Three-dimensional Reconstruction of Space Targets from Multi-view ISAR Images Using Differentiable Voxel Reconstruction Network

Abstract

Inverse Synthetic Aperture Radar (ISAR) is an effective remote sensing technique to obtain valuable 3D information of targets such as satellites using multi-view ISAR images. The special imaging mechanism of ISAR makes the target features vary greatly with the view angles. The silhouette, although more robust than scattered point features, relies on accurate projection information for 3D reconstruction of the target. This paper introduces a differentiable voxel reconstruction network that uses a differentiable projection operator to guarantee the backward propagation of the neural network. The view angle is set as a learnable parameter, so that 3D reconstruction can be achieved even for silhouettes with view angle noise. Experiments on simulation data demonstrate that the proposed method are much better than other traditional silhouette-based 3D reconstruction methods under the view angle noise condition.