A neural network-based integrated method for super-resolution imaging and detection in scanning radar.

Abstract

This invention discloses an integrated super-resolution imaging and detection method for scanning radar based on neural networks. The method constructs an objective function based on the sparsity of the target object, the low-rank nature of the background, and convolutional inversion theory. The ADMM algorithm is used to solve the objective function, and the solution process is expanded into a parameter-learnable neural network. The parameters in the model are optimized and learned through a forward gradient iterative algorithm, thus transforming the problem of simultaneous super-resolution and target detection into an image loss optimization problem. Finally, through iterative training, the mapping relationship between the real-beam image, the super-resolution result, and the target detection result is obtained, thereby simultaneously achieving super-resolution imaging and target detection. This method simultaneously achieves super-resolution imaging of the scene and detection of sparse targets of interest. It can adaptively determine parameters based on information from different scenes, eliminating the step of manually setting parameters, effectively improving the adaptability of the method, and also improving processing efficiency.

Description

Technical Field

[0001] This invention belongs to the field of radar imaging technology, specifically relating to an integrated method for super-resolution imaging and detection of scanning radar based on neural networks. Background Technology

[0002] Compared to optical and infrared imaging and detection technologies, radar imaging and detection can operate around the clock and in all weather conditions, and has extremely wide applications in the civilian sector. However, due to limitations in imaging principles, existing Synthetic Aperture Radar (SAR) imaging and Doppler Beam Sharpening (DBS) suffer from slow Doppler gradient changes and Doppler symmetry blurring in the forward-looking imaging region, resulting in a forward-looking imaging blind zone.

[0003] Airborne forward-looking scanning radars can obtain real-beam images of the forward-looking area by scanning antenna beams. However, due to antenna size limitations, the azimuth resolution of real-beam images is usually low. How to enhance the resolution of real-beam images and perform target detection has been extensively studied in recent years.

[0004] The literature “Y. Zhang, Y. Zhang, Y. Huang, et al. A Sparse bayesian approach for forward-looking superresolution radar imaging. Sensors, 2017, 17(6): 1353” combines Bayesian theory to study a forward-looking superresolution imaging method applied to airborne radar, but does not focus on the detection of targets of interest. The literature “Yu W, Wang Y, Liu H, et al. Superpixel-Based CFAR Target Detection for High-Resolution SAR Images. IEEE Geoscience and Remote Sensing Letters, 2016, 13(5): 1-5” proposes a superpixel-based constant false alarm rate (CFAR) target detection algorithm for SAR imaging results. This method first uses a superpixel generation algorithm to segment the image, then adaptively estimates the distribution parameters of clutter for each pixel based on superpixels, and finally uses hierarchical clustering to obtain the target detection results. The methods described above all treat imaging and target detection as two independent steps: first, a high-resolution image is obtained, and then different detection methods are studied based on this image. Separating imaging and target detection reduces processing efficiency, and the performance of the imaging directly affects the performance of the target detection, leading to error propagation.

[0005] The paper "Li W, Zhang W, Zhang Q, et al. Simultaneous Super-Resolution and Target Detection of Forward-Looking Scanning Radar via Low-Rank and SparsityConstrained Method. IEEE Transactions on Geoscience and Remote Sensing, 2020, 58(10): 7085-7095" studies an integrated method for super-resolution and target detection of forward-looking scanning radar by introducing low-rank sparse matrix factorization theory and optimization algorithm. However, it has the problems of difficult parameter tuning and long iterative calculation time. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes an integrated method for super-resolution imaging and detection of scanning radar based on neural networks.

[0007] The technical solution adopted in this invention is: an integrated method for super-resolution imaging and detection of scanning radar based on neural networks, the specific steps of which are as follows:

[0008] S1. Scan the radar to obtain echo data, perform pulse compression processing on the echo data to obtain the pulse-compressed echo signal matrix;

[0009] The scanning radar antenna moves at a certain angular velocity The forward-looking area is scanned sequentially from left to right (or right to left), transmitting a wide-bandwidth linear frequency modulated signal at a fixed pulse repetition frequency, and receiving and storing the echo signals reflected from the imaging area, storing them in a two-dimensional matrix according to range and azimuth directions. Then, the echo data is pulse-compressed to obtain the pulse-compressed echo signal matrix. ;

[0010] S2. Construct the azimuth convolution measurement matrix using the antenna pattern function;

[0011] Let represent the antenna pattern vector, then the constructed azimuth convolution measurement matrix... The expression is as follows:

[0012] (1)

[0013] in, , Indicates the azimuth beamwidth of the antenna pattern. Indicates angular velocity, Indicates the pulse repetition frequency , This indicates the number of sampling points in the azimuth direction.

[0014] S3. Azimuth echo modeling of real beam scanning radar;

[0015] The azimuth echo model of a real-beam scanning radar is the convolution of the target's scattering coefficient and the antenna pattern plus noise, and its vector multiplication form is:

[0016] (2)

[0017] in, This represents the echo signal matrix after pulse compression. Represents the target scattering coefficient. Indicates noise.

[0018] S4. Based on the model in step S3, construct the objective function using the least squares convolution inversion method and low-rank sparsity theory;

[0019] By using the low-rank and sparse prior information of the scene as penalty constraints to construct the objective function, we can obtain:

[0020] (3)

[0021] in, This represents the sparse target portion. This represents the low-rank background portion. express Norm, The nuclear norm of a matrix is ​​represented by its nucleus. express Norm, , and represents the regularization parameters for the super-resolution part, low-rank part, and sparse part, respectively.

[0022] Equation (3) is transformed into an unconstrained optimization problem by augmenting the Lagrangian function, and the objective function is obtained as follows:

[0023] (4)

[0024] in, Indicates the inner product. Represents the Lagrange multipliers. This represents the penalty factor.

[0025] S5. Based on ADMM, derive the sub-problem model for sparse objectives;

[0026] According to equation (4), the low-rank background is fixed. Target scattering coefficient and Lagrange multipliers Then, the sparse target hierarchical update model is obtained by combining the ADMM algorithm. :

[0027] (5)

[0028] in, Indicates the number of iterations. The soft threshold operator is represented as follows:

[0029] (6)

[0030] in, , Represents a symbolic function. The threshold is represented as follows:

[0031] (7)

[0032] Introduce a learnable parameter To replace the soft threshold operator To participate in the iterative process, its specific expression is as follows:

[0033] (8)

[0034] in, and Indicates the parameters used for calculation Two factors.

[0035] Punishment factor This is transformed into a learnable parameter, specifically expressed as follows:

[0036] (9)

[0037] in, and Indicates the parameters used for calculation Two factors;

[0038] Replace equation (5) with equation (8). , and equation (9) substitution The following model can be obtained:

[0039] (10)

[0040] S6. Based on ADMM, derive the sub-problem model for low-rank background;

[0041] Fixed sparse target Target scattering coefficient and Lagrange multipliers Then, the hierarchical model of the low-rank background part is obtained by solving the problem using the ADMM algorithm. :

[0042] (11)

[0043] in, , , They represent The left unitary matrix, the transpose of the right unitary matrix, and the singular value matrix of singular value decomposition are introduced with a learnable parameter. To replace the soft threshold operator To participate in network updates, the specific expression is as follows:

[0044] (12)

[0045] in, and Indicates the parameters used for calculation The two factors are replaced by equation (11) with equation (12). The following model can be obtained:

[0046] (13)

[0047] S7. Based on ADMM, derive the sub-problem model regarding the target scattering coefficient;

[0048] Fixed sparse target Low-rank background and Lagrange multipliers Afterwards, I received information about The optimization problem yields a hierarchical model of the target scattering coefficients. :

[0049] (14)

[0050] in, Represents the identity matrix. This represents the conjugate transpose operation on a matrix, introducing a learnable parameter. To replace regularization parameters It participates in the update process of the super-resolution layer, and its specific expression is as follows:

[0051] (15)

[0052] in, and Indicates the parameters used for calculation The two factors are replaced by equation (14) with equation (15). , and equation (9) substitution The following model can be obtained:

[0053] (16)

[0054] S8. Update the Lagrange hierarchy;

[0055] Adding equation (9) to the network update yields the following Lagrange hierarchy model:

[0056] (17)

[0057] S9. Define the loss function to achieve super-resolution imaging of the scene and detection of sparse targets of interest;

[0058] The mean squared error (MSE) is used as the loss function.

[0059] (18)

[0060] in, Denotes the Frobenius norm of a matrix. and The weighting parameters represent the loss calculation between the target detection result and the super-resolution result. , This represents the input for each iteration, i.e., the output of the previous iteration; , This represents the output of each iteration.

[0061] When the loss function converges to a certain range , This refers to the target detection results and super-resolution imaging results ultimately output by the network.

[0062] The beneficial effects of this invention are as follows: The method of this invention constructs an objective function based on the sparsity of the target to be detected, the low-rank nature of the background, and convolutional inversion theory. It uses the ADMM algorithm to solve the objective function, expanding the solution process into a parameter-learnable neural network. The parameters in the model are then optimized and learned through a forward gradient iterative algorithm, thereby transforming the problem of simultaneous super-resolution and target detection into an image loss optimization problem. Finally, through iterative training, the mapping relationship between the real-beam image, the super-resolution result, and the target detection result is obtained, thus simultaneously achieving super-resolution imaging and target of interest detection. This invention's method simultaneously achieves super-resolution imaging of the scene and detection of sparse targets of interest. It can adaptively determine parameters based on information from different scenes, eliminating the step of manually setting parameters, effectively improving the method's adaptability and processing efficiency. Attached Figure Description

[0063] Figure 1 This is a flowchart of an integrated super-resolution imaging and detection method for scanning radar based on neural networks, according to the present invention.

[0064] Figure 2 This is a simulation scene diagram used in the embodiments of the present invention.

[0065] Figure 3 This is a schematic diagram of the echo signal after pulse compression in an embodiment of the present invention.

[0066] Figure 4 This is a network structure diagram of the method of the present invention in an embodiment of the present invention.

[0067] Figure 5 This is a super-resolution imaging result obtained in an embodiment of the present invention.

[0068] Figure 6 This is a diagram of the target detection results obtained in an embodiment of the present invention. Detailed Implementation

[0069] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0070] This embodiment mainly uses simulation experiments for verification. All steps and conclusions have been verified to be correct on the Windows 10 operating system platform using Matlab 2018a.

[0071] like Figure 1 The flowchart of the integrated super-resolution imaging and detection method for scanning radar based on neural networks of the present invention is shown below, and the specific steps are as follows:

[0072] S1. Scan the radar to obtain echo data, perform pulse compression processing on the echo data to obtain the pulse-compressed echo signal matrix;

[0073] The scanning radar antenna moves at a certain angular velocity The forward-looking area is scanned sequentially from left to right (or right to left), transmitting a wide-bandwidth linear frequency modulated signal at a fixed pulse repetition frequency, and receiving and storing the echo signals reflected from the imaging area, storing them in a two-dimensional matrix according to range and azimuth directions. Then, the echo data is pulse-compressed to obtain the pulse-compressed echo signal matrix. .

[0074] S2. Construct the azimuth convolution measurement matrix using the antenna pattern function;

[0075] Let represent the antenna pattern vector, then the constructed azimuth convolution measurement matrix... The expression is as follows:

[0076] (1)

[0077] in, , Indicates the azimuth beamwidth of the antenna pattern. Indicates angular velocity, Indicates the pulse repetition frequency , This indicates the number of sampling points in the azimuth direction.

[0078] S3. Azimuth echo modeling of real beam scanning radar;

[0079] Since the Doppler variation in the forward-looking region is negligible, the azimuth echo of a real-beam scanning radar can be modeled as the convolution of the target's scattering coefficient and the antenna pattern plus noise, with the vector multiplication form being:

[0080] (2)

[0081] in, This represents the echo signal matrix after pulse compression. Represents the target scattering coefficient. Indicates noise.

[0082] S4. Based on the model in step S3, construct the objective function using the least squares convolution inversion method and low-rank sparsity theory;

[0083] However, due to the convolution measurement matrix The ill-conditioned nature of the image makes it highly sensitive to noise interference, severely impacting imaging results. To address this issue, the objective function can be constructed by using the low-rank and sparse prior information of the scene as penalty constraints, resulting in:

[0084] (3)

[0085] in, This represents the sparse target portion. This represents the low-rank background portion. express Norm, The nuclear norm of a matrix (the sum of the singular values ​​of the matrix). express Norm, , and represents the regularization parameters for the super-resolution part, low-rank part, and sparse part, respectively.

[0086] Equation (3) is transformed into an unconstrained optimization problem by augmenting the Lagrangian function, and the objective function is obtained as follows:

[0087] (4)

[0088] in, Indicates the inner product. Represents the Lagrange multipliers. This represents the penalty factor.

[0089] S5. Based on ADMM, derive the sub-problem model for sparse objectives;

[0090] According to equation (4), the low-rank background is fixed. Target scattering coefficient and Lagrange multipliers Then, by combining the Alternating Direction Method of Multipliers (ADMM) algorithm, a sparse target hierarchical update model can be obtained. :

[0091] (5)

[0092] in, Indicates the number of iterations. The soft threshold operator is represented as follows:

[0093] (6)

[0094] in, , Represents a symbolic function. The threshold is represented as follows:

[0095] (7)

[0096] Introduce a learnable parameter To replace the soft threshold operator To participate in the iterative process, its specific expression is as follows:

[0097] (8)

[0098] in, and Indicates the parameters used for calculation Two factors.

[0099] Punishment factor It plays a crucial role at each module level, therefore, like other parameters, it is transformed into a learnable parameter, as shown below:

[0100] (9)

[0101] in, and Indicates the parameters used for calculation Two factors. The learnable parameters can be adjusted. and To achieve modification The purpose is to replace equation (5) with equation (8). , and equation (9) substitution The following model can be obtained:

[0102] (10)

[0103] S6. Based on ADMM, derive the sub-problem model for low-rank background;

[0104] Fixed sparse target Target scattering coefficient and Lagrange multipliers Then, the hierarchical model of the low-rank background part is obtained by solving the problem using the ADMM algorithm. :

[0105] (11)

[0106] in, , , They represent The left unitary matrix, the transpose of the right unitary matrix, and the singular value matrix of singular value decomposition are introduced with a learnable parameter. To replace the soft threshold operator To participate in network updates, the specific expression is as follows:

[0107] (12)

[0108] in, and Indicates the parameters used for calculation The two factors are replaced by equation (11) with equation (12). The following model can be obtained:

[0109] (13)

[0110] S7. Based on ADMM, derive the sub-problem model regarding the target scattering coefficient;

[0111] Fixed sparse target Low-rank background and Lagrange multipliers Afterwards, I received information about The optimization problem yields a hierarchical model of the target scattering coefficients. :

[0112] (14)

[0113] in, Represents the identity matrix. This represents the conjugate transpose operation on a matrix, and also introduces a learnable parameter. To replace regularization parameters It participates in the update process of the super-resolution layer, and its specific expression is as follows:

[0114] (15)

[0115] in, and Indicates the parameters used for calculation The two factors are replaced by equation (14) with equation (15). , and equation (9) substitution The following model can be obtained:

[0116] (16)

[0117] S8. Update the Lagrange hierarchy;

[0118] Adding equation (9) to the network update yields the following Lagrange hierarchy model:

[0119] (17)

[0120] S9. Define the loss function to achieve super-resolution imaging of the scene and detection of sparse targets of interest;

[0121] This invention aims to simultaneously achieve super-resolution and target detection of real-beam images, while maintaining good adaptability to data of varying sizes. Therefore, the loss value is calculated based on the average offset of the data after each update. Here, the mean squared error (MSE) is used as the loss function.

[0122] (18)

[0123] in, Denotes the Frobenius norm of a matrix. and The weighting parameters represent the loss calculation between the target detection result and the super-resolution result. , This represents the input for each iteration, i.e., the output of the previous iteration; , This represents the output of each iteration. When the loss function converges to a certain range, , This refers to the target detection results and super-resolution imaging results ultimately output by the network.

[0124] In this embodiment, network initialization and parameter settings are performed. The reconstruction matrix of each layer of the network is initialized to 0, and its dimension is the same as the echo signal matrix after input pulse compression. Same; trainable parameter set The initial values ​​are shown in steps S1-S9 above; the learning rate is set to ε=0.001; as the parameters are continuously optimized and the image is reconstructed, the learning rate is adjusted until the loss value is reached. Decrease to a certain range Updates have ceased.

[0125] In this embodiment, the parameter configuration is shown in Table 1. Figure 2 This embodiment uses a simulation scenario to simulate a forward-looking radar real-beam image, resulting in... Figure 3 The pulse-compressed echo signal shown is used to input real-beam echo data into the network model built by the method of this invention (structure diagram as shown). Figure 4 As shown), the final output is... Figure 5 The super-resolution image shown and Figure 6 The target detection results are shown.

[0126] Table 1

[0127]

[0128] In summary, the method of this invention simultaneously achieves super-resolution imaging of a scene and detection of sparse targets of interest. It can adaptively determine parameters based on information from different scenes, eliminating the need for manual parameter setting and effectively improving the method's adaptability and processing efficiency. Furthermore, the ADMM-based unsupervised training method used in this embodiment of the invention eliminates the need for data labeling, significantly improving network training efficiency and shortening training time.

[0129] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for integrating super-resolution imaging and detection of scanning radar based on neural networks, the specific steps of which are as follows: S1. Scan the radar to obtain echo data, perform pulse compression processing on the echo data to obtain the pulse-compressed echo signal matrix; The scanning radar antenna moves at a certain angular velocity The forward-looking area is scanned sequentially from left to right or from right to left. A wide-bandwidth linear frequency modulated signal is transmitted at a fixed pulse repetition frequency, and the echo signals reflected from the imaging area are received and stored in a two-dimensional matrix according to the range and azimuth directions. Then, the echo data is pulse-compressed to obtain the pulse-compressed echo signal matrix. ; S2. Construct the azimuth convolution measurement matrix using the antenna pattern function; Let represent the antenna pattern vector, then the constructed azimuth convolution measurement matrix... The expression is as follows: （1）； in, , Indicates the azimuth beamwidth of the antenna pattern. Indicates angular velocity. Indicates the pulse repetition frequency , Indicates the number of sampling points in the azimuth direction; S3. Azimuth echo modeling of real beam scanning radar; The azimuth echo model of a real-beam scanning radar is the convolution of the target's scattering coefficient and the antenna pattern plus noise, and its vector multiplication form is: （2）； in, This represents the echo signal matrix after pulse compression. Represents the target scattering coefficient. Indicates noise; S4. Based on the model in step S3, construct the objective function using the least squares convolution inversion method and low-rank sparsity theory; By using the low-rank and sparse prior information of the scene as penalty constraints to construct the objective function, we can obtain: （3）； in, This represents the sparse target portion. This represents the low-rank background portion. express Norm, The nuclear norm of a matrix is ​​represented by its nucleus. express Norm, , and represent the regularization parameters for the super-resolution part, the low-rank part, and the sparse part, respectively; Equation (3) is transformed into an unconstrained optimization problem by augmenting the Lagrangian function, and the objective function is obtained as follows: （4）； in, Indicates the inner product. Represents the Lagrange multipliers. Indicates the penalty factor; S5. Based on ADMM, derive the sub-problem model for sparse objectives; According to equation (4), the low-rank background is fixed. Target scattering coefficient and Lagrange multipliers Then, the sparse target hierarchical update model is obtained by combining the ADMM algorithm. : （5）； in, Indicates the number of iterations. The soft threshold operator is represented as follows: （6）； in, , Represents a symbolic function. The threshold is represented as follows: （7）； Introduce a learnable parameter To replace the soft threshold operator To participate in the iterative process, its specific expression is as follows: （8）； in, and Indicates the parameters used for calculation Two factors; Punishment factor This is transformed into a learnable parameter, specifically expressed as follows: （9）； in, and Indicates the parameters used for calculation Two factors; Replace equation (5) with equation (8). , and equation (9) substitution The following model can be obtained: （10）； S6. Based on ADMM, derive the sub-problem model for low-rank background; Fixed sparse target Target scattering coefficient and Lagrange multipliers Then, the hierarchical model of the low-rank background part is obtained by solving the problem using the ADMM algorithm. : （11）； in, , , They represent The left unitary matrix, the transpose of the right unitary matrix, and the singular value matrix of singular value decomposition are introduced with a learnable parameter. To replace the soft threshold operator To participate in network updates, the specific expression is as follows: （12）； in, and Indicates the parameters used for calculation The two factors are replaced by equation (11) with equation (12). The following model can be obtained: （13）； S7. Based on ADMM, derive the sub-problem model regarding the target scattering coefficient; Fixed sparse target Low-rank background and Lagrange multipliers Afterwards, I received information about The optimization problem yields a hierarchical model of the target scattering coefficients. : （14）； in, Represents the identity matrix. This represents the conjugate transpose operation on a matrix, introducing a learnable parameter. To replace regularization parameters It participates in the update process of the super-resolution layer, and its specific expression is as follows: （15）； in, and Indicates the parameters used for calculation The two factors are replaced by equation (14) with equation (15). , and equation (9) substitution The following model can be obtained: （16）； S8. Update the Lagrange hierarchy; Adding equation (9) to the network update yields the following Lagrange hierarchy model: （17）； S9. Define the loss function to achieve super-resolution imaging of the scene and detection of sparse targets of interest; The mean squared error (MSE) is used as the loss function. （18）； in, Denotes the Frobenius norm of a matrix. and The weighting parameters represent the loss calculation between the target detection result and the super-resolution result. , This represents the input for each iteration, i.e., the output of the previous iteration; , This represents the output of each iteration. When the loss function converges to a certain range , This refers to the target detection results and super-resolution imaging results ultimately output by the network.

Images