文章编号：1002-2082(2024)02-0430-08基于MCRASN的遥感影像变化检测谢国波，廖文康，林志毅，张家源（广东工业大学计算机学院，广东广州510000）摘要：为了提升经配准高分辨率遥感影像对变化检测的精度，基于ChangeFormer提出了一种将移动卷积与相对注意力相结合的孪生网络(mobileconvolutionandrelativeattentionSiamesenetwork,MCRASN)。该网络以垂直布局结合移动卷积和相对注意力，构建多阶段组合编码器替换原网络编码器，高效地捕捉所需的多尺度细节特征和像素间相互关系信息，改进差异模块为1个可学习的距离度量模块进行距离计算，同时通过引入EFL（equalizedfocalloss）损失函数解决数据集正负样本失衡的问题以实现精确的变化检测。实验结果表明，所提出的MCRASN算法在LEVIR-CD数据集上具有更好的变化检测性能，其精确率、召回率、F1得分和总体精度分别为93.94%、89.26%、91.54%和99.18%，优于先前的多种检测方法。关键词：变化检测；孪生网络；移动卷积；相对注意力；距离计算中图分类号：TN919.81;TP391.4文献标志码：ADOI：10.5768/JAO202445.0203005RemotesensingimageschangedetectionbasedonMCRASNXIEGuobo，LIAOWenkang，LINZhiyi，ZHANGJiayuan（SchoolofComputerScience,GuangdongUniversityofTechnology,Guangzhou510000,China）Abstract：Inordertoimprovetheaccuracyofchangedetectioninco-registeredhigh-resolutionremotesensingimages,aSiamesenetworkcombiningmobileconvolutionandrelativeattention(MCRASN)wasproposedbasedonChangeFormer.Amulti-stagecombinedencoderwasconstructedtoreplacetheoriginalnetworkencoderbyusingverticallayoutcombinedwithmobileconvolutionandrelativeattentiontoefficientlycapturetherequiredmulti-scaledetailedfeaturesandpixelcorrelationinformation,andthedifferencemodulewasimprovedtobealearnabledistancemetricmodulefordistancecalculation.Atthesametime,theequalizedfocalloss(EFL)lossfunctionwasintroducedtosolvetheproblemofimbalancebetweenpositiveandnegativesamplesinthedatasettoachieveaccuratechangedetection.TheexperimentalresultsshowthattheproposedMCRASNmethodhasbetterchangedetectionperformanceontheLEVIR-CDdataset,withprecision,recall,F1scoreandoverallaccuracyof93.94%,89.26%,91.54%and99.18%,respectively,whichissuperiortopreviousmethods....
