image super-resolution using very deep residual channel attention networks使用非常深残留通道注意网络图像超分辨率.pdf

ImageSuper-ResolutionUsingVeryDeep

ResidualChannelAttentionNetworks

YulunZhang1[0000−0002−2288−5079],KunpengLi1[0000−0001−5805−793X],Kai

Li1[0000−0002−9027−0914],LichenWang1[0000−0002−3741−9492],Bineng

Zhong1[0000−0003−3423−1539],andYunFu1,2[0000−0002−8588−5084]

1DepartmentofECE,NortheasternUniversity,USA

2CollegeofComputerandInformationScience,NortheasternUniversity,USA

{yulun100,li.kai.gml,wanglichenxj}@,

bnzhong@,{kunpengli,yunfu}@

.Convolutionalneuralnetwork(CNN)depthisofcrucialim-

portanceforimagesuper-resolution(SR).However,weobservethat

deepernetworksforimageSRaremoredifficulttotrain.Thelow-

resolutioninputsandfeaturescontainabundantlow-frequencyinforma-

tion,whichistreatedequallyacrosschannels,hencehinderingtherep-

resentationalabilityofCNNs.Tosolvetheseproblems,weproposethe

verydeepresidualchannelattentionnetworks(RCAN).Specifically,we

proposearesidualinresidual(RIR)structuretoformverydeepnetwork,

whichconsistsofseveralresidualgroupswithlongskipconnections.Each

residualgroupcontainssomeresidualblockswithshortskipconnec-

tions.Meanwhile,RIRallowsabundantlow-frequencyinformationtobe

bypassedthroughmultipleskipconnections,makingthemainnetwork

focusonlearninghigh-frequencyinformation.Furthermore,weproposea

channelattentionmechanismtoadaptivelyrescalechannel-wisefeatures

byconsideringinterdependenciesamongchanneltensiveexperiments

showthatourRCANachievesbetteraccuracyandvisualimprovements

againststate-of-the-artmethods.

Keywords:Sup