基于改进RT-DETR的小目标检测算法

打开文本图片集

SmallobjectdetectionalgorithmbasedonimprovedRT-DETR

WANGKang1，WANGXiaolin²，LIUXinzhi²，DENGJianzhi1 （1.UniversityofTechnology，Guilin541o4，China; 2.GuangxiDatengxiaWaterControlProjectDevelopmentCo.，Ltd.，Nanning53oo21，China）

Abstract：Smallobjectdetectionoften faceschallengessuchasmisseddetectionsandfalsepositivesduetothesmal proportionoftheobjecttotheimageandthelimitedsemanticinformation.Inviewof this，animprovedRT-DETRbasedsmall objectdetectionmodelisproposedtoenhancedetectionperformancewhileensuringreal-timeperformance.Thebackbone networkof theRT-DETRmodelismodifiedbydesigningapartiallre-parameterizedconvolutionmoduletoimprovefeature extractioneficiencyAniientmulti-scaleattention （EMA）mchanismisitroducedtoagggatespatialandcrossspatial information.TheHiLoatentionmechanismisemployedintheAIFIencodertoreducethecomputationalcostsandenhancethe robustnessofthedetectionalgorithm.ExperimentswereconductedontheFloW-Imgdatasetofthesmallobjectsonwater.The resultsshowthatboththemisseddetectionrateandthefalsepositiverateof themodelbasedontheimprovedRT-DETRare reduced in comparison with the baseline model RT-DETR.On the test set， the mAP@0.5 of the proposed algorithm achieves 0.841and itsmAP@0.5：0.95is0.394，representingimprovementsof5.5%and3.7%，respectivelyoverthebaselinemodelRDETR.The detection performance surpasses thoseof both the baseline model and the object detection models of YOLO series.

Keywords：small object detection; RT-DETR model; PCov; Transformer; EMA; HiLo attention mechanism

0 引言

目标检测作为计算机视觉领域中的重要问题，其主要任务包括识别图像中的目标类别和确定目标位置。（剩余9816字）