II. PROPOSED CSGAN ARCHITECTURE
数据集 X ∈ { ( A i ) , ( B i ) } i = 1 n X\in\left \{ (A_i), (B_i) \right \}_{i=1}^n X∈{(Ai),(Bi)}i=1n，包含 n n n个样本，每个样本包含来自domain A A A和 B B B的2幅paired images
学习目标是2个生成器：
G
A
B
:
A
→
B
G_{AB}: A\rightarrow B
GAB:A→B，
G
B
A
:
B
→
A
G_{BA}: B\rightarrow A
GBA:B→A
G
A
B
G_{AB}
GAB利用real image
R
A
R_A
RA生成synthesized image
S
y
n
B
Syn_B
SynB
G
B
A
G_{BA}
GBA利用real image
R
B
R_B
RB生成synthesized image
S
y
n
A
Syn_A
SynA
S
y
n
B
=
G
A
B
(
R
A
)
(
1
)
S
y
n
A
=
G
B
A
(
R
B
)
(
2
)
\begin{aligned} &Syn_B=G_{AB}(R_A) \qquad(1) \\ &Syn_A=G_{BA}(R_B) \qquad(2) \end{aligned}
SynB=GAB(RA)(1)SynA=GBA(RB)(2)
两个判别器
D
A
,
D
B
D_A, D_B
DA,DB
D
A
D_A
DA负责判别
R
A
R_A
RA和
S
y
n
A
Syn_A
SynA
D
B
D_B
DB负责判别
R
B
R_B
RB和
S
y
n
B
Syn_B
SynB
生成图像
S
y
n
A
Syn_A
SynA和
S
y
n
B
Syn_B
SynB被用于二次生成，得到
C
y
c
A
Cyc_A
CycA和
C
y
c
B
Cyc_B
CycB
C
y
c
A
=
G
B
A
(
S
y
n
B
)
=
G
B
A
(
G
A
B
(
R
A
)
)
(
3
)
C
y
c
B
=
G
A
B
(
S
y
n
A
)
=
G
A
B
(
G
B
A
(
R
B
)
)
(
4
)
\begin{aligned} &Cyc_A=G_{BA}(Syn_B)=G_{BA}(G_{AB}(R_A)) \qquad(3) \\ &Cyc_B=G_{AB}(Syn_A)=G_{AB}(G_{BA}(R_B)) \qquad(4) \end{aligned}
CycA=GBA(SynB)=GBA(GAB(RA))(3)CycB=GAB(SynA)=GAB(GBA(RB))(4)
模型框架图如Fig.2所示
A. Proposed Cyclic-Synthesized Loss
Cyclic-Synthesized Loss的思想是，使用同一个生成器生成的Synthesized Image和Cycled Image必须近似，具体定义如下
L
C
S
A
=
∥
S
y
n
A
−
C
y
c
A
∥
1
(
5
)
L
C
S
B
=
∥
S
y
n
B
−
C
y
c
B
∥
1
(
6
)
\begin{aligned} &\mathcal{L}_{CS_A}=\left \| Syn_A-Cyc_A \right \|_1 \qquad(5) \\ &\mathcal{L}_{CS_B}=\left \| Syn_B-Cyc_B \right \|_1 \qquad(6) \end{aligned}
LCSA=∥SynA−CycA∥1(5)LCSB=∥SynB−CycB∥1(6)
B. CSGAN Objective Function
CSGAN的objective function定义如下
L
(
G
A
B
,
G
B
A
,
D
A
,
D
B
)
=
L
L
S
G
A
N
A
+
L
L
S
G
A
N
B
+
λ
A
L
c
y
c
A
+
λ
B
L
c
y
c
B
+
μ
A
L
C
S
A
+
μ
B
L
C
S
B
(
7
)
\begin{aligned} \mathcal{L}&\left ( G_{AB}, G_{BA}, D_A, D_B \right )=\mathcal{L}_{LSGAN_A}+\mathcal{L}_{LSGAN_B}\\ &+\lambda_A\mathcal{L}_{cyc_A}+\lambda_B\mathcal{L}_{cyc_B}+\mu_A\mathcal{L}_{CS_A}+\mu_B\mathcal{L}_{CS_B} \qquad(7) \end{aligned}
L(GAB,GBA,DA,DB)=LLSGANA+LLSGANB+λALcycA+λBLcycB+μALCSA+μBLCSB(7)
其中
L
L
S
G
A
N
A
,
L
L
S
G
A
N
B
\mathcal{L}_{LSGAN_A}, \mathcal{L}_{LSGAN_B}
LLSGANA,LLSGANB是least square adversarial loss，
L
c
y
c
A
,
L
c
y
c
B
\mathcal{L}_{cyc_A}, \mathcal{L}_{cyc_B}
LcycA,LcycB是Cycle-consistency Loss，
L
C
S
A
,
L
C
S
B
\mathcal{L}_{CS_A}, \mathcal{L}_{CS_B}
LCSA,LCSB是提出的Cyclic-Synthesized Loss
【局限性】
必须要样本是paired，才能使用Cyclic-Synthesized Loss
817
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
