NGI-lir-N C:i/U - l/NG DUNG ffng dung mang Neural truyan thang viec nhan dang mat ngtfoii ThS PHAM THANH DANH, ThS PHAM HQU Lpc Trudng Dai hoc Cdng nghiep TP Hd Chi Minh TOM TAT De nhan dang mot doi tuong, ta co nhieu phuong phap khac nhau, n h u DCT, Gabor Wavelets, PCA, ICA, Fisherface, mang neural Sudiang mang neural viec nhan dang da dupc ung dung nhieu linh vuc Ket qua dat dupc phu thupc rat nhieu vao qua trinh thu thap d\x Ueu, cung nhu kinh nghiem ciia nguoi su dung qua trinh huan luyen Trong tuong lai, viec iing diing ket qua nhan dang vao dieu khien la rat can thiet Bai bao muon gioi thieu phuong phap trich dac trung doi tupng bang PCA va nhan dang dtiing mang nexual truyen thang Day la mang tuong doi don gian cac mang neurad, nhung neu van dung thich hpp se cd ket qua kha tot Viec su dung mang neural voi tap du Ueu thich hpp se cho ket qua tuong doi chinh xac va nhanh chong I Dat van de Viec nghien cifu linh vtfc nhin dang ddi tucmg (mat ngUcii) tren the gidi da bat dau tif kha lau, nhung ket qua dat duoc van edn rat ban ehe Bii toin nhan dang mat ngifdi cd the ip dung rdng rai nhieu Ung dung thuc te khac Dd chinh li ly mi bii toan bap din rat nhieu nhdm nghien eiiu thdi gian dai Cic ting dung lien quan den nhan dang mat ngudi cd thci' ke nhu: He thdng phit hien tdi pham; He thdng theo ddi nhan su mdt don vi Vi vay, viec nhan dang mat ngueii ludn dat eho moi ngudi hucing tdi mpt giai phap kha thi vi hieu qua nhit ILCdsdlythuyit Mang neural truyen thang (Feedforward) Cau tnic mpt mang lejp dupc biei'u dien nhu hinh dudi: al= tansig(fWl,l *pl-F bl) a2 = purelin(LW2,l •-al -^ b2) Input f—y Output Layer Hidden Layer / ^ r~ ^-Hi^ ->[jzy / J s^ Hinh 1: Mo hinh mang hai Icfp Mang thucmg ed mpt hoac nhieu leip an chiia Website: www.tapchicongnghlep.vn neural sigmoid va mot Idp xuit tuyen tinh Mang da Idp neural vdi cac ham truycin phi tuyen cd nang khio sat bit cii mdi quan he tuyen tinh hay phi en nio giUa cic veeto vao va Ldp xuat li tuyen tinh thi gia tri tao d ngd nam giua -1 v-a Neu mudn ngd giiia va thi ldp xuat nen dung mdt him sigmoid (vi du him logsig) Trong mang nhieu ldp ta thudng dung sd ciia cic ldp de qui dinh sd eho he sd weight Mang niy cd the dUOc sii dung nhU mot ham xap xi tdng quat, nd cd the xip xi bat cif ham nao cd sd lin gian doan hifu han Luat hpc Widrow-Hoff (thuat toan lan truyen ngupc) Budc 1: Khdi tao gia tri cho ma tran sd W vi ddi b Budc 2: Ngd vao d budc thti i dupc tinh nhu sau: - X'ector ngd ra: Oi = F(P, Wi, bi) - Sai sd: Ei = 01 - Ti Budc 3: Hieu chinh sd AWi va dp ddi Abi theo edng thiic: W:i-^l \V| + AWj vdi AVV| = Lr.Ej.pT b; -I- Ab; vdi Ab, = Lr.Ei 3i+l Budc 4: Lap lai tif budc cho den dat duoc sai sd thich hpp Trong dd: - P li vector ngd v-ao - Oi la vector ngd thii i Khoa hpc & Cong ngh# [10/2009] i'l'I'U'H:!!'!! 45 NGMlf-N Cl/U - l/NG DUNG - Tj la vector dich tuong iing cho ngd i - Lj li ti le hoc hdi - F li ham truyen Xay dung tap dit lieu huan luyen (cac Eigenface) dung phuong phap PCA Thuit loan niy- gdm cic bUdc: - Mdt ma trin inh l(x,y) hai chieu NxN dupe chuyen thinh mpt vector cot Fi ed chieu dai N-^ Nhu vay, mdt tap anh mau gdm M inh mat ngUdi cd kich thudc NxN duoe chuyci'n thinh mdt day cac vector F,, F,, F i' tuong img mo ta dupc nhieu dac trung rieng ctia anh mat hon Cac eigenfaces cd tri rieng nhd cd the bd qua, vi chimg chi the hien mpt phan nhd cac dac trung ciia anh mat Trong luan van nay, cac eigenfaces cd tri rieng nhd thi dupc dua ve gia tri bing - M eigenfaces ui dupe lUu gid, dd li cac dac trung dupc dua qua mang neural de huan luyen nhan dang mat Mdt vi du ve eigenfaces dupe the hien nhU sau: Tinh vector inh mat 'trung binh' bang edng thtic: Sr Y M~ (1) - Tinh loan su sai khac eiia mdi vector inh mat tip dU lieu so vdi anh trung binh V luu vio *i Nhu vay * li mot ma Iran cd kich thudc N2xm f&i = * i - W (2) - Tinh ma tran liiep phuong sai C ciia i Ma tran C ed kich thudc li N'^xN, vi C ^ f f i i f t ' i i (3) K - Tinh vector rieng (eigenfaces) va tri rieng ciia ma tran hiep phUOng sai, phan tinh toan khdng trinh bay d diy-, ta sti dung him cd sin thu vien toan cua Matlab Tuy nhien, ma tran C cd kich thudc li N2xN2 tbi se cd N2 vector rieng vi tri rieng, vi du ddi vcfi anh mat cd kich thude 2.56x256 thi ta phai tinh ma trin 65536x65536 va 65536 eigenfaces Cieh tinh niy khdng hieu qui va hau het cae eigenfaces niy khdng ed lpi cho giai thuat nhan dang Do dd, theo nghien ciiu eiia Turk va Pentland (1991), thay vi tinli vector rieng va tri rieng ciia ma tran C, thi tinh cho ma tran L nhu sau: , M i,