基于SimAM注意力机制的近岸船舶检测方法.pdfVIP

基于SimAM注意力机制的近岸船舶检测方法

摘要:基于视觉图像的船舶目标检测中由于图像背景复杂,无关干扰较多,导致船舶目标检测的难度增大。并且多类别船

舶检测数据集现有数量较少且存在样本不均衡的问题使得船舶目标检测性能较低。针对复杂背景干扰检测,本文通过引入

SimAM注意力机制对YOLOv3模型进行改进,利用该机制加强船舶目标在提取特征中的权重并抑制背景干扰权重,从而提升

模型检测性能;同时,采用强实时数据增强以改善样本尺度分布不均衡的问题,结合迁移学习提升在样本数量受限情况下的

船舶检测精度。提取特征的可视化结果显示改进模型对无关背景特征干扰进行了抑制,增强了模型对于船舶特征的提取能

力。在SeaShips数据集上,提出的改进模型在不引入额外可学习参数的情况下mAP.5、mAP.75分别达到了96.93%、

71.49%,检测速度达到了66frame/s,在检测精度与运行效率方面保持了均衡。与Saliency-awareCNN、eYOLOv3相比更有效

地优化了目标特征,使得mAP.5分别提高了9.53%、9.19%。改进模型在新加坡海事数据集上在船舶类型目标检测的mAP.5

达到了81.81%,验证了模型具有较好的泛化能力。

关键词:船舶检测;注意力机制;数据增强;改进YOLOv3;迁移学习

NearshoreshipdetectionmethodbasedonSimAMattentionmechanism

Abstract:Duetothecomplexbackgroundofshiptargetsandmuchirrelevantinterferenceinvisualimages,itis

difficulttoconductshipdetection.Inaddition,therearefewdatasetsformulti-categoryshipdetectionandthe

samplesareoftenunbalanced,whichmakestheshiptargetdetectionperformancedegraded.Consideringtheship

detectionbackgroundinterference,animprovedYOLOv3modelwasproposedbyintroducingSimAMattention

mechanism,whichwasusedtoenhancetheweightoftheshiptargetintheextractedfeaturesandsuppressthe

weightofbackgroundinterference,thusimprovingthemodeldetectionperformance.Meanwhile,strongreal-time

dataaugmentationwasappliedtoimprovetheunbalanceddistributionofsamplescales,andtransferlearningwas

combinedtoimprovetheshipdetectionaccuracyintheconditionofarestrictednumberofsamples.The

visualizationresultsofextractedfeaturesshowthattheimprovedmodelcouldsuppressirrelevantbackground

features,andtheabilitiesoffeatureextractionandtargetlocalizationwereenhanced.Withoutintroducingadditional

learnableparameters,theproposedmodelachieved96.93%and71.49%formAP.5andmAP.75ontheSeaShips

dataset,anddetectionspeedreached66framespersecond,indicatingagoodbalancebetweendetectionaccuracy

andefficiency.Theimprovedmod