北大中文核心期刊国外电子测量技术DOI:10.19652/j.cnki.femt.2204455基于深度学习的红外过采样扫描图像小目标跟踪算法*姚迎乐1赵娟2(1.郑州工业应用技术学院信息工程学院郑州451150;2.武汉轻工大学数学与计算机学院武汉430023)摘要:红外小目标跟踪过程中由于背景、外界杂波等干扰,导致跟踪精确度和实时性欠佳,为此,提出基于深度学习的红外过采样扫描图像小目标跟踪算法。首先构建了红外过采样扫描图像模型,通过背景估计、形态学开运算,对图像中背景以及外界杂波进行多级滤除;然后增加设计特征融合模块和区域选取模块来改进孪生网络,生成融合特征图输入目标区域,通过分类和回归计算提高图像的特征表征能力和跟踪精度;最后建立损失函数训练孪生网络,输出红外过采样扫描图像小目标跟踪结果。实验结果表明,利用所提算法进行图像滤除后,信噪比能够高达35dB,所提算法的区域重叠率较高、跟踪精度高,且算法的实时性强,帧率达到200fps以上,整体跟踪效果好。关键词:多级滤波;改进孪生网络;特征融合;区域选取;红外小目标跟踪中图分类号:TP391.41文献标识码:A国家标准学科分类代码:520.6SmallobjecttrackingalgorithmforinfraredoversampledscanningimagesbasedondeeplearningYaoYingle1ZhaoJuan2(1.DepartmentofInformationEngineering,InstituteofZhengzhouIndustrialApplicationTechnology,Zhengzhou451150,China;2.DepartmentofMathematicsandComputerScience,WuhanPolytechnicUniversity,Wuhan430023,China)Abstract:Intheprocessofinfraredsmallobjecttracking,thetrackingaccuracyandreal-timeperformancearepoor.Therefore,thealgorithmofinfraredover-samplingscanningimagesmallobjecttrackingbasedondeeplearningisproposed.Firstly,theinfraredoversamplingscanningimagemodelisconstructed,usedtofilterthebackgroundandexternalclutter,thenaddthedesignfeaturefusionmoduleandareaselectionmoduletoimprovethetwinnetwork,generatethefusionfeaturemapinputtargetarea,andimprovethefeaturerepresentationabilityandtrackingaccuracythroughclassificationandregressioncalculation.Finally,thelossfunctionisestablishedtotrainthetwinnetworkandoutputthesmalltargettrackingresultsofinfraredoversampledscanningimages.Theexperimentalresultsshowthattheproposedalgorithmcanbeupto35dB,theproposedalgorithmhashighregionaloverlaprate,hi...
