教学文库网 - 权威文档分享云平台
您的当前位置:首页 > 文库大全 > 专业资料 >

基于皮肤模板和改进HMM的自动人脸识别系统(2)

来源:网络收集 时间:2026-08-04
导读: 第25卷 videsfaceimagesintoa seriesofoverlappingblocks,namelythe observationsequencesoftheHMM.Dueto itsc

第25卷

videsfaceimagesintoa

seriesofoverlappingblocks,namelythe

observationsequencesoftheHMM.Dueto

itscapability

todeal

withilluminationandposechanges,thefacerecognitionsystem

was

developed

on

basisof

this

approach.However。unlike

Neftfill’smethod。whichuses

discretecosin

transform(DCT)for

observation

vector

extraction,theproposed

algorithmuses

dis

crete

wavelet

transfonn(DWT)to

extract

observation

vectors.

Due

to

thesuperiorityofDWT

over

DCTinrepresentingimages,

thesystemhasachievedbetter

performance.

Accuratefacelocationbyskinmasking

Thealgorithmproposed

in[3]isinitiallyappliedforface

detection.However,asseen

fromFig1,one

cRn

observethatthe

faceimagedetectedisnot

accurate.Bothhairandbackgroundin-

formation

are

inchdedintlle

detectedfaceimage.andinsome

cases,shouldersare

alsoincluded.Askin

maskingmodule

is

necessarytorefinetheface

area

so

thatthebackground

andhair

&Yeas

can

beremoved.

(a)Originalimage(b)Detected

face

Fig.1

Originalinputimageanddetectedface

image图1原始图像和检测到的人脸

1.1

ColoranalysisItiswidely

acceptedthatthecolorofhumanskinisdistinc-

fivefromthecolorofmanyothernaturalobjects.Analyzing

the

skincolorstatisticsitisobservedthatskincolors

ate

distributed

overa

small

area

inthechrominanceplaneandthe

major

differ-

ence

betweenskin

tones

isintensity[….To

use

theimportantfea-

tureofcolorforfacedetection,allimageisfirstconvertedintoh minancechannelsandchrominancechannelsinYCbCrcolor

spaces.

Letwg

2[CbFCrF]denoteavector

composedofthechromi—

nance

componentsc6andCrfor

pixel(f,J).The

class—condi—

tionalpdfof

wi

belonging

to

theskinclass菇ismodeledby

2D

Gaussian[121:

p(1‘,gI鬈)=(2订)一1

I∑I一1以e一÷‘”F1“7∑。1‘”#1力

(1)

wherethe

meanvector//,andthecovariancematrix三areesti—

matedfromthetrainingset.Thecontour

ofthepdfshownabove

definesanellipseinthecbcrdomain.whose

center

andprincipal

axis

aredeterminedby弘ande。争”g叫】7∑_1[”#-ttl,respectively.

Fig2showsthedistributionofskincolorsinCb,Crdomain.Afterbuildingtheskincolormodel,theoriginalcolorimage

can

be

easilyconverted

to

skinprobabilityimagePusingequation(2).

TheimagePindicatestheprobabilityofeachimagepixel

belong

to

万 

方数据theskinclass菇:

p(id)一P(既I算)

(2)

Fig3showstheinputcolorimageIandtheskinprobabilityimage

Prespectively.

Fig.2

DistributionofskincolorsinCbandCrdomain

图2皮肤颜色在Cb和。平面分布

(aJ

Faceimage

∞Skinprobability(c)Ma.sked

face

Fig.3

Skin

maskingusingskinprobability

图3皮肤概率图和人脸掩模

1.2

FacemaskingandheadorientationestimationOncelllefaceregionisextractedfromtllevideoframe.the

headorientation

canbeobtainedfromthesecondmomentsofthe

skinprobabilityimagen¨.Assume

face

imagel(x,,,)with

size

W×Hisdetectedbythefacedetectionmodule,and户(茗,,, …… 此处隐藏:1498字,全部文档内容请下载后查看。喜欢就下载吧 ……

基于皮肤模板和改进HMM的自动人脸识别系统(2).doc 将本文的Word文档下载到电脑,方便复制、编辑、收藏和打印
本文链接:https://www.jiaowen.net/wenku/268230.html(转载请注明文章来源)
Copyright © 2020-2025 教文网 版权所有
声明 :本网站尊重并保护知识产权,根据《信息网络传播权保护条例》,如果我们转载的作品侵犯了您的权利,请在一个月内通知我们,我们会及时删除。
客服QQ:78024566 邮箱:78024566@qq.com
苏ICP备19068818号-2
Top
× 游客快捷下载通道(下载后可以自由复制和排版)
VIP包月下载
特价:29 元/月 原价:99元
低至 0.3 元/份 每月下载150
全站内容免费自由复制
VIP包月下载
特价:29 元/月 原价:99元
低至 0.3 元/份 每月下载150
全站内容免费自由复制
注:下载文档有可能出现无法下载或内容有问题,请联系客服协助您处理。
× 常见问题(客服时间:周一到周五 9:30-18:00)