Enhancing small target detection for marine ships using an MS-YOLOv10s-based efficient recognition method

Accurately detecting small maritime vessels under complex meteorological conditions remains a significant challenge. To address this, we propose MS-YOLOv10s, a real-time detection model. A Shifted-Window Multi-Head Self-Attention (SW-MSA) module captures multi-scale features in complex backgrounds, enhancing adaptability and small-target detection. The C3k2 module efficiently extracts multi-layer features across receptive fields. Integrated with YOLOv10, the model leverages C2f and SCDown for fine-grained extraction and C3k2 for hierarchical feature fusion. We also constructed a dataset of 34,000 small-vessel images across diverse spatiotemporal conditions, including variations in time, location, lighting, and angle. Experimental results show that MS-YOLOv10s achieves 96.5% mAP, a 6.1% improvement over YOLOv10s, and outperforms state-of-the-art models in both accuracy and inference speed..