AD-YOLO: An Adaptive Detection YOLO Architecture for Ship Detection in SAR Imagery

Abstract

Deep learning has demonstrated great potential in ship detection within synthetic aperture radar (SAR) imagery. However, challenges, such as background clutter, varying target sizes, and scale inconsistencies continue to impede performance. In this letter, we present AD-you only look once (YOLO), a novel detection framework designed to address these challenges. It integrates key components such as AD2SF and efficient shared multiscale detail head (ES2DH), alongside the contextual feature calibration for clutter removal (CFC_CRB) and scale fusion consistency guidance (SFC_G2) modules, to improve small-target detection in SAR images. The AD2SF module enhances fine-grained feature extraction and global context modeling, significantly reducing background clutter and improving detection accuracy. The ES2DH module replaces the standard detection head, utilizing shared convolution and detail-enhanced convolution (DEConv) to reduce computational complexity while maintaining detailed feature representations, which improves multiscale detection. The CFC_CRB and SFC_G2 modules resolve multiscale feature fusion inconsistencies, optimizing target alignment, particularly for P3 and P5 layers, thus boosting performance in complex environments. Extensive experiments on SSDD and HRSID datasets show that AD-YOLO achieves $\mathrm{mAP}_{50:95}$ scores of 73.0% and 68.7%, outperforming the YOLOv11 baseline by 1.3% and 1.4%, respectively. Moreover, AD-YOLO strikes an ideal balance between accuracy, speed, and computational efficiency, making it a highly effective and efficient solution for SAR ship detection.