Self-Supervised Sparse ISAR Imaging Using Untrained Convolutional Generator

In Inverse Synthetic Aperture Radar (ISAR), highquality imaging with sparse data remains challenging. Traditional Compressed Sensing (CS) methods rely on manual priors and optimization, while deep learning-based methods lack robustness in unseen scenarios. This paper proposes a novel approach using an untrained convolutional generator. An untrained CNN directly reconstructs complex ISAR images, with the image forwardprojected into the echo domain for self-supervised learning with sparse echoes. The method benefits from the inherent bias of untrained networks, producing images with prominent features and reduced noise. The self-supervised framework removes the need for dataset-specific training, enabling robust imaging across scenarios. Experiments show improvements of 41.6% and 19.5% in PSNR and SSIM over untrained CS methods.