12 October 2020 in IEEE Sensors Journal
Deep learning based inverse scattering (DL-IS) methods attract much attention in recent years due to advantages of fast speed and high-quality reconstruction. The loss functions of neural networks in DL-IS methods are commonly based on a pixel-wise mean squared error (MSE) between the reconstructed image and its reference one. In this paper, we introduce a structural similarity (SSIM) loss function to combine with the MSE loss for reconstructing dielectric targets under a DL-IS framework. The SSIM loss imposes a further regularization on the target at the perceptual level. Numerical tests for both synthetic and experimental data verify that this new perceptually-inspired loss function can effectively improve the imaging quality and the generalization capability of the trained model.
30 June 2021 in IEEE Transactions on Computational Imaging
In this work, we introduce a learning-based method to achieve high-quality reconstructions for inverse scattering problems (ISPs). Particularly, the proposed method decouples the full-wave reconstruction model into two steps, including coarse imaging of dielectric profiles by the back-propagation scheme, and a resolution enhancement of coarse results as an image-to-image translation task solved by a novel perceptual generative adversarial network (PGAN). A perceptual adversarial (PA) loss, which is defined as a perceptual loss for the generator network using hidden layers from the discriminator network, is employed as a structural regularization in PGAN. The PA loss is further combined with the pixel-wise loss, and also possibly the adversarial loss, to enforce a multi-level match between the reconstructed image and its reference one. The adversarial training of the generator and discriminator networks ensures that the structural features of targets are dynamically learned by the generator. Numerical tests on both synthetic and experimental data verify that the proposed method is highly efficient and it achieves superior imaging results compared to other data-driven methods. The validation of the proposed PGAN on ISPs also provides a fast and high-precision way for solving other physics-related imaging problems.
10 January 2022 in IEEE Transactions on Antennas and Propagation
In this article, a unified learning-based approach is introduced to solve inverse scattering problems (ISPs) with mixed boundary conditions (BCs). The scattering behavior of hybrid dielectric and perfect electric conductors (PEC) scatterers is modeled by the T-matrix method. A rough image of the zero-order T-matrix coefficients for unknown scatterers is first reconstructed by the backpropagation (BP) method, which is then refined by an attention-assisted pix2pix generative adversarial network (GAN). The spatial attention mechanism is utilized to enforce the generator network to learn salient features of the unknown scatterers instead of the background. The adversarial training of the generator and the discriminator further enables the reconstructed image to be constrained by high-level features of reference scatterers. Numerical tests on both synthetic and experimental data verify the superior performance of the proposed method for ISP reconstructions with hybrid scatterers. It effectively expands the application scope of learning-based ISP methods to reconstruct scatterers without knowing the BCs of scatterers in advance.