Radar Imaging With Highly Sparse Range Profiles Based on Data Completion Neural Network

Data sparsity is the dilemma that high-resolution imaging radar often encounters in practice. Recently, sparse imaging algorithms based on compressive sensing (CS) theory have emerged as commonly adopted methods to address this problem. However, CS-based methods are inherently constrained by the assumption of scene sparsity, which limits further improvements in the quality of image reconstruction. Moreover, these methods are proven ineffective when confronting with highly sparse data scenarios. In this article, we propose a sparse imaging framework that comprehensively utilizes both the data generation capability of deep neural networks (DNN) and a noncoherent tomography algorithm with the more relaxed condition on azimuthal sampling. First, a DNN for data completion is proposed to effectively extrapolate limited-view radar range profiles, generating full-view data with high quality. Subsequently, by incorporating a fast noncoherent imaging algorithm, our method could effectively reconstruct the target image using the completed data from the previous step. Experimental results demonstrate that the proposed approach excels in reconstructing optimal images from highly sparse data (5% degree of sparsity), where other traditional methods even fail. The structure similarity index and image correlation index of reconstructed images increased by 63.8% and 263%, respectively.