the l-1th layer of the LI-Net network l-1 And the approximate error 2u l-1 ;

[0026] According to the target image result X output by the l-1 layer of the LI-Net network (l-1) , two approximate values ​​of imaging error r are calculated by formula (4) l-1 and u l-1 , jointly measure the imaging effect;

[0027]

[0028] Where, l = 1, 2, ..., L;

[0029] Step 2: Calculate the hidden state h of the gated recurrent unit GRU of the lth layer of the LI-Net network l ;

[0030] According to the imaging error approximation r l-1 and u l-1 And the hidden state h of the GRU in the l-1 layer of the LI-Net networkl-1 , get the hidden state h of the l-th layer GRU l ;

[0031] h l =GRU(h l-1 ,r l-1 ,u l-1 )(5)

[0032] Among them, GRU(·) is a gate recurrent operation;

[0033] Step 3: Calculate the threshold θ of the lth layer of the LI-Net network l,X and step size γ l,X ;

[0034] Using the softplus function and the relu function, according to formula (6), the hidden state h of the GRU in the lth layer of the LI-Net network is l Processing is performed to obtain the threshold θ of the lth layer of the LI-Net network l,X and step size γ l,X ;

[0035]

[0036] Among them, for any real number z, softplus(z) = log(1+exp(z)), log(·) is the natural logarithm, exp(·) is the natural exponential, relu(z) = max(0,z), max(·) is the maximum value, and are the transformation matrices of the threshold and step size of the lth layer of the LI-Net network, and are the transformation vectors of the threshold and step size of the lth layer of the LI-Net network, and are the bias vectors of the threshold and step size of the lth layer of the LI-Net network, and They are the bias constants for the threshold and step size of the lth layer of the LI-Net network respectively;

[0037] Step 4: Calculate the target image result X output by the first layer of the LI-Net network (l) ;

[0038] Based on the MIMO radar signal matrix S, the distance measurement matrix Φ1, the angle measurement matrix Φ2, the distance optimal matrix B1, the angle optimal matrix B2, the threshold θ l,X and step size γ l,X , calculate the target image result X output by the lth layer of the LI-Net network (l) ;

[0039]

[0040] Among them, X (l-1) Represents the output of the l-1th layer of the network, and the sparseness-induced penalty parameter η l,X =1 / (4θ l,X ),

[0041] is the proximal operator of the minimum-maximum concave penalty function, calculated by formula (8);

[0042]

[0043] Where z represents any complex number, sgn(·) represents a complex sign function;

[0044] Step 5: According to steps 1 to 4, calculate from the 1st layer to the Lth layer to obtain the initial target image X LI , that is, X LI =X (L) ;

[0045] Let the output of LI-Net layer L-1 be X (L-1) , the threshold and step size of the Lth layer are θ L,X and γ L,X , calculate the output X of the Lth layer of LI-Net (L) , get the initial target image X LI ;

[0046]

[0047] Step 3: Output image X of LI-Net LI Perform channel processing, that is, the single channel complex valued image X LI Split into real and imaginary parts and convert into a two-channel real-valued image X IN ;

[0048] X IN =[Re(X LI );Im(X LI )](10)

[0049] Where Re(·) represents the real part of the matrix, and Im(·) represents the imaginary part of the matrix;

[0050] Step 4: Construct MIMO radar learning imaging network IF-Net;

[0051] IF-Net is constructed based on convolutional neural network CNN, maximum pooling MaxPool layer, unmaximum pooling MaxUnpool layer, skip connection, etc., which is used to convert a partially defocused low-resolution target image into a fully focused high-resolution image. Specifically, the following steps are included:

[0052] Step 1: Output X of LI-Net after channel processing IN Perform encoding processing to obtain the encoding output X EN ;

[0053] X IN Input to the first convolutional neural network (CNN) module for processing to obtain the output of the first CNN module The CNN module consists of a convolutional layer Conv, a batch normalization layer BN, and a ReLU layer;

[0054] Will Input to the first MaxPool layer for processing to obtain the first MaxPool layer output and the first largest index result

[0055] Will Input to the second CNN module for processing and obtain the output of the second CNN module Will Input to the second MaxPool layer for processing, and obtain the second MaxPool layer output and the second largest index result Will Input to the third CNN module for processing to obtain the encoded output X EN ;

[0056] Step 2: Encode the output X EN Perform decoding processing to obtain the decoded output X DE ;

[0057] X EN Input to the first CNN module containing the deconvolution layer ConvTrans for processing, and the result is the same as Perform dot product multiplication to obtain the output of the first CNN module containing the deconvolution layer Will and Input to the first MaxUnpool layer for processing, the result is the same as Perform dot product multiplication to obtain the first MaxUnpool layer output Will Input to the second CNN module containing ConvTrans for processing, and the results are the same as Perform dot product multiplication to obtain the output of the second CNN module containing the deconvolution layer Will and The input is processed by the second MaxUnpool layer, and the result is the same as Perform dot product multiplication to obtain the second MaxUnpool layer output Will Input to the third CNN module containing ConvTrans for processing to obtain the decoded output X DE ;

[0058] Step 3: X IN and X DE Perform stitching and input into the generator to obtain the final target image;

[0059] The encoder input X is connected by channel cascading IN With the decoder output X DE The splicing is performed in the third dimension, and the result is input into the generator composed of CNN modules to obtain the final target image;

[0060] Step 5: Train the MIMO radar self-focusing learning imaging network;

[0061] The self-focusing learning imaging network AFLI-Net, which consists of two sub-networks, LI-Net and IF-Net, was trained using the "pre-training" and "re-training" methods. In the "pre-training" stage, the learning imaging network LI-Net and the image focusing network IF-Net were trained independently in turn. In the "re-training" stage, the two network parameters learned in the previous stage were used as the initial parameters of AFLI-Net, and the entire network was trained end-to-end.

[0062] The specific steps include:

[0063] Step 1: Pre-train LI-Net;

[0064] Given a training dataset And the number of network layers L, using the back propagation method, minimize the loss function L of LI-Net LI-Net , get the optimal network parameters of LI-Net That is, the transformation matrix of the threshold of each layer and the transformation matrix of the step size Transformation vector for threshold and step size and Bias vector for threshold and step size and Bias constants for threshold and step size and And the GRU hidden state h l (l=1,2,...,L);

[0065]

[0066] in, Indicates As input, Θ LI-Net is the parameterized LI-Net output, Θ LI-Net Represents LI-Net network parameters;

[0067] Step 2: Pre-train IF-Net;

[0068] Given a training dataset in Based on the optimal network parameters The resulting LI-Net network output The abbreviation of , using the back propagation method to minimize the loss function L of IF-Net IF-Net , get the optimal network parameters of IF-Net That is, the weights and biases of each Conv layer and the parameters of each BN layer

[0069]

[0070] in, is the oth LI-Net output image after channel processing, For is the IF-Net output of the input, Θ IF-Net represents the IF-Net network parameters; solve equation 12 to obtain the optimal parameters;

[0071] Step 3: End-to-end training of AFLI-Net;

[0072] Load the pre-trained parameters of LI-Net and IF-Net as the initialization network parameters of AFLI-Net, and use the back-propagation method to minimize the AFLI-Net loss function L AFLI-Net , perform end-to-end training on AFLI-Net to obtain its optimal network parameters;

[0073]

[0074] in, represents the network parameters of AFIT-Net, Indicates As input, Θ AFLI-Net AFLI-Net output for network parameters;

[0075] Step 6: Use the trained AFLI-Net to process the actual MIMO radar signal matrix S to obtain a fully focused, high-resolution target image X.

[0076] In a specific embodiment of the present invention, in step 5 of step 1, the number of targets is randomly selected within a numerical range of 1 to 10, and the amplitude of each target is randomly determined based on a Gaussian distribution with a mean of 0 and a standard deviation of 1.

[0077] In step 4 of another specific embodiment of the present invention:

[0078] Step 1: Output X of LI-Net after channel processing IN Perform encoding processing to obtain the encoding output X EN ;

[0079] X IN Input to the first convolutional neural network (CNN) module for processing to obtain the output of the first CNN module The CNN module consists of a convolutional layer Conv, a batch normalization layer BN, and a ReLU layer. The Conv kernel size is 5×5 and the stride is 1.

[0080] Will Input to the first MaxPool layer for processing to obtain the first MaxPool layer output and the first largest index result The window size of MaxPool is 2×2 and the step size is 1;

[0081] Will Input to the second CNN module for processing and obtain the output of the second CNN module The Conv kernel size is 3×3 and the step size is 1; Input to the second MaxPool layer for processing, and obtain the second MaxPool layer output and the second largest index result Will Input to the third CNN module for processing to obtain the encoded output X EN , the Conv kernel size is 3×3 and the stride is 1;

[0082] Step 3: X IN and X DE Perform stitching and input into the generator to obtain the final target image;

[0083] The encoder input X is connected by channel cascading IN With the decoder output X DEThe concatenation is performed in the third dimension, and the result is input into a generator consisting of a CNN module with a convolution kernel size of 1×1 and a step size of 1 to obtain the final target image.

[0084] The present invention is used to realize high-resolution imaging of MIMO radar targets under the condition of imaging model mismatch, and has the following advantages:

[0085] 1) An Auto-Focused Learning Imaging Network (AFLI-Net) for high-resolution imaging of MIMO radar targets is proposed, which consists of a Learning Imaging Network (LI-Net) and an Image Focusing Network (IF-Net). LI-Net is used to obtain a pre-image of a target with local defocus, and IF-Net is used to focus the defocus in the pre-image, and finally outputs a high-resolution image of the target.

[0086] 2) LI-Net is designed based on the two-dimensional Analytical Learned Proximal Operator Method (ALePOM), and the Gate Recurrent Unit (GRU) is introduced in each layer to adaptively calculate the threshold and step size for each training data, thereby obtaining better sparse recovery performance than the two-dimensional ALePOM method.

[0087] 3) IF-Net is designed based on Stacked Convolutional Auto-Encoder (SCAE), which aims to process the pre-image obtained by LI-Net and convert the partially focused low-resolution image into a fully focused high-resolution image. Skip connection operations are embedded in IF-Net to fuse different features in the entire processing process to improve the performance of image transformation. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 It is a framework for high-resolution self-focusing learning imaging methods for MIMO radar;

[0089] Figure 2 It is a one-dimensional linear array model of MIMO radar imaging;

[0090] Figure 3 It is the MIMO radar learning imaging network (LI-Net) structure;

[0091] Figure 4 It is the MIMO radar image focusing network (IF-Net) structure. DETAILED DESCRIPTION

[0092] The present invention will be further described below in conjunction with the accompanying drawings.

[0093] The present invention provides a MIMO radar high-resolution self-focusing learning imaging method, the framework of which is as follows: Figure 1 As shown, the specific steps include the following steps.

[0094] Step 1: Construct a MIMO radar learning imaging training data set;

[0095] The specific steps include:

[0096] Step 1: Determine the MIMO radar system parameters;

[0097] Establish as Figure 2 In the plane rectangular coordinate system shown in the figure, the positive direction of the X axis points to the east, and the positive direction of the Y axis points to the north. m 、x n are the x-axis coordinates of the mth transmitting array element and the nth receiving array element, respectively. m=1,2,…,M,n=1,2,…,N,M and N represent the number of transmitting array elements and receiving array elements of the MIMO radar, respectively. p is the x-axis coordinate of the pth virtual array element, p = 1, 2, ..., MN ("MN" is the product of the two), r k ,θ k are the distance and angle of the kth target scattering point relative to the radar center, k = 1, 2, ..., K, K represents the number of target scattering points. According to actual needs, determine the number of transmitting array elements M, the number of receiving array elements N, and the position of the transmitting array element (x m ,0), receiving array element position (x n ,0), and the starting frequency f1 and center frequency f of the step-frequency continuous wave signal c , frequency step Δf and frequency number Q.

[0098] Step 2: Determine the MIMO radar imaging space;

[0099] According to the possible location of the target, the range of the MIMO radar imaging [r min ,r max ] and angular imaging range [θ min ,θ max ], where r min and r max are the minimum imaging distance and the maximum imaging distance, θ min and θ max are the minimum imaging angle and the maximum imaging angle respectively.

[0100] Step 3: Divide the imaging space and construct the measurement matrix;

[0101] The MIMO radar imaging space is divided into I×J grids, where I and J are the number of grids divided in distance and angle, respectively, to construct a full sampling measurement matrix Ψ1 in the distance direction and a full sampling measurement matrix Ψ2 in the angle direction in a pseudo-polar coordinate system (known to those skilled in the art and will not be described again).

[0102]

[0103]

[0104] Among them, P = MN is the number of virtual array elements, α i =2r i Δf / c is the i-th distance coordinate in the pseudo-polar coordinate system, β j = -2dsinθ j / λ c is the jth angle coordinate in the pseudo-polar coordinate system, c is the speed of light, λ c =c / f c is the center frequency f c The corresponding wavelength, i = 1, 2, ..., I, j = 1, 2, ..., J, d is the radar virtual array element spacing, r i =r min +(i-1)(r max -r min ) / (I-1) and θ j =θ min +(j-1)(θ max -θ min ) / (J-1) are the i-th discrete distance and j-th discrete angle respectively.

[0105] Randomly select frequency points and virtual array elements, extract row vectors corresponding to the selected frequency points and virtual array elements from the range and angle full sampling measurement matrices Ψ1 and Ψ2 respectively, and construct the range and angle measurement matrices Φ1 and Φ2.

[0106] Step 4: Calculate the optimal matrix based on the measurement matrix;

[0107] Given the distance and angle measurement matrices Φ1 and Φ2, the problem shown in equation (3) is solved using the gradient projection method well known in the art to calculate the distance and angle optimal matrices B1 and B2.

[0108]

[0109] Among them, ||·|| F is the Frobenius norm of the matrix, (·) 2 represents the square, diag(·) is the vector of matrix diagonal elements, 1 I and 1 Jare all-1 vectors of size I and J respectively, (·) H represents the conjugate transpose.

[0110] Step 5: Randomly generate ideal target images, calculate the MIMO radar echo signal matrix, and generate a training data set.

[0111] The number of targets is randomly selected within a certain numerical range (for example, the numerical range is 1 to 10), and the amplitude of each target is randomly determined based on a Gaussian distribution with a certain mean and standard deviation (for example, the mean and standard deviation are 0 and 1 respectively). min ,r max ] and [θ min ,θ max ] to randomly set the position of each target and generate O ideal target images (The specific generation method is well known to those skilled in the art and will not be repeated here) Calculate the MIMO radar echo signal matrix (The specific calculation method is well known to those skilled in the art and will not be described again), where O is the size of the training data set, is the oth target ideal image, is the o-th MIMO radar signal matrix, o=1,2,...,O.

[0112] Step 2: Construct a MIMO radar learning imaging network (LI-Net);

[0113] like Figure 3 As shown in Figure 2, an L-layer deep neural network LI-Net with the same structure in each layer is constructed, and its input is the MIMO radar echo signal matrix S input (S input Indicates any one ), distance measurement matrix Φ1, angle measurement matrix Φ2, distance optimal matrix B1, angle optimal matrix B2, target image initialization result X (0) = 0, the hidden state of the gate recurrent unit is initialized as h0 = 0, and the output is the target image X LI , specifically including the following steps:

[0114] Step 1: Calculate the approximate value 1r of the imaging error of the l-1th layer of the LI-Net network l-1 And the approximate error 2u l-1 ;

[0115] like Figure 3 As shown, according to the target image result X output by the l-1 layer of the LI-Net network (l-1) , two approximate values ​​of imaging error r are calculated by formula (4) l-1 and u l-1 , jointly measure the imaging effect.

[0116]

[0117] Among them, l=1,2,...,L.

[0118] Step 2: Calculate the hidden state h of the gate recurrent unit (GRU) of the lth layer of the LI-Net network l ;

[0119] like Figure 3 As shown, according to the imaging error approximation r l-1 and u l-1 And the hidden state h of the GRU in the l-1 layer of the LI-Net network l-1 , get the hidden state h of the l-th layer GRU l .

[0120] h l =GRU(h l-1 ,r l-1 ,u l-1 )(5)

[0121] Wherein, GRU(·) is a gate recurrent operation, which is well known to those skilled in the art and will not be described again.

[0122] Step 3: Calculate the threshold θ of the lth layer of the LI-Net network l,X and step size γ l,X ;

[0123] like Figure 3 As shown, using the softplus function and the relu function, according to formula (6), the hidden state h of the GRU of the lth layer of the LI-Net network is l Processing is performed to obtain the threshold θ of the lth layer of the LI-Net network l,X and step size γ l,X .

[0124]

[0125] Among them, for any real number z, softplus(z) = log(1+exp(z)), log(·) is the natural logarithm, exp(·) is the natural exponential, relu(z) = max(0,z), max(·) is the maximum value, and are the transformation matrices of the threshold and step size of the lth layer of the LI-Net network, and are the transformation vectors of the threshold and step size of the lth layer of the LI-Net network, and are the bias vectors of the threshold and step size of the lth layer of the LI-Net network, and They are the bias constants of the threshold and step size of the lth layer of the LI-Net network respectively.

[0126] Step 4: Calculate the target image result X output by the first layer of the LI-Net network (l) ;

[0127] like Figure 3 As shown, based on the MIMO radar signal matrix S, the distance measurement matrix Φ1, the angle measurement matrix Φ2, the distance optimal matrix B1, the angle optimal matrix B2, the threshold θ l,X and step size γ l,X , calculate the target image result X output by the lth layer of the LI-Net network (l) .

[0128]

[0129] Among them, X (l-1) Represents the output of the l-1th layer of the network, and the sparseness-induced penalty parameter η l,X =1 / (4θ l,X ),

[0130] is the proximal operator of the minimum-maximum concave penalty function, calculated by formula (8).

[0131]

[0132] Wherein, z represents any complex number, and sgn(·) represents a complex signed function.

[0133] Step 5: According to steps 1 to 4, calculate from the 1st layer to the Lth layer to obtain the initial target image X LI , that is, X LI =X (L) ;

[0134] Let the output of LI-Net layer L-1 be X (L-1) , the threshold and step size of the Lth layer are θ L,X and γ L,X , calculate the output X of the Lth layer of LI-Net (L) , get the initial target image X LI .

[0135]

[0136] Step 3: Output image X of LI-Net LI Perform channel processing, that is, the single channel complex valued image X LISplit into real and imaginary parts and convert into a two-channel real-valued image X IN .

[0137] X IN =[Re(X LI );Im(X LI )](10)

[0138] Among them, Re(·) represents the real part of the matrix, and Im(·) represents the imaginary part of the matrix.

[0139] Step 4: Construct MIMO radar learning imaging network (IF-Net);

[0140] IF-Net is constructed based on Convolutional Neural Networks (CNN), MaxPool layer, MaxUnpool layer, skip connection, etc., which is used to convert a partially defocused low-resolution target image into a fully focused high-resolution image. Specifically, the following steps are included:

[0141] Step 1: Output X of LI-Net after channel processing IN Perform encoding processing to obtain the encoding output X EN ;

[0142] X IN Input to the first Convolutional Neural Networks (CNN) module for processing to obtain the output of the first CNN module The CNN module is composed of a convolutional layer (Conv), a batch normalization layer (BatchNormalization, BN) and a ReLU layer (the specific construction method is well known to those skilled in the art and will not be repeated). In a specific embodiment of the present invention, the Conv kernel size is 5×5 and the step size is 1;

[0143] Will Input to the first MaxPool layer for processing to obtain the first MaxPool layer output and the first largest index result In a specific embodiment of the present invention, the window size of MaxPool is 2×2 and the step size is 1;

[0144] Will Input to the second CNN module for processing and obtain the output of the second CNN module In a specific embodiment of the present invention, the Conv kernel size is 3×3 and the step size is 1; Input to the second MaxPool layer for processing, and obtain the second MaxPool layer output and the second largest index result Will Input to the third CNN module for processing to obtain the encoded output X EN In a specific embodiment of the present invention, the Conv kernel size is 3×3 and the step size is 1.

[0145] Step 2: Encode the output X EN Perform decoding processing to obtain the decoded output X DE ;

[0146] X EN The input is processed by the first CNN module containing the deconvolution layer (ConvTrans), and the result is the same as Perform dot product multiplication to obtain the output of the first CNN module containing the deconvolution layer Will and Input to the first MaxUnpool layer for processing, and the result is the same as Perform dot product multiplication to obtain the first MaxUnpool layer output Will Input to the second CNN module containing ConvTrans for processing, and the results are the same as Perform dot product multiplication to obtain the output of the second CNN module containing the deconvolution layer Will and The input is processed by the second MaxUnpool layer, and the result is the same as Perform dot product multiplication to obtain the second MaxUnpool layer output Will Input to the third CNN module containing ConvTrans for processing to obtain the decoded output X DE .

[0147] Step 3: X IN and X DE Concatenate and input into the generator to obtain the final target image;

[0148] The encoder input X is connected by channel cascading IN With the decoder output X DE Stitching is performed in the third dimension (the stitching method is well known to those skilled in the art and will not be described again), and the result is input into a generator consisting of a CNN module with a convolution kernel size of 1×1 and a step size of 1 to obtain the final target image.

[0149] Step 5: Train the MIMO radar self-focusing learning imaging network;

[0150] The self-focusing learning imaging network (AFLI-Net) consisting of two sub-networks, LI-Net and IF-Net, was trained using the "pre-training" and "re-training" methods. In the "pre-training" stage, the learning imaging network (LI-Net) and the image focusing network (IF-Net) were trained independently in turn. In the "re-training" stage, the two network parameters learned in the previous stage were used as the initial parameters of AFLI-Net, and the entire network was trained end-to-end.

[0151] The specific steps include:

[0152] Step 1: Pre-train LI-Net;

[0153] Given a training dataset and the number of network layers L, using the well-known back-propagation method in the field (Ian Goodfellow et al., ed. Deep Learning [M]. Beijing: Posts and Telecommunications Press, 2021), minimize the loss function L of LI-Net LI-Net , get the optimal network parameters of LI-Net That is, the transformation matrix of the threshold of each layer and the transformation matrix of the step size Transformation vector for threshold and step size and Bias vector for threshold and step size and Bias constants for threshold and step size and And the GRU hidden state h l (l=1,2,...,L).

[0154]

[0155] in, Indicates As input, Θ LI-Net is the parameterized LI-Net output, Θ LI-Net Indicates LI-Net network parameters.

[0156] Step 2: Pre-train IF-Net;

[0157] Given a training dataset in Based on the optimal network parameters The resulting LI-Net network output The abbreviation of , using the back propagation method well known in the art to minimize the loss function L of IF-NetIF-Net , get the optimal network parameters of IF-Net That is, the weights and biases of each Conv layer and the parameters of each BN layer

[0158]

[0159] in, is the oth LI-Net output image after channel processing, For is the IF-Net output of the input, Θ IF-Net Represents the IF-Net network parameters. Solving equation 12 can obtain the optimal parameters.

[0160] Step 3: End-to-end training of AFLI-Net;

[0161] Load the parameters obtained by LI-Net and IF-Net pre-training as the initialization network parameters of AFLI-Net, and use the back propagation method well known in the art to minimize the AFLI-Net loss function L AFLI-Net , end-to-end training is performed on AFLI-Net to obtain its optimal network parameters.

[0162]

[0163] in, represents the network parameters of AFIT-Net, Indicates

[0164] As input, Θ AFLI-Net AFLI-Net output for network parameters.

[0165] Step 6: Use the trained AFLI-Net to process the actual MIMO radar signal matrix S to obtain a fully focused, high-resolution target image X.

Claims

1. A MIMO radar high-resolution self-focusing learning imaging method, characterized in that: The specific steps include: Step 1: Construct a MIMO radar learning imaging training data set; The specific steps include: Step 1: Determine the MIMO radar system parameters; Establish a plane rectangular coordinate system, with the positive direction of the X axis pointing to the east and the positive direction of the Y axis pointing to the north. m 、x n are the x-axis coordinates of the mth transmitting array element and the nth receiving array element, respectively. m=1,2,…,M,n=1,2,…,N,M and N represent the number of transmitting array elements and receiving array elements of the MIMO radar, respectively. p is the x-axis coordinate of the pth virtual array element, p = 1, 2, ..., MN, r k ,θ k are the distance and angle of the kth target scattering point relative to the radar center, k = 1, 2, …, K, K represents the number of target scattering points; According to actual needs, determine the number of transmitting array elements M, the number of receiving array elements N, and the position of the transmitting array element (x m ,0), receiving array element position (x n ,0), and the starting frequency f1 and center frequency f of the step-frequency continuous wave signal c , frequency step Δf and frequency number Q; Step 2: Determine the MIMO radar imaging space; According to the possible location of the target, the range of the MIMO radar imaging [r min ,r max ] and angular imaging range [θ min ,θ max ], where r min and r max are the minimum imaging distance and the maximum imaging distance, θ min and θ max are the minimum imaging angle and the maximum imaging angle respectively; Step 3: Divide the imaging space and construct the measurement matrix; The MIMO radar imaging space is divided into I×J grids, where I and J are the number of grids divided in distance and angle respectively, and the distance full sampling measurement matrix Ψ1 and the angle full sampling measurement matrix Ψ2 in the pseudo-polar coordinate system are constructed; Among them, P = MN is the number of virtual array elements, α i =2r i Δf / c is the i-th distance coordinate in the pseudo-polar coordinate system, β j = -2dsinθ j / λ c is the jth angle coordinate in the pseudo-polar coordinate system, c is the speed of light, λ c =c / f c is the center frequency f c The corresponding wavelength, i = 1, 2, ..., I, j = 1, 2, ..., J, d is the radar virtual array element spacing, r i =r min +(i-1)(r max -r min ) / (I-1) and θ j =θ min +(j-1)(θ max -θ min ) / (J-1) are the i-th discrete distance and j-th discrete angle respectively; Randomly select frequency points and virtual array elements, extract row vectors corresponding to the selected frequency points and virtual array elements from the range and angle full sampling measurement matrices Ψ1 and Ψ2 respectively, and construct the range measurement matrix Φ1 and the angle measurement matrix Φ2; Step 4: Calculate the optimal matrix based on the measurement matrix; Given the distance and angle measurement matrices Φ1 and Φ2, the gradient projection method is used to solve equation (3) to calculate the distance and angle optimal matrices B1 and B2; in,‖·‖ F is the Frobenius norm of the matrix, (·) 2 represents the square, diag(·) is the vector of matrix diagonal elements, 1 I and 1 J are all-1 vectors of size I and J respectively, (·) H represents conjugate transpose; Step 5: Randomly generate an ideal target image, calculate the MIMO radar echo signal matrix, and generate a training data set; The number of targets is randomly selected within a certain numerical range, and the amplitude of each target is randomly determined based on a Gaussian distribution with a certain mean and standard deviation. min ,r max ] and [θ min ,θ max ] to randomly set the position of each target and generate O ideal target images Calculate MIMO radar echo signal matrix Where O is the size of the training dataset, is the oth target ideal image, is the oth MIMO radar signal matrix, o=1,2,...,O; Step 2: Construct the MIMO radar learning imaging network LI-Net; Construct an L-layer deep neural network LI-Net with the same structure in each layer, whose input is the MIMO radar echo signal matrix S input , distance measurement matrix Φ1, angle measurement matrix Φ2, distance optimal matrix B1, angle optimal matrix B2, target image initialization result X (0) =0, the hidden state initialization result of the gate recurrent unit h0=0, S input Indicates any one The output is the target image X LI , specifically including the following steps: Step 1: Calculate the approximate value 1r of the imaging error of the l-1th layer of the LI-Net network l-1 And the approximate error 2u l-1 ; According to the target image result X output by the l-1 layer of the LI-Net network (l-1) , two approximate values ​​of imaging error r are calculated by formula (4) l-1 and u l-1 , jointly measure the imaging effect; Where, l = 1, 2, ..., L; Step 2: Calculate the hidden state h of the gated recurrent unit GRU of the lth layer of the LI-Net network l ; According to the imaging error approximation r l-1 and u l-1 And the hidden state h of the GRU in the l-1 layer of the LI-Net network l-1 , get the hidden state h of the l-th layer GRU l ; h l =GRU(h l-1 ,r l-1 ,u l-1 )(5) Among them, GRU(·) is a gate recurrent operation; Step 3: Calculate the threshold θ of the lth layer of the LI-Net network l,X and step length γ l,X ; Using the softplus function and the relu function, according to formula (6), the hidden state h of the GRU in the lth layer of the LI-Net network is l Processing is performed to obtain the threshold θ of the lth layer of the LI-Net network l,X and step length γ l,X ; Among them, for any real number z, softplus(z) = log(1+exp(z)), log(·) is the natural logarithm, exp(·) is the natural exponential, relu(z) = max(0,z), max(·) is the maximum value, and are the transformation matrices of the threshold and step size of the lth layer of the LI-Net network, and are the transformation vectors of the threshold and step size of the lth layer of the LI-Net network, and are the bias vectors of the threshold and step size of the lth layer of the LI-Net network, and They are the bias constants for the threshold and step size of the lth layer of the LI-Net network respectively; Step 4: Calculate the target image result X output by the first layer of the LI-Net network (l) ; Based on the MIMO radar signal matrix S, the distance measurement matrix Φ1, the angle measurement matrix Φ2, the distance optimal matrix B1, the angle optimal matrix B2, the threshold θ l,X and step length γ l,X , calculate the target image result X output by the lth layer of the LI-Net network (l) ; Among them, X (l-1) Represents the output of the l-1th layer of the network, and the sparseness-induced penalty parameter η l,X =1 / (4θ l,X ), is the proximal operator of the minimum-maximum concave penalty function, calculated by formula (8); Where z represents any complex number, sgn(·) represents a complex sign function; Step 5: According to steps 1 to 4, calculate from the 1st layer to the Lth layer to obtain the initial target image X LI , that is, X LI =X (L) ; Let the output of LI-Net layer L-1 be X (L-1) , the threshold and step size of the Lth layer are θ L,X and γ L,X , calculate the output X of the Lth layer of LI-Net (L) , get the initial target image X LI ; Step 3: Output image X of LI-Net LI Perform channel processing, that is, the single channel complex valued image X LI Split into real and imaginary parts and convert into a two-channel real-valued image X IN ; X IN =[Re(X LI );Im(X LI )](10) Where Re(·) represents the real part of the matrix, and Im(·) represents the imaginary part of the matrix; Step 4: Construct MIMO radar learning imaging network IF-Net; IF-Net is constructed based on convolutional neural network CNN, maximum pooling MaxPool layer, unmaximum pooling MaxUnpool layer, skip connection, etc., which is used to convert a partially defocused low-resolution target image into a fully focused high-resolution image. Specifically, the following steps are included: Step 1: Output X of LI-Net after channel processing IN Perform encoding processing to obtain the encoding output X EN ; X IN Input to the first convolutional neural network (CNN) module for processing to obtain the output of the first CNN module The CNN module consists of a convolutional layer Conv, a batch normalization layer BN, and a ReLU layer; Will Input to the first MaxPool layer for processing to obtain the first MaxPool layer output and the first largest index result Will Input to the second CNN module for processing and obtain the output of the second CNN module Will Input to the second MaxPool layer for processing, and obtain the second MaxPool layer output and the second largest index result Will Input to the third CNN module for processing to obtain the encoded output X EN ; Step 2: Encode the output X EN Perform decoding processing to obtain the decoded output X DE ; X EN Input to the first CNN module containing the deconvolution layer ConvTrans for processing, and the result is the same as Perform dot product multiplication to obtain the output of the first CNN module containing the deconvolution layer Will and Input to the first MaxUnpool layer for processing, and the result is the same as X c 2 onv Perform dot product multiplication to obtain the first MaxUnpool layer output Will Input to the second CNN module containing ConvTrans for processing, and the results are the same as Perform dot product multiplication to obtain the output of the second CNN module containing the deconvolution layer Will and The input is processed by the second MaxUnpool layer, and the result is the same as Perform dot product multiplication to obtain the second MaxUnpool layer output Will Input to the third CNN module containing ConvTrans for processing to obtain the decoded output X DE ; Step 3: X IN and X DE Perform stitching and input into the generator to obtain the final target image; The encoder input X is connected by channel cascading IN With the decoder output X DE The splicing is performed in the third dimension, and the result is input into the generator composed of CNN modules to obtain the final target image; Step 5: Train the MIMO radar self-focusing learning imaging network; The self-focusing learning imaging network AFLI-Net, which consists of two sub-networks, LI-Net and IF-Net, was trained using the "pre-training" and "re-training" methods. In the "pre-training" stage, the learning imaging network LI-Net and the image focusing network IF-Net were trained independently in turn. In the "re-training" stage, the two network parameters learned in the previous stage were used as the initial parameters of AFLI-Net, and the entire network was trained end-to-end. The specific steps include: Step 1: Pre-train LI-Net; Given a training dataset And the number of network layers L, using the back propagation method, minimize the loss function L of LI-Net LI-Net , get the optimal network parameters of LI-Net That is, the transformation matrix of the threshold of each layer and the transformation matrix of the step size Transformation vector for threshold and step size and Bias vector for threshold and step size and Bias constants for threshold and step size and And the GRU hidden state h l (l=1,2,...,L); in, Indicates As input, Θ LI-Net is the parameterized LI-Net output, Θ LI-Net Represents LI-Net network parameters; Step 2: Pre-train IF-Net; Given a training dataset in Based on the optimal network parameters The resulting LI-Net network output The abbreviation of , using the back propagation method to minimize the loss function L of IF-Net IF-Net , get the optimal network parameters of IF-Net That is, the weights and biases of each Conv layer and the parameters of each BN layer in, is the oth LI-Net output image after channel processing, For is the IF-Net output of the input, Θ IF-Net represents the IF-Net network parameters; solve equation 12 to obtain the optimal parameters; Step 3: End-to-end training of AFLI-Net; Load the pre-trained parameters of LI-Net and IF-Net as the initialization network parameters of AFLI-Net, and use the back-propagation method to minimize the AFLI-Net loss function L AFLI-Net , perform end-to-end training on AFLI-Net to obtain its optimal network parameters; in, represents the network parameters of AFIT-Net, Indicates As input, Θ AFLI-Net AFLI-Net output for network parameters; Step 6: Use the trained AFLI-Net to process the actual MIMO radar signal matrix S to obtain a fully focused, high-resolution target image X.

2. The MIMO radar high-resolution self-focusing learning imaging method according to claim 1, characterized in that: In step 5 of step 1, the number of targets is randomly selected in the value range of 1 to 10, and the amplitude of each target is randomly determined based on a Gaussian distribution with a mean of 0 and a standard deviation of 1.

3. The MIMO radar high-resolution self-focusing learning imaging method according to claim 1, characterized in that: In step 4: Step 1: Output X of LI-Net after channel processing IN Perform encoding processing to obtain the encoding output X EN ; X IN Input to the first convolutional neural network (CNN) module for processing to obtain the output of the first CNN module The CNN module consists of a convolutional layer Conv, a batch normalization layer BN, and a ReLU layer. The Conv kernel size is 5×5 and the stride is 1. Will Input to the first MaxPool layer for processing to obtain the first MaxPool layer output and the first largest index result The window size of MaxPool is 2×2 and the step size is 1; Will Input to the second CNN module for processing and obtain the output of the second CNN module The Conv kernel size is 3×3 and the step size is 1; Input to the second MaxPool layer for processing, and obtain the second MaxPool layer output and the second largest index result Will Input to the third CNN module for processing to obtain the encoded output X EN , the Conv kernel size is 3×3 and the stride is 1; Step 3: X IN and X DE Perform stitching and input into the generator to obtain the final target image; The encoder input X is connected by channel cascading IN With the decoder output X DE The concatenation is performed in the third dimension, and the result is input into a generator consisting of a CNN module with a convolution kernel size of 1×1 and a step size of 1 to obtain the final target image.

Images