Chng 1:

Tng quan v mng Nron

1.1 Gii thiu chng : Ngy nay, cc k thut sinh trc hc ngy cng c ng dng rng ri. Trong , nhn dng vn tay c xem l mt trong nhng k thut hon thin v ng tin cy nht xc nhn mt ngi. Gn y, k thut ny c ch nhiu v ngi ta thy rng n thch hp vi nhng ng dng c c s d liu nh, nhng khng thun tin cho nhng ng dng c phm vi ln. a s cc h thng bo mt hin nay c bo v bng password v PIN (Personal Identification Number), nhng cc phng php ny c chng minh l khng hiu qu. Bi v, password l nhng con s kh nh, d qun v d b nh cp. Bng cch s dng vn tay v mt m, vic xc nhn mt ngi c th c thc hin bng mt h thng nhn dng vn tay an ton v thun tin. Hnh 1.1 l cu trc c bn ca h thng nhn dng du vn tay. u tin, du vn tay ca mt ngi cn c ly mu (bng mt thit b c th chp c vn tay Biometric sensor) v lu vo c s d liu (Registration module). Sau , khi cn xc nhn ngi cung cp li mt du vn tay khc, du vn tay ny s c so snh vi du vn tay trong c s d liu quyt nh chp nhn hay t chi da trn mt gi tr ngng i snh.

Hnh 1.1: Cu trc c bn ca h thng nhn dng du vn tay Cc phng php nhn dng vn tay kinh in u da vo vic i snh (matching) cc

im c trng (feature) trn vn tay. C nhiu phng php i snh khc nhau. Trong bi ny, chng ti nghin cu phng php i snh bng mng neural nhn to (Artificial Neural Network). ti gii thiu mt hng nghin cu v ng dng lnh vc nhn dng vn tay vo thc tin. Mt lnh vc kh ph bin trn th gii nhng cn hn ch Vit Nam. 1.2 Mng nron sinh hc T bo thn kinh cn gi l neuron. Nghin cu sinh hc v b no con ngi cho thy rng cc neuron l n v c s m nhim nhng chc ngn x l nht nh trong h thn kinh, bao gm: no, ty sng v cc dy thn kinh . Mi neuron c phn thn v nhn bn trong (gi l soma), mt u thn kinh ra (gi l dendrite). Cc dy thn kinhvo to thnh mt li dy c xung quanh thn t bo, chim din tch khong 0,25mm2, cn dy thn kinh to rathnh trc di c th t 1 cm n hng mt. ng knh nhn t bo thng ch l 10-4 m. trc dy thn kinh ra cng c th phn nhnh theo dng cy ni vi dy thn kinh vo hoc trc tip vi nhn t bo cc neuron khc thng qua cc khp ni (gi l synapse). Thng thng, mi neuron c th gm vi chc cho ti hng trm khp ni ni vi cc neuron khc. Ngi ta c lng rng li cc dy thn kinh ra cng vi cc khp ni bao ph din tch khong 90% b mt neuron.

Hnh 3.1 Cu to mng neural sinh hc Cc tn hiu truyn trong dy thn kinh vo v dy thn kinh ra ca cc neural l tnh hiu in, c thc hin thng qua cc qu trnh phn ng v gii phng cc cht hu c. Cc cht ny c pht ra t cc khp ni dn ti cc dy thn kinh vo s lm tng hay gim in th ca nhn t bo. Khi in th ny t n mt mc ngng no s to ra mt xung in dn ti trc dy thn kinh ra. Xung ny c truyn theo trc, ti cc nhnh r khi chm ti cc khp ni vi cc neuron khc v s gii phng cc cht truyn in. Thng chia khp ni thnh 2 loi: khp ni kch thch (excitatory) v khp ni c ch (inhibitory). Pht hin quan trng nht v b no sinh hc l cc lin kt khp thn kinh kh mm do, c th bin ng v sa i theo thi gian ty thuc vo cc dng kch thch. Hn na, cc neuron c th sn sinh cc lin kt mi vi cc neuron khc; i khi, li cc neuron c th di tr t vng ny sang vng khc trong b no. y l c s quan trng gii thch cho c ch hc ca b no con ngi.

Cc chc nng c bn ca b no bao gm: - B nh c t chc theo cc b thng tin v truy cp theo ni dung. - B no c th tng qut ha, c th truy xut cc tri thc hay cc mi lin kt chung ca cc i tng tng ng vi mt khi nim chung no . - B no c kh nng iu chnh hoc tip tc thc hin ngay khi c nhng sai do thng tin b thiu hay thiu chnh xc. Ngoi ra, b no cn c th pht hin v phc hi cc thng tin b mt da trn s tng t gia cc i tng. - B no c kh nng xung cp v thay th dn. Khi c nhng trc trc ti cc vng no (do chn thng) hoc bt gp nhng thng tin hon ton mi l, b no vn c th c tip tc lm vic. - B no c kh nng hc. Nhn chung, tnh ton s b cho thy rng d b vi x l my tnh in t c th tnh ton nhanh hn hng triu ln so vi neuron ca b no, nhng xt tng th th b no li tnh ton nhanh hn hng t ln. Ngoi ra, cng d thy rng b no con ngi c th lu tr nhiu thng tin hn cc my tnh hin i, d rng iu ny khng phi ng mi mi bi l b no tin ha chm cn b nh my tnh th c nng cp rt nhanh nh nhng tin b ca khoa hc k thut. 1.3 Mng nron nhn to 1.3.1 Khi nim: mng neuron nhn to (Artificial neural network ANN) l mng bao gm cc nt (neuron, n v x l) c ni vi nhau bi cc lin kt neuron. Mi lin kt km theo mt trng s no , c trng cho c tnh kch hot hoc c ch gia cc neuron. C th xem cc trng s l phng tin lu thng tin di hn trong mng neuron, cn nhim v ca qu trnh hun luyn (hc) l cp nht cc trng s khi c thm thng tin v cc mu hc, hay ni cch khc, cc trng s c iu chnh sao cho dng iu vo ra ca n m phng hon ton ph hp mi trng ang xem xt. a. M hnh nhn to

Hnh 3.2 M hnh neural nhn to Mi nueron c ni vi cc neuron khc v nhn c cc tn hiu t chng vi cc trng s wj. - Tng thng tin vo c trng s l: - Net = ?w j s j , y l thnh phn tuyn tnh ca neuron

Hm kch hot g ng vai tr bin i t Nt sang tn hiu u ra out Out = g(Net), y l thnh phn phi tuyn ca mng neuron Mt s dng hm kch hot thng dng trong thc t:

b. Mng neuron nhn to Mng neuron nhn to, sau y gi tt l mng neuron,c xy d ng trn c s mng neuron sinh hc, l h thng bao gm nhi u phn t x l n gin (neuron), hot ng song song. Tnh nng ca h thng ny ty thu c vo c u trc ca h, cc trng s lin kt v cu trc ca chngcho ph hp vi mu hc. trong mng neuron, cc neuron n nhn tn hiu vo g i l neuron vo, cn cc neuron a thng tin ra gi l nueron ra. Cc thng s cu trc mng neuron bao gm: - S tn hiu vo, s tn hiu ra. - S lp neuron - S neuron trn mi lp n - S lng lin kt ca mi neuron (y , b phn, ngu nhin) - Cc trng s lin kt 1.3.2 Phn loi : - Theo kiu lin kt neuron, ta c mng neuron truyn thng (feed-forward neural network) v mng neuron hi qui (recurrent neural network). Trong mng neuron truyn thng, cc lin kt neuron

i theo mt hng nh t nh , khng c chu trnh. Ngc li, mng neuron hi qui cho php cc lin kt neuron t o thnh chu trnh. V cc thng tin ra ca cc neuron c truyn li cho chnh cc neuron nn gp phn kch ho t cho chng v to ra kh nng l u gi trng thi trong ca n di dng cc ngng kch hot ngoi cc trng s lin kt neuron. - Theo s lp, ta c mng neuron mt lp (single -layer) v mng neuron a lp (multi-layer). Trong , thng th ng l p neuron vo ch chu trch nhim truyn a tn hiu vo, khng thc hin mt tnh ton no, nn khi tnh s lp ca mng ta khng tnh lp ny vo. 1.3.3 Cch nhn v mng nron ( Cc phng php hc) : + C th xem mng neuron nh mt cng c ton hc, mt bng tra. Gi s mng neuron NN c m neuron vo v n neuron ra, khi vi mi vector tn hiu vo X= (x1 , x2 ,..., xm ) sau qu trnh tnh ton ti cc neuron n s nhn c kt qu ra Y= ( y1 , y2 ,..., yn ) v ta qui c vit Y = out (X,NN). + Mng neuron nh mt h thng thch nghi, c kh nng hc (hun luyn) tinh chnh cc trng s lin kt cng nh cu trc ca chng sao cho ph hp vi cc mu hc (samples). Thng phn bit 3 k thut hc: + Hc c gim st (supervised learning), cn gi l hc c thy: Mng c cung cp mt tp mu hc {(x,d)} theo ngha x l cc tn hiu vo th kt qu ng ca h phi l d. mi ln hc, vector tn hiu vo x c a vo mng, sau so snh s sai khc gia cc kt qu ra ng d vi kt qu tnh ton Y. Sai s ny c dng hiu chnh li cc trng s lin kt trong mng. qu trnh c tip tc cho n khi tha mn mt tiu chun no . C 2 cch s dng tp mu hc: hoc dng cc mu ln lt ht mu ny n mu khc, hoc s dng ng thi tt c cc mu cng mt lc.

Hnh 3.3 Mo hnh hoc co giam sat + Hc khng c gim st (unsupervised learning), cn gi l hc khng thy: Trong k thu t hc ny, s khng c s hi tip t mi trng cho bi t tn hiu ra yu cu ca mng nn nh th no, hoc chng c ng cha ging nh hc c gim st, m ni chung mng neuron ph i t n pht hin ra bt c mi lin h c lin quan c th tn ti trong gi liu vo (chng hn nh: cc mu, cc c tr ng, cc quy tc, s tng quan) v chuyn mi lin h pht hin ny sang u ra. Mng hc vi c ch ny gi l m ng t t chc. Th d, mng neuron hc khng th y c th cho chng ta bit mt mu u vo mi ng dng nh th no vi mu c trng th y trong qu kh (s ng dng); hoc mt dng neuron khc c th xy dng mt tp nhng ci ru trn c s s tng t ca nhng th d trc (phn tch thnh phn ch yu)v.v

Hnh 3.4 Mo hnh hoc khong co giam sat + Hc tng c ng (Reinforced learning): Nh gii thi u trn, k thut hc c gim st l hiu ch nh dn gi tr tn hiu u ra tng ng vi tng cp mu tn hiu vo-ra. Tuy nhin thc t nhi u khi khng th c c cc thng tin chi tit ny. Do thng phi s dng thut ton hc tng cng. trong hc tng cng , d liu hun luyn rt th v chng ch c lng so snh vi s truyn kin thc hi tip trong hc c gim st.

Hnh 3.5 Mo hnh huan luyen tang cng

+ Cc k thut hc trong mng neuron c th nhm vo vic hiu chnh cc trng s lin kt gi l hc tham s; ho c nhm vo vic i u chnh, sa i cu trc ca mng, bao gm s lp, s neuron, kiu v trng s cc lin kt gi l hc cu trc. 1.4 M hnh mng nron 1.4.1 Mng truyn thng: a. Mng Perceptron n lp

Hnh 3.6 Mang perceptron n lp

- Hm kch hot neuron: a(.) - Mc tiu ca qu trnh hun luyn mng l:

a.1 Qui lut hc Perceptron Tn hiu ramu c dng tuyn tnh ngng, ch nhn gi tr (2.4) ta c:

a.2 Adaline (Adaptive linear Element)

Tnh hiu ra c dng tuyn tnh dc, t (3.4) ta c:

1.4.1.2 Mng truyn thng n lp 1.4.1.3 Mng truyn thng a lp a. Lan truuyn ngc (Back propagation) Thut ton hun luyn lan truyn ngc c mt ngha quan trng trong lch s pht trin mng neuron. Cc mng neuron c hun luyn c hun luyn bng thut ton ny gi l mng lan truyn ngc. vi cc tp hc k=1..p, thut ton lan truyn ngc a ra cc th tc thay i trng s trong mng lan truyn ngc. C s ca vic cp nht cc trng s ny l c ch suy gim gradient. Vi mt cp tnh hiu mu vo-ra tin, mu thu t ton lan truyn ngc thc hi n 2 giai on. u . Sai s gi a

c lan truyn t lp vo n lp ra v cho ra tn hiu tht lp ra l

tn hiu tht lp ra v tn hiu ra mu c lan truyn ngc tr li t lp ra n nhng lp trc cp nht li trng s cho chng. Xt c th mt mng neuron 3 l p lan truyn ngc (hnh 3.7) minh ha thut ton lan truyn ngc v kt qu ny hon ton c th m rng cho nhng mng c s lp nhiu hn.

Hnh 3.7 Mang neural 3 lp lan truyen ngc

Mng neuron 3 lp lan truyn ngc hnh 3.7 c m neuron lp vo, l neuron lp n v n neuron lp ra. u tin, xt cp mu hun luyn (x,d). vi tn hiu mu u vo x lp vo, neuron q lp n s nhn c tn hiu:

Tn hiu vo ca neuron I lp ra: b.Cc h s hc trong thut ton lan truyn ngc - Khi to gi tr trng s lin kt neuron (Intial weights): Trong mng truyn thng, gi tr trng s c khi to ban u c nh hng quan trng n kt qu cui cng. Chng c gn ngu nhin vi gi tr tng i nh, v nu qu ln th hm sigmoid s d bo ha ngay lc bt u, dn n h thng s b nghn ti gi tr a phng nh nht hoc rt n nh ln cn im bt u. Do ,tm hp l ca gi tr trng s khi to ban u thng nm trong khong [-3/ki, 3/ki], trong ki l s u vo ni vo neuron I [Wessels and Barnard, 1992] - Hng s hc (learning constant): Mt h s quan trng khc cng lm nh hng n hiu qu v s hi t ca thut ton lan truyn ngc, l hng s hc . S khng c mt gi tr hng s hhc c nh cho cc trng hp hun luyn khc nhau, m thng chng c chn th nghim cho t ng bi ton c th. Mt hng s hc c gi tr ln c th lm gia tng tc hi t, nhng kt qu cng c th s b c ng i u; Trong khi mt hng s hc c gi tr nh h n th c tc dng ngc li. tm gi tr hng s hc thng dao ng trong khong t 10-3 n 10. Mt vn khc cng c t ra l hng s hng s c tt nht lc bt u hun luyn, tuy nhin s khng cn tt na sau vi ln hun luyn. Do , tt nht l dng hng s hc thch nghi. Phng php trc gic xc nh hng s hc ny l kim sot ring l qu trnh cp nht trng s lm gim hm sai s; nu khng th gim dn chng nu kt qu cng iu; hoc trong tr ng hp khi nhiu bc lp u c s suy gim hm sai s dn n bi ton qu hi t, th nn tng dn hng s

hc ln. C th nht, hng s hc nn c cp nht theo cc quy lut sau 1.4.2 Mng hi quy 1.4.2.1 Mng hi tip n lp a. Mng Hopfield ri rc b. Mng Hopfield vi b nh t kt hp c. Mng Hopfield vi b nh 2 chiu kt hp 1.4.2.2 Mng hi quy lan truyn ngc

Chng 2:

K Thut Vn Tay

2.1 Gii thiu chng 2.2 Lch s ca vn tay 2.3 H thng nhn dng v xc thc vn tay 2.3.1 H thng nhn dng 2.3.2 H thng xc thc 2.4 Cc li ca h thng kim tra v nhn dng Chng 3: Nhn Dng Du Vn Tay 3.1 Gii thiu chng : 3.2 Phn tch cu trc vn tay 3.3 Cc im c trng ca nh vn tay 3.3.1 Cc im Singurality: Trn cc nh vn tay c cc im c trng (l nhng im c bit m v tr ca n khng trng lp trn cc vn tay khc nhau) c phn thnh hai loi: singularity v minutiae Singularity: Trn vn tay c nhng vng c cu trc khc thng so vi nhng vng bnh thng khc (thng c cu trc song song), nhng vng nh vy goi l singularity. C hai loi singularity l core v delta.

Hnh 2.1: Cc im singularity core v delta Core thng c mt s dng nh sau:

Hnh 2.2: Mt s loi core thng gp. 3.3.2 Cc im Minutae Khi d theo tng ng vn ta s thy c nhng im ng vn kt thc (Ridge Ending) hoc r nhnh (Bifurcation), nhng im ny c gi chung l minutiaae.

2.3: Cc im minutiae Ridge Ending (im kt thc) v Bifurcation (im r nhnh) 3.4 Biu din hnh nh vn tay 3.5 c lng hng vn cc b 3.6 Tng cng nh 3.7 Lm ni nh vn tay: Cc nh vn tay thng c ly bng hai phng php: t mc hoc t cc sensor. Cc nh vn tay c ly t mc thng c cht lng thp v khng ng u. Phn ny s gii thiu phng php dng b lc Gabor ci thin cht lng ca nh vn tay [8], [13], [14]. Hm Gabor l mt cng c hu dng cho vic x l nh. N c c tnh chn lc trong min khng gian ln tn s. Hm Gabor 2_D thc c dng nh sau:

trong : l hng ca b lc T l chu k ca hm cos (thng c chn t thc nghim c gi tr [0,1]) x , y l cc lch chun (thng c chn t thc nghim c gi tr [0,4]) Cc bc thc hin: 1. Chun ha mc xm: nu I(x,y) l mc xm ti im (x,y) ca nh I th mc xm chun ha Ni(x,y) c xc nh rheo cng thc sau:

trong : M0, V0 l mean v variance mong mun (thng c chn l 100) Mi, Vi l mean v variance ca nh I Ch : nu mc xm ca cc vng khc nhau trn nh I khng ng u th c th chia I thnh cc khi nh v chun ho theo tng khi.

Hnh 2.12: nh I v nh chun ha ca n (Hm normalize.m thc hin chun ha mc xm c gii thiu ph lc) 2. Xc nh trng nh hng theo phng php gii thiu trn 3. S dng hm lc Gabor cho nh chun ha trong min tn s Chia nh cn lc thnh tng khi nh kch thc WxW Xc nh hng ca khi (da vo trng nh hng) Hng ca b lc l hng ca khi S dng php bin i FFT v php bin i IFFT cho tng khi nh v hm Gabor

nh 2.13: Kt qu lc bng hm gabor_filter.m (ph lc) vi T = 0.6, x = 1, y = 2 3.8 Trch cc im c trng Bng cc phng php x l nh ta c th tm c v tr cc im c trng trn cc nh vn tay. 2.2.1 Trch cc im singularity

a. Trng nh hng (orientation field) nh vn tay l nh nh hng, cc ng vn l cc ng cong theo cc hng xc nh. Gc hp bi phng ca mt im trn ng vn vi phng ngang c gi l hng ca im . Tp hp cc hng ca cc im trn nh vn tay gi l trng nh hng ca nh vn tay .

Hnh 2.4: nh vn tay (a) v trng nh hng ca n (b) Phng php xc nh trng nh hng nh sau [5], [14]: Chia nh vn tay thnh cc khi nh hn kch thc WxW. Tnh gradient theo hai hng x, y l Gx, Gy ti mi im (pixel) trong khi Khi hng ca im chnh gia ca khi c xc nh theo cng thc:

Hm orientation.m thc hin tnh trng nh hng c gii thiu trong phn ph lc. b. Xc nh cc im singularity bng ch s Poincare (Poincare index) [3] Gi s (i,j) l mt im bt k trn nh vn tay, C l mt ng cong khp knh xung quanh (i,j) th ch s Poincare ti (i,j) l tng i s cc sai lch hng ca cc im lin k nhau trn ng cong C.

Trong : Np l tng s im trn ng cong s C (x,y) l hng ti im (x,y) Da vo ch s Poincare ta c th xc nh cc im singularity nh sau:

Hnh 2.5 minh ha cch tnh ch s poincare ti im (i,j) vi s im trn ng cong s Np =8

Hnh 2.5: Cch tnh ch s poincare ti im (i,j) vi Np = 8 Hm poincare.m thc hin vic tnh ch s Poincare theo thut ton trn v hm singularity.m xc nh cc im singularity da vo ch s Poincare. 2.2.2. Trch cc im minutiae C hai phng php chnh tm cc im minutiae: trch cc im minutiae t nh binary v trch cc im minutiae trc tip t nh xm. a. Trch cc im minutiae t nh binary [5]

Hnh 2.6: S m t thut ton trch cc im minutiae t nh binary

tng chnh ca phng php ny l t nh xm ban u ta s dng cc b lc thch hp pht hin v lm mnh ng vn di dng mt pixel (ridge detection), bin i nh xm ban u thnh nh binary (c gi tr l 0 hoc 1) tng ng. Sau , cc im minutiae s c trch nh sau: gi s (x,y) l mt im trn ng vn c lm mnh v N0, N1, , N7 l 8 im xung quanh n th

Hnh 2.7 cc kt qu ca thut ton D theo ng vn (Ridge line following) Gi s I l mt nh xm c kch thc l mxn v nu coi chiu th ba z l mc xm ti im (i,j) th b mt ca nh vn tay I c dng nh sau:

Hnh 2.8: B mt ca nh vn tay vi cc ng vn (ridge) v cc rnh (ravine) Theo quan im ton hc th ng vn l tp hp cc im cc i dc theo mt hng xc nh. Vic xc nh cc im minutiae trc tip t nh xm da vo thut ton d theo ng vn. Thut ton ny da vo vic xc nh cc im cc i dc theo hng ca ng vn. Xc nh im cc i

v im cc i c th c xc nh bng cch so snh mc xm gia cc im trong

Hnh 2.9: Thit din ca ng vn ti Tm li vic tm cc im minutiae bng thut ton d theo ng vn c thc hin nh sau (chi tit xem ti liu tham kho[1]): Ly mt im bt k (is,js) trn nh I Tm hng s ti im (is,js) Tm im cc i (ic,jc) gn (is,js) nht

Hnh 2.10: im cc i (ic,jc) tng ng vi (is,js) Tm hng c ti im (ic,jc) Dch chuyn theo hng c mt on Tinh chnh li im cc i (ic,jc) v hng c Tip tc qu trnh ny d theo ng vn (ridge following) cho n khi khng pht hin c im cc i (ic,jc) th l im Ridge Ending hoc chm vo mt ng vn khc th l im Bifurcation (mi ng vn sau khi c d s c gn nhn) Tip theo chn mt im (is,js) khc v thc hin li qu trnh trn cho n khi d ht tt c cc ng vn.

Hnh 2.11: Dch chuyn theo ng vn tng on 3.9 i snh : Hu ht cc phng php nhn dng vn tay u da vo vic i snh v tr cc im c trng. Gn y, mt s tc gi kt hp thm mt s c tnh khc ca nh vn tay nng cao hiu qu i snh nh: Orientation field [9] hoc Density map [10]. Chi tit xem ti liu tham kho, y ti xin gii thiu phng php i snh v tr cc im c trng m ti s dng, phng php ny gn ging vi cc phng php c nu [4] v [11]. Hm matching.m (ph lc) thc hin i snh hai nh vn tay theo phng php ny. Gi s I v I ln lt l cc nh vn tay mu v nh vn tay cn trng c xc nh bi ta (x,y) v hng . m c

trong : m, n ln lt l s im c trng ca I v I. K hi , im 1 c coi l ging vi im mI n u sai l c h v

khng gian v sai lch v hng nh hn cc gi tr ngng r 0 v 0 :

Nu

th I c coi l ging I. Trong T l phn trm s im sai lch cho php.