Travel & Places

A novel f-ART Model for Visual Lipreading based on Active Lip Model

Description
Visual Lipreading, a perception of speech is purely based on observing the lip movements and under noisy conditions. For listeners with hearing impairment visual speech information play an important role. The researches in Lipreading are targeted
Published
of 3
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Transcript
  A novel f-ART Model for Visual Lipreading based on Active Lip Model B Sujatha † *  and T Santhanam †† †  sujatha_basdas@yahoo.co.in  ††    santhanam_dgvc@yahoo.com † Research Scholar, MTWU, Kodaikanal, Meenakshi College for Women, Chennai, India . ††    PG and Research De! of Com"!er Science, D#G# $aishna% College, Chennai, India  Abstract: Visual Lipreading a perception of speech is purel! based on observing the lip movements and under nois! conditions" #or listeners $ith hearing impairment visual speech information pla! an important role" The researches in Lipreading are targeted onl! the recogni%er model and there is no model to refer to the classification" &n this paper a ne$ feature e'traction techni(ue Active Lip Model )ALM* is used for Visual onl! isolated digit fu%%! ART )f-ART* classifier is presented" #irstl! the video images are subdivided into frames and for selected frames the four lip features are e'tracted using the ALM" The e'tracted features are given as a input to f-ART for recogni%ing +one, to +four, digits" The same method is applied for TL&.S /atabase" 0'perimental results sho$ that the proposed method gives superior recognition rate" 1e!$ords2 #u%%! ART  Tulips /atabase  Active Lip Model  mugshot /atabase" & 3TR4/5T&43 &i'reading aroaches can (e classified in!o image'(ased and model'(ased s)s!ems# Image' (ased s)s!ems "se gre) le%el informa!ion from an image region con!aining !he lis ei!her direc!l) or af!er some rocessing as seech fea!"res# In general, fea!"re e!rac!ion me!hods con!ain !hree ca!egories+ !he firs! is shae'(ased fea!"re e!rac!ion me!hod, hich ass"mes !ha! mos! seech informa!ion is con!ained in !he con!o"rs of !he seaker-s lis, or more generall) in !he face con!o"rs, e#g# li, .a and cheek /01# Geome!ric'!)e fea!"re as idel) "sed in earl) !ime, s"ch as mo"!h id!h, heigh!, and area , 2o"rier and image momen! descri!ors of !he li con!o"rs, la!er on , e!rac!ing !he arame!ers of li shae model as aid more a!!en!ion, s"ch as 3c!i%e Shae Model 43SM56 !he second is aearance'(ased me!hod# In con!ras!, aearance'(ased me!hod ass"mes !ha! all iels in R7I are informa!i%e seech fea!"res# To red"ce high dimensionali!) of fea!"re %ec!or, some linear !ransforma!ion are in!rod"ced, s"ch as Princial Comonen! 3nal)sis 4PC35, Discre!e Cosine Transforma!ion 4DCT5, &D3, Discre!e Wa%ele! Transforma!ion 4DWT5, e!c6 !he !hird is h)(rid me!hod, hich conca!ena!e !he former !o fea!"res in!o a .oin! shae and aearance fea!"re %ec!or, e#g# ac!i%e aearance model 433M5 is learned on s"ch %ec!ors# /81# 3SM is "sed in li reading ("! !he fea!"re e!rac!ion me!hod s"or!ed in /91 is differen! from !he !echni:"e roosed in o"r ork# In !his aer, Sec!ion 0 3&M is disc"ssed and 2ea!"re e!rac!ion !echni:"e is elained in !he ne! sec!ion# 2'3RT s!r"c!"re is elained in sec!ion 9# The las! sec!ion disc"ss a(o"! !he eerimen!al res"l!s and concl"sion of !he ork# &"A 5T&V0  L &.  M 4/0L In /;1 3SM is "sed in li reading ("! !he fea!"re e!rac!ion me!hod s"or!ed in !his s!"d) is differen! from o!her roosed !echni:"es# 3SM is idel) "sed in facial recogni!ion and mos! of !he researchers "se !his model onl) !o e!rac! !he li region and (ranch !oards o!her areas, no! foc"sing on !he li reading# Geome!rics of o"!er and inner lis are fo"nd from 3SM i!h o!imal' oin! li model# Inner li geome!r) is a s"('se! !o  ro%ide addi!ional informa!ion# Us"al !raining and !es!ing of classifiers like <MM and o!her ne!ork models (ecome !ime cons"ming and more erra!ic hen da!a(ase si=e increases i!h more ords# These classifiers are loaded hea%il) in se:"ence comarison of indi%id"al frame fea!"res (elonging !o a ord, hile !raining i!# 7nce !he ("rden is shared () 3<P, simle f"==) classifier makes decision a(o"! a ar!ic"lar ord, (elonging !o !he !es! frame'se:"ence# &&"# 0ATR0  0 6TRA5T&43  T 0573&80 2or images, lo le%el %is"al fea!"res are color, !e!"re, shae and sa!ial locali=a!ion# <oe%er, among !hese fea!"res, shae is !he mos! imor!an! (eca"se i! reresen!s significan! regions or rele%an! o(.ec!s in images# Ideall), shae segmen!a!ion o"ld (e a"!oma!ic and efficien!, ("! i! is ei!her imossi(le or diffic"l! i!h he!erogeneo"s images# />1  #igure 92 #lo$ diagram for feature e'traction Shae is a %er) imor!an! fea!"re !o h"man  erce!ion# <"man (eings !end !o ercei%e scenes as (eing comosed of indi%id"al o(.ec!s, hich can (e (es! iden!ified () !heir shaes# ?esides, as far as :"er) is concerned, shae is simle for "ser !o descri(e, ei!her () gi%ing eamle or () ske!ching# 7nce images or scenes are (roken don in!o indi%id"al o(.ec!s, !he) can  (e eloi!ed !o facili!a!e o(.ec! recogni!ion /@1# 3SM is a shae'cons!rained i!era!i%e fi!!ing algori!hm, and i! can onl) deform !o fi! !he !arge! o(.ec! in a)s consis!en! i!h !he !raining se!# The shae cons!rain! comes from !he "se of a s!a!is!ical shae model !ha! is o(!ained from !he s!a!is!ics of hand la(eled !raining da!a /A1# 3earance Model con!ains a s!a!is!ical model of !he shae and gra)'le%el aearance of !he o(.ec! of in!eres! hich can generali=e !o almos! an) %alid eamle# 3l) 33M !o sim"l!aneo"sl) e!rac! !he e!ernal and in!ernal li con!o"rs# Using !he fig#0 flo diagram !he fea!"res are e!rac!ed for o"r s)s!em# &&&"# ::;  ART 3 0T<4R1  2"==) 3da!i%e Resonance Theor) is a fas!, in!erac!i%e, incremen!al, s"er%ised learning s)s!em for analog in"!s# f'3RT "ses simle f"==) learning r"les like ma and min for ac!i%a!ion and selec!ion of ne"rons# The f"==) r"les minimi=e !he com"!a!ion re:"ired for learning and i! learns e%er) a!!ern i!h %er) fe i!era!ions# This ne!ork s!ar!s i!h no connec!ion eigh!s, gros in si=e !o s"i! !he ro(lem, "ses simle learning e:"a!ions, and has onl) "ser'selec!a(le arame!er knon as %igilance arame!er# The fas! learning caa(ili!) is made ossi(le () a series of rocessing s!ages+ in"! crea!ion, in"! forma!!ing for o"!"! node ac!i%a!ion, a!!ern ma!ching, and ca!egorical maing/B1#   #igure =2 Bloc> diagram of f-ART Model Simle f'3RT con!ains !o la)ers+ an in"! and ano"!"! la)er# 3 (lock diagram of !he ne!ork highligh!ing !he main archi!ec!"re is shon in fig#8# In"! in!o !he ne!ork m"s! (e normali=ed !o a %al"e from  !o0#<ence a s"i!a(le normali=a!ion %al"e m"s! (e chosen so !ha! no in"! ill fall o"!side !he %alid range# 3 comlimen! coder normali=es !he in"! and also ro%ides !he f"==) comlimen! for each %al"e# This eanded in"! 4I5 is !hen assed !o !he in"! la)er# Weigh!s 45 from each o"!"! node samle !he in"! la)er, making !he eigh!s !o'don# Training (egins i!h ."s! one hidden node hose eigh!s are se! e:"al !o !he firs! record and redic!ion is se! e:"al !o !he class of !he firs! record# Similarl), hene%er a ne class is enco"n!ered a ne node is crea!ed# The node, hose eigh!s (es! ma!ch !he c"rren! in"!, s"lies !he redic!ion ro%ided, !he degree of !he ma!ch eceeds !he %igilance !hreshold %al"e# If !his   redic!ion is correc!, !he eigh!s of !his inning node are ad."s!ed !oards !his in"!# If !he  redic!ion is rong or %igilance !hreshold is no! achie%ed, a ne node is crea!ed i!h eigh!s and  redic!ion e:"al !o !his record# &V"0 6.0R&M03TAL  R  0SLTS   The e!rac!ed fea!"res for "ser crea!ed da!a(ase and !he T"lis da!a(ase are lis!ed in Ta(le 0 and Ta(le 8 resec!i%el)# Table 92 0'tracted #eatures for Mugshot /atabase Word7&WR$&DR&7WRI&WR one0#B>B80#E8>#BBA99#88>90#EB;0#;B@#BBEA9#9@0!o0#A>80#0EB#E;88#@8@80#;0E0#>#99A9E#@0880#;8A0#>#9880>#E;@0#E9;;0#BEB#90>B@#89EE!hree0#>8>A#B9E@#99;B9#;;EB0#;;90#8B;#>90>>#0AA;0#B>B80#0;0#@0@#;@0#@#BE98#98>B>#BE@Bfo"r0#>@8;0#8@#8EAE@#BE900#>;@#BA;8#8B88A#09>0#@>A9#BB08#8EA@@#;@>00#9EE80#89E#8AE@#@;0 Table =2 0'tracted #eatures for TL&.S /atabase Word7&WR$&DR&7WRI&WR one0#>>AE#9E8>#E;#;@EA0#AAA;#;@EA#;@EA8#EE0#9@E@#@0>9#;0@;9#0;EA0#A>E0#B8EA#;B;E9#>0;!o0#>9E#@8B>#;EEB9#E;0#>>B#>E>E#;8>E9#EB@0#AAA;#;0>9#A8A0>#>@0#E@>80#>#A9009#>890!hree0#;>@00#>#;A;B9#80;>0#@08#@909#;@;98#BA90#>9@#@>@#EE;9#BB0#8@B9#9E>A#;;;>#@8A9fo"r0#9A@0#BAEE#;AE>9#;0>0#AEA9#EEEB#;B8A9#@E990#;@@00#>0;#;;@>9#9A0#AE#AAA;#;@8;9#9A V"5 435LS&43 3 no%el fea!"re e!rac!ion !echni:"e 3c!i%e &i Model 43&M5 is "sed for $is"al onl) isola!ed digi! f"==) 3RT 4f'3RT5 classifier is resen!ed in !his aer# 2irs!l) !he %ideo images are s"(di%ided in!o frames and for selec!ed frames !he fo"r li fea!"res are e!rac!ed "sing !he 3&M# The e!rac!ed fea!"res are gi%en as a in"! !o f'3RT for recogni=ing Fone- !o Ffo"r- digi!s# The same !echni:"e is alied for !he T"lis Da!a(ase and !he classifica!ion ra!e is s"erior for f'3RT !han !he 3<P (ased 2"==) classifier in / ;1# R  0#0R0350S /01 "ergen &"e!!in0,8, Heil 3# Thacker0, S!e%e W# ?ee!0, SPJJC<RJ3DIHG USIHG S<3PJ 3HD IHTJHSIT IH27RM3TI7HL /81 "n <e , 2ea!"re J!rac!ion in SeechreadingL , 7URH3& 72 S72TW3RJ, $7&# @, H7# ;, U& 80#/91 www.medwelljournals.com/fulltext/?doi=ijscomp.2010.13.18 />1 T"Venu ?opal V"1ama>shi .rasad  , 3 Ho%el 3roach !o Shae ?ased Image Re!rie%al In!egra!ing 3da!ed 2o"rier Descri!ors and 2reeman CodeL, ICSHS In!erna!ional o"rnal of Com"!er Science and He!ork Sec"ri!), $7&#E Ho#A, "ne 8E#/@1 Takeshi Sai!oh, Ka="!oshi Morishi!a and R)os"ke Konishi, 3nal)sis of Jfficien! &i Reading Me!hod for $ario"s &ang"agesL#/A1 ?# S".a!ha and T# San!hanam , LClassical 2lei(le &i Model ?ased Rela!i%e Weigh! 2inder for ?e!!er &i Reading U!ili=ing M"l!i 3sec! &i Geome!r)L, o"rnal of Com"!er Science A 405+ 09;'0>0, 80 ISSH 0@>B'9A9A  80 Science P"(lica!ions#/;1 Tomoe Jn!ani and <ideo Tanaka , Modified In!er%al Glo(al Weigh!s in 3<PL# /E1 J%angelos Trian!ah)llo" , S!"ar! <# Mann, Using !he 3nal)!ic <ierarch) Process for decision making in engineering alica!ions+ some challengesL, In!erna!ional o"rnal of Ind"s!rial Jngineering+ 3lica!ions and Prac!ice, $ol# 8, Ho# 0, # 9@'>>, 0BB@#/B1 $enka!esh P and S"resh M &, LClassifica!ion of Renal 2ail"re "sing Simlified 2"==) 3RT MaL, In!erna!ional o"rnal of Com"!er Science and He!ork Sec"ri!) , $ol# B  Ho#00,Ho% 8B#  
Search
Tags
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks