Shadow-Aware Moving Target Detection in ViSAR: A Multiscale CNN–Transformer Hybrid Detection Framework

Based on the all-weather imaging capability of synthetic aperture radar (SAR), video synthetic aperture radar (ViSAR) enables dynamic ground target monitoring through high-frame-rate continuous observation. Since moving targets' shadows remain stable and unaffected by defocusing in ViSAR images, this property is widely leveraged for ViSAR moving target detection. However, ViSAR moving target detection faces challenges including low contrast in moving targets' shadows, complex background clutter interference, and the difficulty of simultaneously detecting small and medium-sized shadow targets. Existing deep learning-based detection methods for ViSAR struggle to address these challenges effectively. To overcome these limitations, we propose a multiscale CNN–Transformer hybrid detection framework, which consists of three key components. Firstly, in the pre-processing stage of ViSAR images, an improved low-rank representation algorithm is used to suppress background clutter and stationary targets' shadows, and enhance the contrast of moving targets' shadows. Secondly, the CNN's feature representation capability for small and medium-sized shadow targets is enhanced by improving the backbone and designing a novel feature pyramid network structure. This addresses the challenge of detecting small and medium-sized shadow targets simultaneously. Finally, within the Transformer architecture, a novel proposal generation module and a contrastive denoising training strategy are integrated into the Deformable Transformer to mitigate ambiguous semantic encoding and accelerate convergence in DETR- like detectors. Experimental results on the real-world ViSAR dataset demonstrate that the proposed framework achieves state-of-the-art performance.