Radar sparse imaging method based on data completion network

Abstract

The invention provides a sparse imaging method based on a data completion network, a radar sparse imaging framework combining the data completion network and an incoherent imaging algorithm is provided, a loss function fusing focusing reconstruction loss and WGAN-GP confrontation loss is adopted for training, sparse range profile data of a limited angle is completed into full-view data, and the full-view data is obtained. And more target details can be recovered during imaging, and the imaging effect is better.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal processing and electromagnetic imaging, and particularly relates to a radar sparse imaging method based on a data completion network. Background Art

[0002] Radar imaging technology is widely used in military and civilian fields. It achieves range resolution by transmitting broadband signals and azimuth resolution by forming a synthetic aperture through multi-angle observations. However, in practical applications, due to the instability of the observation platform or external interference, data sparsity problems often occur. The existing radar sparse imaging methods mainly fall into two categories, one is based on compressive sensing, and the other is a combination of compressive sensing and neural networks.

[0003] The method based on compressive sensing models the radar sparse imaging problem as a signal recovery problem and uses various optimization algorithms to obtain high-quality imaging results. For example, the Orthogonal Matching Pursuit (OMP) algorithm and the Alternating Direction Method of Multipliers (ADMM) algorithm.

[0004] The method combining compressive sensing and neural networks incorporates the neural model into the compressive sensing algorithm by expanding the traditional iterative solution steps into cascaded blocks and replacing the non-linear components with neural layers. Through a large amount of training, data priors can be automatically encoded into the deep neural network, and all parameters can be discriminatively learned. For example, the ADMM-Net algorithm.

[0005] The existing technologies have the following defects:

[0006] Compressive sensing-based algorithms: rely on the assumption of scene sparsity, resulting in poor imaging quality; rely on strict random sampling conditions, and the effect significantly degrades or even fails under the condition of highly sparse azimuth data; require time-consuming iterative optimization.

[0007] Algorithms combining compressive sensing and neural networks: rely on the assumption of scene sparsity, resulting in poor imaging quality; overfitting of the network, poor generalization ability, and the effect significantly degrades or even fails under the condition of highly sparse azimuth data. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide a radar sparse imaging method based on a data completion network, which can solve the problem of poor imaging quality under the condition of highly sparse data.

[0009] A radar sparse imaging method includes:

[0010] Step 1: Use the Full-view Range Image Completion Network FVRPC-Net to process the input local sparse observation data, specifically as follows:

[0011] The FVRPC-Net network structure includes a generator structure and a discriminator structure;

[0012] The generator structure is composed of a coarse generator and a fine generator in cascade, both of which are encoder-decoder structures composed of convolutional layers and dilated convolutional layers; the input of the coarse generator is the local sparse observation data matrix processed by the input mask matrix, and the rough completion result is output; the fine generator processes the rough completion result and generates a fine result after complete completion;

[0013] The discriminator structure includes a global discriminator and a local discriminator; the global discriminator takes the fine result after complete completion as the input and evaluates the global structural integrity of the image; the local discriminator extracts and takes the completed area of the fine result after complete completion as the input and evaluates the continuity within the completed area;

[0014] Step 2: Construct a generator loss function based on the evaluation result output by the generator structure, and construct a discriminator loss function based on the evaluation result output by the discriminator structure;

[0015] Step 3: Train the FVRPC-Net network structure based on the loss function; use the trained FVRPC-Net network structure to complete the limited-angle sparse range image data of the aircraft target into full-view range image data;

[0016] Step 4: Based on the full-view range image data, uniformly "back-project" the measurement values of each projection angle along the original path into the image space, and the projections of all angles are superimposed to form a reconstructed image.

[0017] Preferably, in the input mask matrix, the element 0 represents the part that needs to be completed by the network, and the element 1 represents the known sparse data.

[0018] Preferably, the loss function of the generator structure is as follows:

[0019]

[0020] Where, and respectively represent the rough generator focusing reconstruction loss and the fine generator focusing reconstruction loss; represents taking the expectation; ⊙ is the matrix Hadamard product, that is, element-wise multiplication, M is the attention mask matrix, z is the local sparse observation data matrix processed by the input mask matrix, G coarse (·) and G refine (·) respectively represent the output results of the coarse generator and the fine generator, xr is the true value of the full-view range image data, and || ||1 represents the calculation of the 1-norm.

[0021] Preferably, the pixel values of each row h in the attention mask matrix M are determined by the following formula:

[0022] M(h,:) = γ |H / 2-h| ;

[0023] where H is the total height of the set mask matrix M, and γ is a set coefficient factor related to balancing the convergence speed and reconstruction quality.

[0024] Preferably, the loss function of the discriminator structure is as follows:

[0025]

[0026] where, and represent the local adversarial loss and the global adversarial loss respectively; represents the truncation operation on the completed region;

[0027] and D global (x r ) represent the discrimination losses of the local and global discriminators on the true value image respectively;

[0028] and D global (G refine (G coarse (z))) represent the discrimination losses of the local and global discriminators on the completed image respectively; and represent the discrimination losses of the local and global discriminators on the interpolation samples respectively; λ gp is the gradient penalty coefficient, is the interpolation sample in the image space.

[0029] Preferably, the interpolation sample in the image space is calculated by the following formula:

[0030]

[0031] where x g is the generated image finally output by the FVRPC-Net network, and ∈ is sampled from a uniform distribution between 0 and 1.

[0032] Preferably, the total loss function for training the FVRPC-Net network structure is:

[0033]

[0034] Among them, λ g , λ l , λ c , λ r are the weighting coefficients of the global adversarial loss, the local adversarial loss, the rough generator focusing reconstruction loss, and the refined generator focusing reconstruction loss respectively; in the loss function, min G represents taking the minimum value of the loss of the generator structure; max D represents taking the maximum value of the loss of the discriminator structure.

[0035] Preferably, the weighting coefficient λ g = λ l = 1, λ c = 1.2, λ r = 1.

[0036] Preferably, the method for obtaining the dataset for training the FVRPC-Net network structure includes:

[0037] Collect 3D models of various different types of aircraft, and use CST Studio Suite to simulate the full-view echo data of these models; perform an augmentation operation on the dataset to obtain full-view range images of different sizes.

[0038] Preferably, the augmentation operation includes adding random noise to the full-view range image data and randomly scaling the image scale in the range dimension.

[0039] The present invention has the following beneficial effects:

[0040] The present invention proposes a sparse imaging method based on a data completion network, proposes a radar sparse imaging framework combining a data completion network and an incoherent imaging algorithm, and trains using a loss function that fuses the focusing reconstruction loss and the WGAN-GP adversarial loss, complements the sparse range image data with limited angles into full-view data, can restore more target details during imaging, and has a better imaging effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is the flowchart of the technical solution of the present invention;

[0042] Figure 2 is the structure diagram of FVRPC-Net;

[0043] Figure 3 is the global-local discriminator;

[0044] Figure 4 is the simulation model photo and the corresponding full-view range image;

[0045] Figure 5 is the data augmentation strategy;

[0046] Figure 6 This is the imaging result of the present invention. Detailed implementation manners

[0047] The present invention will be described in detail below with reference to the accompanying drawings and by way of examples.

[0048] The technical solution of the present invention is as Figure 1 shown. To illustrate the source of sparse observation data in radar imaging, in the Figure 1 upper part, taking an aircraft target as an example, the process of obtaining sparse data using an electromagnetic simulation software is shown. After obtaining 360° full-view full-sampling radar data through full-aircraft electromagnetic simulation, we intercept part of the angular range data for downsampling to generate local sparse observation data for target sparse imaging. Figure 1 The lower part is the sparse data imaging scheme in the present invention. It mainly includes two steps, namely full-view range profile completion and incoherent BP imaging. The specific implementation manners are as follows.

[0049] Step 1: Completing sparse data to full-view data

[0050] For the highly sparse range profile sequence data, use the Full-View RangeProfile Completion Network (FVRPC-Net) to complete it into 360° full-view range profile data. Range profile completion is to extrapolate the echo data of a limited angle to the full-view echo data by capturing the correlation between data. Mathematically, this process can be modeled as learning a mapping function f between the limited-angle and full-view range profiles:

[0051] f:X LVRP →X FVRP ;

[0052] where X LVRP represents the data matrix of the limited-view range profile (LVRP), and X FVRP represents the data matrix of the full-view range profile (FVRP).

[0053] The present invention adopts a deep learning method based on a data completion network to implement this range profile completion process. The network structure is as Figure 2 shown. The following will provide the specific implementation details of the generator and discriminator of FVRPC-Net and the network training details.

[0054] The FVRPC-Net network structure includes a generator structure and a discriminator structure:

[0055] Generator Structure: To obtain a more refined completion result, the generator adopts a cascaded "coarse generation - fine generation" structure. The cascaded structure consists of two-stage convolutional neural networks, and each stage is an encoder-decoder structure composed of convolutional layers and dilated convolutional layers. The first-stage network is called the coarse generator, which is responsible for generating a full-view coarse completion result from limited-angle data; the second-stage network is called the refinement generator, which is responsible for refining the coarse completion result of the first stage into a high-quality refined completion result. The networks of both stages are trained by monitoring the generation effect through the reconstruction loss, and the output of the second-stage network is used as the final output and is also trained by monitoring the image difference through the adversarial loss.

[0056] For the forward propagation of the first-stage network, the input of the coarse generator is a data matrix z processed by the input mask matrix, and the output is the coarse completion result G coarse (z). The 0 values in the mask matrix represent the parts that need to be completed by the network, and the 1 values represent the known sparse data. Its function is to indicate which parts of the data need to be completed and which data do not need to be completed. For the forward propagation of the second-stage network, the fine generator directly processes the coarse completion result and generates a refined result G refine (G coarse (z)) with improved details. The two-stage design effectively separates the responsibilities of the generator. The coarse generator focuses on modeling the correlation of adjacent angular distance images, while the refinement generator can focus on smaller-scale features and is dedicated to restoring fine details.

[0057] Discriminator Structure: To ensure both the global integrity and local consistency of the completed FVRP simultaneously, the discriminator adopts a combined "global - local" discriminator structure, and the complete structure is as Figure 3 shown. The global discriminator takes the completed FVRP result G refine (G coarse (z)) as the input to evaluate the global structural integrity of the image; the local discriminator only focuses on the completed region in the FVRP, extracts the completed region of the network in G refine (G coarse (z)) and uses it as the input to evaluate the continuity within the completed region. The intercepted data is represented by , representing the operation of intercepting the completed region. Both discriminators are based on the convolutional neural network structure and compress the image into a low-dimensional feature vector. Finally, the outputs of the two networks are fused and a final scalar value is output, reflecting the overall probability that the completion result is judged as a real image. The mathematical expectation of this probability is also regarded as a loss function and is used as part of the adversarial loss. The global and local discriminator losses can be mathematically expressed as:

[0058]

[0059] Step 2, Loss function design:

[0060] (a) Focused reconstruction loss. To guide the generator to pay more attention to the movement trends of scatterers at different observation angles during the learning process and reduce the interference of irrelevant information, a focused reconstruction loss based on the attention mechanism is proposed. Compared with the original reconstruction loss, the focused reconstruction loss can prompt the network to focus on the middle region of the range image data matrix that contains more information, thus accelerating the network convergence.

[0061] Specifically, the focused reconstruction loss weights the original loss function with the attention mask matrix M. The pixel value of the h-th row of the mask matrix M is determined by the following formula:

[0062] M(h,:) = γ |H / 2-h| ;

[0063] where H is the total height of the mask matrix, and γ is a coefficient factor related to balancing the convergence speed and reconstruction quality. During training, the attention mask matrix M is multiplied by the ground truth x r as well as the output results of the coarse network and the fine network respectively. Then, the two calculation results are subtracted and the L1 norm of the difference matrix is obtained. Finally, the reconstruction loss is obtained by calculating the expectation over all training data. The reconstruction loss is expressed as follows:

[0064]

[0065] where, and represent the losses of the coarse generator and the fine generator respectively; represents taking the expectation; ⊙ is the matrix Hadamard product, that is, element-wise multiplication, M is the mask matrix, z is the input of the completion network, G coarse () and G refine () represent the output results of the coarse generator and the fine generator respectively, x r is the ground truth of the full-view range image data, and || ||1 represents taking the 1-norm.

[0066] (b) Adversarial loss. In the present invention, WGAN-GP is used as the adversarial loss. The WGAN loss with gradient penalty can be expressed as the following formula, where the first two terms calculate the "Wasserstein distance" between the generated data and the real data distribution, and the last term is the gradient penalty term:

[0067]

[0068] where, and represent the local and global losses respectively; z is the input of the completion network, x r is the ground truth, and D global (x r ) respectively represent the discrimination losses of the local and global discriminators for the ground truth images; and D global (G refine (G coarse (z))) respectively represent the discrimination losses of the local and global discriminators for the completed images; and respectively represent the discrimination losses of the local and global discriminators for the interpolated samples; λ gp is the gradient penalty coefficient, is the interpolated sample in the image space and can be calculated by the following formula:

[0069]

[0070] where x g is the generated image finally output by the completion network, sampled from a uniform distribution between 0 and 1. It can be seen that the interpolated samples are obtained by linearly interpolating the real image and the generated image. Finally, the total loss can be expressed as:

[0071]

[0072] where λ g , λ l , λ c , λ r are the weighting coefficients of the global adversarial loss, local adversarial loss, rough generator focused reconstruction loss, and refinement generator focused reconstruction loss respectively, and the quality of the generated samples is controlled by balancing each weight.

[0073] Step 3, Training of the FVRPC-Net network:

[0074] A) Training dataset: To train the FVRPC-Net, a set of high-quality training datasets containing limited-angle, sparse range image data of the target and corresponding full-view range image data is required. Since there is no publicly available high-quality dataset currently, we create a synthetic dataset to implement network training. We collected 3D models of six different types of aircraft (two fixed-wing aircraft, two airliners, and two fighter jets) and used CST Studio Suite to simulate the full-view echo data of these models. Photos of the aircraft models and full-view range images are as Figure 4 shown.

[0075] And, we performed some necessary augmentation operations on the dataset. We first added random noise to the full-view range image data, controlling its signal-to-noise ratio range between 15 dB and 30 dB. Subsequently, we randomly scaled the image scale in the range dimension to obtain full-view range images of different sizes. The data augmentation strategy is asFigure 5 as shown

[0076] B) Training process: First, some preprocessing is performed on the training data. The full-view data obtained by simulation consists of 128 distance-dimensional sampling points and 7200 azimuth-dimensional sampling points (azimuth sampling interval 0.05°, azimuth angle range 360°). First, we downsample the data in the azimuth dimension by 20 times (azimuth sampling interval 1°) and perform data normalization to adjust the data dynamics to [-1, 1]. Subsequently, we select partial angle-distance images within an angle range of 30° and fill them with 0 elements to form a data matrix. Finally, the data matrix and the input mask matrix are stacked together as the input data for the data completion network, as Figure 2 shown. After passing through FVRPC-Net, the network outputs the completed full-view distance image.

[0077] When training the network, set the hyperparameters λ g = λ l = 1, λ c = 1.2, λ r = 1, γ = 0.97. The network parameters are optimized using the Adam algorithm, and the learning rate is set to l r = 0.001, Adam momentum parameters β1 = 0.5, β2 = 0.9.

[0078] After sufficient training, FVRPC-Net can complete the limited-angle sparse distance image data of the aircraft target into full-view distance image data.

[0079] Step 4: Incoherent backprojection imaging:

[0080] Backprojection means uniformly "backprojecting" the measurement values of each projection angle along the original path into the image space, and the projections of all angles are superimposed to form the reconstructed image. The mathematical form can be expressed as:

[0081] where r = xcosθ + ysinθ.

[0082] The imaging result of the aircraft model in the present invention is as Figure 6 shown.

[0083] The method of the present invention has the following technical effects:

[0084] 1. Imaging ability for high-sparsity data: It can obtain high-quality imaging at an extremely high sparsity (angle sampling interval 1°). When traditional coherent imaging methods process sparse data, due to non-compliance with the spatial sampling condition in the azimuth direction, azimuth aliasing will occur in the imaging result. The sampling condition can be described by the following formula:

[0085]

[0086] where λ c is the electromagnetic wave wavelength at the center frequency, and Y max is the lateral distance size of the imaging interval. Taking the millimeter wave band as an example, when imaging a target with a length of about 10 m, an angular sampling interval of about 0.06° is required to avoid azimuth aliasing. Some sparse imaging methods based on compressive sensing can handle imaging tasks under certain sparse conditions. However, under highly sparse conditions (such as an undersampling ratio of 5%), they have been proven to be infeasible.

[0087] The incoherent imaging method used in the present invention has much lower restrictions on spatial sampling than the coherent imaging method. According to the imaging theory, when the condition that the target observation angle is greater than 180° is satisfied, the incoherent backprojection algorithm only needs to meet the sampling condition:

[0088]

[0089] where △r is the radial distance resolution. Assuming that the radial distance resolution is 0.15 m, when imaging a target with a length of 10 m, the angular sampling interval only needs to be less than 1.72°, greatly reducing the dependence on high-sampling-rate data. At the same time, since the 180° observation angle condition has been achieved by step 1 of the present invention, by combining the two processes, imaging of highly sparse data can ultimately be achieved.

[0090] 2. Good imaging quality: The present invention uses a neural network encoded with the prior of the full-view range image of an aircraft target to complete the sparse range image data with a limited angle into full-view data, and more target details can be restored during imaging, resulting in a better imaging effect.

[0091] In summary, the above are only the preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A radar sparse imaging method, characterized in that, Including: Step 1: Process the input local sparse observation data using the Full-View Range Profile Completion Network (FVRPC-Net). Specifically: The FVRPC-Net network structure includes a generator structure and a discriminator structure; The generator structure is composed of a coarse generator and a fine generator in cascade. The input of the coarse generator is the local sparse observation data matrix processed by the input mask matrix, and the output is the rough completion result; The fine generator processes the rough completion result and generates the refined result after complete completion; The discriminator structure includes a global discriminator and a local discriminator. The global discriminator takes the refined result after complete completion as the input and evaluates the global structural integrity of the image. The local discriminator extracts the completed area of the refined result after complete completion and takes it as the input to evaluate the continuity within the completed area; Step 2: Construct a generator loss function based on the evaluation result output by the generator structure, and construct a discriminator loss function based on the evaluation result output by the discriminator structure; Step 3: Train the FVRPC-Net network structure based on the loss function; use the trained FVRPC-Net network structure to complete the limited-angle sparse range profile data of the aircraft target into full-view range profile data; Step 4: Based on the full-view range profile data, uniformly "back-project" the measurement values of each projection angle along the original path into the image space, and the projections of all angles are superimposed to form a reconstructed image.

2. The radar sparse imaging method according to claim 1, characterized in that In the input mask matrix, the element 0 represents the part that needs to be completed by the network, and the element 1 represents the known sparse data.

3. A radar sparse imaging method according to claim 2, characterized in that The loss function of the generator structure is as follows: Among them, and respectively represent the rough generator focus reconstruction loss and the fine generator focus reconstruction loss; represents taking the expectation; ⊙ is the matrix Hadamard product, that is, element-wise multiplication. M is the attention mask matrix, z is the local sparse observation data matrix processed by the input mask matrix, G coarse (·) and G refine (·) represent the output results of the coarse generator and the fine generator respectively. x r is the true value of the full-view range image data, and ||||1 represents the calculation of the 1-norm.

4. The radar sparse imaging method according to claim 3, characterized in that The pixel value of each row h in the attention mask matrix M is determined by the following formula: M(h,:) = γ |H / 2-h| ; where H is the total height of the set mask matrix M, and γ is a set coefficient factor related to balancing the convergence speed and reconstruction quality.

5. A radar sparse imaging method according to claim 3 or 4, characterized in that The loss function of the discriminator structure is as follows: Among them, and represent the local adversarial loss and the global adversarial loss respectively; represents the operation of intercepting the completed region; and D global (x r ) respectively represent the discrimination losses of the local and global discriminators for the true-value images; and D global (G refine (G coarse (z))) denote the discrimination losses of the local and global discriminators for the completed images respectively; and denote the discrimination losses of the local and global discriminators for the interpolated samples respectively; λ gp is the gradient penalty coefficient, is the interpolated sample in the image space.

6. A radar sparse imaging method according to claim 5, characterized in that Interpolated samples in the image space Calculated by the following formula: Among them, x g is the generated graph finally output by the FVRPC-Net network, and ∈ is sampled from a uniform distribution between 0 and 1.

7. A radar sparse imaging method according to claim 6, characterized in that, The total loss function for training the FVRPC-Net network structure is: Among them, λ g , λ l , λ c , λ r are the weighting coefficients of the global adversarial loss, the local adversarial loss, the rough generator focused reconstruction loss, and the refined generator focused reconstruction loss respectively; in the loss function, min G represents taking the minimum value of the loss of the generator structure; max D represents taking the maximum value of the loss of the discriminator structure.

8. A radar sparse imaging method according to claim 7, characterized in that Weighting coefficient λ g = λ l = 1, λ c = 1.2, λ r = 1.

9. A radar sparse imaging method according to claim 5, characterized in that, The method for obtaining the dataset used to train the FVRPC-Net network structure includes: Collect 3D models of various different types of aircraft, and use CST Studio Suite to simulate the full-view echo data of these models; perform an expansion operation on the dataset to obtain full-view range profiles of different sizes.

10. A radar sparse imaging method according to claim 9, characterized in that, The expansion operation includes adding random noise to the full-view range profile data and randomly scaling the image scale in the range dimension.

Images