28 November 2023 in IEEE Transactions on Geoscience and Remote Sensing
This paper presents a novel early fusion framework (EFF) of deep learning (DL) for solving electromagnetic (EM) inverse scattering problems (ISPs) in real-time with high accuracy. The EFF integrates the scattered field encoder (SFE) and the backpropagation encoder (BPE). On the one hand, the nonlinear relationships constrained by the multiple scattering phenomenon, are introduced in the SFE, which guarantees better external scattering manifestations. On the other hand, the pre-reconstructed backpropagation approximation (BP) distributions provide prior information in ISPs via the BPE, which is in accordance with the internal scattering manifestations. The novel EFF concatenates the features extracted from the scattered field and the BP approximation, achieving more scattering manifestations in the solving process. Furthermore, the multiple scattering information and BP distributions regularize each other’s encoding process effectively, which achieves faster convergence and more stable inversion results. It is expected that the EFF exploits a general framework for solving ISPs, which can be adopted to different scenarios by importing different kinds of prior information according to the fusion approach. The superior veracity, stability and generalization ability of the EFF have been demonstrated by the synthetic and the experimental results.
30 November 2023 in IEEE Transactions on Antennas and Propagation
In this paper, a machine learning-assisted surrogate modeling (MLSM) method is presented to facilitate the circularly polarized (CP) antenna array synthesis with mutual coupling considered. In the premise that the vectorial active element pattern (Vec. AEP) of an array element is only susceptible to its surrounding region, the MLSM, primarily comprised of two deep neural networks (DNNs), is employed to predict the Vec. AEPs of all the array elements sequentially based on their corresponding subarray features. Subsequently, the co-polarization (CoP) pattern, cross-polarization (XP) pattern, and axial ratio (AR) of the CP array are computed accordingly. In this manner, the MLSM enables the fast performance prediction of arrays with arbitrary geometries and adds an extra degree of freedom, namely element position, to CP antenna array design. The trained MLSM is also employed to predict the performances of arrays with any given element numbers exploiting the ideology of heterogeneous transfer learning. To verify the effectiveness of the proposed framework, the MLSM is employed to predict the performance of arrays in serval CP array syntheses. Numerical results demonstrate that the proposed MLSM is a feasible and efficient method for the fast prediction of CP arrays with arbitrary geometries and any given element numbers.
29 April 2019 in IEEE Transactions on Antennas and Propagation
The coupling of two antennas can be described with mutual impedance. This paper proposes computing the mutual impedance by reaction theorem, which involves integrating vector products of near fields over a sphere surrounding the second antenna. We express the near fields of both antennas by the vector Spherical Harmonic Transform (SHT), evaluate the near fields of the first antenna on the eccentric sphere by a new FFT/interpolation method efficiently, and then obtain the mutual impedance by the reaction integrals. This new method has a computation complexity of O(N3), where N represents the truncation order in SHT. The comparison with the Higher Order Moment Method (HOMOM) indicates that the proposed method is a fast, accurate, stable and universal method to analyze the coupling of any two antennas.
23 June 2023 in IEEE Geoscience and Remote Sensing Letters
A robust hierarchical domain constraint (HDC) strategy is proposed for the multi-scale contrast objects reconstruction in this letter. The Monte Carlo method and point evaluation are adopted and modified for the classification of background and scatterers cells, the parameters of the background cells are set as the same as the actual background and removed from the inversion. The major work of the HDC strategy is to divide the inversion process into several stages, and the classification of cells is evaluated separately at each stage, making the classification reversible. And the HDC can effectively classify the inversion domain of low-contrast objects without negligence in the multi-scale contrast objects case, as well as the continuously varying inhomogeneous objects’ low-contrast parts. Another advantage of HDC is the threshold determining the classification of cells can be flexibly chosen. This robust constraint of the inversion domain compresses the solution space so that the ill-posedness of the inverse problem is mitigated, and better accuracy thus is achieved. In this letter, the HDC strategy is applied to the Variational Born Iterative Method (VBIM), numerical results validate the accuracy, robustness, and antinoise ability of the proposed VBIM-HDC.
26 February 2024 in IEEE Transactions on Geoscience and Remote Sensing
Based on the assumption that the pseudo-random characteristic of VBIM, the retrieved contrast parameter distribution in the discrete cells complies with the Gaussian mixture model (GMM). This article presents a constrained optimization framework that combines GMM and a hierarchical domain constraint (HDC) strategy for solving inverse scattering problems (ISPs). This GMM-HDC framework divides the whole inversion process into several stages, each stage operating within a reduced domain and subject to lower and upper bound constraints. The reduced inversion domain is determined by the GMM’s cluster analysis, thereby eliminating the need for manual threshold selection. And the HDC strategy offers a re-classification mechanism to avoid misclassification of scatterer cells. By narrowing the inversion domain, this approach effectively reduces the solution’s dimensionality, and mitigating the ill-posedness of ISPs. The bound constraint is determined by a hypothesis testing method. When there is prior information of contrast bound, the solution during iteration may not be within the exact bound due to the ill-posedness of ISPs. Therefore, the flexibility-bound constraint can provide more stability. When the prior information is not available, the determined upper bound is also guaranteed to converge probabilistically to the exact upper bound in case of degradation. In each iteration, a quadratic programming (QP) formula is introduced to solve the inverse subproblem in the VBIM, which enables the incorporation of the contrast range determined by the GMM-HDC framework. The formula includes both Tikhonov and total variation regularization terms. Several typical models, including synthetic and experimental measurement data, are presented to verify the performance of the proposed method.
20 January 2025 in IEEE Transactions on Antennas and Propagation
This article proposes an adaptive damping Anderson acceleration (ADAA) method that augments the variational Born iterative method (VBIM) for solving the electromagnetic inverse scattering problems (ISPs). Traditionally, the Anderson acceleration (AA) method accelerates convergence in fixed point iterations by extrapolating solutions from a linear combination of previous iterations. However, different from the fixed point iteration, whose operator is definite, the mapping operator in VBIM is approximate and non-contractive. A necessary modification is made by substituting the gradient-acceptance-based strategy in standard AA to the total-quantity-acceptance-based strategy. Additionally, considering the later work in AA demonstrates a damping factor boosts the convergence speed of AA, we present an adaptive damping strategy to further augment the inversion’s convergence rate. The dual enhancements significantly improve both the speed and robustness of convergence. From a mathematical perspective, the VBIM-ADAA method seeks an exact data equation solution within a subspace spanned by previous regularization solutions during each iteration. Furthermore, the robustness facilitated by VBIM-ADAA permits a broader range of choices for the regularization parameter. By leveraging QR factorization (via the Gram-Schmidt process), we analyze the VBIM-ADAA’s characteristics under different mapping operators. Extensive numerical experiments, involving both synthetic and real-world data, verify the superior convergence, robustness and better accuracy of the VBIM-ADAA compared to conventional VBIM. Moreover, numerical experiments demonstrate that the computational time consumed by the ADAA is negligible, which makes it a practical and efficient method for solving ISPs.