<html xmlns="http://www.w3.org/1999/xhtml">
<head>
<meta charset="utf-8">
<meta name="generator" content="pdf2htmlEX">
<meta http-equiv="X-UA-Compatible" content="IE=edge,chrome=1">
<link rel="stylesheet" href="https://static.pudn.com/base/css/base.min.css">
<link rel="stylesheet" href="https://static.pudn.com/base/css/fancy.min.css">
<link rel="stylesheet" href="https://static.pudn.com/prod/directory_preview_static/6247201d62b5053d3c2a43f8/raw.css">
<script src="https://static.pudn.com/base/js/compatibility.min.js"></script>
<script src="https://static.pudn.com/base/js/pdf2htmlEX.min.js"></script>
<script>
try{
pdf2htmlEX.defaultViewer = new pdf2htmlEX.Viewer({});
}catch(e){}
</script>
<title></title>
</head>
<body>
<div id="sidebar" style="display: none">
<div id="outline">
</div>
</div>
<div id="pf1" class="pf w0 h0" data-page-no="1"><div class="pc pc1 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://static.pudn.com/prod/directory_preview_static/6247201d62b5053d3c2a43f8/bg1.jpg"><div class="t m0 x1 h2 y1 ff1 fs0 fc0 sc0 ls0 ws0">A Multi-Algorithmic Face Recognition System </div><div class="t m0 x2 h3 y2 ff1 fs1 fc0 sc0 ls1 ws1">#Soumitra Kar, Swati Hiremath, Dilip G. Joshi, Vinod.K.Chadda and<span class="_ _0"></span> </div><div class="t m0 x3 h4 y3 ff1 fs2 fc0 sc0 ls2 ws2">1</div><div class="t m0 x4 h3 y2 ff1 fs1 fc0 sc0 ls3 ws3">Apurva Bajpai </div><div class="t m0 x5 h5 y4 ff2 fs3 fc0 sc0 ls4 ws4">EISD, BARC, Mumbai-400 085,<span class="_ _0"></span> India </div><div class="t m0 x6 h5 y5 ff2 fs3 fc0 sc0 ls5 ws2">skar@barc.go<span class="_ _0"></span>v.in </div><div class="t m0 x7 h6 y6 ff3 fs3 fc0 sc0 ls6 ws2">Abstract </div><div class="t m0 x8 h5 y7 ff2 fs3 fc0 sc0 ls7 ws5">The impo<span class="_ _1"></span>rtance of<span class="_ _1"></span> utilisi<span class="_ _1"></span>ng biomet<span class="_ _1"></span>rics t<span class="_ _1"></span>o establish<span class="_ _1"></span> personal </div><div class="t m0 x8 h5 y8 ff2 fs3 fc0 sc0 ls8 ws6">authentic<span class="_ _0"></span>ity and to de<span class="_ _0"></span>tect impostors<span class="_ _0"></span> is growing in <span class="_ _0"></span>the present </div><div class="t m0 x8 h5 y9 ff2 fs3 fc0 sc0 ls9 ws7">scenario of global secur<span class="_ _0"></span>ity concern. Developm<span class="_ _0"></span>ent of a </div><div class="t m0 x8 h5 ya ff2 fs3 fc0 sc0 lsa ws8">biometric system <span class="lsb ws9">for personal identification, which fulfills the </span></div><div class="t m0 x8 h5 yb ff2 fs3 fc0 sc0 lsc wsa">requirements for access control of secured areas and other </div><div class="t m0 x8 h5 yc ff2 fs3 fc0 sc0 ls5 wsb">applicati<span class="_ _0"></span>ons like identity val<span class="_ _0"></span>idation for social we<span class="_ _0"></span>lfare, crime </div><div class="t m0 x8 h5 yd ff2 fs3 fc0 sc0 lsd wsc">detection, ATM access, comp<span class="_ _1"></span>uter securit<span class="_ _1"></span>y, etc. is felt to be the </div><div class="t m0 x8 h5 ye ff2 fs3 fc0 sc0 lse wsd">need of the day<span class="_ _0"></span>. Face recogn<span class="_ _0"></span>ition has bee<span class="_ _0"></span>n evolving as<span class="_ _0"></span> a </div><div class="t m0 x8 h5 yf ff2 fs3 fc0 sc0 lsf wse">convenie<span class="_ _0"></span>nt biometric mode for human authe<span class="_ _0"></span>ntication<span class="_ _0"></span> for </div><div class="t m0 x8 h5 y10 ff2 fs3 fc0 sc0 ls8 wsf">more tha<span class="_ _0"></span>n last two dec<span class="_ _0"></span>ades. Several ven<span class="_ _0"></span>dors aroun<span class="_ _0"></span>d the </div><div class="t m0 x8 h5 y11 ff2 fs3 fc0 sc0 ls10 ws10">world clai<span class="_ _1"></span>m the s<span class="_ _1"></span>uccessf<span class="_ _1"></span>ul wor<span class="_ _1"></span>king of<span class="_ _1"></span> their fac<span class="_ _1"></span>e recogni<span class="_ _1"></span>tion </div><div class="t m0 x8 h5 y12 ff2 fs3 fc0 sc0 ls11 ws11">systems. However, the Face Recognition Vendor Test </div><div class="t m0 x8 h5 y13 ff2 fs3 fc0 sc0 ls8 ws12">(FRVT) condu<span class="_ _0"></span>cted by the Nationa<span class="_ _0"></span>l Institute of St<span class="_ _0"></span>andards and<span class="_ _0"></span> </div><div class="t m0 x8 h5 y14 ff2 fs3 fc0 sc0 ls12 ws13">Technology (<span class="_ _0"></span>NIST), USA, indic<span class="_ _0"></span>ates that the commercial face </div><div class="t m0 x8 h5 y15 ff2 fs3 fc0 sc0 ls13 ws14">recognition systems do not <span class="ls14 ws15">perform up to t<span class="_ _0"></span>he mark under the </span></div><div class="t m0 x8 h5 y16 ff2 fs3 fc0 sc0 ls15 ws16">variations ub<span class="_ _0"></span>iquitously pre<span class="_ _0"></span>sent in a real-<span class="_ _0"></span>life situation.<span class="_ _0"></span> </div><div class="t m0 x8 h5 y17 ff2 fs3 fc0 sc0 ls16 ws17">Availabilit<span class="_ _1"></span>y of a largely accepted robust face reco<span class="_ _1"></span>gnition </div><div class="t m0 x8 h5 y18 ff2 fs3 fc0 sc0 ls17 ws18">system has proved elusi<span class="_ _1"></span>ve so far. Keep<span class="_ _1"></span>ing in view th<span class="_ _1"></span>e </div><div class="t m0 x8 h5 y19 ff2 fs3 fc0 sc0 ls18 ws19">importance of indige<span class="_ _0"></span>nous developmen<span class="_ _0"></span>t of biometric systems </div><div class="t m0 x8 h5 y1a ff2 fs3 fc0 sc0 ls19 ws1a">to cater to the requirements at<span class="_ _1"></span> BARC and elsewhere in th<span class="_ _1"></span>e </div><div class="t m0 x8 h5 y1b ff2 fs3 fc0 sc0 ls1a ws1b">country, the work was <span class="_ _0"></span>started <span class="_ _0"></span>on the devel<span class="_ _0"></span>opment of a face-</div><div class="t m0 x8 h5 y1c ff2 fs3 fc0 sc0 ls1b ws1c">based bi<span class="_ _0"></span>ometric authentic<span class="_ _0"></span>ation system.<span class="_ _0"></span> In this paper, we<span class="_ _0"></span> </div><div class="t m0 x8 h5 y1d ff2 fs3 fc0 sc0 ls12 ws1d">discuss our efforts in developing a face recognition syst<span class="_ _0"></span>em </div><div class="t m0 x8 h5 y1e ff2 fs3 fc0 sc0 ls1c ws1e">that functions<span class="_ _0"></span> successfully <span class="_ _0"></span>under a reasonably c<span class="_ _0"></span>onstrained </div><div class="t m0 x8 h5 y1f ff2 fs3 fc0 sc0 ls1d ws1f">set-up for facial image acqu<span class="lsc ws20">isition. The <span class="ls1e ws21">prototype system </span></span></div><div class="t m0 x8 h5 y20 ff2 fs3 fc0 sc0 ls1f ws22">built in our l<span class="_ _1"></span>ab finds facia<span class="_ _1"></span>l match by util<span class="_ _1"></span>izing multi<span class="_ _1"></span>-</div><div class="t m0 x8 h5 y21 ff2 fs3 fc0 sc0 ls1b ws23">algorithmic<span class="_ _0"></span> multi-biome<span class="_ _0"></span>tric technique, co<span class="_ _0"></span>mbining gray level<span class="_ _0"></span> </div><div class="t m0 x8 h5 y22 ff2 fs3 fc0 sc0 lsf ws24">statistical co<span class="_ _0"></span>rrelation me<span class="_ _0"></span>thod with Princ<span class="_ _0"></span>ipal Component<span class="_ _0"></span> </div><div class="t m0 x8 h5 y23 ff2 fs3 fc0 sc0 ls20 ws25">Analysis (PC<span class="_ _0"></span>A) or Discrete Cosine <span class="_ _0"></span>Transform (DCT<span class="_ _0"></span>) <span class="_ _2"></span> </div><div class="t m0 x8 h5 y24 ff2 fs3 fc0 sc0 ls4 ws26">techniques in order to boost<span class="_ _0"></span> our system performance. Afte<span class="_ _0"></span>r </div><div class="t m0 x8 h5 y25 ff2 fs3 fc0 sc0 lsf ws27">automatic de<span class="_ _0"></span>tection of the<span class="_ _0"></span> face in the image an<span class="_ _0"></span>d its gross </div><div class="t m0 x8 h5 y26 ff2 fs3 fc0 sc0 ls1d ws28">scale correction, its PCA and DCT signature<span class="_ _0"></span>s are extracte<span class="_ _0"></span>d. </div><div class="t m0 x8 h5 y27 ff2 fs3 fc0 sc0 ls1c ws29">Based on a comparison of the ex<span class="_ _0"></span>tracted signature with the set </div><div class="t m0 x8 h5 y28 ff2 fs3 fc0 sc0 lsb ws2a">of references, the se<span class="_ _0"></span>t of top five hits are selected.<span class="_ _0"></span> Exact scale </div><div class="t m0 x8 h5 y29 ff2 fs3 fc0 sc0 ls21 ws2b">of the face<span class="_ _0"></span> is ascertaine<span class="_ _0"></span>d w.r.t. each <span class="_ _0"></span>of these hits by f<span class="_ _0"></span>irst </div><div class="t m0 x8 h5 y2a ff2 fs3 fc0 sc0 lsf ws2c">locating t<span class="_ _0"></span>he eyes employi<span class="_ _0"></span>ng template ma<span class="_ _0"></span>tching tech<span class="_ _0"></span>nique and </div><div class="t m0 x8 h5 y2b ff2 fs3 fc0 sc0 ls13 ws2d">then finding the inter-<span class="_ _0"></span>ocular distance. After interpolating the </div><div class="t m0 x8 h5 y2c ff2 fs3 fc0 sc0 ls13 ws2e">face to the exa<span class="_ _0"></span>ct scale, matching scores <span class="_ _0"></span>are computed base<span class="_ _0"></span>d </div><div class="t m0 x8 h5 y2d ff2 fs3 fc0 sc0 ls8 ws2f">on gray level corre<span class="_ _0"></span>lation of a number <span class="_ _0"></span>of features on the<span class="_ _0"></span> face. </div><div class="t m0 x8 h5 y2e ff2 fs3 fc0 sc0 ls22 ws30">Final iden<span class="_ _0"></span>tification dec<span class="_ _0"></span>ision is taken amongs<span class="_ _0"></span>t this set of five </div><div class="t m0 x8 h5 y2f ff2 fs3 fc0 sc0 ls1f ws31">faces, dependin<span class="_ _1"></span>g on the highest score. We have test<span class="_ _1"></span>ed the </div><div class="t m0 x8 h5 y30 ff2 fs3 fc0 sc0 ls23 ws32">techniqu<span class="_ _1"></span>e on a set of 109<span class="_ _1"></span> images bel<span class="_ _1"></span>onging to<span class="_ _1"></span> 43 subject<span class="_ _1"></span>s, <span class="_ _3"></span> </div><div class="t m0 x8 h5 y31 ff2 fs3 fc0 sc0 ls24 ws33">both male and f<span class="_ _1"></span>emale. The result on th<span class="_ _1"></span>is image-set indicates<span class="_ _1"></span> </div><div class="t m0 x8 h5 y32 ff2 fs3 fc0 sc0 ls12 ws34">89% success rate of <span class="_ _0"></span>our technique.</div><div class="t m0 x8 h7 y33 ff1 fs4 fc0 sc0 ls2 ws2">1</div><div class="t m0 x9 h8 y34 ff1 fs3 fc0 sc0 ls25 ws35">Deputed to BARC by ECIL for this project. </div><div class="t m0 xa h6 y35 ff3 fs3 fc0 sc0 ls4 ws36">Index Terms </div><div class="t m0 xb h8 y36 ff1 fs3 fc0 sc0 ls26 ws37"> <span class="_ _4"></span>Face recognition, I<span class="_ _0"></span>dentifi<span class="ls2 ws38">cation, Multi-algorithmic, Multi-</span></div><div class="t m0 xb h8 y37 ff1 fs3 fc0 sc0 ls1a ws39">biometric, Discrete C<span class="_ _0"></span>osine Transform, Principa<span class="_ _0"></span>l Component </div><div class="t m0 xb h8 y38 ff1 fs3 fc0 sc0 ls1b ws3a">Analysis, Correlati<span class="_ _0"></span>on. </div><div class="t m0 xc h6 y39 ff3 fs3 fc0 sc0 ls27 ws2">1.<span class="_ _5"> </span>INTRODUCTION </div><div class="t m0 xb h8 y3a ff1 fs3 fc0 sc0 lse ws3b"> <span class="_ _4"></span>Biometrics m<span class="_ _0"></span>akes autom<span class="_ _0"></span>ated use of<span class="_ _0"></span> the unique per<span class="_ _0"></span>sonal </div><div class="t m0 xb h8 y3b ff1 fs3 fc0 sc0 ls20 ws3c">features to est<span class="_ _0"></span>ablish the iden<span class="_ _0"></span>tity of a person. It is<span class="_ _0"></span> a tool for </div><div class="t m0 xb h8 y3c ff1 fs3 fc0 sc0 ls23 ws3d">positiv<span class="_ _1"></span>e identifi<span class="_ _1"></span>cation of<span class="_ _1"></span> a human subjec<span class="_ _1"></span>t as bio<span class="_ _1"></span>metric </div><div class="t m0 xb h8 y3d ff1 fs3 fc0 sc0 ls4 ws3e">signatures cannot <span class="_ _0"></span>be stolen, for<span class="_ _0"></span>gotten, lost or comm<span class="_ _0"></span>unicated </div><div class="t m0 xb h8 y3e ff1 fs3 fc0 sc0 ls22 ws3f">to another, as is p<span class="_ _0"></span>ossible in the case<span class="_ _0"></span> of authent<span class="_ _0"></span>ication </div><div class="t m0 xb h8 y3f ff1 fs3 fc0 sc0 ls1a ws39">employing cards, keys or pa<span class="_ _0"></span>sswords, so common in day-to-</div><div class="t m0 xb h8 y40 ff1 fs3 fc0 sc0 ls28 ws40">day use. In the pre<span class="_ _0"></span>sent scenario of increase<span class="_ _0"></span>d security concern, </div><div class="t m0 xb h8 y41 ff1 fs3 fc0 sc0 ls29 ws41">the necessity and rele<span class="_ _0"></span>vance of making use of biome<span class="_ _0"></span>trics to </div><div class="t m0 xb h8 y42 ff1 fs3 fc0 sc0 ls2a ws42">establish persona<span class="_ _0"></span>l identity and to detect impostors a<span class="_ _0"></span>re </div><div class="t m0 xb h8 y43 ff1 fs3 fc0 sc0 ls1c ws43">assuming significanc<span class="_ _0"></span>e. Biometric tec<span class="_ _0"></span>hniques are based on </div><div class="t m0 xb h8 y44 ff1 fs3 fc0 sc0 lsc ws44">either physiological char<span class="lsa ws45">acteristics (like finger print, i<span class="_ _1"></span>ris, etc.) </span></div><div class="t m0 xb h8 y45 ff1 fs3 fc0 sc0 ls29 ws46">or behavioral traits (li<span class="_ _0"></span>ke voice dynamics, gait, etc.). </div><div class="t m0 xd h8 y46 ff1 fs3 fc0 sc0 ls29 ws47">Depending upo<span class="_ _0"></span>n the suitability <span class="_ _0"></span>in a particular </div><div class="t m0 xb h8 y47 ff1 fs3 fc0 sc0 ls1c ws48">application, one ha<span class="_ _0"></span>s to choose a<span class="_ _0"></span> particular <span class="_ _0"></span>biometric [15]<span class="_ _0"></span> to </div><div class="t m0 xb h8 y48 ff1 fs3 fc0 sc0 ls5 ws49">be used as the<span class="_ _0"></span> basic signatur<span class="_ _0"></span>e for recogniti<span class="_ _0"></span>on. We selected </div><div class="t m0 xb h8 y49 ff1 fs3 fc0 sc0 ls15 ws4a">the face base<span class="_ _0"></span>d approach on the<span class="_ _0"></span> considerati<span class="_ _0"></span>ons that facia<span class="_ _0"></span>l </div><div class="t m0 xb h8 y4a ff1 fs3 fc0 sc0 ls2b ws4b">imaging, bein<span class="_ _0"></span>g non-intrusive<span class="_ _0"></span>, has easy client ac<span class="_ _0"></span>ceptance, </div><div class="t m0 xb h8 y4b ff1 fs3 fc0 sc0 ls2c ws4c">apart fro<span class="_ _1"></span>m the fact<span class="_ _1"></span> that fac<span class="_ _1"></span>e recogni<span class="_ _1"></span>tion is th<span class="_ _1"></span>e most natural<span class="_ _1"></span> </div><div class="t m0 xb h8 y4c ff1 fs3 fc0 sc0 ls1a ws4d">means of biometric identif<span class="_ _0"></span>ication for human beings. The </div><div class="t m0 xb h8 y4d ff1 fs3 fc0 sc0 ls2d ws4e">present circumstances around us demand increased<span class="_ _1"></span> level of </div><div class="t m0 xb h8 y4e ff1 fs3 fc0 sc0 ls5 ws4f">security and, theref<span class="_ _0"></span>ore, machine capa<span class="_ _0"></span>bility of persona<span class="_ _0"></span>l </div><div class="t m0 xb h8 y4f ff1 fs3 fc0 sc0 ls23 ws50">identifi<span class="_ _1"></span>cation from t<span class="_ _1"></span>he faci<span class="_ _1"></span>al image is c<span class="_ _1"></span>onsidered<span class="_ _1"></span> invaluabl<span class="_ _1"></span>e. </div><div class="t m0 xb h8 y50 ff1 fs3 fc0 sc0 ls28 ws51">Face-based biometric syste<span class="_ _0"></span>ms have the potentia<span class="_ _0"></span>l to fulfill the </div><div class="t m0 xb h8 y51 ff1 fs3 fc0 sc0 ls12 ws52">requirements for <span class="_ _0"></span>access control of secure<span class="_ _0"></span>d areas, surve<span class="_ _0"></span>illance, </div><div class="t m0 xb h8 y52 ff1 fs3 fc0 sc0 ls12 ws53">social welfare, law e<span class="_ _0"></span>nforcement, etc.<span class="_ _0"></span> Although there are a </div><div class="t m0 xb h8 y53 ff1 fs3 fc0 sc0 ls1c ws1e">few commercial face<span class="_ _0"></span> recognition s<span class="_ _0"></span>ystems availa<span class="_ _0"></span>ble around </div><div class="t m0 xb h8 y54 ff1 fs3 fc0 sc0 ls14 ws54">the globe, their perform<span class="_ _0"></span>ance is not up to the mark under the </div><div class="t m0 xb h8 y55 ff1 fs3 fc0 sc0 lsb ws55">practical variabilities [5]. This encouraged us to ta<span class="_ _0"></span>ke up face </div><div class="t m0 xb h8 y56 ff1 fs3 fc0 sc0 ls1c ws56">recognition system deve<span class="_ _0"></span>lopment. </div><div class="t m0 xd h8 y57 ff1 fs3 fc0 sc0 ls1c ws57">For enhanci<span class="_ _0"></span>ng the accuracy of <span class="_ _0"></span>biometric system<span class="_ _0"></span>s, </div><div class="t m0 xb h8 y58 ff1 fs3 fc0 sc0 ls14 ws58">multi-biometric technique wit<span class="_ _0"></span>h the ability to utilise either </div><div class="t m0 xb h8 y59 ff1 fs3 fc0 sc0 ls2b ws59">multiple biom<span class="_ _0"></span>etric modalit<span class="_ _0"></span>ies or multiple i<span class="_ _0"></span>nstances wi<span class="_ _0"></span>thin a </div><div class="t m0 xb h8 y5a ff1 fs3 fc0 sc0 ls14 ws5a">modality and/or<span class="_ _0"></span> multiple algorithms, prior to making a </div><div class="t m0 xb h8 y5b ff1 fs3 fc0 sc0 ls19 ws5b">specific verificati<span class="ls11 ws5c">on/identification d<span class="ls2e ws5d">ecisi<span class="_ _1"></span>on, has been </span></span></div><div class="t m0 xb h8 y5c ff1 fs3 fc0 sc0 ls9 ws5e">suggested [16]. We have de<span class="_ _0"></span>veloped a prototype fa<span class="_ _0"></span>ce </div><div class="t m0 xb h8 y5d ff1 fs3 fc0 sc0 ls1c ws5f">recognition system<span class="_ _0"></span> utilising mul<span class="_ _0"></span>ti-algorithmic m<span class="_ _0"></span>ulti-</div><div class="t m0 xb h8 y5e ff1 fs3 fc0 sc0 ls2f ws60">biomet<span class="_ _1"></span>ric te<span class="_ _1"></span>chnique t<span class="_ _1"></span>o boost th<span class="_ _1"></span>e syst<span class="_ _1"></span>em performanc<span class="_ _1"></span>e. It i<span class="_ _1"></span>s </div><div class="t m0 xb h8 y5f ff1 fs3 fc0 sc0 ls1c ws61">based on combining Correlation with Principal Com<span class="_ _0"></span>ponent </div><div class="t m0 xb h8 y60 ff1 fs3 fc0 sc0 ls1a ws62">Analysis (PCA) or Discr<span class="_ _0"></span>ete Cosine Transform (DCT) </div><div class="t m0 xb h8 y61 ff1 fs3 fc0 sc0 ls1a ws63">technique for the purpose of deciding a facial match. </div><div class="t m0 xe h9 y62 ff4 fs5 fc0 sc0 ls2 ws2">1-4244-0716-8/06/$20.00 ©2006 IEEE.<span class="_ _6"> </span>321</div><div class="t m0 xf ha y63 ff5 fs6 fc0 sc0 ls2 ws2"><span class="fc1 sc0">Authorized licensed use limited to: Harbin University of Science and Technology. Downloaded on April 01,2010 at 22:49:46 EDT from IEEE Xplore. Restrictions apply. </span></div></div><div class="pi" data-data='{"ctm":[1.568627,0.000000,0.000000,1.568627,0.000000,0.000000]}'></div></div>
</body>
</html>
<div id="pf2" class="pf w0 h0" data-page-no="2"><div class="pc pc2 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://static.pudn.com/prod/directory_preview_static/6247201d62b5053d3c2a43f8/bg2.jpg"><div class="t m0 x10 h8 y64 ff1 fs3 fc0 sc0 ls9 ws64">The outline of the paper is as fol<span class="_ _0"></span>lows. In the present </div><div class="t m0 x8 h8 y65 ff1 fs3 fc0 sc0 ls14 ws65">section, the<span class="_ _0"></span> topic biometrics is <span class="_ _0"></span>introduced a<span class="_ _0"></span>nd the choice of </div><div class="t m0 x8 h8 y66 ff1 fs3 fc0 sc0 ls28 ws66">face as the <span class="_ _0"></span>underlying signatur<span class="_ _0"></span>e for biometric system<span class="_ _0"></span> </div><div class="t m0 x8 h8 y67 ff1 fs3 fc0 sc0 ls8 ws67">developme<span class="_ _0"></span>nt is jus<span class="_ _0"></span>tified. In secti<span class="_ _0"></span>on II, we<span class="_ _0"></span> make a br<span class="_ _0"></span>ief </div><div class="t m0 x8 h8 y68 ff1 fs3 fc0 sc0 ls12 ws68">literature s<span class="_ _0"></span>urvey on face recognition.<span class="_ _0"></span> In section III, we </div><div class="t m0 x8 h8 y69 ff1 fs3 fc0 sc0 ls13 ws69">describe our technique of face <span class="ls30 ws6a">recognition. This is followed </span></div><div class="t m0 x8 h8 y6a ff1 fs3 fc0 sc0 ls1c ws6b">by the presentati<span class="_ _0"></span>on of performance r<span class="_ _0"></span>esults of our prototype<span class="_ _0"></span> </div><div class="t m0 x8 h8 y6b ff1 fs3 fc0 sc0 ls31 ws6c">system in section IV, an<span class="_ _1"></span>d con<span class="ls32 ws6d">cluding remarks in section<span class="_ _1"></span> V. </span></div><div class="t m0 x11 h6 y6c ff3 fs3 fc0 sc0 ls1e ws6e">2. BACKGROUND OF THE WORK </div><div class="t m0 x8 h8 y6d ff1 fs3 fc0 sc0 lsd ws6f">For over t<span class="_ _1"></span>wo decades [4] face <span class="_ _1"></span>recognition<span class="_ _1"></span> has drawn att<span class="_ _1"></span>ention </div><div class="t m0 x8 h8 y6e ff1 fs3 fc0 sc0 ls33 ws70">of the rese<span class="_ _1"></span>arch communit<span class="_ _1"></span>y. Face id<span class="_ _1"></span>entificatio<span class="_ _1"></span>n from a single </div><div class="t m0 x8 h8 y6f ff1 fs3 fc0 sc0 ls21 ws71">image is a challe<span class="_ _0"></span>nging task bec<span class="_ _0"></span>ause of variable<span class="_ _0"></span> factors like<span class="_ _0"></span> </div><div class="t m0 x8 h8 y70 ff1 fs3 fc0 sc0 ls25 ws72">alterations in scale, location,<span class="_ _0"></span> pose, facial express<span class="_ _0"></span>ion, </div><div class="t m0 x8 h8 y71 ff1 fs3 fc0 sc0 ls34 ws73">occlusion, lighting conditions<span class="_ _0"></span> and overall<span class="ls14 ws74"> appearance <span class="_ _0"></span>of the </span></div><div class="t m0 x8 h8 y72 ff1 fs3 fc0 sc0 ls5 ws75">face. With the<span class="_ _0"></span> synergy of efforts from<span class="_ _0"></span> researchers in dive<span class="_ _0"></span>rse </div><div class="t m0 x8 h8 y73 ff1 fs3 fc0 sc0 ls14 ws76">fields including com<span class="_ _0"></span>puter engineering, mathem<span class="_ _0"></span>atics, </div><div class="t m0 x8 h8 y74 ff1 fs3 fc0 sc0 ls29 ws77">neuroscienc<span class="_ _0"></span>e and psychophysics,<span class="_ _0"></span> different framew<span class="_ _0"></span>orks have </div><div class="t m0 x8 h8 y75 ff1 fs3 fc0 sc0 ls35 ws78">evolved for sol<span class="_ _1"></span>ving the<span class="_ _1"></span> problem of face recogni<span class="_ _1"></span>tion. Amon<span class="_ _1"></span>g </div><div class="t m0 x8 h8 y76 ff1 fs3 fc0 sc0 ls2b ws79">these, the pr<span class="_ _0"></span>ominent a<span class="_ _0"></span>pproaches ar<span class="_ _0"></span>e those based o<span class="_ _0"></span>n Principal </div><div class="t m0 x8 h8 y77 ff1 fs3 fc0 sc0 ls12 ws7a">Component Ana<span class="_ _0"></span>lysis (PCA), Loca<span class="_ _0"></span>l Feature Ana<span class="_ _0"></span>lysis (LFA), <span class="_ _3"></span> </div><div class="t m0 x8 h8 y78 ff1 fs3 fc0 sc0 ls12 ws7b">Template Matching, <span class="_ _0"></span>Neural Network, <span class="_ _0"></span>Model Matching, </div><div class="t m0 x8 h8 y79 ff1 fs3 fc0 sc0 ls4 ws7c">Partitioned Iterated Func<span class="_ _0"></span>tion System (PIFS), Wavelets and </div><div class="t m0 x8 h8 y7a ff1 fs3 fc0 sc0 ls36 ws2a">Discrete Cosine Transform<span class="_ _0"></span> (DCT). The<span class="ls32 ws7d"> choice of a particular </span></div><div class="t m0 x8 h8 y7b ff1 fs3 fc0 sc0 ls1c ws7e">solution is governed by its suita<span class="_ _0"></span>bility in a particular </div><div class="t m0 x8 h8 y7c ff1 fs3 fc0 sc0 ls21 ws2">applica<span class="_ _0"></span>tion. </div><div class="t m0 x10 h8 y7d ff1 fs3 fc0 sc0 ls6 ws7f">In PCA met<span class="_ _1"></span>hod, also<span class="_ _1"></span> known as Eig<span class="_ _1"></span>enface meth<span class="_ _1"></span>od, </div><div class="t m0 x8 h8 y7e ff1 fs3 fc0 sc0 ls37 ws80">face image<span class="_ _1"></span>s are proj<span class="_ _1"></span>ected onto t<span class="_ _1"></span>he so call<span class="_ _1"></span>ed eigen<span class="_ _1"></span>space [6] </div><div class="t m0 x8 h8 y7f ff1 fs3 fc0 sc0 ls18 ws81">that best enc<span class="_ _0"></span>odes the varia<span class="_ _0"></span>tions among known facial c<span class="_ _0"></span>lasses, </div><div class="t m0 x8 h8 y80 ff1 fs3 fc0 sc0 ls25 ws82">and recognition is achie<span class="_ _0"></span>ved by carrying out matc<span class="_ _0"></span>h of these </div><div class="t m0 x8 h8 y81 ff1 fs3 fc0 sc0 ls2b ws60">projected fea<span class="_ _0"></span>ture vectors. The advan<span class="_ _0"></span>tage of this metho<span class="_ _0"></span>d is </div><div class="t m0 x8 h8 y82 ff1 fs3 fc0 sc0 ls38 ws83">real-time recognition, <span class="_ _1"></span>but the m<span class="ls39 ws84">ethod in itself is sensitive to </span></div><div class="t m0 x8 h8 y83 ff1 fs3 fc0 sc0 ls5 ws85">change in il<span class="_ _0"></span>lumination,<span class="_ _0"></span> facial orienta<span class="_ _0"></span>tion and its size<span class="_ _0"></span>. LFA [8] </div><div class="t m0 x8 h8 y84 ff1 fs3 fc0 sc0 ls15 ws86">method of rec<span class="_ _0"></span>ognitio<span class="_ _0"></span>n is based on ana<span class="_ _0"></span>lysing the face <span class="_ _0"></span>in terms </div><div class="t m0 x8 h8 y85 ff1 fs3 fc0 sc0 ls28 ws87">of local feat<span class="_ _0"></span>ures, e.g., eyes, nose, etc.<span class="_ _0"></span> by what is referred to as </div><div class="t m0 x8 h8 y86 ff1 fs3 fc0 sc0 ls36 ws88">LFA kernels. LFA <span class="_ _0"></span>technique offers better robustness a<span class="_ _0"></span>gainst </div><div class="t m0 x8 h8 y87 ff1 fs3 fc0 sc0 ls3a ws89">local va<span class="_ _1"></span>riations on<span class="_ _1"></span> the facial<span class="_ _1"></span> image in<span class="_ _1"></span> carrying<span class="_ _1"></span> out a match<span class="_ _1"></span>, </div><div class="t m0 x8 h8 y88 ff1 fs3 fc0 sc0 ls15 ws3d">but does no<span class="_ _0"></span>t account for<span class="_ _0"></span> global fac<span class="_ _0"></span>ial attribu<span class="_ _0"></span>tes. Face </div><div class="t m0 x8 h8 y89 ff1 fs3 fc0 sc0 ls3b ws8a">recognit<span class="_ _0"></span>ion based on<span class="_ _0"></span> template m<span class="_ _0"></span>atching [13] repr<span class="_ _0"></span>esents a </div><div class="t m0 x8 h8 y8a ff1 fs3 fc0 sc0 ls20 ws8b">face in terms of a template consis<span class="_ _0"></span>ting of several masks<span class="_ _0"></span> </div><div class="t m0 x8 h8 y8b ff1 fs3 fc0 sc0 lsf ws8c">enclosing t<span class="_ _0"></span>he prominent featur<span class="_ _0"></span>es e.g. the eyes, the nose and </div><div class="t m0 x8 h8 y8c ff1 fs3 fc0 sc0 lsd ws8d">the mouth. Matchin<span class="_ _1"></span>g is usually carried out<span class="_ _1"></span> by a correlati<span class="_ _1"></span>on </div><div class="t m0 x8 h8 y8d ff1 fs3 fc0 sc0 ls18 ws8e">score computed from the <span class="_ _0"></span>pixel intensities<span class="_ _0"></span> of these masks<span class="_ _0"></span> </div><div class="t m0 x8 h8 y8e ff1 fs3 fc0 sc0 ls5 ws8f">taken from th<span class="_ _0"></span>e reference, with t<span class="_ _0"></span>he query image<span class="_ _0"></span>. However, </div><div class="t m0 x8 h8 y8f ff1 fs3 fc0 sc0 ls9 ws90">this method is se<span class="_ _0"></span>nsitive to scale, orie<span class="_ _0"></span>ntation, and nonlinea<span class="_ _0"></span>r </div><div class="t m0 x8 h8 y90 ff1 fs3 fc0 sc0 ls3b ws91">illuminat<span class="_ _0"></span>ion changes of t<span class="_ _0"></span>he face. Re<span class="_ _0"></span>cognition <span class="_ _0"></span>by Neural </div><div class="t m0 x8 h8 y91 ff1 fs3 fc0 sc0 ls22 ws92">Network [11]<span class="_ _0"></span> are based on learn<span class="_ _0"></span>ing of the faces in<span class="_ _0"></span> an </div><div class="t m0 x8 h8 y92 ff1 fs3 fc0 sc0 ls4 ws93">‘Example Se<span class="_ _0"></span>t’ by the machine <span class="_ _0"></span>in the ‘Training Phase’<span class="_ _0"></span> and </div><div class="t m0 x8 h8 y93 ff1 fs3 fc0 sc0 ls3c ws94">carrying out reco<span class="_ _1"></span>gnition in the ‘Gen<span class="_ _1"></span>eralization Ph<span class="_ _1"></span>ase’. But to </div><div class="t m0 x8 h8 y94 ff1 fs3 fc0 sc0 ls1c ws95">succeed in a<span class="_ _0"></span> practical set-up, the exa<span class="_ _0"></span>mples should be </div><div class="t m0 x8 h8 y95 ff1 fs3 fc0 sc0 ls22 ws96">adequately l<span class="_ _0"></span>arge in number<span class="_ _0"></span> to account for variat<span class="_ _0"></span>ions in real </div><div class="t m0 x8 h8 y96 ff1 fs3 fc0 sc0 ls1a ws97">life situat<span class="_ _0"></span>ions. Model Matching methods of face r<span class="_ _0"></span>ecognition </div><div class="t m0 x8 h8 y97 ff1 fs3 fc0 sc0 ls28 ws98">(like Hidden Ma<span class="_ _0"></span>rkov Model (H<span class="_ _0"></span>MM) [12]) train a model <span class="_ _0"></span>for </div><div class="t m0 x8 h8 y98 ff1 fs3 fc0 sc0 ls3b ws99">every person dur<span class="_ _0"></span>ing model learni<span class="_ _0"></span>ng and choose the bes<span class="_ _0"></span>t </div><div class="t m0 x8 h8 y99 ff1 fs3 fc0 sc0 ls10 ws9a">matching<span class="_ _1"></span> model, gi<span class="_ _1"></span>ven a query<span class="_ _1"></span> image. Succ<span class="_ _1"></span>ess of these </div><div class="t m0 x8 h8 y9a ff1 fs3 fc0 sc0 ls9 ws9b">methods largely <span class="_ _0"></span>depends on building realistic repr<span class="_ _0"></span>esentative </div><div class="t m0 xb h8 y9b ff1 fs3 fc0 sc0 ls9 ws9c">models. Recognition <span class="_ _0"></span>technique formula<span class="_ _0"></span>ted on Partitioned </div><div class="t m0 xb h8 y9c ff1 fs3 fc0 sc0 ls20 ws9d">Iterated Function Sys<span class="_ _0"></span>tem (PIFS) [9] ma<span class="_ _0"></span>kes use of the fact </div><div class="t m0 xb h8 y9d ff1 fs3 fc0 sc0 ls1a ws9e">that human face shows reg<span class="ls36 ws9f">ion-wise<span class="_ _0"></span> (fractal) self-similarity, </span></div><div class="t m0 xb h8 y9e ff1 fs3 fc0 sc0 ls9 wsa0">which is utilised f<span class="_ _0"></span>or encoding the face to generate<span class="_ _0"></span> the PIFS </div><div class="t m0 xb h8 y9f ff1 fs3 fc0 sc0 ls8 wsa1">code. Reco<span class="_ _0"></span>gnition is perfor<span class="_ _0"></span>med by matc<span class="_ _0"></span>hing these PIFS </div><div class="t m0 xb h8 ya0 ff1 fs3 fc0 sc0 ls4 wsa2">codes. Although P<span class="_ _0"></span>IFS code extracts t<span class="_ _0"></span>he signature of the face </div><div class="t m0 xb h8 ya1 ff1 fs3 fc0 sc0 ls3d wsa3">efficiently<span class="_ _1"></span>, the techniqu<span class="_ _1"></span>e is not rob<span class="_ _1"></span>ust against<span class="_ _1"></span> the facial </div><div class="t m0 xb h8 ya2 ff1 fs3 fc0 sc0 ls34 wsa4">variations occurring from instance <span class="_ _0"></span>to instance. Selection of </div><div class="t m0 xb h8 ya3 ff1 fs3 fc0 sc0 ls20 wsa5">the prominent c<span class="_ _0"></span>oefficients from the wave<span class="_ _0"></span>let transform (WT) </div><div class="t m0 xb h8 ya4 ff1 fs3 fc0 sc0 ls8 wsa6">is effective in extr<span class="_ _0"></span>acting the signat<span class="_ _0"></span>ure of a face and </div><div class="t m0 xb h8 ya5 ff1 fs3 fc0 sc0 ls14 wsa7">eliminating the r<span class="_ _0"></span>edundancies. Moreover, WT <span class="_ _0"></span>provides a </div><div class="t m0 xb h8 ya6 ff1 fs3 fc0 sc0 ls16 wsa8">inherent handle t<span class="_ _1"></span>o deal the data at<span class="_ _1"></span> different scales. Ref<span class="_ _1"></span>. [12] </div><div class="t m0 xb h8 ya7 ff1 fs3 fc0 sc0 lsf wsa9">deals with WT employed <span class="_ _0"></span>to carry out face r<span class="_ _0"></span>ecognition. D<span class="_ _0"></span>CT-</div><div class="t m0 xb h8 ya8 ff1 fs3 fc0 sc0 ls15 wsaa">based recogniti<span class="_ _0"></span>on technique [7]<span class="_ _0"></span> depends on the capabi<span class="_ _0"></span>lities of </div><div class="t m0 xb h8 ya9 ff1 fs3 fc0 sc0 ls18 wsab">the discrete cosine<span class="_ _0"></span> transform to extract the facial signature in </div><div class="t m0 xb h8 yaa ff1 fs3 fc0 sc0 ls10 wsac">terms of a<span class="_ _1"></span> few DCT coeffici<span class="_ _1"></span>ents. Recog<span class="_ _1"></span>nition is<span class="_ _1"></span> achieved by </div><div class="t m0 xb h8 yab ff1 fs3 fc0 sc0 ls36 wsad">matching the DCT signatures. Both wavelet and DCT have<span class="_ _0"></span> </div><div class="t m0 xb h8 yac ff1 fs3 fc0 sc0 ls12 wsae">promising future as the underlying techniques<span class="_ _0"></span> for </div><div class="t m0 xb h8 yad ff1 fs3 fc0 sc0 ls20 wsaf">implementat<span class="_ _0"></span>ion of successful <span class="_ _0"></span>face recognition s<span class="_ _0"></span>ystems. </div><div class="t m0 xb h8 yae ff1 fs3 fc0 sc0 ls3c wsb0">Fusing the scores o<span class="_ _1"></span>f several dif<span class="_ _1"></span>ferent cl<span class="_ _1"></span>assifie<span class="_ _1"></span>rs applied on </div><div class="t m0 xb h8 yaf ff1 fs3 fc0 sc0 ls1b ws4c">the same data is a very pr<span class="_ _0"></span>omising approa<span class="_ _0"></span>ch to improve the </div><div class="t m0 xb h8 yb0 ff1 fs3 fc0 sc0 ls33 wsb1">overall ac<span class="_ _1"></span>curacy of th<span class="_ _1"></span>e biometri<span class="_ _1"></span>c systems [18<span class="_ _1"></span>]. We mak<span class="_ _1"></span>e use </div><div class="t m0 xb h8 yb1 ff1 fs3 fc0 sc0 ls2b wsb2">of such intram<span class="_ _0"></span>odal fusion [14]<span class="_ _0"></span> combining cor<span class="_ _0"></span>relation-base<span class="_ _0"></span>d </div><div class="t m0 xb h8 yb2 ff1 fs3 fc0 sc0 ls1a wsb3">template matching wit<span class="_ _0"></span>h PCA or DCT methods applied <span class="_ _0"></span>on the </div><div class="t m0 xb h8 yb3 ff1 fs3 fc0 sc0 ls33 wsb4">same faci<span class="_ _1"></span>al dat<span class="_ _1"></span>a to decid<span class="_ _1"></span>e the a<span class="_ _1"></span>uthenti<span class="_ _1"></span>city of a<span class="_ _1"></span> subj<span class="_ _1"></span>ect. In </div><div class="t m0 xb h8 yb4 ff1 fs3 fc0 sc0 ls12 wsb5">this paper, we present a prototype system im<span class="_ _0"></span>plementing our </div><div class="t m0 xb h8 yb5 ff1 fs3 fc0 sc0 ls39 wsb6">technique of face recognition. </div><div class="t m0 x12 h6 yb6 ff3 fs3 fc0 sc0 ls32 ws6d">3.<span class="_ _5"> </span>MULTI-ALGORITHMIC FACE </div><div class="t m0 x13 h6 yb7 ff3 fs3 fc0 sc0 ls3e ws2">RECOGNITION </div><div class="t m0 xb h8 yb8 ff1 fs3 fc0 sc0 ls12 wsb7"> <span class="_ _7"></span>Earlier we developed a tec<span class="_ _0"></span>hnique [1] of verifying human </div><div class="t m0 xb h8 yb9 ff1 fs3 fc0 sc0 ls2b wsb8">face by matc<span class="_ _0"></span>hing against <span class="_ _0"></span>templates re<span class="_ _0"></span>trieved from t<span class="_ _0"></span>he </div><div class="t m0 xb h8 yba ff1 fs3 fc0 sc0 ls4 wsb9">reference database c<span class="_ _0"></span>reated during registration process. In this </div><div class="t m0 xb h8 ybb ff1 fs3 fc0 sc0 ls28 wsba">technique the matching is car<span class="_ _0"></span>ried out in terms of a set of </div><div class="t m0 xb h8 ybc ff1 fs3 fc0 sc0 ls3f wsbb">correlation scores corresponding to different are<span class="_ _0"></span>as of interest </div><div class="t m0 xb h8 ybd ff1 fs3 fc0 sc0 ls2 wsbc">(rectangular bit-maps represe<span class="_ _0"></span>nting different regions of the </div><div class="t m0 xb h8 ybe ff1 fs3 fc0 sc0 ls3b wsbd">face) and <span class="_ _0"></span>their Eucli<span class="_ _0"></span>dean distance<span class="_ _0"></span>s measured i<span class="_ _0"></span>n pixels. Thi<span class="_ _0"></span>s </div><div class="t m0 xb h8 ybf ff1 fs3 fc0 sc0 ls3a wsbe">techniqu<span class="_ _1"></span>e can b<span class="_ _1"></span>e extend<span class="_ _1"></span>ed for <span class="_ _1"></span>identi<span class="_ _1"></span>fication<span class="_ _1"></span> by matchi<span class="_ _1"></span>ng the </div><div class="t m0 xb h8 yc0 ff1 fs3 fc0 sc0 ls28 wsbf">input face agains<span class="_ _0"></span>t every registered ide<span class="_ _0"></span>ntity to choose the one<span class="_ _0"></span> </div><div class="t m0 xb h8 yc1 ff1 fs3 fc0 sc0 ls29 wsc0">which gives the best correlat<span class="_ _0"></span>ion scores and minim<span class="_ _0"></span>um </div><div class="t m0 xb h8 yc2 ff1 fs3 fc0 sc0 ls27 wsc1">distance error. Unfortunately, correlation is compute-</div><div class="t m0 xb h8 yc3 ff1 fs3 fc0 sc0 ls14 wsc2">intensive, and to match a face against a<span class="_ _0"></span> large number of </div><div class="t m0 xb h8 yc4 ff1 fs3 fc0 sc0 lsb ws2a">reference faces using correlation software ta<span class="_ _0"></span>kes unacceptably </div><div class="t m0 xb h8 yc5 ff1 fs3 fc0 sc0 lsf wsc3">long duratio<span class="_ _0"></span>n for a practical a<span class="_ _0"></span>pplicatio<span class="_ _0"></span>n. </div><div class="t m0 xb h5 yc6 ff2 fs3 fc0 sc0 ls25 wsc4">A PCA-based Approach </div><div class="t m0 xb h8 yc7 ff1 fs3 fc0 sc0 ls14 ws74">A survey [2,4] on the availa<span class="_ _0"></span>ble literature reveale<span class="_ _0"></span>d the main </div><div class="t m0 xb h8 yc8 ff1 fs3 fc0 sc0 ls33 wsc5">techniqu<span class="_ _1"></span>es of face reco<span class="_ _1"></span>gnition whi<span class="_ _1"></span>ch are men<span class="_ _1"></span>tioned in<span class="_ _1"></span> </div><div class="t m0 xb h8 yc9 ff1 fs3 fc0 sc0 ls14 wsc2">section II. We decided to <span class="_ _0"></span>implement the PCA method of </div><div class="t m0 xb h8 yca ff1 fs3 fc0 sc0 ls9 ws15">recognition first. In this met<span class="_ _0"></span>hod as well as those descr<span class="_ _0"></span>ibed in </div><div class="t m0 xb h5 ycb ff1 fs3 fc0 sc0 ls1b wsc6">subsections 3<span class="ff2 ls2 ws2">B</span><span class="ls22 wsc7"> an<span class="_ _0"></span>d 3<span class="ff2 ls2 ws2">C</span><span class="ls35 wsc8">, our prototype reco<span class="_ _1"></span>gnition sy<span class="_ _1"></span>stem </span></span></div><div class="t m0 xb h5 ycc ff1 fs3 fc0 sc0 ls35 wsc9">(describ<span class="_ _1"></span>ed in the l<span class="_ _1"></span>ater part<span class="_ _1"></span> of subs<span class="_ _1"></span>ection 3<span class="ff2 ls2 ws2">C</span><span class="ls13 wsca">) makes use of </span></div><div class="t m0 xb h8 ycd ff1 fs3 fc0 sc0 lsf wscb">reasonably c<span class="_ _0"></span>onstrained im<span class="_ _0"></span>aging set-u<span class="_ _0"></span>p with facial im<span class="_ _0"></span>age </div><div class="t m0 xb h8 yce ff1 fs3 fc0 sc0 ls20 ws8b">grabbe</div><div class="t m0 x14 h8 ycf ff1 fs3 fc0 sc0 ls20 ws8b">d in frontal ge<span class="_ _0"></span>ometry under<span class="_ _0"></span> sufficient illum<span class="_ _0"></span>ination </div><div class="t m0 xb h8 yd0 ff1 fs3 fc0 sc0 ls2b wscc">level in order<span class="_ _0"></span> to minimise the variations<span class="_ _0"></span> in the acquired </div><div class="t m0 xb h8 yd1 ff1 fs3 fc0 sc0 ls2f wscd">image. We ma<span class="_ _1"></span>de use<span class="_ _1"></span> of the open so<span class="_ _1"></span>urce routi<span class="_ _1"></span>ne [17] for </div><div class="t m0 x15 h9 y62 ff4 fs5 fc0 sc0 ls2 ws2">322</div><div class="t m0 xf ha y63 ff5 fs6 fc0 sc0 ls2 ws2"><span class="fc1 sc0">Authorized licensed use limited to: Harbin University of Science and Technology. Downloaded on April 01,2010 at 22:49:46 EDT from IEEE Xplore. Restrictions apply. </span></div></div><div class="pi" data-data='{"ctm":[1.568627,0.000000,0.000000,1.568627,0.000000,0.000000]}'></div></div>
