A Lightweight Attention-Driven Multi-Scale Feature Fusion Network for Ship Detection in UAV Aerial Imagery

Ship detection in UAV aerial imagery is critical for maritime surveillance and coastal security. However, dynamic sea clutter, large intra-class scale variance, and limited onboard computational resources pose significant challenges to existing detectors. This paper proposes EBW-YOLO, a lightweight detection network built upon YOLOv8n, with improvements across feature extraction, fusion, and regression. Concretely, the Efficient Multi-Scale Attention (EMA) module enhances discriminative feature representation against cluttered backgrounds; the Bi-directional Feature Pyramid Network (BiFPN) replaces PANet for adaptive multi-scale feature aggregation; and the Wise-IoU (WIoU) loss refines bounding box regression by suppressing low-quality samples. Experiments on a UAV maritime dataset demonstrate that EBW-YOLO achieves 82.6% mAP@0.5, outperforming YOLOv8n by 4.3 pp, while maintaining only 3.2M parameters and 8.2 GFLOPs, confirming its practicality for real-time UAV deployment.