前景与背景交互融合网络用于伪装目标检测

打开文本图片集

关键词：伪装目标检测；前景与背景信息；非局部注意力；特征交互；特征融合DOI：10.15938/j. jhust.2025.02.006中图分类号：TP391.4 文献标志码：A 文章编号：1007-2683（2025）02-0053-11

Abstract：Aiming atthe problem of incompletedetectionresultsand blurrededgedetailsincurrentcamouflagedobjectdetection （COD）methods，anovelForegroundand BackgroundInteractiveFusionNetwork（FBIFNet）wasproposed tofurtherimprovethe performanceofCODthroughjointexplorationofforegroundandbackgroundregions.FBFNetcontainsakeyBilateralInteractiveusion module（BIF），whichusesapairofcomplementaryatentionstoguidethenetwork tojointlyreasonaboutcamouflagedobjectsfrom bothforegroudandbackgrounddirectionsndalsoutilizesaninteractionstrategybasedonthebidirectionalatentionmechanismanda weighted fusionstrategytoleancomplementaryiformationbetweenforegroudandbackgroundIndition，anAtentionalCascaded Positioning module（ACP）isincluded，whichcanlocalizecamouflagedbjectsfromaglobalperspectiveandprovidemoreacurate foregroundandbackgroundguidanceforBIF.Withthetwoproposedmodules，FBIFNetcanmoreaccuratelydetectcamouflaged objects.Extensive experimentsonthree publicdatasets（CAMO，CODlOK，and NC4K）demonstrate thatthe proposednetwork outperforms state-of-the-art methods in related fields on four evaluation metrics.

Keywords：camouflagedbjectdetection；foregroundandbackgroundinformation;non-localatention；feature interaction；featur fusion

0 引言

伪装是大自然中猎物为了躲避捕食者所进化出的独特能力。（剩余16994字）