An au-net-based ISAR enhanced imaging method

Abstract

An ISAR enhanced imaging method based on AU-Net, comprising the following steps: step 1, generating ISAR echo data according to a random scattering point model simulation, and constructing an ISAR image training set through random noise adding, downsampling and imaging preprocessing; step 2, constructing an AU-Net imaging network for ISAR sparse high-resolution imaging by introducing an attention mechanism into a U-Net structure, and initializing network model parameters; step 3, designing an ISAR imaging loss function according to the constructed AU-Net imaging network; step 4, formulating an ISAR enhanced imaging training strategy for training, updating AU-Net imaging network model parameters to obtain an optimal ISAR high-resolution imaging model; step 5, acquiring simulated / measured target ISAR echo data, and performing the same preprocessing operation as step 1 to construct an ISAR image test set, and the optimal ISAR high-resolution imaging model can be used to realize ISAR enhanced imaging. The present application can improve the ISAR imaging performance under the conditions of low signal-to-noise ratio and data loss, and further improve the recovery ability for weak scattering points.

Description

Technical Field

[0001] This invention belongs to the field of radar imaging technology, specifically relating to an ISAR enhanced imaging method based on AU-Net. Background Technology

[0002] Inverse Synthetic Aperture Radar (ISAR) technology can acquire high-resolution images of non-cooperative targets under all-weather, all-day conditions. Traditional ISAR imaging is generally based on the Range Doppler (RD) algorithm, while in recent years, Compressed Sensing (CS) algorithms have attracted much research and have been widely applied to ISAR imaging. Compared to the RD algorithm, the CS algorithm can obtain high-resolution, low-sidelobe reconstructed images even when radar data is missing, but its imaging quality is limited by the sparse observation model. In addition, the CS algorithm has high computational complexity and long iterative imaging time. With the rapid development of Deep Learning (DL) technology, its excellent feature learning and fitting representation capabilities provide a new technical means to break through the constraints of traditional ISAR imaging methods.

[0003] Since 2019, researchers have been applying deep learning (DL) technology to ISAR imaging, using different network frameworks such as CV-CNN, FCNN, and ResNet for ISAR learning and imaging, achieving better performance than the CS algorithm. These studies demonstrate that DL technology can explore the complex nonlinear mapping relationship between training data and reconstructed images, establishing implicit imaging models and further improving ISAR imaging quality and efficiency under conditions of missing radar data. Due to the powerful image reconstruction capabilities of the U-Net network, researchers have proposed enhanced imaging methods combining the U-Net architecture to improve ISAR imaging quality. Although these DL methods have made significant progress in improving ISAR imaging quality, similar to CS methods, the performance of DL-based ISAR imaging methods is significantly affected by noise and missing data. Furthermore, most of these methods are based on minimizing the mean square error (MSE), which can easily lead to overly smooth reconstructed images and the loss of image details such as weak scattering points. Summary of the Invention

[0004] In order to overcome the shortcomings of the existing technology, the present invention aims to provide an AU-Net-based ISAR enhancement imaging method that can improve ISAR imaging performance under conditions of low signal-to-noise ratio and missing data, and further improve the ability to recover weak scattering points.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] An ISAR enhancement imaging method based on AU-Net includes the following steps;

[0007] Step 1: ISAR echo data is generated based on the random scattering point model simulation, and an ISAR image training set is constructed through random noise addition, downsampling, and imaging preprocessing.

[0008] Step 2: Construct the AU-Net imaging network for ISAR sparse high-resolution imaging by introducing an attention mechanism into the U-Net structure, and initialize the network model parameters.

[0009] Step 3: Design the ISAR imaging loss function based on the constructed AU-Net imaging network;

[0010] Step 4: Develop an ISAR enhanced imaging training strategy for the constructed image training set and imaging loss function, train the model, and update the AU-Net imaging network model parameters to obtain the optimal ISAR high-resolution imaging model.

[0011] Step 5: Obtain simulated / measured target ISAR echo data and perform the same preprocessing operations as in Step 1 to construct an ISAR image test set. ISAR enhanced imaging can be achieved through the obtained optimal ISAR high-resolution imaging model.

[0012] Step 1 specifically includes the following steps;

[0013] (1a) Under radar parameter conditions, 100 to 1000 random scattering points following a Gaussian distribution are randomly generated using Matlab. The scattering coefficient of each scattering point is randomly distributed in the interval (0, 1). The ideal scattering point model of each scene is used as the label sample X in the training set. L ;

[0014] (1b) Suppose the target consists of K scattering points. The ISAR imaging model simulates and generates the corresponding radar echo S for each scene after motion compensation and range compression: (The random scattering point model is the target to be imaged, but it does not have imaging capabilities. The imaging model in this sub-step refers to the simulation of the imaging scene. The relationship between the two is that the random scattering point model needs to be simulated through the imaging model in order to generate the corresponding radar echo.)

[0015]

[0016] Among them, f r and t m These represent the range frequency and azimuth time slow, respectively, and ω represents the target's rotational angular velocity. σ iT represents the scattering coefficient at the i-th scattering point. p γ = B / T p λ = c / f c These represent pulse width, modulation frequency, wavelength, B, and f, respectively. c These represent bandwidth and carrier frequency, respectively. i ,y i () represents the original coordinates of the i-th scattering point on the target;

[0017] (1c) Add random Gaussian white noise with different signal-to-noise ratios to the echo S, and randomly sample data of different proportions in the azimuth and range directions to obtain the echo Y:

[0018] Y = ASD + N (Formula 2)

[0019] Where, Y∈C M′×N′ A∈R M′×M D∈R N×N′ S∈C M×N and N∈C M′×N′ Let N represent the sparsely denoised echo signal matrix, azimuth downsampling matrix, range downsampling matrix, original echo signal matrix, and noise matrix, respectively; N represents the number of range cells, M represents the number of azimuth cells, M′ represents the number of azimuth cells after downsampling, and N′ represents the number of range cells after downsampling.

[0020] (1d) The range-Doppler (RD) algorithm is used to preprocess the sparsely noisy echo Y to obtain the ISAR image X to be reconstructed. This result is used as the input sample of the training set. The overall enhancement imaging process is as follows:

[0021]

[0022] Where, N t Let W represent the batch sample size, W represent the imaging parameters, and X, X' represent the batch sample size. L These represent the ISAR image to be reconstructed and the labeled image, respectively.

[0023] In step 2, the AU-Net network used for ISAR sparse high-resolution imaging includes four parts: ISAR image feature extraction module m1, ISAR image restoration module m2, ISAR attention feature fusion module m3, and ISAR image output layer m4.

[0024] Step 2 specifically includes the following steps;

[0025] (2a)m1 contains four encoders to extract ISAR image features from X;

[0026] The first three encoders each consist of two 3×3 convolutional layers (conv) and a 2×2 max pooling layer (maxpool), and each conv is followed by an activation layer (LeakyReLU);

[0027] The last encoder is basically the same as the first three structures, but it does not contain a max pooling layer;

[0028] (2b) m2 contains four decoders, which are used to map the ISAR image features extracted by m1 to the ideal scattering point model;

[0029] Each decoder consists of one 2×2 upsampling layer (upconv), an ISAR attention feature fusion layer m4, two 3×3 convolutional layers (conv), and an activation layer (LeakyReLU);

[0030] An attention mechanism is introduced into the skip connections of the U-Net structure, forming an ISAR attention feature fusion layer. This layer extracts key features of the ISAR image from the ISAR features output by each encoder through the attention feature fusion module and fuses them with the ISAR features output by the corresponding decoder.

[0031] (2c) m3 is distributed between each encoder and decoder to extract the attention features in m1 and fuse them with the features obtained from m2;

[0032] The attention feature fusion module is a cascade of the ISAR channel attention module and the ISAR spatial attention module, which learns the weights of ISAR feature maps at different stages from both channel and spatial dimensions.

[0033] The ISAR channel attention module uses max pooling and average pooling to convert global ISAR image information into channel information. Then, fully connected layers (FC) learn the channel information and perform information fusion. After a sigmoid activation operation, an ISAR channel attention feature map is generated. This feature map is multiplied by the ISAR input feature map F to generate the input feature F required by the ISAR spatial attention module. c ;

[0034] The ISAR spatial attention module outputs the feature map F from the ISAR channel attention module. cAs the input feature map for this module, max pooling and average pooling operations are performed respectively, and the results are concatenated based on the channel dimension. Then, a 3×3 convolution operation is used to reduce the number of channels to 1, followed by a sigmoid function to generate an ISAR spatial attention feature map. Finally, this feature map is compared with the input feature map F. c Multiplying these together yields the final ISAR attention feature map F. s ;

[0035] (2d)m4 is a 1×1 convolutional layer (conv), which performs channel fusion on the ISAR feature map finally obtained from m2 to output the ISAR high-resolution imaging result;

[0036] F inm2 =r×F outm1 Formula 4

[0037] Where r ~ Bernoulli(p) is used to generate the probability vector, F outm1 F is the ISAR output feature map of module m1. inm2 This is the ISAR input feature map for the m2 module;

[0038] Furthermore, the input feature F c Expressed as:

[0039]

[0040] Where σ is the sigmoid operation and F represents the input feature map. These represent the ISAR channel feature maps after max pooling and average pooling, respectively, with W0 and W1 being the weight parameters of the fully connected layer.

[0041] ISAR attention feature map F s Expressed as:

[0042]

[0043] Among them, f 3×3 This indicates a convolution operation with a 3×3 kernel. F c ISAR spatial feature maps obtained through max pooling and average pooling;

[0044] (2e) Connect the above modules. First, input the ISAR image X to be reconstructed into module m1. When passing through the first three encoders, it needs to be input into the next encoder and the attention fusion layer m3 at the same time. Then, after passing through the fourth encoder and the dropout layer, it is input into module m2. When passing through the first three decoders, it is fused with the results obtained from each of the m3 modules. Finally, after passing through the fourth decoder, it is input into module m4 to obtain the final high-resolution ISAR image X. T Thus, the AU-Net imaging network for ISAR sparse high-resolution imaging was constructed, and the network parameters W were randomly initialized.

[0045] In step 3, the imaging loss function is defined as:

[0046]

[0047] in, The MSE loss function is... For L1 regularization terms, For L2 regularization; X T =WX,X L N represents the reconstructed image and the labeled image, respectively. t This represents the batch size, and W represents the imaging network parameters. Let represent the first norm and the second norm, respectively, and α, β, and γ be the hyperparameters.

[0048] The specific implementation of step 4 is as follows:

[0049] (4a) Input the preprocessed ISAR image X into the AU-Net imaging network, and calculate layer by layer according to the network cascade order to obtain the prediction result X. T =WX;

[0050] (4b) The Adam algorithm is used to update and optimize the network parameters W. The update formula is as follows:

[0051] W new =W-ηΔL Formula 8

[0052] Among them, W new Here, W represents the network parameters obtained after the update, η represents the network parameters before the update, η is the learning rate (set to 0.001), and ΔL is the partial derivative of the loss function with respect to W.

[0053] (4c) Use the updated weight W new Repeat the above calculation process, perform multiple iterations, and save the network parameter W that minimizes the loss during each iteration. best This yields the optimal ISAR high-resolution imaging model.

[0054] The specific implementation of step 5 is as follows:

[0055] (5a) Randomly add noise and randomly downsample the simulated / measured data to obtain the echo data S. test ;

[0056] (5b) Apply the RD algorithm to S test Imaging preprocessing is performed to obtain the ISAR images X to be reconstructed in the test set. test ;

[0057] (5c) X test By inputting the data into the pre-trained optimal imaging model AU-Net, the ISAR enhanced imaging result X can be obtained. T :

[0058] X T =W best X test Formula 9.

[0059] The beneficial effects of this invention are:

[0060] This invention introduces an attention mechanism into the U-Net structure to form an ISAR attention feature fusion layer, which enhances the fusion of key features of ISAR images; it removes the batch normalization (BN) layer to improve the detail representation capability of the ISAR imaging model, while adding a dropout layer to improve the generalization performance of the ISAR imaging model, and using the LeakyReLU activation function to replace the ReLU activation function to improve the imaging accuracy of the ISAR imaging model.

[0061] This invention adds L1 and L2 regularization terms to optimize the MSE loss function, thereby constructing an ISAR imaging combined loss function to train the ISAR sparse high-resolution imaging model, further improving the recovery performance for weak scattering points.

[0062] This invention utilizes the prior conditions of signal noise and data sparsity to enable ISAR enhanced imaging under conditions of low signal-to-noise ratio and missing data, and has low computational complexity. Attached Figure Description

[0063] Figure 1 This is a flowchart of the ISAR enhanced imaging method based on AU-Net according to the present invention.

[0064] Figure 2 This is a schematic diagram of the geometric structure of the ISAR imaging system.

[0065] Figure 3 This is a diagram of the AU-Net imaging network structure constructed in this invention.

[0066] Figure 4 This is a schematic diagram of the dataset samples. Figure 4 (a) is the input sample. Figure 4 (b) is a label sample.

[0067] Figure 5 This is a schematic diagram of the target model used in the simulation experiment.

[0068] Figure 6 This is a schematic diagram showing the imaging results of different algorithms when the signal-to-noise ratio (SNR) is 0dB and the sparsity ratio (SR) is 50% under simulated data experiments.

[0069] Figure 7 This is a schematic diagram of the SNR curves of different algorithms when SR=50% under simulation data experiment.

[0070] Figure 8 This is a schematic diagram of the imaging results of different algorithms under noise-free conditions of SR=25% in the simulation data experiment.

[0071] Figure 9 This is a schematic diagram of the SR curves of different algorithms when SNR=0dB in the simulation data experiment.

[0072] Figure 10 This is a schematic diagram showing the imaging results of different algorithms under experimental data. Detailed Implementation

[0073] The present invention will now be described in further detail with reference to the accompanying drawings.

[0074] The specific implementation steps of this invention are as follows:

[0075] Step 1: ISAR echo data is generated based on the random scattering point model simulation, and an ISAR image training set is constructed through random noise addition, downsampling, and imaging preprocessing.

[0076] Table 1

[0077]

[0078] The specific implementation of this step is as follows:

[0079] (1a) Under the radar parameters given in Table 1, 2000 sets of random scattering points of 100 to 1000 that follow a Gaussian distribution were randomly generated using Matlab. The scattering coefficient of each scattering point was randomly distributed in the interval (0, 1). The ideal scattering point model of each scene was used as the label sample X in the dataset. L ;

[0080] (1b) Assuming the target consists of K scattering points, according to... Figure 2 The ISAR imaging model shown in the simulation generates the corresponding radar echo S for each scene after motion compensation and range compression:

[0081]

[0082] Among them, f r and t m These represent the range frequency and azimuth time slow, respectively, and ω represents the target's rotational angular velocity. σ i T represents the scattering coefficient at the i-th scattering point. p γ = B / T p λ = c / f c These represent pulse width, modulation frequency, wavelength, B, and f, respectively. c These represent bandwidth and carrier frequency, respectively. i ,y i () represents the original coordinates of the i-th scattering point on the target;

[0083] (1c) Add random Gaussian white noise with different signal-to-noise ratios to the echo S, and randomly sample data of different proportions in the azimuth and range directions to obtain the echo Y:

[0084] Y = ASD + N (Formula 2)

[0085] Where Y∈C M′×N′ A∈R M′×M D∈R N×N′ S∈C M×N and N∈C M′×N′ Let N represent the sparsely noise-added echo signal matrix, azimuth downsampling matrix, range downsampling matrix, original echo signal matrix, and noise matrix, respectively. N represents the number of range cells, M represents the number of azimuth cells, M′ represents the number of downsampled azimuth cells, and N′ represents the number of downsampled range cells.

[0086] (1d) The range-Doppler (RD) algorithm is used to preprocess the sparsely noisy echo Y to obtain the ISAR image X to be reconstructed. This result is used as the input sample of the dataset, and the overall enhancement imaging process is transformed into an optimization problem:

[0087]

[0088] Where, N t Let W represent the batch sample size, W represent the imaging parameters, and X, X' represent the batch sample size. L These represent the ISAR image to be reconstructed and the labeled image, respectively. To obtain the optimal imaging parameters W, a reasonable imaging network structure and loss function need to be designed.

[0089] Step 1 is characterized by the entire process of constructing the dataset, including generating a random scattering point model and random noise addition, downsampling, and preprocessing steps.

[0090] Step 2: Construct the AU-Net network for ISAR sparse high-resolution imaging by introducing an attention mechanism into the U-Net structure, and initialize the network model parameters.

[0091] The specific implementation of this step is as follows:

[0092] like Figure 3 As shown, the AU-Net network for ISAR sparse high-resolution imaging of the present invention mainly consists of four parts: ISAR image feature extraction module m1, ISAR image restoration module m2, ISAR attention feature fusion module m3, and ISAR image output layer m4.

[0093] (2a) m1 contains four encoders used to extract ISAR image features from X. The first three encoders each include two 3×3 convolutional layers (conv) and a 2×2 max pooling layer (maxpool), with an activation layer (LeakyReLU) following each conv. The last encoder is basically the same as the first three, but does not include a max pooling layer. In particular, the encoder of this invention omits the BN layer in the U-Net structure and replaces the ReLU activation function with the LeakyReLU function, which can improve the detail representation capability of the ISAR image.

[0094] (2b) m2 contains four decoders to map the ISAR image features extracted by m1 to the ideal scattering point model. Each decoder includes a 2×2 upsampling layer (upconv), an ISAR attention feature fusion layer m4, two 3×3 convolutional layers (conv), and an activation layer (LeakyReLU). In particular, this invention introduces an attention mechanism into the skip connections of the U-Net structure, forming an ISAR attention feature fusion layer. That is, the ISAR features output by each encoder are used to extract key features of the ISAR image through the attention feature fusion module, and then fused with the ISAR features output by the corresponding decoder. This can reduce the loss of weak scattering point information during downsampling and further improve imaging accuracy.

[0095] (2c) m3 is distributed between each encoder and decoder to extract attention features from m1 and fuse them with the features obtained from m2. The attention feature fusion module is a cascade of the ISAR channel attention module and the ISAR spatial attention module. It can learn the weights of ISAR feature maps at different stages from both channel and spatial dimensions, focusing on feature information that is more conducive to refined ISAR imaging. For the ISAR channel attention module, max pooling and average pooling are used to convert global ISAR image information into channel information. Then, the channel information is learned and fused through a fully connected layer (FC). After a sigmoid activation operation, an ISAR channel attention feature map is generated. This feature map is multiplied with the ISAR input feature map F to generate the input feature F required by the ISAR spatial attention module. c The above process can be expressed as follows:

[0096]

[0097] Where σ is the sigmoid operation and F represents the input feature map. W0 and W1 represent the ISAR channel feature maps after max pooling and average pooling, respectively, where W0 and W1 are the weight parameters of the fully connected layer.

[0098] The ISAR spatial attention module outputs the feature map F from the ISAR channel attention module. c As the input feature map for this module, max pooling and average pooling operations are performed respectively, and the results are concatenated based on the channel dimension. Then, a 3×3 convolution operation is used to reduce the number of channels to 1, followed by a sigmoid function to generate an ISAR spatial attention feature map. Finally, this feature map is compared with the input feature map F. c Multiplying these together yields the final ISAR attention feature map F. s The above process can be expressed as:

[0099]

[0100] Among them, f 3×3 This indicates a convolution operation with a 3×3 kernel. F c ISAR spatial feature maps obtained through max pooling and average pooling.

[0101] (2d) m4 is a 1×1 convolutional layer (conv) that performs channel fusion on the ISAR feature map obtained from m2 to output the high-resolution ISAR imaging result. Furthermore, this invention adds a dropout layer after the ISAR image feature extraction module m1 and before the ISAR image restoration module m2 to improve the generalization performance of the ISAR imaging model.

[0102] F inm2 =r×F outm1 Formula 4

[0103] Where r ~ Bernoulli(p) is used to generate the probability vector, F outm1 F is the ISAR output feature map of module m1. inm2 This is the ISAR input feature map for the m2 module.

[0104] (2e) Connect the above modules. First, input the ISAR image X to be reconstructed into module m1. When passing through the first three encoders, it needs to be input into the next encoder and the attention fusion layer m3 at the same time. Then, after passing through the fourth encoder and the dropout layer, it is input into module m2. When passing through the first three decoders, it is fused with the results obtained from each of the m3 modules. Finally, after passing through the fourth decoder, it is input into module m4 to obtain the final high-resolution ISAR image X. T This leads to the construction of, for example Figure 3 The AU-Net imaging network is shown for ISAR sparse high-resolution imaging, and the network parameters W are randomly initialized.

[0105] Step 2 is characterized by the overall construction and connection method of the AU-Net network.

[0106] Step 3: Design the ISAR imaging loss function based on the constructed AU-Net imaging network.

[0107] The imaging loss function designed in this invention is defined as follows:

[0108]

[0109] in, The MSE loss function is... For L1 regularization terms, This is an L2 regularization term. X T =WX,X L N represents the reconstructed image and the labeled image, respectively. t This represents the batch size, and W represents the imaging network parameters. Let L1 and L2 represent the first and second norms, respectively, and α, β, and γ be the hyperparameters, which are empirically set to 1000, 0.005, and 0.1, respectively. The loss function designed in this invention can improve the recovery performance of weak scattering points while enhancing the accuracy of ISAR imaging.

[0110] Step 3 is characterized by the design of the imaging loss function.

[0111] Step 4: Develop an ISAR enhanced imaging training strategy based on the constructed training set and imaging loss function, train the model, and update the AU-Net imaging network model parameters to obtain the optimal ISAR high-resolution imaging model.

[0112] The specific implementation of this step is as follows:

[0113] (4a) Input the preprocessed ISAR image X into the AU-Net imaging network, and calculate layer by layer according to the network cascade order to obtain the prediction result X. T =WX;

[0114] (4b) The Adam algorithm is used to update and optimize the network parameters W. The update formula is as follows:

[0115] W new =W-ηΔL Formula 8

[0116] Among them, W new Here, W represents the network parameters obtained after the update, η represents the network parameters before the update, η is the learning rate (set to 0.001), and ΔL is the partial derivative of the loss function with respect to W.

[0117] (4c) Use the updated weight W new Repeat the above calculation process, perform multiple iterations, and save the network parameter W that minimizes the loss during each iteration. best This yields the optimal ISAR high-resolution imaging model.

[0118] Step 4 is characterized by the training and updating strategy and parameter settings of the imaging model.

[0119] Step 5: Acquire simulated / measured target ISAR echo data and perform the same preprocessing operations as in Step 1 to construct an ISAR test set. ISAR enhanced imaging can then be achieved using the optimal ISAR high-resolution imaging model. The key difference in Step 5 lies in the method of generating the test data, which requires the same preprocessing as in Step 1.

[0120] The specific implementation of this step is as follows:

[0121] (5a) Randomly add noise and randomly downsample the simulated / measured data to obtain the echo data S. test ;

[0122] (5b) Apply the RD algorithm to S test Imaging preprocessing is performed to obtain the ISAR images X to be reconstructed in the test set. test ;

[0123] (5c) X test By inputting the data into the pre-trained optimal imaging model AU-Net, the ISAR enhanced imaging result X can be obtained. T :

[0124] X T =W best X test Formula 9

[0125] The effects of this invention can be further illustrated by the following experiments:

[0126] Simulation data experiment:

[0127] 1. Simulation conditions:

[0128] The simulation was implemented using Matlab software, and all relevant simulation parameters are shown in Table 1. The target model used in the simulation is as follows: Figure 5 As shown, the model is an aircraft model composed of 114 scattering points. The comparison algorithms include: RD, OMP, SBL, FCNN, and U-Net. In addition, to quantitatively evaluate ISAR imaging quality, this invention employs five imaging performance evaluation metrics: Root Mean Square Error (RMSE), Target to Clutter Ratio (TCR), Image Entropy (IE), Image Contrast (IC), and Imaging Time.

[0129] 2. Simulation content:

[0130] Simulation 1:

[0131] In this simulation experiment, 128 pulses were randomly selected from 256 continuous pulses to simulate sparse aperture data. These 128 sparse aperture data were used for imaging, and Gaussian white noise was added to the echo to make the SNR = -5dB. Figure 6 Imaging results from different algorithms are presented for comparison. Figure 6 It can be seen that the AU-Net algorithm of the present invention has the best imaging effect.

[0132] The evaluation index results of imaging results of different algorithms are shown in Table 2.

[0133] Table 2

[0134]

[0135] As shown in Table 2, the AU-Net algorithm of this invention outperforms other algorithms in all five evaluation metrics for reconstructing ISAR images, meaning that the AU-Net algorithm has the best reconstruction performance at low signal-to-noise ratios.

[0136] Simulation 2:

[0137] To further verify the imaging performance of the AU-Net algorithm of this invention under different SNR conditions, Figure 7 The relationship curves of RMSE, TCR, IE, IC, and SNR are presented when SR is 50%. The results show that the AU-Net algorithm has a higher TCR, lower IE, and higher IC, and a smaller RMSE under low signal-to-noise ratio conditions. The experimental results further demonstrate the effectiveness and superiority of the AU-Net algorithm.

[0138] Simulation 3:

[0139] In this simulation experiment, 64 pulses were randomly selected from 256 continuous pulses to simulate sparse aperture data, and these 64 sparse aperture data were used for imaging. Figure 8 Imaging results from different algorithms are presented for comparison. Figure 8 It can be seen that the AU-Net algorithm of the present invention has the best imaging effect.

[0140] The evaluation metrics for imaging results from different algorithms are shown in Table 3.

[0141] Table 3

[0142]

[0143]

[0144] As shown in Table 3, the AU-Net algorithm of this invention outperforms other algorithms in all five evaluation metrics for reconstructing ISAR images, meaning that the AU-Net algorithm has the best reconstruction performance under sparse aperture conditions.

[0145] Simulation 4:

[0146] To further verify the imaging performance of the AU-Net algorithm of this invention under different SR conditions, Figure 9 The relationship curves of RMSE, TCR, IE, IC and SR are presented when SNR is 0dB. The results show that the AU-Net algorithm consistently has a smaller RMSE, higher TCR, lower IE and higher IC under different SR conditions. The experimental results further demonstrate the superiority and effectiveness of the AU-Net algorithm under sparse aperture conditions.

[0147] Experimental data:

[0148] 1. Experimental conditions:

[0149] Table 4

[0150]

[0151] This invention uses measured data from ground-based radar imaging of a Boeing 727 to verify the algorithm's performance, and employs IE, IC, and IT metrics to evaluate the imaging effect. The Boeing 727 data is typical echo data published by VCChen, and the data parameters are shown in Table 4.

[0152] 2. Experiment Content:

[0153] refer to Figure 10 It can be seen that, as Figure 10 As shown in the red highlighted portions of the imaging results, the U-Net network has limited performance in recovering weak scattering points; while FCNN can recover weak scattering points, its image contrast and focus are poor; SBL performs well in recovering weak scattering points, but it also generates many false scattering points and has poor focusing; while the AU-Net algorithm of this invention outperforms the other algorithms in both weak scattering point recovery and image focusing. To quantitatively illustrate the quality of the imaging results of the six algorithms, Table 5 provides the evaluation metrics for each imaging result.

[0154] Table 5

[0155]

[0156] As shown in Table 5, the AU-Net algorithm of this invention has the lowest signal-to-clutter ratio, the lowest image entropy value, the highest image contrast, and the shortest imaging time. This demonstrates the effectiveness and superiority of the AU-Net algorithm on the measured data.

[0157] like Figure 1The figure shows the overall flowchart of the ISAR enhancement imaging method based on AU-Net. This process mainly includes four parts: training set, supervised training, test set, and focused imaging. First, random scattering points and their backscattering coefficients are generated according to a Gaussian random distribution within a specific area. The corresponding sample labels in the training set are composed of these ideal scattering point models. Then, the above scattering points are simulated using a signal model to obtain radar echo signals after motion compensation and range compression. Random noise addition and downsampling are then performed on the obtained radar echoes. After imaging preprocessing, the input samples in the training set are obtained. Next, these input samples are fed into the AU-Net network and sample labels for supervised training to obtain the final imaging model. Finally, the same preprocessing operation is performed on the simulated / measured targets to be reconstructed in the test set, and high-resolution ISAR imaging can be achieved through the trained AU-Net imaging network.

[0158] like Figure 2 As shown: This figure is the ISAR imaging geometric model used in step 1 to simulate radar target echoes. The origin O is the target's centroid, the line-of-sight direction from the radar to the origin is defined as the Y-axis, and the direction perpendicular to the Y-axis on the imaging plane is defined as the X-axis, establishing a Cartesian coordinate system. Since the coherent processing time of ISAR imaging is generally short, it is assumed that the target rotates uniformly around its center at a speed of ω. R0 represents the distance from the origin to the radar, and R... i (t m ) represents t m The distance r from the i-th scattering point to the radar at time i. i θ represents the distance from the i-th scattering point to the origin, and θ0 represents the initial angle.

[0159] like Figure 3 As shown: This figure shows the AU-Net network structure for ISAR sparse high-resolution imaging constructed in step 2 of this invention.

[0160] like Figure 4 As shown: This figure is a sample dataset constructed in step 1 of this invention. Figure 4 (a) is the input sample. Figure 4 (b) is a label sample;

[0161] like Figure 5 As shown: This figure is the aircraft target model constructed in step 5 of the present invention for simulation experiments, which contains 114 simulated scattering points.

Claims

1. An ISAR enhanced imaging method based on AU-Net, characterized in that, Includes the following steps; Step 1: ISAR echo data is generated based on the random scattering point model simulation, and an ISAR image training set is constructed through random noise addition, downsampling, and imaging preprocessing. Step 2: Construct the AU-Net imaging network for ISAR sparse high-resolution imaging by introducing an attention mechanism into the U-Net structure, and initialize the network model parameters. Step 3: Design the ISAR imaging loss function based on the constructed AU-Net imaging network; Step 4: Develop an ISAR enhanced imaging training strategy for the constructed image training set and imaging loss function, train the model, and update the AU-Net imaging network model parameters to obtain the optimal ISAR high-resolution imaging model. Step 5: Obtain simulated / measured target ISAR echo data and perform the same preprocessing operations as in Step 1 to construct an ISAR image test set. ISAR enhanced imaging can be achieved through the obtained optimal ISAR high-resolution imaging model. Step 2 specifically includes the following steps; (2a) m1 contains four encoders used to extract ISAR image features from X; The first three encoders each include two A convolutional layer and a The max pooling layer, with an activation layer following each convolutional layer; The last encoder is basically the same as the first three structures, but it does not contain a max pooling layer; (2b) m2 contains four decoders, which are used to map the ISAR image features extracted by m1 to the ideal scattering point model; Each decoder includes 1 The upsampling layer, the ISAR attention feature fusion layer m4, and 2 Convolutional layers and activation layers; An attention mechanism is introduced into the skip connections of the U-Net structure, forming an ISAR attention feature fusion layer. This layer extracts key features of the ISAR image from the ISAR features output by each encoder through the attention feature fusion module and fuses them with the ISAR features output by the corresponding decoder. (2c) m3 is distributed between each encoder and decoder to extract the attention features in m1 and fuse them with the features obtained from m2; The attention feature fusion module is a cascade of the ISAR channel attention module and the ISAR spatial attention module, which learns the weights of ISAR feature maps at different stages from both channel and spatial dimensions. The ISAR channel attention module uses max pooling and average pooling to convert global ISAR image information into channel information. Then, fully connected layers learn the channel information and perform information fusion. After a sigmoid activation operation, an ISAR channel attention feature map is generated. This feature map is multiplied by the ISAR input feature map F to generate the features required by the ISAR spatial attention module. ; The ISAR spatial attention module will output the features from the ISAR channel attention module. The input feature maps for this module are also subjected to max pooling and average pooling operations, and the results are concatenated based on the channel dimension, then processed through a... The convolution operation reduces the number of channels to 1, then a sigmoid function is used to generate an ISAR spatial attention feature map. Finally, this feature map is compared with the input feature map. Multiplying them together yields the final ISAR attention feature map. ; (2d)m4 is The convolutional layer (conv) performs channel fusion on the ISAR feature map finally obtained from m2 to output the high-resolution ISAR imaging result; Official 4 in, Used to generate probability vectors This is the ISAR output feature map of module m1. This is the ISAR input feature map for the m2 module; (2e) Connect the above modules. First, input the ISAR image X to be reconstructed into module m1. When passing through the first three encoders, it needs to be input into the next encoder and the attention fusion layer m3 at the same time. Then, after passing through the fourth encoder and the dropout layer, it is input into module m2. When passing through the first three decoders, it is fused with the results obtained from each of the m3 modules. Finally, after passing through the fourth decoder, it is input into module m4 to obtain the final high-resolution ISAR image. Thus, the AU-Net imaging network for ISAR sparse high-resolution imaging was constructed, and the network parameters W were randomly initialized. Input features Expressed as: Official 5 in, For sigmoid operation, F represents the input feature map. , These represent the ISAR channel feature maps after max pooling and average pooling, respectively. , These are the weight parameters for the fully connected layer; ISAR attention feature map Expressed as: Official 6 in, Indicates that the convolution kernel is Convolution operation, , They represent ISAR spatial feature maps obtained through max pooling and average pooling; In step 3, the imaging loss function is defined as: Official 7 in, The MSE loss function is... For L1 regularization, For L2 regularization terms; , These represent the reconstructed image and the labeled image, respectively. This represents the batch sample size, and W represents the imaging parameters. , Let them represent the first norm and the second norm, respectively. , , These are hyperparameters.

2. The ISAR enhanced imaging method based on AU-Net according to claim 1, characterized in that, Step 1 specifically includes the following steps; (1a) Under radar parameter conditions, 100 to 1000 random scattering points following a Gaussian distribution are randomly generated using Matlab. The scattering coefficient of each scattering point is randomly distributed in the interval (0, 1). The ideal scattering point model is used as the label sample in the training set. ; (1b) Assume the target consists of K scattering points. The ISAR imaging model simulates and generates the corresponding radar echoes for each scene after motion compensation and range compression. : Official 1 in, and These represent the range frequency and the azimuth time, respectively. Indicates the angular velocity of the target rotation. , Indicates the first i The scattering coefficient at each scattering point , , These represent pulse width, modulation frequency, and wavelength, respectively. and These represent bandwidth and carrier frequency, respectively. Indicates the first target i The original coordinates of the scattering points; (1c) Add random Gaussian white noise with different signal-to-noise ratios to the echo S, and randomly sample data of different proportions in the azimuth and range directions to obtain the echo Y: Official 2 in, , , , and These represent the sparsely noise-added echo signal matrix, azimuth downsampling matrix, range downsampling matrix, original echo signal matrix, and noise matrix, respectively; N represents the number of range elements, M represents the number of azimuth elements, and so on. Indicates the number of azimuth cells after downsampling. This indicates the number of distance cells after downsampling; (1d) The range-Doppler (RD) algorithm is used to preprocess the sparsely noisy echo Y to obtain the ISAR image X to be reconstructed. This result is used as the input sample of the training set. The overall enhancement imaging process is as follows: Official 3 in, This represents the batch size, and W represents the imaging parameters. , These represent the ISAR image to be reconstructed and the labeled image, respectively.

3. The ISAR enhanced imaging method based on AU-Net according to claim 2, characterized in that, In step 2, the AU-Net imaging network used for ISAR sparse high-resolution imaging includes four parts: ISAR image feature extraction module m1, ISAR image restoration module m2, ISAR attention feature fusion module m3, and ISAR image output layer m4.

4. The ISAR enhanced imaging method based on AU-Net according to claim 1, characterized in that, The specific implementation of step 4 is as follows: (4a) Input the preprocessed ISAR image X into the AU-Net imaging network and calculate layer by layer according to the network cascade order to obtain the prediction result. ; (4b) The Adam algorithm is used to update and optimize the network parameters W. The update formula is as follows: Official 8 in, W represents the updated network parameters, and W represents the network parameters before the update. It's the learning rate. It is the partial derivative of the loss function with respect to W, i.e. ; (4c) Use the updated weights Repeat the above calculation process, perform multiple iterations, and save the network parameters that minimize the loss during the iterations. This yields the optimal ISAR high-resolution imaging model.

5. The ISAR enhanced imaging method based on AU-Net according to claim 4, characterized in that, The specific implementation of step 5 is as follows: (5a) Randomly add noise and randomly downsample the simulated / measured data to obtain echo data. ; (5b) Using the RD algorithm to... Imaging preprocessing is performed to obtain the ISAR images to be reconstructed in the test set. ; (5c) will By inputting the data into the pre-trained optimal imaging model AU-Net, the ISAR enhanced imaging results can be obtained. : Official 9.

Images