[20]，作者總結(jié)出特征金字塔方法性能不足的原因便是金字塔中每個特征圖主要由單級特征組成，所以設(shè)計融合了骨干網(wǎng)絡(luò)中的多尺度特征從而得到基礎(chǔ)特征，將基礎(chǔ)特征送入一組交替連接的簡化U型模塊和特征融合模塊組中，得到的多尺度特征度輸出作為輸入送到下一層中進行等尺度特征組合得到多層次特征金字塔。單階段目標檢測算法雖然在起步時間上落后于兩階段目標檢測算法，但是其模型結(jié)構(gòu)的簡單以及計算檢測性能的高效吸引了很多學者和研究者的目光，無論是在現(xiàn)在還是未來其發(fā)展都是越來越好的。早期的單階段目標檢測算法常常以檢測速度快為主要優(yōu)點，但在檢測精度方面卻與兩階段檢測算法差距較大。隨著計算機視覺領(lǐng)域的不斷發(fā)展進步，如今的單階段目標檢測模型的速度和精度性能都得到了很大的提升。參考文獻LiuL,OuyangW,WangX,etal.DeepLearningforGenericObjectDetection:ASurvey[J].SpringerUS,2018P.Felzenszwalb,D.McAllester,andD.Ramanan.Adiscriminativelytrained,multiscale,deformablepartmodel[C].2008IEEEConferenceonComputerVisionandPatternRecognition.IEEE,2008.UIJLINGSJR,VANDSKEA,GEVERST,etal.Selectivesearchforobjectrecognition[J].InternationalJournalofComputerVision,2013,104(2):154-171.PAPAGEORGIOUCP,ORENM,POGGIOT.Ageneralframeworkforobjectdetection[C].SixthInternationalConferenceonComputerVision，1998：555-562.A.Krizhevsky,I.Sutskever,andG.E.Hinton,“Imagenetclassificationwithdeepconvolutionalneuralnetworks,”inAdvancesinneuralinformationprocessingsystems,2012,pp.1097–1105.ZITNICKCL,DOLLARP.Edgeboxes:Locatingobjectproposalsfromedges[C].EuropeanConferenceonComputerVision.Springer,Cham,2014:391-405.HUQ,ZHAIL.RGB-Dimagemulti-targetdetectionmethodbasedon3DDSFR-CNN[J].InternationalJournalofPatternRecognitionandArtificialIntelligence,2019,33(8):1954026.R.Girshick,J.Donahue,T.Darrell,andJ.Malik,“Richfeaturehierarchiesforaccurateobjectdetectionandsemanticsegmentation,”inProceedingsoftheIEEEconferenceoncomputervisionandpatternrecognition,2014,pp.580–587.HeK,ZhangX,RenS,etal.SpatialPyramidPoolinginDeepConvolutionalNetworksforVisualRecognition[J].IEEETransactionsonPatternAnalysis&MachineIntelligence,2014,37(9):1904-16.R.Girshick.FastR-CNN[J].ComputerScience,2015.S.Ren,K.He,R.Girshick,andJ.Sun.FasterR-CNN:TowardsReal-TimeObjectDetectionwithRegionProposalNetworks[J].IEEETransactionsonPatternAnalysis&MachineIntelligence,2017,39(6):1137-1149.HEK，GKIOXARIG，DOLLáRP，etal.MaskR-CNN[C].ProceedingsoftheIEEEInternationalConferenceonComputerVision，2017：2961-2969.T.-Y.Lin,P.Dollár,R.B.Girshick,K.He,B.Hariharan,andS.J.Belongie.FeaturePyramidNetworksforObjectDetection[C].2017IEEEConferenceonComputerVis