On Classification of Distorted Images with Deep Convolutional Neural Networks

0.96 MB·

by  Yiren Zhou

ONCLASSIFICATIONOFDISTORTEDIMAGESWITHDEEPCONVOLUTIONALNEURAL
NETWORKS
YirenZhou,SiboSong,Ngai-ManCheung
SingaporeUniversityofTechnologyandDesign
ABSTRACT Some previous works have studied the effect of image distor-
Imageblurandimagenoisearecommondistortionsduringim- tion[10].FocusingonDNN,Basuetal.[11]proposedanewmodel
7 ageacquisition. Inthispaper,wesystematicallystudytheeffectof modifiedfromdeepbeliefnetstodealwithnoisyinputs. Theyre-
1 imagedistortionsonthedeepneuralnetwork(DNN)imageclassi- portedgoodresultsonanoisydatasetcalledn-MNIST,whichcon-
0 fiers. First,weexaminetheDNNclassifierperformanceunderfour tainsGaussiannoise,motionblur,andreducedcontrastcomparedto
2 typesofdistortions. Second,weproposetwoapproachestoallevi- originalMNISTdataset. Recently,DodgeandKaram[12]reported
n ate the effect of image distortion: re-training and fine-tuning with the degradation due to various image distortions in several DNN.
a noisy images. Our results suggest that, under certain conditions, Comparedtotheseworks,weperformaunifiedstudytoinvestigate
J fine-tuningwithnoisyimagescanalleviatemucheffectduetodis- effectofimagedistortionon(i)hand-writtendigitclassificationand
8 tortedinputs,andismorepracticalthanre-training. (ii) natural image classification. Moreover, we examine using re-
Index Terms— Image blur; image noise; deep convolutional trainingandfine-tuningwithnoisyimagestoalleviatetheeffect.
] In classification of “clean” images (i.e., without distortion),
neuralnetworks;re-training;fine-tuning
V some previous work has attempted to introduce noise to the train-
C 1. INTRODUCTION ing data [13, 14]. In these works, their purpose is to use noise to
. Recently,deepneuralnetworks(DNNs)haveachievedsuperiorre- regularizethemodelinordertopreventoverfittingduringtraining.
s sults on many computer vision tasks [1]. In image classification, On the contrary, our goal is to understand the benefits of using
c
[ DNN approaches such as Alexnet [2] have significantly improved noisytrainingdatainclassificationofdistortedimages. Ourresults
theaccuracycomparedtopreviouslyhand-craftedfeatures. Further alsosuggestthat, undercertainconditions, fine-tuningusingnoisy

1 worksonDNN[3,4]continuetoadvancetheDNNstructuresand imagescanbeaneffectiveandpracticalapproach.
v improvetheperformance.
2. DEEPARCHITECTURE
4 Inpracticalapplications,varioustypesofdistortionsmayoccur
2 inthecapturedimages. Forexample,imagescapturedwithmoving In this section, we briefly introduce the idea of deep neural net-
9 cameras may suffer from motion blur. In this paper, we system- work (DNN). There are many types of DNN, here we mainly in-
1 atically study the effect of image distortion on DNN-based image troducedeepconvolutionalneuralnetwork(DCNN),adetailedin-
0 classifiers.Wealsoexaminesomestrategytoalleviatetheimpactof troductionforDNNcanbefoundin[15].
.
1 imagedistortionontheclassificationaccuracy. DNNisamachinelearningarchitecturethatisinspiredbyhu-
0 Two main categories of image distortions are image blur and mans’centralnervoussystems. ThebasicelementinDNNisneu-
7 imagenoise[5].Theyarecausedbyvariousissuesduringimageac- ron. InDNN,neighborhoodlayersarefullyconnectedbyneurons,
1 quisition. Forexample,defocusbluroccurswhenthecameraisout andoneDNNcanhavemultipleconcatenatedlayers. Thoselayers
: offocus. Motionbluriscausedbyrelativemovementbetweenthe togetherformaDNN.
v
cameraandtheview,whichiscommonforsmartphone-basedimage DNNhasachievedgreatperformanceforproblemsonsmallim-
i
X analysis[6,7,8]. Imagenoiseisusuallycausedbypoorillumina- ages[16]. However, forproblemswithlargeimages, conventional
r tionand/orhightemperature,whichdegradetheperformanceofthe DNN need to use all the nodes in the previous layer as inputs to
a chargecoupleddevice(CCD)insidethecamera. thenextlayer,andthisleadtoamodelwithaverylargenumberof
WhenweapplyaDNNclassifierinapracticalapplication,itis parameters,andimpossibletotrainwithalimiteddatasetandcom-
possiblethatsomeimageblurandnoisewouldoccurintheinputim- putationsources. Theideaofconvolutionalneuralnetwork(CNN)
ages. ThesedegradationswouldaffecttheperformanceoftheDNN is to make use of the local connectivity of images as prior knowl-
classifier. Our work makes several contributions to this problem. edge,thatanodeisonlyconnectedtoitsneighborhoodnodesinthe
First, we study the effect of image distortion on the DNN classi- previouslayer. Thisconstraintsignificantlyreducesthesizeofthe
fier. WeexaminetheDNNclassifierperformanceunderfourtypes model,whilepreservingthenecessaryinformationfromanimage.
ofdistortionsontheinputimages: motionblur,defocusblur,Gaus- Foraconvolutionallayer,eachnodeisconnectedtoalocalre-
siannoiseandacombinationofthem. Second,weexaminetwoap- gion in the input layer, which is called receptive field. All these
proachestoalleviatetheeffectofimagedistortion.Inoneapproach, nodesformanoutputlayer. Forallthesenodesintheoutputlayer,
were-trainthewholenetworkwiththenoisyimages. Wefindthat they have different kernels, but they share the same weights when
thisapproachcanimprovetheaccuracywhenclassifyingdistorted calculatingactivationfunction.
images.However,re-trainingrequireslargetrainingdatasetsforvery Fig.1showsthearchitectureofLeNet-5,whichisusedfordigit
deepnetworks.Inspiredby[9],inanotherapproach,wefine-tunethe imageclassificationonMNISTdataset[17].Fromthefigurewecan
firstfewlayersofthenetworkwithdistortedimages.Essentially,we seethatthemodelhastwoconvolutionallayersandtheircorrespond-
adjustthelow-levelfiltersoftheDNNtomatchthecharacteristics ingpoolinglayers.Thisistheconvolutionalpartforthemodel.The
ofthedistortedimages. followingtwolayersareflattenandfullyconnectedlayers,theselay-
Infine-tuning,westartfromthepre-trainedmodeltrainedwith
the original dataset (i.e., images without distortion). We fine-tune
the first N layers of the model on a distorted dataset while fixing
the parameters in the remaining layers. The reason to fix param-
eters in the last layers is that image blur and noise are considered
tohavemoreeffectonlow-levelfeaturesinimages, suchascolor,
edge,andtexturefeatures. However,thesedistortionshavelittleef-
Fig.1.StructureofLeNet-5.
fectonhigh-levelinformation,suchasthesemanticmeaningsofan
image[21].Therefore,infine-tuning,wefocusonthestartinglayers
ofaDNN,whichcontainmorelow-levelinformation. Asanexam-
ple, for LeNet-5 we have 4 layers with parameters, that means N
israngingfrom1to4. Wedenotefine-tuningmethodsasfirst-1to
first-4.
In re-training, we train the whole network with the distorted
Fig.2.StructureofCIFAR10-quickmodel. datasetfromscratchanddonotusethepre-trainedmodel.Wedenote
there-trainingmethodasre-training.
For re-training LeNet-5, we set the learning rate to 10−3, and
ersareinheritedfromconventionalDNN.
thenumberofepochsto20. Forfine-tuning,wesetlearningrateto
3. EXPERIMENTALSETTINGS 10−5(1%ofthere-traininglearningrate),andnumberofepochsto
We conduct experiment on both relatively small datasets [17, 18] 15.Eachepochtakesabout1minute,sothetrainingproceduretakes
andalargeimagedataset,ImageNet[19].Weexaminedifferentfull about20minutesforre-training,and15minutesforfine-tuning.
training/fine-tuningconfigurationsonsomesmalldatasetstogain CIFAR-10 dataset consists of 60000 32×32 color images in
insight into their effectiveness. We then examine and validate our 10classes,with6000imagesperclass. 50000aretrainingimages,
approachonImageNetdataset. and10000aretestimages.Tomakethetrainingfaster,weuseafast
We conduct the experiment using MatConvNet [20], a MAT- model provided in MatConvNet [20]. The structure of CIFAR10-
LAB toolbox which can run and learn convolutional neural net- quickmodelisshowninFig.2.
works. AlltheexperimentsareconductedonaDellT5610Work- SimilartopreviousapproachesforMNIST,weusefine-tuning
StationwithIntelXeonE5-2630CPU. andre-trainingforCIFARdistorteddataset. Thereare5layerswith
parameters in CIFAR10-quick model, so we have first-1 to first-5
as fine-tuning methods. The re-training method is denoted as re-
n training.
o
otiur Forre-trainingCIFAR10-quick,wesetthenumberofepochsto
Mbl 45. Learningrateissetto5×10−2 forfirst30epochs,5×10−3
for the following 10 epochs, and 5×10−4 for the last 5 epochs.
Forfine-tuning,wesetthenumberofepochsto30. Learningrateis
s
cu 5×10−4forfirst25epochs,and5×10−5forlast5epochs. Each
o
efur epochtakesabout3minutes,sothetrainingproceduretakesabout
Dbl 135minutesforre-training,and90minutesforfine-tuning.
n
a
Gaussinoise Motionblur
e
bin us
m c
Allco Defoblur
Fig. 3. Example MNIST images after different amount of motion
blur,defocusblur,Gaussiannoise,andallcombined.
n
a
Deep architectures and datasets: In this evaluation we con- ssie
us
siderthreewell-knowndataset: MNIST[17], CIFAR-10[18], and Ganoi
ImageNet[19].
MNISTisahandwrittendigitsdatasetwith60000trainingim-
ages and 10000 test images. Each image is a 28×28 greyscale ne
image,belongingtoonedigitclassfrom’0’to’9’. ForMNIST,we bi
m
useLeNet-5[17]forclassification.ThestructureofLeNet-5weuse Allco
isshowninFig.1.Thisnetworkhas6layersand4ofthemhavepa-
Fig.4.ExampleCIFAR-10imagesafterdifferentamountofmotion
rameterstotrain:thefirsttwoconvolutionallayers,flattenandfully
blur,defocusblur,Gaussiannoise,andallcombine.
connectedlayers.
Weconsidertwoapproachestodealwithdistortedimages:fine- Here we also present an evaluation on ILSVRC2012 dataset.
tuningandre-trainingwithnoisyimages. ILSVRC2012[22]isalarge-scalenaturalimagedatasetcontaining
morethanonemillionimagesin1000categories. Theimagesand 4. EXPERIMENTALRESULTSANDANALYSIS
categoriesareselectedfromImageNet[19].Tounderstandtheeffect
Fig.6and7showtheresultsofourexperiment.Wecompare3meth-
of limited data in many applications, we randomly choose 50000
ods: notrainmeansthatthemodelistrainedonthecleandataset,
imagesfromtrainingdatasetfortraining, andusevalidationsetof
whiletestedonthenoisydataset. first-N meansthatwefine-tuning
ILSVRC2012,whichcontains50000images,fortesting.
thefirstNlayerswhilefixingtheremaininglayersinthenetwork.
Weusefine-tuningmethodforILSVRC2012validationsetwith ForLeNet-5network,thereare4trainablelayers,sowehavefirst-1
apre-trainedAlexnetmodel[2]. Wedonotusere-trainingmethod tofirst-4,forCIFAR10-quicknetwork,wehavefirst-1tofirst-5.
here,becausere-trainingAlexnetusingonlysmallpartofthetrain- ResultsonMNIST:Fig.6showstheresultsonMNISTdataset.
ingsetofILSVRC2012wouldcauseoverfitting. Wefine-tunethe FormotionblurandGaussiannoise,theeffectofdistortionisrela-
first3layersofAlexnet,whilefixingtheremaininglayers.Forfine- tivelysmall(notethatthescalesofdifferentplotsaredifferent).De-
tuningprocess,thenumberofepochsissetto20. Thelearningrate focusblurandcombinednoisehavemoreeffectonerrorrate. This
issetto10−8to10−10fromepoch1toepoch20,decreasesbylog result is consistent with the observation on Fig. 3, that the motion
space.Wealsouseaweightdecayof5×10−4.Approximatetrain- blurandGaussiannoiseimagesaremorerecognizablethandefocus
ingtimeis90minutesforeachepoch,and30hoursfortotalprocess. blurandcombinednoise.MNISTdatasetcontainsgreyscaleimages
Regardingthecomputationtime,fine-tuningtakeslesstimethan withhandwrittenstrokes, soedgesalongthestrokesareimportant
re-trainingontheMNISTandCIFAR-10dataset.ForILSVRC2012 features. In our experiment, the stroke after defocus blur covers a
validation set, we also need to use fine-tuning method in order to widerarea, whileweakens theedgeinformation. The motionblur
preventoverfitting. also weakens edge information, but not as severe as defocus blur.
Thisisbecause,underthesameparameter,theareaofmotionbluris
smallerthanthedefocusblur. Gaussiannoisehaslimitedeffecton
theedgeinformation,sotheerrorratehaslittleincrease. Combined
noisehavemuchimpactontheerrorrate.
Both fine-tuning and re-training methods can significantly re-
duceerrorrate. first-3andfirst-4haveverysimilarresults,indicat-
ingthatdistortionhaslittleeffectonthelastseverallayers. When
thedistortionissmall,fine-tuningbyfirst-3andfirst-4achievecom-
parableresultswithre-training.Whenthedistortionlevelincreases,
(a) (b)
Fig. 5. Example images from ImageNet validation set. (a) is the re-trainingachievesabetterresult.
originalimage.(b)isthedistortedimage. ResultsonCIFAR-10:FromFig.4wecanseethedistortionsin
CIFAR-10notonlyaffecttheedgeinformation,butalsohaveeffect
Typesofblurandnoise: Inthisexperiment,weconsidertwo oncolorandtextureinformation.Therefore,all3typesofdistortion
typesofblur: motionbluranddefocusblur,andonetypeofnoise: canmaketheimagesdifficulttorecognize. Thisisconsistentwith
Gaussiannoise. the results shown in Fig. 7. Different from the results on MNIST
Motionblurisatypicaltypeofblurusuallycausedbycamera dataset, all 3 types of distortion significantly worsen the error rate
shakingand/orfast-movingofthephotographedobjects. Wegener- onnotrainresult.
ate the motion blur kernel using random walk [23]. For each step Usingbothfine-tuningandre-trainingmethodscansignificantly
size, we move the motion blur kernel in a random direction by 1- reducetheerrorrate. first-3tofirst-5givesimilarresults,indicating
pixel.Thesizeofthemotionblurkernelissampledfrom[0,4]. that the distortion mainly affects the first 3 layers. When the dis-
Defocus blur happens when the camera loses focus of an im- tortionlevelislow,fine-tuningandre-traininghavesimilarresults.
age.Wegeneratethedefocusblurbyuniformanti-aliaseddisc.The However,whenthedistortionlevelishighorundercombinednoise,
radiusofthediscissampledfrom[0,4]. re-traininghasbetterresultsthanfine-tuning.
From both figures we can observe that when we fine-tune the
Aftergeneratingamotionoradefocusblurkernelforoneimage,
first 3 layers, the results are very similar to fine-tuning the whole
weusethiskernelforconvolutionoperationonthewholeimageto
networks.Thisresultindicatesthatimagedistortionhasmoreeffect
generateablurredimage.
onthelow-levelinformationoftheimage, whileithaslittleeffect
Gaussiannoiseiscausedbypoorilluminationand/orhightem-
onhigh-levelinformation.
perature,whichpreventsCCDinacamerafromgettingcorrectpixel
Analysis: To gain some insight into the effectiveness of fine-
values. WechooseGaussiannoisewithzeromeans,andwithstan-
tuning and re-training on distorted data, we look into the statistics
darddeviationσsampledfrom[0,4]onacolorimagewithaninteger
ofthefeaturemapinsidethemodel. Inspiredby[24], wefindthe
valuein[0,255].
meanvarianceofimagegradientmagnitudetobeausefulfeature.
Finally,weconsideracombinationofalltheabovethreetypes
Instead of calculating the image gradient, we calculate the feature
of distortions. The value of each noise is sampled from [0,4], re-
mapgradient. Then,wecalculatethemeanvarianceoffeaturemap
spectively.
gradientmagnitude.
Fig.3and4showtheexampleimagesofblurandnoiseeffectsin Given a feature map fm as input, we first calculate gradient
MNISTandCIFAR-10,respectively.Eachrowofimagesrepresents alonghorizontal(x)andverticaldirectionsusingSobelfilters
onetypeofdistortion.Forthefirst3rows,onlyonetypeofdistortion 1(cid:0)10−1(cid:1) 1(cid:0) 1 2 1 (cid:1)
is applied, and for the last row, we apply all 3 types of distortion sx = 4 2100−−21 ,sy = 4 −01−02−01 (1)
on one single image. As we see each row from left to right, the
Thenwehavegradientmagnitudeoffmatlocation(m,n)as
distortionlevelincreasesfrom0to4.
Fig.5showsanexampleinILSVRC2012validationset. When (cid:112)
g (m,n)= (fm⊗s )2(m,n)+(fm⊗s )2(m,n) (2)
wegeneratethedistorteddataset,eachimageintrainingandtesting fm x y
sethasrandomdistortionvaluessampledfrom[0,4]forall3types
ofdistortion.
1.8 25 1.05 40
no train
)%( etaR rorrE111...2461 ffffriiiierrrrsssstrtttta----1234in )%( etaR rorrE1120505 )%( etaR rorrE0.09.591 )%( etaR rorrE123000
0.85
0.8 0 0 1 2 3 4 0
0 1 2 3 4 0 1 2 3 4 < 0 1 2 3 4
Motion Blur Size Defocus Blur Size Gaussian Noise Combined Noise
(a) (b) (c) (d)
Fig.6.ErrorratesforLeNet-5modelonMNISTdatasetunderdifferentblursandnoises.
60 no train 80 70 100
)%( etaR rorrE23450000 fffffriiiiierrrrrssssstrttttta-----12345in )%( etaR rorrE246000 )%( etaR rorrE2345600000 )%( etaR rorrE 24680000
10
10 0 0 1 2 3 4 0
0 1 2 3 4 0 1 2 3 4 < 0 1 2 3 4
Motion Blur Size Defocus Blur Size Gaussian Noise Combined Noise
(a) (b) (c) (d)
Fig.7.ErrorratesforCIFAR10-quickmodelonCIFARdatasetunderdifferentblursandnoises.
1.1 1.1 agescanbeclosetotheresultsonoriginalimages.Whenweretrain
edutingam1.051 edutingam1.051 tthoertemdoddaetlaosent,daisntdorttheedfiemaatugrees,mwaeptrreyptroesfietnttahteioDnNisNnmotondeecleosnsadriilsy-
tneidarg fo ecnairaV00..0089..55890 nfffffriiiiierrrrrossssst rtttttta-----r112345ainin 2 3 4 tneidarg fo ecnairaV00..0089..55890 nfffffriiiiierrrrrossssst rtttttta-----r112345ainin 2 3 4 cTalnaodbselfienteo1er.-rttohuArencrreacedtupermra(ec%osyed)nectloaomticnoploenIarmairongiafsigonoenarNlidgbmieiestnttoowavdrlaeetlieelimddnaaptgireoecfisn-l.ntesreaae-nittn.uendeddAismlteoxorndteeedtlmodel
Motion Blur Size Defocus Blur Size
(a) (b) data data data data
Fig.8. Meanvarianceoffeaturemapgradientmagnitudeforconv top-1error 42.9 53.1 42.9 47.7
layer 3 of CIFAR10-quick model. (a): motion-blur. (b): defocus top-5error 20.1 28.2 20.4 23.6
blur.
ResultsonImagenet: Wealsoexaminetheefficiencyoffine-
tuningonalargedatasetandaverydeepnetwork. Forexperiment
Afterwehavethegradientmagnitudeg forfeaturemapfm, onthetrainingandvalidationsetofILSVRC2012,wegeneratedthe
fm
wecalculatethevarianceofgradientmagnitude:v =var(g ). distorteddatabycombiningall3typesofblur/noise. Foreachim-
fm fm
age,andforeachtypeofdistortion,thedistortionlevelisuniformly
When we apply defocus blur or motion blur on an image, the
sampledfrom[0,4].Afterobtainingthedistorteddata,wefine-tune
clear edges are smeared out into smooth edges, thus the gradient
thefirst3layersofapre-trainedAlexnetmodel[2]. Table1shows
magnitudemapbecomessmooth, andhaslowervariance. Feature
mapswithhighergradientvariancevaluev areconsideredtohave the accuracy comparison between the original pre-trained Alexnet
fm
model and the fine-tuned model. Compared with the original pre-
moreedgeandtextureinformation,thusmorehelpfulforimagerep-
resentation. While lower v value indicates that the information trained model, the fine-tuned model increases the performance on
fm
distorteddata,whilekeepingtheperformanceoncleandata. When
insidethefeaturemapislimited,thusnotsufficientforimagerepre-
we want to use a large DNN model like Alexnet on a limited and
sentation.
distorteddataset,fine-tuningonfirstfewlayerscanincreasemodel
Fig. 8 shows the mean variance of feature map gradient mag-
accuracyondistorteddata,whilemaintainingtheaccuracyofclean
nitude for conv3 layer(the last conv layer) of the CIFAR10-quick
data.
model. From the two figures we observe that: (1) When apply-
ingoriginalmodelondistortedimages,themeanvariancedecreases 5. CONCLUSIONS
comparedtoapplyingthemodelonoriginalimages(seenotrain),
Fine-tuningandre-trainingthemodelusingnoisydatacanincrease
suggesting that edge or texture information is lost because of the
the model performance on distorted data, and re-training method
distortion. (2) When applying fine-tuning method to the distorted
usually achieves comparable or better accuracy than fine-tuning.
images,themeanvariancemaintainssimilarasthatoforiginalim-
However,thereareissuesweneedtoconsider:
ages,suggestingthatbyfine-tuningondistortedimages,themodel
• The size of the distorted dataset: If the model is very deep
canextractusefulinformationfromthedistortedimages. (3)When
andthesizeofdistorteddatasetissmall,trainingthemodel
applyingretrainingmethodondistortedimages,themeanvariance
onthelimiteddatasetwouldleadtooverfitting. Inthiscase,
ishigherthanapplyingthemodelonoriginalimages. Itmeansthat
wecanfine-tunethemodelbyfirstNlayerswhilefixingthe
the retrained model fits the distorted image dataset. These results
remaininglayerstopreventoverfitting.
suggestthatwhenwefine-tunethemodelondistortedimages, we
trytomakethefeaturemaprepresentationofdistortedimagesclose • The distortion level of noise: When the distortion level is
tooriginalimages,sothattheclassificationresultsondistortedim- high, re-training on distorted data has better results. When
thedistortionlevelislow,bothre-trainingandfine-tuningcan [14] Yixin Luo and Fan Yang, “Deep learning with noise,”
achievegoodresults. Andinthiscase,fine-tuningisprefer- http://www.andrew.cmu.edu/user/fanyang1/
ablebecauseitconvergesfaster,whichmeanslesscomputa- deep-learning-with-noise.pdf,2014.
tiontime,andisapplicabletolimitedsizedistorteddatasets.
[15] “DeepLearning documentation,” http://
deeplearning.net/tutorial/contents.html,
6. REFERENCES 2016.
[16] Geoffrey E Hinton and Ruslan R Salakhutdinov, “Reducing
[1] AliRazavian,HosseinAzizpour,JosephineSullivan,andSte-
thedimensionalityofdatawithneuralnetworks,”Science,vol.
fanCarlsson, “Cnnfeaturesoff-the-shelf:anastoundingbase-
313,no.5786,pp.504–507,2006.
lineforrecognition,” inProceedingsoftheIEEEConference
onComputerVisionandPatternRecognitionWorkshops,2014, [17] Yann LeCun, Le´on Bottou, Yoshua Bengio, and Patrick
pp.806–813. Haffner, “Gradient-basedlearningappliedtodocumentrecog-
nition,” ProceedingsoftheIEEE,vol.86, no.11, pp.2278–
[2] AlexKrizhevsky,IlyaSutskever,andGeoffreyEHinton,“Ima-
2324,1998.
genetclassificationwithdeepconvolutionalneuralnetworks,”
in Advances in neural information processing systems, 2012, [18] AlexKrizhevskyandGeoffreyHinton,“Learningmultiplelay-
pp.1097–1105. ersoffeaturesfromtinyimages,”2009.
[19] J.Deng, W.Dong, R.Socher, L.-J.Li, K.Li, andL.Fei-Fei,
[3] K. Simonyan and A. Zisserman, “Very deep convolutional
networks for large-scale image recognition,” CoRR, vol. “ImageNet: ALarge-ScaleHierarchicalImageDatabase,” in
CVPR09,2009.
abs/1409.1556,2014.
[20] A.VedaldiandK.Lenc, “Matconvnet–convolutionalneural
[4] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun,
networksformatlab,” .
“Deepresiduallearningforimagerecognition,”arXivpreprint
arXiv:1512.03385,2015. [21] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang,
“Deeplearningfaceattributesinthewild,” inProceedingsof
[5] Antoni Buades, Bartomeu Coll, and Jean-Michel Morel, “A
theIEEEInternationalConferenceonComputerVision,2015,
reviewofimagedenoisingalgorithms,withanewone,”Multi-
pp.3730–3738.
scaleModeling&Simulation,vol.4,no.2,pp.490–530,2005.
[22] OlgaRussakovsky, JiaDeng, HaoSu, JonathanKrause, San-
[6] Hossein Nejati, V Pomponiu, Thanh-Toan Do, Yiren Zhou,
jeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy,
S Iravani, and Ngai-Man Cheung, “Smartphone and mobile
Aditya Khosla, Michael Bernstein, Alexander C. Berg, and
imageprocessingforassistedliving,” IEEESignalProcessing
LiFei-Fei, “ImageNetLargeScaleVisualRecognitionChal-
Magazine,pp.30–48,2016.
lenge,” InternationalJournalofComputerVision(IJCV),vol.
[7] V Pomponiu, H Nejati, and Ngai-Man Cheung, “Deepmole: 115,no.3,pp.211–252,2015.
Deepneuralnetworksforskinmolelesionclassification,” in
[23] MichalHradisˇ,JanKotera,PavelZemc´ık,andFilipSˇroubek,
Proc. IEEE International Conference on Image Processing
“Convolutionalneuralnetworksfordirecttextdeblurring,” in
(ICIP),2016.
ProceedingsofBMVC,2015,pp.2015–10.
[8] Thanh-ToanDo,YirenZhou,HaitianZheng,Ngai-ManChe-
[24] Zhong Zhang and Shuang Liu, “Gmvp: gradient magnitude
ung,andDawnKoh, “Earlymelanomadiagnosiswithmobile
andvariancepooling-basedimagequalityassessmentinsensor
imaging,”inProc.36thAnnualInternationalConferenceofthe
networks,” EURASIP Journal on Wireless Communications
IEEEEngineeringinMedicineandBiologySociety(EMBC),
andNetworking,vol.2016,no.1,pp.1,2016.
2014.
[9] JasonYosinski, JeffClune, YoshuaBengio, andHodLipson,
“Howtransferablearefeaturesindeepneuralnetworks?,” in
Advancesinneuralinformationprocessingsystems,2014,pp.
3320–3328.
[10] YirenZhou,Thanh-ToanDo,HZheng,Ngai-ManCheung,and
LuFang, “Computationandmemoryefficientimagesegmen-
tation,” IEEETransactionsonCircuitsandSystemsforVideo
Technology,2016.
[11] SaikatBasu,ManoharKarki,SangramGanguly,RobertDiB-
iano, Supratik Mukhopadhyay, and Ramakrishna Nemani,
“Learning sparse feature representations using probabilistic
quadtreesanddeepbeliefnets,” inProceedingsoftheEuro-
peanSymposiumonArtificialNeuralNetworks,ESANN,2015.
[12] Samuel Dodge and Lina Karam, “Understanding how im-
age quality affects deep neural networks,” arXiv preprint
arXiv:1604.04004,2016.
[13] Salah Rifai, Xavier Glorot, Yoshua Bengio, and Pascal Vin-
cent, “Adding noise to the input of a model trained with a
regularizedobjective,” arXivpreprintarXiv:1104.3250,2011.