Multi-baseline SAR 3D reconstruction of vehicle from very sparse aspects: A generative adversarial network based approach

Reconstruction of anisotropic targets using SAR from multiple observation aspects is an important research direction. Conventionally, to obtain reconstruction of high quality using compressed sensing (CS) or back-projection (BP) methods, it is essential to acquire signals from plenty of aspects. However, limited by the flight trajectory and other factors, only few aspects are available. Besides, massive signal data from abundant aspects brings huge burden in computation. Thus, a new framework to generate 3D reconstruction of anisotropic targets from few aspects with low computational cost is necessary. To tackle this problem, we propose a new framework, which combines conventional CS algorithm and neural network. The CS algorithm is an imaging module to transform data from frequency domain into spatial domain with high resolution in each aspect to generate incomplete outputs. The sparse-aspects-completion network (SACNet) based on GAN principle is originally introduced to predict integral structures from the incomplete results. To evaluate the effectiveness and robustness of our framework, the simulation data (Civilian Vehicle Demo) is used to train while the measured data (GOTCHA) is for validating. Extensive experiments are conducted under condition of {4,6,8,10} aspects to estimate the performance with respect to the number of aspects. The proposed framework achieves the highest IoU and the lowest BCE metrics compared with the conventional algorithms and the SOTA networks used in similar optical tasks in both simulation and measured datasets under various number of aspects, which proves the effectiveness and robustness of our framework. The source code is available at: https://gitee.com/WshongCola/sar_3D_multi_aspect.