LGM-Det: A Lightweight Geometry-Aware Multiscale Detector for Oriented Ship Target Detection in SAR Images

Ship target detection is of great significance in Synthetic Aperture Radar (SAR) image applications. However, balancing detection accuracy and efficiency remains a significant challenge in practical SAR ship detection tasks. Ship targets in SAR images often exhibit significant multi-scale variations and arbitrary orientations, and are frequently embedded in complex background clutter. These characteristics demand robust geometric modeling and feature representation capabilities. Meanwhile, deployment in resource-constrained environments places higher demands on the lightweight design of detection models. To address these challenges, this paper proposes a lightweight geometry-aware multi-scale detector (LGM-Det) for oriented ship target detection in SAR images. Specifically, the backbone stage employs a Lightweight Feature Perception Backbone (LFP-Backbone), which incorporates Multi-scale Gated Context-Aware (MSGCA) module and Attention-based Intra-scale Feature Interaction (AIFI) module to enhance scale sensitivity and directional modeling capacity. In the feature fusion stage, we introduce an Efficient Long-Short Attention Coupling (ELSAC) module, which mitigates the imbalance between local and global semantic modeling inherent in traditional attention mechanisms by coupling long-range and short-range attention, thereby improving feature representation. In the detection head, we construct the Cross-Scale Task-Coupled Head (CSTC-Head), which leverages scale- wise normalization and independent batch normalization strategies to effectively address the inconsistency of multi-scale information and further enhance the model's perception of oriented ship targets. LGM-Det achieves AP50 scores of 98.8% on SSDD+ and 97.4% on RSDD-SAR, while maintaining a compact architecture with just 1.5 million parameters and 5.1 GFLOPs, demonstrating both high accuracy and efficiency.