minP,B,S,<W,b>s.t.||PX−BS||2F+λ1||S||22+λ2||X−PTPX||2F+2λ3∑c=1CL(S,yc,wc,bc)PPT=I,∑i=1rB2i,j≤1,∀j=1,...,K(3)(4)
(3)
min
P
,
B
,
S
,
<
W
,
b
>
|
|
P
X
−
B
S
|
|
F
2
+
λ
1
|
|
S
|
|
2
2
+
λ
2
|
|
X
−
P
T
P
X
|
|
F
2
+
2
λ
3
∑
c
=
1
C
L
(
S
,
y
c
,
w
c
,
b
c
)
(4)
s
.
t
.
P
P
T
=
I
,
∑
i
=
1
r
B
i
,
j
2
≤
1
,
∀
j
=
1
,
.
.
.
,
K
其中,L(⋅)
L
(
·
)
为SVM优化模型,∀c
∀
c
写作,
L(S,y,w,b)=||w||22+θ∑i=1Nl(si,yi,w,b)
L
(
S
,
y
,
w
,
b
)
=
|
|
w
|
|
2
2
+
θ
∑
i
=
1
N
l
(
s
i
,
y
i
,
w
,
b
)
上式l(si,yi,w,b)
l
(
s
i
,
y
i
,
w
,
b
)
为Hinge损失,但实验证明采用其二次形式精度更高,其二次型可写作,
l(si,yi,w,b)={max(0,1−yci(wTcsi+bc))}2
l
(
s
i
,
y
i
,
w
,
b
)
=
{
m
a
x
(
0
,
1
−
y
i
c
(
w
c
T
s
i
+
b
c
)
)
}
2
也即是,
l(si,yi,w,b)=⎧⎩⎨||yci(wTcsi+bc))−1||22,0,if 1−yci(wTcsi+bc))>0else
l
(
s
i
,
y
i
,
w
,
b
)
=
{
|
|
y
i
c
(
w
c
T
s
i
+
b
c
)
)
−
1
|
|
2
2
,
if
1
−
y
i
c
(
w
c
T
s
i
+
b
c
)
)
>
0
0
,
else
模型求解
Step 1. 低维空间字典学习
等价目标,
minB,Ss.t.||PX−BS||2F+λ1||S||22+2λ3∑c=1CL(S,yc,wc,bc)∑i=1rB2i,j≤1,∀j=1,...,K(144)(145)
(144)
min
B
,
S
|
|
P
X
−
B
S
|
|
F
2
+
λ
1
|
|
S
|
|
2
2
+
2
λ
3
∑
c
=
1
C
L
(
S
,
y
c
,
w
c
,
b
c
)
(145)
s
.
t
.
∑
i
=
1
r
B
i
,
j
2
≤
1
,
∀
j
=
1
,
.
.
.
,
K
① 固定B
B
,求S
当HLS⊆∅
H
L
S
⊆
∅
,(参考论文式4.16~4.19)
S=(BTB+λ1I)−1BTPX
S
=
(
B
T
B
+
λ
1
I
)
−
1
B
T
P
X
当HLS
H
L
S
不为空集时,等价于对每个
minsis.t.||PX−BS||2F+λ1||S||22+2θλ3∑i,c∈HLS∑c=1C||yci(wTcsi+bc))−1||22∑i=1rB2i,j≤1,∀j=1,...,K(146)(147)
(146)
min
s
i
|
|
P
X
−
B
S
|
|
F
2
+
λ
1
|
|
S
|
|
2
2
+
2
θ
λ
3
∑
i
,
c
∈
H
L
S
∑
c
=
1
C
|
|
y
i
c
(
w
c
T
s
i
+
b
c
)
)
−
1
|
|
2
2
(147)
s
.
t
.
∑
i
=
1
r
B
i
,
j
2
≤
1
,
∀
j
=
1
,
.
.
.
,
K