Computer Science

Hao Shen, Xiaofeng Cong, Henghui Ding, Yulun Zhang, Xudong Jiang

2026.2.10Visual Intelligence

DOI: 10.1007/s44267-026-00108-2

tlooto Summary

A synergic expert modulation network is constructed by inserting SEM-based building blocks into the U-Net architecture to increase efficiency and achieve state-of-the-art performance on multiple datasets for the image dehazing task while incurring lower computational costs.

Abstract

Recentdevelopmentsincontextmodulationmechanismshaveachievedsignificantimprovementsinperformance, aswellasbettertrade-offsbetweenmodelaccuracyandefficiency.Thesemechanismsoperateoninputthrough contextmodelingandthenleveragethesecontextstomodulateprojectedinputfeatures.However,existing methodshavelimitations.First,theycannotadaptivelylearntheextractedhierarchicalcontextorignorethe complementarityofthecross-scalecontext.Second,thesemethodsdonotadequatelyaddresstheunique frequencycharacteristicsofhazyimages.Inresponse,weproposeasynergicexpertmodulation(SEM)mechanism toexplicitlymodelcontextinformation.Specifically,theSEMconsistsprimarilyoftwomixtureofspatialexperts (MSE)modulesthathandlefeaturesofdifferentscalesandonemixtureoffrequencyexperts(MFE)modulethat operateswithinthefrequencydomain.TheMSElearnshierarchicalfeaturesofvariousgranularitiesinanadaptive manner,guidedbymultiplegatingexpertsandaroutingnetwork.TheMFEspecializesinminingfrequencycontexts guidedbymultiplefrequencyexpertsandaroutingnetwork.Atthemicrolevel,eachfrequencyexpertoperatesin twostages:spectralfilteringandspectrallearning.Theformerperformsmaskfilteringtoenhancetheweightsof low-frequencycomponents,andthelatterperformsFourieramplitudeandphasedecoupledlearning,thus promotingtheremovalofhazeinformationandglobalcontextlearning.Finally,theobtainedcontextsareintegrated tomodulatetheprojectedfeature,therebysignificantlyenhancingcross-domainfeaturesynergies.Theproposed network,referredtoasthesynergicexpertmodulationnetwork,isconstructedbyinsertingSEM-basedbuilding blocksintotheU-Netarchitecturetoincreaseefficiency.Extensiveexperimentsdemonstratethatournetwork achievesstate-of-the-artperformanceonmultipledatasetsfortheimagedehazingtaskwhileincurringlower computationalcosts.

Citation format

SHEN, Hao, et al. Efficient image dehazing with synergic expert modulation. Visual Intelligence, 2026, 4.