7 2004 7 ACTA ELECTRONICA SINICA Vol. 32 No. 7 July 2004 1,2,Ngo Cong2Wa 3, 1,2 (11, 100871 ;21, 100871 ; 31, ) :..,,. :,.,Kun2Munkre,.,.. : ; ; Kun2Munkre ; : TP301 : A : 037222112 (2004) 0721135205 An Appro ac for Sot Retrieal by Optimal Matcing in te Bipartite Grap PENG Yu2xin 1,2,NGO Cong2Wa 3,XIAO Jian2guo 1,2 (11 Intitute of Computer Science and Tecnology, Peking Unierity, Beijing 100871, Cina ; 21 National Key Laboratory of Text Proceing Tecnology, Beijing 100871, Cina ; 31 Dept. of Computer Science, City Unierity of Hong Kong, Kowloon, Hong Kong, Cina) Abtract : Sot retrieal play a critical role in content2baed ideo retrieal. Motiated by te teory of optimal matcing in bi2 partite grap,we propoe a noel approac baed on te Kun2Munkre algoritm for ot retrieal. In contrat to exiting algoritm, te propoed approac empaize one2to2one mapping among frame between two ot for effectie imilarity meaure. A weigted bi2 partite grap i contructed to model te imilarity between two ot :eery ertex in a bipartite grap repreent one frame in a ot, and te weigt of eery edge repreent te imilarity alue for a pair of frame between two ot. Ten Kun2Munkre algoritm i employed to compute te maximum weigt of a contructed bipartite grap a te imilarity alue between two ot by guaranteeing te one2to2one mapping among frame. To improe te peed efficiency,we alo propoe two improed algoritm. Experimental reult in2 dicate tat te propoed approac aciee uperior performance tan ome exiting metod. Key word : Content2baed ot retrieal ;optimal matcing ; Kun2Munkre algoritm ;improed algoritm 1,,.,., 90,,MPEG27.,,,,. [1 ], ;,.,. [2,3 ],,[ 4 ], (nearet feature line, NFL). :,. [5, 6 ] :2003201228 ; :2003211205
1136 2004,(ubot),. [5 ],,,, (tatic), (pan,tilt) (moaic,panoramic image), (zoom), (Sim ( i, j ) = 1 2 { M ( i, j ) + ^M ( i, j ) }, ^M ( i, j ) i j ). [6 ],, (Sim( a, b) = max (Sim( a i, b j ) ),Sim( a i, b j ) a i, j i b j ). [5,6 ],,., [7 ],, (dy2 namic programming),. :,.,,.,. :,. Kun2Munkre,. [5 ] HSV :,,,. [5 ],, [5 ],. [5,6 ], [7 ]., [7 ] (dynamic programming),,. [7 ],,., :. 3. 2 211 Kun2Munkre, :n x 1, x 2,, x n m y 1, y 2,, y m, 1,e ij = ( x i, y j ) x i y j, e ij ij x i y j,,. :? 1, :X n x 1, x 2,, x n, Y m y 1, y 2,, y m,e ij = ( x i, y j ) x i y j, ij x i y j, 1. x i y j,:? X Y,,,.,,.,,. G = { X, Y, E} (,V = X Y, E = { e ij } ), x i y j ij e ij. ij. ij = Interec t ( x i, y j ) = 1 A ( x i, y j ) A ( x i, y j ) = min min{ H i (,, ), H j (,, ) } (1) H i (,, ), H j (,, ) H i (,, ), H j (,, ) HSV, H,S,V 18 3 3, 162. Interect ( x i, y j ),, A ( x i, y j ) Interect ( x i, y j ) 0,1. Kun2Munkre. X Y ij G = { X, Y, E} e ij,g ( G m n ). Kun2Munkre [8 ] : (1) l ( x i ) = max ij, l ( y j ) = 0, j i, j = 1,2,, t, t = max( n, m) ; (2) E l = { ( x i, y j ) l ( x i ) + l ( y j ) = ij } G l = ( X, Y k, E l ) G l M ( M Α E, M ) ; (2)
7 : 1137 (3) M X, M G, ; (4) X M x 0,A ω{ x 0 }, B ω <, A, B ; (5) N Gl ( A) = B,(9),, N Gl ( A) Α Y k A ; (6) y N Gl ( A) - B ; (7) y M, y z, A ω A { z}, B ω B { y} (5) ; (8) x 0 y P, M ω M g E ( P), (3) ; : (9) a : a = min { l ( x i ) + l ( y j ) - ij }, x A i y N (A) j Gl l ( ) = l ( ) - a, A l ( ) + a, B l ( ), l E l G l ; (10) l ω l, G l ω G l, (6). Kun2Munkre O ( p 4 ), p = n + m. G = { X, Y, E} M,M e ij ij, G = { X, Y, E}. X Y Similarity ( X, Y) = min( n, m). min( n, m) Similarity( X, Y) 0,1,, X Y. 212 Kun2Munkre,,, Kun2Munkre.,.. : (1).,.,,. [5 ]. 1 : 1 Motion Type Static Pan or tilt Zoom Multiple motion Indeterminitic Action Select one frame Form a new panoramic image Select firt and lat frame Recontruct background Select one frame 1,,, Kun2 Munkre. : (, ), Kun2Munkre,.,,, (2). (2).,,. Kun2Munkre, :,(1).,. 3, [5 ],,. 3 2002,41,777,62132.,. 7,, 2. 2 7 4 : (1) : ; (2) ; (3) [5 ]; (4). 4, HSV 162,,,. 3,, 1,,,. 3 (022430. bmp ), (1),.,.
1138 2004 MPEG27 : ANMRR ( Aerage Normalized Modified Retrieal Rank) AR(Aerage Recall). AR ( Recall ), ANMRR ( Preciion ) ( ). ANMRR ;AR, K( K 2 (1) (2) (3) (4) AR ANMRR AR ANMRR AR ANMRR AR ANMRR 11 0. 9412 0. 1441 0. 9412 0. 1059 0. 9412 0. 1343 0. 8824 0. 1882 2. 0. 8556 0. 2356 0. 7333 0. 3775 0. 7333 0. 3775 0. 7889 0. 2873 3. 0. 9400 0. 1597 0. 9600 0. 1520 0. 9400 0. 1522 0. 9200 0. 2223 4. 0. 7750 0. 3157 0. 7750 0. 3731 0. 7250 0. 3963 0. 8500 0. 2746 5. 0. 8125 0. 2102 0. 8125 0. 2566 0. 8125 0. 2677 0. 7500 0. 2976 6. 0. 8214 0. 1871 0. 7857 0. 1929 0. 8214 0. 2049 0. 7500 0. 2788 7. 0. 8421 0. 2503 0. 8421 0. 2503 0. 8421 0. 2293 0. 7632 0. 2783 0. 8554 0. 2147 0. 8357 0. 2440 0. 8308 0. 2517 0. 8149 0. 2610 ). 2 4 7 AR ANMRR. 2,AR, ANM2 RR,,. (2),,,Kun2Munkre,. (1) 3 (2), (1) AR ANMRR (2).,PIII CPU 1G,256M PC, (1) 1514, (2) 916, 777,.,,, 3. 4. : (1). :,, Kun2 Munkre,.,. (2),.., HSV, :,.,,. [ 6] ANMRR (Aerage Normalized Modified Retrieal Rank) AR (Aerage Recall). Q,, ( Ground Trut). q ng (q). q,k min{ 4 ng ( q),2 GTM},, GTM
7 : 1139,GTM = max{ ng ( q) }. q, K nr( q),m ( q) = ng ( q) - nr( q). R ( q) = nr( q) / ng ( q). (rank) r ( i), i = 1,, ng ( q). K r ( i), r( i) K + 1. q, ng ( q) r( i) ARR( q) = ng ( q),mrr( q) = ARR( q) - ng ( q) - 015 2 i =1 MRR( q) [0,1 ], NMRR( q) : NMRR( q) = MRR( q) ng ( q) K - + 015 2 Q q NMRR( q) R ( q), ANMRR AR : : ANMRR = 1 Q Q NMRR( q),ar = 1 q =1 Q Q R ( q) q =1 [ 1 ] Ngo C W,Pong T C,Cin R T. Video partitioning by temporal lice co2 erency [ J ]. IEEE Tranaction on Circuit and Sytem for Video Tecnology,2001,11 (8) :941-953. [ 2 ] Liu X M,Zuang Y T,Pan Y H. A new approac to retriee ideo by example ideo clip[a]. Proceeding of ACM Multimedia[ C]. Orlando : ACM,1999. 41-44. [ 3 ] Wu Y, Zuang Y T, Pan Y H. Content2baed ideo imilarity model [ EB/ OL ]. ttp :/ / www. acm. org/ ig/ igmm/ MM2000/ ep/ wu/ wu. pdf. [ 4 ] Zao L,Qi W,Li S Z,et al. Key2Frame extraction and ot retrieal u2 ing nearet feature line (NFL) [ A ]. Proceeding of te International Workop on Multimedia Information Retrieal, in Conjunction wit ACM Multimedia Conference 2000[ C]. Lo Angele :ACM Multimedia conference 2000. [ 5 ] Ngo C W,Pong T C,Zang H J. Motion2baed ideo repreentation for cene cange detection[j ]. Internal Journal of Computer Viion,2002, 50 (2) :127-143. [ 6 ],,,. [J ].,2002,13 (8) :1577-1585. [ 7 ] Cen L P, Cua T S. A matc and tiling approac to content2baed ideo retrieal [ A ]. Proceeding of IEEE International Conference on Multimedia and Expo[ C]. Tokyo :IEEE,2001. 417-420. [ 8 ]. [M]. :,1993. 134-142. :,1974,2003 6,,,. Email :peng - yuxin @ict. pku. edu. cn. Ngo Cong2Wa(),1971,,2000 8,,,.