[1]朱世伟,杭仁龙,刘青山.基于类加权YOLO网络的水下目标检测[J].南京师范大学学报(自然科学版),2020,43(01):129-135.[doi:10.3969/j.issn.1001-4616.2020.01.019]
　ZhuShiwei,HangRenlong,LiuQingshan.UnderwaterObjectDetectionBasedontheClass-WeightedYOLONet[J].JournalofNanjingNormalUniversity(NaturalScienceEdition),2020,43(01):129-135.[doi:10.3969/j.issn.1001-4616.2020.01.019]
点击复制

基于类加权YOLO网络的水下目标检测()
分享到：

《南京师范大学学报》（自然科学版）[ISSN:1001-4616/CN:32-1239/N]

卷:
第43卷
期数:
2020年01期
页码:
129-135
栏目:
·计算机科学与技术·
出版日期:
2020-03-15

文章信息/Info

Title:
UnderwaterObjectDetectionBasedontheClass-WeightedYOLONet
文章编号:
1001-4616(2020)01-0129-07
作者:
朱世伟; 杭仁龙; 刘青山
南京信息工程大学自动化学院,江苏南京210044
Author(s):
ZhuShiwei; HangRenlong; LiuQingshan
CollegeofAutomation,NanjingUniversityofInformationScienceandTechnology,Nanjing210044,China
关键词:
水下目标; YOLO; 类加权损失; 自适应维度聚类
Keywords:
underwaterobject; YOLO; class-weightedloss; dimensionadaptiveclustering
分类号:
TP391
DOI:
10.3969/j.issn.1001-4616.2020.01.019
文献标志码:
A
摘要:
由于水下目标检测面临着图像模糊、尺度多样化、复杂背景等问题,给水下目标检测应用带来很多挑战.本文提出了一种基于类加权YOLO网络的水下目标检测方法,主要思想是在深度网络YOLO的基础上,构造了类加权损失函数,来平衡样本难易程度以获得更好的效果,并引入了目标框自适应维度聚类方法,进一步提升了检测性能.实验结果表明,本文算法与传统的YOLO网络模型相比,在每幅图片包含近20个目标的密集目标检测任务中,能够将平均准确率从71.2%提升至74.1%,召回率由71.1%提升到78.3%.
Abstract:
Underwaterobjectsdetectionexistmanyissues,suchasblurimage,variousobjectscales,complexbackgroundandsoon.Inthispaper,weproposeaclass-weightedYOLOnetforunderwaterobjectdetection,inwhichaclass-weightedlossisdesignedtobalancesampleofdifficultysoastoacquirebettereffect.Moreover,adimensionadaptiveclusteringofobjectboxisintroducedtopromotethedetectionperformance.TheexperimentalresultsshowthattheproposedmethodoutperformstothetraditionalYOLOnet,withtheincreasingofthemAPfrom71.2%to74.1%andtherecallfrom71.1%to78.3%,inthetaskofdenseobjectdetectionwhicheveryimagenearlycontained20objects.

参考文献/References:

[1]OLMOSA,TRUCCOE.Detectingman-madeobjectsinunconstrainedsubseavideos[C]//IEEEInternationalConferenceonBritishMachineVision.BritishPublisher:BMVAPress,2002:517-526.
[2]王猛,杨杰,白洪亮.基于区域分割的水下目标实时识别系统[J].计算机仿真,2005,22(8):101-105.
[3]RIZONM,HANIZAY,PUTEHS,etal.Objectdetedtionusinggeometricinvariantmoment[J].Americanjournalofappliedsciences,2006,3(6):1876-1878.
[4]张铭钧,尚云超,杨杰.基于灰度-梯度不变矩的水下目标识别系统[J].哈尔滨工程大学学报,2009,30(6):653-657.
[5]SHIHAVUDDINASM,GRACIASN,GARCIAR.Onlinesunflickerremovalusingdynamictextureprediction[C]//InternationalConferenceonComputerVisionTheoryandApplications.Barcelona,Spain,2012:161-167.
[6]GIRSHICKR,DONAHUEJ,DARREL1T,etal.Richfeaturehierarchiesforaccurateobjectdetectionandsemanticsegmentation[J].IEEEtransactionspatternanalysisandmachineintelligence,2015,38(1):142-158.
[7]RENS,HEK,GIRSHICKR,etal.Fasterr-cnn:Towardsreal-timeobjectdetectionwithregionproposalnetworks[J].IEEEtransactionsonpatternanalysisandmachineintelligence,2015,39(6):1137-1149.
[8]REDMONJ,DIVVALAS,GIRSHICKR,etal.Youonlylookonce:unified,real-timeobjectdetection[C]//IEEEConferenceonComputerVisionandPatternRecognition.LasVegas,America,2016:779-788.
[9]REDMONJ,FARHADIA.YOLO9000:Better,Faster,Stronger[C]//IEEEConferenceonComputerVisionandPatternRecognition.Hawaii,America,2017:7263-7271.
[10]LINTY,DOLLARP,GIRSHICKR,etal.Featurepyramidnetworksforobjectdetection[C]//IEEEConferenceonComputerVisionandPatternRecognition.Hawaii,America,2016:2117-2125.
[11]HEK,ZHANGX,RENS,etal.Deepresiduallearningforimagerecognition[C]//IEEEConferenceComputerVisionandPatternRecognition.LasVegas,America,2016:770-778.
[12]乔小妮,张明新,史变霞.一种基于密度的K-means算法[J].电脑开发与应用,2008,21(10):9-11.
[13]KANGK,LIH,XIAOT,etal.Objectdetectioninvideowithtubeletproposalnetworks[C]//IEEEConferenceonComputerVisionandPatternRecognition.Hawaii,America,2017:727-735.

备注/Memo

备注/Memo:
收稿日期:2018-09-21.
基金项目:江苏省高校自然科学研究面上项目(18KJB520032)、江苏省青年基金项目(BK20180786).
通讯作者:刘青山,教授,博士生导师,研究方向:模式识别与计算机视觉.E-mail:qsliu@nuist.edu.cn
更新日期/Last Update: 2020-03-15