Inverse Scattering Solver Based on Deep Neural Network with Total-Variation Regularization

Deep neural network (DNN) techniques have been applied to solve nonlinear electromagnetic inverse scattering problems (ISP) and shown potentials in superior imaging performances over traditional methods. In DNN ISP solvers, data features are learned from imaging samples to accurately represent the contrast distribution through the optimization of hyper-parameters of neural network, while the loss function which often quantifies discrepancies between the predicted imaging result and the ground truth is targeted to be minimized in the training process. In this paper, to further make use of the (piece-wise) smoothness of the contrast distribution, the desired solution is regularized by additionally penalized its total variation (TV) in the loss function. Effects of the TV regularization on the DNN ISP solver are studied and summarized with various simulation settings and the improvement in imaging performance is verified based on experiment data.