KD-ISAR: A Complex-Domain Multilevel Knowledge Distillation Framework for Robust Sparse-Aperture ISAR Imaging

Abstract

Sparse-aperture inverse synthetic aperture radar (ISAR) imaging suffers from severe underdetermined reconstruction under low sampling rates and residual phase errors, which is further aggravated by insufficient priors of lightweight networks for edge deployment. This letter proposes knowledge distillation (KD)-ISAR, a novel complex-domain multilevel KD framework. It adopts a dual-channel real–imaginary representation to preserve phase-related information. The core insight is that sparse-aperture reconstruction is fundamentally an ill-posed inverse problem, and the structural priors implicitly encoded by a high-SR teacher—target topology, scattering texture, and global radar cross section (RCS) distribution—can serve as multiscale regularization constraints to stabilize the underdetermined recovery at low SR. To this end, multilevel semantic field distillation (MSFD) transfers complementary priors across three response-field levels: shallow scattering texture, bottleneck target topology, and output scattering-energy distribution, providing hierarchical regularization for low-SR reconstruction. Experiments on simulated and Yak-42 measured data demonstrate competitive reconstruction performance, robustness to residual phase errors, and efficient student inference under the evaluated conditions.