Synthetic aperture assisted multi-channel radar forward-looking learning imaging method

Abstract

The application discloses a kind of synthetic aperture assisted multi-channel radar forward-looking learning imaging method, first, the echo data of the region to be imaged is acquired, the data obtained is compressed in distance direction pulse, then the echo data after pulse compression is imaged using the synthetic aperture assisted depth unfolding iterative shrinkage threshold algorithm ISTA network, i.e. in real aperture dimension, the imaging problem is converted into the estimation problem of target scattering coefficient, the target function is solved using the depth unfolding ISTA network, then the azimuth focusing is carried out using back projection BP algorithm in synthetic aperture dimension, then the synthetic aperture dimension result and the ISTA network result are fused to obtain the left-right unambiguous super-resolution imaging result.The method of the application overcomes the problem of left-right ambiguity and low resolution of single-base radar forward-looking imaging, realizes multi-channel radar forward-looking high-resolution imaging, and through depth learning, the parameter setting step is saved, the efficiency is improved, and the problem of signal-to-noise ratio sensitivity is overcome.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar imaging, and particularly relates to a multi-channel radar forward-looking learning imaging method assisted by synthetic aperture. BACKGROUND

[0002] Radar forward-looking imaging has important applications in autonomous landing, autonomous navigation and other fields. However, due to the problems of Doppler symmetry ambiguity and small angle change, the existing single-base synthetic aperture radar (SAR) or Doppler beam sharpening (DBS) technology has a blind area in forward-looking imaging.

[0003] The document "G. Krieger, J. Mittermayer, M. Wendler, F. Witte, and A. Moreira, Sirev-sector imaging radar for enhanced vision. Proceedings of the 2nd International Symposium on Image and Signal Processing and Analysis. 2001" proposes that by using one or more transmitting channels and multiple receiving channels on a single platform, an aperture can be formed in the azimuth direction, which has the potential for forward-looking imaging. However, due to the limitation of platform size, the forward-looking azimuth resolution of multi-channel radar is usually low. In order to improve the azimuth resolution, different algorithms have been developed in recent years to realize super-resolution imaging of multi-channel radar. The document "Zhang Jie, Wu Di, Zhu Daoyin. An airborne / munition array radar forward-looking super-resolution imaging algorithm. Radar Science and Technology, 2018, 16(2): 6" proposes to use the MUSIC algorithm for forward-looking super-resolution imaging, but this algorithm requires prior knowledge of the number of sources, and usually requires multiple snapshot data to obtain good performance. The document "Lu Jingyue, Zhang Lei, Wang Guanyong. Forward-looking multi-channel synthetic aperture radar deblurring imaging method. Electronics and Information Journal, 2018, 40(12): 2820-2825" proposes a multi-channel forward-looking synthetic aperture radar imaging scheme. First, the platform motion information is used to obtain a forward-looking image with left / right ambiguity, and then the data of multiple channels are combined to resolve the left / right ambiguity. However, in the forward-looking observation area, the azimuth resolution is still poor due to the small change in viewing angle. The document "Ren Lingyun, Wu Di, Zhu Daoyin, et al. Forward-looking imaging based on iterative super-resolution estimation of airborne multi-channel radar. Journal of Radar: 1-13 [2023-12-11]" proposes a multi-channel radar forward-looking imaging algorithm based on single-shot iterative super-resolution processing, which can obtain accurate position and amplitude information of the target through iterative spectrum estimation of a single spatial snapshot, and then realize high-resolution forward-looking azimuth imaging through non-coherent accumulation of multiple pulses. However, this method has the problems of large amount of calculation, difficulty in adjusting iterative parameters, and sensitivity to signal-to-noise ratio. SUMMARY

[0004] To solve the above technical problems, the application provides a synthetic aperture assisted multi-channel radar forward-looking learning imaging method, aiming to overcome the problems of parameter adjustment difficulty, large amount of calculation and noise sensitivity of existing super-resolution algorithms, and realize multi-channel radar forward-looking super-resolution imaging.

[0005] The technical scheme adopted by the application is as follows: a synthetic aperture assisted multi-channel radar forward-looking learning imaging method, the specific steps are as follows:

[0006] S1, acquiring multi-channel radar forward-looking echo data;

[0007] S2, range compression of the data acquired in step S1;

[0008] S3, selecting a high-resolution SAR image for cropping and selecting a randomly generated image with only point targets to obtain 150 64x64 images as original scenes, generating multi-channel echo data based on steps S1 and S2, constructing a data set, and forming echo-image pairs with the real aperture data and the original scenes, and dividing them into a training set and a test set;

[0009] Among them, the training set includes 150 original scenes, and the test set randomly selects images that do not exist in the training set.

[0010] S4, building a synthetic aperture assisted multi-channel radar forward-looking imaging network and training;

[0011] S5, acquiring the multi-channel radar pulse compression echo signal of the region to be imaged according to steps S1-S2, and then inputting it into the multi-channel radar forward-looking imaging network trained in step S4 to obtain the synthetic aperture assisted multi-channel radar forward-looking imaging result.

[0012] Further, the step S1 is specifically as follows:

[0013] The multi-channel radar forward-looking imaging adopts a single-transmit multi-receive channel configuration, in an xyz coordinate system, O represents the origin of the coordinate system, the platform motion direction is the X axis, each channel is uniformly arranged on the Y axis, and the height direction is the Z axis.

[0014] The multi-channel radar platform flies at a speed v r , a height h along the X axis direction at a constant speed, forming a synthetic aperture with a length L s , on the platform, each channel is uniformly arranged along the Y axis direction, the transmitting channel T x transmits signals at a pre-set pulse repetition frequency PRF, and each channel receives signals simultaneously.

[0015] Let the imaging region exist a point target P1(x0, y0, 0), and each channel coordinate is expressed as (v r t a ,y m ,h), then the transmission distance history of each channel to the target P1 and the receiving distance history are expressed as follows:

[0016]

[0017] wherein t a represents the slow time, y m represents the azimuth coordinate of the mth receiving channel.

[0018] Then for the point target, the distance history is expressed as follows:

[0019]

[0020] Let the transmission signal be a linear frequency modulation pulse, then the echo signal is expressed as follows:

[0021]

[0022] wherein β0represents the scattering coefficient of the point target, t r represents the fast time, ω r represents the distance window, K r represents the frequency modulation slope of the transmission signal, λ represents the wavelength, and c represents the light speed.

[0023] Further, the step S2 is specifically as follows:

[0024] After the echo signal passes through the range matching filter, the expression is as follows:

[0025]

[0026] The results of all channel echo pulse compression are obtained using formula (4), and noise is added, then the synthetic aperture echo expression of a given channel is as follows:

[0027]

[0028] At a given slow time, the instantaneous data of multiple channels form a snapshot or real aperture data S2(y m ,t r ), and the expression is as follows:

[0029]

[0030] The synthetic aperture and real aperture two-dimensional echo signals are obtained, and noise is added.

[0031] ​​Further, the step S4 is specifically as follows:

[0032] S41, build a deep unfolding network ISTA-Net;

[0033] According to the knowledge of array signal processing, if the antenna array is composed of M elements, and N narrowband signals are incident on the spatial array, the received signal s of the mth channel in a range cell is m The expression is as follows:

[0034]

[0035] Wherein, σ i represents the scattering coefficient of the ith target, f0 represents the center frequency of the transmitted signal, n m represents the noise of the mth channel, τ mi represents the delay of the ith signal to the mth channel relative to the reference element, i.e. the wave path difference, the same grid is used to represent the position P(ρcosγ i ,ρsinγ i ) of the ith target, and the position of each channel of the first snapshot is represented as (0,y m , h), and the expression is as follows:

[0036]

[0037] Wherein, ρ represents the slant range, γ i represents the azimuth angle of the ith target, and y0 represents the middle channel.

[0038] The echo expression of one range cell of different elements is as follows:

[0039]

[0040] Equation (9) is written in matrix form, and the expression is as follows:

[0041] s=A·σ+n (10)

[0042] Wherein, s represents the echo of one range cell, n represents the noise vector, and A represents the steering matrix.

[0043] In the case of given input echo s, the estimation result is first initialized by ISTA-Net Then the estimation result is obtained through the L-layer network.

[0044] Wherein, [·] T represents the transpose of the matrix.

[0045] Then the difference between the estimation result and the label is used to design a loss function, and the value of the loss function is used as a feedback signal to continuously adjust the weight of each layer in the network through back propagation.

[0046] According to the existing ISTA iterative formula, the ISTA-Net is divided into two modules: a linear module R and a nonlinear module N at the l(th) (1, 2, 3…L) layer, and the expression is as follows:

[0047]

[0048] wherein r (l) represents the linear reconstruction result of the l(th) layer, β (l) represents the step length that can be learned in the l(th) layer, and the regularization coefficient κ and β are combined into a parameter V, i.e. V (l) = κ (l) β (l) , represents the learnable threshold.

[0049] In the case of a non-sparse scene, a nonlinear transformation function F(·) is used, which represents a matrix that reflects the inherent characteristics of the observed scene, including edges and contours. Then, the unknown The calculation expression is as follows:

[0050]

[0051] wherein, represents the left inverse of F(·), such that represents the Hadamard product, and I represents the unit operator.

[0052] Both F and F are learned and updated in the training process of the network, and F includes two linear convolution operators separated by ReLU.

[0053] wherein the number of filters is preset as 2N f , and the size of each filter is ω f × ω f , the first convolution operator corresponds to N f filters, and the second convolution operator corresponds to another group of N f filters.

[0054] Considering that the radar echo data is complex-valued, the matrix vector multiplication is decomposed, and the expression is as follows:

[0055]

[0056] wherein Re(·) and Im(·) represent the real part and the imaginary part of a complex number, respectively.

[0057] Using average euclidean distance loss function The expression is as follows:

[0058]

[0059] F(·) and Symmetric constraint loss function is applied The expression is as follows:

[0060]

[0061] The deep unfolding network ISTA-Net end-to-end loss function expression is as follows:

[0062]

[0063] Wherein, The difference loss function is represented by, The symmetric constraint loss function is represented by G, the total number of training samples, L represents the total number of network layers, and η represents the balance And The weight parameters of and.

[0064] Using the designed loss function, the trained ISTA-Net is obtained, which is used for real aperture super-resolution imaging to obtain the imaging result f ISTA_Net (ρ,γ).

[0065] S42, synthetic aperture dimension processing;

[0066] According to the echo model formula (5), the back projection BP algorithm is used to process the synthetic aperture data, the radar echo data is projected onto each unit of the imaging area, and the values of different directions are coherently superimposed to realize the azimuth focusing.

[0067] According to the BP imaging principle, the imaging result f bp-i (ρ,γ,y i ) of the i-th channel is expressed as follows:

[0068]

[0069] Wherein, ρ represents the slant range, and γ represents the azimuth angle.

[0070] S43, fusion processing;

[0071] On the basis of steps S41 and S42, the synthetic aperture processing result and the real aperture ISTA-Net super-resolution result are fused to obtain the final imaging result. And the imaging result is displayed on the same grid, then the fusion is realized by multiplication, and the non-blurred forward-looking imaging result f SAP_ISTA_Net (ρ,γ) is obtained, and the expression is as follows:

[0072] f SAP_ISTA_Net (p, g) = f bp-i (p, g) f ISTA_Net (p, g) (19)

[0073] The method of the present application first acquires echo data of the region to be imaged, performs range direction pulse compression on the acquired data, and then uses a synthetic aperture assisted deep unfolding iterative shrinkage threshold value algorithm (ISTA) network to image the echo data after pulse compression, that is, converts the imaging problem into an estimation problem of the target scattering coefficient in the real aperture dimension, solves the target function by using a deep unfolding ISTA network, uses a back projection (BP) algorithm to perform azimuth focusing in the synthetic aperture dimension, then fuses the results in the synthetic aperture dimension and the results of the ISTA network to obtain a super-resolution imaging result without left-right blur. The method of the present application overcomes the problems of left-right blur and low resolution in single-base radar forward-looking imaging, realizes multi-channel radar forward-looking high-resolution imaging, and through deep learning self-adaptive determination of parameters, the step of manually setting parameters is saved, the efficiency of the method is effectively improved, and the problem of sensitivity to signal-to-noise ratio is overcome to some extent. BRIEF DESCRIPTION OF DRAWINGS

[0074] Figure 1 A flowchart of a synthetic aperture assisted multi-channel radar forward-looking learning imaging method of the present application.

[0075] Figure 2 A geometric model diagram of multi-channel radar forward-looking imaging in an embodiment of the present application.

[0076] Figure 3 A multi-channel real aperture ISTA-Net imaging result diagram of an observed scene in an embodiment of the present application.

[0077] Figure 4 A whole framework schematic diagram of a synthetic aperture assisted multi-channel radar forward-looking imaging network in an embodiment of the present application.

[0078] Figure 5 An ISTA-Net structure diagram in an embodiment of the present application.

[0079] Figure 6 A single-channel synthetic aperture forward-looking imaging result diagram of an observed scene in an embodiment of the present application.

[0080] Figure 7 A schematic diagram of an observed scene in an embodiment of the present application.

[0081] Figure 8 An imaging result diagram of an observed scene in an embodiment of the present application. DETAILED DESCRIPTION

[0082] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0083] like Figure 1 The flowchart of a synthetic aperture-assisted multi-channel radar forward-looking learning imaging method of the present invention is shown below, and the specific steps are as follows:

[0084] S1. Acquire multi-channel radar forward-looking echo data;

[0085] S2. Compress the distance pulse of the data obtained in step S1;

[0086] S3. Select high-resolution SAR images for cropping and randomly generated images containing only point targets to obtain 150 64x64 images as the original scene. Based on steps S1 and S2, generate multi-channel echo data, construct a dataset, and form echo-image pairs with the real aperture data and the original scene, and divide them into training set and test set.

[0087] The training set consists of 150 original scene images, while the test set randomly selects other images that do not exist in the training set.

[0088] S4. Build and train a synthetic aperture-assisted multi-channel radar forward-looking imaging network.

[0089] S5. Obtain the echo signal after multi-channel radar pulse compression from the area to be imaged according to steps S1-S2, and then input it into the multi-channel radar forward-looking imaging network trained in step S4 to obtain the synthetic aperture-assisted multi-channel radar forward-looking imaging result.

[0090] In this embodiment, step S1 is specifically as follows:

[0091] In this embodiment, the geometric model of multi-channel radar forward-looking imaging is as follows: Figure 2 As shown in Table 1, the parameters of the multi-channel radar are as follows. The forward-looking imaging of the multi-channel radar adopts a single-transmitter, multi-receiver channel configuration. In an xyz spatial coordinate system, O represents the origin of the coordinate system, the platform movement direction is the X-axis, each channel is evenly arranged on the Y-axis, and the height direction is the Z-axis.

[0092] Table 1

[0093] Parameter Symbol Value Unit Platform flying height h 5000 m Platform flying speed v r ]]> 400 m / s Pulse width [CAT r ]]> 1 μs Signal bandwidth B 60 MHz Array antenna length [[ L e ]]> 3.00 m Distance direction sampling points

[0014] N r ]]> 256 Channel number

[0010] N a ]]> 16 Synthetic aperture dimension sampling points

[0010] N sa ]]> 128 Pulse repetition frequency PRF 400 Hz Transmit signal wavelength λ 0.0315 m

[0094] Multi-channel radar platform at speed v r =400m / s, altitude h=5000m, flying at a constant speed in a straight line along the X-axis, forming a formation of length L s =80m synthetic aperture, on the platform, the channels are evenly arranged along the Y-axis, the emission channel T xThe signal is transmitted at a preset pulse repetition frequency PRF = 400Hz, and each channel receives the signal simultaneously.

[0095] Among them, R xi This represents the i-th receiving channel.

[0096] Suppose there exists a point target P1(x0,y0,0) in the imaging region, and the coordinates of each channel are represented as (v r t a ,y m If h), then the historical transmission distance from each channel to target P1 is... and reception distance history The expression is as follows:

[0097]

[0098] Among them, t a Indicates slow time, y m This represents the azimuth coordinates of the m-th receiving channel.

[0099] For a point target, the distance history expression is as follows:

[0100]

[0101] If the transmitted signal is set to a linear frequency modulated pulse, the echo signal expression is as follows:

[0102]

[0103] Where β0 represents the point target scattering coefficient, t r ω represents fast time. r K represents the distance window. r λ represents the frequency modulation slope of the transmitted signal, λ represents the wavelength, and c represents the speed of light.

[0104] like Figure 3 As shown, the multi-channel real aperture ISTA-Net imaging results of the observed scene in this embodiment are as follows: the azimuth angle of the scene is -8.5° to 8.5°, and the range length is set to 250m. For each channel, the historical transmission distance can be calculated using Equation (1) based on the number of sampling points in the synthetic aperture dimension and the number of sampling points in the range dimension. and reception distance history Then, the echo signal is calculated using equation (3). The middle channel is selected to synthesize the aperture echo, and the results of the first snapshot of all channels are used to form the real aperture echo.

[0105] In this embodiment, step S2 is specifically as follows:

[0106] The embodiment sets the target scene center position as (5000m, 0m, 0m), calculates the reference transmitting distance and receiving distance by formula (1), and obtains the fast time t r and pulse compression matching function

[0107] Considering the echo data generated in step S1 in the embodiment, after the echo signal passes through the range matching filter, the expression is as follows:

[0108]

[0109] The results of all channel echo pulse compression are obtained by using formula (4), and noise with SNR=20dB is added. Then, the synthetic aperture echo expression of a given channel is as follows:

[0110]

[0111] At a given slow time, the instantaneous data of multiple channels form a snapshot or real aperture data S2(y m ,t r ), and the expression is as follows:

[0112]

[0113] The synthetic aperture and real aperture two-dimensional echo signals are obtained, and noise with SNR=20dB is added.

[0114] In the embodiment, the step S4 is specifically as follows:

[0115] The overall architecture of the synthetic aperture assisted multi-channel radar forward-looking imaging network is as shown in Figure 4 .

[0116] S41, build a depth unfolding network ISTA-Net;

[0117] The basic idea of ISTA-Net is to map the iterative steps of the iterative shrinkage threshold algorithm (ISTA) to a deep network composed of a fixed number of layers, and each layer corresponds to an iteration in the existing ISTA. The structure of ISTA-Net is as shown in Figure 5 .

[0118] According to the knowledge of array signal processing, if an antenna array is composed of M elements, and N narrowband signals are incident on the spatial array, the received signal s m of the mth channel in a range cell is as follows:

[0119]

[0120] Where, σ idenotes the scattering coefficient of the i-th target, f0denotes the center frequency of the transmitted signal, n m denotes the noise of the m-th channel, τ mi denotes the delay of the i-th signal relative to the reference element when it reaches the m-th channel, i.e., the wave path difference. Similar to the BP algorithm, the same grid can be used to represent the position P(ρcosγ i ,ρsinγ i ) of the i-th target, and the position of each channel of the first snapshot is represented as (0,y m ,h), then the expression is as follows:

[0121]

[0122] wherein ρ denotes the slant range, γ i denotes the azimuth angle of the i-th target, and y0denotes the middle channel.

[0123] Then the echo expression of one distance unit of different elements is as follows:

[0124]

[0125] Equation (9) is written in matrix form, and the expression is as follows:

[0126] s=A·σ+n (10)

[0127] wherein s denotes the echo of one distance unit, n denotes the noise vector, and A denotes the steering matrix.

[0128] In the case of a given input echo s, the estimation result is first initialized by ISTA-Net Then the estimation result is obtained through an L-layer network.

[0129] wherein [·] T denotes the transpose of a matrix.

[0130] Then a loss function is designed according to the difference between the estimation result and the label (true data), and the value of the loss function is taken as a feedback signal to continuously adjust the weight of each layer in the network through back propagation.

[0131] According to the existing ISTA iterative formula, ISTA-Net is divided into two modules: a linear module R and a nonlinear module N at the l-th (1, 2, 3... L) layer, and the expression is as follows:

[0132]

[0133] wherein r (l) denotes the linear reconstruction result of the l-th layer, β (l)denotes the step size that can be learned in the l-th layer, the regularization coefficients κ and β are combined into one parameter V, i.e. V (l) = κ (l) β (l) denotes the learnable threshold. Compared with the predefined parameters such as κ, β and V in the existing ISTA, the parameters in ISTA-Net are all learnable.

[0134] In the case of non-sparse scenarios, a nonlinear transformation function F(·) is used, which represents a matrix that observes the inherent characteristics of the scene, including edges and contours. Then, the unknown The calculation expression is as follows:

[0135]

[0136] wherein, denotes the left inverse of F(·), such that denotes the Hadamard product, and I denotes the unit operator.

[0137] and F are both learned and updated in the training process of the network, and F includes two linear convolution operators separated by ReLU.

[0138] wherein, the number of filters is preset to 2N f , and the size of each filter is ω f × ω f The first convolution operator corresponds to N f filters, and the second convolution operator corresponds to another group of N f filters.

[0139] Considering that the radar echo data is complex-valued, the matrix vector multiplication is decomposed, and the expression is as follows:

[0140]

[0141] wherein, Re(·) and Im(·) respectively represent the real part and the imaginary part of a complex number.

[0142] Generally, the network needs to be trained to obtain learnable parameters, and the loss function is an indispensable part. In order to reduce the difference between the label and the reconstruction result, the average Euclidean distance loss function The expression is as follows:

[0143]

[0144] At the same time, in order to ensure symmetric constraint loss function is imposed on F(·) and to satisfy the reversibility of the nonlinear transformation, The expression is as follows:

[0145]

[0146] The end-to-end loss function expression of the deep unfolding network ISTA-Net is as follows:

[0147]

[0148] wherein, represents the difference loss function, represents the symmetry constraint loss function, G represents the total number of training samples, L represents the total number of network layers, and η represents the balance and weight parameters of and. By using the designed loss function, a trained ISTA-Net is obtained, which is used for real-aperture super-resolution imaging to obtain an imaging result f ISTA_Net (ρ,γ).

[0149] In this embodiment, the network loss function is set according to formulae (15), (16) and (17), and the specific training parameters are shown in Table 2, and the training code is based on the Pytorch deep learning framework.

[0150] Table 2

[0151]

[0152] According to the parameters in Table 2, the ISTA-Net is trained, and the multi-channel real-aperture echo of formula (6) and the steering matrix A are input into the trained ISTA-Net, and the ISTA-Net super-resolution result is as shown in Figure 3 .

[0153] S42, synthetic aperture dimension processing;

[0154] According to the echo model formula (5), the back projection (BP) algorithm is used to process the synthetic aperture data, the radar echo data is projected onto each unit of the imaging area, and the values of different azimuths are coherently superimposed to realize azimuth focusing.

[0155] According to the BP imaging principle, the expression of the imaging result f bp-i (ρ,γ,y i ) of the i-th channel is as follows:

[0156]

[0157] wherein, ρ represents the slant range, and γ represents the azimuth angle. Similar to the existing single-station radar forward-looking imaging, the imaging result of one channel exists left-right ambiguity.

[0158] For the synthetic aperture echo of a certain channel of formula (5), the BP algorithm is used to perform coherent accumulation according to formula (18) to obtain the azimuth focusing result, as shown in Figure 6 At this time, since only single-channel data is processed, there is left-right blur in the imaging result, but this will greatly reduce the calculation amount.

[0159] S43, fusion processing;

[0160] On the basis of steps S41 and S42, the synthetic aperture processing result and the real aperture ISTA-Net super-resolution result are fused to obtain the final imaging result. Since the imaging results are displayed on the same grid, the fusion can be realized by multiplication to obtain the non-blurred forward-looking imaging result f SAP_ISTA_Net (ρ,γ), and the expression is as follows:

[0161] f SAP_ISTA_Net (ρ,γ)=f bp-i (ρ,γ)·f ISTA_Net (ρ,γ) (19)

[0162] In addition, in practice, in order to avoid the loss of weak target information, the synthetic aperture result can be displayed in logarithm.

[0163] In this embodiment, in step S5, the observed scene is as shown in Figure 7 The multi-channel radar pulse pressure echo signal of the to-be-imaged area (to-be-tested target scene) is obtained according to steps S1-S2, and then input into the multi-channel radar forward-looking imaging network trained in step S4 to obtain the multi-channel radar forward-looking imaging result assisted by synthetic aperture, as shown in Figure 8 .

[0164] In summary, the method of the present application overcomes the problems of left-right blur and low resolution of single-base radar forward-looking imaging, and the problems of difficulty in adjusting parameters and large calculation amount of existing super-resolution algorithms, realizes multi-channel radar forward-looking high-resolution imaging, and through deep learning, the parameters are adaptively determined, the step of manually setting the parameters is saved, the efficiency of the method is effectively improved, and the problem of sensitivity to signal-to-noise ratio is overcome to some extent.

[0165] Those skilled in the art will appreciate that the embodiments described herein are intended to aid the reader in understanding the principles of the present application and should not be understood to limit the scope of the present application to such specific embodiments and examples. The present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the scope of the claims of the present application.

Claims

1. A synthetic aperture assisted multi-channel radar forward-looking learning imaging method, the specific steps being as follows: S1, acquiring multi-channel radar forward-looking echo data; S2, range direction pulse compression on the data acquired in step S1; S3, selecting a high-resolution SAR image for cropping and selecting a randomly generated image with only point targets to obtain 150 64x64 images as original scenes, generating multi-channel echo data based on steps S1 and S2, constructing a data set, and forming echo-image pairs with the real aperture data and the original scenes, and dividing them into a training set and a test set; wherein The training set includes 150 original scenes, and the test set randomly selects images that do not exist in the training set; S4, building a synthetic aperture assisted multi-channel radar forward-looking imaging network and training it; The step S4 is specifically as follows: S41, building a deep unfolding network ISTA-Net; According to the knowledge of array signal processing, if the antenna array is composed of array elements, and a narrowband signal is incident on the spatial array, then the received signal of the channel in a range cell is expressed as follows: ​ （1）； in, Indicates the first The scattering coefficient of a target, Indicates the center frequency of the transmitted signal. Indicates the first Noise in each channel, Indicates the first The signal arrived at the first The delay relative to the reference array element at the th channel, i.e., the path difference, is represented using the same grid. The location of the target And the position of each channel in the first snapshot is represented as The expression is as follows: （2）； wherein, denotes the slant range, denotes the azimuth angle of the target, denotes the intermediate channel; The echo expression of a distance unit of different array elements is as follows: （3）； Equation (3) is written in matrix form, and the expression is as follows: （4）； wherein, represents an echo of a distance unit, represents a noise vector, represents a steering matrix; In the case of a given input echo , the ISTA-Net first initializes the estimation result , and then obtains the estimation result through layer network; wherein denotes the transpose of a matrix; Then, a loss function is designed using the difference between the estimation result and the label, and the value of the loss function is used as a feedback signal to continuously adjust the weight of each layer in the network through back propagation; According to the existing ISTA iteration formula, ISTA-Net is in the first layer divided into two modules: linear module and nonlinear module , expressed as follows: （5）； （6）； wherein, represents the linear reconstruction result of the layer, represents the linear reconstruction result of the layer, and are merged into one parameter , i.e. , represents the learnable threshold value; If the scene is not sparse, a nonlinear transformation function is used , which represents the intrinsic characteristics of the observed scene, including edges and contours. Then, the unknown The calculation expression is as follows: （7）； wherein denotes the left inverse of , denotes the Hadamard product, denotes the identity operator; and are both learned and updated during the training process of the network, and include two linear convolution operators with ReLU separation; wherein the number of filters is predetermined as and the size of each filter is the first convolution operator corresponds to filters and the second convolution operator corresponds to another group of filters; Considering that the radar echo data is complex, the matrix vector multiplication is decomposed, and the expression is as follows: （8）； wherein and denote the real and imaginary parts of a complex number, respectively; Using an average euclidean distance loss function with the following expression: （9）； To and Apply symmetric constraint loss function , the expression is as follows: （10）； The end-to-end loss function expression of the deep unfolding network ISTA-Net is as follows: （11）； wherein, represents a difference loss function, represents a symmetric constraint loss function, represents the total number of training samples, represents the total number of network layers, represents a balance and a weight parameter of With the designed loss function, the trained ISTA-Net is obtained, which is used for real-aperture super-resolution imaging to obtain an imaging result ; S42, synthetic aperture dimension processing; According to a synthetic aperture echo model of a channel , a back projection BP algorithm is used to process the synthetic aperture data, radar echo data is projected onto each unit of the imaging area, and values of different azimuths are coherently superimposed to realize azimuth focusing. According to the principle of BP imaging, the imaging result of the i-th channel The expression is as follows: （12）； wherein denotes the slant range, denotes the azimuth angle; S43, fusion processing; On the basis of steps S41 and S42, the synthetic aperture processing result and the real aperture ISTA-Net super-resolution result are fused to obtain a final imaging result; and the imaging result is displayed on the same grid, and the fusion is realized by multiplication to obtain a non-blurred forward-looking imaging result The expression is as follows: （13）； S5, the multi-channel radar pulse compression echo signal of the region to be imaged is obtained according to steps S1-S2, and then input into the multi-channel radar forward-looking imaging network trained in step S4 to obtain the synthetic aperture assisted multi-channel radar forward-looking imaging result.

2. The synthetic aperture assisted multichannel radar forward looking learning imaging method according to claim 1, characterized in that, The step S1 is specifically as follows: The multi-channel radar forward-looking imaging adopts a single-transmit multi-receive channel configuration, and in an xyz space coordinate system, O represents the origin of the coordinate system, the platform motion direction is the X axis, the channels are uniformly arranged on the Y axis, and the height direction is the Z axis; A multi-channel radar platform flies at a constant speed along the X-axis direction , and forms a synthetic aperture with a length of . On the platform, the channels are uniformly arranged along the Y-axis direction, and the transmitting channels transmit signals at a pre-set pulse repetition frequency , and each channel receives signals simultaneously. Setting imaging region presence point target , each channel coordinate is expressed as , each channel to the target emission distance history and reception distance history expressions are as follows: （14）； wherein, denotes the slow time, denotes the azimuth coordinate of the mth receiving channel; For a point target, the distance history expression is as follows: （15）； The transmission signal is set to be a linear frequency modulation pulse, and the echo signal expression is as follows: （16） wherein, denotes the point target scattering coefficient, denotes the fast time, denotes the range window, denotes the frequency modulation slope of the transmitted signal, denotes the wavelength, denotes the speed of light.

3. The method of claim 1, wherein, The step S2 is specifically as follows: After the echo signal passes through the range matching filter, the expression is as follows: （17）； The results of all channel echo pulse compression are obtained using equation (17), and noise is added, and the synthetic aperture echo expression of a given channel is as follows: （18）； At a given slow time, the instantaneous data of the multiple channels form a snapshot or real-aperture data The expression is as follows: （19）； The synthetic aperture and real aperture two-dimensional echo signals are obtained, and noise is added.

Images