Inverse Scattering via Cascaded Neural Network

Abstract

Traditional nonlinear solutions for inverse scattering problem have the drawback of high computational complexity. Its approximated alternative, back-projection (BP) algorithm, can achieve a good trade-off between imaging quality and complexity. However, in the case of limited frequency samples, BP suffers from the image quality degeneration. In order to achieve high-resolution imaging, this work turns to deep learning (DL) based approach and proposes an end-to-end cascaded neural network structure, namely a convolutional neural network (CNN) followed by a UNet network. Firstly, the equivalence between the fully connected network and the BP algorithm is derived. Secondly, to increase the learning ability of the network and avoid overfitting, a CNN is used to replace the fully connected network. By directly focusing the raw scattered radar echoes using the network, a coarse radar imagery of the region under investigation can be obtained. Then, a UN et network is further cascaded to suppress the clutter and improve the image quality of the coarse focused radar imagery. Finally, EM simulations using the MINST dataset are conducted to train the proposed network. The results show that the reconstruction using the trained cascaded network outperforms the BP algorithm under the condition that the computational complexity of the proposed algorithm and the BP algorithm is close. A better focusing performance is achieved as expected.