Fast computational imaging method for near-field metasurface antennas based on deep convolutional neural networks

Abstract

The present invention relates to a near-field metasurface antenna fast computational imaging method based on a deep convolutional neural network, including: reflecting the mapping relationship between echo measurement values ​​and scene target scattering coefficient distribution by constructing a measurement matrix; constructing a near-field MSACI-Net imaging network based on a convolutional neural network, training the near-field MSACI-Net imaging network, and reconstructing a receiving matrix; designing a network training algorithm according to the near-field MSACI-Net imaging network, using a mean square error to calculate the loss between the network prediction value and the true value, and using the Adam optimization algorithm to minimize the loss of the near-field MSACI-Net imaging network; using measured field data to train the near-field MSACI-Net imaging network, and reconstructing high-quality original target information in real time and with high precision after the training is completed. The present invention is based on a deep convolutional neural network, and the constructed metasurface antenna computational imaging network quickly realizes near-field metamaterial aperture calculation, and can quickly and accurately image point-sized targets and more complex targets.

Description

Technical Field

[0001] The present invention relates to the technical field of near-field computational imaging of metasurface antennas, and in particular to a near-field fast computational imaging method of metasurface antennas based on a deep convolutional neural network. Background Art

[0002] Metasurface antennas have made great progress in both antenna design and computational imaging (CI) applications due to their ability to sculpt complex radiation wavefronts and couple energy from a reference wave to a desired radiation pattern. The overall imaging performance of a metasurface antenna computational imaging (MSACI) system is closely related to the optimized design of the front-end antenna and the robust back-end computational imaging algorithm.

[0003] When designing CI metasurface antennas, the key factors affecting CI performance include radiating elements, feeding structures, and reconstruction algorithms. In the current CI framework, the complex interaction of these factors leads to some inherent contradictions between maximizing spatial frequency sampling and maintaining the necessary signal-to-noise ratio: namely, measurement quantity and quality, radiation losses, and the loaded Q factor of radar elements. In order to achieve the main goal of CI for metasurface antennas, the trade-offs between the above inconsistencies need to be properly balanced in imaging applications. In most current MSACI demonstrations, the constructed measurement matrix is ​​usually not subject to direct matrix inversion, and then the scene reflectivity is retrieved using computational techniques (mainly including least squares solvers or compressive reconstruction techniques). These methods either directly adopt the matched filter (MF) method or use the MF result as the initial estimate and use iterative optimization methods to refine the estimate. The main challenge of this type of traditional compressed sensing (CS) reconstruction techniques is that the success of the reconstruction process depends heavily on the randomness and condition number of the measurement matrix. Furthermore, when it comes to imaging problems with extremely large dimensionality, the computational cost will increase significantly considering the memory and storage usage required for matrix inversion and iteration, and compression reconstruction techniques using ill-conditioned measurement matrices will also suffer from high signal-to-noise ratios. Summary of the invention

[0004] The object of the present invention is to provide a near-field metasurface antenna fast computational imaging method based on a deep convolutional neural network, which can quickly and accurately image point-sized targets and more complex targets, reconstruct ordinary sparse targets, reconstruct extended targets, and obtain high-quality reconstructed images.

[0005] To achieve the above object, the present invention adopts the following technical solution: a near-field metasurface antenna fast computational imaging method based on a deep convolutional neural network, the method comprising the following steps in order:

[0006] (1) The mathematical principle of near-field imaging of metasurface antennas is modeled. The probe of the metamaterial aperture antenna receives the echo measurement value scattered by the target scene to be detected, and a measurement matrix is ​​constructed to reflect the mapping relationship between the echo measurement value and the scattering coefficient distribution of the target scene.

[0007] (2) A near-field MSACI-Net imaging network is constructed based on a convolutional neural network. The entire near-field MSACI-Net imaging network consists of a fully connected layer and a convolutional layer. The measurement values ​​collected by the receiving antenna and the estimation results of the MF method together constitute the training data set and test set of the MSACI-Net network. The near-field MSACI-Net imaging network is trained and the receiving matrix is ​​reconstructed.

[0008] (3) Design a network training algorithm based on the near-field MSACI-Net imaging network, use the mean square error to calculate the loss between the network prediction value and the true value, and use the Adam optimization algorithm to minimize the loss of the near-field MSACI-Net imaging network by setting the hyperparameters of the number of iterations, exponential decay rate, and learning rate;

[0009] (4) The near-field MSACI-Net imaging network is trained using measured field data. After training, the original target information can be reconstructed in real time with high precision and high quality.

[0010] The step (1) specifically comprises the following steps:

[0011] (1a) The echo measurement value g received by the probe of the metamaterial aperture antenna contains all the scattering information of the near-field scene target. Under the first-order Born approximate scattering model, the mapping relationship between the echo measurement value g and the scene target scattering coefficient σ is described by constructing the measurement matrix H. When reconstructing the measurement matrix, it is necessary to calculate the linear phase terms introduced by the pitch and azimuth scene resolution units contained in each distance slice at different distance slices in the scene, and then arrange and combine them to form the total measurement matrix.

[0012] (1b) The measurement matrix H is defined as:

[0013]

[0014] In formula (1), represents the vector pointing into free space, Indicates the scene The size of the radiation field at Indicates that the probe is in scene space The physical meaning of H is the size of the radiation field at the transmitting antenna U 0TX and receiving antenna U 0RX Aperture at the target location At , the product of the two radiation field sizes is in the form of:

[0015]

[0016] The integral equation in formula (1) is written in the form of discrete matrix measurement equation:

[0017]

[0018] That is, the mathematical model of metasurface antenna near-field imaging is:

[0019] g=Hσ+n (4)

[0020] In formula (4), is the echo measurement value received by the probe, M is the number of effective radiation patterns of the metamaterial aperture antenna, is the scattering coefficient of the scene target to be resolved, N is the number of discrete units in the scene, is the measurement noise term, Defined as a measurement matrix or measurement pattern.

[0021] The step (2) specifically comprises the following steps:

[0022] (2a) Based on the near-field imaging mathematical model of the metasurface antenna g = Hσ + n, a metasurface antenna computational imaging system is constructed using multiple fully connected layers and convolutional layers. Based on the existing metasurface antenna scene target echo measurement value g, an approximate scene target estimation result is obtained through the traditional MF method. The training dataset that constitutes the near-field MSACI-Net imaging network;

[0023] (2b) The near-field MSACI-Net imaging network consists of three fully connected layers and four convolutional layers. Each fully connected layer contains N neurons. When the echo measurement value g is input, the estimated value of σ is obtained after passing through three fully connected layers. After the reshape operation, The shape changes from N×1 to k x ×k y , where k x and k y Represents horizontal and vertical near-field scan points, and the size of the feature maps generated by all fully connected layers is consistent with the number of voxels to be imaged;

[0024] (2c) The fully connected layer is followed by four convolutional layers. The first and third convolutional layers use convolutional kernels to reduce network parameters, and the second and last convolutional layers use convolutional kernels to generate feature maps as the output of the network. Each of the four convolutional layers performs a ReLu zero-padding operation to keep the imaging system consistent with the size of the feature map.

[0025] (2d) After four convolutional layers, the reconstructed target image is finally obtained, completing the reconstruction operation of the entire image.

[0026] The step (3) specifically comprises the following steps:

[0027] (3a) The training data set containing N training samples is represented as {(σ1,g1),(σ2,g2),...,(σ N ,g N )}, find the error based on the training data set The smallest H, denoted as H f :

[0028]

[0029] In formula (5), σ=[σ1,σ2,...,σ N ], is the original scene target data of the entire near field, is the reconstructed scene target obtained by inverting the received echo measurement value, that is, in g = [g1, g2, ..., g N ], n and H are known, and it is calculated according to the mathematical model of near-field imaging of metasurface antenna as g = Hσ + n; where, is the echo measurement value received by the probe, M is the number of effective radiation patterns of the metamaterial aperture antenna, is the scattering coefficient of the scene target to be resolved, is the measurement noise term, Defined as a measurement matrix or measurement pattern;

[0030] (3b) Set the basic parameters of the near-field MSACI-Net imaging network, including the randomly initialized measurement matrix H, the random objective function f(·) with parameters, the maximum number of iterations T, the exponential decay rate of the moment estimate ρ1 = 0.9, ρ2 = 0.999, and the numerical stability constant ε = 10 -8 , the learning rate of the gradient optimization algorithm is δ = 10 -3 ;

[0031] (3c) Use the mean square error function MSE to calculate the loss L(H) between the network prediction value and the true value, that is:

[0032]

[0033] In the formula, σ i ∈σ, represents the imaging data of a target in the near field, represents imaging data of a near-field target calculated by a near-field imaging mathematical model of a metasurface antenna;

[0034] (3d) Use the Adam optimization algorithm to train the network and calculate the gradient of the measurement matrix H The weight coefficients of the convolutional layer are updated during forward and backward propagation, updating the first-order moment vector and the second-order moment vector Finally update the objective function parameters Until the end of the training cycle, the system's loss function tends to converge.

[0035] The step (4) specifically comprises the following steps:

[0036] (4a) Prepare the training dataset (g tr ,σ tr ) and the test set g t , set the maximum number of iterations T for training the near-field MSACI-Net imaging network, and randomly initialize the weight parameter W of each layer of the near-field MSACI-Net imaging network;

[0037] (4b) In each round of training, a random batch of g is drawn from the training dataset. tr and σ tr They are used as the input and output of the near-field MSACI-Net imaging network respectively; the MSE function is used as the loss function of the near-field MSACI-Net imaging network to train the weight parameter W of each layer in the near-field MSACI-Net imaging network;

[0038] (4c) After T rounds of iterations, the loss value of the near-field MSACI-Net imaging network approaches a minimum value, the near-field MSACI-Net imaging network training is completed, and the trained near-field MSACI-Net imaging network is saved;

[0039] (4d) The near-field MSACI-Net imaging network after training is used to perform high-precision imaging of the target. t Input the trained near-field MSACI-Net imaging network to obtain the reconstructed target σ t , reconstruct high-quality original target information in real time and with high precision.

[0040] It can be seen from the above technical scheme that the beneficial effects of the present invention are: first, the metasurface antenna computational imaging network constructed by the present invention based on the deep convolution neural network (DCNN) method can quickly realize near-field metamaterial aperture computational imaging, and can quickly and accurately image point-sized targets and more complex targets, and can effectively reduce the inevitable computational burden and the necessary large operating frequency band; second, the DCN method is used to realize near-field metasurface antenna rapid imaging, which can not only reconstruct ordinary sparse targets, but also reconstruct extended targets, and obtain high-quality reconstructed images; third, compared with traditional sparse reconstruction algorithms such as matched filtering MF and sparse Bayes SBL, the present invention does not require a long number of iterations, has a short imaging time and low complexity, and can call the network to perform offline test imaging results after the network training is completed. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the method of the present invention;

[0042] Figure 2 Schematic diagram of the architecture of the near-field MSACI-Net imaging network in the present invention;

[0043] Figure 3 It is a training flow chart of the near-field MSACI-Net imaging network in the present invention;

[0044] Figure 4 , Figure 5 All of them are reconstruction result diagrams of the MSACI network under different signal-to-noise ratios in Example 1 of the present invention;

[0045] Figure 6 This is the MSE variation curve of MSACI-MNIST at different scene information sampling rates in Example 1 of the present invention;

[0046] Figure 7 This is the PSNR variation curve of MSACI-MNIST at different scene information sampling rates in Example 1 of the present invention;

[0047] Figure 8 This is the MSE variation curve of MSACI-FMNIST at different scene information sampling rates in Example 1 of the present invention;

[0048] Fig. 9 The PSNR variation curve of MSACI-FMNIST at different scene information sampling rates provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0049] like Figure 1As shown, a method for fast computational imaging of near-field metasurface antenna based on deep convolutional neural network includes the following steps in sequence:

[0050] (1) The mathematical principle of near-field imaging of metasurface antennas is modeled. The probe of the metamaterial aperture antenna receives the echo measurement value scattered by the target scene to be detected, and a measurement matrix is ​​constructed to reflect the mapping relationship between the echo measurement value and the scattering coefficient distribution of the target scene.

[0051] (2) A near-field MSACI-Net imaging network is constructed based on a convolutional neural network. The entire near-field MSACI-Net imaging network consists of a fully connected layer and a convolutional layer. The measurement values ​​collected by the receiving antenna and the estimation results of the MF method together constitute the training data set and test set of the MSACI-Net network. The near-field MSACI-Net imaging network is trained and the receiving matrix is ​​reconstructed. The test set is generated together with the training data set. 70% of the 60,000 sets of data are divided into training data sets and 30% are test sets.

[0052] (3) Design a network training algorithm based on the near-field MSACI-Net imaging network, use the mean square error to calculate the loss between the network prediction value and the true value, and use the Adam optimization algorithm to minimize the loss of the near-field MSACI-Net imaging network by setting the hyperparameters of the number of iterations, exponential decay rate, and learning rate;

[0053] (4) The near-field MSACI-Net imaging network is trained using measured field data. After training, the original target information is reconstructed in real time with high precision and high quality.

[0054] The step (1) specifically comprises the following steps:

[0055] (1a) The echo measurement value g received by the probe of the metamaterial aperture antenna contains all the scattering information of the near-field scene target. Under the first-order Born approximate scattering model, the mapping relationship between the echo measurement value g and the scene target scattering coefficient σ is described by constructing the measurement matrix H. When reconstructing the measurement matrix, it is necessary to calculate the linear phase terms introduced by the pitch and azimuth scene resolution units contained in each distance slice at different distance slices in the scene, and then arrange and combine them to form the total measurement matrix.

[0056] (1b) The measurement matrix H is defined as:

[0057]

[0058] In formula (1), represents the vector pointing into free space, Indicates the scene The size of the radiation field at Indicates that the probe is in scene space The physical meaning of H is the size of the radiation field at the transmitting antenna U 0TX and receiving antenna U 0RX Aperture at the target location At , the product of the two radiation field sizes is in the form of:

[0059]

[0060] The integral equation in formula (1) is written in the form of discrete matrix measurement equation:

[0061]

[0062] That is, the mathematical model of metasurface antenna near-field imaging is:

[0063] g=Hσ+n (4)

[0064] In formula (4), is the echo measurement value received by the probe, M is the number of effective radiation patterns of the metamaterial aperture antenna, is the scattering coefficient of the scene target to be resolved, N is the number of discrete units in the scene, is the measurement noise term, Defined as a measurement matrix or measurement pattern.

[0065] like Figure 2 As shown, the step (2) specifically includes the following steps:

[0066] (2a) Based on the near-field imaging mathematical model of the metasurface antenna g = Hσ + n, a metasurface antenna computational imaging system is constructed using multiple fully connected layers and convolutional layers. Based on the existing metasurface antenna scene target echo measurement value g, an approximate scene target estimation result is obtained through the traditional MF method. The training dataset that constitutes the near-field MSACI-Net imaging network;

[0067] (2b) The near-field MSACI-Net imaging network consists of three fully connected layers and four convolutional layers. Each fully connected layer contains N neurons. When the echo measurement value g is input, the estimated value of σ is obtained after passing through three fully connected layers. After the reshape operation, The shape changes from N×1 to k x ×k y , where k x and k y Represents horizontal and vertical near-field scan points, and the size of the feature maps generated by all fully connected layers is consistent with the number of voxels to be imaged;

[0068] (2c) The fully connected layer is followed by four convolutional layers. The first and third convolutional layers use convolutional kernels to reduce network parameters, and the second and last convolutional layers use convolutional kernels to generate feature maps as the output of the network. Each of the four convolutional layers performs a ReLu zero-padding operation to keep the imaging system consistent with the size of the feature map.

[0069] (2d) After four convolutional layers, the reconstructed target image is finally obtained, completing the reconstruction operation of the entire image.

[0070] like Figure 3 As shown, the step (3) specifically includes the following steps:

[0071] (3a) The training data set containing N training samples is represented as {(σ1,g1),(σ2,g2),...,(σ N ,g N )}, find the error based on the training data set The smallest H, denoted as H f :

[0072]

[0073] In formula (5), σ=[σ1,σ2,...,σ N ], is the original scene target data of the entire near field, is the reconstructed scene target obtained by inverting the received echo measurement value, that is, in g = [g1, g2, ..., g N ], n and H are known, and it is calculated according to the near-field imaging mathematical model of the metasurface antenna as g = Hσ + n; where, is the echo measurement value received by the probe, M is the number of effective radiation patterns of the metamaterial aperture antenna, is the scattering coefficient of the scene target to be resolved, is the measurement noise term, Defined as a measurement matrix or measurement pattern;

[0074] (3b) Set the basic parameters of the near-field MSACI-Net imaging network, including the randomly initialized measurement matrix H, the random objective function f(·) with parameters, the maximum number of iterations T, the exponential decay rate of the moment estimate ρ1 = 0.9, ρ2 = 0.999, and the numerical stability constant ε = 10 -8 , the learning rate of the gradient optimization algorithm is δ = 10 -3 ;

[0075] (3c) Use the mean square error function MSE to calculate the loss L(H) between the network prediction value and the true value, that is:

[0076]

[0077] In the formula, σ i ∈σ, represents the imaging data of a target in the near field, represents imaging data of a near-field target calculated by a near-field imaging mathematical model of a metasurface antenna;

[0078] (3d) Use the Adam optimization algorithm to train the network and calculate the gradient of the measurement matrix H The weight coefficients of the convolutional layer are updated during forward and backward propagation, updating the first-order moment vector and the second-order moment vector Finally update the objective function parameters Until the end of the training cycle, the system's loss function tends to converge.

[0079] The step (4) specifically comprises the following steps:

[0080] (4a) Prepare the training dataset (g tr ,σ tr ) and the test set g t , set the maximum number of iterations T for training the near-field MSACI-Net imaging network, and randomly initialize the weight parameter W of each layer of the near-field MSACI-Net imaging network;

[0081] (4b) In each round of training, a random batch of g is drawn from the training dataset. tr and σ tr They are used as the input and output of the near-field MSACI-Net imaging network respectively; the MSE function is used as the loss function of the near-field MSACI-Net imaging network to train the weight parameter W of each layer in the near-field MSACI-Net imaging network;

[0082] (4c) After T rounds of iterations, the loss value of the near-field MSACI-Net imaging network approaches a minimum value, the near-field MSACI-Net imaging network training is completed, and the trained near-field MSACI-Net imaging network is saved;

[0083] (4d) The near-field MSACI-Net imaging network after training is used to perform high-precision imaging of the target. t Input the trained near-field MSACI-Net imaging network to obtain the reconstructed target σ t , reconstruct high-quality original target information in real time and with high precision.

[0084] Embodiment 1

[0085] In order to verify the superiority of the present invention in near-field metasurface antenna computational imaging, a set of simulation experiments are conducted to compare the present invention with the traditional method. The simulation test hardware platform parameters are shown in Table 1, and the software platform parameters are shown in Table 2:

[0086] Table 1 Hardware platform parameters

[0087] CPU Intel(R)i7-11700 Memory 32GB Graphics card model NVIDIAGeForceRTX 3070Ti Graphics card memory 8GB

[0088] Table 2 Software platform parameters

[0089] operating system Windows 10 64-bit Compiler Pycharm 2021.3.1 CUDA Version 11.3.1

[0090] Simulation results and analysis

[0091] A two-dimensional parallel plate waveguide mid-surface antenna is used, and the specific parameters are as follows: the antenna panel size is 250×250mm 2 , the dielectric constant is 3.66, the loss tangent is 0.003, the substrate thickness between the copper ground plane and the conductive copper metamaterial hole is 0.5mm, the upper conductor of the waveguide uses a 125×125cELC metamaterial resonator, and the Q factor of each resonator is between 50 and 60. The imaging experiment is based on the simulated mid-surface antenna radiation field pattern data and imaging scenarios. Among them, the working bandwidth of the antenna is 33 to 37GHz, the frequency sampling interval is 10mHz, and the size of the original pattern T is 400×(61×61).

[0092] During the simulation, the learning rate is 0.001 and the batch size is 64. The model is trained for 100 epochs. After each training epoch, 20% of the training data is randomly selected to generate a validation set to monitor the network performance. The training of each dataset network takes about 20 hours.

[0093] In order to test the noise resistance and robustness of the algorithm, additive noise of 0dB, 5dB and 10dB was added when generating the data set. In addition, the scene information sampling rate in the MSACI-MNIST and MSACI-FMNIST data sets was set to 0.1. The imaging reconstruction results are shown in Figure 4 As shown in Figure 3, the results show that under different signal-to-noise ratios and fixed scene compression ratios, the proposed algorithm can not only reconstruct common sparse objects, but also reconstruct extended objects. In addition, the proposed algorithm can also obtain high-quality reconstructed images.

[0094] In order to further evaluate the performance of the proposed algorithm, the experimental conditions under different compression ratios were set to 0.1, 0.05 and 0.025 respectively. The scattering coefficients of the two scene targets are between 0 and 1 and obey uniform distribution. The traditional matched filter reconstruction algorithm was also used for these two groups of targets, and a comparative experiment was conducted with the proposed algorithm under the same conditions. The experimental results of the two groups are shown in Figure 2. Figure 5 shown.

[0095] like Figure 6 , 7 As shown in Figures 8 and 9, MSE and PSNR are used to quantitatively evaluate the accuracy of imaging reconstruction and the quality of imaging results. As the scene information sampling rate increases, the mean square error of the proposed algorithm gradually decreases and is numerically lower than the MF and SBL algorithms in this standard. Compared with the commonly used CS method, the MSACI network algorithm shows better imaging capabilities because the end-to-end neural network can adaptively adjust the imaging error. Under the same scene information sampling rate, the image quality reconstructed by the algorithm is significantly improved in PSNR.

[0096] In addition, the time taken by the two algorithms to reconstruct the scene targets was recorded. The image reconstruction method based on deep convolutional neural network is easy to parallelize. Therefore, the reconstruction time of the MSACI network algorithm was recorded using CPU and GPU. Table 3 shows the average running time of the two algorithms for 100 rounds of testing on different targets. The proposed algorithm takes less time than the MF and SBL algorithms. Therefore, the proposed method is superior in imaging efficiency and can even achieve near real-time imaging reconstruction.

[0097] Table 3 Running time

[0098] MF algorithm 2.1756s SBL Algorithm 4.2184s MSACI Network (CPU) 0.2785s MSACI Network (GPU) 0.0584s

[0099] In summary, the present invention is based on a deep convolutional neural network to construct a computational imaging network for metasurface antennas, which can quickly realize near-field metamaterial aperture computational imaging. The advantage is that it can quickly and accurately image point-sized targets and more complex targets, and can effectively reduce the inevitable computational burden and the necessary large operating band; the DCN method is used to achieve near-field metasurface antenna rapid imaging, which can not only reconstruct ordinary sparse targets, but also reconstruct extended targets, and obtain high-quality reconstructed images; compared with traditional sparse reconstruction algorithms such as matched filtering MF and sparse Bayes SBL, the present invention does not require a long number of iterations, the imaging time is short, the complexity is low, and it can be realized that after the network training is completed, the network can be called to perform offline testing imaging results.

Claims

1. A fast computational imaging method for near-field metasurface antenna based on deep convolutional neural network, characterized in that: The method comprises the following steps in order: (1) The mathematical principle of near-field imaging of metasurface antennas is modeled. The probe of the metamaterial aperture antenna receives the echo measurement value scattered by the target scene to be detected, and a measurement matrix is ​​constructed to reflect the mapping relationship between the echo measurement value and the scattering coefficient distribution of the target scene. (2) A near-field MSACI-Net imaging network is constructed based on a convolutional neural network. The entire near-field MSACI-Net imaging network consists of a fully connected layer and a convolutional layer. The measurement values ​​collected by the receiving antenna and the estimation results of the MF method together constitute the training data set and test set of the MSACI-Net network. The near-field MSACI-Net imaging network is trained and the receiving matrix is ​​reconstructed. (3) Design a network training algorithm based on the near-field MSACI-Net imaging network, use the mean square error to calculate the loss between the network prediction value and the true value, and use the Adam optimization algorithm to minimize the loss of the near-field MSACI-Net imaging network by setting the hyperparameters of the number of iterations, exponential decay rate, and learning rate; (4) Use the measured field data to train the near-field MSACI-Net imaging network, and reconstruct high-quality original target information in real time and with high precision after training; The step (1) specifically comprises the following steps: (1a) The echo measurement value g received by the probe of the metamaterial aperture antenna contains all the scattering information of the near-field scene target. Under the first-order Born approximate scattering model, the mapping relationship between the echo measurement value g and the scene target scattering coefficient σ is described by constructing the measurement matrix H. When reconstructing the measurement matrix, it is necessary to calculate the linear phase terms introduced by the pitch and azimuth scene resolution units contained in each distance slice at different distance slices in the scene, and then arrange and combine them to form the total measurement matrix. (1b) The measurement matrix H is defined as: In formula (1), represents the vector pointing into free space, Indicates the scene The size of the radiation field at Indicates that the probe is in scene space The physical meaning of H is the size of the radiation field at the transmitting antenna U 0TX and receiving antenna U 0RX Aperture at the target location At , the product of the two radiation field sizes is in the form of: The integral equation in formula (1) is written in the form of discrete matrix measurement equation: That is, the mathematical model of metasurface antenna near-field imaging is: g=Hσ+n (4) In formula (4), is the echo measurement value received by the probe, M is the number of effective radiation patterns of the metamaterial aperture antenna, is the scattering coefficient of the scene target to be resolved, N is the number of discrete units in the scene, is the measurement noise term, Defined as a measurement matrix or measurement pattern; The step (2) specifically comprises the following steps: (2a) Based on the near-field imaging mathematical model of the metasurface antenna g = Hσ + n, a metasurface antenna computational imaging system is constructed using multiple fully connected layers and convolutional layers. Based on the existing metasurface antenna scene target echo measurement value g, an approximate scene target estimation result is obtained through the traditional MF method. The training dataset that constitutes the near-field MSACI-Net imaging network; (2b) The near-field MSACI-Net imaging network consists of three fully connected layers and four convolutional layers. Each fully connected layer contains N neurons. When the echo measurement value g is input, the estimated value of σ is obtained after passing through three fully connected layers. After the reshape operation, The shape changes from N×1 to k x ×k y , where k x and k y Represents horizontal and vertical near-field scan points, and the size of the feature maps generated by all fully connected layers is consistent with the number of voxels to be imaged; (2c) The fully connected layer is followed by four convolutional layers. The first and third convolutional layers use convolutional kernels to reduce network parameters, and the second and last convolutional layers use convolutional kernels to generate feature maps as the output of the network. Each of the four convolutional layers performs a ReLu zero-padding operation to keep the imaging system consistent with the size of the feature map. (2d) After four convolutional layers, the reconstructed target image is finally obtained, completing the reconstruction operation of the entire image; The step (3) specifically comprises the following steps: (3a) The training data set containing N training samples is represented as {(σ1,g1),(σ2,g2),...,(σ N ,g N )}, find the error based on the training data set The smallest H, denoted as H f : In formula (5), σ=[σ1,σ2,...,σ N ], is the original scene target data of the entire near field, is the reconstructed scene target obtained by inverting the received echo measurement value, that is, in g = [g1, g2, ..., g N ], n and H are known, and it is calculated according to the near-field imaging mathematical model of the metasurface antenna as g = Hσ + n; where, is the echo measurement value received by the probe, M is the number of effective radiation patterns of the metamaterial aperture antenna, is the scattering coefficient of the scene target to be resolved, is the measurement noise term, Defined as a measurement matrix or measurement pattern; (3b) Set the basic parameters of the near-field MSACI-Net imaging network, including the randomly initialized measurement matrix H, the random objective function f(·) with parameters, the maximum number of iterations T, the exponential decay rate of the moment estimate ρ1 = 0.9, ρ2 = 0.999, and the numerical stability constant ε = 10 -8 , the learning rate of the gradient optimization algorithm is δ = 10 -3 ; (3c) Use the mean square error function MSE to calculate the loss L(H) between the network prediction value and the true value, that is: In the formula, σ i ∈σ, represents the imaging data of a target in the near field, represents imaging data of a near-field target calculated by a near-field imaging mathematical model of a metasurface antenna; (3d) Use the Adam optimization algorithm to train the network and calculate the gradient of the measurement matrix H The weight coefficients of the convolutional layer are updated during forward and backward propagation, updating the first-order moment vector and the second-order moment vector Finally update the objective function parameters Until the end of the training cycle, the system's loss function tends to converge.

2. The method for fast computational imaging of near-field metasurface antenna based on deep convolutional neural network according to claim 1, characterized in that: The step (4) specifically comprises the following steps: (4a) Prepare the training dataset (g tr ,σ tr ) and the test set g t , set the maximum number of iterations T for training the near-field MSACI-Net imaging network, and randomly initialize the weight parameter W of each layer of the near-field MSACI-Net imaging network; (4b) In each round of training, a random batch of g is drawn from the training dataset. tr and σ tr They are used as the input and output of the near-field MSACI-Net imaging network respectively; the MSE function is used as the loss function of the near-field MSACI-Net imaging network to train the weight parameter W of each layer in the near-field MSACI-Net imaging network; (4c) After T rounds of iterations, the loss value of the near-field MSACI-Net imaging network approaches a minimum value, the near-field MSACI-Net imaging network training is completed, and the trained near-field MSACI-Net imaging network is saved; (4d) The near-field MSACI-Net imaging network after training is used to perform high-precision imaging of the target. t Input the trained near-field MSACI-Net imaging network to obtain the reconstructed target σ t , reconstruct high-quality original target information in real time and with high precision.

Images