Semantic information-assisted 3D reconstruction method for satellite-borne multi-view SAR

To address the problem of low three-dimensional (3D) reconstruction accuracy of ship targets caused by their complex structure and scattering characteristics in spaceborne synthetic aperture radar (SAR) images, a semantic-assisted multi-view SAR 3D reconstruction technology is proposed. This technology overcomes the limitation of the fundamental matrix in expressing the relationship between targets and projection points by deriving an intrinsic matrix model. Additionally, a multi-level Laplacian Gaussian Blob detection algorithm is adopted to identify significant feature points in SAR images, and the normalized cross-correlation algorithm is used to perform image registration, effectively addressing the interference from side lobes. After completing image segmentation and feature extraction, the extracted semantic information is used to systematically eliminate incorrect reconstruction points, thereby enhancing the quality of ship target 3D reconstruction. Experimental results show that the relative error of obtained ship structure result is under 4%, which demonstrates the effectiveness of the proposed method.