电子测量技术ELECTRONICMEASUREMENTTECHNOLOGY第45卷第23期2022年12月DOI:10.19651/j.cnki.emt.2210042基于改进YOLOX的红外目标检测算法*谌海云余鸿皓王海川黄忠义(西南石油大学电气信息学院成都610500)摘要:针对红外目标图像分辨率低,缺少纹理细节,存在复杂背景干扰导致检测精度低的问题,提出一种基于改进YOLOX的红外目标检测算法。首先,设计了一种有效的空间通道混合注意力模块,将其引入在特征提取主干网络CSP-Darknet53中,以减少网络由于远距离传输造成的精度损失;其次,为了进一步提升红外目标的检测精度,在原本加强特征提取网络PANet的基础上提出一种改进的路径特征融合方法;最后,为了解决红外目标中小物体预测精度低的问题,在YOLOX输出检测头处进行反卷积操作扩大输出特征图。在FLIR红外公开数据集上进行实验,实验结果表明,所提算法识别的平均精度均值(mAP)达91.00%,相比于基准YOLOX网络的平均精度提升了5.04个百分点,对于提升红外目标的检测精度是有效的。关键词:卷积神经网络;红外目标检测;YOLOX;注意力机制;特征融合中图分类号:TP391文献标识码:A国家标准学科分类代码:520.20ObjectdetectionalgorithmofthermalinfraredimagesbasedonimprovedYOLOXShenHaiyunYuHonghaoWangHaichuanHuangZhongyi(SchoolofElectricalEngineeringandInformation,SouthwestPetroleumUniversity,Chengdu610500,China)Abstract:Tosolvetheproblemoflowresolutionofinfraredtargetimages,lackoftexturedetails,andlowdetectionaccuracycausedbycomplexbackgroundinterference,aninfraredtargetdetectionalgorithmbasedonimprovedYOLOXisproposed.First,aneffectivespatialchannelmixedattentionmoduleisintroducedintothefeatureextractionbackbonenetworkCSP-Darknet53toreducetheaccuracylossofthenetworkduetolong-distancetransmission;secondly,inordertofurtherimprovethedetectionaccuracyofinfraredtargets,basedontheoriginalenhancedfeatureextractionnetworkPANet,animprovedpathfeaturefusionmethodisproposed;finally,inordertosolvetheproblemoflowrecognitionrateofsmallobjectsininfraredtargets,adeconvolutionoperationisperformedattheYOLOXoutputdetection-headtoexpandtheoutputfeaturemap.ExperimentsarecarriedoutontheFLIRinfraredpublicdataset.TheexperimentalresultsshowthatthemeanAveragePrecision(mAP)oftheproposedalgorithmrecogn...
