A Learning-Based Variational Backpropagation Method for Inverse Scattering Problems

Abstract

The fusion of deep learning techniques with conventional methods has garnered significant attention within the field of electromagnetic inverse scattering. The utilization of a traditional noniterative method for acquiring a presolution, followed by an enhancement procedure via neural networks, presents notable benefits such as simplicity and fast computational speed. However, the accuracy of this approach is usually impacted by the precision of the presolution, especially for strong scatterers. To alleviate this limitation, this research introduces a novel learning-based variational backpropagation method (VBPM). Through the utilization of variational operations, the proposed method refines the initial induced current obtained by the backpropagation (BP) method. Subsequently, an appropriate neural network is constructed to establish the relationship between the refined presolution and the true solution. Compared with the BP scheme (BPS) without variational operations, the proposed approach effectively enhances the solution accuracy with almost the same inversion time.