Mraissi58@yahoo.com homayoun@ce.aut.ac.ir SVM SVM SVM VQ ...

grizzlybearcroatianΤεχνίτη Νοημοσύνη και Ρομποτική

16 Οκτ 2013 (πριν από 3 χρόνια και 8 μήνες)

110 εμφανίσεις



 
            !" #$
%&'  ()) !$&

 



 


Mraissi58@yahoo.com
homayoun@ce.aut.ac.ir

 
     
 .
  !  "# $ %&'
 ( ) ( *    +"#


 

SVM
￿
  
        .   
SVM
   
 !"  #
$ %&'( )*
+,    - $/   0  . 1 0  "2 34 ''!(     
SVM
%   0$ 5 1 657 8 2 $  .
9$( 1 0 8 %&2   52:75
VQ
￿
;15
<
1 =  >7   15
<   75?7 -( 2  5@$

A2 0  0  "2 %&2 .  0  "2 5
& B ( 657  C  657 A&( 0 :
15
< 0   :
D ( 0 E >7 F 
1< 
SVM
A2 0 2 . G 7  5(  1 0 " 2 A" 5( 
0
< A2
VQ
H I J  15
<   75?7 $ 2, 1 K 4
$ 
L"I :7   %&2 8    
=  F   %&2 ;15
<
1  . M8N 1 0  "2  0 
< A2 G 7  5( 
C 
MFCC
0$$ $ 2 0 ?  < F '

SVM
 I  7
7  5( 
25 & B ( OP/RS % U8 2 $ V
A# A2 0$ 5  1 657 .

:
,#- /0 "1 2/  3 #- 4  2  / - 5   67  .

8  #- 1 2-  6 90) $ '  # $  #   :;  ##< 6< #(0<) 0" / - 5  67   , . <-
( / <" , < * 1 :<- 6< , < 6< # =<" # < #- 1  6!<( #(> / -  67   , 8   #  
< ?7 < ( /   - @ ,  1  , - # A / 6 :B' :-  6 8 ) 26C , ( ,( <) <- @
( <CCD 4<- ?<) > ( E&<F 2 < G"< # H IJ $ %&' 0<"
<  3 :<- < # ( <( 
<  < - #
#<J- 67K <- A<    / - =" 67   2 "#J- C L
C  /-  :+# # :" - ;  %&'
A . # M'  - 6 "- 81 N 4"O ( (  , :%P / -.


1
Support Vector Machine
2
Vector Quantization

Q  #RSTS
House
"#" #   / - 67   ,- U# M1  Q  / + ( . Q  #RSVW ,# 8 
@
Leonard
+ #   6X( #- /% :-  C) 5" - 5 , F(J  :1  :7  / + ( # . Q  #RSVY
@ 6 
Cimarusti
- , :CF ; 1 ,&1 5   :1  N' RWW 6X( , #- ( I%
LPC
#  
#1 6 .
Sugiyama
Q  # RSSW , BB 
VQ
6X( Z7[  :-
LPC
#" #     / - 6   A F .
Q  #RSS\ 0"
Muthusamy
- ,- E&1 A F 6I% , :B  #   ( ]0- 6B ") , ( :- ") 8^
A#  / - =" _L # 6! " .  ` P" / - 5   67   a :- ,#  $ !D b 0" J / D  #
 2 6%I 6I% , :B  #   "  6" " , ( ( A ,- 6 cCL Q  ( M1  6L Q
A  : )  +-  / - =" 67  
.[1]


SVM
# 6 O #   # , - ,&1 ,- ZCd :1  - 6 ,F H " ,#  e( . 4 D( f) e(
Q  #  -RSTS ` " :- 6( ,"# @
Vapnik
gb 
[2]
Q  # / ,( 6C $K ^ ;  RSSh ## .
# 6C M
SVM
1 6 F B  6bJ $ :-  ## :1 A , :>  - 4 )  . # ZCd # /;
2#" F B  6bJ $ :- / 6"  ##
SVM
:  a" / ^  , :   #   - 2, :CF ; ,
## " , F ( K - !- - 67 i) :- :"" /#- # 6! 7 , :   ( 6 , :  .
 - " :F bJ j-  6C> , 8 # /# )  ?B - 6I% , :B M&J - /  3  #- 4  . 6)' 
 -  J  /#( A- , 1 ,&1   :- A"  1 ,    3 4 # ?D# :- /  3 #- 4  ,- :
 - 6  "  6bJd .  - 6 ,0  1 ,&1 6%" /  3 #- 4  : P"  4k
[3]
4 )  # 6! (
6  :"" #! - ## 6 , :"" 4-  cCL Q ( M1  6L Q 4+D 2  - 6 e ?- O 81
 - 
#    , :""   " 2
# 
## Q  N 6D ; lO# 4 L a :- 6.
8#   / - =" 67   A F /  3  #- 4   #   IO 6^ ^>  1 4 # . " - #  #   ,
 - 6 61 / - ( 6-% / - 26C" / - 2 6 ) / - $   . `  6  - / -  /#" Q ,- :B" ?D# :-
:F - ( 8 - 6 (-( 6 , :""   ,#  8PN - ZCd G3 2# : ) a" # / 6B ") ( 67( $ IJ
e  /   :- ,K - H "
SVM
 #   # F ,#B( # 2
SVM
5B  #   -  - 2
VQ
, :"" #!
# 6 ## 9 1 8  6 .  #   - f ) e( # # $ % :-
VQ
:-  / -  m   6X( , #-
:J 4 ?^n 01 ( #" ,- :^' 67 :J8 6 a" # / - / m   6  , #- /% :- . #   -
# 6 oN 8  67 1 Q N 4% # ( -  6 9 1 $ :- 8  e /   f) e(  .
9L- #  - :D ^ :# # Y , <- ,<&1 e( <- 6D <F ,(<
SVM
A<# 8 <J . 9<L- # G*<\ e(
# 6 ## g 6 ^^>  1 4 #  #   . 9L- #p #( ,# <3 8 < ,( <- < ` <P" $   q "
9L- A " # ( A h##3  :D ^ , :P " :- ..


  
 !"
 
 
 
 , ## 1 r) , F( ?B 2 "   `1  6 F(J sCb # ^ ( 6
),(
ii
yx
:Bb-  -  
n
i
Rx ∈
( Li,...,2,1=(
)1,1( +−∈
i
y
 - 6 ., i) #
d
#- 5 - / 6   :> -  1t ,!- 
/(
w
E- A-n ^ 5 (
b
#" u!  ?B :- :

0.=+
bxw


)R (

E&1 , :"" " - :1  - 6 a" , :> - 4 )  /  4 #R + E&1 , :""   R   F  " . :> - :-
# 6 :  1 0   :> - f) .:b^"  A 6 -
x
:D#! # # O :>- ,( :1
0.=+
bxw
f
1 6
.[2,4]
1 0   j-  51 :- ( J :D# ! 51 :-  f) E&1 (#
).()(
bxwsignxf
+=
 / 6
#" F B . - ,( Q( E&1 6X( ,  #- (  - t3 67F E&1 (# , ## 1 r) # $ % :-
:>
+
H
:> - ,( `(# E&1 6X( ,  #- (

H
" jO( .$ > -
+
H
(

H
$  :- / 6 
#" u!  :
1.:
1.:
−=+
+=+

+
bxwH
bxwH

)Y (


$ > -  :1 1 :F
+
H
(

H
 - 6 , .$ > - ,( :1 67 :""
+
H
(

H
#- 2" 6 O
" 6  " /  3 .:> -  (# 4- :C )
+
H
(

H
- -- :1
w
2
 6 , :N "  : N   - 6 .
0- ,0 :N " 4 # :1 A z[( - 6  -Cb  .6  - a 4-
w
2
6 % :-  ( ## 801 
w
?ON
# . - :"" H " n (  " ,- :^' 6 # :-  :"" HK(  6N' 6  - 6"  3 #- 4  A / 6 G3
 { J  ( ,(" O , :N "  . :"" `  | - A) 1 f) @ | [ ,-
1).(
≥+
bxwy
ii
 - . ` 
#" :&J   :b- # / 6  f) @:







=≥+ LiforbxwytoSubject
wMinimize
ii
,...,2,11).(
2


)\(

s> ,  : - :C} 5 f) :C}
￿
`(# :F# 
￿
 - 6 .-X"K j-  f) :C} ?N ,
￿
 / ( ## ?B   
81 6 ?N:

=
+−=
L
i
iii
bxwywbwL
1
2
).(
2
1
),,( αα

)p(

l  /## O  ,(  -
L
:- A"
w
(
b
(
α
K#∆! :- sαF :b^" # 8 6  $v



=
=
=
=
L
i
ii
L
i
iii
y
xyw
1
1
0
α
α

)h(

# f) # ^ O - (
),,(
α
bwL
8 6 M( / (# :C} :- :


3

Convex

Optimization
4

Quadratic
5

Lagrange Function
)~ (







==≥
−=

∑ ∑∑
=
= = =
L
i
iii
L
i
L
i
L
j
jijijii
yandLifortoSubject
xxyywMaximize
1
1 1 1
0,...,2,10
).()(
αα
αααα

K Z7[ / (# f) :C} ?N  `1  :B' :- 2  # 6  :-   X"
i
α
, :""  6B - m  
i
x
 - 6 .
, D
i
x
- m   :1 
0

i
α
/  3  #- 2 - 6
￿
i
sv
 " 6 . 1 0   :> - j-  / 6 $I #
A"  $ :- 
[2,4]
:
)T (
)).(()).(()(
11
∑∑
==
+=+=
SvSv
N
i
iii
N
i
iii
bsvxySignbxxySignxf
αα

  
            !" #
 ?% # ZCd  - 6" t3 / B 6 N-  1 0   :> - 5 4 ) . % - " 6 ,O 0" 5 Q € /% :-
# &1 /# ) 8 ( .  -  O-  @ :1 , :"- #" #   #1 ,    / 6 # :" 4  #
[2]
:

0
,...,2,11).(

=−≥+
i
iii
Liforbxwy
ξ
ξ

)V(

# :"" / - #  :>  - 5 - i) / # :1 #- 67 i) :-  ## e#3 93 Q % - / 6 6bJ d AD N 
#" F B   .,#(( #- :1 #" #   6" ^  , :   / 6 a 4 ,-
n
,!-
x
#- :- 
n


,!-
z
1 6 ? .

nn
RR

→:φ

)S(
)(x
φ
 - t3 67F F 6X( , i) , #- :1 ## s L " , :"- 6  - .
Vapnik
, :  :1 #" A- n
" ^  :iO @ :1 6
Mercer
" #   a 4 ,- 6-J :- " 6 2 - : #  . , :   67 :""
 - 6  $ :- / ^ :
).tanh(),(
)
2
exp(),(
,.....3,2)1.(),(
2
2
θ
σ
+=

−=
=+=
yxyxk
yx
yxk
pyxyxk
p

)١٠ (


 :- j- 4 :- 6 5D-* A"X"  :  26 :  2, :CF ; :  Z . AD N # / (# ,  : - :C}
#-  J  $  :- 6bJ  d ( t3 " 67F:
)RR (







=≤≤
−=

∑ ∑∑
=
= = =
L
i
iii
L
i
L
i
L
j
jijijii
yCtoSubject
xxyywMaximize
1
1 1 1
0,0
)().(
2
1
)(
αα
φφαααα



6
Support Vector

 $ :- 0" /( #-# 6 u ! :
)RY (

=
=
L
i
iii
xyw
1
)(φα

#" u!  $ :- / 6 0"  1 0   :> - j- :
)))().((()(
1

=
+=
Sv
N
i
iii
bsvxySignxf φφα
)R\ (

- 2 - 6 t3 67F AD N # :C} ?N :-  6bJ d ( t3 " 67F AD N # :C} ?N # 6   :1 b" 
,- :1 $(  4
i
α
,K - /1
C
/ ^  , :  ( A   : ) a" #
φ
,∆σαD ,(
x
(
i
sv
A  Q % .
 # / X"K Z7[ :1 ,#( ( , :"" AD > # :1 1 :F:b-
C
i
≤≤
α
0
/  3 , #- 1 f
 - 6 8 
[2,4]
.
  
  
 
 l ^> 4 # p# #   uC L
" -0- / ## . 2$# E ) / ## 6 ) / - ,- #   # / ##
6C d / ## 6C" / -  ,- #   # / ##
TIMIT
jF / ## 61 ( 6- % / - ,- #   # / ## (
 6 ^^> 8 @  ,(6   ,       $ %&' ,( ) ( * 1 6  B"# , 
 -  1  6 !  "# . e#3 ,     , - # 6  ( 8  AO 5   , 6 X( {L 
 - 6   5   ."( 6  B-  1 4 ,- 6(   ( =  , (    6 IL {L  " M :1
 -    5  e#3 # a" # M ,- :1 A
[5]
. ,- `K $ IJ 6  - 2 s L " 6X(
" - €1N :-  67 " 8  AO# ( A% 2 ## 8PN ?ON - ( : #  a" # M . /#- /  d ?D- Q 
 1 /   $ :-    Q  HK! 2/  6' # / ,  ‚J  (  
7
6X( ( ##  O 0D " #
1 6 {L    / , . Q' - , ) :- $ Q  6^^ >  1 4 #Yh :B' :- A  8 ^ :" n 6C
" :- ( P , )Rh"# 6" * B - :" n 6C . $ %&' oN a :- ) ,- 6" * , B-
 - 6 ) , " .1  93  C) 51 :- $ Q  ,- 8)  G3
1
1)(

−= zzH
α

K - , " 1) 2
7 " 6 AJB  Q  u' ( #" A^  Q 8 . 6^^>  1 4  #STh/W
=
α
A  ## O
[5]
.  G3
8 #" s[ „ P3 5 #   8)  2 ) , " (  - # 6 3 " n 9 1 a :- Q  1  93 .
Z[ l^ > 4 # #   #
X(
8
MFCC
2
9
LPCC
(
MFCC

 -  .
      
2 / - :^ # :1 b" 
SVM
 / 6 6 N-  #( 6  1 :- D , - ,&1 ,- ZCd
SVM
,-
#" #   0" :&1 ; AD N .:^- ?- ^ # 5 e( 2 ( 4  6B
￿￿
- 6 
[6]
. # 6! J e( #
:&1 (# :C} ; :- :&1 ; :C} ?B 4- 28 #  e#J  d &1 6^-   &1  `1  , F
# 6 ? . &1 #!  tD
n
:- { N 2-
n
1 ,- : #
SVM
8# . …&1 2  a" # #- 1 #p E&1


7
Short-Time Analysis
8
Mel Frequency Cepstral Coefficient
9
Linear Prediction Cepstral Coefficint
10
One Versus the Rest

:- {  N f) , &1 ,&1 ,- :P " # 2 - 6 6-% ( 61 2 6C" 2 6 ) , " - - m   :1 ## #F(p
1 6 F " -     " -   6 B "   `1 :1 8# , - 1 ,- : # .  a " ,- : # :1 A 4 -
, - 1
SVM
: # :- A"  / - / - m   6 , :"" 6X( , #-  , - # 2/ - U - m  
, -
1 ,-
SVM
1 `&% #J :- lC!  " -   - m   .
,- 2#F / ## :- :F - 2/ -  Q e~WWWW #-   J #  #- . ' :- f) 6 , #-
" s L " / ## # #F $&F 4-  6O  . ^- <   < " -   <+ 2 < <
<- : #  `  :1 P" 
 / - :- lC!  :1 :""  " :- 2   F " - Z; - "#- a" #R + ,(  b- :1 :""  # " :- (
Z;- "#-  s L " # , " - R  ## . /#- #  ?D-~WWWW <- ( / <- < Q< e ,- 6X( #-
# Q( 
:   #  
SVM
Q :1 #  
SVM
"  D^! /  # ## 6 . Q<
<- :< <P" 
e( :<- / <- Q e /  /#- "K' :- :F - 0" ( A  " # / - /  ) 
## / - U /#"
SVM

<"  <  e( :<- < - 2 < 
< ##  8^  #  
P- : 8 ) 8I 2Q^! /   # / /"  (
 e(
VQ
8 # 9   ## 8PN 2 . e(  #   - ( ,#  , ##  #   -
VQ
8 <D  H <IL (
LBG

8#" # P 1 s 1 5 / -  ,- . e< :<CN # < , #- /% :- / -  1 s 1 ?^n 01
SVM

A) O #   # . #   - A<)  9 <1 8 < e< / <  6 F ?- O /0 :- f) e(  .  #  < <-
VQ

 6
:"" 8PN~WWWW :- ~Rpp ( YWYW ( A<) $< D e  F , ## - (  ## 9  :""
A) $ Q e DO ?- O /   $ # (  Q Z 4- .D N #  #   A~Rpp ,-  :""
6! ## 4  6" 2/ -  Q e\WTY Z;- - :"" R + `< < <# / <- :<  ( <a" # / - RWYp
Z;- - :""R

# #    " # 6 , ## # 8" /!-.

 
 
  
          !  
" # .
  
$% $% &$%

 
/'(% */'&% *+/'&
MFCC

/'(% ,/'+% ,/'+%
LPCC

/'(% /'(% */'&%
MFCC+

MFCC


/ - U :- ?O 9L- #  :7 8 
Visual C++
G" ) - * 
( - RWWW - 0   RYV :a) N A - 
RAM

  # 3 . / - , :"" ,- 2$# E ) / ##   ) / - 
:""
-  / - &O : b" 
6 Cd / ##  6C"
TIMIT
 6-% ( 61 / - 6 , :"" ,- ( # @  ,( jF / ##
6  ,       #   1  6 !  "# $ %&' ,( ) ( * 1 6  B"# , 
 . !- - 6X( , #-  #   -  ` P" $   q "RY Z[  ?B  (
MFCC
(
LPCC
, D(F # R (
Y  A  . Q(F :- :F -R( Y# 6 8  / " 90) % - A , :"" /  Q' 90) :1 A _L .
# 6 8  / " 90) % - / -  Q# ! 1 s 1 " 90) A- n 6X ( 5 | - 4k .
Q(F #  - AO# R 90)  :  #  / " 4 k A ## 8PN ) )~ - ( :" n ( : #  ZF
A : " "; n  7  # - # X( ="  # ( # O 8    J # )  $ %&' . : b"  tD
| - ##  :aN&~  X( ="  - A ## :" n
MFCC
:-
LPCC
(
MFCC

MFCC+
# "; 
## " :aN& 8  /  " . 6 90)  8  e /   6 , :"" #! 90) - :1 A 1† :- ` K
-  .A  _L , :CF ; :  j-   #   `  # f) : B" ."" , #- /#1 0D "  #   ( 6 :
( e /   6 F ?- O /0 :- / 6 2 " /#" 5;1 ( N( ,#% - " 8^ - A :"" , #-
,&1
SVM
## 9 1  . Q(F # , :CF ; :  j-   #   - 8  ,  # 3 q "\A  .

& 
      
 
 &+&+    !  
" # .
   
$% $% &$%

 
,/'*% &/,'% -/,-%
MFCC

*/'&% /,% /,%
LPCC

Q(F #pA  :7  , :CF; ( 6 :   #   - 8  e /    6L . 8  # 1 :F
-  2A" #F( 8  67 / B  2/0D " ` P" /(- , :CF ; :   #   AD N #  6N'
A)  9 1 $ :- 8  e /  2,#  , #- ( e , #- /#"  0D " . Q(F #h  6L 0"
 ,- `K / A   #( 6   , #- A .  " # /   # ZCd # :1 A"  61 N  A- q "
 -  9  ## /     / - =" 4 ! (  
## ,&1 ,- . Q(F : ^R Q(F ( \ / , 
"  81 6  ## 8PN | - :1 AY:" n :C F ; :  j-  :- A" , - / " :- P 6 :   #  
## 6 ,.

* 
      
 
   0 1#  2   !  
" # .
   
$% $% &$%

 
/'(% */'&% ,/'+%
MFCC

       !
    03  0  4 0  
 54 3 6
" 78 3!$9 4 : !  ;3< 
=

   7>?  0 1#  2   ! 
" # .
 

        
& =$@ 
*=$@ 0 1#  2

(  A
" B #
 C 3 7! 78 3!$9 4 : !  ;3< 
:  .
      ! "#$#  

 3#4*$% &$%

 3#4($% $%

 3#4,$%  $%
# 9  ## 8PN (
& /  /#- : ^ ?- O #F( -  #   - !O(
#-  ,- / 6 2 f) Q(F
#" j
#  N   ?+  !  (  $O3
*  .
  # 3 # uC L , ( : ^ a :- K!
N 2/ ## =" ?O   : )  B- @ /#- / B :- A -    e #3
#-  ( e
## 8P
#" :F / "  (  #   * 1 6#3 $ IL 2: )  +- X( =" 29  . : ^ a :- 0" P #
#  ` P" l ^> :P " -  l ^> 4 q " 2 / -   -
- (   - l^> 4 # #   # e(
   jFR8#" : ^ .A   #   :- 
X( ( / ##  l^> (#  # .   N : ^ 4 q "
e(  - 7 


SVM
e( - : ^ #
GMM
 -   / - 4!
- . e(  #   AD N #
SVM
/  "
 K -  A-SW %  AD N # ( Q  # 
GMM
/ " ~/VW %A   A-.

  
# 3  / - 5   67   ,- /  3  #- 4  ( ,#- /0 "1  ?B   6D :D ^ 4  #
# .:"" ,( ,#- /0 "1  #    M  / - :1 b"   Q# ! , D 4 J  ,- 6 ,
 - 6 e /  9 1 2/  3  #- 4  @ / - .  6N' 8  Z  67 1  61 N  A- q "
 - 6 .## a" # / " /0 :- 6 - 1 s 1 " 8a A 6 - .s 1  " 90) - A z[(  1
-  6 90) 8  /  " / - . :C N , :"" #!   0- / -  1 s 1 " : #" AO#   - 4k
 - / - A . 6 90) 8  /  " 2 A ## 8PN 90) - :1 A / 6   A- q " :- :F - 4k
-  .#   # # a" :- ,([ X( , #- /#" 0D " 2e /  9  a :- , :CF ; :  j-  
## 9 1  8  ˆ 3 ( 67 /  ( #" #   f) e(  / 6 0" 6 :  j-  ,- ;   6 . #
,- : # 67 A% : A / 6 6C1 AD N  1
SVM
1 ,- : #   1 , :C F ; :  j-  -
SVM
 - 6 6 :  j-  - . /   Q' ( / -  Q e  ,- #   # 1 s 1 " 90) - 4 k
-  6 90) 8  / " A .

ا
[1]
> 2 " 6i : B; 23 /  , " / # q , " -  6i!- 5  67  " :"K < G"<1 4  2
‚ 2/ * 1 4P" .pWR\SY  2R\VW.


[2]

C.

Cortes, V. Vapnik, "Support-Vector Networks", Machine Learning, Vol.

20, pp.

273-297, 1995.
[3]

V. Wan, Speaker Verification using Support Vector Machines, June 2003.
[4] C.

J.

C. Burges, "A Tutorial on Support Vector Machines for Pattern Recognition", Data Mining and
Knowledge Discovery, Vol. 2, pp.121-167, 1998.
[5]
4 - 26O ,#3   23 /  ,  > R\VR .

[6] B. Schölkopf, A.J. Smola, Learning with Kernels, MIT Press, 2002.