saguar1

YANAI Lab.

電気通信大学 総合情報学科/大学院 総合情報学専攻 メディア情報学コース 柳井研究室
電気通信大学 > 情報工学科 > コンピュータ学講座 > 柳井研究室 > 研究紹介  

$B0lHLJ*BNG'<1$K$*$1$k0LCV>pJs$NMxMQ$K4X$9$k8&5f(B

2010$BG/(B $BH,=E3_7CB@(B

1 $B$O$8$a(B$B$K(B

1.1 $BGX7J(B

$B:#F|!$%G%8%?%k%+%a%i$NIa5Z$K$h$j!$%G%8%?%k$B$O(B WWW (WorldWideWeb)$B>e$KBgNL$KB8:_$9$k$h$&$K$J$C$?!%BgNL$NpJs$NIa5Z$K$b$+$+$o$i$:!$$=$l$i$r<+F0$G@0M}!&J,N`$7!$%f!<%6!<$NJ$/$3$H$OL$$@:$Fq$J2]Bj$G$"$jB3$1$F$$$k!%e$r?^$k$K$"$?$j!$pJs$r$$$+$K8zN(E*$KAH$_9g$o$;$k$+$,5a$a$i$l$k(B$B!%(B

1.2 $BL\E*(B

$BK\8&5f$NL\E*$O!$2hA|G'<1$K$*$1$k0LCV>pJs$NM-8z@-$,9b$$G'<1%+%F%4%j$H!$$=$&$G$J$$%+%F%4%j$rL@3N$K6hJL$9$k$3$H$K$"$k!%u67$rI=$9>pJs8;$H$7$F!$9R6upJs$rMQ$$$k!%l=j$NCOM}E*$J%3%s%F%-%9%H$rH?1G$7$?G'<1$,2DG=$H$J$k$H9M$($F$$$k!%G'<1pJsFCD'$N$I$A$i$,$I$l$@$1M-8z$K:nMQ$7$?$+$K$D$$$F!$$h$jL@3N$KH=CG$9$k$Y$/!$5!3#3X=,$NCJ3,$K$*$$$F%^%k%A%+!<%M%k3X(B$B=,(B (MultipleKernel Learning)$B$rMQ$$$F!$pJs$NM-8z@-$rI>2A$9$k(B$B!%(B

2 $B

$B2f!9$NG'<1$B?^(B 1$B$K<($9MWNN$G9T$o$l$k!%A4BN$NN.$l$H$7$F$O(B$B!$(B Flickr$B$h$j<}=8$7$?0LCV>pJsIU$-2hA|$+$iFCD'Cj=P$7$?$b$N$r!$5!3#3X=,$9$k$3$H$GG'<1@:EY$r8!>Z$9$k$b$N$G$"$k!%(B

$B5!3#3X=,$NCJ3,$K$*$$$F!$2hA|$H0LCV>pJs$,$I$N$h$&$KM-8z$KMxMQ$5$l$F$$$k$+$I$&$+$r9M;!$9$kI,MW$,$"$k!%$3$l$K$"$?$j=i4|$N$BK!(B [6, 7]$B$G$O!$0LCV>pJs$H9R6u$B??(B ($B$J$$$7$O0LCV>p(B$BJs(B )$B$rC1=c$K7k9g$7$?FCD'NL$r3X=,$5$;!$7k9g$9$kFCD'$NZ$7$F$$$?!%$?$@$7AH$_9g$o$;$4$H$N@:EY$rHf3S$9$k$N$_$G$O!$2hA|$H0LCV>pJs$H$NM-8z@-$r=@Fp$KH=CG$9$k>e$G$OIT==J,$G$"$C$?!%K\8&5f$G$O!$8e=R$9(B$B$k(B MKL-SVM$B$rMQ$$$F!$G'<1@:EY$N$[$+$K2hA|$H9R6upJs!$2hA|(B$B$N(B HSV$B?'FCD'$NM-8z@-$r=E$_$N?dDj$K$h$j8!>Z$9$k(B$B!%(B

2.1 $B2hA|$+$i$NFCD'Cj=P(B

$B2hA|$NFCD'$r5-=R$9$k$B$k(B SIFT$BFC(B$BD'(B [2]$B$r4J7i$K5-=R$9$k$Y(B$B$/(B bag of keypoints$BI=(B$B8=(B [1]($B0J(B$B2<(B BoK)$B$KJQ49$9$k$3$H$G%Y%/%H%kNL;R2=$7$?%G!<%?$H(B$B!$(B HSV$B?'6u4V$K$h$k%+%i!<%R%9%H%0%i(B$B%`(B ($B0J(B$B2<(B HSV)$B$rMQ$$$k!%%Y%/%H%kNL;R2=$O!$%f!<%/%j%C%I5wN%$r<\EY$H$7(B$B!$(B k-Means$BK!(B ($BK\8&5f$G(B$B$O(B k = 1000)$B$rMQ$$$F9T$&(B.

$B2hA|$+$i(B$B$O(B BoK$B$H(B HSV$B!$9R6u$B$O(B BoK$B$N$_$rCj=P$9$k(B$B!%(B

2.2 $B%a%?%G!<%?$+$i$NFCD'Cj=P(B

$B2hA|$KBP1~$9$k%a%?%G!<%?$+$i!$;~4V>pJs$H<~JU%F%-%9%H>p(B$BJs(B ($B0J2$B%H(B )$B$r!$5!3#3X=,$N$?$a$NFCD'NL$H$7$FCj=P$9$k!%$=$l$>$l2hA|$N;#1F0LCV$N0^EY!&7PEY(B$B$r(B Yahoo!$B%m!<%+%k%5!<(B$B%A(B API$B$KM?$(!$La$jCM$+$i0LCV<~JU$N7zJ*!&;\@_!&%i%s%I%^!<%/$J$I$K4X$9$k%F%-%9%H>pJs(B$B$r(B Chasen[8]$B$G7ABVAG2r@O$7!$Cj=P$7$?A4C18l$N=P8=IQEY$N>e(B$B0L(B 2000$B8l$KBP$7$F%R%9%H%0%i%`$r:n@.$7!$COM}%F%-%9%HFCD'$H$9$k!%;~4V>pJs$K$D$$$F$O!$7n$K$D$$(B$B$F(B 12$B%S%s$H;~4V$K$D$$(B$B$F(B 24$B%S%s$rMQ0U$7!$3:Ev$9$k7n!&;~9o$H$=$NNY@\$K$=$l$>(B$B$l(B 0.5$B!$(B0.25$B$rEjI<$7!$%R%9%H%0%i%`$K=`$:$kI=8=$H$9$k(B$B!%(B

2.3 $B3X=,$HJ,N`(B

$BG'<1@:EY$N8!>Z$KEv$?$C$F$O!$=>Mh$N$B%s(B (SVM)$B$rMQ$$$k$3$H$K2C$(!$2hA|$H9R6u$B!%(B SVM$B$N%+!<%M%k4X?t$H$7$FHs@~7A(B$B$N(B 2$B%+!<%M%k$rMQ$$$k(B$B!%(B

2.4 $B%^%k%A%+!<%M%k3X=,(B

$BFCD'$rE}9g$7$FFCD'NL$NM-8z@-$r<1JL$9$k$?$a$K!$J#?t$NFCD'NL$N%+!<%M%k$r@~7A7k9g$9$k$3$H$K$h$jE}9g%+!<%M%k$r:n@.$7!$$=$l(B$B$r(B SVM$B$KE,MQ$7$FFCD'E}9g$K$h$k2hA|G'<1$r$B%k(B ($B%+!<%M%k$r=E$_$D$-$G@~7A7k9g$7$?%+!<%M(B$B%k(B )$B$N%5%V%+!<%M%k$KBP$9$k=E(B$B$_(B j$B$r3X=,$9$k!%$3$l$O%^%k%A%+!<%M%k3X(B$B=,(B (MKL) [3]$BLdBj$H8F$P$l!$E}9g%+!<%M%k$O0J2<$GDj<02=$5$l$k(B$B!%(B

$B:G6a$N8&5f$G$O!$$3(B$B$N(B MKL$BLdBj$rFLLL:GE,2=LdBj$H$7$F8z2LE*$K2r$/J}K!$,Ds0F$5$l$F$$(B$B$k(B [4]$B!%%^%k%A%+!<%M%k3X=,(B$B$O(B SVM$B$N$_$rA0Ds$H$7$?$b$N$G$O$J$$$,(B$B!$(B SVM$B$N%U%l!<%`%o!<%/$G2r$/J}K!$,0lHLE*$G(B$B!$(B MKL-SVM$B$H8F$P$l$k$3$H$b$"$k!%K\8&5f$G$O!$(BSHOGUN[5]$B%D!<%k%-%C%H$rMQ$$$F$B$?(B MKL-SVM$B$r;HMQ$7$F$B!%(B

3 $B

$B$B%s(B (SVM)$B$G3X=,$5$;!$J,N`7k2L$K$h$j@:EY$rH=Dj$9$k$3$H$K$h$j9T$&(B$B!%(B SVM$B$N%U%l!<%`%o!<%/$N2<$G!$%^%k%A%+!<%M%k3X=,$G$B$k(B (MKL-SVM)$B$3$H$K$h$j!$2hA|$H0LCV>pJs$J$I$NB>$NFCD'NL$,$=$l$>$l$I$l$@$1M-MQ$G$"$k$+$rH=CG$9$k(B$B!%(B

3.1 $B%G!<%?%;%C%H(B

$B3X=,$HJ,N`$O@52r2hA|$HIT@52r2hA|(B$B$N(B 2$B%/%i%9$G9T$&!%@52r%G!<%?%;%C%H$K$D$$$F$O!$(B$BI=(B 2$B$K<((B$B$9(B 8$B%8%c%s(B$B%k(B 28$B$B$G(B 200$BKg$:$DA*Dj$7$?!%IT@52r%G!<%?%;%C%H$O(B$B!$(B Flickr$B$+$i<}=8$7$?2hA|%G!<%?$NCf$+$i!$@52r%G!<%?$KMQ$$$i$l$F$$$J$$$b$N$r%i%s%@%`$KA*Dj$9$k$3$H(B$B$G(B 200$BKg=`Hw$7$?!%9R6u$B$O(B 4$B$B$7(B ($B?^(B 3)$B!$;#1F0LCV$,Cf1{$K$/$k$h$&$J@5J}7A$K2C9)$7$?$b$N$rMxMQ$9$k!%(B

$BI=(B 1: MKL-SVM$B$K$*$1$kJ?6QE,9gN($N7W;;7k2L$H(B$B!$(B MKL$B$K$h$k=E$_?dDj7k2L(B$B!%(B 2$B$B$k(B 1$B!A(B4$B!$(BG$B!$(BT$B!$(BH$B$O!$$=$l$>$l9R6u$B!$(B HSV$B$r<($9!%(B


3.2 $BFCD'NL$N

$B2hA|(B$B$N(B bag of keypoints$BI=8=!$9R6u$B$N(B bag of keypoints$BI=(B$B8=(B (4$B%l%Y%k$G(B$B$N(B 4$B$BN`(B )$B!$<~JU>pJs$K4X$9$k%R%9%H%0%i%`!$;#1F;~9o$K4X$9$k%R%9%H%0%i%`I=8=(B$B!$(B HSV$B%+%i!<%R%9%H%0%i%`$N!$(B$B7W(B 8$B$B!$(B MKL$B$KM?$($k$K$"$?$j(B$BI=(B 3$B$N$h$&$KAH$_9g$o$;$k(B$B!%(B

3.3 $BI>2A(B

$B3X=,$HJ,N`(B$B$O(B 5-fold$B$N%/%m%9%P%j%G!<%7%g%s$G9T$&!%G'<1@:EY$O3FJ,N`7k2L$K$*$1(B$B$k(B SVM$B$N=PNOCM$KBP$7$F!$J?6QE,9gN($r7W;;$9$k$3$H$GI>2A$9$k!%J?6QE,9gN((B$B$O(B recall-precision$B%0%i%U$NLL@Q$KAjEv$9(B$B$k(B ($B?^(B 2)$B!%(B

3.4 $B7k(B$B2L(B

28$B%+%F%4%j(B$B$r(B 8$B%8%c%s%k$K$^$H$a$?3FJ?6Q$K$D$$$F(B$B!$(B MKLSVM$B$NA4$F$NAH$_9g$o$;$K$h$kG'<1@:EY$H(B$B!$(B SVM$B$K$h$k=E$_?dDj7k2L$N$&$A(B$B!$(B 2$B$BN`(B ($B%Z(B$B%"(B )$B$K$h$k7k2L$r>JN,$7$?$b$N$r(B$BI=(B 1$B$K<($9!%(B

$B$B!$(B 28$B$B$?(B 8$B$B$,(B 81.05%$B$H!$2hA|(B$B$N(B bag of keypoints$BI=8=C1BN(B$B$N(B 77.05%$B$r(B 4%$B2~A1$7(B$B!$(B HSV$B?'FCD'C1BN(B$B$N(B 75.40%$B$r(B 6%$B$7$N$$$@!%K\pJs%F%-%9%H$+$i@8@.$G$-$kFCD'NL$O!$9R6u$B$N(B visual words$B$r(B MKL$B$N=E$_$GN?2o$7$?!%(B

$BA4BNE*798~$r=E$_$H@:EY$+$iH=CG$9$k$K$O!$B?MM$JFCD'NL(B $B$rM;9g$7$?>l9g$K%F%-%9%HFCD'$,M-Mx$G$"$j!"$3$l$O0UL#$,A/L@(B $BL@$J0LCV>pJs%F%-%9%H$NJ}$,9R6u$B!$(B 5$B\:Y$J9R6u$BA|(B BoK$B$K\:Y$J9R6uDj$9$l$P!$9R6u$HHf3S$7$F;~4V$N=E$_$,6/$$$,!$;~4V$NFCD'NL<+BN$OA4BNE*$K=E$_$GB>$rN?$$$G$$$kLu$G$O$J$$!%(B

$B%i%s%I%^!<%/$dFCDj$N%$%Y%s%H$K4X$7$F$O0LCV>pJs$K$h$C$F@:EY$,8~>e$7$?$,!$CO7A$dCOM}9=@.J*$K$D$$$F$O>\:Y$JCOM}>pJs$rpJs$K$h$C$F$"$^$j@:EY$,>e$,$i$J$+$C$?$3$H$,$o$+$C$?(B$B!%(B

4 $B$*$o$j(B$B$K(B

4.1 $B$^$H$a(B

$BK\8&5f$G$O!$0LCV>pJsIU$-pJs$rIU2CE*$J2hA|FCD'NL$H$7$FMxMQ$9$kZ$9$k$H$H$b$K!$%^%k%A%+!<%M%k3X(B$B=,(B (MKL,Multiple Kernel Learning)$B$rF3F~$9$k$3$H$G!$FCD'NL$N$B!%(B 28$B$B!$(B 8$B$B$N(B bag of keypoints$BI=8=(B$B$H(B HSV$B?'FCD'$N$=$l$>$lC1BN$G$NG'<1@:EY$r2~A1$7$?(B$B!%(B

4.2 $B:#8e$N2]Bj(B

$B0LCV>pJs$NG'<1@:EY$r$5$i$K8!>Z$7!$LrN)$F$k$K$O!$BgNL$+$DNIpJsIU$-

$B;29MJ8(B$B8%(B

[1] G. Csurka,C. Bray,C. Dance, and L.Fan.Visual categorization with bags

of keypoints. Proc.of ECCVWorkshop on Statistical Learningin Computer Vision, pp. 5974, 2004.

[2] D. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. InternationalJournalof ComputerVision,Vol. 60, No. 2, pp. 91110, 2004.

[3] G. R.G. Lanckriet,N. Cristianini,P. Bartlett, L. E. Ghaoui, and M. I.Jor

dan. Learning the kernel matrix with semidenite programming. Journal of Machine Learning Research,Vol. 5, pp. 2772, 2004.

[4]S. Sonnenburg,G.R$B!/(Batsch,C.Sch$B!/(Bafer, andB.Scha$B!/(Bolkopf. Large scale mul

tiple kernel learning. Journal of Machine Learning Research, Vol. 7, pp. 15311565, 2006.

[5] Shogun -A Large Scale Machine Learning Toolbox. http://www.shoguntoolbox.org/

[6] K.Yaegashi and K.Yanai. Can Geotags Help Image Recognition?. Proc. of

thePacic-Rim Symposium on Image andVideoTechnology, pp. 361373, 2009.

[7] $BH,=E3_7C(B$BB@(B ,$BLx0f7<(B$B;J(B :$B;#1F0LCV$N>pJs$rMQ$$$?0lHL2hA|G'<1$N2DG=@-$N8!(B$BF$(B ,$B>pJs=hM}3X(B$B2q(B CVIM$B8&5f2q(B, pp.1522, CVIM163-3, (2008)

[8] $BCcd%(B. http://chasen.naist.jp/hiki/ChaSen/.