Electromagnetic inverse scattering imaging method based on SERU-Net model

Abstract

The present invention relates to an electromagnetic inverse scattering imaging method based on a SERU-Net model. The method obtains scattering field information based on antennas arranged around an imaging area to obtain an initial image of an actual target; then, based on the MNIST handwritten digital picture set, a virtual target in the imaging area is generated to obtain an initial image of a virtual target; then, the virtual target is used as a SERU-Net deep learning network label value, and the initial image of the virtual target is used as a SERU-Net input value to construct an electromagnetic imaging prediction model based on the SERU-Net model; finally, the initial image of the actual target is input into the electromagnetic imaging prediction model to obtain an actual target prediction imaging result. The present invention not only handles the imaging problem under the condition of low signal-to-noise ratio in a short time and efficiently, but also significantly improves the problem that the classical network has poor imaging effect on weak scatterers and may have residual images on strong scatterers, thereby greatly improving the imaging accuracy.

Description

Technical Field

[0001] The invention relates to the field of electromagnetic computing, and in particular to an electromagnetic inverse scattering imaging method. Background Art

[0002] Nonlinear electromagnetic inverse scattering imaging is a technology that uses the complex nonlinear relationship between electromagnetic waves and matter to reconstruct images. This technology first arranges multiple antennas around the target area. When the electromagnetic waves emitted by the transmitting antenna shine on the target area, the scatterers in the target area will generate a scattering field. The receiving antenna will capture the scattering field data and establish a functional relationship between the scattering field and the medium distribution in the target area. Then, the inverse scattering algorithm is used to calculate the dielectric constant, conductivity and other parameters in the target area, and then these parameters are used to reconstruct the image of the scatterer to show its position, shape and material properties. Finally, through iterative optimization or deep learning technology, the imaging accuracy and resolution are improved.

[0003] Compared with traditional tomography theory, nonlinear inverse scattering imaging can not only provide higher resolution in more complex media, but also is more sensitive to the nonlinear characteristics and multiple scattering effects inside objects, and can present details that are difficult to capture with traditional imaging technology. Therefore, it has a very broad application scenario, such as: early tumor detection and cancer diagnosis; non-destructive detection of defects such as tiny cracks and pores in composite materials, etc.

[0004] However, the nonlinearity of the electromagnetic inverse scattering problem makes it difficult to analyze the relationship between the scattered signal and the scatterer. The pathological nature means that the solution of the equation is very sensitive to small changes in the observed data. A large number of high-dimensional numerical calculations and iterative optimizations require powerful computing power and efficient algorithms to process. These factors combined make electromagnetic inverse scattering imaging face huge challenges in practical applications, and more effective algorithms need to be developed to improve imaging accuracy and efficiency.

[0005] The key to quantitative analysis of nonlinear inverse scattering imaging problems lies in solving the uniqueness, stability and nonlinearity of the inverse problem. The traditional algorithms for solving these problems are divided into iterative and non-iterative methods. Classic non-iterative inversion methods include Born approximation (BA), Litov approximation (RA), extended Born approximation (EBA), back propagation (BP), etc. Iterative inversion methods include Born iteration method (BIM), subspace optimization algorithm (SOM), contrast source inversion (CSI), etc.

[0006] The Born approximation (BA) is applicable when the dielectric constant of the scatterer is close to that of the background medium and the size of the scatterer is of the same order of magnitude as the wavelength or much smaller than the wavelength. The secondary effect of the scattered field is ignored, only the primary scattering is considered, and the total field is approximately equal to the incident field.

[0007] The Litov approximation (RA) is applicable to the case where the dielectric constant of the scatterer is close to that of the background medium. It takes into account the influence of the scattered field to a certain extent. By making specific transformations to the integral equation, it further linearizes the inverse scattering problem and is applicable to slightly stronger scattering than the Born approximation (BA).

[0008] The extended Born approximation (EBA) is a further improvement on the Born approximation (BA). It modifies the first-order approximation by including secondary or even higher-order scattering effects, making it perform better when dealing with stronger scattering. The back propagation method (BP) assumes that the induced current is linearly related to the scattered field, and derives the distribution of the total field and the relative dielectric constant of the scatterer.

[0009] These non-iterative methods have a common weakness, which is that they are only applicable when the dielectric constant of the scatterer is similar to that of the background medium (the scatterer is weaker). Since high-order scattering effects are ignored, non-iterative methods have low accuracy when inverting complex structures and cannot accurately reconstruct the details and high-contrast areas of the object. They are also sensitive to noise and measurement errors and can be easily affected by inaccurate data, resulting in unstable inversion results.

[0010] The Born iteration method (BIM) is based on the Born approximation (BA). In the initial step, the higher-order scattering effects are ignored and only one scattering is considered. Then, based on the scattering field obtained by the initial estimate, the relative dielectric constant distribution of the scatterer is gradually corrected through repeated iterations, considering more scattering effects and expanding the scope of application of the Born approximation (BA). However, for strong scatterers or obvious multiple scattering effects, the accuracy and convergence are still poor.

[0011] The subspace optimization algorithm (SOM) gradually approaches the optimal solution by optimizing in a low-dimensional subspace. It first performs a rough search on a global scale to determine a rough solution, and then accurately optimizes in the low-dimensional subspace, reducing the amount of calculation and improving the convergence speed. This method is suitable for high-dimensional and complex scattering problems, especially when prior information is limited, but it relies heavily on initial estimates and may fall into local optimal solutions.

[0012] The contrast source inversion (CSI) method transforms complex nonlinear problems into linear inversion of contrast sources, thereby continuously updating the contrast sources and dielectric constant distribution in iterations. This method has strong robustness and can handle strong scattering and multiple scattering effects. The inversion results are stable and reliable.

[0013] It can be seen that the above iterative methods have a wider range of application scenarios and significantly improve imaging accuracy compared to non-iterative methods, but they all consume more computing resources and time. Summary of the invention

[0014] The purpose of the present invention is to avoid the shortcomings of the prior art and provide an electromagnetic inverse scattering imaging method based on the SERU-Net model, which can eliminate the "artifact" phenomenon that occurs in imaging and effectively improve the imaging accuracy of weak scatterers. The SERU-Net model proposed in the present invention is designed based on the U-Net model; SE represents the added attention mechanism (Squeeze-and-Excitation block); R represents the added residual structure.

[0015] To achieve the above object, the technical solution adopted by the present invention is: an electromagnetic inverse scattering imaging method based on the SERU-Net model, comprising the following steps:

[0016] Step 1: Use N i transmit antennas and N r The actual target scattering field data of the imaging area is obtained by using N receiving antennas. The actual target scattering field data is r ×N i The matrix is ​​then used to obtain the actual target initial image using the back propagation algorithm, where N r ≥N i And N i 、N r Take a positive integer;

[0017] Step 2: Generate a virtual target in the imaging area based on the MNIST handwritten digit image set, and calculate the scattering field of the virtual target based on the moment method to obtain N r The scattered field data at the location of the receiving antennas is composed of N dimensions. r ×N i The virtual target scattering field data matrix is ​​obtained, and then the initial image of the virtual target is obtained by using the back propagation algorithm;

[0018] Step 3: Based on the virtual scattering field data matrix, the virtual target is used as the label value of the SERU-Net model, and the initial image of the virtual target is used as the input value of the SERU-Net model to construct an electromagnetic imaging prediction model based on the SERU-Net model;

[0019] Step 4: input the actual target initial image into the electromagnetic imaging prediction model to obtain the actual target prediction imaging result.

[0020] Furthermore, the step 1 comprises:

[0021] With the center point of the imaging area as the center, N pixels are evenly arranged on a circle with a radius of R. i transmit antennas and N r There are N receiving antennas, i The transmitting antennas transmit transverse magnetic waves in turn, and the actual target scattering field data are all captured by N r The data collected by the receiving antennas is N r ×N i The scattering field data matrix is ​​obtained, and the initial image of the actual target with M×M pixels is obtained using the back propagation algorithm, where M≥16.

[0022] Furthermore, the step 2 comprises:

[0023] First, at least 400 MNIST handwritten digit images are randomly selected from the MNIST handwritten digit image set. The MNIST handwritten digit images are discretized into M×M images with M≥16 pixels using an interpolation algorithm. The grayscale value of each image is extracted, and the pixel value is converted into an image with a pixel value range of [1,5] through a linear transformation. The image with a pixel value range of [1,5] is equivalent to a virtual target relative dielectric constant distribution map with a range of [1,5]. The image with a pixel value range of [1,5] is the virtual target.

[0024] Then, simulate N i The transmitting antennas transmit transverse magnetic waves in turn, and the virtual target scattering field is calculated based on the moment method to obtain N r The scattered field data at the location of the receiving antennas is composed of N dimensions. r ×N i Virtual scattered field data matrix of ;

[0025] Finally, the back propagation algorithm is used to obtain the initial image of the virtual target with M×M pixels, M≥16.

[0026] Furthermore, the neural network of the electromagnetic imaging prediction model established in step 3 is composed of an input layer, an encoder, a decoder and an output layer;

[0027] The encoder includes multiple convolution blocks, each of which includes a 3×3 convolution layer, batch normalization and ReLU activation function, and a pooling layer is added between the two layers to reduce the size of the feature map; as the encoder goes deeper, the number of channels in each layer will gradually increase, and the resolution of the feature map will decrease accordingly until it reaches the middle layer, generating a feature map with a pixel number of (M / 4)×(M / 4);

[0028] The decoder includes multiple deconvolution blocks, each of which includes a 3×3 deconvolution layer, batch normalization and ReLU activation function, which is used to gradually restore the feature map with the number of pixels (M / 4)×(M / 4) to a picture with the same number of pixels as the input; after each upsampling in the decoder part, the feature map of the corresponding layer of the encoder is spliced ​​with the upsampling result of the decoder through a jump connection, which is used to enhance the utilization efficiency of the encoder feature map;

[0029] Next, the cost function is used to calculate the network error and perform regression to optimize the network parameters. After multiple iterative optimizations, the electromagnetic imaging prediction model is finally established.

[0030] Furthermore, the electromagnetic imaging prediction model also includes:

[0031] The feature map of the initial image of the virtual target is subjected to feature extraction and dimensionality reduction processing through the convolution layer. The first and second convolution layers are followed by residual connections to solve the gradient vanishing problem in deep neural networks. Each convolution layer is followed by an SE block to enhance the correlation between modeling channels to enhance the network's discrimination ability. At the same time, the input is directly passed to the output through a jump connection to make the network easier to train.

[0032] Furthermore, the residual connection includes a main path and a shortcut path.

[0033] Among them, the main path is composed of the transposed convolution, batch normalization and ReLU activation function, which is used to extract features and perform nonlinear transformation; the shortcut path is to pass the input to the next layer of the main path, or adjust the number of channels through 1×1 convolution, which is used to match the output dimension of the main path, increase the depth of the output feature map, and maintain the efficiency and effect of training.

[0034] Furthermore, the cost function is:

[0035] ,

[0036] in, is the pixel value of the predicted image of the virtual target, x is the pixel value of the real image of the virtual target, represents the pixel value of x located in the i-th row and j-th column, is the regularization parameter, Therefore, the first term of the cost function quantifies the difference between the predicted value and the true value, and the second term is the total variation (TV) regularization constraint.

[0037] Furthermore, the actual target prediction imaging result obtained in step 4 is a relative dielectric constant distribution map.

[0038] The beneficial effects of the present invention are as follows: the present invention obtains scattering field information based on antennas arranged around the imaging area, uses a back propagation algorithm (BP) to generate an actual target initial image, and then inputs it into a trained deep learning network (SERU-Net) electromagnetic imaging model to obtain a relative dielectric constant distribution map of the target scatterer, thereby realizing real-time imaging of the target scatterer.

[0039] The SERU-Net-based electromagnetic imaging model provided by the present invention not only handles the imaging problem under the condition of low signal-to-noise ratio in a short time and efficiently, but also significantly improves the problem that the classical network has poor imaging effect on weak scatterers and possible afterimages on strong scatterers, thereby greatly improving the imaging accuracy.

[0040] At the same time, the present invention adds residuals to the traditional U-Net network to improve the network training efficiency, accelerate the network training convergence, enhance the prediction accuracy and generalization ability, add a channel attention mechanism, enhance the network's attention to important features, improve the discriminability of feature maps, reduce model parameters, and greatly reduce the computational cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is the network structure of SERU-Net proposed by the present invention;

[0042] Figure 2 This is a schematic diagram of the inversion results of MNIST handwritten digits;

[0043] Figure 3 This is a schematic diagram of the polygon-like inversion result;

[0044] Figure 4 This is a schematic diagram of the Austrian ring inversion result;

[0045] Figure 5 It is a schematic diagram of the inversion results of experimental data. DETAILED DESCRIPTION

[0046] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0047] In order to achieve the above object, the present invention provides the following specific implementation methods:

[0048] Example 1: Figure 1 As shown, an electromagnetic inverse scattering imaging method based on the SERU-Net model includes the following steps:

[0049] S01, with the center point of the imaging area as the center, evenly arrange N i transmit antennas and N r There are N receiving antennas, iThe transmitting antennas transmit transverse magnetic waves in turn, and the actual target scattering field data are all captured by N r The data collected by the receiving antennas is N r ×N i The scattered field data matrix of

[0050] S02. Use the back propagation algorithm to obtain the initial image of the actual target with a pixel number of M×M, where M≥16 and N r ≥N i And N i 、N r Take a positive integer;

[0051] S03, randomly extract at least 400 MNIST handwritten digit images from the MNIST handwritten digit image set, and use the interpolation algorithm to discretize the MNIST handwritten digit images into images with M×M pixels, where M≥16;

[0052] S04, extracting the grayscale value of each image, and converting the pixel value into an image with a pixel value range of [1,5] through linear transformation, and equating the image with a pixel value range of [1,5] to a virtual target relative dielectric constant distribution map with a range of [1,5], then the image with a pixel value range of [1,5] is the virtual target;

[0053] S05, Simulation N i The transmitting antennas transmit transverse magnetic waves in turn, and the virtual target scattering field is calculated based on the moment method to obtain N r The scattered field data at the location of the receiving antennas is composed of N dimensions. r ×N i Virtual scattered field data matrix of ;

[0054] S06. Obtain an initial image of a virtual target with M×M pixels by using a back propagation algorithm, where M≥16.

[0055] S07. Based on the virtual scattering field data matrix, the virtual target is used as the label value of the SERU-Net model, and the initial image of the virtual target is used as the input value of the SERU-Net model to construct an electromagnetic imaging prediction model based on the SERU-Net model. The neural network of the electromagnetic imaging prediction model is composed of an input layer, an encoder, a decoder and an output layer, such as Figure 1 As shown;

[0056] The encoder of the electromagnetic imaging prediction model includes multiple convolutional blocks, each of which contains a 3×3 convolutional layer, batch normalization and ReLU activation function. A pooling layer is added between the two layers to reduce the size of the feature map. As the encoder goes deeper, the number of channels in each layer will gradually increase, while the resolution of the feature map will decrease accordingly until it reaches the middle layer, generating a feature map with a pixel number of (M / 4)×(M / 4).

[0057] The decoder of the electromagnetic imaging prediction model includes multiple deconvolution blocks, each of which contains a 3×3 deconvolution layer, batch normalization, and ReLU activation function, which is used to gradually restore the feature map with the number of pixels (M / 4)×(M / 4) to the same number of pixels as the input. After each upsampling, the decoder part concatenates the feature map of the corresponding layer of the encoder with the upsampling result of the decoder through a jump connection to enhance the utilization efficiency of the encoder feature map.

[0058] Among them, the feature map of the initial image of the virtual target is subjected to feature extraction and dimensionality reduction processing through the convolution layer. The first and second convolution layers are followed by residual connections to solve the gradient vanishing problem in deep neural networks. Each convolution layer is followed by an SE block to enhance the correlation between modeling channels to enhance the network's discrimination ability. At the same time, the input is directly passed to the output through the jump connection to make the network easier to train.

[0059] Among them, the residual connection includes the main path and the shortcut path:

[0060] The main path consists of the transposed convolution, batch normalization, and ReLU activation function, which are used to extract features and perform nonlinear transformations; the shortcut path passes the input to the next layer of the main path, or adjusts the number of channels through 1×1 convolution to match the output dimension of the main path, increase the depth of the output feature map, and maintain the efficiency and effect of training.

[0061] S08. Calculate the network error using the cost function and perform regression to optimize the network parameters. After multiple iterations of optimization, an electromagnetic imaging prediction model is finally established.

[0062] Among them, the cost function is:

[0063] ,

[0064] in, is the pixel value of the predicted image of the virtual target, x is the pixel value of the real image of the virtual target, represents the pixel value of x located in the i-th row and j-th column, is the regularization parameter, , therefore, the first term of the cost function quantifies the difference between the predicted value and the true value, and the second term is the total variation (TV) regularization constraint;

[0065] S09. Inputting the actual target initial image into the electromagnetic imaging prediction model to obtain the actual target prediction imaging result, and the obtained actual target prediction imaging result is a relative dielectric constant distribution map.

[0066] In order to further illustrate the technical solutions and technical effects of the present invention, the following specific examples are provided:

[0067] Specific example 1:

[0068] S01. Label data generation: 3,000 images are randomly extracted from the MNIST handwritten digit image set. The extracted images are preprocessed. The images are discretized into 64×64 pixels using an interpolation algorithm. The grayscale values ​​of the images are extracted and the pixel values ​​are converted into images with a pixel value range of [1,5] through a linear transformation. This image is equivalent to a virtual target relative dielectric constant distribution map with a range of [1,5], which is the label data of the SERU-Net model.

[0069] S02, input data generation: set the antenna operating frequency to 4GHz, then the wavelength , with size The center point of the imaging area D is the center of the circle, and the radius is 36 transmitting antennas and 36 receiving antennas are evenly arranged on the circle. The 36 transmitting antennas transmit transverse magnetic waves in sequence. In each transmission, the scattering field of the target in the imaging area D is collected by the 36 receiving antennas, so a 36×36 electromagnetic scattering field data matrix is ​​obtained. The back propagation algorithm (BP) is used to generate the initial image of the virtual target with a pixel number of 64×64. For the above 3000 virtual targets with a relative dielectric constant range of [1,5], the steps of generating the initial image of the virtual target are repeated to obtain 3000 initial images of the virtual target, which are the input values ​​of the SERU-Net model.

[0070] S03. Training of electromagnetic imaging model based on SERU-Net: 1000 virtual target label data and input data are randomly selected as training set, and the remaining 2000 are used as test set. Training is performed on a workstation with 1.67 TB memory, 2.1 GHz CPU and NVIDIA RTX 6000 graphics card. The initial learning rate is set to 5×10 -6 , the training is halved every 20 times, and the training process is terminated after 150 iterations to obtain the SERU-Net electromagnetic imaging model.

[0071] S04. Prediction of imaging results based on SERU-Net electromagnetic imaging model: Input the actual target initial image into the trained SERU-Net electromagnetic imaging model. The model output value is the predicted target scatterer relative dielectric constant distribution map, that is, the target scatterer imaging result.

[0072] like Figure 2 , Figure 3 , Figure 4 , Figure 5As shown, four example results are obtained below through the process of specific example 1 to illustrate the advantages of the present invention.

[0073] Example result 1:

[0074] In order to verify that the SERU-Net electromagnetic imaging model proposed in the present invention has significantly improved imaging accuracy compared with the traditional U-Net electromagnetic imaging model, the 2000 virtual target scatterers in the above test set are input into the SERU-Net electromagnetic imaging model, and the predicted target scatterer relative dielectric constant distribution map is output. It is compared with the virtual target initial image and the prediction result of the traditional U-Net electromagnetic imaging model, and analyzed from two aspects: imaging result comparison diagram and statistical data.

[0075] Five images were randomly selected from 2000 images as cases to compare the SERU-Net and U-Net electromagnetic imaging models. The imaging results are compared as shown in the figure below. Figure 2 As shown by Figure 2 From the imaging results of the handwritten number "1", it can be found that the SERU-Net electromagnetic imaging model has a significantly better optimization effect on weak scatterers than the traditional U-Net electromagnetic imaging model. And from the imaging results of the handwritten number "3", it can be found that the SERU-Net electromagnetic imaging model has a more significant effect on eliminating "afterimages". Figure 2 The scale on the right side of the figure represents the range of relative dielectric constant, and 20-60 in the figure is the number of pixels in each picture.

[0076] Then, a statistical analysis was performed on the 2000 prediction results, and the mean square error (MSE) and structural similarity (SSIM) were used as evaluation indicators. The statistical results of different methods are shown in Table 1. Table 1 also shows that SERU-Net has the lowest MSE and the highest SSIM, which is significantly better than the traditional U-Net electromagnetic imaging model.

[0077] Table 1

[0078]

[0079] Example result 2:

[0080] In order to verify the generalization ability of the SERU-Net electromagnetic imaging model proposed in this paper, the SERU-Net electromagnetic imaging model is used to predict the class polygons. The prediction results are as follows: Figure 3 As shown in the figure, it can be found that the effect of using the SERU-Net electromagnetic imaging model to predict the relative dielectric constant distribution map of the target scatterer is significantly better than the traditional U-Net electromagnetic imaging model, which proves that the SERU-Net electromagnetic imaging model has good generalization ability. Figure 3The scale on the right side of the figure represents the range of relative dielectric constant, and 20-60 in the figure is the number of pixels in each picture.

[0081] Example Result 3:

[0082] In order to verify the ability of the present invention to handle complex targets and strong scatterers, the SERU-Net electromagnetic imaging model is used to predict the Austrian ring pattern. The relative dielectric constant range of the ring is set to [1.5, 5], with a step size of 0.5. The prediction results are as follows: Figure 4 As shown, Figure 4 It can be found that when the relative dielectric constant is small, the SERU-Net electromagnetic imaging model predicts that the distribution map of the relative dielectric constant of the target scatterer is closer to the true value; when the relative dielectric constant is large, the inversion result deteriorates, but the SERU-Net electromagnetic imaging model is obviously better at handling "afterimages" than the traditional U-Net electromagnetic imaging model. The reconstruction capabilities of the present invention for both strong and weak scatterers are greatly improved.

[0083] Example Result 4:

[0084] In order to further verify the present invention, the SERU-Net electromagnetic imaging model was used to predict the experimental data provided by the Fresnel Institute in Marseille, France. The prediction results are as follows: Figure 5 As shown, Figure 5 It can be found that the SERU-Net electromagnetic imaging model has a much better effect in predicting the relative permittivity distribution map of the target scatterer than the traditional U-Net electromagnetic imaging model, and is closer to the true value. Figure 5 The scale on the right represents the range of relative permittivity.

[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An electromagnetic inverse scattering imaging method based on the SERU-Net model, characterized in that: The following steps are involved: Step 1: Use N i transmit antennas and N r The actual target scattering field data of the imaging area is obtained by using N receiving antennas. The actual target scattering field data is r ×N i The matrix is ​​then used to obtain the actual target initial image using the back propagation algorithm, where N r ≥N i And N i 、N r Take a positive integer; Step 2: Generate virtual targets in the imaging area based on the MNIST handwritten digit image set, and simulate N based on the moment method i transmit antennas and N r The receiving antennas obtain the scattering field data of the virtual target. The scattering field data of the virtual target is N r ×N i The matrix is ​​then used to obtain the initial image of the virtual target using the back propagation algorithm; Step 3: Based on the virtual scattering field data matrix, the virtual target is used as the label value of the SERU-Net model, and the initial image of the virtual target is used as the input value of the SERU-Net model to construct an electromagnetic imaging prediction model based on the SERU-Net model; Among them, the neural network of the established electromagnetic imaging prediction model is composed of an input layer, an encoder, a decoder and an output layer; The encoder includes multiple convolution blocks, each of which includes a 3×3 convolution layer, batch normalization and ReLU activation function, and a pooling layer is added between the two layers to reduce the size of the feature map; as the encoder goes deeper, the number of channels in each layer will gradually increase, and the resolution of the feature map will decrease accordingly until it reaches the middle layer, generating a feature map with a pixel number of (M / 4)×(M / 4); The decoder includes multiple deconvolution blocks, each of which includes a 3×3 deconvolution layer, batch normalization and ReLU activation function, which is used to gradually restore the feature map with the number of pixels (M / 4)×(M / 4) to a picture with the same number of pixels as the input; after each upsampling in the decoder part, the feature map of the corresponding layer of the encoder is spliced ​​with the upsampling result of the decoder through a jump connection, which is used to enhance the utilization efficiency of the encoder feature map; Next, the cost function is used to calculate the network error and perform regression to optimize the network parameters. After multiple iterations of optimization, an electromagnetic imaging prediction model is finally established. The electromagnetic imaging prediction model also includes: The feature map of the initial image of the virtual target is subjected to feature extraction and dimensionality reduction processing through the convolution layer. The first and second convolution layers are followed by residual connections to solve the gradient vanishing problem in deep neural networks. Each convolution layer is followed by SEblock to enhance the correlation between modeling channels to enhance the network's discrimination ability. At the same time, the input is directly passed to the output through the jump connection to make the network easier to train. The residual connection includes a main path and a shortcut path, The main path consists of transposed convolution, batch normalization, and ReLU activation function, which are used to extract features and perform nonlinear transformations. The shortcut path passes the input to the next layer of the main path, or adjusts the number of channels through 1×1 convolution to match the output dimension of the main path, increase the depth of the output feature map, and maintain the efficiency and effect of training. Step 4: input the actual target initial image into the electromagnetic imaging prediction model to obtain the actual target prediction imaging result.

2. The electromagnetic inverse scattering imaging method based on the SERU-Net model according to claim 1, characterized in that: The step one comprises: With the center point of the imaging area as the center, N pixels are evenly arranged on a circle with a radius of R. i transmit antennas and N r There are N receiving antennas, i The transmitting antennas transmit transverse magnetic waves in turn, and the actual target scattering field data are all captured by N r The data collected by the receiving antennas is N r ×N i The scattering field data matrix is ​​obtained, and the initial image of the actual target with M×M pixels is obtained using the back propagation algorithm, where M≥16.

3. The electromagnetic inverse scattering imaging method based on the SERU-Net model according to claim 1, characterized in that: The step 2 comprises: First, at least 400 MNIST handwritten digit images are randomly selected from the MNIST handwritten digit image set. The MNIST handwritten digit images are discretized into M×M images with M≥16 pixels using an interpolation algorithm. The grayscale value of each image is extracted, and the pixel value is converted into an image with a pixel value range of [1,5] through a linear transformation. The image with a pixel value range of [1,5] is equivalent to a virtual target relative dielectric constant distribution map with a range of [1,5]. The image with a pixel value range of [1,5] is the virtual target. Then, simulate N i The transmitting antennas transmit transverse magnetic waves in turn, and the virtual target scattering field is calculated based on the moment method to obtain N r The scattered field data at the location of the receiving antennas is composed of N dimensions. r ×N i Virtual scattered field data matrix of ; Finally, the back propagation algorithm is used to obtain the initial image of the virtual target with M×M pixels, M≥16.

4. The electromagnetic inverse scattering imaging method based on the SERU-Net model according to claim 1, characterized in that: The cost function is: ， in, is the pixel value of the predicted image of the virtual target, x is the pixel value of the real image of the virtual target, and x i,j represents the pixel value of x located in the i-th row and j-th column, is the regularization parameter, , therefore, the first term of the cost function quantifies the difference between the predicted value and the true value, and the second term is the total variation regularization constraint.

5. The electromagnetic inverse scattering imaging method based on the SERU-Net model according to any one of claims 1 to 4, characterized in that: The actual target prediction imaging result obtained in step 4 is a relative dielectric constant distribution map.

Images