基于改进CascadeR-CNN的探地雷达管线目标检测①来鹏飞,李伟,高尧,丁健刚,袁博,杨明(长安大学信息工程学院,西安710064)通信作者:高尧,E-mail:gaoyao3499@126.com摘要:针对人工识别探地雷达管线图像时效率低、误差大和成本高昂等问题,本文提出了一种基于改进CascadeR-CNN的管线目标智能化检测方法.首先对探地雷达管线图像数据集进行预处理,提升数据质量.然后采用ResNeXt代替ResNet作为主干网络提取目标特征信息,并添加多尺度特征融合模块FPN使高层特征向低层特征融合,增强低层特征表达能力.其次,使用高斯形式的非极大值抑制方法Soft-NMS得到更加精准的候选框,使用Smooth_L1作为损失函数,加速了模型收敛并且降低了训练中发生梯度爆炸的概率.最后,对于管线目标特殊的形状特征,设置合适的锚框长宽比和大小,提高锚框的生成质量.实验结果表明,本文方法解决了复杂特征的地下管线目标智能化检测,对地下管线目标检测的平均精度达到94.7%,比CascadeR-CNN方法提高了10.1%.关键词:探地雷达;地下管线;深度学习;CascadeR-CNN;FPN;Soft-NMS;目标检测引用格式:来鹏飞,李伟,高尧,丁健刚,袁博,杨明.基于改进CascadeR-CNN的探地雷达管线目标检测.计算机系统应用,2023,32(2):102–110.http://www.c-s-a.org.cn/1003-3254/8945.htmlGPRPipelineTargetDetectionBasedonImprovedCascadeR-CNNLAIPeng-Fei,LIWei,GAOYao,DINGJian-Gang,YUANBo,YANGMing(SchoolofInformationEngineering,Chang’anUniversity,Xi’an710064,China)Abstract:Asmanualidentificationofground-penetratingradar(GPR)pipelineimagesfacestheproblemsoflowefficiency,largeerrors,andhighcosts,thisstudyproposesanintelligentpipelinetargetdetectionmethodbasedonimprovedCascadeR-CNN.First,theGPRpipelineimagedatasetispreprocessedtoimprovedataquality.ResNeXtisusedinsteadofResNetasthebackbonenetworktoextracttargetfeatureinformation,andamulti-scalefeaturefusionmoduleFPNisaddedtofusehigh-levelfeaturestolow-levelfeaturestoenhancetheexpressivenessoflow-levelfeatures.Secondly,theGaussiannon-maximumsuppression(NMS)methodSoft-NMSisusedtoobtainmoreaccuratecandidateboxes,andSmooth_L1istakenasthelossfunction,whichacceleratesmodelconvergenceandreducestheprobabilityofgradientexplosionduringtraining.Finally,forthespecialshapefeaturesofthepipelinetarget,theap...
