轻量级注意力机制与跨尺度融合的船舶目标检测  

Ship Object Detection with Lightweight Attention Mechanism and Cross-Scale Fusion

作　　者：李冬琴 彭琪 吴洋 LI Dongqin;PENG Qi;WU Yang(School of Naval Architecture and Intelligent Manufacturing,Jiangsu Maritime Institute,Nanjing 211000,China;School of Marine and Offshore Engineering,Jiangsu University of Science and Technology,Zhenjiang 212000,China)

机构地区：[1]江苏海事职业技术学院船舶与智能制造学院,南京211000 [2]江苏科技大学船舶与海洋工程学院,江苏镇江212000

出　　处：《电光与控制》2025年第4期52-57,共6页Electronics Optics & Control

基　　金：国家自然科学基金(52101315);江苏海事职业技术学院科研启动基金(2025BSKY01)。

摘　　要：针对船舶上算力资源有限导致的检测算法速度慢和检测率低问题,基于YOLOv5s算法,提出了轻量级注意力机制与跨尺度融合的船舶目标检测算法。在主干网络中引入轻量级注意力机制SimAM,并与颈部网络跨尺度进行融合,提升算法的检测精度;引入轻量化卷积C3Ghost和GhostConv降低检测算法的参数量,实现船舶的实时检测;对于边界框回归损失,采用自适应参数提高锚框的适应性以及鲁棒性。最后,与其他流行算法在SeaShips数据集上进行对比和消融实验,结果验证了所提算法的有效性。A lightweight attention mechanism and cross-scale fusion based ship target detection algorithm is proposed to address the issues of slow detection speed and low detection rates caused by limited computing power resources onboard.Based on YOLOv5s algorithm,a lightweight attention mechanism of SimAM is introduced into the backbone network and fused cross-scale with the neck network,thereby improving the detection accuracy of the algorithm.Lightweight convolutions of C3Ghost and GhostConv are incorporated to reduce the parameters of the detection algorithm,enabling real-time ship detection.For bounding box regression loss,adaptive parameters are employed to enhance the adaptability and robustness of anchor box.Finally,comparative and ablation experiments with mainstream algorithms are conducted on the SeaShips dataset.The experimental results validate the effectiveness of the proposed algorithm.

关 键 词：船舶检测 YOLOv5s 轻量级注意力机制 轻量化卷积 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]