23 September 2022 in IEEE Transactions on Microwave Theory and Techniques
To alleviate the extremely intrinsical ill-posedness and nonlinearity of electromagnetic inverse scattering under high contrast and low signal to noise ratio (SNR), we propose a deep unfolding network based on generative adversarial network (GAN) under contrast source inversion (CSI) framework, termed UCSI-GAN. The method solves inverse scattering problems (ISPs) using end-to-end generating confrontation way by incorporating a physical model together with its iterative updating formulation into the internal architecture of GAN. First, the nonlinear iterative scheme is extended to a deep unfolding generator network, and the contrast source and contrast updates are mapped to each module of the generator network. Second, to stabilize the imaging process, we add a refinement network to each variable update. Finally, the discriminator network is employed to ensure the authenticity of reconstructed images. The generator network and discriminator network are alternately trained with a generative adversarial learning strategy to reconstruct the properties of the medium object. Numerical experiments demonstrated that the performance of UCSI-GAN is better than traditional CSI and state-of-the-art learning approaches under high contrast and low SNR condition.
27 February 2024 in IEEE Transactions on Microwave Theory and Techniques
Numerous learning-based methods have been proposed to address the challenging task of solving inverse scattering problems (ISPs), with the requirement of ample labeled training datasets. To overcome the difficulty of collecting fully labeled datasets, we present a physically guided deep unfolding parallel network with self-supervised learning strategy to solve nonlinear ISPs, called SSISP-Net. The proposed method incorporates a physical model and the iterative optimization algorithm of ISPs into the network architecture and, thus, forms a deep unfolding network. First, here, two subsets of scattered fields are generated from fully measured scattered field data using receiver antenna positions. Second, we construct a parallel network by using two deep unfolding networks and then train this parallel network using only these two subsets of the scattered field. Finally, to constrain the two unfolding networks in the parallel network, we design two scattered field crossing losses and a contrast difference loss in the network. In this way, the reconstructed model can be properly trained using only scattered field data. The numerical and experimental results demonstrate that the proposed self-supervised learning method outperforms traditional and several other supervised learning methods under similar conditions. Moreover, it achieves competitive reconstruction performance compared with supervised learning methods known for their superior reconstruction capabilities.