A forward-looking imaging method for airborne radar based on alternating direction multiplier method network

Abstract

The present invention discloses an airborne radar forward-looking imaging method based on an alternating direction multiplier method network. First, a single channel is used to receive airborne radar echo signals. After pulse compression and range migration, the obtained real-beam scanning results, namely complex image data, are input into an ADMMN training network. Second, an airborne radar forward-looking imaging network based on the ADMMN is constructed, comprising n stages, each of which contains three sublayers: a reconstruction layer, a nonlinear transformation layer, and a multiplier update layer. The reconstruction layer ultimately outputs the results. A network training method based on a high-fidelity synthetic aperture radar image simulation sample-label set is then proposed to achieve high-precision training of the network. Finally, a low-resolution simulated image in azimuth and a measured radar image are input to obtain a high-resolution reconstructed image. While maintaining the convergence of the alternating direction multiplier method, the present invention achieves adaptive adjustment of hyperparameters, effectively improving the azimuth resolution of radar real-beam scanning data.

Description

Technical Field

[0001] The present invention belongs to the technical field of airborne radar deconvolution signal processing, and in particular relates to an airborne radar forward-looking imaging method based on an alternating direction multiplier method network. Background Art

[0002] Radar imaging technology, as a key means of information acquisition, is widely used in numerous military and civilian applications, including precision terminal guidance, low-altitude obstacle avoidance, and airdrops. Forward-looking radar imaging, a key branch of radar imaging technology, has been a research hotspot in recent years. With the increasing demand for radar imaging accuracy, achieving super-resolution imaging of the forward-looking area of ​​a moving platform is becoming a challenging technical challenge in these applications.

[0003] Traditional Doppler beam sharpening (DBS) technology and synthetic aperture radar (SAR) technology both have forward-looking imaging "blind spots" due to problems such as Doppler ambiguity and small Doppler gradient. In view of this situation, researchers have successively carried out research on forward-looking imaging and achieved some valuable results. The main methods used include: (1) real beam imaging technology, which obtains the real beam echo of the forward-looking imaging area through the azimuth scanning method of the real aperture antenna, and arranges the echo amplitude according to the corresponding distance and azimuth to form the forward-looking radar image; (2) bistatic SAR forward-looking imaging technology, bistatic SAR places the transmitter and receiver on different moving platforms (such as aircraft-aircraft, aircraft-missile, satellite-aircraft, etc.), and can achieve forward-looking imaging through the different flight directions of the transmitting and receiving platforms; (3) single pulse imaging technology, under the condition of high pulse repetition frequency, the angle of each pulse is measured to obtain The amplitude value of the mouth is superimposed by angle at different distances to obtain the forward-looking area image; (4) Array radar airspace super-resolution forward-looking imaging technology, a real array in the direction of the track is installed on the flight platform, the Doppler bandwidth in the azimuth direction is obtained by arranging the antenna array, and the forward-looking imaging can be realized by using the processing method of synthetic aperture imaging; (5) Imaging technology based on deconvolution theory, the echo signal output by the radar sensor can be regarded as the convolution of the antenna beam and the target angle information in the azimuth domain, so the accurate information about the target can be obtained through the idea of ​​deconvolution, thereby realizing super-resolution forward-looking imaging.

[0004] All of the above methods are feasible in practice, and the performance of some algorithms has been verified in actual systems. However, these methods also have their own problems in actual processing, which mainly include: (1) The azimuth resolution of real beam imaging technology is very limited due to the antenna aperture constraint; (2) The overall structural model of bistatic SAR forward imaging technology is complicated due to the different locations of transmission and reception. The algorithm used in this method often requires complex processing of a large amount of data. In addition, the imaging algorithm variability caused by the synchronization problem of the two stations and the spatial geometric relationship of the two stations makes it difficult to implement; (3) Single pulse imaging technology is limited by the spatial freedom degree, and it is difficult to distinguish targets that are close to each other in the beam at the same time, and angular glint will occur; (4) Array radar spatial super-resolution forward imaging technology is also limited by the antenna size. In addition, the imaging effect is also related to the number of receiving channels and the complexity of the beam scene, and the resolution improvement is limited; (5) Although the imaging technology based on deconvolution theory can transform the ill-conditioned problem faced by single-channel deconvolution into a well-conditioned one under certain conditions, it also faces the problem of signal-to-noise ratio loss. Furthermore, the deconvolution process requires a large amount of computation, resulting in a time-consuming calculation process and difficulty in parameter selection. These issues indicate that existing imaging modes still need improvement and room for performance enhancement.

[0005] In recent years, deep neural networks (DNNs) have attracted increasing attention from scholars and are being used to process complex, high-dimensional data and learn effective features. After the AlexNet model, designed using deep learning algorithms, won the 2012 ImageNet image classification competition, Hinton's team won first place, becoming a hot topic in the scientific community. Since then, convolutional neural networks have been used in various fields, including image detection, classification, and recognition. Several outstanding models have been proposed, such as VGGNet, GoogleNet, and ResNet. The tremendous success of deep learning in other fields has provided new research ideas for radar. In 2015, Morgan demonstrated the effectiveness and applicability of CNNs in radar image processing, including their superior feature extraction capabilities, by effectively applying them to SAR beacon slice recognition. In 2017, Yazici's team achieved passive radar imaging by constructing recurrent neural networks. In 2020, Hua Qinglong applied complex-domain convolutional neural networks (CV-CNNs) to SAR ship target recognition, demonstrating through simulation and field data that complex-domain networks outperform real-domain networks. In 2021, Ding Zegang's team proposed a parameterized high-resolution imaging network based on the SAR mixed echo model. They mapped the multi-component ADMM reconstruction optimization algorithm into an unfolded complex-domain deep neural network. By training the imaging network, the network's complex-domain parameters are learned independently, achieving high imaging quality and efficiency. After DNN training, its internal parameters are optimized, and its rapid processing speed is also a key reason for its popularity. These algorithms have, to a certain extent, solved radar imaging problems in certain situations, and their performance often surpasses that of traditional algorithms. However, the structural design of DNNs is a challenge, and therefore their current applications in forward-looking imaging are rare, making it necessary to explore applications in this area. Summary of the Invention

[0006] Purpose of the invention: In order to overcome the problem of improving the azimuth resolution in radar forward imaging with Doppler ambiguity and small Doppler frequency gradient, the present invention proposes an airborne radar forward imaging method based on an alternating direction multiplier method network.

[0007] Technical solution: The present invention proposes an airborne radar forward imaging method based on an alternating direction multiplier method network, which specifically includes the following steps:

[0008] (1) A single channel is used to receive the airborne radar echo signal. After pulse compression and range migration, the obtained real beam scanning result, i.e., the complex image data, is input into the alternating direction multiplier method network for training.

[0009] (2) Utilizing the generalization learning capability of high-dimensional features in deep learning, an airborne radar forward-looking imaging network based on ADMMN is constructed; the network comprises n levels, each level comprising three sublayers, namely, a reconstruction layer X, a nonlinear transformation layer Z, and a multiplier update layer β, with the reconstruction layer outputting the result; in any level, the input data of the previous level is passed through the reconstruction layer to obtain a reconstructed forward-looking image, and then through the nonlinear transformation layer to obtain a sparse representation of the forward-looking image, and finally the multiplier update layer is used to implement the update of the Lagrange multiplier;

[0010] (3) A training method based on a high-fidelity SAR image simulation sample-label set completes high-precision training of the airborne radar forward-looking imaging network;

[0011] (4) Input the low-resolution simulated image and the radar measured image in azimuth to obtain a high-resolution reconstructed image.

[0012] Furthermore, the implementation process of step (1) is as follows:

[0013] For the airborne radar echo signal received by a single channel, after pulse compression and range migration, the real beam scanning result, i.e., the complex image data, is input into the ADMMN for step-by-step training. The gradients of all levels except the xth level are fixed, and the parameters of the xth level are pre-optimized separately. The gradients of levels 1 to x-1 are released, and the parameters of the first x levels are optimized end-to-end.

[0014] Furthermore, the explicit formula for reconstructing layer X in step (2) is:

[0015] x (n+1) =(H T H+ρ (n+1) I) -1 (H T y+ρ (n+1) (z n -β n ))

[0016] Where H is the antenna pattern matrix, x (n+1) is the reconstruction result of the n+1th level, y is the input sample data, ρ n+1 is the trainable penalty parameter of the n+1th level, and Z (0) and β (0) Initialized to 0.

[0017] Furthermore, the explicit formula of the nonlinear transformation layer in step (2) is:

[0018] z (n+1) =S(x n+1 +β n ;λ (n+1) / ρ (n+1) )

[0019] In the Z layer, z (n+1) is the sparse representation result of the n+1th level, and the trainable parameter is the regularization parameter λ of the n+1th level (n+1) With penalty parameter ρ n+1 , β (0) It needs to be initialized to 0, S(·) is the soft threshold operation function, and its specific form is as follows:

[0020]

[0021] Among them, x, λ, and ρ are the output result of the reconstruction layer X, the regularization parameter, and the penalty parameter, respectively.

[0022] Furthermore, the explicit formula for the multiplier update layer β in step (2) is:

[0023] β (n+1) =β n +η (n+1) (x (n+1) -z (n+1) )

[0024] In the β layer, β (n+1) is the Lagrange multiplier of the n+1th level, and the trainable parameter is the step size parameter η of the n+1th level (n+1) .

[0025] Furthermore, the implementation process of step (3) is as follows:

[0026] High-resolution SAR images are used as ground scattering scenes and real-beam echo simulation technology is used to generate two-dimensional echo samples as input data for the airborne radar forward-looking imaging network. The echo sample data is iterated over a long period of time by the traditional ADMM algorithm to generate well-reconstructed data which will be used as labels for the input samples.

[0027] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: the airborne radar forward-looking imaging network based on ADMMN constructed by the present invention is composed of a cascade of multiple explicitly constructed hidden layers. The number of hidden layers can be determined as needed, and the output of each lower layer is used as the input of the higher layer. The true form of the relationship between end-to-end can be discovered, and benign parameters can be learned from a large number of sample-label pairs. Due to the increase in the network hidden layers, it has powerful fitting and generalization capabilities. Moreover, ADMMN does not contain convolutional layers, which effectively avoids the problem of poor interpretability of convolutional neural networks, and can solve the problem of difficult parameter selection faced by existing imaging algorithms based on deconvolution theory. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of the present invention;

[0029] Figure 2 It is the airborne radar forward imaging echo model diagram;

[0030] Figure 3 This is a network diagram of the alternating direction multiplier method;

[0031] Figure 4 This is a schematic diagram of the process of constructing a high-fidelity SAR image simulation sample-label set;

[0032] Figure 5 Figure 1 is a graph of simulation data and super-resolution processing results; (a) is the original scene image of the simulated radar scan; (b) is the echo image of the simulated radar real beam scanning; (c) is the IRN forward-looking super-resolution result image; (d) is the ADMMN forward-looking super-resolution result image;

[0033] Figure 6 are the measured data and super-resolution processing result images; among them, (a) is a set of measured data images; (b) is the IRN super-resolution result image of the measured data; (c) is the ADMMN super-resolution result image of the measured data. DETAILED DESCRIPTION

[0034] The present invention will be described in further detail below with reference to the accompanying drawings.

[0035] The present invention proposes an airborne radar forward imaging method based on an alternating direction multiplier method network, such as Figure 1 As shown, the following steps are included:

[0036] The airborne radar forward imaging echo model is as follows: Figure 2 As shown. Define the carrier aircraft's flight direction as the Y axis, the flight speed as V, the flight altitude as H, and the radar antenna working in the forward-looking mode, sweeping across the imaging area at an angular velocity ω, emitting a linear frequency modulation signal with a fixed pulse repetition frequency (PRF). θ is the azimuth angle, is the pitch angle, and φ is the scan angle. When the aircraft flies from position L1 to L2, its distance from the center point P of the imaging area also changes from R0 to R(t). For high-resolution imaging of the range of radar echoes, a wide bandwidth signal can be emitted and pulse compression can be used. The resolution of the range direction after pulse compression is:

[0037]

[0038] Where B is the signal bandwidth and c is the propagation speed of electromagnetic waves.

[0039] For airborne scanning radars flying at low speeds, the Doppler shift in the forward-looking region is minimal. The following expression directly represents the echo signal received by the receiver after eliminating azimuth Doppler modulation and range shift. It is expressed as the convolution kernel formed by the antenna pattern and the range pulse modulation function, convolved with the target backscatter coefficient:

[0040]

[0041] Where R is the distance in the range direction, h(θ) is a function of the antenna pattern, and f pm (·) represents the pulse modulation function, represents the convolution operation, A r is the amplitude constant, and x(R,θ) represents the backscattering coefficient of the target area.

[0042] Ignore the constant term in the above equation and assume that A r is 1, and R is a fixed value, then the above formula can be rewritten as:

[0043]

[0044] Due to the existence of noise n(θ), the actual echo expression is:

[0045]

[0046] From the perspective of matrix analysis, convolution operations in the time domain and frequency domain are corresponding, and the above formula can be simplified to matrix vector form:

[0047] y=Hx+n

[0048] As can be seen, the azimuth echo y is constructed as the product of the antenna pattern matrix H and the target scattering coefficient vector x, superimposed with noise n. Given the known echo vector and antenna measurement matrix, the problem of improving azimuth resolution can be transformed into a corresponding inversion process. Theoretically, the inverse of H can be used to recover the target and improve azimuth resolution. However, the inherent characteristics of the convolution matrix make the inversion process ill-conditioned. From the perspective of matrix analysis, the condition number of H is large. In an environment with small perturbations, the value obtained by inversion will deviate significantly from the true value, resulting in an invalid solution. Therefore, forward-looking super-resolution imaging can be equivalent to solving an ill-conditioned inverse problem, and DNNs are one of the effective ways to overcome this ill-conditioned problem and restore the original signal.

[0049] In light of this principle, the present invention proposes an airborne radar forward-looking imaging method based on an alternating direction multiplier network. The ADMMN architecture consists of a cascade of multiple explicitly constructed hidden layers. The number of hidden layers can be determined as needed. The output of each lower layer serves as the input to the higher layer, allowing the true form of end-to-end relationships to be discovered and benign parameters to be learned from a large number of sample-label pairs. This paper focuses on the structural analysis of the ADMMN, designs a suitable network architecture based on the specific problem of forward-looking super-resolution imaging, and inputs a sample-label set of simulated single-channel, highly realistic complex SAR images into the network.

[0050] A single channel is used to receive the airborne radar echo signal. After pulse compression and range migration, the real beam scanning result, i.e., the complex image data, is input into ADMMN for training.

[0051] After pulse compression and range migration, the airborne radar echo signal received by a single channel is used to obtain the real beam scanning result, i.e., the complex image data, which is then fed into the ADMMN for level-by-level training. Taking level x as an example, the gradients of all levels except level x are first fixed, and the parameters of level x are pre-optimized. Then, the gradients of levels 1 to x-1 are released, and the parameters of the first x levels are optimized end-to-end, and so on.

[0052] like Figure 3 As shown in the figure, the generalized learning capability of deep learning high-dimensional features is used to construct an airborne radar forward imaging network based on ADMMN. It consists of 10 levels, each of which contains three sublayers: reconstruction layer X, nonlinear transformation layer Z, and multiplier update layer β. After the network receives low-resolution two-dimensional real beam data, the reconstruction layer finally outputs a high-resolution reconstruction result. The explicit formula of the reconstruction layer is:

[0053] x (n+1) =(H T H+ρ (n+1) I) -1 (H T y+ρ (n+1) (z n -β n ))

[0054] In the above formula, H is the antenna pattern matrix, x (n+1) is the reconstruction result of the n+1th level, y is the input sample data, ρ n+1 is the trainable penalty parameter of the n+1th level, and Z (0) and β (0) Initialized to 0.

[0055] The explicit formula of the nonlinear transformation layer is:

[0056] z (n+1) =S(x n+1 +β n ;λ (n+1) / ρ (n+1) )

[0057] In the Z layer, the trainable parameter is the regularization parameter λ of the n+1th level (n+1) With penalty parameter ρ n+1 , β (0) It needs to be initialized to 0, S(·) is the soft threshold operation function, and its specific form is as follows:

[0058]

[0059] The explicit formula for the multiplier update layer β is:

[0060] β (n+1) =β n +η (n+1) (x (n+1) -z (n+1) )

[0061] In the β layer, the trainable parameter is the step size parameter η of the n+1th level (n+1) .

[0062] At any level, the input data from the previous level passes through the reconstruction layer to obtain a reconstructed front-view image. A sparse representation of the front-view image is then obtained through the nonlinear transformation layer. Finally, the multiplier update layer is used to update the Lagrange multipliers. By training the learnable parameters of the network layer by layer, adaptive adjustment of the three parameters at each level is achieved. Furthermore, a step-wise learning rate decay is employed, with a higher learning rate initially selected and then decreasing over time, enabling the network to reach the optimal solution more quickly.

[0063] The high-precision training of the airborne radar forward-looking imaging network is completed based on the training method of high-realistic SAR image simulation sample-label set. Compared with other research fields, radar has a unique military background, which makes it difficult to obtain public measured data samples that can be used for network training, and it is even more difficult to obtain the corresponding measured labels. In view of this situation, a method of using high-realistic SAR image simulation sample-label set to implement network training is proposed. First, based on the radar parameters of multiple sets of measured data and the geometric relationship of in-flight data acquisition, high-resolution SAR images can be used as ground scattering scenes and real beam echo simulation technology to generate two-dimensional echo data as samples. Correspondingly, the echo sample data is well-reconstructed after a long period of iteration of the traditional ADMM algorithm and used as the label of the input sample, such as Figure 4 The trained network model parameters are imported into ADMMN, and the low-resolution simulated image and the radar measured image are input to obtain the high-resolution reconstructed image.

[0064] In order to test the performance of ADMMN, the ADMMN-based airborne radar forward-looking imaging network designed by the present invention is used to perform forward-looking super-resolution processing on a set of airborne radar real-beam scanning simulation data and a set of airborne radar measured data. In order to compare with the existing methods, the iterative reweighted norm (IRN) method is also used to process the data. Figure 5 (a) is the original simulation scene, which shows the environment of the airport. In fact, the beam imaging results are as follows Figure 5 As shown in (b), the runway and buildings in the airport cannot be distinguished due to the low azimuth resolution. Figure 5 The ADMMN super-resolution effect shown in (d) is better than Figure 5The IRN method in (c) effectively improves the azimuth resolution, and the runway and building outlines are clearer. In order to further verify the superiority of ADMMN, Figure 6 As shown in (a), a set of measured data containing five isolated ship targets was selected, and it can be seen that the azimuth resolution is poor. Figure 6 As for the super-resolution results of IRN in (b), Figure 6 The ADMMN super-resolution result in (c) is more improved and has smaller sidelobes, and also shows better performance when there are many isolated strong scattering points.

Claims

1. An airborne radar forward imaging method based on an alternating direction multiplier method network, characterized in that: The following steps are involved: (1) A single channel is used to receive the airborne radar echo signal. After pulse compression and range migration, the obtained real beam scanning result, i.e., the complex image data, is input into the alternating direction multiplier method network for training. (2) Utilizing the generalization learning capability of high-dimensional features in deep learning, an airborne radar forward-looking imaging network based on ADMMN is constructed; the network comprises n levels, each level comprising three sublayers, namely, a reconstruction layer X, a nonlinear transformation layer Z, and a multiplier update layer β, with the reconstruction layer outputting the result; in any level, the input data of the previous level is passed through the reconstruction layer to obtain a reconstructed forward-looking image, and then through the nonlinear transformation layer to obtain a sparse representation of the forward-looking image, and finally the multiplier update layer is used to implement the update of the Lagrange multiplier; The explicit formula for the reconstruction layer X is: x (n+1) =(H T H+r (n+1) I) -1 (H T y+r (n+1) (z n -b n )) Where H is the antenna pattern matrix, x (n+1) is the reconstruction result of the n+1th level, y is the input sample data, ρ n+1 is the trainable penalty parameter of the n+1th level, and Z (0) and β (0) is initialized to 0; The explicit formula of the nonlinear transformation layer is: z (n+1) =S(x n+1 +b n ;l (n+1) / r (n+1) ) In the Z layer, z (n+1) is the sparse representation result of the n+1th level, and the trainable parameter is the regularization parameter λ of the n+1th level (n+1) With penalty parameter ρ n+1 , β (0) It needs to be initialized to 0, S(·) is the soft threshold operation function, and its specific form is as follows: Among them, x, λ and ρ are the output result, regularization parameter and penalty parameter of the reconstruction layer X respectively; The explicit formula for the multiplier update layer β is: b (n+1) =b n +n (n+1) (x (n+1) -z (n+1) ) In the β layer, β (n+1) is the Lagrange multiplier of the n+1th level, and the trainable parameter is the step size parameter η of the n+1th level (n+1) ; (3) A training method based on a high-fidelity SAR image simulation sample-label set completes high-precision training of the airborne radar forward-looking imaging network; (4) Input the low-resolution simulated image and the radar measured image in azimuth to obtain a high-resolution reconstructed image.

2. The airborne radar forward imaging method based on the alternating direction multiplier method network according to claim 1, characterized in that: The implementation process of step (1) is as follows: For the airborne radar echo signal received by a single channel, after pulse compression and range migration, the real beam scanning result, i.e., the complex image data, is input into the ADMMN for step-by-step training. The gradients of all levels except the xth level are fixed, and the parameters of the xth level are pre-optimized separately. The gradients of levels 1 to x-1 are released, and the parameters of the first x levels are optimized end-to-end.

3. The airborne radar forward imaging method based on the alternating direction multiplier method network according to claim 1, characterized in that: The implementation process of step (3) is as follows: High-resolution SAR images are used as ground scattering scenes and real-beam echo simulation technology is used to generate two-dimensional echo samples as input data for the airborne radar forward-looking imaging network. The echo sample data is iterated over a long period of time by the traditional ADMM algorithm to generate well-reconstructed data which will be used as labels for the input samples.

Images