面向油田图像分割的空-频双域融合深度网络

打开文本图片集

关键词：语义分割；空频双域融合；多尺度频域金字塔；混合注意力；石油生产场景中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2026）04-031-1222-07doi：10. 19734/j.issn. 1001-3695.2025.07.0260

Spatial-frequency dual-domain fusion deep network for oilfield image segmentation

Lu Xiaolinabt，Ji Xiaofengab，Wang Yana，b （a.QingdaoIstuteofSoftare，b.CollgeofomputerSciencendTecologyChina UniversityfPetrolem（EastChina），Qingdao Shandong 266580，China）

Abstract：Petroleum production involves complex scenarios where fires and leaks occurred frequently，posing severe safety risks.Toachieveaccurateanomalyrecognitionandlocalization，thisstudyproposedaspatial-frequencydual-domain fusion network（SFDFNet）to enhancesemanticsegmentation in detail preservation andglobal featuremodeling.Themethod employedamulti-resolutionconvolutionalneuralnetwork inthespatialdomaintoextractmulti-levelimagefeatures.Itintroducedamulti-scale frequency-domainpyramid（MFDP）module tocapture edgesandglobal textures，enriching highfrequencydetails.Ahybridatention fusion module（HAFM）adaptively integrated spatialand frequencyfeatures.Experimental results show that SFDFNet achieves 70.8% mIoU on the petroleum production dataset and reaches 78.5% and 78.1% mIoU on Cityscapes and CamVid datasets.The method significantly outperforms mainstream approaches，which demonstrates its effectiveness.

Keywords：semanticsegmentation；spatial-frequencydual-domain fusion；multi-scale frequency-domain pyramid；hybrid attention；petroleum production scenarios

0引言

石油生产场景具有高度复杂性与多变性，尤其陆地生产场景中常伴随火灾、烟雾、刺漏等突发异常情况。（剩余18180字）