基于MSAM-YOLOv5的内河航道船舶识别方法  

Ship detection method based on MSAM-YOLOv5 for inland waterways

作　　者：萧筝[1] 王继业 夏叶亮 XIAO Zheng;WANG Jiye;XIA Yeliang(School of Mechanical and Electronic Engineering,Wuhan University of Technology,Wuhan 430070,China;Wuhan Marine Electric Propulsion Device Research Institute,Wuhan 430070,China)

机构地区：[1]武汉理工大学机电工程学院,湖北武汉430070 [2]武汉船用电力推进装置研究所,湖北武汉430070

出　　处：《华中科技大学学报（自然科学版）》2023年第5期67-73,118,共8页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(51905397).

摘　　要：针对内河航道上无人船识别目标时受背景复杂性和分布多样性影响而存在漏检的问题,提出一种基于YOLOv5(you only look once)的算法.首先,提出一种注意力模块MSAM(多尺度注意力模块),可对带有大量空间信息的浅层特征图和带有丰富语义信息的深层特征图进行注意力融合,使得融合后的特征图具有更强的特征;然后,研究MSAM模块的不同位置的影响;最后,优化锚框参数,使得锚框形状更加符合内河船舶的形状.在船舶数据集上进行实验,结果表明:本算法的召回率提高了1.12%,三个mAP(平均精度均值)指标分别提高了0.87%,5.00%和2.07%,FPS(帧率)指标提高了3,漏检率降低,整体检测准确性和检测速度均得到提升.An algorithm based on YOLOv5(you only look once)was proposed to solve the problem of missing detection when unmanned surface ships identifying targets on inland waterways due to the influence of background complexity and distribution diversity.First,an attention module MSAM(multiscale attention module)was proposed.The MSAM attentionally fused the shallow feature map with a large amount of spatial information and the deep feature map with rich semantic information,so that the fused feature map could have stronger features.Then,the influence of different positions of MSAM was studied.Finally,the parameters of the anchor were optimized to make the shape of the anchor more consistent with the shape of inland ships.Experiments on ship data sets show that the recall of the proposed algorithm increases by 1.12%,the three indicators of mAP(mean average precision)increase by 0.87%,5.00%and 2.07%,respectively,and the FPS(frames per second)index increases by 3,the missed detection rate decreases,and the overall detection accuracy and the detection speed are improved.

关 键 词：船舶检测 内河航道 多尺度注意力模块 YOLOv5 注意力模块位置 

分 类 号：U675.79[交通运输工程—船舶及航道工程]