文档名：基于改进DPGN的少样本图像分类算法研究
摘要：DPGN(distributionpropagationgraphnetwork)是基于深度学习的少样本图像分类算法,在数据稀疏的条件下可以顺利完成图像分类,但其分类的准确率仍需进一步提升.以DPGN算法为研究对象,提出SFOD_DPGN(SinAM_FRN_layer_ODConv_DM&EMD_distributionpropagationgraphnetwork)算法.在骨干神经网络Resnet12的残差块中融入注意力机制;将Res-net12网络中批量归一化与ReLu激活函数搭配使用的方式改为滤波器响应归一化与阈值线性单元激活函数搭配使用的方式;在分类器模块中选用全维动态卷积替换普通卷积;使用马氏距离和推土机距离替换L2距离度量函数.在CUB-200-2011数据集上的实验表明,在5way-1shot和5way-5shot分类任务下,SFOD_DPGN算法比DPGN算法的准确率提升约7.97%和2.66%.
Abstract：Thedistributionpropagationgraphnetwork(DPGN)isafew-shotimageclassificationalgorithmbasedondeeplearning.Unfortunately,theDPGNalgorithmcompletelyignoressemanticinformation,whichisimportantforfine-grainedclassification.Therefore,itdeliverspoorclassificationperformances.ThispaperproposesanewFew-shotlearningalgorithmbasedontheDPGNalgorithm,SinAM-FRN_layer-ODConv-DM&EMD_DistributionPropagationGraphNetwork(SFOD_DPGN).First,toaddresstheinabilitytoextractimagefeaturesbythefeatureextractionmoduleoftheDPGNalgorithm,theSimAMattentionmechanismisintegratedintofourresidualblocksofthefeatureextractionnetworkResNet12.TheSimAMattentionmechanismcangeneratethree-dimensionalweightsforfeaturemapsfrombothspatialandchanneldimensions,andthenaggregatesthegeneratedweightswiththefeaturemapstoenabletheimprovedResNet12tolearnmoreandricherimagefeatures;Second,inviewthatthenormalizationmethodoftheResNet12isaffectedbythenumberofimagesselectedintraining,thecombinationofbatchnormalizationandtheReLuactivationfunctioninthemainpathofeachresidualblockoftheResNet12ischangedtothecombinationofthefilterresponsenormalization(FRN)andthethresholdlinearunitactivationfunction(TLU).BecauseoftheFRNwithoutmeanoperation,iteasilyleadstoactivationwitharbitrarybiasfarfromzero.IftheFRNcombineswiththeReLuactivationfunction,thisbiashasadverseeffectsontraining.ThispaperemploystheTLUaftertheFRNtoaddresstheproblem.TheSFOD_DPGNalgorithmimprovestheclassificationaccuracyandensuresitsinferencespeed.Then,itoptimizestheclassifiermoduleoftheDPGNalgorithm.Tosolvepoorclassificationperformanceoftheclassifiermodule,thefull-dimensionaldynamicconvolution(ODConv)isselectedtoreplacethecommonconvolutionintheclassifiermodule.TheODconvemploysalinearcombinationofnconvolutionalkernelsandparallelstrategiestointroducemultidimensionalattentionmechanismsfordynamicweighting,makingtheconvolutionoperationdependentontheinput.TheODconvimprovestherobustnessoftheSFOD_DPGNalgorithm.Finally,theDPGNalgorithmusestheL2distancemeasurementmethodintheclassifiermodule,easilycausingerrorsincalculatingthedistancebetweensamples.Basedonthecharacteristicsofdistancemeasurementmethods,theMahalanobisDistance(MD)issuitableforcalculatingthedistancebetweensamples(pointgraphs).TheEarthMoves'sDistance(EMD)distanceismoresuitableforcalculatingthedistancebetweendistributiongraphs.ThispaperusestheMDandEMDtoreplacetheL2inordertoimprovetheabilityoftheclassifiertomeasurethedistancebetweensamples.ItimprovestheclassificationaccuracyoftheSFOD_DPGNalgorithm.ExperimentsontheCUB-200-2011datasetshowstheSFOD_DPGNalgorithmissuperiortotheDPGNalgorithmover5way-1shotand5way-5shotclassificationtasks.Theaccuracyimprovesby7.97%and2.66%respectively.Meanwhile,ablationexperimentsareperformedforeachparttoverifytheeffectoftheimprovedResNet12andtheclassifiermodule.ComparedtotheDPGNalgorithm,aftertheSimAMattentionmechanismisintegratedintotheResNet12,theaccuracyimprovesby2.77%and1.16%over5way-1shotand5way-5shotclassificationtasksrespectively.Furthermore,aftertheimprovingthenormalizationmethodandactivationfunctionoftheResNet12,theaccuracyis5.00%and2.04%higherrespectivelyover5way-1shotand5way-5shotclassificationtasks.AfterthefurtherreplacementofthecommonconvolutionwiththeODconv,theaccuracyisupby7.25%and2.42%respectivelyover5way-1shotand5way-5shotclassificationtasks.OurexperimentalresultsdemonstrateallimprovementsareeffectivetoimproveclassificationaccuracyoftheSFOD_DPGNalgorithm.
作者：王玲 孙莹 王鹏 白燕娥Author：WANGLing SUNYing WANGPeng BAIYan'e
作者单位：长春理工大学计算机科学技术学院,长春130022
刊名：重庆理工大学学报
Journal：JournalofChongqingInstituteofTechnology
年，卷(期)：2024, 38(3)
分类号：TP391
关键词：深度学习 少样本图像分类 注意力机制 全维动态卷积 马氏距离 推土机距离
Keywords：deeplearning few-shotimageclassification attentionmechanism omni-dimensionaldynamicconvolution mahalanobisdistance earthmover'sdistance
机标分类号：TP391.4SX591
在线出版日期：2024年3月25日
基金项目：吉林省自然科学基金基于改进DPGN的少样本图像分类算法研究[
期刊论文] 重庆理工大学学报--2024, 38(3)王玲 孙莹 王鹏 白燕娥DPGN(distributionpropagationgraphnetwork)是基于深度学习的少样本图像分类算法,在数据稀疏的条件下可以顺利完成图像分类,但其分类的准确率仍需进一步提升.以DPGN算法为研究对象,提出SFOD_DPGN(SinAM_FRN_layer_O...参考文献和引证文献
参考文献
引证文献
本文读者也读过
相似文献
相关博文
基于改进DPGN的少样本图像分类算法研究 Research on image classification algorithm with few-shot based on improved DPGN
基于改进DPGN的少样本图像分类算法研究.pdf
- 文件大小:
- 1.53 MB
- 下载次数:
- 60
-
高速下载
|
|
