基于AdaBoost的人脸检测赵楠.doc

Size: px

Start display at page:

Download "基于AdaBoost的人脸检测赵楠.doc"

诲妈公叔
5 years ago
Views:

1 AdaBoo Face Deeco Baed o AdaBoo

2 Abrac AdaBoo 995 AdaBoo PCA Boog Adapve Boog AdaBoo AdaBoo AdaBoo FáDè Vola Keyword AdaBoo Boog PCA II

3 A- G- C- 0 h derao dedcaed o A,, G, C, ad 0, he buldg block of ellgece. ad o he poeer ucoverg he foudao of ellgece. III

4 Coe ABSRAC... II CONENS... IV LIS OF FIGURES...VII LIS OF ABLES... VIII NEURAL NEWORK EIGENFACE EXAMPLE-BASED MEHODS SUPPOR VECOR MACHINE SVM HIDDEN MARKOV MODEL HMM ADABOOS IV

5 4.2 PAC BOOSING RECANGLE FEAURE INEGRAL IMAGE ADABOOS WEAK CLASSIFER fx θ p V

6 6.3 SRONG CLASSIFIER ACKNOWLEDGMENS REFERENCES Ver 0.76 VI

7 L of Fgure ROWLEY A x, y x, y A x, y y D ADABOOS A A B B VII

8 VIOLA 5/ 0/ VIOLA 7/ / / 2/ / / / 0/ // / 0/ / 0/ / 0/ / 0/ / 0/ / 0/ L of able , AB VIII

9 AdaBoo. face deeco face recogo

10 AdaBoo

11 AdaBoo SVM

12 MI Yale CMU 6 27 MI fp://whechapel.eda..edu/pub/age FERE hp:// AdaBoo 4

UMIS Uv. of Ber Yale A& Olve Harvard M2VS Purdue AR 20 564 hp://age.ee.u.ac.uk/day/daabae.hl 30 300 0 fp://afp.ube.ch/pub/age/faceage/ 50 5 hp://cvc.yale.

13 UMIS Uv. of Ber Yale A& Olve Harvard M2VS Purdue AR hp://age.ee.u.ac.uk/day/daabae.hl fp://afp.ube.ch/pub/age/faceage/ 50 5 hp://cvc.yale.edu 40 0 hp:// fp://fp.hrl.harvard.edu/pub/face hp://poedo.cd.auh.gr/2v/dex.hl 3276 hp://rvl.ec.purdue.edu/~alex/alex_face_db.hl 3 AdaBoo 5

3 MI 20 20.3.2 Face Recogo echologyfere 4.

14 3 MI Face Recogo echologyfere 4. Deeco Rae 2. Fale Deeco 00% AdaBoo 6

15 00% 0 4 FERE AdaBoo 7

16 2 []. Kowledge-baed 2. Feaure vara 3. eplae achg 4. Appearace-baed [2] 2. [3] AdaBoo 8

17 AdaBoo

18 AdaBoo AdaBoo AdaBoo 0

19 2.4 x p x face p x oface Bayea clafcao axu lkelhood. x 2. p x face p x oface 3. p x face p x oface for ulodal aural paraeerzed p x face p x oface AdaBoo

20 3 3. Neural Nework Rowley 996 [4] Rowley 7 AdaBoo 2

21 7 Rowley 3.2 Egeface Kohoe [5] PCA, Belhueur FherFace [6] BP PCA, Prcpal Copoe Aaly AdaBoo 3

22 3.3 Exaple-Baed Mehod MI Sug Poggo [7] k Sug boorap 8 AdaBoo 4

23 3.4 Suppor Vecor Mache SVM Suppor Vecor Mache SVM BoerGuyoVapk COL-92 Oua [8] Sug [7] 30 Heele [9] SVM SVM SVM 3.5 Hdde Markov Model HMM Hdde Markov ModelHMM HMM HMM HMM Nefa [0] Karhue-Loève Bau-Welch Verb HMM AdaBoo 5

24 4 AdaBoo 4. PAC PAC Boog 995 Freud Schapre AdaBoo [] AdaBoo Adapve Boog AdaBoo Boog Boog adapvely AdaBoo Boog AdaBoo AdaBoo AdaBoo Boog 4.2 PAC 4.2. [2] AdaBoo 6

25 PACProbably Approxaely Correc Vala984 [3] 2 3 PAC PAC PAC 2 3 v, K, v } {0,} { k 4, x H L Px x, K, x L h x h x Lx 2 aple coplexy copuaoal coplexy AdaBoo 7

26 [ x ], h x, K, x, h x ] [ { A } A A :[ ] H, x A x, h x, K, x, h x H, x x, K, x L δ δ h ε L p, / ε,/ δ p, / ε,/ δ L PAC PAC ε ε δ δ % AdaBoo 8

27 Kear [4] Kear Vala [5] 4.4 Boog Boog Boog Kear Vala [5] 990 Schapre [6] Boog Freud Boog [7] 993 Drucker Schapre [8] Boog OCR Boog 995 Freud Schapre Adaboo Boog AdaBoo AdaBoo 9

28 5 5. AdaBoo 5 Vola egral age [9] 5.2 Recagle Feaure ,000 AdaBoo 20

29 AdaBoo 2

30 A x, y B x 2, y2 0 A x, y B x 2, y2,, 3 Paul Vola Mchael J. Joe booed cacade [9][20][2] ,00080,00045,396 AdaBoo 22

31 x 2 y [/] [/] [/] [/] [ ] A x, y x {,2, L,, } y {,2, L,, } 2 A B 0 x Χ x, x 2, L, x p, x p } 2 { y Υ y, y 2, L, y q, y q } 2 { x y p q Χ p Υ q, Ω, x y x y x x p q x y y y L L ,, AdaBoo 23

32 ,, 2 2,, 3 3, 2, 2 3, 5 Ω 5, Ω Ω,2 Ω 2, Ω,3 Ω 3, Ω 2, Ω 2 Ω Ω Ω, 2 2, 3 2, Ω 24, L L 2 2 L Ω 24, ,200 86, L L 5 Ω 24 2, ,600 55, L L 3 3 L L L 2 2 L2 2 L ,736 Ω 24 62, AdaBoo 24

33 , ,725 62,336 78,460 32, Iegral Iage 5.3. f x dx b F x f x dx f x dx F b F a a b a fx a b A x, y x, y 2 AdaBoo 25

34 x', y' x', y' x, y x', y' x' x, y' y 0~255 x, y x, y x, y x,0 0 0, y 0 x, y x, y x, y x, y x, y x, y y x, y 2 y' y x, y', x x x, y y A x, y y 2 A x, y x, y A x, y y AdaBoo 26

35 AdaBoo 27,,,,,,,,,, 2,,, 2,,, 0,,,,,, 0,,, 2,,,,,, 2,,, 0,,,2,,2,2,2,2,2 0,,2,2 0,2,2,,0,,,,,,0,, 2,, 2, 2,0 2, 2,, 2,,,0,, 0,, L M O M K D D A, x y A B C D 2 4 3

36 2 A B 3 A C 4 A B C D D A B A B AdaBoo 28

37 , x 2 3 A B y 4 e coa AdaBoo 29

38 6 AdaBoo AdaBoo [9] A. x, y, x, y, K, x, y 0 y 2 2 y 4 B. w D, C., K,. q, w, j w, j 2. f h x, f, p, θ q ε f 4 D D D l 2 2l l AdaBoo 30

39 ε f q h x, f, p, θ y 3. x ε ε f p, θ h q h x, f, p, θ y q h x, f, p, y, θ h x h x, f, p, θ w w,, β e 0 e x e x β ε ε D. α h x α C x 2 0 α log β AdaBoo 3

40 AdaBoo AdaBoo 32

41 6.2 Weak Clafer h x, f, p, θ f θ p pf x < pθ h x, f, p, θ fx fx k ,460 2,706 4, k 0 6 MI AdaBoo 33

42 6 78,460 A B 7 2,706 4,38 7 FáDè A B x2 AdaBoo 34

43 7 8 A B 2,706 4,38-2,706 4,38-330,953-37,43 468,096-4,480,530 -,60,750-2,878, , , ,33 23,503-9,370 8,763 4, % 5.5% -5% 93.4% 59.4% 34% 6 AB A B 7 A %5.5% 8-5% A 8 B % B B 93.4% 0 34% B 8 A B 0 0 AdaBoo 35

44 A , -5, 9, 2, A A AdaBoo 36

45 B , 2-6,,, B B AdaBoo 37

46 6.2.2 θ p θ p 50% 50% Boog f f f f S 4 S 9 AdaBoo 38

47 F F k j 0 e S S, S S k j I k F k S S Ik 2 I F k 2 F F k F k j k j Ik Ik k f F 9 0 AdaBoo 39

48 6.3 Srog Clafer 6.3. h x, L, h α h x α C x 2 0 x α log β ε log ε log ε α 2 α α ε 2 ε ε AdaBoo 40

49 AdaBoo 4 [ ] [ ],,,,,,,,, e w e q w e w w w e w w w ε β β β β β L x x,, [ ] w, ε β x h 2 α α x h x h x h log log β β α 2 / log log 2 2 β β α 2 / e β β 2 e D w, β 3 23

50 AdaBoo 42 2 / 2 / : :,, e e D w w β ε β 4 : e D ε 4 [ ] 2 / 2 2 ε ε ε ε ε β ε β ε β β ε ε γ exp 4 l 2 exp 4 l exp 4 l exp 4 γ γ γ γ γ ε L l L x x ε ε 0 0 γ γ exp γ ε 40 0., ε ε 98%

51 7 7. AdaBoo.3. FERE MI 7,087 2,706 4, AdaBoo 43

52 ,460 7, ,046,020 PIII 850MHz52Mb uged log 556,046, ,224,84,080 Bye2.07Gbye PC 66Mb/ ,520 7,940 AdaBoo 44

53 46,953 40% MI Phoohop AdaBoo 45

54 k S ks k k k 3 k /3 /5 AdaBoo 46

55 Vola [9] Vola 22 Vola 5/ 0/ 5 AdaBoo 47

56 Vola 23 Vola 7/ / 4 AdaBoo 48

57 / 2/ 25 6/ / 4 AdaBoo 49

58 26 33/ 0/ / 5/ 4 AdaBoo 50

59 28 3/ 0/ / 0/ / 0/ AdaBoo 5

60 3 2/ 0/ 0 32 / 0/ / 0/ 45 AdaBoo 52

61 7.3.3 Vola PIII 700MHz 5 / PIII 850MHz Vola AdaBoo 53

62 Ackowledge Referece [] Yag Mg-HuaDavd J. K.Naredra A.Deecg Face Iage: A Survey IEEE raaco o Paer Aaly ad Mache Iellgece Vol 24:34-58, 2002/0 [2] Hjelå E.Low B. K.Face Deeco: A SurveyCopuer Vo ad Iage Uderadg 83,236~ [3] 2003, Vol 8:4-9 [4] Rowley H. A. e al Neural Nework-Baed Face Deeco IEEE raaco o Paer Aaly ad Mache Iellgece Vol 20:23-38, 998/0 AdaBoo 54

63 [5]. KohoeSelf-Orgazao ad Aocave MeorySprger 989 [6] Belhueur P.N.Hepaha J. P. Krega D. J. Egeface v. Fherface: Recogo Ug Cla Specfc Lear ProjecoIEEE ra. Paer Aaly ad Mache Iellgece, vol. 97, 7-720, July 997 [7] K. K. Sug. PoggoExaple-Baed Learg for Vew-Baed Hua Face Deeco echcal Repor AI Meo 52, Maachue I. of echology AI Lab, 994 [8] E. OuaR. Freudad F. Grorag Suppor Vecor Mache: A Applcao o Face DeecoProc. IEEE Cof. Copuer Vo ad Paer Recogo, 30-36, 997 [9] B. Heele SerreM Pol e alcopoe-baed Face DeecoProc IEEE Copuer Socey Coferece o Copuer Vo ad Paer Recogo, 200, :I657-I662 [0] Nefa A.Haye M.Hdde Markov Model for Face RecogoIEEE Ieraoal Coferece o Acouc, Speech ad Sgal Proceg, Seale, Wahgo, 998: [] Freud Y., Schapre R. E. A Deco-heorec Geeralzao of O-Le Learg ad a Applcao o Boog. Joural of Copuer ad Sye Scece, 997, 55:9-39 [2] Vol 75, , 2000/0 [3] Vala L.G.A heory of he LearableCoucao of he ACMVol 27, 34-42, 984/ [4] Kear M.he Copuaoal Coplexy of Mache LeargCabrdge: MI Pre, 990 [5] Kear M. Vala L. G. Crypographc Lao o Learg Boolea Forulae ad Fe Auoaa Joural of he ACM, 4:67-95, 994/0 [6] Schapre R. E.he Sregh of Weak LearablyMache Learg, 990, 52: AdaBoo 55

64 [7] Freud Y. Boog a Weak Learg Algorh by Majory Iforao ad Copuao, 995, 22: [8] Drucker H.Schapre R. E.Sard P.Boog Perforace Neural NeworkIeraoal Joural of Paer Recogo ad Arfcal Iellgece, 993, 74: [9] Vola P.Joe M. J.Robu Real-e Face DeecoIeraoal Joural of Copuer Vo 572, 37-54, 2004 [20] Vola P.Joe M. J.Rapd Objec Deeco ug a Booed Cacade of Sple FeaureCopuer Vo ad Paer Recogo, 200Volue, 8-4 [2] Vola P.Joe M. J.Robu Real-e Objec DeecoCabrdge Reearch Laboraory, echcal Repor SereCRL 200/0 Powered by a Vvd Logo of FáDè Boog Ver Copyrgh 2005 Zhao Na, vea@2c.coall Rgh Reerved Powered by veasudo, 2005FáDè Boog Vero.00 AdaBoo 56

% %! # % & ( ) % # + # # % # # & & % ( #,. %

!!! # #! # % & % %! # % & ( ) % # + # # % # # & & % ( #,. % , ( /0 ) %, + ( 1 ( 2 ) + %, ( 3, ( 123 % & # %, &% % #, % ( ) + & &% & ( & 4 ( & # 4 % #, #, ( ) + % 4 % & &, & & # / / % %, &% ! # #! # # #

基于AdaBoost的人脸检测 赵楠.doc

基于AdaBoost的人脸检测赵楠.doc