11.07.2015 Views

Fast Detection of Multi-View Face and Eye Based on Cascaded ...

Fast Detection of Multi-View Face and Eye Based on Cascaded ...

Fast Detection of Multi-View Face and Eye Based on Cascaded ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

4) Output the classifier Ksign f k( x) k 1A weak classifier c<strong>on</strong>sists <str<strong>on</strong>g>of</str<strong>on</strong>g> a threshold value <str<strong>on</strong>g>and</str<strong>on</strong>g> asimple feature, as shown in Fig. 1. The specific valuesare determined according to an iterative learning. Thesimple features are designed to detect an edge or a line<str<strong>on</strong>g>of</str<strong>on</strong>g> the face easily. Several supplementary features areadded to Viola <str<strong>on</strong>g>and</str<strong>on</strong>g> J<strong>on</strong>es’ features [10].In the learning stage, all possible positi<strong>on</strong>s, sizes, <str<strong>on</strong>g>and</str<strong>on</strong>g>types <str<strong>on</strong>g>of</str<strong>on</strong>g> features are c<strong>on</strong>sidered within a 2424 window.There are a total number <str<strong>on</strong>g>of</str<strong>on</strong>g> 117,400 features with somerestricti<strong>on</strong>s to their freedom such as minimum area. Aface/n<strong>on</strong>-face str<strong>on</strong>g classifier is applied to the inputimage in all possible positi<strong>on</strong>s with all possible sizes inorder to detect all the faces with various positi<strong>on</strong>s <str<strong>on</strong>g>and</str<strong>on</strong>g>sizes.However, accordingaswerotatethesimplefeature<str<strong>on</strong>g>of</str<strong>on</strong>g>adetectorwith90°orperform amirroringoperati<strong>on</strong><strong>on</strong>thesimplefeature, wegetotherdetectors.Forexample,aleft-view detectorcanbearight-view detector.Also,12 fr<strong>on</strong>tal-view detectorscanbemadefrom <strong>on</strong>ly2 detectors.Astheresult,thenumber<str<strong>on</strong>g>of</str<strong>on</strong>g>facedetectorstobetrainedisshowninTableI.3.2 The Structure <str<strong>on</strong>g>of</str<strong>on</strong>g> <str<strong>on</strong>g>Multi</str<strong>on</strong>g>-<str<strong>on</strong>g>View</str<strong>on</strong>g> <str<strong>on</strong>g>Face</str<strong>on</strong>g> DetectorInordertodetectmulti-view face, whenweuse12 or72 facedetectorsindependently, weneed12 timesor72times<str<strong>on</strong>g>of</str<strong>on</strong>g>computati<strong>on</strong>comparingtoasinglefacedetector.Ifweadoptaposeestimator, therehappensaposeestimator’errorproblemaswementi<strong>on</strong>ed.Weproposeanew multi-view detectorbased <strong>on</strong> a specialcascadedclassifier structure using coarse-to-fine search, simple-to-complexsearch, <str<strong>on</strong>g>and</str<strong>on</strong>g> paralel-to-separated searchasshowninFig.2.(a)(b)Fig. 1. Haar-like simple features: (a) edge features;(b)line features.3 <str<strong>on</strong>g>Multi</str<strong>on</strong>g>-<str<strong>on</strong>g>View</str<strong>on</strong>g> <str<strong>on</strong>g>Face</str<strong>on</strong>g> <str<strong>on</strong>g>Detecti<strong>on</strong></str<strong>on</strong>g>3.1 Detectable Rotati<strong>on</strong> AngleFor x-axis rotati<strong>on</strong>, we need <strong>on</strong>ly 2 detectors. Theyare a down-view face detector covering [-60, -20]<str<strong>on</strong>g>and</str<strong>on</strong>g>afr<strong>on</strong>tal/upwardfacedetectorcovering[-20, 50].Fory-axisrotati<strong>on</strong>, weneed3 detectorsincludingaleft-viewfacedetector, afr<strong>on</strong>talfacedetector<str<strong>on</strong>g>and</str<strong>on</strong>g> aright-viewface detector.Theircovering anglesare [-90, -20],[-20, 20], [20, 90], respectively.Forz-axisrotati<strong>on</strong>, wedealwithalrotati<strong>on</strong>covering[-180, 180].However, duringst<str<strong>on</strong>g>and</str<strong>on</strong>g>ing, apers<strong>on</strong>canlean his/her head with [-45, 45].W e cal“Basicmode”<str<strong>on</strong>g>of</str<strong>on</strong>g>z-axisrotati<strong>on</strong>as[-45, 45]<str<strong>on</strong>g>and</str<strong>on</strong>g>“Extendedmode”<str<strong>on</strong>g>of</str<strong>on</strong>g>z-axisrotati<strong>on</strong>as[-180, 180].W hen wedesign adetectorcovering 30°<strong>on</strong> z-axisrotati<strong>on</strong>, 12 detectorscoverstheextendedmode, [-180°,180°].Forthe basic mode, 3 detectorsare sufficient.Thisiscaledas“MethodI.”W henwedesignadetectorcovers45°, 8detectorscancoverstheextendedmode.Forthebasicmode, 2 detectorsaresuficient.Thisiscaled as“Method II.”W hen wec<strong>on</strong>siderx, y, z-axisrotati<strong>on</strong>, thenumber<str<strong>on</strong>g>of</str<strong>on</strong>g>facedetectorsneededisshowninTableI.TableI.Thenumbers<str<strong>on</strong>g>of</str<strong>on</strong>g>individualfacedetectorsneededinmulti-view facedetecti<strong>on</strong>.MethodIMethodIINumber<str<strong>on</strong>g>of</str<strong>on</strong>g>faceDetectorsneeded.Number<str<strong>on</strong>g>of</str<strong>on</strong>g>facedetectorstobetrainedBasicmode 18=233 10=25Extendedmode72=2312 10=25Basicmode 12=232 6=23Extendedmode48=238 6=23(a) (b) (c)Fig.2.Threemethodsrenderingacascadedclassifierusedin the multi-view face detector;(a)Coarse-to-fine search; (b) Simple-to-complexsearch;(c)Paralel-to-separatedsearch.Coarse-to-finesearchisthatawhole-view classifierislocatedinearlystage<str<strong>on</strong>g>and</str<strong>on</strong>g>narrower-view classifiersarelocated in latestage.Simple-to-complex search isthateasierclassifiersare located in early stage <str<strong>on</strong>g>and</str<strong>on</strong>g> morecomplexclassifiersarelocatedinlatestage.Usingthesetwostages, wecanspeedamulti-view detectorupsincemostn<strong>on</strong>-facesareeliminatedinearlystages.Paralel-to-separated search isthataldetectorsarearranged in paraleluntilK stage<str<strong>on</strong>g>and</str<strong>on</strong>g> each detectorisarrangedseparatelyfrom K+1stage.Usualy, whenaninputisentered<str<strong>on</strong>g>and</str<strong>on</strong>g>afaceisdetectedsuccessfulyinacertainstage<str<strong>on</strong>g>of</str<strong>on</strong>g>aview, theinputmovest<strong>on</strong>extstage<str<strong>on</strong>g>of</str<strong>on</strong>g>thesameview tobetested.Paralelarrangementmeansthatwhenafaceisnotdetected, theinputmovestothesamestage<str<strong>on</strong>g>of</str<strong>on</strong>g>nextview.Separatedarrangementmeansthatwhenafaceisnotdetected, nomoreprocedureisneeded.Wejustdecidetheinputisan<strong>on</strong>-face.Usingthismethod, aninputimageisdecidedasacertainviewinearlystage, <str<strong>on</strong>g>and</str<strong>on</strong>g>thenwec<strong>on</strong>centrate<strong>on</strong>whethertheinputisafaceornotinlatestage.InFig.2, s st<str<strong>on</strong>g>and</str<strong>on</strong>g>sforsuccessinfacedetecti<strong>on</strong><str<strong>on</strong>g>and</str<strong>on</strong>g>f st<str<strong>on</strong>g>and</str<strong>on</strong>g>sforfailure.We combine three cascading methods into amulti-view facedetectorasshowninFig.3.In1~2 stage,whole-view faceisdetected.In3~4stage, facegroupssuchasuprightface, left-leanedface, right-leanedfacearedetected.In5~M stage, alview facearedetected.Also, paralelsearchesare implemented untilK stage<str<strong>on</strong>g>and</str<strong>on</strong>g>separatedsearchesareimplementedfrom K+1stage.InFig.3, NF st<str<strong>on</strong>g>and</str<strong>on</strong>g>sforn<strong>on</strong>-face, s st<str<strong>on</strong>g>and</str<strong>on</strong>g>sforsuccessinfacedetecti<strong>on</strong>, <str<strong>on</strong>g>and</str<strong>on</strong>g>f st<str<strong>on</strong>g>and</str<strong>on</strong>g>sforfailure.117

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!