Accurate image super-resolution using very deep convolutional networks

《Accurate image super-resolution using very deep convolutional networks》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、AccurateImageSuper-ResolutionUsingVeryDeepConvolutionalNetworksJiwonKim,JungKwonLeeandKyoungMuLeeDepartmentofECE,ASRI,SeoulNationalUniversity,Koreafj.kim,deruci,kyoungmug@snu.ac.krAbstract37.6VDSR(Ours)37.4Wepresentahighlyaccuratesingle-imagesuper-resolution(S

2、R)method.Ourmethodusesaverydeepcon-37.2volutionalnetworkinspiredbyVGG-netusedforImageNetclassification[19].Wefindincreasingournetworkdepth37PSNR(dB)showsasignificantimprovementinaccuracy.Ourfinal36.8modeluses20weightlayers.BycascadingsmallfiltersSRCNNmanytimesinade

3、epnetworkstructure,contextualinfor-36.6A+mationoverlargeimageregionsisexploitedinanefficientSelfExRFLway.Withverydeepnetworks,however,convergencespeed36.410210110010-110-2becomesacriticalissueduringtraining.Weproposeasim-slowrunningtime(s)fastpleyeteffectivetra

4、iningprocedure.Welearnresidualsonly4Figure1:OurVDSRimprovesPSNRforscalefactor2onanduseextremelyhighlearningrates(10timeshigherdatasetSet5incomparisontothestate-of-the-artmethods(SR-thanSRCNN[6])enabledbyadjustablegradientclipping.CNNusesthepublicslowerimpleme

5、ntationusingCPU).VDSROurproposedmethodperformsbetterthanexistingmeth-outperformsSRCNNbyalargemargin(0.87dB).odsinaccuracyandvisualimprovementsinourresultsareeasilynoticeable.end-to-endmanner.Theirmethod,termedSRCNN,doesnotrequireanyengineeredfeaturesthataretyp

6、icallyneces-1.Introductionsaryinothermethods[25,26,21,22]andshowsthestate-of-the-artperformance.Weaddresstheproblemofgeneratingahigh-resolution(HR)imagegivenalow-resolution(LR)image,commonlyWhileSRCNNsuccessfullyintroducedadeeplearningreferredassingleimagesupe

7、r-resolution(SISR)[12],[8],techniqueintothesuper-resolution(SR)problem,wefind[9].SISRiswidelyusedincomputervisionapplicationsitslimitationsinthreeaspects:first,itreliesonthecon-rangingfromsecurityandsurveillanceimagingtomedicaltextofsmallimageregions;second,trai

8、ningconvergestooimagingwheremoreimagedetailsarerequiredondemand.slowly;third,thenetworkonlyworksforasinglescale.arXiv:1511.04587v1[cs.CV]14Nov2015ManySISRmethodshavebeenstudiedinth