Scholarlink

<p>Ship detection presents significant challenges, including limited adaptability to complex maritime environments, the trade-off between lightweight design and accuracy, and the degradation of generalization caused by model compression. To address these issues, we propose ELSD-YOLO, a lightweight ship detection algorithm built upon the YOLO11n architecture. A GhostNet-based backbone is constructed to ensure efficient computation while preserving feature representation, depthwise separable convolutions are employed to accelerate inference, and the StarNet block is integrated to enhance nonlinear modeling, thereby improving feature extraction for small and densely distributed ships. In addition, a SimHead module is designed to eliminate redundant computations while maintaining channel-wise parallelism, achieving a balance between efficiency and accuracy. Structural simplification and quantization further enable deployment on resource-constrained edge devices for real-time operation. Experimental results on the SeaShips dataset demonstrate that ELSD-YOLO not only achieves a 1.1% improvement in mAP50-95 compared with the baseline YOLO11n, alongside an 11% reduction in model parameters and a 19% reduction in computational complexity, but also exhibits superior robustness in detecting small vessels and handling complex background interference. These results validate ELSD-YOLO as an efficient and accurate framework, offering strong technical support for maritime traffic safety.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ELSD-YOLO: an enhanced lightweight ship detection framework based on YOLO11n

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Xue Meng,
  • Zhenbo Bi,
  • Lei Jia,
  • Tianyuan Wang,
  • Xuejian Meng

摘要

Ship detection presents significant challenges, including limited adaptability to complex maritime environments, the trade-off between lightweight design and accuracy, and the degradation of generalization caused by model compression. To address these issues, we propose ELSD-YOLO, a lightweight ship detection algorithm built upon the YOLO11n architecture. A GhostNet-based backbone is constructed to ensure efficient computation while preserving feature representation, depthwise separable convolutions are employed to accelerate inference, and the StarNet block is integrated to enhance nonlinear modeling, thereby improving feature extraction for small and densely distributed ships. In addition, a SimHead module is designed to eliminate redundant computations while maintaining channel-wise parallelism, achieving a balance between efficiency and accuracy. Structural simplification and quantization further enable deployment on resource-constrained edge devices for real-time operation. Experimental results on the SeaShips dataset demonstrate that ELSD-YOLO not only achieves a 1.1% improvement in mAP50-95 compared with the baseline YOLO11n, alongside an 11% reduction in model parameters and a 19% reduction in computational complexity, but also exhibits superior robustness in detecting small vessels and handling complex background interference. These results validate ELSD-YOLO as an efficient and accurate framework, offering strong technical support for maritime traffic safety.