文章编号:1002-2082(2024)02-0346-08基于改进FasterR-CNN的红外目标检测算法汪西晨1,彭富伦2,李业勋3,张俊举1(1.南京理工大学电子工程与光电技术学院,江苏南京210094;2.西安应用光学研究所,陕西西安710065;3.江苏北方湖光光电有限公司,江苏无锡214100)摘要:为提升红外目标的检测精度,提出了一种引入频域注意力机制的FasterR-CNN红外目标检测算法。首先,针对红外图像边缘模糊和噪声问题,设计了一种并行的图像增强预处理结构;其次,在FasterR-CNN中引入频域注意力机制,设计了一种新型红外目标检测主干网络;最后,引入路径增强金字塔结构,融合多尺度特征进行预测,利用底层网络丰富的位置信息,提升检测精度。在红外飞机的数据集上进行实验,结果表明,改进后的FasterR-CNN目标检测框架比以ResNet50为主干的算法的AP提升了7.6%。此外,与目前主流算法对比,本文算法提高了红外目标的检测精度,验证了算法改进的有效性。关键词:红外目标检测;图像增强;FasterR-CNN;频域注意力机制;多尺度特征融合中图分类号:TN206;TP391.4文献标志码:ADOI:10.5768/JAO202445.0202001InfraredtargetdetectionalgorithmbasedonimprovedFasterR-CNNWANGXichen1,PENGFulun2,LIYexun3,ZHANGJunju1(1.SchoolofElectronicandOpticalEngineering,NanjingUniversityofScienceandTechnology,Nanjing210094,China;2.Xi'anInstituteofAppliedOptics,Xi'an710065,China;3.JiangsuNorthHuguangPhotoelectricCo.,Ltd.,Wuxi214100,China)Abstract:Inordertoimprovethedetectionaccuracyofinfraredtargets,aFasterR-CNNinfraredtargetdetectionalgorithmintroducingafrequencydomainattentionmechanismwasproposed.Firstly,aparallelimageenhancementpreprocessingstructurewasdesignedtoaddresstheissuesofedgeblurandnoiseininfraredimages.Secondly,afrequencydomainattentionmechanismwasintroducedintoFasterR-CNN,andanewinfraredtargetdetectionbackbonenetworkwasdesigned.Finally,apathenhancedpyramidstructurewasintroducedtofusemulti-scalefeaturesforprediction,andtherichlocationinformationoftheunderlyingnetworkwasutilizedtoimprovedetectionaccuracy.Theexperimentwasconductedonadatasetofinfraredaircraft.TheresultsshowthattheAPofimprovedFasterR-CNNtargetdetectionframeworkis7.6%higherthanthatofthealgorithmwithResN...