ISSN 1000-9825, CODEN RUUEW E-mal jos@scasaccn Journal of Sofware, Vol17, No12, December 2006, pp2529 2536 hp//wwwjosorgcn DOI 101360/jos172529 Tel/Fax +86-10-62562563 2006 by Journal of Sofware All rghs reserved 1+, 2, 1 1, 1 (, 710049 2 (, 116026 Super-Resoluon Reconsrucon for Face Images Based on Parcle Flers Mehod HUANG Hua 1+, FAN n 2, QI Chun 1, ZHU Sh-Hua 1 1 (School of Elecroncs and Informaon Engneerng, an Jaoong Unversy, an 710049, Chna 2 (School of Informaon Engneerng, Dalan Marme Unversy, Dalan 116026, Chna + Correspondng auhor Phn +86-29-82668772, Fax +86-29-83237910, E-mal huanghua@xjueducn Huang H, Fan, Q C, Zhu SH Super-Resoluon reconsrucon for face mages based on parcle flers mehod Journal of Sofware, 2006,17(122529 2536 hp//wwwjosorgcn/1000-9825/17/2529hm Absrac Super-Resoluon (SR reconsrucon s posed as a Bayesan esmaon of he locaon and appearance parameers of a face model Image regsraon and mage fuson, he wo seps for SR reconsrucon, are combned no one unfed probablsc framewor, n whch he pror nformaon abou facal appearance and gray from he face model s ncorporaed no boh of he seps In addon, a parcle fler based algorhm s proposed o acheve he esmaon, e SR reconsrucon The proposed approach avods he nheren dlemma of he mos radonal mehods, n whch demands a hgh-resoluon mage o ge an accurae and robus esmaon of he moon feld, whle reconsrucng a hgh-resoluon mage requres he accurae and robus esmaon of moon feld Expermens performed on synheszed fronal face sequences show ha he proposed approach gans superor performance boh n regsraon and reconsrucon Key words super-resoluon reconsrucon; face mage; parcle fler,,,,,,,,, ; TN391 A Suppored by he Naonal Hgh-Tech Research and Developmen Plan of Chna under Gran No2004AA1Z2312 ( (863 Receved 2006-04-21; Acceped 2006-08-17
2530 Journal of Sofware Vol17, No12, December 2006 (super resoluon, SR (low resoluon, LR (hgh resoluon, HR,, [1,2] [1] LR (,,,, [3 5],,, [6],, LK [7],,, ;, [1],, 4 (,,, 1, ; ;,,, 1, 4 4 (=1,,4 4 (, =(l,a, l, l=( x, y,s; a PCA, T a T =T 0 +Φ a (1,T 0,Φ PCA PCA [8] 1 4 E,j,, ψ(, j =ψ l (l,l j ψ a (a,a j (2 j ; ψ l, ; ψ a, [9]
2531 1 2 E 1,3 E 2,3 3 E 3,4 4 Fg1 The graphcal srucure of par-based face model 1, a (1, l,, LR,SR 2, Y =(Y 1,,Y 4 =( 1, 2, 3, 4, Y Y ( Y, Y Y Y (3,Y,,,Y p ( Y = Y (4 (3,,,, [10,11], ψ (5 ( (, j, j E, (5 2 (3, N (3 (3, 3 [12], N s, w } = 1 N s {, (3 Y,, (3 p (6 ( Y Y w 1, ( (7 ~ p, w w = w Y (8 (7 (8,
2532 Journal of Sofware Vol17, No12, December 2006 N { s, w } = 1 p, Y (5 (6, (3 p ( Y Y ψ (, j w ( (9, j E (9 (6 (7 p ( ( (, { ~ ( ( N s } = 1 p (10, (, ( =l,a l (,a ( (11 l,a l (,a ( =a l,l (,a ( l l (,a ( (12 l l (, a (, l l (,a ( =l l ( (13,, l l ( N s } = 1 { l l ( a ( l, a a l,l (,a ( =N(a (,Q ( (14, Q (, (l l (, ;,, 10~20,,, (14, a0 ~ N( a0, P 0, P T = ( Φ Φ + ( Q( 1 + P ( (15 T a P Φ Y + ( Q P a (16 = ( ( 1 + ( ( (9 (6, (8 w = w 1 Y (, j = w Y ψ l ( l, l j ψ a ( a, a j, j E, j E ψ (17, ψ ( l, l ψ ( a, a Y l j a j 1 2 Y exp T T 0 Φ a (18 2 Σ, T Y f, f, x 2 Σ T = x Σ,Σ, x (13 (15~ (17 N { s, w } = 1 LR l (19 Y, l a a (1, l,
2533 4 AR 143, 286 4, 4 T 0,Φ I,Σ 10, 1 l=( x, y,s, p=[l x,l y ] T p = [ l, ] T f( x l y p =f(p=s [l x,l y ] T +[ x, x ] T (20,,,,,,,,, { l0} = 1 ; T, Φ T T 0 N a, P a, P 0 ( 0 0 0 Σ, (13 (15~ (17, HR LK 2 05, LK 2,,LK ;,, LK, [13],,LK,LK,,,,,LK ;, (14,, LK 3, 18,, LK,,LK ;,,,, LK Mean error 03 02 01 0 Proposed approach LK algorhm 1 5 10 15 20 Translaon Mean error 06 04 02 0 Proposed approach ( N s LK algorhm 1 12 14 16 18 2 Fg2 The mean esmaon errors obaned by Fg3 The effecs of scale varaons on he proposed approach and LK algorhm Scale esmaon accuracy 2 3
2534 Journal of Sofware Vol17, No12, December 2006 4 4 LR 4(b LR 3, 4(c,,,, 4(d 4(c,,,, ;, [7] (a The orgnal mages (a (b The HR mage obaned by b-cubc nerpolaon (b 3 (c The reconsruced HR mages of he facal componens esmaed by he proposed approach (c (d The reconsruced HR mages esmaed by he proposed approach (d Fg4 The resuls of reconsrucng LR face mages 4
2535,,,,, [14], 5,,,,,,,,,,,,,,,, LK Malab, References [1] Baer S, Kanade T Lms on super-resoluon and how o brea hem IEEE Trans on Paern Analyss and Machne Inellgence, 2002,29(91167183 [2] Sung CP, Mn KP, Moon GK Super-Resoluon A echncal overvew IEEE Sgnal Processng Magazne, 2003,20(521 36 [3] Capel D, Zsserman A Super-Resoluon from mulple vews usng learn mage models In Proc of he IEEE Conf Compuer Vson and Paern Recognon Hawa IEEE CS Press, 2001 627 634 [4] Lu C, Shum HY, Zhang C A wo-sep approach o hallucnang faces Global paramerc model and local nonparamerc model In Proc of he IEEE Conf on Compuer Vson and Paern Recognon Hawa IEEE CS Press, 2001 19298 [5] Wang, Tang Hallucnang face by egenransformaon IEEE Trans on Sysems, Man and Cybernecs, Par C, 2005,35(3 425 434 [6] Robnson D, Mlanfar P Fundamenal performance lms n mage regsraon IEEE Trans on Image Processng, 2004,13(9 1185199 [7] Lucas BD, Kanade T An erave mage regsraon echnque wh an applcaon o sereo vson In Proc of he Imagng Undersandng Worshop San Francsco Morgan Kaufmann Publshers, 1981 12130 [8] Moghaddam B, Penland A Probablsc vsual learnng for objec represenaon IEEE Trans on Paern Analyss and Machne Inellgence, 1997,19(7696 710 [9] Ban ZQ, Zhang G Paern Recognon Bejng Tsnghua Unversy Press, 2000 (n Chnese [10] Chang C, Ansar R, Khohar A Cyclc arculaed human moon racng by sequenal ancesral smulaon In Proc of he 2004 IEEE Compuer Socey Conf on Compuer Vson and Paern Recognon Los Alamos IEEE CS Press, 2004 45 52 [11] Khan Z, Balch T, Dellaer F MCMC-Based parcle flerng for racng a varable number of neracng arges IEEE Trans on Paern Analyss and Machne Inellgence, 2005,27(111805818 [12] Arulampalam S, Masell S, Gordon NJ, Clapp T A uoral on parcle flers for on-lne non-lnear/non-gaussan bayesan racng IEEE Trans on Sgnal Processng, 2002,50(217488 [13] Baer S, Mahews I Lucas-Kanade 20 years on A unfyng framewor In l Journal of Compuer Vson, 2004,56(3221 255 [14] Freeman WT, Paszor EC, Carmchael OT Learnng low-level vson In l Journal of Compuer Vson, 2000,20(125 47
2536 Journal of Sofware Vol17, No12, December 2006 [9],,2000 (1975,,,,, (1955,,,,,, (1977,,,, (1950,,,, 2 Web (SWON 2007 Web SWON 2007 2007 9 14~16 Web Web Web Web Web Web Web hp//wwwruceducn/wsa2007/ hp//wwwneueducn/wsa2007 1 2007 4 1 2 2007 4 20 3 2007 5 10