A method for SAR interference suppression of image intelligent decomposition

Abstract

The application discloses a SAR interference suppression method for image intelligent decomposition, and mainly solves the problem of poor SAR imaging interference suppression effect under a complex electromagnetic environment in the prior art.The implementation scheme is as follows: measured data is acquired to generate interference signals; an imaging algorithm is used to image the measured data, the interference signals and mixed data composed of the interference and the measured data, respectively, and a training set and a test set are divided; an image decomposition interference suppression network including an image decomposition network and an interference suppression network is constructed, and the network is iteratively trained by using the training set through back propagation; and the test set is input into the trained image decomposition interference suppression network to acquire an image after interference suppression.The application can fully suppress interference in a specific scene, improves the utilization rate of image data and the ability of suppressing interference in the specific scene, and can be used for suppressing various interference caused by radar imaging under a complex electromagnetic environment in target detection.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of radar signal processing, and particularly relates to a synthetic aperture radar (SAR) interference suppression method, which can be used for effectively suppressing various interferences to radar imaging in a complex electromagnetic environment in target detection. BACKGROUND

[0002] With the rapid development of electronic technology and communication technology, radars have been widely used in many fields. However, radars are extremely susceptible to various interferences when working in a complex electromagnetic environment, which seriously affects the accuracy of radar target detection. In modern warfare and civil fields, the electromagnetic environment is becoming increasingly complex, and various intentional or unintentional interference sources are emerging, which poses a serious threat to the normal operation of radars. For example, in some electromagnetic-intensive areas such as city centers and industrial clusters, electromagnetic signals generated by civil devices can also interfere with radars. In recent years, in order to solve the problem of radar interference, researchers have proposed numerous methods and achieved some results.

[0003] The patent document with the patent application number CN202410783996.1 proposes a "multi-band radar signal analysis and processing method", which uses a multi-band radar receiving antenna to simultaneously collect the returned multi-band radar signals; uses a variational frequency band decomposition model to decompose the multi-band radar signals into several sub-bands; synchronously squeezes wavelet transform for each sub-band; after nonlinear Fourier transform of each synchronous squeezing wavelet transform result, the synchronous adaptive entropy is calculated, and then the synchronous adaptive entropy is subjected to multi-scale adaptive phase adjustment to obtain the multi-scale adaptive phase adjustment result; based on all the multi-scale adaptive phase adjustment results, signal reconstruction is performed to extract the target signal in the multi-band radar signal. Although this method can effectively process multi-band signals and improve the accuracy of time-frequency representation, it cannot effectively separate the interference signals and radar signals when the interference signal frequency band is extremely close to or even coincides with the radar signal frequency band, so that the interference signals are still mixed in the radar signal frequency band and participate in the subsequent processing, resulting in poor interference suppression effect for some interference signals with frequencies close to the target signal.

[0004] The patent document with the patent application number CN202110056256.4 proposes a "radar slice storage and forwarding type interference suppression method based on optimized deep model", and the implementation scheme is: obtaining radar target echo signal and radar slice storage type interference signal, and preprocessing them to obtain training set and verification set; constructing a deep U-shaped network model realizing radar interference suppression function, and adjusting the hyperparameters of the deep U-shaped network model; training the U-shaped network model using the training set; using the optimized deep model to suppress interference on the radar target echo signal and the radar slice storage type interference signal, and completing the radar slice storage and forwarding type interference suppression based on deep model optimization. This method can only realize the interference suppression of the radar slice storage and forwarding type interference, and has limited adaptability when dealing with complex and variable interference scenes. SUMMARY

[0005] The present application aims to overcome the shortcomings of the prior art and provide a SAR interference suppression method for image intelligent decomposition to effectively deal with complex interference situations, improve the accuracy of radar target detection imaging, and reduce the time and effort spent on interference suppression.

[0006] The technical key to achieving the above-mentioned purpose is: by constructing and training an image decomposition interference suppression network, the interference image is decomposed into real measurement data image and pure interference image, the interference is directly suppressed at the imaging level, avoiding complex suppression processing of interference signals in the frequency domain and adjustment of specific interference signal types and frequency band parameters, and using GPU for parallel processing of interference images to improve the efficiency of interference suppression and reduce the time and effort spent on interference suppression.

[0007] According to the above idea, the implementation steps of the present application include the following:

[0008] (1) Create training set and test sample set:

[0009] Obtain synthetic aperture radar real measurement data from a public website, generate interference signals according to the real measurement data, signal-to-interference ratio and interference type; add interference signals to the real measurement data to obtain mixed data;

[0010] Image the real measurement data, interference signals and mixed data respectively, crop them into sub-images of a specified size, and then divide these sub-images into training set and test set according to a 7:3 ratio;

[0011] (2) Construct an image decomposition interference suppression network:

[0012] 2a) Construct an image decomposition network including an image feature extraction module and a feature decomposition module, which is used to decompose the interfered SAR image into real measurement data image and interference image; the image feature extraction module is connected in parallel with two feature decomposition modules;

[0013] 2b) constructing an interference suppression network comprising an encoding link, a decoding link and a feature fusion module;

[0014] The encoding link is sequentially connected with an image patch linear embedding module, a first feature extraction module, a first down-sampling module, a second feature extraction module and a second down-sampling module;

[0015] The decoding link is sequentially connected with a first up-sampling module, a third feature extraction module, a second up-sampling module, a fourth feature extraction module and an image merging module;

[0016] The feature fusion module is connected between the first down-sampling module and the third feature extraction module;

[0017] 2c) connecting two output ports of the image decomposition network with two interference suppression networks respectively to form an image decomposition interference suppression network, and taking a mean square error function as a loss function Loss of the network MSE ;

[0018] (3) inputting a training set into the image decomposition interference suppression network, and iteratively training the same by using data driving to obtain a trained image decomposition interference suppression network, and taking an Adam algorithm as an optimization algorithm;

[0019] (4) inputting a test set into the trained image decomposition interference suppression network to obtain an image decomposition interference suppression result.

[0020] Compared with the prior art, the present application has the following advantages:

[0021] Firstly, the present application suppresses interference at the image level, so it does not require prior knowledge of specific interference signal bandwidth, frequency, etc., and has stronger robustness and flexibility when facing complex and variable interference environments, and can work more stably;

[0022] Secondly, the present application adds interference signals to the original echo signals and performs imaging to generate SAR images affected by multiple types and multiple interference levels, so it can perform sufficient interference suppression training for specific scenes and improve the interference suppression ability for specific scenes;

[0023] Thirdly, the present application constructs an image decomposition network comprising an image decomposition module and an image feature extraction module, which can decompose the interfered SAR image into a measured data image and an interference image, and improve the utilization of the measured data image and the interference image in a data-driven manner, thereby improving the interference suppression ability.

[0024] Fourthly, the present application builds an interference suppression network including an encoding link, a feature fusion module and a decoding link, splits and embeds a feature image into a feature image block matrix through an image splitting linear embedding module, introduces a multi-head self-attention mechanism through a feature extraction module, extracts image features in layers through a down-sampling module and an up-sampling module, and fuses multi-layer image features through a feature fusion module, which not only improves the network's interference suppression capability, but also enhances the network's image feature extraction capability. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is the implementation flowchart of the present application;

[0026] Figure 2 is the image decomposition interference suppression network structure schematic diagram constructed in the present application;

[0027] Figure 3 is the interference suppression network structure schematic diagram in Figure 2

[0028] Figure 4 is the flowchart of iterative training of the image decomposition interference suppression network in the present application;

[0029] Figure 5 is the interference suppression effect diagram of the present experiment. DETAILED DESCRIPTION

[0030] In order to make the personnel in the technical field better understand the present application scheme, the technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, other embodiments obtained by the person skilled in the art without creative labor should belong to the protection scope of the present application.

[0031] It should be noted that the step labels in the present application specification and claims are only for clearly describing the embodiments of the present application, for the convenience of understanding, and the sequence of the serial numbers is not limited.

[0032] Referring to Figure 1 , the implementation steps of the present example include the following:

[0033] Step 1, set a training set and a test sample set.

[0034] 1.1) Obtain the measured data of ESR2 satellite from a public website, extract the carrier frequency f c and bandwidth B of the radar signal from the measured data, and convert the measured data into a two-dimensional complex matrix DATA with a size of Na*Nr;

[0035] ​1.2) According to the carrier frequency f of the radar signal c and the bandwidth B, a filter with the same passband as the radar signal is designed;

[0036] 1.3) A complex Gaussian white noise signal with size Na*Nr is generated using simulation software, and the generated Gaussian white noise is filtered using the designed filter to obtain a noise signal NGWN;

[0037] 1.4) Set the signal-to-interference ratio JSR, calculate the average power P signal of the measured data and the average power P jammer of the noise signal;

[0038] 1.5) According to the results of step 1.4), the amplitude coefficient J of the interference signal is calculated,

[0039]

[0040] 1.6) According to the amplitude coefficient J of the interference signal and the interference type, the following four different interference signals are generated:

[0041] Noise product interference: J1 = J * NGWN;

[0042] Noise amplitude modulation interference: J2 = J * abs(NGWN)

[0043] Noise frequency modulation interference: J3 = J * cos(2πf c *t + 2π * NGWN)

[0044] Noise phase modulation interference: J4 = J * imag(NGWN)

[0045] abs(NGWN) is the amplitude matrix of each point of the noise signal, f c is the carrier frequency of the radar signal, t is the time of the radar signal, and imag(NGWN) is the imaginary part of the noise signal;

[0046] 1.7) The measured data matrix DATA is respectively added to the four kinds of interference signals to obtain product interference mixed data D1, amplitude modulation interference mixed data D2, frequency modulation interference mixed data D3 and phase modulation interference mixed data D4:

[0047] D1 = DATA + J1

[0048] D2 = DATA + J2

[0049] D3 = DATA + J3

[0050] D4 = DATA + J4;

[0051] 1.8) The range-Doppler algorithm imaging is performed on the measured data, interference signals and mixed data respectively, and the measured data image, interference signal image and mixed data image are each cropped to a sub-image with a specified size of 256*256;

[0052] 1.9) The sub-images are divided into a training set and a test set according to a ratio of 7:3.

[0053] Step 2: Constructing an image decomposition interference suppression network.

[0054] Reference Figure 2 and Figure 3 The implementation of this step is as follows:

[0055] 2.1) Constructing an image decomposition network:

[0056] 2.1.1) Establishing an image feature extraction module composed of a first convolutional layer, a second convolutional layer, a third convolutional layer and a nonlinear activation layer connected in sequence, wherein: the first convolutional layer has 3 convolutional kernels, each with a size of 3*3, a step of 1*1 and a padding of 1; the second layer has 8 convolutional kernels, each with a size of 3*3, a step of 1*1 and a padding of 1; the third layer has 16 convolutional kernels, each with a size of 3*3, a step of 1*1 and a padding of 1; and the nonlinear activation layer is implemented by a ReLU function;

[0057] 2.1.2) Establishing a first image feature decomposition module composed of a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer and a nonlinear activation layer connected in sequence, wherein: the fourth convolutional layer has 3 convolutional kernels, each with a size of 3*3, a step of 1*1 and a padding of 1; the fifth layer has 8 convolutional kernels, each with a size of 3*3, a step of 1*1 and a padding of 1; the sixth layer has 16 convolutional kernels, each with a size of 3*3, a step of 1*1 and a padding of 1; and the nonlinear activation layer is implemented by a ReLU function;

[0058] 2.1.3) Establishing a second image feature decomposition module with the same structure and parameters as the first image feature decomposition module;

[0059] 2.1.4) Connecting the image feature extraction module in parallel with the first image feature decomposition module and the second image feature decomposition module to form an image decomposition network;

[0060] 2.2) Constructing an encoding link:

[0061] 2.2.1) Establish an image patch linear embedding module comprising an image segmentation window and an image patch embedding layer, wherein the image segmentation window is a sliding window with a size of 16*16 and a moving step of 16, which is used to segment the image into 16*16 image patches; the image patch embedding layer constructs an image patch matrix according to the position of the image patch in the image before segmentation;

[0062] 2.2.2) Establish a first feature extraction module and a second feature extraction module, both of which comprise a convolution layer and a multi-head self-attention layer, the convolution layer is used to realize feature extraction of the feature image, and the multi-head self-attention layer is used to realize extraction of the feature dependency relationship, wherein the convolution layer and the multi-head self-attention layer are existing;

[0063] 2.2.3) Establish a first down-sampling module composed of a down-sampling first convolution layer, a down-sampling second convolution layer, a down-sampling first nonlinear activation function layer and a down-sampling first pooling layer connected in sequence, wherein: the down-sampling first convolution layer has 8 convolution kernels with a size of 3*3 and a step of 1*1 and a padding of 1; the down-sampling second convolution layer has 16 convolution kernels with a size of 3*3 and a step of 1*1 and a padding of 1; the down-sampling first nonlinear activation function layer is realized by a ReLU function; and the down-sampling second pooling layer is realized by a MaxPool function;

[0064] 2.2.4) Establish a second down-sampling module composed of a down-sampling third convolution layer, a down-sampling fourth convolution layer, a down-sampling second nonlinear activation function layer and a down-sampling second pooling layer connected in sequence, wherein: the second down-sampling module is used to reduce the size of the image and increase its dimension; the down-sampling third convolution layer has 32 convolution kernels with a size of 3*3 and a step of 1*1 and a padding of 1; the down-sampling fourth convolution layer has 64 convolution kernels with a size of 3*3 and a step of 1*1 and a padding of 1; the down-sampling second nonlinear activation function layer is realized by a ReLU function; and the down-sampling second pooling layer is realized by a MaxPool function;

[0065] 2.2.5) Connect the image patch linear embedding module, the first feature extraction module, the first down-sampling module, the second feature extraction module and the second down-sampling module in sequence to form an encoding link;

[0066] 2.3) Construct a decoding link:

[0067] 2.3.1) Establish a first up-sampling module composed of a first up-sampling transposed convolution layer, a second up-sampling transposed convolution layer and a first up-sampling nonlinear activation function layer connected in turn, wherein: the first up-sampling transposed convolution layer has 64 convolution kernels, the size of the transposed convolution kernel is 3*3, the step is 1*1, and the padding is 1; the second up-sampling transposed convolution layer has 32 convolution kernels, the size of the transposed convolution kernel is 3*3, the step is 1*1, and the padding is 1; the first up-sampling nonlinear activation function layer is realized by a ReLU function;

[0068] 2.3.2) Establish a second up-sampling module composed of a third up-sampling transposed convolution layer, a fourth up-sampling transposed convolution layer and a second up-sampling nonlinear activation function layer connected in turn, wherein: the third up-sampling transposed convolution layer has 8 convolution kernels, the size of the convolution kernel is 3*3, the step is 1*1, and the padding is 1; the fourth up-sampling transposed convolution layer has 3 convolution kernels, the size of the convolution kernel is 3*3, the step is 1*1, and the padding is 1; the second up-sampling nonlinear activation function layer is realized by a ReLU function;

[0069] 2.3.3) Establish a third feature extraction module and a fourth feature extraction module, both of which include a convolution layer and a multi-head self-attention layer, the convolution layer is used to realize feature extraction of the feature image, and the multi-head self-attention layer is used to realize extraction of the feature dependency relationship, wherein the convolution layer and the multi-head self-attention layer are existing;

[0070] 2.3.4) Establish an image merging module realized by a mapping layer, which is used to merge the image patches into a complete image according to the positions of the image patches in the image patch matrix;

[0071] 2.3.5) Connect the first up-sampling module, the third feature extraction module, the second up-sampling module, the fourth feature extraction module and the image merging module in turn to form a decoding link;

[0072] 2.4) Construct a feature fusion module for connecting the first down-sampling module and the third feature extraction module, which fuses the feature images processed by the first down-sampling module and the feature images processed by the third feature extraction module through a fusion layer, and performs tensor addition on the corresponding channel feature images;

[0073] 2.5) Connect the encoding link and the decoding link, and connect the feature fusion module between the first down-sampling module of the encoding link and the third feature extraction module of the decoding link to form a first interference suppression network;

[0074] 2.6) Establish a second interference suppression network with the same structure and parameters as the first interference suppression network;

[0075] 2.7) the image decomposition network is connected in parallel with the first interference suppression network and the second interference suppression network respectively to form an image decomposition interference suppression network.

[0076] Step 3: iteratively train the image decomposition interference suppression network model.

[0077] Reference Figure 4 This step realizes the following:

[0078] 3.1) set the network parameters, the learning rate is 0.0001, the batch size is 16, the optimization algorithm is Adam, the loss function is MSE, the number of training rounds is 100, and the preset weights of the interference loss value are: {β1, β2,... β i ...β n}, i = 1, 2...n, each β i value corresponds to the number of rounds;

[0079] 3.2) input the mixed data image in the training set into the image decomposition interference suppression network model, and obtain a preliminary feature image through the image feature extraction module;

[0080] 3.3) the preliminary feature image is respectively input into the first image decomposition module and the second image decomposition module to obtain a decomposed measured data image and a decomposed interference signal image;

[0081] 3.4) the decomposed measured data image is input into the first interference suppression network to obtain a predicted measured data image:

[0082] 3.4.1) the decomposed measured data image is processed through the image block linear embedding module to obtain a decomposed measured data image block matrix;

[0083] 3.4.2) the first feature extraction module extracts features from the decomposed measured data image block matrix to obtain a first feature extraction measured image block matrix;

[0084] 3.4.3) the first downsampling module performs downsampling processing on each image block in the first feature extraction measured image block matrix to obtain a first downsampling measured image block matrix;

[0085] 3.4.4) the second feature extraction module extracts features from the first downsampling feature measured image block matrix to obtain a second feature extraction measured image block matrix;

[0086] 3.4.5) the second downsampling module performs downsampling processing on each image block in the second feature extraction measured image block matrix to obtain a second downsampling measured image block matrix;

[0087] 3.4.6) The first up-sampling module performs up-sampling processing on each image patch in the second down-sampled measured image patch matrix to obtain a first up-sampled measured image patch matrix;

[0088] 3.4.7) The third feature extraction module performs feature extraction on the first up-sampled measured image patch matrix to obtain a third feature-extracted measured image patch matrix;

[0089] 3.4.8) The feature fusion module performs splicing fusion on the image patches at corresponding positions in the first down-sampled measured image patch matrix and the third feature-extracted measured image patch matrix to obtain a measured feature-fused image patch matrix;

[0090] 3.4.9) The second up-sampling module performs processing on each image patch in the feature-fused image patch matrix to obtain a second up-sampled measured image patch matrix;

[0091] 3.4.10) The fourth feature extraction module performs feature extraction on the second up-sampled measured image patch matrix to obtain a fourth feature-extracted measured image patch matrix;

[0092] 3.4.11) The image merging module performs merging processing on the fourth feature-extracted measured image patch matrix to obtain a predicted measured data image;

[0093] 3.5) The decomposed interference data image is subjected to a second interference suppression network to obtain a predicted interference data image;

[0094] 3.5.1) The decomposed interference data image is subjected to an image patch linear embedding module to obtain a decomposed interference data image patch matrix;

[0095] 3.5.2) The first feature extraction module performs feature extraction on the decomposed interference data image patch matrix to obtain a first feature-extracted interference image patch matrix;

[0096] 3.5.3) The first down-sampling module performs down-sampling processing on each image patch in the first feature-extracted interference image patch matrix to obtain a first down-sampled interference image patch matrix;

[0097] 3.5.4) The second feature extraction module performs feature extraction on the first down-sampled feature interference image patch matrix to obtain a second feature-extracted interference image patch matrix;

[0098] 3.5.5) The second down-sampling module performs down-sampling processing on each image patch in the second feature-extracted interference image patch matrix to obtain a second down-sampled interference image patch matrix;

[0099] 3.5.6) The first up-sampling module performs up-sampling processing on each image patch in the second down-sampled interference image patch matrix to obtain a first up-sampled interference image patch matrix;

[0100] 3.5.7) The third feature extraction module performs feature extraction on the first up-sampled interference image patch matrix to obtain a third feature extracted interference image patch matrix;

[0101] 3.5.8) The feature fusion module splices and fuses the image patches at corresponding positions in the first down-sampled interference image patch matrix and the third feature extracted interference image patch matrix to obtain an interference feature fused image patch matrix;

[0102] 3.5.9) The second up-sampling module processes each image patch in the interference feature fused image patch matrix to obtain a second up-sampled interference image patch matrix;

[0103] 3.5.10) The fourth feature extraction module performs feature extraction on the second up-sampled interference image patch matrix to obtain a fourth feature extracted interference image patch matrix;

[0104] 3.5.11) The image merging module performs merging processing on the fourth feature extracted interference image patch matrix to obtain a predicted interference data image;

[0105] 3.6) The loss value between the predicted measured data image output by the network model and the measured data image in the training set is calculated as the measured loss value loss1, and the loss value between the predicted interference signal image output by the network model and the interference signal image in the training set is calculated as the interference loss value loss2; the measured loss value and the interference loss value are weighted to obtain the total loss value loss = loss1 + β i ·loss2;

[0106] 3.7) The total loss value loss is back-propagated to optimize the network parameters;

[0107] 3.8) Repeat steps 3.2) to 3.7) for a total of 100 rounds;

[0108] 3.9) Update the interference loss value weight to β i+1 , repeat steps 3.2) to 3.8) until all preset weight values {β1, β2,... β i ...β n} are traversed to obtain a trained network model.

[0109] Step 4 obtains the image decomposition interference suppression result.

[0110] 4.1) input the mixed data image in the test set into the trained image decomposition interference suppression network model, which obtains a preliminary feature image through an image feature extraction module;

[0111] 4.2) the preliminary feature image is processed through an image decomposition module to obtain a decomposed measured data image;

[0112] 4.3) the decomposed measured data image is processed through an image block linear embedding module to obtain a decomposed measured data image block matrix;

[0113] 4.4) a first feature extraction module extracts features from the decomposed measured data image block matrix to obtain a first feature extraction image block matrix;

[0114] 4.5) a first downsampling module performs downsampling processing on each image block in the first feature extraction image block matrix to obtain a first downsampling feature image block matrix;

[0115] 4.6) a second feature extraction module extracts features from the first downsampling feature image block matrix to obtain a second feature extraction image block matrix;

[0116] 4.7) a second downsampling module performs downsampling processing on each image block in the second feature extraction image block matrix to obtain a second downsampling feature image block matrix;

[0117] 4.8) a first upsampling module performs upsampling processing on each image block in the second downsampling feature image block matrix to obtain a first downsampling feature image block matrix;

[0118] 4.9) a third feature extraction module extracts features from the first upsampling feature image block matrix to obtain a third feature extraction image block matrix;

[0119] 4.10) a feature fusion module splices and fuses image blocks at corresponding positions in the first downsampling feature image block matrix and the third feature extraction image block matrix to obtain a feature fusion image block matrix;

[0120] 4.11) a second upsampling module processes each image block in the feature fusion image block matrix to obtain a second upsampling feature image block matrix;

[0121] 4.12) a fourth feature extraction module extracts features from the second upsampling feature image block matrix to obtain a fourth feature extraction image block matrix;

[0122] 4.13) an image merging module merges the fourth feature extraction image block matrix to obtain a predicted measured data image, which is the interference suppressed image output result.

[0123] The effects of the present application can be further illustrated by the following simulation experiment.

[0124] I. Simulation conditions

[0125] ERS2 satellite data is downloaded from the ASF Data Serch website;

[0126] In this experiment, MATLAB R2022a software platform is used to extract data, generate interference signals, add interference signals and image the measured data, such as Figure 5 (a);

[0127] The network model is established and trained using the PyTorch framework.

[0128] II. Simulation results

[0129] Under the above conditions, the SAR image containing noise amplitude modulation interference as shown in Figure 5 (a) is suppressed by the present application, and the results are as shown in Figure 5 (b).

[0130] As can be seen from Figure 5 , before interference suppression, the SAR image can be observed to have obvious interference components, and after interference suppression, the interference components in the SAR image can be observed to be greatly suppressed, obtaining a clear SAR image.

Claims

1. A method for suppressing SAR interference through intelligent image decomposition, characterized in that, Includes the following steps: (1) Create training and test sample sets: Synthetic Aperture Radar (SAR) measured data is obtained from publicly available websites. Interference signals are generated based on the measured data, signal-to-interference ratio (SIR), and interference type. Interference signals are then added to the measured data to obtain mixed data. The measured data, interference signals, and mixed data are imaged separately, cropped into sub-images of a specified size, and then these sub-images are divided into training set and test set in a 7:3 ratio; (2) Construct an image decomposition interference suppression network: 2a) Construct an image decomposition network including an image feature extraction module and a feature decomposition module to decompose the interfered SAR image into a measured data image and an interfering image; the image feature extraction module is connected to the two feature decomposition modules in parallel; 2b) Construct an interference suppression network that includes an encoding link, a decoding link, and a feature fusion module; The encoding link consists of a sequentially connected image block linear embedding module, a first feature extraction module, a first downsampling module, a second feature extraction module, and a second downsampling module; The decoding link consists of a first upsampling module, a third feature extraction module, a second upsampling module, a fourth feature extraction module, and an image merging module connected in sequence. This feature fusion module is connected between the first downsampling module and the third feature extraction module; 2c) Connect the two output ports of the image decomposition network to two interference suppression networks respectively to form an image decomposition interference suppression network, and use the mean square error function as the loss function of this network. MSE ; (3) Input the training set into the image decomposition interference suppression network, and use data-driven iterative training to obtain the trained image decomposition interference suppression network. The Adam algorithm is used as the optimization algorithm. (4) Input the test set into the trained image decomposition interference suppression network to obtain the image decomposition interference suppression results.

2. The method according to claim 1, characterized in that, Step (1) Generates an interference signal based on measured data, signal-to-interference ratio, and interference type. The steps include the following: 1a) Use the program provided on the website to extract the carrier frequency f of the radar signal. c And bandwidth B, and convert the measured data into a two-dimensional complex matrix DATA of size Na*Nr; 1b) Design a filter with the same passband as the radar signal frequency band based on the carrier frequency and bandwidth of the radar signal; 1c) Use simulation software to generate a complex Gaussian white noise signal of size Na*Nr, and use a pre-designed filter to filter the generated Gaussian white noise to obtain the noise signal NGWN; 1d) Calculate the average power P of the measured data. signal and the average power P of the noise signal jammer The amplitude coefficient J of the interference signal is calculated based on the set signal-to-interference ratio JSR. 1e) Based on the amplitude coefficient J and interference type of the interference signal, four types of interference signals are generated: noise product interference J1, noise amplitude modulation interference J2, noise frequency modulation interference J3, and noise phase modulation interference J4. J1=J*NGWN J2 = J * abs(NGWN) J3=J*cos(2πf c *t+2π*NGWN) J4 = J * imag(NGWN) Its abs(NGWN) is the amplitude matrix of the noise signal at each point, f c t is the carrier frequency of the radar signal, imag(NGWN) is the time of the radar signal, and imag(NGWN) is the imaginary part of the noise signal.

3. The method according to claim 1, characterized in that, Step (1) involves imaging the measured data, interference signal, and mixed data separately. This is achieved by using a range Doppler imaging algorithm to perform range pulse compression, range migration correction, and azimuth pulse compression on the measured data, interference signal, and mixed data respectively to obtain the measured data image, interference signal image, and mixed data image.

4. The method according to claim 1, characterized in that, The two modules that constitute the image decomposition network in step 2a) have the following structure and function: The image feature extraction module, used to extract image features, consists of a first convolutional layer, a second convolutional layer, a third convolutional layer, and a nonlinear activation layer connected sequentially. Specifically: the first convolutional layer has 3 kernels, each 3x3 in size, with a stride of 1x1 and padding of 1; the second convolutional layer has 8 kernels, each 3x3 in size, with a stride of 1x1 and padding of 1; the third convolutional layer has 16 kernels, each 3x3 in size, with a stride of 1x1 and padding of 1; the nonlinear activation layer is implemented using the ReLU function. The image feature decomposition module is used to decompose an image. It consists of a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, and a nonlinear activation layer connected in sequence. Specifically: the fourth convolutional layer has 3 convolutional kernels, each with a kernel size of 3*3, a stride of 1*1, and padding of 1; the fifth convolutional layer has 8 convolutional kernels, each with a kernel size of 3*3, a stride of 1*1, and padding of 1; the sixth convolutional layer has 16 convolutional kernels, each with a kernel size of 3*3, a stride of 1*1, and padding of 1; the nonlinear activation layer is implemented using the ReLU function.

5. The method according to claim 1, characterized in that, The structure of each module constituting the encoding link in step 2b) is as follows: The image block linear embedding module includes: an image segmentation window and image block embedding, wherein: the image segmentation window is a sliding window with a size of 16*16 and a movement step of 16, used to segment the image into 16*16 image blocks; the image block embedding is to embed the segmented image blocks into an image block matrix. Both the first feature extraction module and the second feature extraction module include a convolutional layer and a multi-head self-attention layer. The convolutional layer is used to extract features from the feature image, and the multi-head self-attention layer is used to extract feature dependencies. The first downsampling module consists of a downsampling first convolutional layer, a downsampling second convolutional layer, a downsampling first nonlinear activation function layer, and a downsampling first pooling layer connected sequentially, wherein: The first convolutional layer of this downsampling layer has 8 convolutional kernels, a kernel size of 3*3, a stride of 1*1, and padding of 1. The second convolutional layer of this downsampling layer has 16 convolutional kernels, a kernel size of 3*3, a stride of 1*1, and padding of 1. The first nonlinear activation function layer of this downsampling is implemented using the ReLU function; The second pooling layer for downsampling is implemented using the MaxPool function; The second downsampling module consists of a downsampling third convolutional layer, a downsampling fourth convolutional layer, a downsampling second nonlinear activation function layer, and a downsampling second pooling layer connected sequentially, wherein: This second downsampling module is used to reduce the size of the image and increase its dimensionality; The third convolutional layer of this downsampling layer has 32 convolutional kernels, a kernel size of 3*3, a stride of 1*1, and padding of 1. The fourth convolutional layer of this downsampling layer has 64 convolutional kernels, a kernel size of 3*3, a stride of 1*1, and padding of 1. The second nonlinear activation function layer of this downsampling is implemented using the ReLU function; The second pooling layer for downsampling is implemented using the MaxPool function.

6. The method according to claim 1, characterized in that, The structure and function of each module in the decoding link in step 2b) are as follows: Both the third and fourth feature extraction modules include a convolutional layer and a multi-head self-attention layer. The convolutional layer extracts features from the feature image, and the multi-head self-attention layer extracts feature dependencies. Both the first upsampling module and the second upsampling module include a transposed convolutional layer and a nonlinear activation function layer; The first upsampling module consists of an upsampling first transposed convolutional layer, an upsampling second transposed convolutional layer, and an upsampling first nonlinear activation function layer connected in sequence. The upsampling first transposed convolutional layer has 64 convolutional kernels, a kernel size of 3*3, a stride of 1*1, and padding of 1. The upsampling second transposed convolutional layer has 32 convolutional kernels, a kernel size of 3*3, a stride of 1*1, and padding of 1. The upsampling first nonlinear activation function layer is implemented using the ReLU function. The second upsampling module consists of an upsampling third transposed convolutional layer, an upsampling fourth transposed convolutional layer, and an upsampling second nonlinear activation function layer connected in sequence. The upsampling third transposed convolutional layer has 8 convolutional kernels, a kernel size of 3*3, a stride of 1*1, and padding of 1. The upsampling fourth transposed convolutional layer has 3 convolutional kernels, a kernel size of 3*3, a stride of 1*1, and padding of 1. The upsampling second nonlinear activation function layer is implemented using the ReLU function. The image merging module is used to merge the image block matrix into a complete image.

7. The method according to claim 1, characterized in that, In step 2b), the feature fusion module achieves feature fusion by splicing features at the same level in the network, which is used to fuse shallow and deep features in the network.

8. The method according to claim 1, characterized in that, Step 2c) The image decomposition interference suppression network loss function Loss MSE , means as follows: Where (i,j) is the position of the pixel, I target (i,j) represents the pixel value at position (i,j) in the measured data image, I generated (i,j) represents the pixel value at position (i,j) in the predicted measured data image, where m and n are the total number of pixels in the row and column, respectively.

9. The method according to claim 1, characterized in that, In step (3), the image decomposition interference suppression network is iteratively trained using data-driven methods. The steps include the following: 3a) Assume the network training learning rate is 0.0001, the batch size is 16, the optimization algorithm is Adam, and the interference loss value is preset with weights β as {β1, β2, ... β}. n Each β value is trained for 100 epochs. 3b) The input mixed data image is processed by the network model to obtain the predicted measured data image and the predicted interference signal image. The loss value between the predicted measured data image and the real measured data image is calculated as the measured loss value loss1, and the loss value between the predicted interference signal image and the real interference signal image is calculated as the interference loss value loss2. The total loss value loss = loss1 + β·loss2 is obtained by weighting the measured loss value and the interference loss value. 3c) The backpropagation total loss value is used to optimize the network parameters; 3d)β i After training is complete, update the weights of the perturbation loss values ​​to β. i+1 Repeat steps 3b) and 3c) until all preset weight values ​​{β1, β2, ... β} have been traversed. n }, thus obtaining the trained network model.

10. The method according to claim 1, characterized in that, In step (4), the test set is input into the trained image decomposition interference suppression network to obtain the image decomposition interference suppression result. The steps include the following: 4a) The mixed images in the test set are decomposed into feature images of the measured data by an image decomposition network; 4b) The feature images of the decomposed measured data are processed by the interference suppression network to obtain the final output of the network model.

Images