一种应用在视频监控场景中的显著目标识别和分割方法.pdfVIP

长，信息通信

DOI:10.20153/j.issn.2096-9759.2025.03.010

一种应用在视频监控场景中的显著目标识别和分割方法

张军山，郝源，罗振,郭志彪，郑斌

中国移动通信集团广东有限公司，广东广州510623

ASignificantObjectRecognitionandSegmentationMethodAppliedin

VideoSurveillanceScenarios

ZHANGJunshan,HAOYuan,LUOZhen,GUOZhibiao,ZHENGBin

ChinaMobileGroupGuangdongCompanyLimited,GuangdongGuangzhou510623,China

摘要：深度神经网络（DNN)在显著性检Abstract:Theperformanceofdeepneuralnetworks(DNNs)insaliencydetectionfarexceeds

测方面的性能远远超过传统方法。然thatoftraditionalmethods.However,manyDNNmodelsbasedonsegmentationnetworksfor

而，基于分割网络的显著性检测的许多saliencydetectionaredesignedfortheoptimizationofhigh-levelfeatures,whichareadaptedto

DNN模型被设计为高级特征的优化，这saliencydatasets.Thesemethodsrelyonlarge-scaledatasetsfortraining,andtheirsaliencyinfer-

些特征适应于显著性数据集。这些方法encessometimesconflictwiththeresultsofhumanvisionandcognitivefocus.Tosolvethese

对于训练依赖于大规模数据集，并且它problems,weproposeasaliencyfunctioncloselyrelatedtohumanunderstandingofsaliencyfor

们的显著性推断有时与人类视觉和认知objectproposalclassificationinsaliencydetection.ThesaliencyfunctioncanbeusedasaDNN

焦点的结果存在冲突。为了解决这些问layer,explicitlyprovidingthecriteriafordiscoveringsaliencyandreducingthedependenceon

题，文章提出了一个与人类对显著性理thesizeofthetrainingset.Incxperimentalproofs,weverifiedourmethodonfivebenchmark

解密切相关的显著性函数，用于显著性datasetsandcompareditwithtenstate-of-the-artobjectproposalclassificationmethodsforsali-

检测的对象提议分类。显著性函数可以cncydetection.OurmethodoutperformsthecorrespondingmethodsintermsofaccuracyandF-

作为DNN层，明确提供发现显著性的标measure,improvingtheaccuracybyupto4.27%andtheF-measureby2.80%.Wealsodem-

准，减少对训练集规模的依赖性。在实onstratethatbyintroducingthesaliencyfunction,evenwhentrainingonasmall-scaledataset

验证明，我们在五个基准数据集上验证containing