<s>
Multi-focus	B-Algorithm
image	I-Algorithm
fusion	I-Algorithm
is	O
a	O
multiple	B-General_Concept
image	I-General_Concept
compression	I-General_Concept
technique	O
using	O
input	O
images	O
with	O
different	O
focus	O
depths	O
to	O
make	O
one	O
output	O
image	O
that	O
preserves	O
all	O
information	O
.	O
</s>
<s>
In	O
recent	O
years	O
,	O
image	B-General_Concept
fusion	I-General_Concept
has	O
been	O
used	O
in	O
many	O
applications	O
such	O
as	O
remote	O
sensing	O
,	O
surveillance	O
,	O
medical	O
diagnosis	O
,	O
and	O
photography	O
applications	O
.	O
</s>
<s>
Two	O
major	O
applications	O
of	O
image	B-General_Concept
fusion	I-General_Concept
in	O
photography	O
are	O
fusion	O
of	O
multi-focus	O
images	O
and	O
multi-exposure	O
images	O
.	O
</s>
<s>
The	O
main	O
idea	O
of	O
image	B-General_Concept
fusion	I-General_Concept
is	O
gathering	O
important	O
and	O
the	O
essential	O
information	O
from	O
the	O
input	O
images	O
into	O
one	O
single	O
image	O
which	O
ideally	O
has	O
all	O
of	O
the	O
information	O
of	O
the	O
input	O
images	O
.	O
</s>
<s>
The	O
research	O
history	O
of	O
image	B-General_Concept
fusion	I-General_Concept
spans	O
over	O
30	O
years	O
and	O
many	O
scientific	O
papers	O
.	O
</s>
<s>
Image	B-General_Concept
fusion	I-General_Concept
generally	O
has	O
two	O
aspects	O
:	O
image	B-General_Concept
fusion	I-General_Concept
methods	O
and	O
objective	O
evaluation	O
metrics	O
.	O
</s>
<s>
In	O
visual	B-Device
sensor	I-Device
networks	I-Device
(	O
VSN	O
)	O
,	O
sensors	O
are	O
cameras	O
which	O
record	O
images	O
and	O
video	O
sequences	O
.	O
</s>
<s>
Therefore	O
,	O
just	O
the	O
object	O
located	O
in	O
the	O
focal	B-Algorithm
length	I-Algorithm
of	O
camera	O
is	O
focused	O
and	O
clear	O
,	O
and	O
other	O
parts	O
of	O
the	O
image	O
are	O
blurred	O
.	O
</s>
<s>
Due	O
to	O
the	O
large	O
amount	O
of	O
data	O
generated	O
by	O
cameras	O
compared	O
to	O
other	O
sensors	O
such	O
as	O
pressure	O
and	O
temperature	O
sensors	O
and	O
some	O
limitations	O
of	O
bandwidth	B-Algorithm
,	O
energy	O
consumption	O
and	O
processing	O
time	O
,	O
it	O
is	O
essential	O
to	O
process	O
the	O
local	O
input	O
images	O
to	O
decrease	O
the	O
amount	O
of	O
transmitted	O
data	O
.	O
</s>
<s>
Much	O
research	O
on	O
multi-focus	B-Algorithm
image	I-Algorithm
fusion	I-Algorithm
has	O
been	O
done	O
in	O
recent	O
years	O
and	O
can	O
be	O
classified	O
into	O
two	O
categories	O
:	O
transform	O
and	O
spatial	O
domains	O
.	O
</s>
<s>
Commonly	O
used	O
transforms	O
for	O
image	B-General_Concept
fusion	I-General_Concept
are	O
Discrete	B-General_Concept
cosine	I-General_Concept
transform	I-General_Concept
(	O
DCT	B-General_Concept
)	O
and	O
Multi-Scale	O
Transform	O
(	O
MST	O
)	O
.	O
</s>
<s>
Recently	O
,	O
Deep	B-Algorithm
learning	I-Algorithm
(	O
DL	O
)	O
has	O
been	O
thriving	O
in	O
several	O
image	O
processing	O
and	O
computer	B-Application
vision	I-Application
applications	O
.	O
</s>
<s>
Huang	O
and	O
Jing	O
have	O
reviewed	O
and	O
applied	O
several	O
focus	O
measurements	O
in	O
the	O
spatial	O
domain	O
for	O
the	O
multi-focus	B-Algorithm
image	I-Algorithm
fusion	I-Algorithm
process	O
,	O
suitable	O
for	O
real-time	O
applications	O
.	O
</s>
<s>
They	O
mentioned	O
some	O
focus	O
measurements	O
including	O
variance	O
,	O
energy	O
of	O
image	B-Algorithm
gradient	I-Algorithm
(	O
EOG	O
)	O
,	O
Tenenbaum	O
's	O
algorithm	O
(	O
Tenengrad	O
)	O
,	O
energy	O
of	O
Laplacian	O
(	O
EOL	O
)	O
,	O
sum-modified-Laplacian	O
(	O
SML	O
)	O
,	O
and	O
spatial	O
frequency	O
(	O
SF	O
)	O
.	O
</s>
<s>
Image	B-General_Concept
fusion	I-General_Concept
based	O
on	O
the	O
multi-scale	O
transform	O
is	O
the	O
most	O
commonly	O
used	O
and	O
promising	O
technique	O
.	O
</s>
<s>
Laplacian	O
pyramid	B-Algorithm
transform	O
,	O
gradient	O
pyramid-based	O
transform	O
,	O
morphological	B-Algorithm
pyramid	B-Algorithm
transform	O
and	O
the	O
premier	O
ones	O
,	O
discrete	O
wavelet	B-Algorithm
transform	I-Algorithm
,	O
shift-invariant	O
wavelet	B-Algorithm
transform	I-Algorithm
(	O
SIDWT	O
)	O
,	O
and	O
discrete	B-General_Concept
cosine	I-General_Concept
harmonic	O
wavelet	B-Algorithm
transform	I-Algorithm
(	O
DCHWT	O
)	O
are	O
some	O
examples	O
of	O
image	B-General_Concept
fusion	I-General_Concept
methods	O
based	O
on	O
multi-scale	O
transform	O
.	O
</s>
<s>
For	O
example	O
,	O
multi-focus	B-Algorithm
image	I-Algorithm
fusion	I-Algorithm
methods	O
based	O
on	O
DWT	O
require	O
a	O
lot	O
of	O
convolution	B-Language
operations	O
,	O
so	O
they	O
take	O
more	O
time	O
and	O
energy	O
to	O
process	O
.	O
</s>
<s>
Moreover	O
,	O
these	O
methods	O
are	O
not	O
very	O
successful	O
along	O
edges	O
,	O
due	O
to	O
the	O
wavelet	B-Algorithm
transform	I-Algorithm
process	O
missing	O
the	O
edges	O
of	O
the	O
image	O
.	O
</s>
<s>
Due	O
to	O
the	O
aforementioned	O
problems	O
in	O
the	O
multi-scale	O
transform	O
methods	O
,	O
researchers	O
are	O
interested	O
in	O
multi-focus	B-Algorithm
image	I-Algorithm
fusion	I-Algorithm
in	O
the	O
DCT	B-General_Concept
domain	O
.	O
</s>
<s>
DCT-based	O
methods	O
are	O
more	O
efficient	O
in	O
terms	O
of	O
transmission	O
and	O
archiving	O
images	O
coded	O
in	O
Joint	O
Photographic	O
Experts	O
Group	O
(	O
JPEG	O
)	O
standard	O
to	O
the	O
upper	O
node	O
in	O
the	O
VSN	O
agent	O
.	O
</s>
<s>
In	O
the	O
encoder	O
,	O
images	O
are	O
divided	O
into	O
non-overlapping	O
8×8	O
blocks	O
,	O
and	O
the	O
DCT	B-General_Concept
coefficients	O
are	O
calculated	O
for	O
each	O
.	O
</s>
<s>
Since	O
the	O
quantization	O
of	O
DCT	B-General_Concept
coefficients	O
is	O
a	O
lossy	B-Algorithm
process	O
,	O
many	O
of	O
the	O
small-valued	O
DCT	B-General_Concept
coefficients	O
are	O
quantized	O
to	O
zero	O
,	O
which	O
corresponds	O
to	O
high	O
frequencies	O
.	O
</s>
<s>
DCT-based	O
image	B-General_Concept
fusion	I-General_Concept
algorithms	O
work	O
better	O
when	O
the	O
multi-focus	B-Algorithm
image	I-Algorithm
fusion	I-Algorithm
methods	O
are	O
applied	O
in	O
the	O
compressed	O
domain	O
.	O
</s>
<s>
After	O
implementation	O
of	O
the	O
image	B-General_Concept
fusion	I-General_Concept
operations	O
,	O
the	O
output	O
fused	O
images	O
must	O
again	O
be	O
encoded	O
.	O
</s>
<s>
DCT	B-General_Concept
domain-based	O
methods	O
do	O
not	O
require	O
complex	O
and	O
time-consuming	O
consecutive	O
decoding	O
and	O
encoding	O
operations	O
.	O
</s>
<s>
Therefore	O
,	O
the	O
image	B-General_Concept
fusion	I-General_Concept
methods	O
based	O
on	O
DCT	B-General_Concept
domain	O
operate	O
with	O
much	O
less	O
energy	O
and	O
processing	O
time	O
.	O
</s>
<s>
Recently	O
,	O
a	O
lot	O
of	O
research	O
has	O
been	O
carried	O
out	O
in	O
the	O
DCT	B-General_Concept
domain	O
.	O
</s>
<s>
DCT+Variance	O
,	O
DCT+Corr_Eng	O
,	O
DCT+EOL	O
,	O
and	O
DCT+VOL	O
are	O
some	O
prominent	O
examples	O
of	O
DCT	B-General_Concept
based	O
methods	O
.	O
</s>
<s>
Nowadays	O
,	O
the	O
deep	B-Algorithm
learning	I-Algorithm
is	O
utilized	O
in	O
image	B-General_Concept
fusion	I-General_Concept
applications	O
such	O
as	O
multi-focus	B-Algorithm
image	I-Algorithm
fusion	I-Algorithm
.	O
</s>
<s>
were	O
the	O
first	O
researchers	O
that	O
used	O
CNN	B-Architecture
for	O
multi-focus	B-Algorithm
image	I-Algorithm
fusion	I-Algorithm
.	O
</s>
<s>
submitted	O
MSCNN	O
method	O
that	O
obtains	O
the	O
initial	O
segmented	O
decision	O
map	O
with	O
image	B-Algorithm
segmentation	I-Algorithm
between	O
the	O
focused	O
and	O
unfocused	O
patches	O
through	O
the	O
multi-scale	O
convolution	B-Language
neural	B-Architecture
network	I-Architecture
.	O
</s>
<s>
introduced	O
the	O
pixel-wise	O
convolution	B-Language
neural	B-Architecture
network	I-Architecture
(	O
p-CNN	O
)	O
for	O
classification	O
of	O
the	O
focused	O
and	O
unfocused	O
patches	O
.	O
</s>
<s>
All	O
of	O
these	O
CNN	B-Architecture
based	O
multi-focus	B-Algorithm
image	I-Algorithm
fusion	I-Algorithm
methods	O
have	O
enhanced	O
the	O
decision	O
map	O
.	O
</s>
<s>
Therefore	O
,	O
satisfaction	O
of	O
their	O
final	O
fusion	O
decision	O
map	O
depends	O
to	O
use	O
vast	O
post-processing	O
algorithms	O
such	O
as	O
Consistency	O
Verification	O
(	O
CV	O
)	O
,	O
morphological	B-Algorithm
operations	I-Algorithm
,	O
watershed	O
,	O
guiding	O
filters	O
,	O
and	O
small	O
region	O
removal	O
on	O
the	O
initial	O
segmented	O
decision	O
map	O
.	O
</s>
<s>
Along	O
with	O
the	O
CNN	B-Architecture
based	O
multi-focus	B-Algorithm
image	I-Algorithm
fusion	I-Algorithm
methods	O
,	O
fully	O
convolutional	B-Architecture
network	O
(	O
FCN	O
)	O
is	O
also	O
utilized	O
in	O
multi-focus	B-Algorithm
image	I-Algorithm
fusion	I-Algorithm
.	O
</s>
<s>
The	O
Convolutional	B-Architecture
Neural	I-Architecture
Networks	I-Architecture
(	O
CNNs	B-Architecture
)	O
based	O
multi-focus	B-Algorithm
image	I-Algorithm
fusion	I-Algorithm
methods	O
have	O
recently	O
attracted	O
enormous	O
attention	O
.	O
</s>
<s>
In	O
the	O
method	O
of	O
ECNN	O
,	O
a	O
novel	O
CNNs	B-Architecture
based	O
method	O
with	O
the	O
help	O
of	O
the	O
ensemble	B-Algorithm
learning	I-Algorithm
is	O
proposed	O
.	O
</s>
<s>
It	O
is	O
very	O
reasonable	O
to	O
use	O
various	O
models	O
and	O
datasets	B-General_Concept
rather	O
than	O
just	O
one	O
.	O
</s>
<s>
The	O
ensemble	B-Algorithm
learning	I-Algorithm
based	O
methods	O
intend	O
to	O
pursue	O
increasing	O
diversity	O
among	O
the	O
models	O
and	O
datasets	B-General_Concept
in	O
order	O
to	O
decrease	O
the	O
problem	O
of	O
the	O
overfitting	O
on	O
the	O
training	O
dataset	B-General_Concept
.	O
</s>
<s>
It	O
is	O
obvious	O
that	O
the	O
results	O
of	O
an	O
ensemble	O
of	O
CNNs	B-Architecture
are	O
better	O
than	O
just	O
one	O
single	O
CNNs	B-Architecture
.	O
</s>
<s>
Also	O
,	O
the	O
proposed	O
method	O
introduces	O
a	O
new	O
simple	O
type	O
of	O
multi-focus	O
images	O
dataset	B-General_Concept
.	O
</s>
<s>
It	O
simply	O
changes	O
the	O
arranging	O
of	O
the	O
patches	O
of	O
the	O
multi-focus	O
datasets	B-General_Concept
,	O
which	O
is	O
very	O
useful	O
for	O
obtaining	O
the	O
better	O
accuracy	O
.	O
</s>
<s>
With	O
this	O
new	O
type	O
arrangement	O
of	O
datasets	B-General_Concept
,	O
the	O
three	O
different	O
datasets	B-General_Concept
including	O
the	O
original	O
and	O
the	O
Gradient	O
in	O
directions	O
of	O
vertical	O
and	O
horizontal	O
patches	O
are	O
generated	O
from	O
the	O
COCO	O
dataset	B-General_Concept
.	O
</s>
<s>
Therefore	O
,	O
the	O
proposed	O
method	O
introduces	O
a	O
new	O
network	O
that	O
three	O
CNNs	B-Architecture
models	O
which	O
have	O
been	O
trained	O
on	O
three	O
different	O
created	O
datasets	B-General_Concept
to	O
construct	O
the	O
initial	O
segmented	O
decision	O
map	O
.	O
</s>
<s>
These	O
ideas	O
greatly	O
improve	O
the	O
initial	O
segmented	O
decision	O
map	O
of	O
the	O
proposed	O
method	O
which	O
is	O
similar	O
,	O
or	O
even	O
better	O
than	O
,	O
the	O
other	O
final	O
decision	O
map	O
of	O
CNNs	B-Architecture
based	O
methods	O
obtained	O
after	O
applying	O
many	O
post-processing	O
algorithms	O
.	O
</s>
<s>
The	O
obtained	O
experimental	O
results	O
indicate	O
that	O
the	O
proposed	O
CNNs	B-Architecture
based	O
network	O
is	O
more	O
accurate	O
and	O
have	O
the	O
better	O
decision	O
map	O
without	O
post-processing	O
algorithms	O
than	O
the	O
other	O
existing	O
state	O
of	O
the	O
art	O
multi-focus	O
fusion	O
methods	O
which	O
used	O
many	O
post-processing	O
algorithms	O
.	O
</s>
<s>
The	O
pro	O
-	O
posed	O
method	O
introduces	O
a	O
new	O
architecture	O
which	O
uses	O
an	O
ensemble	O
of	O
three	O
CNNs	B-Architecture
trained	O
on	O
three	O
different	O
datasets	B-General_Concept
.	O
</s>
<s>
Also	O
,	O
the	O
proposed	O
method	O
prepares	O
a	O
new	O
simple	O
type	O
of	O
multi	O
-	O
focus	O
image	O
datasets	B-General_Concept
for	O
achieving	O
the	O
better	O
fusion	O
performance	O
than	O
the	O
other	O
popular	O
multi-focus	O
image	O
datasets	B-General_Concept
.	O
</s>
