基于MSAM-YOLOv5的内河航道船舶识别方法.pdfVIP

基于MSAM-YOLOv5的内河航道船舶识别方法

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(youonlylookonce)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.

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

M-YOLOv5forinlandwaterways

ShipdetectionmethodbasedonMSAN

AbstractAnalgorithmbasedonYOLOv5(youonlylookonce)wasproposedtosolvetheproblemofmissingdetectionwhen

unmannedsurfaceshipsidentifyingtargetsoninlandwaterwaysduetotheinfluenceofbackgroundcomplexityanddistribution

diversity.First,anattentionmoduleMSAM(multiscaleattentionmodule)wasproposed.TheMSAMattentionallyfusedthe

shallowfeaturemapwithalargeamountofspatialinformationandthedeepfeaturemapwithrichsemanticinformation,sothatthe

fusedfeaturemapcouldhavestrongerfeatures.Then,theinfluenceofdifferentpositionsofMSAMwasstudied.Finally,the

parametersoftheanchorwereoptimizedtomaketheshapeoftheanchormoreconsistentwiththeshapeofinlandships.

Experimentsonshipdatasetsshowthattherecalloftheproposedalgorithmincreasesby1.12%,thethreeindicatorsofmAP(mean

averageprecision)increaseby0.87%,5.00%and2.07%,respectively,andtheFPS(framespersecond)indexincreasesby3,the

misseddetectionratedecreases,andtheoveralldetectionaccuracyandthedetectionspeedareimproved.

sshipdetection;inlandwaterways;multiscalesattentionmodule;YOLOv5;attentionmoduleposition

Keywords

近年来，无论在民用领域还是在军用领域，无地提升性能，检测模型辅助无人船最大限度地规避

人船的战略地位都十分重要．无人船是自主性非风险．当无人船检测船舶时，漏检直接影响其后续

常高的设备，自主航行能力是无人船的核心能力之的避障决策和航行规划等任务