HF-Head: A Lightweight Head Design for SAR Small-Ship Detection with Algorithm–Hardware Co-Optimization

Synthetic aperture radar (SAR) ship detection remains challenging in complex maritime scenes because small targets usually exhibit weak contours and limited textures and suffer from strong background interference, while edge deployment further demands low complexity and high efficiency. To address these challenges, this paper proposes HF-Head, a lightweight detection head for SAR small-ship detection with algorithm–hardware co-optimization. Built on the YOLO framework, HF-Head reconstructs the detection head with a unified depthwise separable design to improve deployment regularity and reduce computational cost, and further introduces a lightweight dual-branch enhancement mechanism, where a high-frequency branch strengthens edge and local-detail representation and a semantic branch enlarges the receptive field for contextual modeling.Aspatially gated injection strategy is used for adaptive feature fusion, and a re-parameterized folding scheme maintains stronger training-time representation while preserving single-path inference at deployment. Experiments on HRSID, SAR-Ship, and SSDD show that HF-Head reduces parameters by 13.2% and GFLOPs by 28.0% while maintaining or improving detection accuracy; specifically, mAP@0.5:0.95 is improved by 6.5%, 3.9%, and 3.4%, reaching 0.754, 0.554, and 0.661, respectively. FPGA RTL evaluation of the P3-level head under a near-iso-DSP budget further demonstrates a 3.81× latency speedup and a 58.5% reduction in energy per frame, reaching 240.66 mJ, indicating that HF-Head achieves a favorable accuracy–efficiency trade-off for edge-oriented SAR ship detection.