(a) (b) (c) (d) (e) (f) (g)
Figure 2.12: Comparativedemonstration ofskinmaps ofsomesampleimages: a) Originalimage, b)
Bayes lassier[13℄,)FASS [14℄,d) DSPF[9℄,e)SASS [15℄,f)MMSC[16℄, andg) ProposedDSM.
olour similarity.
Image SPM (inverted)
DSM Image SPM DSM Cost map
(inverted) Cost map
Figure 2.13: Unsatisfatoryostmaps for some sampleimagesseleted fromECU dataset.
non-skin pixels. Additionally, a dynami region growing (DRG) method is employed in our
method. The DRG allows skin regions to grow dynamially based on a feedbak input in
the form of false detetion. The main advantage of our proposed method is that the most
disriminantfeatures are extrated from animageitself, and hene,our methodis suitablefor
dierent bakgrounds. However insome ofthe instanes, skinregions have altogetherdierent
olour harateristis due to poor lighting onditions and/or oloured light soures. In these
spei situations, SPM fails to identify the skin pixels. Also, the proposed ADA may fail to
auratelydetetskinregionswhenthenon-skinregionshavealmostsimilarolourandtexture
harateristisasthatofskinregions. Figure2.13shows someofsuhaseswheretheproposed
method doesnot produe satisfatoryresults.
3
Detetion of Skin using Image Pixel
Distribution Information
The previous method an not perfetly handle non-uniform illumination as improper illumi-
nation makes skin regions to appear darker than its atual tone. Also, the derivation of an
aurate disriminative spae map of an image is possible only if the largest skin-like oloured
region inan image belongtoskin. To overomethese limitations of thepreviousmethod, anew
skindetetionmethodisproposedwhihisbasedonaloalskinmodeladaptionshemebyutiliz-
ing image pixel distribution information. The distribution image pixels in a given olour spae
is approximated by a Gaussian mixture model (GMM), whih is termed as image distribution
model (IDM). In this method, a loal skin distribution model (LSDM) and a loal bakground
distribution model (LBDM)are derived by nding the similarities of the Gaussian omponents
of IDMto agivenreferene skindistributionmodel. Thereferene skinmodelisderivedfrom a
setoffaialskinpixels,anditistermedas faialskindistributionmodel(FSDM).Subsequently,
a loal skin probability map (LSPM) is derived by using the LSDM and the LBDM. Finally,
the LSPM is fused with a skin probability map obtained from a global skin model by following
a fusion rule, and a fusion-based skin probability map (FSPM) is derived. The proposed DRG
algorithm in Chapter 2 is applied on the FSPM for additional redution in detetion errors.
Experimental results show that the proposed FSPM an better disriminate skin regions from
non-skin regions as ompared to the state-of-the-art methods.
3.1 Introdution
Theaurayofolour-basedskinsegmentationalgorithmsarelimitedduetothepreseneof
someoloursinthe bakground,whihare similartohumanskinolour,andpoorillumination
onditions. Skinlike-oloursinthe bakgroundinreases falsepositivedetetion. On the other
hand, poor illuminationmay altogether hange the hrominane properties of the atual skin
olours, and the atual skin olour will be dierent from the image olour. Many authors
like [99,120122℄ onverted the olour spae from RGB spae to another spae, dropped the
luminane omponent, and used only the hrominane omponents in order to ompensate
the brightness variations in the image. However, Storring et al. [123℄ showed that the skin
reetane lous and the illuminantlous are diretly related. This means that the pereived
olour depends onsene illumination.
To overome these problems, many researhers proposed loal adaptation shemes for a
global skin detetion model by utilizing loal hromati and/or textural informations. It is
observed that faialskin olour resembles the overall skin tone of a person. Motivated by this
fat, some researhers extrated pixels of the faial regions in and image and used them as
a referene skin tone for loal adaptation of the global skin detetion model. For example,
Kawulok et al.[19℄ proposed a dynami skinmodel, where the globalpixel statistisare fused
withloalstatsitisoffaialskinpixels. Yogarajahetal.[20℄proposedadynamithresholding-
based pproah by using hromati properties of faial skin pixels. Tan et al. [10℄ fused a
smoothed olour histogram and a Gaussian skin model with the help of faial skin pixels as
referene. Kawulok et al. [21℄ used a loal skin distribution model derived from faial skin
pixels toobtain seed regions fortheir propagation-based skin segmentation method. However,
skin appearane over a body hanges on aount of non-uniform illuminations. Therefore, a
skin model obtained only from the faial skin pixels annot generalize the overall skin olour
distribution of a person due toinsuient trainingskin samples.
In view of the above-mentioned issues relating to aurate skin region detetion, a novel
skin detetion algorithm is proposed by utilising the information of the distribution of image
pixels. We onsidered the priniplethat an image an be lustered into a numberof Gaussian
distributed lusters [124℄. So, the pixel distributionan be modelledas a mixtureof Gaussian
funtions. Theimagepixeldistributionmodelistermedasimagedistributionmodel (IDM).So,
the Gaussiandistributedlusters ofskin pixelsshouldbestatistiallyloser toarefereneskin
pixel distribution in a given olour spae. The referene skin pixel luster ould be either ob-
tainedfromaskinmodelderivedfromasetofskinpathesbelongingtodierentpeopleorfrom
the faial pixelsof aperson present inthe image. Subsequently, a new LoalSkin Probability
Map(LSPM) is derived by measuring the similaritybetween IDMand the referene skinpixel
distribution model. Finally, a fusion-based probability map is derived by ombining a global
skinprobabilitymap(GSPM)withtheproposedLSPM.Subsequently, atheproposeddynami
region growing (DRG) methodis applied inorder to improve the segmentation auray. The
proposed method isdisussed indetails inthe setiontofollow.