AMSDet: Anisotropic Multiscale Detector With Cross-Interactive Rotated Head for SAR Ship Object Detection

Abstract

Synthetic aperture radar (SAR) ship detection plays a vital role in maritime surveillance, yet it remains challenging due to extreme scale variations, anisotropic scattering characteristics, and orientation ambiguity in complex near-shore environments. Existing detectors relying on fixed convolutional kernels and independent detection heads often struggle to capture directional structural features of ships while incurring heavy computational overhead. To address these limitations, this letter proposes AMSDet, an anisotropic multiscale detector built upon an oriented bounding box (OBB) framework. In particular, a multireceptive downsampling block (MRDB) replaces standard strided convolutions in the backbone, employing parallel depthwise convolutions with diverse kernel sizes to preserve fine-grained multiscale features. A C2f-strip module embeds asymmetric strip convolutions into the C2f architecture, enabling the network to capture anisotropic directional scattering patterns inherent to ship targets. Finally, a multiscale cross-interactive rotated head (MSCI-RHead) is devised, where a unified lightweight convolutional stack is shared across all detection scales, enabling cross-scale parameter interaction while substantially reducing model complexity. Extensive experiments on RSDD and SRSDD datasets demonstrate that AMSDet achieves 96.4% and 49.6% mAP@0.5. Remarkably, these gains are attained with merely 8.93 M parameters and 25.0 GFLOPs, which corresponds to a 21.8% reduction over the baseline while delivering 85.31 FPS, validating its effectiveness for efficient and accurate rotated ship detection in SAR imagery.