Deep-Unfolded Fusion Optimization With Motion Correlation Learning for ViSAR Multi-GMTIm

Video synthetic aperture radar (ViSAR) has emerged as a critical technology for ground moving target imaging (GMTIm) due to its high-resolution and high-frame-rate monitoring capabilities. However, ViSAR multiple ground moving target imaging (multi-GMTIm) faces significant challenges: 1) targets exhibit spatially variant motion-induced defocusing, making it difficult to achieve simultaneous focused imaging for multiple targets and 2) it remains a critical yet unresolved problem to effectively leverage strong temporal motion correlations across consecutive ViSAR frames to enhance motion error compensation of targets. To address these issues, this article proposes a novel learning imaging framework integrating multiframe motion compensation and multitarget fusion optimization. First, we formulate the motion phase error (MPE) estimation of targets as a joint sparse recovery problem, where a long short-term memory (LSTM) network is innovatively embedded to learn multiframe motion dependencies, significantly improving estimation accuracy compared to conventional framewise processing. Subsequently, the compensated multitarget imaging in each frame is fused through a kernel-based consensus alternating direction method of multipliers (CADMM)-Net. This architecture is designed by unfolding iterative optimization procedures into a layer-fixed network with trainable parameters, which further promotes the computing time and reconstruction accuracy. Comprehensive experiments on simulated and real datasets validate the method’s superiority in terms of both efficiency and image focusing quality.