HILoF-DETR: A Lightweight Framework for SAR Ship Detection with Spatial Frequency Enhancement and Dynamic Alignment

Synthetic Aperture Radar (SAR) ship detection faces multiple challenges. The inherent speckle noise and low resolution of SAR images cause the ship features to become blurred. Meanwhile, ship targets in nearshore scenes are often mixed with complex backgrounds such as waves and islands, further interfering with effective detection. The traditional pyramid feature fusion method suffers from multi-scale semantic gaps and spatial misalignment, leading to an increase in the missed detection rate of small targets. The existing models have high computational complexity and are difficult to meet the real-time requirements of edge devices. To this end, this paper proposes an efficient detection framework for High Low Frequency Detection Transformer(HiLoF-DETR): a lightweight MobileNetV4 backbone network is used to achieve fast feature extraction, a High Low Frequency Encoder (HiLoF-Encoder) is designed to suppress noise and enhance details, and a multi-scale alignment mechanism guided by local similarity in FreqFusion is introduced to eliminate feature fusion bias. The experimental results show that HiLoF-DETR achieves AP50 accuracy of 96.7% and 90.5% on the SSDD and HRSID datasets, respectively, while the model parameter size is only 16.8M, achieving a balanced optimization of detection accuracy and computational efficiency.