Deep Learning-Based Microwave Inverse Scattering Imaging Method for Human Tissue

Abstract

This invention proposes a deep learning-based microwave backscattering imaging method for human tissue. The steps are as follows: acquiring scattering parameter images and training and testing sample sets from human tissue slices; constructing a reconstruction network model and iteratively training it; and obtaining the image reconstruction results. During the training of the reconstruction network model, the decoder fuses the multi-scale feature maps extracted by the spatial pyramid module and the preliminary feature maps extracted by the encoder, followed by upsampling. This allows for the extraction of effective features from the original scattering parameter matrix, which is often noisy. This reduces the dependence on the accuracy of the scattering data and avoids the impact of biases in the inversion results on imaging accuracy in existing technologies. Furthermore, the imaging process only uses the trained model for forward propagation, eliminating the need for backprojection updates, resulting in lower computational resource consumption and improved real-time imaging performance.

Description

Technical Field

[0001] This invention belongs to the field of microwave imaging technology and relates to a microwave backscattering imaging method, specifically a deep learning-based microwave backscattering imaging method for human tissue, which can be applied to fields such as medical imaging. Background Technology

[0002] Microwave backscattering imaging technology refers to obtaining information about the object's internal structure by emitting microwaves towards it and measuring the signals scattered back from its surface. Using the scattering data, backscattering algorithms can reconstruct the object's shape or dielectric constant distribution, thereby achieving high-precision imaging of the object's internal properties.

[0003] Currently, widely used medical imaging techniques in clinical practice include X-ray imaging, microwave backscattering imaging, computed tomography (CT), magnetic resonance imaging (MRI), and ultrasound imaging. Compared to other imaging technologies, microwave backscattering imaging has advantages such as non-ionization, safety, portable equipment, and low cost. However, due to the multiple scattering of incident electromagnetic waves at the target, a complex nonlinear relationship exists between the scattered field and the incident field. Traditional methods for solving the microwave backscattering problem mainly include linearization methods and optimization methods. Linearization methods are computationally fast but lack accuracy; optimization methods, while improving accuracy, have high computational complexity and are prone to getting trapped in local optima. The solution to the backscattering problem lacks systematic methods and stable algorithms, and the spatial resolution and dielectric parameter reconstruction accuracy need improvement.

[0004] The emergence of machine learning methods, especially the development of deep learning technology, has provided new solutions. Deep learning can utilize large-scale data for training, thereby overcoming the shortcomings of traditional methods and improving imaging accuracy and efficiency.

[0005] To achieve microwave inverse scattering imaging reconstruction of human tissues, patent application CN118330634A, entitled "Inverse Scattering Iterative Imaging Method and Device Based on Reduced-Order Model," discloses an inverse scattering iterative imaging method based on a reduced-order model. This method uses a multi-transmitter, multi-receiver architecture to transmit excitation waves, collect scattered field data, and construct a reduced-order model, thereby reducing the nonlinearity of the inverse scattering problem and improving inversion stability. In each iteration, the target parameter distribution is updated using Born data and the Jacobian matrix until the final imaging result is obtained. However, the accuracy of this method depends on the approximate accuracy of the reduced-order model, which may lead to deviations in the inversion results. Furthermore, due to its complex calculation process, especially during back-projection updates, it consumes significant computational resources and cannot meet the requirements of real-time imaging. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and propose a microwave inverse scattering imaging method based on deep learning to solve the technical problems of low imaging accuracy and poor real-time performance in the prior art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0008] (1) Obtain the scattering parameter image:

[0009] R antennas arranged in a circular, periodic pattern transmit sequentially. Electromagnetic waves of different frequencies affect the computational domain. Human tissue sections were irradiated, and the results were recorded for every two antennas. and Scattering parameters between the antenna pairs Among them, antenna For transmitting antenna, antenna The scattering parameters measured when using a receiving antenna are as follows: With antenna For transmitting antenna, antenna The scattering parameters measured when using a receiving antenna are as follows: Then construct the first Each frequency corresponds to For row vectors, with The scattering parameter matrix is ​​a column vector. This scattering parameter matrix is ​​then preprocessed to obtain... A three-channel scattering parameter image consisting of the real part, the imaginary part of the scattering parameters, and the microwave frequency, wherein... , , ;

[0010] (2) Obtain the training sample set and the test sample set:

[0011] From Randomly selected from the amplitude scattering parameter image The dielectric constant distribution image corresponding to the amplitude is used as a label, and the image is then labeled. The training sample set consists of amplitude scattering parameter images and their labels. , the remaining Amplitude scattering parameter images constitute the test sample set ,in , No. The labels corresponding to the amplitude scattering parameter images are ;

[0012] (3) Constructing a reconstruction network model :

[0013] Construct a reconstruction network model that includes a cascaded encoder, a spatial pyramid module, and a decoder, with the encoder's output also connected to the decoder's input. The spatial pyramid module includes multiple cascaded dilated convolutional blocks with different dilation rates, global average pooling blocks, and feature integration convolutional blocks, with the input of the global average pooling block also connected to the output of the encoder; the decoder includes parallel channel dimensionality reduction convolutional layers and a first interpolation upsampling module, as well as a cascaded feature extraction layer, a second interpolation upsampling module, and a channel transformation convolutional layer.

[0014] (4) Reconstructing the network model Perform iterative training:

[0015] training sample set As a model for reconstructing networks The input is used to iteratively train the model to obtain a trained reconstruction network model. ;

[0016] (5) Obtain the reconstruction results:

[0017] Test sample set As a trained reconstruction network model The input is propagated forward to obtain the reconstructed dielectric constant images corresponding to all test samples.

[0018] Compared with existing technologies, the present invention has the following advantages:

[0019] In the process of training the reconstruction network model, the decoder fuses the multi-scale feature map extracted by the spatial pyramid module and the preliminary feature map extracted by the encoder and then upsamples it. This allows for the extraction of effective features from the original scattering parameter matrix, which has a lot of noise. This reduces the dependence on the accuracy of the scattering data and avoids the impact of deviations in the inversion results on the imaging accuracy of existing technologies. Furthermore, the imaging process only uses the trained model for forward propagation, without the need for back-projection updates, resulting in lower computational resource consumption and improved real-time imaging performance. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0021] Figure 2 This is a schematic diagram of the scattering parameter image imaging system used in this invention;

[0022] Figure 3 This is a schematic diagram of the reconstructed network model of the present invention;

[0023] Figure 4 This is a schematic diagram of the decoder structure of the present invention. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Referring to Figure 1, the present invention includes the following steps:

[0026] Step 1) Obtain the scattering parameter image:

[0027] The imaging system used in this embodiment to acquire the scattering parameter image is as follows: Figure 2 As shown, the system includes a computational domain, a slice of human tissue to be tested, and an antenna array. The computational domain refers to the region used to solve the electromagnetic wave propagation and scattering problem. Numerical calculations are performed within this region to obtain scattering parameter data. It is defined as a uniformly distributed square with a side length of 400 mm, and its center is set as the origin of the coordinate system. The human tissue slice can be approximated as a circular thin slice with a diameter ranging from 180 mm to 190 mm. The background dielectric constant of the imaging region is set to 1, while the dielectric constant of the human tissue ranges from 2 to 55. The antennas in the antenna array are dipole antennas with discrete surface ports, and the input excitation is a power source excitation. The input excitation enters the dipole arm through a feed line and is transformed into a radiated wave illuminating the phantom. Each antenna can function as both a transmitting and receiving antenna. Through multiple transmissions and receptions, comprehensive coverage of the human tissue slice within the computational domain is achieved.

[0028] The steps to obtain the scattering parameter image are as follows: through a circularly periodically arranged... Each antenna transmits in sequence. Electromagnetic waves of different frequencies affect the computational domain. Human tissue sections were irradiated, and the results were recorded for every two antennas. and Scattering parameters between the antenna pairs Then construct the first Each frequency corresponds to For row vectors, with The scattering parameter matrix is ​​given by columns, and then the scattering parameter matrix is ​​preprocessed to obtain... A three-channel scattering parameter image consisting of the real part, the imaginary part, and the microwave frequency of the scattering parameters. , , In this embodiment, , Selection of electromagnetic wave frequency range , , .

[0029] The preprocessing of each scattering parameter matrix involves the following steps: performing preprocessing on each scattering parameter matrix based on the extreme values ​​of the scattering parameters in each matrix. and Normalization yields the normalized real and imaginary scattering parameter matrices. Simultaneously, the electromagnetic wave... Each frequency is analyzed based on frequency extrema. and The normalization process generates a frequency matrix whose elements are the normalized frequency values ​​and have the same size as the scattering parameter matrix. The normalized real part scattering parameter matrix, imaginary part scattering parameter matrix, and frequency matrix are then standardized to obtain the standardized dielectric constant. and frequency Linearly map its numerical range to . and The calculation formula is:

[0030] ,

[0031] ,

[0032] The normalized real scattering parameter matrix, imaginary dielectric constant distribution matrix, and frequency matrix are mapped to the red, green, and blue color channels of the image, respectively, to obtain a three-channel scattering parameter image.

[0033] Step 2) Obtain the training sample set and the test sample set:

[0034] From Randomly selected from the amplitude scattering parameter image The dielectric constant distribution image corresponding to the amplitude is used as a label, and the image is then labeled. The training sample set consists of amplitude scattering parameter images and their labels. , the remaining Amplitude scattering parameter images constitute the test sample set ,in , No. The labels corresponding to the amplitude scattering parameter images are ;

[0035] The method for constructing the dielectric constant distribution image is as follows: For each scattering parameter image, acquire the grayscale image of the human tissue slice in the test area during each irradiation. Based on the color intensity of each pixel in the image, convert the human tissue slice image into a numerical matrix. Assign values ​​to the elements in the matrix using the real and imaginary parts of the dielectric constant corresponding to each pixel in the human tissue slice image, obtaining the real part dielectric constant distribution matrix and the imaginary part dielectric constant distribution matrix. For each dielectric constant distribution matrix, normalize the assigned dielectric constant distribution matrix using the maximum value in each matrix element, and linearly map the numerical range of the normalized matrix to... The standardized real and imaginary dielectric constant distribution matrices are obtained and assigned to the red and green color channels of the image, respectively. The blue channel is left empty, resulting in a two-channel dielectric constant distribution image.

[0036] Step 3) Constructing the reconstruction network model Its structure is as follows Figure 3 As shown:

[0037] Construct a reconstruction network model that includes a cascaded encoder, a spatial pyramid module, and a decoder, with the encoder's output also connected to the decoder's input. The spatial pyramid module includes multiple cascaded dilated convolutional blocks with different dilation rates, global average pooling blocks, and feature integration convolutional blocks, with the input of the global average pooling block also connected to the output of the encoder; the decoder includes parallel channel dimensionality reduction convolutional layers and a first interpolation upsampling module, as well as a cascaded feature extraction layer, a second interpolation upsampling module, and a channel transformation convolutional layer.

[0038] The encoder section utilizes a ResNet-50 network to extract features from the input image. Its main structure includes cascaded convolutional layers, pooling layers, and multiple residual modules. The convolutional layers use kernels with a size of... The convolution operation extracts basic features, with an output channel count of 64 and a stride of 2. The feature map obtained from the convolution operation is then processed by a convolution kernel of size [missing value]. A downsampling max pooling layer with a stride of 2 is used to reduce the spatial dimension of the feature map. The residual block part extracts increasingly higher-level features in a hierarchical structure. This part includes four main stages: The first stage contains 3 residual blocks, each consisting of 3 convolutional layers with output channels of 64, 64, and 256 respectively, increasing the number of channels in the feature map to 256; the second stage contains 4 residual blocks, each with output channels of 128, 128, and 512 respectively, increasing the number of channels in the feature map to 512; the third stage contains 6 residual blocks, each with output channels of 256, 256, and 1024 respectively, increasing the number of channels in the feature map to 1024; the fourth stage contains 3 residual blocks, each with output channels of 512, 512, and 2048, extracting the highest-level features and increasing the number of channels in the feature map to 2048, reaching the maximum depth of the network. After the above processing, the encoder part extracts the preliminary feature map of the input image. .

[0039] In the spatial pyramid module, a dilated convolution method based on the DeepLabV3 model is used to generate the initial feature map output by the encoder. The convolution operation uses multiple dilated convolutional blocks with different dilation rates. The basic principle of dilated convolution is to introduce holes into the convolutional kernel, allowing the convolutional operation to cover a larger receptive field. This enables the network to capture feature information at different scales without increasing computational complexity. Multi-scale feature extraction is achieved by applying convolutional operations with different dilation rates. The dilation rate allows the convolutional kernel to skip some positions on the input feature map, thereby increasing the receptive field. The spatial pyramid module applies multiple... Hollow convolutional layers with dilation rates set to 6, 12, and 18 were used to capture features at different scales.

[0040] ,

[0041] in, The height and width position indices of the output feature map. It is the location of the output feature map after the convolution operation. The value at that location. The input feature map is the source data for the convolution operation. For convolution kernel, The height and width indexes of the convolution kernel refer to their specific positions within the convolution kernel matrix. The width of the convolution kernel. is the height of the convolution kernel. The dilation rate of the dilated convolution kernel determines the stride of the kernel as it slides across the input feature map. During dilated convolution, the stride at each position of the kernel... By multiplying by the corresponding dilation position on the input feature map The value of the feature map is calculated by summing all the results and determining the output feature map. The value at that location.

[0042] The spatial pyramid module also includes a global average pooling block and a feature integration convolutional block. The global average pooling layer performs average pooling on each feature map before feeding it into the feature integration convolutional block. This module contains a size of... The convolutional kernels and linear interpolation upsampling operations reduce the number of channels to 256, and the feature fusion convolutional block upsamples the feature maps to the same spatial size as the input feature maps through bilinear interpolation. After all these feature maps are fused, a multi-scale feature map is generated. .

[0043] The decoder section includes parallel channel dimensionality reduction convolutional layers and a first interpolation upsampling module, as well as cascaded feature extraction layers, a second interpolation upsampling module, and channel transformation convolutional layers. The channel dimensionality reduction convolutional layers perform dimensionality reduction operations on the preliminary feature maps output by the encoder, using... The convolutional kernel reduces the number of channels to 256 to improve computational efficiency and accommodate subsequent processing. The first interpolation upsampling block uses bilinear interpolation to upsample the multi-scale feature map output by the spatial pyramid module to the same spatial size as the dimensionality-reduced feature map, thus obtaining the feature map. The feature extraction layer will reduce the dimensionality of the feature map. and upsampled multi-scale feature maps To splice together, and through two sets Convolutional layers are used to further process the upsampled feature maps. Each convolutional layer includes batch normalization and ReLU activation function. After the convolution operation, a fused feature map is generated. .

[0044] The second interpolation upsampling module processes the fused feature map. Bilinear interpolation upsampling is performed to restore the same spatial resolution as the input feature map, resulting in the feature map. The final channel transformation convolutional layer is a 1x1 convolutional layer, and the shape of the input feature map is... The shape of the output feature map is For the network of this invention, the number of input channels... Number of output channels The specific implementation of dimensionality reduction for the output channel, reducing it to the dielectric parameter matrices of the real and imaginary parts of two channels, is as follows: using the `nn.Conv2d` function, setting the convolution kernel size to... The step size is 1. The formula for convolution is as follows:

[0045] ,

[0046] Wherein, input is the feature map. , shape `output` is the output feature map after the convolution operation, with the shape of... , yes The weights of the convolution kernel, It is a bias term. The channel transforms the convolutional layer on the feature map. The channel number is converted to two channels to output the real and imaginary parts of the dielectric constant for each pixel, thus generating the final dielectric constant distribution image. .

[0047] Step 4) Reconstruct the network model Perform iterative training:

[0048] Step 4a) Use the training sample set as input to iteratively train the reconstructed network model, initializing the number of iterations to be... The maximum number of iterations is , , No. Reconstructing the network model in the next iteration The parameters are , and order ;

[0049] Step 4b) The encoder performs preliminary feature extraction for each training sample; the spatial pyramid module extracts the preliminary feature map output by the encoder. Multi-scale feature extraction is performed; the decoder extracts multi-scale feature maps from the spatial pyramid module and the encoder extracts preliminary feature maps. After fusion, upsampling is performed to obtain the first... Dielectric constant distribution image corresponding to each training sample ;

[0050] The spatial pyramid module performs multi-scale feature extraction on the preliminary feature map output by the encoder. The steps are as follows:

[0051] Step 4b1) The encoder performs convolution operations on each of the extracted preliminary feature maps using multiple dilated convolution blocks with different dilation rates to obtain multiple feature maps of different scales.

[0052] Step 4b2) The global average pooling block performs pooling processing on the initial feature map output by the encoder and the multiple feature maps of different scales output by the dilated convolution block, and then concatenates the pooled feature maps to obtain a global feature map containing multiple feature maps.

[0053] Step 4b3) The feature integration convolutional block integrates the features of multiple feature maps contained in the global feature map to obtain a multi-scale feature map.

[0054] The decoder fuses the multi-scale feature maps extracted by the spatial pyramid module and the preliminary feature maps extracted by the encoder, and then performs upsampling. Its structure is as follows: Figure 4 As shown, the implementation steps are as follows:

[0055] Preliminary feature map of encoder output from channel-reduced convolutional layers. Dimensionality reduction is performed, and the first interpolation upsampling module simultaneously modifies the multi-scale feature map output by the pyramid module. Interpolation and upsampling are performed; the feature extraction layer outputs a low-dimensional preliminary feature map from the channel-reducing convolutional layer. The size output by the first interpolation upsampling module and Matching multi-scale feature maps The concatenated feature maps are then subjected to a convolution operation to obtain a fused feature map. The second interpolation upsampling module extracts the fused feature map from the feature extraction layer. Interpolation upsampling is performed; the channel conversion convolutional layer outputs a feature map with the same size as the dielectric constant distribution map label from the binary interpolation upsampling module. The channel number is converted to 2 to predict the real and imaginary parts of the dielectric constant at each pixel location, thus obtaining the dielectric constant distribution image. .

[0056] Step 4c) Employ the root mean square error loss function and analyze it using the dielectric constant distribution image. and their corresponding tags Computational reconstruction of network models loss value Then through Parameters of the network model , The network model is updated to obtain the reconstruction network model for this iteration. ;

[0057] Reconstructing the network model loss value The calculation formula is:

[0058] ,

[0059] in, This indicates a summation operation.

[0060] Through the mean square error loss function For the model The update is performed using the following formula:

[0061] ,

[0062] ,

[0063] in, , They represent , The update results It's the learning rate. , They represent right , The partial derivative operation.

[0064] Step 4d) Judgment If true, then a well-trained reconstruction network model is obtained. Otherwise , Then proceed to step (4b).

[0065] Step 5) Obtain the reconstruction results:

[0066] Test sample set As a trained reconstruction network model The input is propagated forward to obtain the reconstructed dielectric constant images corresponding to all test samples.

[0067] Based on simulation results, the proposed deep learning-based microwave inverse scattering imaging method for human tissue achieves a mean square error (MSE) of 15.38 and a root mean square error (RMSE) of 3.91 on the training set, while on the test set, the mean MSE is 14.03 and the RMSE is 3.68. Simulation results demonstrate that this method effectively reduces errors in microwave inverse scattering imaging of human tissue, and the deep learning-based approach offers higher resolution and enables real-time imaging.

Claims

1. A microwave inverse scattering imaging method based on deep learning, characterized in that, Includes the following steps: (1) Obtain the scattering parameter image: Arranged in a circular, periodic pattern Each antenna transmits in sequence. Electromagnetic waves of different frequencies affect the computational domain. Human tissue sections were irradiated, and the results were recorded for every two antennas. and Scattering parameters between the antenna pairs Among them, antenna For transmitting antenna, antenna The scattering parameters measured when using a receiving antenna are as follows: With antenna For transmitting antenna, antenna The scattering parameters measured when using a receiving antenna are as follows: Then construct the first Each frequency corresponds to For row vectors, with The scattering parameter matrix is ​​a column vector. This scattering parameter matrix is ​​then preprocessed to obtain... A three-channel scattering parameter image consisting of the real part, the imaginary part, and the microwave frequency of the scattering parameters. , , ; (2) Obtain the training sample set and the test sample set: From Randomly selected from the amplitude scattering parameter image The dielectric constant distribution image corresponding to the amplitude is used as a label, and the image is then labeled. The training sample set consists of amplitude scattering parameter images and their labels. , the remaining Amplitude scattering parameter images constitute the test sample set ,in , No. The labels corresponding to the amplitude scattering parameter images are , ; (3) Constructing a reconstruction network model : Construct a reconstruction network model that includes a cascaded encoder, a spatial pyramid module, and a decoder, with the encoder's output also connected to the decoder's input. The spatial pyramid module includes multiple cascaded dilated convolutional blocks with different dilation rates, global average pooling blocks, and feature integration convolutional blocks, with the input of the global average pooling block also connected to the output of the encoder; the decoder includes parallel channel dimensionality reduction convolutional layers and a first interpolation upsampling module, as well as a cascaded feature extraction layer, a second interpolation upsampling module, and a channel transformation convolutional layer. (4) Reconstructing the network model Perform iterative training: training sample set As a model for reconstructing networks The input is used to iteratively train the model to obtain a trained reconstruction network model. ; (5) Obtain the reconstruction results: Test sample set As a trained reconstruction network model The input is propagated forward to obtain the reconstructed dielectric constant images corresponding to all test samples.

2. The method according to claim 1, characterized in that, The preprocessing described in step (1) is implemented as follows: For each scattering parameter matrix, normalization is performed based on the extreme values ​​of the scattering parameters in each matrix, yielding normalized real and imaginary scattering parameter matrices. Simultaneously, the electromagnetic wave... Normalization is performed on each frequency based on its frequency extrema to generate a frequency matrix whose elements are the normalized frequency values ​​and are the same size as the scattering parameter matrix. The normalized real part scattering parameter matrix, imaginary part scattering parameter matrix, and frequency matrix are then standardized. The standardized real part scattering parameter matrix, imaginary part dielectric constant distribution matrix, and frequency matrix are then mapped to the red, green, and blue color channels of the image, respectively, to obtain a three-channel scattering parameter image.

3. The method according to claim 1, characterized in that, The dielectric constant distribution image mentioned in step (2) is constructed as follows: For each scattering parameter image, acquire the grayscale image of the human tissue slice in the test area at each irradiation time. The grayscale image of human tissue slices is based on the color intensity of each pixel in the image. Converted to a numerical matrix and processed using human tissue slice images. The real and imaginary parts of the dielectric constant of the biological tissue corresponding to each pixel are assigned to the elements in the matrix to obtain the real part dielectric constant distribution matrix and the imaginary part dielectric constant distribution matrix. For each dielectric constant distribution matrix, the extreme values ​​in each matrix element are used to normalize the assigned dielectric constant distribution matrix. The normalized matrix values ​​are then amplified to the range of 0 to 255 to obtain the standardized real part dielectric constant distribution matrix and the imaginary part dielectric constant distribution matrix. These are then mapped to the red and green color channels of the image to obtain the dielectric constant distribution image of the two channels.

4. The method according to claim 1, characterized in that, The encoder described in step (3) includes cascaded convolutional layers, pooling layers and multiple residual modules.

5. The method according to claim 1, characterized in that, The iterative training of the reconstructed network model described in step (4) is implemented as follows: (4a) Initialize the number of iterations to be The maximum number of iterations is , , No. Reconstructing the network model in the next iteration The parameters are , and order ; (4b) The encoder performs preliminary feature extraction for each training sample; the spatial pyramid module extracts the preliminary feature map output by the encoder. Perform multi-scale feature extraction; The decoder extracts multi-scale feature maps from the spatial pyramid module and the encoder extracts preliminary feature maps. After fusion, upsampling is performed to obtain the first... Dielectric constant distribution image corresponding to each training sample ; (4c) The root mean square error loss function is used, and the dielectric constant distribution image is used as the basis for the loss function. and its corresponding tags Computational reconstruction of network models loss value Then through Parameters of the network model , The network model is updated to obtain the reconstruction network model for this iteration. ; (4d) Judgment If true, then a well-trained reconstruction network model is obtained. Otherwise , Then proceed to step (4b).

6. The method according to claim 5, characterized in that, The spatial pyramid module described in step (4b) generates the preliminary feature map output by the encoder. Multi-scale feature extraction is performed, and the steps are as follows: (4b1) Multiple dilated convolutional blocks with different dilation rates are used by the encoder to process each extracted preliminary feature map. Perform convolution operations to obtain feature maps of different scales. ; (4b2) Preliminary feature map of encoder output by global average pooling block Multiple feature maps of different scales output by dilated convolutional blocks Pooling is performed separately for each feature map, and the pooled feature maps are then concatenated to obtain a global feature map containing multiple feature maps. ; (4b3) Feature integration convolutional blocks on global feature maps Multiple feature maps are integrated to obtain a multi-scale feature map. .

7. The method according to claim 5, characterized in that, The decoder described in step (4b) fuses the multi-scale feature map extracted by the spatial pyramid module and the preliminary feature map extracted by the encoder, and then performs upsampling. The steps are as follows: Preliminary feature map of encoder output from channel-reduced convolutional layers. Dimensionality reduction is performed, and the first interpolation upsampling module simultaneously modifies the multi-scale feature map output by the pyramid module. Perform interpolation upsampling; features Extraction layer: Low-dimensional preliminary feature map output from channel-reduction convolutional layer The size output by the first interpolation upsampling module and Matching multi-scale feature maps The concatenated feature maps are then subjected to a convolution operation to obtain a fused feature map. The second interpolation upsampling module extracts the fused feature map from the feature extraction layer. Interpolation upsampling is performed; the channel conversion convolutional layer outputs a feature map with the same size as the dielectric constant distribution map label from the binary interpolation upsampling module. The channel number is converted to 2 to predict the real and imaginary parts of the dielectric constant at each pixel location, thus obtaining the dielectric constant distribution image. .

8. The method according to claim 5, characterized in that, The reconstruction of the network model described in step (4c) loss value The calculation formula is: ， in, This indicates the summation operation.

9. The method according to claim 5, characterized in that, In step (4c), the mean square error loss function is used. For the model The update is performed using the following formula: ， ， in, , They represent , The update results It's the learning rate. , They represent right , The partial derivative operation.

Images