• Tidak ada hasil yang ditemukan

4.3 Experimental Analysis

4.3.2 Experimental validation

At rst, we examine the eet of

φ max

on detetion results. For this, detetion errors are

alulated by varying

φ max

, and the results are given in Table 4.1. It is observed that when

φ max = 0

, average false negative error

δ f n,avg

is maximum. For

φ max = 0

, the parameter

ξ = 1 8

. This implies that some of the lusters orresponding to the true skin pixels in

G I

are

inluded inLBDM,and/orexluded fromMSDM.In eitherof theseonditions, lusterswhose

entroids are very lose to that of FSDM are only seleted as skin lusters. An inrease in

φ max

value makes more skin-like lusters to be exluded from the bakground model, and/or

inluded inMSDM.Thus, falsenegativeerrorredues withaninrease in

φ max

value. However,

this inreases thehane offalseaeptaneerror. Someofthe lustersofskin-likebakground

Table 4.3: Comparative analysisof dierent methods

Method

δ f p,avg

(%)

δ f n,avg

(%)

δ t,avg

(%) Avg. A.

Bayes lassier[13℄ 19.69 34.22 53.91 0.7842

SDDMA[128℄ 4.71 58.08 62.80 0.8710

ASSC [21℄ 8.57 25.85 34.42 0.8919

DSPF [9℄ 21.67 22.37 44.04 0.7832

SASC [22℄ 11.70 45.15 56.84 0.8470

FSPM [23℄ 16.85 13.17 30.02 0.8373

FSDM + FBDM 5.33 15.14 20.47 0.9345

FSDM + MSDM + FBDM 5.78 8.19 13.97 0.9404

δ f p =

falsepositiveerror rate;

δ f n =

false negativeerror rate;

δ t =

total detetion error rate

= δ f p + δ f n

;A. = Auray

pixelsan beinluded inMSDM and/orexludedfromLBDM ifthe

φ max

isinreased further.

For example, the average false positive error

δ f p,avg

inreases signiantly for an inrease in

φ max

from

20

to

40

. On the other hand, the average false negative error

δ f n,avg

dereases

signiantlyfor aninrease in

φ max

from

0

to

20

. It isalsoobserved that for

φ max ≥ 20

, the

average total detetion error

δ t,avg

for all the videos beomes the lowest. Hene,

φ max

value is

xed at

20

for the remaininganalysis.

A omparative analysis is also performed between the original Bhattaharyya distane

(OBD) and the proposed Modied Bhattaharyya distane (MBD). Detetion errors are al-

ulated by both OBD and MBD. As mentioned in Eq. 4.8, the maximum value of MBD is

xed by OBD. The omparison between the OBD and the MBD inalulating SPM is Shown

in Table 4.2. In ontrast to OBD, detetion errors redue due the appliation of MBD. This

validates the eay of the proposed MBD inskin detetion.

Finally,theproposedmethodisomparedwiththestate-of-the-artmethods,suhasBayes

lassier[13℄,Fast propagation-basedskinsegmentation(FPSS)[14℄,Adaptiveseed-basedskin

lassiation (ASSC) [21℄, Skin detetion by dual maximization of detetors agreement (SD-

DMA) [128℄, Disriminative Skin-Presene Feature (DSPF) [9℄, Staked Autoenoders-based

skin lassiation(SASC) [22℄, and Fusion-based SkinProbabilityMap (FSPM) [23℄. The de-

tetionresultsobatinedbydierentmethodsaregiveninTable4.3. TheBayeslassiermethod

video #

1 2 3 4 5 6 7 8 9

δ t

0%

20%

40%

60%

80%

100%

120%

140%

Jones & Rehg (2002) ASSC (2013) DSPF (2014) SASC (2017) FSPM (2017) FSDM + FBDM FSDM + MSDM + FBDM

(a)

video #

1 2 3 4 5 6 7 8 9

Accuracy

0.2 0.4 0.6 0.8 1 1.2

1.4 Jones & Rehg (2002)

ASSC (2013) DSPF (2014) SASC (2017) FSPM (2017) FSDM + FBDM FSDM + MSDM + FBDM

(b)

Figure 4.11: Comparative barplots for dierentvideos: (a)

δ t

for dierent videos,and (b)Auray

for dierent videos.

δ fn (%)

0 1 2 3 4 5 6 7

δ fp (%)

0 20 40 60 80

100 Video # 1

Jones & Rehg (2002) ASSC (2013) DSPF (2014) FSPM (2017) FSDM + FBDM FSDM + MSDM + FBDM

(a)

δ fn (%)

0 1 2 3 4 5 6 7

δ fp (%)

0 10 20 30 40

50 Video # 2

Jones & Rehg (2002) ASSC (2013) DSPF (2014) FSPM (2017) FSDM + FBDM FSDM + MSDM + FBDM

(b)

δ fn (%)

0 5 10 15 20 25 30

δ fp (%)

0 10 20 30 40 50 60

70 Video # 3

Jones & Rehg (2002) ASSC (2013) DSPF (2014) FSPM (2017) FSDM + FBDM FSDM + MSDM + FBDM

()

δ fn (%)

0 5 10 15 20 25 30 35 40

δ fp (%)

0 20 40 60 80

100 Video # 4

Jones & Rehg (2002) ASSC (2013) DSPF (2014) FSPM (2017) FSDM + FBDM FSDM + MSDM + FBDM

(d)

δ fn (%)

0 20 40 60 80 100

δ fp (%)

0 20 40 60 80

100 Video # 5

Jones & Rehg (2002) ASSC (2013) DSPF (2014) FSPM (2017) FSDM + FBDM FSDM + MSDM + FBDM

(e)

δ fn (%)

0 5 10 15

δ fp (%)

20 30 40 50 60 70 80 90

100 Video # 6

Jones & Rehg (2002) ASSC (2013) DSPF (2014) FSPM (2017) FSDM + FBDM FSDM + MSDM + FBDM

(f)

Figure 4.12: ROC plots for dierent videos(16).

proposedbyJones andRehg[13℄isonsideredasabenhmark. Fortraining,themethodneeds

a set of skin and non-skin pixel samples, whih are obtained globally. However, the auray

of this method is totally dependent on the training sample set. The Jones & Rehg's method

δ fn (%)

5 10 15 20 25 30

δ fp (%)

0 10 20 30 40

50 Video # 7

Jones & Rehg (2002) ASSC (2013) DSPF (2014) FSPM (2017) FSDM + FBDM FSDM + MSDM + FBDM

(a)

δ fn (%)

0 5 10 15 20

δ fp (%)

0 10 20 30 40 50 60 70

80 Video # 8

Jones & Rehg (2002) ASSC (2013) DSPF (2014) FSPM (2017) FSDM + FBDM FSDM + MSDM + FBDM

(b)

δ fn (%)

0 5 10 15 20 25 30

δ fp (%)

0 10 20 30 40 50 60

70 Video # 9

Jones & Rehg (2002) ASSC (2013) DSPF (2014) FSPM (2017) FSDM + FBDM FSDM + MSDM + FBDM

()

Figure 4.13: ROC plots for dierent videos(79).

produes more false negatives as ompared to our proposed method. In order to ompare the

SDDMA methodwith the proposed method, itis trained with

N T F

number of labelled initial

frames of9videos (Intotal,

9 × N T F

frames). However, SDDMAfailssigniantlyindeteting

true skin regions. It produes highest false negatives among allthe benhmark methods. The

DSPF method an give omparatively better results than the standard SPM, and it gives a

disriminative spae-based skin map. The DSPF method mostly relies on SPM derived from

global training samples, and thus it is not adaptive to loal environmental onditions of an

image.

As human faialolouralmostresembles the overallskin olourof aperson, faedetetion-

basedskinmodeladaptationtehniqueangivesuperiorskindetetionperformane. TheASSC

method uses adaptive seeds for skin region growing. The adaptive seeds are derived from a

V# 1, F# 230 V# 2, F# 230 V# 3, F# 185

V# 4, F# 146

V# 5, F# 181 V# 6, F# 65

(a) (b) (c) (d) (e) (f) (g) (h)

V# 7, F# 199

V# 8, F# 219

V# 9, F#242

Figure 4.14: Comparative detetion results for minimum

δ t,avg

: (a) Original frames, (b) Jones and

Rehg [13℄,()ASSC[21℄,(d)DSPF[9℄,(e) SASC[22℄,(f)FSPM[23℄,(g)Proposedmethod,and(h)

Groundtruth. Here, white olour represents true positive, blak olour represents true negative, red

olour representsfalse positive,andgreen olour representsfalse negative.

loal skin model (derived from faial skin pixels). The ASSC method relies on a standard

SPM for region growing, and hene, it is unable to detet many true skin pixels. So, this

method produes higher false negative errors than the proposed method. The FSPM-based

method gives better detetion auray by utilizing image pixel distribution information to

derive a loalSPM (LSPM). The LSPM is later fused with the original or global SPM to get

the FSPM.The FSPM isabletodetetmore trueskin pixelsasomparedtoother benhmark

methods. In our method, the ombination of FSDM and FBDM gives a stati skin detetion

modelforvideos. TheinorporationofMSDMmakestheskindetetionmodeladaptivetoloal

illuminationhanges, and a dynami skin detetion model for videos is obtained. The stati

(a)

OQ R ST U S VU S WU SXU S YU S ZU

V# 7,

F# 199

V# 8, F# 219

V# 9, F#242 V# 1, F# 230 V# 2, F# 230 V# 3, F# 185

V# 4, F# 146

V# 5, F# 181

V# 6, F# 65

Figure 4.15: Comparative detetion results for maximum auray: (a) Original frames, (b) Jones

andRehg[13℄,()ASSC[21℄,(d)DSPF[9℄,(e)SASC[22℄,(f)FSPMetal.[23℄,(g)Proposedmethod,

and(h)Groundtruth. Here,whiteolourrepresentstruepositive,blakolourrepresentstruenegative,

red olour represents falsepositive,and greenolour representsfalse negative.

model isprone toprodue more false rejetions as ompared to itsdynami ounterpart. The

dynamiadaptation ofthe proposed MSDM forloalilluminationhangesmakesthe proposed

system more robust to unonstrained illumination and bakground onditions. In Fig. 4.12

4.13, omparative ROC plots are shown for dierent videos. In Fig. 4.144.15, the detetion

results are shown for allthe test videos. It is evident from allthe experiential results that the

proposedmethodan detetskinregionsinavideomoreaurately thantheexisting standard

methodseven inthe presene of loalolour deformations.