Three-Dimensional Geometry Reconstruction Method from Multi-View ISAR Images Utilizing Deep Learning

Abstract

The three-dimensional (3D) geometry reconstruction method utilizing ISAR image sequence energy accumulation (ISEA) shows great performance on triaxial stabilized space targets but fails when there is unknown motion from the target itself. The orthogonal factorization method (OFM) can solve this problem well under certain assumptions. However, due to the sparsity and anisotropy of ISAR images, the extraction and association of feature points become very complicated, resulting in the reconstructed geometry usually being a relatively sparse point cloud. Therefore, combining the advantages of the above methods, an extended factorization framework (EFF) is proposed. First, the instance segmentation method based on deep learning is used for the extraction and association of a number of key points on multi-view ISAR images. Then, the projection vectors between the 3-D geometry of the space target and the multi-view ISAR images are obtained, using the improved factorization method. Finally, the 3D geometry reconstruction problem is transformed into an unconstrained optimization problem and solved via the quantum-behaved particle swarm optimization (QPSO) method. The proposed framework uses discretely observed multi-view range–Doppler ISAR images as an input, which can make full use of the long-term data of space targets from multiple perspectives and which is non-sensitive to movement. Therefore, the proposed framework shows high feasibility in practical applications. Experiments on simulated and measured data show the effectiveness and robustness of the proposed framework.