Meta-Learning-Assisted Untrained Neural Network for Electromagnetic Inverse Scattering Problems

Abstract

Untrained neural network (UNN) has shown promising potential for solving inverse scattering problems (ISPs) with high flexibility and no need of training data. However, iterative optimization of UNN model parameters is time-consuming, and its reconstruction quality also strongly depends on the loss constraints that guide the optimization. In this article, a meta-learning strategy is introduced to obtain proper model parameter initialization, which can accelerate the convergence of an untrained deep unrolling network of subspace optimization, called SOM-Net. The untrained SOM-Net (uSOM-Net) equipped with the meta-learned initialization is referred to as Meta-uSOM. In addition, an edge-preserving total variation (EPTV) loss is introduced to enhance the reconstruction of Meta-uSOM by protecting edges from oversmoothness in conventional TV loss. The superiority of the proposed method is validated on both synthetic and experimental data, which demonstrate a significant improvement in the convergence and reconstruction quality of existing UNNs.