SASI-Net: A Sparse Adaptive SAR Imaging Network with Complex-Valued Mask-Aware Completion and Adaptive Truncated Penalty

Near-field millimeter-wave synthetic aperture radar (MMW-SAR) imaging provides high-resolution sensing capability for short-range target detection and characterization, but dense spatial sampling imposes considerable storage and processing burdens. To enable reliable imaging from sparsely sampled measurements, this paper proposes a sparse adaptive imaging network, termed SASI-Net, which sequentially combines complex-domain echo completion with physics-guided image-domain sparse refinement. In the first stage, a complex-valued mask-aware completion network (CMAC-Net) jointly exploits the zero-filled echo and its binary sampling mask to recover missing measurements while promoting consistency with the available observations. The completed echo is then transformed into an initial complex image through a norm-preserving imaging operator. In the second stage, an adaptive truncated penalty sparse network (ATPS-Net) unfolds a magnitude-domain nonconvex optimization model while explicitly retaining the phase of the initial complex image. The corresponding minimax concave penalty (MCP)-based proximal update suppresses weak and diffuse responses while reducing excessive shrinkage of dominant scatterers. In addition, the regularization strength is estimated from the current magnitude features at each unfolding stage, enabling input- and stage-adaptive sparse refinement without manual parameter tuning. Experiments using real-measured near-field MMW radar data from the public 3DRIED dataset demonstrate that CMAC-Net improves complex echo and phase recovery, while ATPS-Net enhances scattering concentration and target-background separation. Under sampling rates as low as 10%, SASI-Net preserves identifiable target structures and maintains robust reconstruction performance. Additional noise experiments further verify the effectiveness of the proposed refinement strategy under measurement perturbations. The current evaluation focuses on near-field two-dimensional planar-scan MMW-SAR, and the applicability of SASI-Net to conventional strip-map SAR remains to be further investigated. These results demonstrate the potential of SASI-Net for computationally efficient sparse-aperture near-field MMW imaging.