Learning to Structure an
Image with Few Colors
2020 8 3
● Learning to Structure an Image with Few Colors
❖ Yunzhong Hou, Liang Zheng, Stephen Gould
(Australian National University)
❖ CVPR2020
2
31bit(2 )
※ OCTree
Q.
41bit(2 )
※ OCTree
A.
51bit(2 )
OCTree ColorNet
6
1
2N N =
K-means
Median Cut
OCTree
4bit
↓
16
1bit
↓
2
8bit
↓
256
2bit
↓
4
※ K-means
CNN
7
1bit(2 )
CNN
◆
◆
CNN+AutoEncode
8
1~2bit
ColorCNN
9
ColorCNN
activation
[ ⋅ ]i : i
∙ :
1,2bit
C-channel
one pixel
0.2
0.0
0.7
0.1
{c = 4
activation
❶
[t(x)]c =
∑(u,v)
[x]u,v ⋅ [m(x)]u,v,c
∑(u,v)
[m(x)]u,v,c
❷
˜x =
∑
c
[t(x)]c ⋅ [m(x)]c
❸
Umap
ψ⋆
= arg min
ψ
∑
(x,y)∈D
ℒ
(
y, fθ⋆
(gψ(x)))
+ γR
10
ψ⋆
: ColorCNN
fθ⋆ :
R :
γ :
gψ(x) :
AlexNet
11
epoch: 60, batch size: 128, optimizer: SGD
❷
ColorCNN
❶ ColorCNN
❸
ImageNet: A Large-Scale Hierarchical Image Database
CIFAR-100 (Canadian Institute for Advanced Research)
AlexNet
12
epoch: 60, batch size: 128, optimizer: SGD
❷
ColorCNN
❶ ColorCNN
❸
ImageNet: A Large-Scale Hierarchical Image Database
CIFAR-100 (Canadian Institute for Advanced Research)
6bit

【論文読み】Learning to Structure an Image with Few Colors