Aggregated-Refined Feature Pyramid Network for Small Vehicle Target Detection in Monostatic/Bistatic SAR Images

Spaceborne and airborne synthetic aperture radar (SAR) systems have now reached a highly mature stage of development, they still encounter challenges in real-time response speed, operational costs, and low-altitude detail capture. To address these issues, we have independently developed a low-cost and highly flexible monostatic MiniSAR system, followed by a bistatic MiniSAR system. Leveraging these systems, we conducted the imaging mission over small ground vehicle targets. However, arbitrary observation angles result in high intra-class variation and low inter-class variation among small ground vehicle targets, compounded by complex background clutter in SAR images. These two issues make precise detection of small ground vehicle targets particularly challenging. To overcome these challenges, this paper proposes a novel small target detection method based on an enhanced feature pyramid network (FPN). First, spatial deformable convolution is incorporated into the backbone network to improve feature extraction capability for vehicle targets of varying shapes. Second, we propose an aggregated-refined feature pyramid network (AR-FPN) that adaptively aggregates multi-level semantic information and refines them via global context modeling, significantly enhancing feature representation. Finally, to address the small-size characteristics of ground vehicle targets, we introduce an additional detection head for small targets and adopt the Normalized Wasserstein Distance (NWD) for loss function optimization. The effectiveness of the proposed algorithm is validated on our self-developed monostatic/bistatic SAR datasets, and its generalization capability is further confirmed on public dataset. Experimental results demonstrate that our method outperforms other state-of-the-art (SOTA) detectors in detecting small vehicle targets in SAR images.