摘要
为适应复杂水面环境与资源受限的水面无人艇(unmanned surface vehicle,USV)平台,满足水面目标检测对高精度和轻量化算法的需求,本文提出了一种改进YOLOv11n的轻量级水面目标检测算法YOLO-WaterLite.首先,在主干网络中引入C3K2-SCG模块,该模块优化了特征提取过程,进一步避免特征冗余,同时增强了全局与局部特征的表达能力,保持了网络的轻量化特性.其次,针对多尺度特征融合问题,设计了一种多尺度特征聚合-扩散机制,提升了网络的多尺度上下文信息整合能力,从而提高检测精度.最后,提出联合任务动态检测头,通过共享特征提取器增强分类与定位任务的交互,显著提升模型对水面小目标的检测鲁棒性与精确性.实验结果表明,YOLO-WaterLite在水面目标检测数据集(water surface object detection dataset,WS-ODD)和水面废弃物检测数据集(FloW-Img)上的mAP@0.5分别较基准模型YOLOv11n提高了5.4%和2.8%,召回率分别提高了3.4%和3.9%.此外,YOLO-WaterLite的模型参数量仅为2.4 M,相较于其他主流轻量级检测算法,在性能与效率上均表现出显著优势.
To accommodate the complex surface environment and resource-constrained unmanned surface vehicle(USV)platforms,and to meet the demand for high-precision and lightweight algorithms in surface target detection,this paper proposes an improved lightweight surface target detection algorithm,YOLO-WaterLite,based on YOLOv11n.First,the C3K2-SCG module is introduced into the backbone network to optimize the feature extraction process,further eliminating feature redundancy while enhancing the representation of both global and local features,thus maintaining the lightweight nature of the network.Second,a multi-scale feature aggregation-diffusion mechanism is designed to address the issue of multi-scale feature fusion,improving the network’s ability to integrate multi-scale contextual information and consequently enhancing detection accuracy.Finally,a joint task dynamic detection head is proposed,which enhances the interaction between classification and localization tasks through a shared feature extractor,significantly improving the model’s robustness and accuracy in detecting small surface targets.Experimental results show that YOLO-WaterLite achieves a 5.4%and 2.8%improvement in mAP at 0.5 over the baseline model YOLOv11n on the water surface object detection dataset(WSODD)and FloW-Img,respectively,with recall rates improving by 3.4% and 3.9%.Additionally,YOLO-WaterLite has only 2.4 M parameters,demonstrating significant advantages in both performance and efficiency compared to other mainstream lightweight detection algorithms.
作者
冀越文
张涛
JI Yuewen;ZHANG Tao(School of Instrument Science and Engineering,Southeast University,Nanjing 210096,China;Key Laboratory of Micro-Inertial Instrument and Advanced Navigation Technology,Ministry of Education,Southeast University,Nanjing 210096,China;Advanced Ocean Institute of Southeast University,Nantong 226000,China)
出处
《全球定位系统》
2025年第4期20-30,共11页
Gnss World of China
基金
东南大学南通海洋高等研究院基金重点项目(KP202403)。
作者简介
冀越文(1999-),男,硕士,研究方向为无人载体视觉感知技术.E-mail:jyw1090285259@163.com;通信作者:张涛(1980-),男,教授,研究方向为惯性导航与自主水下航行器定位.E-mail:zhangtao22@seu.edu.cn。
