Abstract:In electromagnetic inverse scattering, the goal is to reconstruct object permittivity using scattered waves. While deep learning has shown promise as an alternative to iterative solvers, it is primarily used in supervised frameworks which are sensitive to distribution drift of the scattered fields, common in practice. Moreover, these methods typically provide a single estimate of the permittivity pattern, which may be inadequate or misleading due to noise and the ill-posedness of the problem. In this paper, we propose a data-driven framework for inverse scattering based on deep generative models. Our approach learns a low-dimensional manifold as a regularizer for recovering target permittivities. Unlike supervised methods that necessitate both scattered fields and target permittivities, our method only requires the target permittivities for training; it can then be used with any experimental setup. We also introduce a Bayesian framework for approximating the posterior distribution of the target permittivity, enabling multiple estimates and uncertainty quantification. Extensive experiments with synthetic and experimental data demonstrate that our framework outperforms traditional iterative solvers, particularly for strong scatterers, while achieving comparable reconstruction quality to state-of-the-art supervised learning methods like the U-Net.

What problem does this paper attempt to address?

This paper attempts to address several key challenges in the electromagnetic inverse scattering problem: 1. **Unsupervised Framework**: Traditional deep - learning methods usually require a supervised framework, that is, both the scattering field and the target dielectric constant pattern need to be provided during training. However, these methods are very sensitive to changes in the scattering field distribution, which is a common problem in practical applications. This paper proposes a data - driven framework that is trained only with the target dielectric constant pattern and does not require scattering field data. This means that the model can be used in any experimental configuration and has stronger robustness. 2. **Multiple - Solution Estimation and Uncertainty Quantification**: Due to the high ill - posedness of the inverse scattering problem, a single point estimate is often insufficient to reflect the real physical situation. This paper introduces a Bayesian framework that can generate estimates of multiple dielectric constant patterns and perform uncertainty quantification, thereby providing more abundant information. 3. **Efficient Regularization**: The ill - posedness and non - linearity of the inverse scattering problem make effective regularization techniques crucial. This paper utilizes deep generative models (especially injective flows) to learn low - dimensional manifolds as regularizers to constrain the solution space and improve the reconstruction quality. ### Main Contributions of the Paper 1. **Data - Driven Framework**: A data - driven framework based on deep generative models for the electromagnetic inverse scattering problem is proposed. This framework only requires the target dielectric constant pattern for training and does not need scattering field data, which improves the generalization ability and robustness of the model. 2. **Bayesian Framework**: A Bayesian framework is introduced, which can generate estimates of multiple dielectric constant patterns and perform uncertainty quantification. This helps to better understand the diversity and reliability of solutions. 3. **Use of Injective Flows**: Injective flows are used as generative priors. By optimizing the latent codes in the low - dimensional latent space, dielectric constant patterns consistent with the measurement data are generated. The low - dimensional latent space of injective flows provides effective regularization and solves the highly ill - posed inverse scattering problem. ### Formula Summary - **Forward Scattering Model**: \[ \nabla\times(\nabla\times E_t(\mathbf{r})) - k_0^2\epsilon_r(\mathbf{r})E_t(\mathbf{r}) = i\omega\mu_0J(\mathbf{r}), \] where \(E_t\) is the total electric field, \(k_0=\omega\sqrt{\mu_0\epsilon_0}\) is the wavenumber of the background medium, \(J(\mathbf{r})=\chi(\mathbf{r})E_t(\mathbf{r})\) is the contrast current density, and \(\chi(\mathbf{r})=\epsilon_r(\mathbf{r}) - 1\) is the contrast. - **Data Equation**: \[ E_s = G_s\chi E_t+\delta, \] where \(E_s\) is the scattering electric field, \(G_s\) is the scattering matrix, and \(\delta\) is the measurement noise. - **Unified Forward Model**: \[ E_s = G_s\chi(I - G_d\chi)^{-1}E_i+\delta, \] where \(G_d\) is the state matrix and \(E_i\) is the incident electric field. - **Maximum A Posteriori Estimation**: \[ x_{\text{MAP}}=\arg\min_x\frac{1}{2}\|y - A(f(z))\|^2_2-\lambda\log p_X(f(z)), \] where \(y\) is the measurement data, \(A\) is the forward scattering operator, \(f(z)\) is the generative model, and \(\lambda\) is the regularization parameter. - **Bayesian Posterior Modeling**: \[ \psi^*=\arg\min_\psi\text{KL}(q_Z(z,\p

Deep Injective Prior for Inverse Scattering

Enhanced Deep Learning Approach Based on the Deep Convolutional Encoder–Decoder Architecture for Electromagnetic Inverse Scattering Problems

Physics Embedded Deep Neural Network for Solving Full-Wave Inverse Scattering Problems

Electromagnetic Inverse Scattering Based on Deep Learning

A Reconstruction Method of Electromagnetic Scattering Target Based on Diffusion Model

Unrolled Optimization with Deep Learning-based Priors for Phaseless Inverse Scattering Problems

Deep Learning: A Powerful Framework for the Real-Time Solution of Inverse Scattering Problems

A Priori Knowledge-Based Physics-Informed Neural Networks for Electromagnetic Inverse Scattering

GAN-driven Electromagnetic Imaging of 2-D Dielectric Scatterers

DeepNIS: Deep Neural Network for Nonlinear Electromagnetic Inverse Scattering

A Physics-Induced Deep Learning Scheme for Electromagnetic Inverse Scattering

An Improved Deep Learning Scheme for Solving 2-D and 3-D Inverse Scattering Problems

Towards Intelligent Electromagnetic Inverse Scattering Using Deep Learning Techniques and Information Metasurfaces

Physics-informed Unsupervised Deep Learning Framework for Solving Full-Wave Inverse Scattering Problems

A More General Electromagnetic Inverse Scattering Method Based on Physics-Informed Neural Network

A Direct Sampling-Based Deep Learning Approach for Inverse Medium Scattering Problems

Nonlinear Electromagnetic Inverse Scattering Using DeepNIS

A deep learning enhanced inverse scattering framework for microwave imaging of piece-wise homogeneous targets

Physics-guided Loss Functions Improve Deep Learning Performance in Inverse Scattering

Generalization Capabilities Of Deep Learning Schemes In Solving Inverse Scattering Problems

Deep Learning with Attention Mechanism for Electromagnetic Inverse Scattering