SAR interference intelligent reconstruction method and system based on complex signal domain coding and decoding network

Abstract

The application discloses a SAR interference intelligent reconstruction method and system based on a complex signal domain coding and decoding network. The method comprises the following steps: S1, constructing a SAR interference intelligent reconstruction model based on a complex signal domain coding and decoding network, which comprises an input preprocessing module, a coding feature extraction module, an improved connection module, a decoding feature fusion module and an image restoration module; S2, constructing a data set for training the model; S3, training the SAR interference intelligent reconstruction model based on the complex signal domain coding and decoding network by using the training set, and saving the weight parameters after the training; S4, loading the pre-stored weight parameters to obtain the SAR interference intelligent reconstruction model based on the complex signal domain coding and decoding network after the training; and S5, inputting a SAR echo signal to the SAR interference intelligent reconstruction model based on the complex signal domain coding and decoding network after the training for interference reconstruction. The application realizes the extraction, reconstruction and cancellation of the interference signal in the SAR signal.

Description

Technical Field

[0001] The present invention relates to the technical field of computer neural networks and radar signal processing, and more specifically, to a SAR interference intelligent reconstruction method and system based on a complex signal domain encoding and decoding network. Background Art

[0002] Synthetic Aperture Radar (SAR) is a widely used microwave remote sensing device that actively radiates electromagnetic waves to produce high-resolution images of the target area. As electromagnetic environments become increasingly complex, SAR is susceptible to various coherent interference signals during operation, obscuring the true target signal in the SAR image and severely degrading the image quality. Interference suppression technology, as a radar defense measure, reduces the impact of interference signals on imaging, facilitating subsequent image interpretation and other tasks.

[0003] In the field of optical image processing, deep neural networks can adaptively extract deep features from training data to perform classification or regression tasks. SAR (Radio Frequency Amplification) radar images are obtained by actively transmitting microwaves, receiving their reflected signals, and then compressing and phase-correcting them. These images contain rich electromagnetic scattering information. Therefore, SAR images are a unique type of remote sensing data. Their imaging principles, data characteristics, and information presentation differ significantly from those of optical images. This makes neural network models or methods that excel in optical image processing inappropriate for SAR image processing. Therefore, when a SAR system is subject to coherent forwarding interference, performing neural network processing directly on radar images does not conform to the relevant principles of SAR signal processing. Instead, processing should be performed on the pre-compressed echoes.

[0004] In summary, it is necessary to design a corresponding neural network model or optimize its structure according to the echo characteristics of SAR to realize the reconstruction of interference echoes and lay the foundation for further interference suppression. Summary of the Invention

[0005] The purpose of the present invention is to provide a SAR interference intelligent reconstruction method and system based on a complex signal domain coding and decoding network to overcome the defects of the prior art.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] The SAR interference intelligent reconstruction method based on the complex signal domain encoding and decoding network includes the following steps:

[0008] S1. Construct a SAR interference intelligent reconstruction model based on a complex signal domain encoding and decoding network. The model includes an input preprocessing module, a coding feature extraction module, an improved connection module, a decoding feature fusion module, and an image restoration module.

[0009] S2. Constructing a data set for training the SAR interference intelligent reconstruction model based on the complex signal domain encoding and decoding network;

[0010] S3, using the training set to train the coding feature extraction module, the improved connection module, and the decoding feature fusion module of the SAR interference intelligent reconstruction model based on the complex signal domain coding and decoding network, and saving the weight parameters of the trained coding feature extraction module, the improved connection module, and the decoding feature fusion module;

[0011] S4. Load the pre-stored weight parameters to obtain the trained SAR interference intelligent reconstruction model based on the complex signal domain encoding and decoding network;

[0012] S5. Input the SAR echo signal to the trained SAR interference intelligent reconstruction model based on the complex signal domain encoding and decoding network to perform interference reconstruction.

[0013] Furthermore, in step S1:

[0014] An input preprocessing module is used to obtain an input data set, divide the data set into a training set and a test set, randomly scatter and slice the training set and the test set, extract amplitude and phase information from the randomly scatter and sliced ​​training set and the test set, and perform normalization and size adjustment;

[0015] The encoding feature extraction module is used to downsample the input image using convolution and pooling operations to extract the feature matrix;

[0016] An improved connection module, configured to perform two convolution operations on the feature matrix using a convolution kernel and then connect the convolution operation to a Dropout layer;

[0017] A decoding feature fusion module is used to upsample the extracted feature matrix to the input resolution for signal sampling point level decomposition, and perform improved concatenation and fusion of high-resolution features of different scales from the encoder;

[0018] The image restoration module is used to recompress the echo information into an image and calculate the corresponding parameters to obtain the final result.

[0019] Furthermore, the step S3 specifically includes:

[0020] Step 1: Use the Relu function as the activation function in training. The Relu function has the following formula, where x is the input sample data value:

[0021]

[0022] Step 2: Perform forward propagation on the sample data to obtain the predicted value and calculate the model loss. The Huber function is used as the loss function to calculate the loss. The formula is:

[0023]

[0024] Where y and f(x) are the true value and predicted value of the sample respectively, and δ is a hyperparameter;

[0025] Step 3: Use the optimizer to update the weight parameters of the model;

[0026] Step 4: Perform training iterations according to the training process from step 1 to step 3, and store the final weight parameters of each module.

[0027] The present invention also provides a system for implementing the above-mentioned SAR interference intelligent reconstruction method based on the complex signal domain encoding and decoding network, comprising:

[0028] The first building module is used to build a SAR interference intelligent reconstruction model based on a complex signal domain encoding and decoding network. The model includes an input preprocessing module, a coding feature extraction module, an improved connection module, a decoding feature fusion module, and an image restoration module;

[0029] A second construction module is used to construct a data set for training the SAR interference intelligent reconstruction model based on the complex signal domain encoding and decoding network;

[0030] A training module, configured to train the coding feature extraction module, the improved connection module, and the decoding feature fusion module of the SAR interference intelligent reconstruction model based on the complex signal domain coding and decoding network using the training set, and save the weight parameters of the trained coding feature extraction module, the improved connection module, and the decoding feature fusion module;

[0031] A loading module is used to load pre-stored weight parameters to obtain a trained SAR interference intelligent reconstruction model based on a complex signal domain encoding and decoding network;

[0032] The input module is used to input the SAR echo signal to the trained SAR interference intelligent reconstruction model based on the complex signal domain encoding and decoding network for interference reconstruction.

[0033] Compared with the prior art, the present application has the advantages that: the present application inputs the uncompressed SAR complex echo signal, uses the coding and decoding network to combine the low-resolution information after multiple down-sampling and the high-resolution information directly transmitted from the coding layer to the same height decoding layer through the improved splicing operation, realizes the interference reconstruction, and finally compresses the reconstructed interference signal into an image by using the echo signal to cancel out, realizing the extraction, reconstruction and cancellation of the interference signal in the SAR signal. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0035] Figure 1 The overall flowchart of the present application;

[0036] Figure 2 The logical structure diagram of the system constructed in step S1 of the present application;

[0037] Figure 3 The partial image of the MiniSAR data set used in the present application (only the processed SAR gray scale image is shown)

[0038] Figure 4 The partial image of the FuCheng data set used in the present application (only the processed SAR gray scale image is shown)

[0039] Figure 5 The partial image of the COSMO data set used in the present application (only the processed SAR gray scale image is shown)

[0040] Figure 6 The typical sample of the SAR echo signal after fixed frequency modulation and retransmission used in the second step of the present application for training the system (only the processed SAR gray scale image is shown)

[0041] Figure 7 The typical sample of the SAR echo signal after distance direction intermittent sampling modulation and retransmission used in the second step of the present application for training the system (only the processed SAR gray scale image is shown)

[0042] Figure 8 The typical sample of the SAR echo signal after azimuth direction intermittent sampling modulation and retransmission used in the second step of the present application for training the system (only the processed SAR gray scale image is shown)

[0043] Figure 9This is a typical sample of the SAR echo signal after two-dimensional intermittent sampling modulation and forwarding used for training the system in the second step of the present invention (only the processed SAR grayscale image is shown);

[0044] Figure 10 This is a typical sample of the SAR echo signal after frequency micro-modulation forwarding used for training the system in the second step of the present invention (only the processed SAR grayscale image is shown);

[0045] Figure 11 This is a typical sample of the SAR echo signal after intermittent sampling + frequency micro-modulation forwarding used for training the system in the second step of the present invention (only the processed SAR grayscale image is shown);

[0046] Figure 12 This is a typical sample of the SAR echo signal after intermittent sampling + random frequency shift modulation forwarding used for training the system in the second step of the present invention (only the processed SAR grayscale image is shown);

[0047] Figure 13 This is a typical sample of the SAR echo signal after azimuth intermittent sampling + stepped frequency shift modulation forwarding for the training system in the second step of the present invention (only the processed SAR grayscale image is shown);

[0048] Figure 14 This is a typical sample of the SAR echo signal after forwarding the micro-frequency + step-shift frequency modulation used for training the system in the second step of the present invention (only the processed SAR grayscale image is shown);

[0049] Figure 15 This is the structure diagram of the encoding and decoding network;

[0050] Figure 16 Typical test results of the complex signal domain encoding and decoding neural network proposed in this invention;

[0051] Figure 17 Typical test results for interference reconstruction and image restoration;

[0052] Figure 18 This is a system framework diagram of the present invention. DETAILED DESCRIPTION

[0053] The preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more precise definition of the protection scope of the present invention.

[0054] See Figure 1 As shown, this embodiment discloses a SAR interference intelligent reconstruction method based on a complex signal domain encoding and decoding network, comprising the following steps:

[0055] Step S1: construct a SAR interference intelligent reconstruction model based on a complex signal domain encoding and decoding network, which includes an input preprocessing module, a coding feature extraction module, an improved connection module, a decoding feature fusion module and an image restoration module. Figure 2 shown.

[0056] In this embodiment, the input preprocessing module is used to obtain an input dataset, randomly shuffle the samples to enhance the model's generalization and robustness, and divide the dataset into training and test sets. Furthermore, this module enhances the model's generalization and robustness by randomly slicing the samples. It then extracts the sample's amplitude and phase information and performs normalization, resizing, and other operations to facilitate network processing.

[0057] The encoding feature extraction module is also called the compression path or downsampling path. After the SAR complex echo data to be processed of size 1928×220×2 is fed into the network, two convolution layers of size 3×3×64 are used, followed by a maximum pooling operation using a convolution kernel of size 2×2. This process is defined as the first stage. The subsequent network structure then repeats the first stage operation process four times. During the repetition process, the number of convolution kernels begins to change after each pooling operation, and the number increases exponentially, from the initial 64 to 512. When the first stage is repeated for the fourth time, a Dropout layer is inserted between the convolution layer and the pooling layer. This process is the encoding stage of the network. In general, the input image is downsampled through a series of convolution and pooling operations to extract the feature matrix.

[0058] The improved connection module is located between the encoder and the decoder, and performs two convolution operations using a convolution kernel of size 3×3×1024, followed by a Dropout layer. Since the feature matrices of the paths on both sides of the encoding and decoding are of different scales, the improvement is mainly reflected in the normalization of the feature matrix of the encoding path by clipping it to the feature matrix of the same size as the decoding path, that is, for the adjustment of the size of the SAR complex echo signal, the specific approach is to add a resize layer in front of each node of the decoding module. The role of this module is to further extract and fuse features to provide more useful information for the decoder, especially for the problem that deep neural networks may "forget" certain features when passing information through consecutive layers. The improved connection can reintroduce these features to make learning more powerful. This idea was introduced in the residual network (ResNet), and it also shows the advantages of improved classification and smoother learning gradients. Inspired by this mechanism, the improved connection is added to the encoding and decoding network of the present invention so that each decoder contains the feature matrix of its corresponding encoder.

[0059] The decoding feature fusion module is also called the expansion path or upsampling path, which is symmetrical with the encoder part in data size. The extracted global depth features are upsampled to the input resolution for signal sampling point level decomposition, and the high-resolution features of different scales from the encoder are fused with improved connections to alleviate the spatial information loss caused by downsampling. After each upsampling step, the feature matrix of the corresponding level of the encoder is spliced ​​with the feature matrix of the current level of the decoder through improved connections. After the above encoding operation is completed, a feature matrix of 120×13×1024 is obtained, and then the expansion path operation is started. This part consists of 4 modules. Before each module starts, the size of the feature matrix is ​​doubled by a 2×2×1024 transposed convolution layer, and the number of convolution kernels is halved. Then, the feature matrix of the corresponding layer of the decoding path is merged through the improved connection, and then two 3×3 convolution layers are used. The resulting feature matrix size returns to 1928×220×2. This process is defined as the second stage in this paper. The entire decoding path needs to repeat the second stage four times. Similar to the first stage, the number of convolution kernels begins to change after each splicing operation, and its number decreases exponentially, from the initial 512 to 64. After the second stage is completed, two convolution kernels of size 1*1 are required to perform convolution operations, and the entire encoding and decoding network is completed.

[0060] The image restoration module recompresses the echo information and images it, calculates the corresponding parameters, and obtains the final result.

[0061] Step S2: constructing a data set for training the SAR interference intelligent reconstruction model based on the complex signal domain encoding and decoding network.

[0062] In this example, we constructed a coherent forwarding jamming signal model and selected nine modulation forwarding jamming patterns for the countermeasure scenario. Compared to conventional approaches, we innovatively considered five single modulation methods and also selected four complex modulation methods, which are more challenging to identify. The specific principles of each modulation method are shown in Table 1.

[0063] Table 1 Nine types of modulation and forwarding interference echo signal forms

[0064]

[0065] Among them, t r For fast time, t a is the slow time, A0 is the echo amplitude, ω r (t r ) represents the range antenna pattern function, ω a (t a) represents the range antenna pattern function, R(t r ,t a ) represents the distance from the SAR to the target, c represents the speed of light, rect(·) represents the rectangular signal, t ac is the reference time of slow time, f c is the carrier frequency of the SAR signal, K r is the range modulation frequency of the linear frequency modulation signal transmitted by SAR. For frequency shift interference, f d 、f sh is the frequency shift amount; for intermittent sampling forwarding interference, g(t) is the intermittent sampling pulse signal, its sampling time is τ and the repetition period is T s '; For micro-modulation interference, A m is the modulation amplitude, f m is the modulation frequency, is the initial phase of the modulation signal.

[0066] According to the above principles of this embodiment, the MiniSAR dataset, FuCheng dataset and COSMO dataset are used as input, and the typical scenarios are as follows: Figure 3-5 As shown. Through computer simulation, the three data sets mentioned above without the implementation of coherent modulation forwarding technology are used to generate coherent modulation forwarding signals to form interference data sets composed of coherent modulation forwarding signals. The typical samples are shown as follows: Figure 6-Figure 14 As shown in Figure 1. Furthermore, to enhance the generalization capability of the network, five modulation and forwarding positions are set for each coherent modulation and forwarding technology, with nine interference states and four interference-to-signal ratios at each point. The parameters of the modulation and forwarding technology are randomly generated. The interference-to-signal ratio is defined as the ratio of the modulation power to the original echo signal power, as shown in Formula (1).

[0067]

[0068] Where, P J Indicates the modulation power, P S Indicates the original echo signal power.

[0069] Step S3: Use the training set to train the coding feature extraction module, improved connection module, and decoding feature fusion module of the SAR interference intelligent reconstruction model based on the complex signal domain coding and decoding network, and save the weight parameters of the trained coding feature extraction module, improved connection module, and decoding feature fusion module. The network structure is as follows: Figure 15 shown.

[0070] In this embodiment, during the training process, the Relu function is used as the activation function, and the Relu function is shown in formula (2). Wherein, x is the input sample data value.

[0071]

[0072] The sample data is then forward propagated to obtain the predicted value, and the model loss is calculated. The loss function is used to measure the degree of inconsistency between the model's predicted value f(x) and the true value y. One loss function is the mean square error (MSE), which is the average of the squared distances between the model's predicted value f(x) and the sample's true value y, as shown in Formula (3). The root mean square error (RMSE) is defined as shown in Formula (4).

[0073]

[0074] Where y i and f(x i ) represent the true value and predicted value of the i-th sample, respectively, and m is the number of samples. If there are outliers in the sample, MSE will give higher weight to the outliers, but will sacrifice the prediction effect of other normal data points, which will reduce the overall performance of the model. The second loss function is the mean absolute error (MAE), which refers to the average distance between the model prediction value f(x) and the true value y of the sample, as shown in formula (5):

[0075]

[0076] Where y i and f(x i ) represent the true value and predicted value of the i-th sample, respectively, and m is the number of samples. The MAE curve is V-shaped and continuous, but it is not differentiable at yf(x) = 0, making it difficult to calculate the derivative using a computer. Furthermore, the MAE gradient is constant in most cases, meaning that even for small loss values, the gradient is large. This is detrimental to function convergence and model learning.

[0077] Based on this, this embodiment uses the Huber function as the loss function to calculate the loss, as shown in formula (6).

[0078]

[0079] HuberLoss is a combination of MSE and MAE, including a hyperparameter δ. The value of δ determines HuberLoss's emphasis on MSE and MAE. When |yf(x)| ≤ δ, it becomes MSE; when |yf(x)| > δ, it becomes similar to MAE. Therefore, HuberLoss combines the advantages of both MSE and MAE, reduces sensitivity to outliers, and achieves universal differentiation.

[0080] Then, the optimizer is used to update the system's weight parameters. The RMSProp optimizer is used to address the Adagrad algorithm's issues by introducing a decay coefficient, allowing historical information to decay exponentially, thus avoiding the problem of a continuous decrease in the learning rate. The exponentially weighted moving average of the squared gradient is used to apply a different learning rate to each parameter, adaptively adjusting it based on the parameter's most recent gradient, making the learning rate adjustment smoother. This is shown in Formula (7).

[0081]

[0082] Among them, g is the gradient of the weight parameter, θ is the weight parameter, r is the cumulative square gradient, ρ is the decay rate, ⊙ represents the element product, ε = 10 -8 is the global learning rate, avoiding division by 0.

[0083] Finally, the training iteration is performed according to the above training process, and the final weight parameters of each module are stored.

[0084] Step S4: load the pre-stored weight parameters to obtain a trained SAR interference intelligent reconstruction model based on the complex signal domain encoding and decoding network.

[0085] Step S5: Input the SAR echo signal to the trained SAR interference intelligent reconstruction model based on the complex signal domain codec network for interference reconstruction, perform inferential restoration on the interfered image, and then complete the recompression imaging of the echo signal in the image recovery module, and evaluate the imaging quality.

[0086] In this embodiment, the image quality is evaluated by comparing the visual error between the de-scrambled image and the true image. In addition to the MSE, the peak signal-to-noise ratio (PSNR) is also used. This is an engineering term that represents the ratio of the maximum possible signal power to the destructive noise power that affects its representation accuracy. It is usually expressed in logarithmic decibels. For a grayscale image, assume that the original image and the de-scrambled image, both of size m×n, are I0 and I1 respectively. p , the mean square error MSE is defined as shown in formula (8).

[0087]

[0088] PSNR is calculated by MSE and is defined as shown in formula (9).

[0089]

[0090] Where, MAX I To represent the maximum value of the image pixel value, the closer the denoised image is to the original image, the smaller the MSE is, and the larger the PSNR is.

[0091] The similarity between the two images is then measured again using SSIM. SSIM evaluates image quality by comparing brightness, contrast, and structure, making it particularly suitable for evaluating the quality of the interference suppression image in this invention. SSIM uses two images: one is the original image x, and the other is the image y after interference removal. SSIM is defined as shown in Equation (10). Here, α, β, and γ are greater than 0.

[0092] SSIM(x,y)=l(x,y) α c(x,y) β s(s,y) γ (10)

[0093] The brightness contrast function is defined as

[0094]

[0095] In formula (11), μ x is the average grayscale, used for brightness measurement estimation, and C1 is a selected constant.

[0096] The contrast function is defined as:

[0097]

[0098] In formula (12), σ x is the standard deviation, defined as Used for contrast estimation, C2 is another selected constant.

[0099]

[0100] The structural contrast function is defined as:

[0101]

[0102] In formula (14), C3 is another selected constant.

[0103] In summary, in order to obtain a simplified form, let α=β=γ=1, C3=C2 / 2, and we can get

[0104]

[0105] Complete the final result.

[0106] The trained codec network demonstrated an average SSIM of 81.1% between the reconstructed interference and the ideal interference, with an average RMSE of only 11.32 and an average PSNR of 20.45. Compared to scenarios without interference reconstruction and cancellation, the similarity between the corresponding scene and the interference-free scene was improved to 85.5%, with an average RMSE of only 4.43 and an average PSNR of 28.85.

[0107] The present invention has been verified through simulation tests to have good interference extraction, reconstruction and cancellation effects for coherent modulation forwarding interference. Its typical training curve is as follows Figure 16 As shown in , the final Loss is 0.02. Figure 17 As shown in the figure, the interference repair effect of SAR complex echo using RMSProp optimizer and Huber loss function is very good.

[0108] See Figure 18 As shown, the present invention also provides a system for implementing the above-mentioned SAR interference intelligent reconstruction method based on the complex signal domain codec network, comprising: a first construction module 1, for constructing a SAR interference intelligent reconstruction model based on the complex signal domain codec network, the model comprising an input preprocessing module, a coding feature extraction module, an improved connection module, a decoding feature fusion module and an image restoration module; a second construction module 2, for constructing a data set for training the SAR interference intelligent reconstruction model based on the complex signal domain codec network; a training module 3, for using the training set to train the coding feature extraction module, the improved connection module and the decoding feature fusion module of the SAR interference intelligent reconstruction model based on the complex signal domain codec network, and save the weight parameters of the trained coding feature extraction module, the improved connection module and the decoding feature fusion module; a loading module 4, for loading pre-stored weight parameters to obtain a trained SAR interference intelligent reconstruction model based on the complex signal domain codec network; and an input module 5, for inputting a SAR echo signal into the trained SAR interference intelligent reconstruction model based on the complex signal domain codec network for interference reconstruction.

[0109] Therefore, it can be seen that the complex signal domain encoding and decoding network can complete the extraction, reconstruction and cancellation functions of SAR composite modulation forwarding interference with higher SSIM and PSNR, and can retain the essential characteristics of the SAR signal and reduce the number of overfitting, verifying that the present invention has a good effect.

[0110] Although the embodiments of the present invention are described in conjunction with the accompanying drawings, the patent owner may make various changes or modifications within the scope of the appended claims. As long as they do not exceed the scope of protection described in the claims of the present invention, they should be within the scope of protection of the present invention.

Claims

1. A SAR interference intelligent reconstruction method based on a complex signal domain encoding and decoding network, characterized in that: The following steps are involved: S1. Construct a SAR interference intelligent reconstruction model based on a complex signal domain encoding and decoding network. The model includes an input preprocessing module, a coding feature extraction module, an improved connection module, a decoding feature fusion module, and an image restoration module. In the step S1: An input preprocessing module is used to obtain an input data set, divide the data set into a training set and a test set, randomly scatter and slice the training set and the test set, extract amplitude and phase information from the randomly scatter and sliced ​​training set and the test set, and perform normalization and size adjustment; The encoding feature extraction module is used to downsample the input image using convolution and pooling operations to extract the feature matrix; An improved connection module, configured to perform two convolution operations on the feature matrix using a convolution kernel and then connect the convolution operation to a Dropout layer; A decoding feature fusion module is used to upsample the extracted feature matrix to the input resolution for signal sampling point level decomposition, and perform improved concatenation and fusion of high-resolution features of different scales from the encoder; Image restoration module, used to re-compress the echo information and image, and calculate the corresponding parameters to obtain the final result; S2. Constructing a data set for training the SAR interference intelligent reconstruction model based on the complex signal domain encoding and decoding network; S3, using the training set to train the coding feature extraction module, the improved connection module, and the decoding feature fusion module of the SAR interference intelligent reconstruction model based on the complex signal domain coding and decoding network, and saving the weight parameters of the trained coding feature extraction module, the improved connection module, and the decoding feature fusion module; S4. Load the pre-stored weight parameters to obtain the trained SAR interference intelligent reconstruction model based on the complex signal domain encoding and decoding network; S5. Inputting the SAR echo signal into the trained SAR interference intelligent reconstruction model based on the complex signal domain encoding and decoding network to perform interference reconstruction; The improved connection module is located between the encoder and decoder, performs two convolution operations using a convolution kernel of size 3×3×1024, and is then connected to a dropout layer. The feature matrix of the encoding path is normalized by cropping it to the same size as the feature matrix of the decoding path. Specifically, this adjusts the size of the SAR complex echo signal by adding a resize layer before each node of the decoding module.

2. The SAR interference intelligent reconstruction method based on complex signal domain coding and decoding network according to claim 1 is characterized in that: The specific steps in step S3 are: Step 1: Use the Relu function as the activation function in training. The Relu function has the following formula, where x is the input sample data value: Step 2: Perform forward propagation on the sample data to obtain the predicted value and calculate the model loss. The Huber function is used as the loss function to calculate the loss. The formula is: Where y and f(x) are the true value and predicted value of the sample respectively, and δ is a hyperparameter; Step 3: Use the optimizer to update the weight parameters of the model; Step 4: Perform training iterations according to the training process from step 1 to step 3, and store the final weight parameters of each module.

3. A system for implementing the SAR interference intelligent reconstruction method based on a complex signal domain coding and decoding network according to any one of claims 1 to 2, characterized in that: include: The first building module is used to build a SAR interference intelligent reconstruction model based on a complex signal domain encoding and decoding network. The model includes an input preprocessing module, a coding feature extraction module, an improved connection module, a decoding feature fusion module, and an image restoration module; A second construction module is used to construct a data set for training the SAR interference intelligent reconstruction model based on the complex signal domain encoding and decoding network; A training module, configured to train the coding feature extraction module, the improved connection module, and the decoding feature fusion module of the SAR interference intelligent reconstruction model based on the complex signal domain coding and decoding network using the training set, and save the weight parameters of the trained coding feature extraction module, the improved connection module, and the decoding feature fusion module; A loading module is used to load pre-stored weight parameters to obtain a trained SAR interference intelligent reconstruction model based on a complex signal domain encoding and decoding network; The input module is used to input the SAR echo signal to the trained SAR interference intelligent reconstruction model based on the complex signal domain encoding and decoding network for interference reconstruction.

Images