GLAM-YOLO: a global-local attention and multi-scale fusion enhanced YOLOv10 framework for robust SAR object detection

Abstract

Synthetic aperture radar (SAR), with its all-weather and all-day imaging capability, has become indispensable in military reconnaissance and remote sensing applications. However, SAR object detection remains challenging due to strong speckle noise, the frequent presence of small-scale targets and sensitivity to target rotation. To address these issues, this study proposes an enhanced detection framework, termed GLAM-YOLO, based on YOLOv10. The framework integrates a Global-Local Attention Module for speckle suppression and feature enhancement, a SAR-Aware Multi-scale Difference Fusion Module (SA-MDFM) to strengthen small-target representation and a Weighted IoU loss to improve localisation accuracy for objects with diverse shapes and aspect ratios. Experiments were conducted on two high-resolution public datasets, SAR-Aircraft-1.0 and HRSID, which involve diverse object scales, orientations and cluttered backgrounds. The results demonstrate that GLAM-YOLO outperforms state-of-the-art detectors, including Faster R-CNN, EfficientDet, DETR and YOLOv10-n. Specifically, it achieves an mAP@0.5 of 0.953 on SAR-Aircraft-1.0 and 0.913 on HRSID, along with improved precision, recall and F1-scores, highlighting its robustness against noise and clutter. Ablation studies further confirm the complementary effects of the proposed modules. Overall, GLAM-YOLO achieves a favourable accuracy-efficiency trade-off among compared SAR detectors, offering a reliable solution for robust and accurate SAR object detection in complex environments.