MASS-Net: Multiaspect SAR Stereo Network for Target 3-D Reconstruction

The reconstruction of the 3-D structure of synthetic aperture radar (SAR) targets is a hot and difficult issue in the field of SAR. Conventional 3-D reconstruction methods based on 2-D SAR images do not consider the inherent characteristics of SAR imaging such as geometric deformation, overlap, and occlusion, and can only reconstruct simple and regular targets. To address this, we propose a convolutional neural network (CNN)-based SAR 3-D reconstruction method called multiaspect SAR stereo network (MASS-Net). Our network is an end-to-end deep learning architecture that can automatically complete dense matching among multiaspect SAR images and calculate the height to obtain height maps by learning prior knowledge. In the network, a feature extractor based on CNN is constructed to extract features from SAR images, which can extract high-dimensional features of 2-D SAR images, and help to capture neighborhood information and overcome the influence of geometric deformation and occlusion. Then, a differentiable SAR projection relationship is established to construct a cost volume that includes the differences in multiaspect image features. This projection relationship ensures the overall differentiability of our pipeline. Meanwhile, the encoding and decoding architecture based on 3-D CNN is utilized to achieve regularization and regression to generate height maps. Finally, we use multiaspect height maps to construct a dense 3-D point cloud of the target. These make MASS-Net efficient and effective. Compared with traditional methods, our method can address issues such as distortion and occlusion, and efficiently reconstruct dense and accurate 3-D point clouds of complex targets. Simulation experiments and actual measurement experiments have been conducted to verify our proposed method.