A Two-Dimensional Hybrid Electromagnetic Reconstruction Scheme for Dielectric Objects Based on Generative Adversarial Network

Electromagnetic inverse scattering problems have been extensively studied due to their wide range of applications in various fields, including medical imaging, non-destructive testing, and remote sensing. However, these problems are inherently nonlinear and ill-posed, which poses a challenge to obtain accurate and efficient reconstructions using traditional methods. Direct reconstruction methods based on deep learning have been proposed to address the limitations of traditional methods, but the lack of a priori information makes it difficult to learn. In this paper, we propose a hybrid electromagnetic reconstruction scheme that combines the diffraction tomography (DT) algorithm and the Generative Adversarial Network (GAN) to solve electromagnetic inverse scattering problems. The DT algorithm based on Born approximation is used to reconstruct the rough image of the target scatterer, which is then used as the input of the generator network of the pixel-based GAN. This approach effectively utilizes a priori information to establish the mapping of the rough image to the target image, making the learning process easier. In addition, the attention mechanism incorporated in the GAN improves the accuracy of the generator network in reconstructing the relative permittivity distribution of target scatterers. The GAN consists of a generator network that generates fake images similar to the target images and a discriminator network that determines whether the input image is real or fake. The discriminator guides the generator to generate a more realistic image by back-propagating the gradient. The generator and discriminator networks are trained in an adversarial way until reaching a Nash equilibrium to learn the features of target scatterers, which enables the GAN to produce high-quality reconstructions. Numerical simulations are included to demonstrate the feasibility and efficiency of the proposed hybrid electromagnetic reconstruction scheme based on GAN.