SALIENCY MAP BASED IMPROVED SEGMENTATION

Presented at IEEE International Conference on Image Processing (ICIP), 2015, Quebec, Canada

1. RESEARCH POSTER PRESENTATION DESIGN © 2012
www.PosterPresentations.com
Inthispaperwepresentanovelapproachforrefiningsegmentation
usingsaliencymap.Toachievethis,wefirstdevelopanewsaliency
detectionmethodbasedoncuesatvariouslevels.Initially
preprocessingstepisdoneusingnon-linearanisotropicdiffusion
filteringinordertopreservetheedgeinformationintheforeground
salientobjectsandsmoothenthebackground.Thenweapplygrab
cutsegmentationusingsaliencymapastheinputtogetimproved
segmentation.Repeatedapplicationoftheschemeisusedformulti-
objectsegmentation.Theexperimentalresultsforthesaliency
techniqueshowhighprecisionandrecallratesagainstthestate-of-
the-artmethods.
ABSTRACT
Saliencyisameasureofthemostconspicuousregionorobjectwhich
standsoutdistinctlyinanimage.Attentionalselectioncanbe
essentiallyviewedassaliencydetectionwhichisbasedoncertain
cuesofthesensoryinformation.Thesevisualattentionmechanisms
canbedrivenbytopdown(memory-driven)orbottomup(memory-
free)influences.
Anobjectischaracterizedbyawelldefinedboundaryandits
relativeconspicuityagainstthebackground.Thus,saliencybecomes
animportantpropertyforsegmentingobjects.Theimprovementin
thesaliencymapismadepossiblebyfusionofvarioussaliencycues
atlocal,globalandraritylevel.Further,toincreasethe
discriminabilityoftheproposedsaliencymap,weapplyan
anisotropicdiffusionfilteringtotheimage.
INTRODUCTION
ASD-1000andSOD-300databaseisusedforexperimentation.ASDdatasetcontains1000imagesfromMSRA(MicrosoftResearchAsia)dataset.
SODdatasetcontains300imagesfromBSD(Berkeleysegmentationdataset).
Fig.1.a)Originalimageb)Non-linearanisotropicdiffusion
filteredimagec)Normalizedcolormapd)Normalizedintensity
mape)Normalizedorientationmapf)Depthmapg)Local
featuresmap
Prerana Mukherjee*, Brejesh Lall, Archit Shah
*e-mail: [email protected]
SALIENCY MAP BASED IMPROVED SEGMENTATION
Department of Electrical Engineering, Indian Institute of Technology, Delhi, India.
RESULTS
Fig.5.a)Originalimageb)Binarisedgroundtruthc)Groundtruth
labeld)MaximumSymmetricSurrounde)GBVSf)Itti-Kochg)Seg
h)FrequencyTunedi)Houj)ContextAwarek)SUNl)Our
Approach
Fig. 6. Performance measures for a) ASD-1000 database b) SOD-300 database
PROPOSED APPROACH
CONCLUSION
•Bettersegmentationresults,weusenon-linearanisotropic
diffusionfilteringwhichsubstantiallyreducesthetimecomplexity
forcomputationofsaliencymapandimprovestheperformance
rates.
•Targetmulti-objectsegmentationusingdifferentlevelsofcuesof
saliency.
•Theproposedsaliencymaptechniquegivesrelativelyhighresults
(precisionandrecallrates)comparedtothepriorstate-of-the-art
methods.
FEATURE MAPS
The scheme consists of the following steps:
Fig.2.a)Meanshiftsegmentedimageb)GlobalContrastmapc)
SpatialSparsitymapd)Globalfeaturesmape)PQFTrarity
mapf)Combinedsaliencymap
1. Preprocess image using non linear anisotropic scale space
filtering.
2. Compute the saliency map of the preprocessed image using the
modified saliency map.
3. Obtain global variance of the saliency map.
4. If var>Threshold Gotostep 5 else Stopno further segmentation
of the region is possible.
5. Binarizethe saliency map and use the two regions as initial
input for grab-cut segmentation.
6. Apply bounding box to the two regions and generate two images,
each containing the pixels lying within the bounding box.
7. Repeat for the two images generated in Step 6.Fig.3.a)Originalimage(ASD)b)Saliencymapc)Binarisedsaliency
mapd)Initialsegmentedimagee)Originalimage(SOD)f)Ground
Truthg)Saliencymap
Fig.4.a)Originalimageb)PQFTc)PFTd)PCT
REFERENCE
1.L.Itti,C.KochandE.Niebur,“AModelofsaliencybasedvisualattentionforrapidsceneanalysis,”InIEEETrans.OnPatternAnalysisandMachine
Intelligence,vol.20,no.11,pp.1254-1259,November1998.
2.R.Achanta,S.Hemami,F.Estrada,andS.Susstrunk,“Frequency-tunedsalientregiondetection,”InCVPR,pp.1597–1604,2009.
3.P.PeronaandJ.Malik,“Scale-spaceandedgedetectionusinganisotropicdiffusion,”IEEETrans.PatternAnal.MachineIntell.,vol.12,no.7,pp.629–
639,1990.