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(3)

来源:网络收集 时间:2026-08-04
导读: theellipse are indicatedby across. Fig.4 Detect facewithdifferentorientations 图4不同方向的人脸 2 ImprovedHMMfor

theellipse

are

indicatedby

across.

Fig.4

Detect

facewithdifferentorientations

图4不同方向的人脸

ImprovedHMMforfacerecognition

2.1

HMMbasedfacerecognition

TheHMMis

finitesetofstates.eachofwhichisassociat-

edwith

multidimensionalprobabilitydistribution[131.Transi.

tionsamongthe

states

are

govemedby

aset

ofprobabilities

called

transitionprobabilities.Ina

particular

state

anoutcome

or

ohser-

vation

can

begeneratedaccording

to

the

associatedprobability

distribution.InimplementationofHMMbasedfacerecognition,a

faceimageis

dividedinto

seriesofoverlappingimageblocks.

%e

observationsequencefortheHMM

can

be

generatedby

con-

catenating

the

observation

vectors

extractedfrom

each

image

block.Nefian

et

al[10]apply2D

DCT

on

each

image

blockand

onlythelowfrequencycoefficientsarie

extractedto

produceohser-

vationvectors.7nleobservationsequenceextractedfrom

atest

im-

ageisinput

toallofthetrainedHMMs

associatedwitheachper-

son

andtheconditionalprobabilitygivenbyeachHMMisalso

calculated.ne

identityoftheinputfaceis

determinedbythe

HMMwhich

gives

thehighest

probability.SinceDWT

iswidely

believedtobeadvantageous

over

DCTinrepresentingfeatures,

DWTwillbeusedforobservationvectorextractioninthe

next

section.2.2

ImprovedHMMusingDWTWaveletahavemany

advantages

over

othermathematicaltransformssuchas

theFourier

transfo咖or

DCT.Inusing

wave-

lets

to

produceobservationvectors,thealgorithmworkbyscan

万 

方数据ningthefaceimagefromleft

to

right

andtoptobottomusing

Lsized

blockandperformingwaveletmulti resolutionanalysis

at

each

imageblock.Theimageblockisdecomposed

to

certain

levelandthecoefficients

at

thelowestlevel

or

theenergyofthe

subbandsare

extracted

to

formobservationvectorsfortheHMM.

2D

J_levelwaveletdecompositiononall

image,[P,L]represents

theimageby3J+1subbands

[口J,{司,孝,霉}.『=1,…,.,],whereD,is

lowresolution

approximationoftheoriginalimage,and钟arc

detail

images

con-

miningthedetailsofthe

image

at

different

scales(巧)and

often—

rations(&).Waveletcoefficients

in谚,毋and霉correspond

to

vertical

hi.ghfrequencies(horizontaledges),horizontalhighfre-

quencies(verticaledges),andhighfrequencies

inbeth

direc-

tions,respectively.Fig5showsthe2levelwaveletdecompositionofanimagerepresenting

an

eye.Forthis2levelsdecomposition

in

total7

subbands{a2,d:,《,前,以,《,d;}are

generated.

Notethattheimagehasbeenrescaledforvisualinspection.Inthispaper,theobservation

vector

is

producedbyconcatenating

each

lOW

ofaJ.Oncetheobservation

vectors

are

extracted,theobservationsequences

are

generatedandgiven

to

theHMMfor

processing.

(a)An

eye

imageblock蚴Wavelet

decomposition

Fig.5 …… 此处隐藏:1581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(3).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)