DIP-SA-ISAR: Training-Free Sparse-Aperture ISAR Imaging via Deep Image Prior
Abstract
Sparse-aperture inverse synthetic aperture radar (SA-ISAR) imaging techniques enable reconstruction of ISAR images of a target with sparse azimuth observations. The existing SA-ISAR imaging approaches mainly include sparse constraint-based optimization methods and end-to-end neural network-based imaging methods. However, sparse constraint-based optimization approaches often exhibit weak enforcement of sparse priors, whereas end-to-end imaging networks are hindered by limited training data and poor generalization capability, both of which ultimately compromise reconstruction quality. To overcome these limitations, we propose a deep image prior (DIP)-based ISAR imaging framework for high-precision SA-ISAR reconstruction that requires no training data. The method leverages the implicit priors embedded in neural network architectures under the DIP paradigm. It incorporates a composite loss function that combines logarithmic norm and total variation regularization to enhance the fidelity of reconstruction. Extensive simulations and real-data experiments demonstrate that the proposed DIP-SA-ISAR achieves high-quality reconstruction of ISAR images across diverse sampling schemes and noise conditions.