SHIPENGINEERING船舶工程Vol.45No.22023总第45卷，2023年第2期—154—结合空洞编码器和特征金字塔的中心点船舶检测熊超1a,2，周海峰1a,2，郑东强1b，林忠华1b，张兴杰1c，关天敏1d（1．集美大学，a.轮机工程学院；b.海洋装备与机械工程学院；c.航海学院；d.海洋信息工程学院，福建厦门361021；2．福建省船舶与海洋工程重点实验室，福建厦门361021）摘要：为了进一步提高基于深度学习的船舶目标检测技术的检测精度，在无锚框中心点检测算法基础上，提出一种结合空洞编码器和特征金字塔的改进中心点船舶检测算法。采用ResNeXt-50网络对船舶图像进行特征提取，引入基于空洞残差的空洞编码器（DE）增大32倍下采样特征图的感受野，生成覆盖多个目标尺度的特征图，并采用特征金字塔网络（FPN）进行上采样，在上采样过程中融合空洞编码器生成的32倍下采样特征图和原16倍、8倍和4倍下采样特征图，从而提取到更丰富的船舶特征信息，提升船舶检测效果。结果表明，改进算法对不同类型和不同尺度下的船舶检测平均精确率相比原算法具有较明显的提升，相比SSD和YOLOv3算法具有更高的精度优势。关键词：无锚框中心点检测；空洞编码器；特征金字塔中图分类号：TN911文献标志码：A【DOI】10.13788/j.cnki.cbgc.2023.02.21CenterNetShipDetectionCombiningDilatedEnconderandFeaturePyramidNetworkXIONGChao1a,2,ZHOUHaifeng1a,2,ZHENGDongqiang1b,LINZhonghua1b,ZHANGXingjie1c,GUANTianmin1d(1.JimeiUniversity,a.CollegeofMarineEngineering;b.SchoolofMarineEquipmentandMechanicalEngineering;c.NauticalInstitute;d.CollegeofOceanInformationEngineering,Xiamen361021,Fujian,China;2.KeyLaboratoryofNavalArchitectureandOceanMarineEngineeringofFujianProvince,Xiamen361021,Fujian,China)Abstract:Inordertofurtherimprovethedetectionaccuracyoftheshiptargetdetectiontechnologybasedondeeplearning,basedontheanchor-freecenternetdetectionalgorithm,animprovedcenternetshipdetectionalgorithmcombiningthedilatedenconderandthefeaturepyramidnetworkisproposed.ResNeXt-50networkisusedtoextractthefeaturesoftheshipimage,dilatedenconder(DE)basedonthedilatedresidualnetworkisintroducedtoincreasethereceptivefieldof32timesdown-samplingfeaturemaptogeneratefeaturemapscoveringmultipletargetscales.Then,thefeaturepyramidnetwork(FPN)isu...
