基于物理解耦與自適應(yīng)特征提取的無監(jiān)督圖像去霧

打開文本圖片集

引用格式：，.基于物理解耦與自適應(yīng)特征提取的無監(jiān)督圖像去霧[J].現(xiàn)代電子技術(shù)，2025，48（17）：77-84.

中圖分類號：TN911.73-34；TP391.41 文獻(xiàn)標(biāo)識碼：A 文章編號：1004-373X（2025）17-0077-08

Unsupervised image dehazing based on physical decoupling and adaptivefeatureextraction

YAN Zaishuang1，2，HE Peng1，2 （1.HubeiKeyLaboratoryofIntellgentVisionBasedMonitoringforHydroelectricEngineering，Yichang443oChina; 2.CollegeofComputerandInformationTechnology，ChinaThreeGorgesUniversityYichang443Oo2，China）

Abstract：Unsupervisedlearning isadoptedtoimprovegeneralization inrealscenarios inmanyimage dehazingalgorithms. However，theexistingunsuperviseddehazingmethodsmostlyrelyontheCycleGAN（cycle-consistent generativeadversarial network）framework andrealizedehazing bysimplecyclesofdehazing andre-hazing，soitlacks efectiveconstraintsonthe generatedimagesandresultsinsuboptimalperformance.Moreover，thesemethodsfailtosuficientlyconsiderthephysical propertiesof haze inthe featureextraction，which causesasevere lossof texturedetails.Therefore，anunsupervised image dehazingnetworkbasedonphysicaldecouplingandadaptivefeatureextractionisproposed，andthenetworkmodelisnamed UPDA-Net.Specificalyaphysicaldecoupling network（PDN）isdesigned toseparateandestimatetheatmosphericlightand transmissionmapinhazyimagesbyestablishing physicalconstraintsbasedontheatmosphericscattering model.Thisdesign allowsformoreaccuraterestorationofimageilumination，andenhancesthephysicalconsistencyandclarityofthedehazing proces.Furthermore，anadaptive featureextraction module （AFEM）isdeveloped tointegrate thephysicalcharacteristicsof atmosphericlightand transmissionmaps.Byapplying theatmosphericscateringmodelinthefeaturespace，thismodule approximatesrelevantcharacteristicsandfuses thepotential physicalfeaturesofclearimages，soastoenhance themodel's featurerepresentationcapabilityandimprovethedetailandtexturequalityoftherestoredimages.Experimentalresults demonstratethattheproposedmethodoutperformssixmainstreamdehazingalgorithmsintermsofobjectiveevaluationandvisual quality on several public real image dehazing datasets.

Keywords：image dehazing;atmosphericscatering model;physical decoupling;parameter estimation；adaptive feature extraction;unsupervised learning;CycleGAN;imagerestoration

0 引言

圖像去霧是計算機視覺領(lǐng)域的一個重要研究課題，近年來受到廣泛關(guān)注。（剩余11993字）