RaNeRF: Neural 3-D Reconstruction of Space Targets From ISAR Image Sequences

Compared to 2-D inverse synthetic aperture radar (ISAR) images of a space target, its 3-D model can provide adequate details and accurate measurement parameters. However, it is challenging to tackle the problem of feature extraction and correlation during 3-D reconstruction of space targets purely based on radar image sequences, due to their lack of clear evidence in imaging similarity compared to optical images. To address this problem, this article proposes radar neural radiance fields (i.e., RaNeRF), which is a novel 3-D reconstruction method using only observed ISAR image sequences. First, the 3-D structure of a target is represented as a continuous 6-D function of space positions and viewing directions using a fully connected deep network. Second, the relationship between the 3-D structure and 2-D ISAR images of the target is constructed to enable differential rendering of ISAR images. Our overall pipeline can thus be trained using the discrepancy between the modulus of rendered and observed ISAR images in a purely self-supervised manner without 3-D supervision. Finally, the 3-D mesh model of the target can be retrieved from the learned density field via marching cube (MC). As a result, the proposed RaNeRF can directly reconstruct the 3-D structure of targets without explicit feature extraction and correlation of ISAR image sequences. Both quantitative and qualitative results verify the effectiveness of the proposed method. Compared to conventional baseline methods using point clouds, our reconstructed structure is more complete and accurate. In addition, the optimized model can synthesize ISAR images at novel observation direction, which can be used for downstream tasks, including data augmentation and target recognition.