<s>
In	O
digital	B-Algorithm
image	I-Algorithm
processing	I-Algorithm
and	O
computer	B-Application
vision	I-Application
,	O
image	B-Algorithm
segmentation	I-Algorithm
is	O
the	O
process	O
of	O
partitioning	O
a	O
digital	B-Algorithm
image	I-Algorithm
into	O
multiple	O
image	B-Algorithm
segments	I-Algorithm
,	O
also	O
known	O
as	O
image	B-Algorithm
regions	I-Algorithm
or	O
image	O
objects	O
(	O
sets	O
of	O
pixels	B-Algorithm
)	O
.	O
</s>
<s>
The	O
goal	O
of	O
segmentation	B-Algorithm
is	O
to	O
simplify	O
and/or	O
change	O
the	O
representation	O
of	O
an	O
image	O
into	O
something	O
that	O
is	O
more	O
meaningful	O
and	O
easier	O
to	O
analyze	O
.	O
</s>
<s>
Image	B-Algorithm
segmentation	I-Algorithm
is	O
typically	O
used	O
to	O
locate	O
objects	O
and	O
boundaries	B-Algorithm
(	O
lines	O
,	O
curves	O
,	O
etc	O
.	O
)	O
</s>
<s>
More	O
precisely	O
,	O
image	B-Algorithm
segmentation	I-Algorithm
is	O
the	O
process	O
of	O
assigning	O
a	O
label	O
to	O
every	O
pixel	B-Algorithm
in	O
an	O
image	O
such	O
that	O
pixels	B-Algorithm
with	O
the	O
same	O
label	O
share	O
certain	O
characteristics	O
.	O
</s>
<s>
The	O
result	O
of	O
image	B-Algorithm
segmentation	I-Algorithm
is	O
a	O
set	O
of	O
segments	O
that	O
collectively	O
cover	O
the	O
entire	O
image	O
,	O
or	O
a	O
set	O
of	O
contours	O
extracted	O
from	O
the	O
image	O
(	O
see	O
edge	B-Algorithm
detection	I-Algorithm
)	O
.	O
</s>
<s>
,	O
such	O
as	O
color	B-Language
,	O
intensity	O
,	O
or	O
texture	O
.	O
</s>
<s>
Adjacent	O
regions	O
are	O
significantly	O
different	O
color	B-Language
respect	O
to	O
the	O
same	O
characteristic(s )	O
.	O
</s>
<s>
When	O
applied	O
to	O
a	O
stack	O
of	O
images	O
,	O
typical	O
in	O
medical	B-Application
imaging	I-Application
,	O
the	O
resulting	O
contours	O
after	O
image	B-Algorithm
segmentation	I-Algorithm
can	O
be	O
used	O
to	O
create	O
3D	B-Algorithm
reconstructions	I-Algorithm
with	O
the	O
help	O
of	O
interpolation	O
algorithms	O
like	O
marching	B-Algorithm
cubes	I-Algorithm
.	O
</s>
<s>
Some	O
of	O
the	O
practical	O
applications	O
of	O
image	B-Algorithm
segmentation	I-Algorithm
are	O
:	O
</s>
<s>
Medical	B-Application
imaging	I-Application
,	O
including	O
volume	O
rendered	O
images	O
from	O
computed	O
tomography	O
and	O
magnetic	O
resonance	O
imaging	O
.	O
</s>
<s>
Several	O
general-purpose	O
algorithms	O
and	O
techniques	O
have	O
been	O
developed	O
for	O
image	B-Algorithm
segmentation	I-Algorithm
.	O
</s>
<s>
To	O
be	O
useful	O
,	O
these	O
techniques	O
must	O
typically	O
be	O
combined	O
with	O
a	O
domain	O
's	O
specific	O
knowledge	O
in	O
order	O
to	O
effectively	O
solve	O
the	O
domain	O
's	O
segmentation	B-Algorithm
problems	O
.	O
</s>
<s>
There	O
are	O
two	O
classes	O
of	O
segmentation	B-Algorithm
techniques	O
.	O
</s>
<s>
Semantic	B-Algorithm
segmentation	I-Algorithm
is	O
an	O
approach	O
detecting	O
,	O
for	O
every	O
pixel	B-Algorithm
,	O
belonging	O
class	O
of	O
the	O
object	O
.	O
</s>
<s>
Instance	O
segmentation	B-Algorithm
is	O
an	O
approach	O
that	O
identifies	O
,	O
for	O
every	O
pixel	B-Algorithm
,	O
a	O
belonging	O
instance	O
of	O
the	O
object	O
.	O
</s>
<s>
Panoptic	O
segmentation	B-Algorithm
combines	O
semantic	O
and	O
instance	O
segmentation	B-Algorithm
.	O
</s>
<s>
Like	O
semantic	B-Algorithm
segmentation	I-Algorithm
,	O
panoptic	O
segmentation	B-Algorithm
is	O
an	O
approach	O
that	O
identifies	O
,	O
for	O
every	O
pixel	B-Algorithm
,	O
the	O
belonging	O
class	O
.	O
</s>
<s>
Unlike	O
semantic	B-Algorithm
segmentation	I-Algorithm
,	O
panoptic	O
segmentation	B-Algorithm
distinguishes	O
different	O
instances	O
of	O
the	O
same	O
class	O
.	O
</s>
<s>
The	O
simplest	O
method	O
of	O
image	B-Algorithm
segmentation	I-Algorithm
is	O
called	O
the	O
thresholding	B-Algorithm
method	O
.	O
</s>
<s>
Several	O
popular	O
methods	O
are	O
used	O
in	O
industry	O
including	O
the	O
maximum	O
entropy	O
method	O
,	O
balanced	B-Algorithm
histogram	I-Algorithm
thresholding	I-Algorithm
,	O
Otsu	B-Algorithm
's	I-Algorithm
method	I-Algorithm
(	O
maximum	O
variance	O
)	O
,	O
and	O
k-means	B-Algorithm
clustering	I-Algorithm
.	O
</s>
<s>
Recently	O
,	O
methods	O
have	O
been	O
developed	O
for	O
thresholding	B-Algorithm
computed	O
tomography	O
(	O
CT	O
)	O
images	O
.	O
</s>
<s>
The	O
key	O
idea	O
is	O
that	O
,	O
unlike	O
Otsu	B-Algorithm
's	I-Algorithm
method	I-Algorithm
,	O
the	O
thresholds	O
are	O
derived	O
from	O
the	O
radiographs	O
instead	O
of	O
the	O
(	O
reconstructed	O
)	O
image	O
.	O
</s>
<s>
In	O
these	O
works	O
decision	O
over	O
each	O
pixel	B-Algorithm
's	O
membership	O
to	O
a	O
segment	O
is	O
based	O
on	O
multi-dimensional	O
rules	O
derived	O
from	O
fuzzy	O
logic	O
and	O
evolutionary	O
algorithms	O
based	O
on	O
image	O
lighting	O
environment	O
and	O
application	O
.	O
</s>
<s>
The	O
K-means	B-Algorithm
algorithm	I-Algorithm
is	O
an	O
iterative	B-Algorithm
technique	O
that	O
is	O
used	O
to	O
partition	B-Algorithm
an	I-Algorithm
image	I-Algorithm
into	O
K	O
clusters	B-Algorithm
.	O
</s>
<s>
In	O
this	O
case	O
,	O
distance	O
is	O
the	O
squared	O
or	O
absolute	O
difference	O
between	O
a	O
pixel	B-Algorithm
and	O
a	O
cluster	O
center	O
.	O
</s>
<s>
The	O
difference	O
is	O
typically	O
based	O
on	O
pixel	B-Algorithm
color	B-Language
,	O
intensity	O
,	O
texture	O
,	O
and	O
location	O
,	O
or	O
a	O
weighted	O
combination	O
of	O
these	O
factors	O
.	O
</s>
<s>
K	O
can	O
be	O
selected	O
manually	O
,	O
randomly	O
,	O
or	O
by	O
a	O
heuristic	B-Algorithm
.	O
</s>
<s>
The	O
quality	O
of	O
the	O
solution	O
depends	O
on	O
the	O
initial	O
set	O
of	O
clusters	B-Algorithm
and	O
the	O
value	O
of	O
K	O
.	O
</s>
<s>
The	O
Mean	B-Algorithm
Shift	I-Algorithm
algorithm	O
is	O
a	O
technique	O
that	O
is	O
used	O
to	O
partition	B-Algorithm
an	I-Algorithm
image	I-Algorithm
into	O
an	O
unknown	O
apriori	O
no	O
.	O
</s>
<s>
of	O
clusters	B-Algorithm
.	O
</s>
<s>
Motion	O
based	O
segmentation	B-Algorithm
is	O
a	O
technique	O
that	O
relies	O
on	O
motion	O
in	O
the	O
image	O
to	O
perform	O
segmentation	B-Algorithm
.	O
</s>
<s>
proposed	O
interactive	O
segmentation	B-Algorithm
.	O
</s>
<s>
They	O
use	O
a	O
robot	O
to	O
poke	O
objects	O
in	O
order	O
to	O
generate	O
the	O
motion	O
signal	O
necessary	O
for	O
motion-based	O
segmentation	B-Algorithm
.	O
</s>
<s>
Interactive	O
segmentation	B-Algorithm
follows	O
the	O
interactive	O
perception	O
framework	O
proposed	O
by	O
Dov	O
Katz	O
and	O
Oliver	O
Brock	O
.	O
</s>
<s>
Another	O
technique	O
that	O
is	O
based	O
on	O
motion	O
is	O
rigid	B-Algorithm
motion	I-Algorithm
segmentation	I-Algorithm
.	O
</s>
<s>
Compression	O
based	O
methods	O
postulate	O
that	O
the	O
optimal	O
segmentation	B-Algorithm
is	O
the	O
one	O
that	O
minimizes	O
,	O
over	O
all	O
possible	O
segmentations	B-Algorithm
,	O
the	O
coding	O
length	O
of	O
the	O
data	O
.	O
</s>
<s>
The	O
connection	O
between	O
these	O
two	O
concepts	O
is	O
that	O
segmentation	B-Algorithm
tries	O
to	O
find	O
patterns	O
in	O
an	O
image	O
and	O
any	O
regularity	O
in	O
the	O
image	O
can	O
be	O
used	O
to	O
compress	O
it	O
.	O
</s>
<s>
This	O
prior	O
is	O
used	O
by	O
Huffman	B-General_Concept
coding	I-General_Concept
to	O
encode	O
the	O
difference	O
chain	B-Algorithm
code	I-Algorithm
of	O
the	O
contours	O
in	O
an	O
image	O
.	O
</s>
<s>
Texture	O
is	O
encoded	O
by	O
lossy	B-Algorithm
compression	I-Algorithm
in	O
a	O
way	O
similar	O
to	O
minimum	O
description	O
length	O
(	O
MDL	O
)	O
principle	O
,	O
but	O
here	O
the	O
length	O
of	O
the	O
data	O
given	O
the	O
model	O
is	O
approximated	O
by	O
the	O
number	O
of	O
samples	O
times	O
the	O
entropy	O
of	O
the	O
model	O
.	O
</s>
<s>
For	O
any	O
given	O
segmentation	B-Algorithm
of	O
an	O
image	O
,	O
this	O
scheme	O
yields	O
the	O
number	O
of	O
bits	O
required	O
to	O
encode	O
that	O
image	O
based	O
on	O
the	O
given	O
segmentation	B-Algorithm
.	O
</s>
<s>
Thus	O
,	O
among	O
all	O
possible	O
segmentations	B-Algorithm
of	O
an	O
image	O
,	O
the	O
goal	O
is	O
to	O
find	O
the	O
segmentation	B-Algorithm
which	O
produces	O
the	O
shortest	O
coding	O
length	O
.	O
</s>
<s>
The	O
distortion	O
in	O
the	O
lossy	B-Algorithm
compression	I-Algorithm
determines	O
the	O
coarseness	O
of	O
the	O
segmentation	B-Algorithm
and	O
its	O
optimal	O
value	O
may	O
differ	O
for	O
each	O
image	O
.	O
</s>
<s>
For	O
example	O
,	O
when	O
the	O
textures	O
in	O
an	O
image	O
are	O
similar	O
,	O
such	O
as	O
in	O
camouflage	O
images	O
,	O
stronger	O
sensitivity	O
and	O
thus	O
lower	O
quantization	B-Algorithm
is	O
required	O
.	O
</s>
<s>
Histogram-based	O
methods	O
are	O
very	O
efficient	O
compared	O
to	O
other	O
image	B-Algorithm
segmentation	I-Algorithm
methods	O
because	O
they	O
typically	O
require	O
only	O
one	O
pass	O
through	O
the	O
pixels	B-Algorithm
.	O
</s>
<s>
In	O
this	O
technique	O
,	O
a	O
histogram	B-Algorithm
is	O
computed	O
from	O
all	O
of	O
the	O
pixels	B-Algorithm
in	O
the	O
image	O
,	O
and	O
the	O
peaks	O
and	O
valleys	O
in	O
the	O
histogram	B-Algorithm
are	O
used	O
to	O
locate	O
the	O
clusters	B-Algorithm
in	O
the	O
image	O
.	O
</s>
<s>
Color	B-Language
or	O
intensity	O
can	O
be	O
used	O
as	O
the	O
measure	O
.	O
</s>
<s>
A	O
refinement	O
of	O
this	O
technique	O
is	O
to	O
recursively	O
apply	O
the	O
histogram-seeking	O
method	O
to	O
clusters	B-Algorithm
in	O
the	O
image	O
in	O
order	O
to	O
divide	O
them	O
into	O
smaller	O
clusters	B-Algorithm
.	O
</s>
<s>
This	O
operation	O
is	O
repeated	O
with	O
smaller	O
and	O
smaller	O
clusters	B-Algorithm
until	O
no	O
more	O
clusters	B-Algorithm
are	O
formed	O
.	O
</s>
<s>
One	O
disadvantage	O
of	O
the	O
histogram-seeking	O
method	O
is	O
that	O
it	O
may	O
be	O
difficult	O
to	O
identify	O
significant	O
peaks	O
and	O
valleys	O
in	O
the	O
image	O
.	O
</s>
<s>
Histogram-based	O
approaches	O
can	O
also	O
be	O
quickly	O
adapted	O
to	O
apply	O
to	O
multiple	O
frames	O
,	O
while	O
maintaining	O
their	O
single	O
pass	O
efficiency	O
.	O
</s>
<s>
The	O
histogram	B-Algorithm
can	O
be	O
done	O
in	O
multiple	O
fashions	O
when	O
multiple	O
frames	O
are	O
considered	O
.	O
</s>
<s>
The	O
histogram	B-Algorithm
can	O
also	O
be	O
applied	O
on	O
a	O
per-pixel	O
basis	O
where	O
the	O
resulting	O
information	O
is	O
used	O
to	O
determine	O
the	O
most	O
frequent	O
color	B-Language
for	O
the	O
pixel	B-Algorithm
location	O
.	O
</s>
<s>
This	O
approach	O
segments	O
based	O
on	O
active	O
objects	O
and	O
a	O
static	O
environment	O
,	O
resulting	O
in	O
a	O
different	O
type	O
of	O
segmentation	B-Algorithm
useful	O
in	O
video	B-Operating_System
tracking	I-Operating_System
.	O
</s>
<s>
Edge	B-Algorithm
detection	I-Algorithm
is	O
a	O
well-developed	O
field	O
on	O
its	O
own	O
within	O
image	B-Algorithm
processing	I-Algorithm
.	O
</s>
<s>
Region	O
boundaries	B-Algorithm
and	O
edges	O
are	O
closely	O
related	O
,	O
</s>
<s>
since	O
there	O
is	O
often	O
a	O
sharp	O
adjustment	O
in	O
intensity	O
at	O
the	O
region	O
boundaries	B-Algorithm
.	O
</s>
<s>
Edge	B-Algorithm
detection	I-Algorithm
techniques	O
have	O
therefore	O
been	O
used	O
as	O
the	O
base	O
of	O
another	O
segmentation	B-Algorithm
technique	O
.	O
</s>
<s>
The	O
edges	O
identified	O
by	O
edge	B-Algorithm
detection	I-Algorithm
are	O
often	O
disconnected	O
.	O
</s>
<s>
To	O
segment	O
an	O
object	O
from	O
an	O
image	O
however	O
,	O
one	O
needs	O
closed	O
region	O
boundaries	B-Algorithm
.	O
</s>
<s>
The	O
desired	O
edges	O
are	O
the	O
boundaries	B-Algorithm
between	O
such	O
objects	O
or	O
spatial-taxons	O
.	O
</s>
<s>
Spatial-taxons	O
are	O
information	O
granules	O
,	O
consisting	O
of	O
a	O
crisp	O
pixel	B-Algorithm
region	O
,	O
stationed	O
at	O
abstraction	O
levels	O
within	O
a	O
hierarchical	O
nested	O
scene	O
architecture	O
.	O
</s>
<s>
Edge	B-Algorithm
detection	I-Algorithm
methods	O
can	O
be	O
applied	O
to	O
the	O
spatial-taxon	O
region	O
,	O
in	O
the	O
same	O
manner	O
they	O
would	O
be	O
applied	O
to	O
a	O
silhouette	O
.	O
</s>
<s>
Segmentation	B-Algorithm
methods	O
can	O
also	O
be	O
applied	O
to	O
edges	O
obtained	O
from	O
edge	O
detectors	O
.	O
</s>
<s>
This	O
method	O
is	O
a	O
combination	O
of	O
three	O
characteristics	O
of	O
the	O
image	O
:	O
partition	O
of	O
the	O
image	O
based	O
on	O
histogram	B-Algorithm
analysis	O
is	O
checked	O
by	O
high	O
compactness	O
of	O
the	O
clusters	B-Algorithm
(	O
objects	O
)	O
,	O
and	O
high	O
gradients	O
of	O
their	O
borders	O
.	O
</s>
<s>
For	O
that	O
purpose	O
two	O
spaces	O
have	O
to	O
be	O
introduced	O
:	O
one	O
space	O
is	O
the	O
one-dimensional	O
histogram	B-Algorithm
of	O
brightness	O
H	O
=	O
H(B )	O
;	O
the	O
second	O
space	O
is	O
the	O
dual	O
3-dimensional	O
space	O
of	O
the	O
original	O
image	O
itself	O
B	O
=	O
B(x,y )	O
.	O
</s>
<s>
On	O
that	O
bitmap	O
a	O
measure	O
has	O
to	O
be	O
defined	O
reflecting	O
how	O
compact	O
distributed	O
black	O
(	O
or	O
white	O
)	O
pixels	B-Algorithm
are	O
.	O
</s>
<s>
Maximum	O
of	O
MDC	O
defines	O
the	O
segmentation	B-Algorithm
.	O
</s>
<s>
Region-growing	B-Algorithm
methods	O
rely	O
mainly	O
on	O
the	O
assumption	O
that	O
the	O
neighboring	O
pixels	B-Algorithm
within	O
one	O
region	O
have	O
similar	O
values	O
.	O
</s>
<s>
The	O
common	O
procedure	O
is	O
to	O
compare	O
one	O
pixel	B-Algorithm
with	O
its	O
neighbors	O
.	O
</s>
<s>
If	O
a	O
similarity	O
criterion	O
is	O
satisfied	O
,	O
the	O
pixel	B-Algorithm
can	O
be	O
set	O
to	O
belong	O
to	O
the	O
same	O
cluster	O
as	O
one	O
or	O
more	O
of	O
its	O
neighbors	O
.	O
</s>
<s>
The	O
method	O
of	O
Statistical	O
Region	O
Merging	O
(	O
SRM	O
)	O
starts	O
by	O
building	O
the	O
graph	B-Application
of	O
pixels	B-Algorithm
using	O
4-connectedness	O
with	O
edges	O
weighted	O
by	O
the	O
absolute	O
value	O
of	O
the	O
intensity	O
difference	O
.	O
</s>
<s>
Initially	O
each	O
pixel	B-Algorithm
forms	O
a	O
single	O
pixel	B-Algorithm
region	O
.	O
</s>
<s>
SRM	O
then	O
sorts	O
those	O
edges	O
in	O
a	O
priority	O
queue	O
and	O
decides	O
whether	O
or	O
not	O
to	O
merge	O
the	O
current	O
regions	O
belonging	O
to	O
the	O
edge	O
pixels	B-Algorithm
using	O
a	O
statistical	O
predicate	O
.	O
</s>
<s>
One	O
region-growing	B-Algorithm
method	O
is	O
the	O
seeded	O
region	B-Algorithm
growing	I-Algorithm
method	O
.	O
</s>
<s>
The	O
regions	O
are	O
iteratively	O
grown	O
by	O
comparison	O
of	O
all	O
unallocated	O
neighboring	O
pixels	B-Algorithm
to	O
the	O
regions	O
.	O
</s>
<s>
The	O
difference	O
between	O
a	O
pixel	B-Algorithm
's	O
intensity	O
value	O
and	O
the	O
region	O
's	O
mean	O
,	O
,	O
is	O
used	O
as	O
a	O
measure	O
of	O
similarity	O
.	O
</s>
<s>
The	O
pixel	B-Algorithm
with	O
the	O
smallest	O
difference	O
measured	O
in	O
this	O
way	O
is	O
assigned	O
to	O
the	O
respective	O
region	O
.	O
</s>
<s>
This	O
process	O
continues	O
until	O
all	O
pixels	B-Algorithm
are	O
assigned	O
to	O
a	O
region	O
.	O
</s>
<s>
Because	O
seeded	O
region	B-Algorithm
growing	I-Algorithm
requires	O
seeds	O
as	O
additional	O
input	O
,	O
the	O
segmentation	B-Algorithm
results	O
are	O
dependent	O
on	O
the	O
choice	O
of	O
seeds	O
,	O
and	O
noise	O
in	O
the	O
image	O
can	O
cause	O
the	O
seeds	O
to	O
be	O
poorly	O
placed	O
.	O
</s>
<s>
Another	O
region-growing	B-Algorithm
method	O
is	O
the	O
unseeded	O
region	B-Algorithm
growing	I-Algorithm
method	O
.	O
</s>
<s>
It	O
starts	O
with	O
a	O
single	O
region	O
—	O
the	O
pixel	B-Algorithm
chosen	O
here	O
does	O
not	O
markedly	O
influence	O
the	O
final	O
segmentation	B-Algorithm
.	O
</s>
<s>
At	O
each	O
iteration	B-Algorithm
it	O
considers	O
the	O
neighboring	O
pixels	B-Algorithm
in	O
the	O
same	O
way	O
as	O
seeded	O
region	B-Algorithm
growing	I-Algorithm
.	O
</s>
<s>
It	O
differs	O
from	O
seeded	O
region	B-Algorithm
growing	I-Algorithm
in	O
that	O
if	O
the	O
minimum	O
is	O
less	O
than	O
a	O
predefined	O
threshold	O
then	O
it	O
is	O
added	O
to	O
the	O
respective	O
region	O
.	O
</s>
<s>
If	O
not	O
,	O
then	O
the	O
pixel	B-Algorithm
is	O
considered	O
different	O
from	O
all	O
current	O
regions	O
and	O
a	O
new	O
region	O
is	O
created	O
with	O
this	O
pixel	B-Algorithm
.	O
</s>
<s>
One	O
variant	O
of	O
this	O
technique	O
,	O
proposed	O
by	O
Haralick	O
and	O
Shapiro	O
(	O
1985	O
)	O
,	O
is	O
based	O
on	O
pixel	B-Algorithm
intensities	O
.	O
</s>
<s>
The	O
mean	O
and	O
scatter	O
of	O
the	O
region	O
and	O
the	O
intensity	O
of	O
the	O
candidate	O
pixel	B-Algorithm
are	O
used	O
to	O
compute	O
a	O
test	O
statistic	O
.	O
</s>
<s>
If	O
the	O
test	O
statistic	O
is	O
sufficiently	O
small	O
,	O
the	O
pixel	B-Algorithm
is	O
added	O
to	O
the	O
region	O
,	O
and	O
the	O
region	O
’s	O
mean	O
and	O
scatter	O
are	O
recomputed	O
.	O
</s>
<s>
Otherwise	O
,	O
the	O
pixel	B-Algorithm
is	O
rejected	O
,	O
and	O
is	O
used	O
to	O
form	O
a	O
new	O
region	O
.	O
</s>
<s>
A	O
special	O
region-growing	B-Algorithm
method	O
is	O
called	O
-connected	O
segmentation	B-Algorithm
(	O
see	O
also	O
lambda-connectedness	B-Algorithm
)	O
.	O
</s>
<s>
It	O
is	O
based	O
on	O
pixel	B-Algorithm
intensities	O
and	O
neighborhood-linking	O
paths	O
.	O
</s>
<s>
A	O
degree	O
of	O
connectivity	O
(	O
connectedness	O
)	O
is	O
calculated	O
based	O
on	O
a	O
path	O
that	O
is	O
formed	O
by	O
pixels	B-Algorithm
.	O
</s>
<s>
For	O
a	O
certain	O
value	O
of	O
,	O
two	O
pixels	B-Algorithm
are	O
called	O
-connected	O
if	O
there	O
is	O
a	O
path	O
linking	O
those	O
two	O
pixels	B-Algorithm
and	O
the	O
connectedness	O
of	O
this	O
path	O
is	O
at	O
least	O
.	O
</s>
<s>
Split-and-merge	B-Algorithm
segmentation	I-Algorithm
is	O
based	O
on	O
a	O
quadtree	B-Data_Structure
partition	O
of	O
an	O
image	O
.	O
</s>
<s>
It	O
is	O
sometimes	O
called	O
quadtree	B-Data_Structure
segmentation	B-Algorithm
.	O
</s>
<s>
Curve	O
propagation	O
is	O
a	O
popular	O
technique	O
in	O
this	O
category	O
,	O
with	O
numerous	O
applications	O
to	O
object	O
extraction	O
,	O
object	O
tracking	O
,	O
stereo	B-Algorithm
reconstruction	I-Algorithm
,	O
etc	O
.	O
</s>
<s>
Such	O
techniques	O
are	O
fast	O
and	O
efficient	O
,	O
however	O
the	O
original	O
"	O
purely	O
parametric	O
"	O
formulation	O
(	O
due	O
to	O
Kass	O
,	O
Witkin	O
and	O
Terzopoulos	O
in	O
1987	O
and	O
known	O
as	O
"	O
snakes	B-General_Concept
"	O
)	O
,	O
is	O
generally	O
criticized	O
for	O
its	O
limitations	O
regarding	O
the	O
choice	O
of	O
sampling	O
strategy	O
,	O
the	O
internal	O
geometric	O
properties	O
of	O
the	O
curve	O
,	O
topology	O
changes	O
(	O
curve	O
splitting	O
and	O
merging	O
)	O
,	O
addressing	O
problems	O
in	O
higher	O
dimensions	O
,	O
etc	O
..	O
Nowadays	O
,	O
efficient	O
"	O
discretized	O
"	O
formulations	O
have	O
been	O
developed	O
to	O
address	O
these	O
limitations	O
while	O
maintaining	O
high	O
efficiency	O
.	O
</s>
<s>
The	O
level-set	B-Algorithm
method	I-Algorithm
was	O
initially	O
proposed	O
to	O
track	O
moving	O
interfaces	O
by	O
Dervieux	O
and	O
Thomasset	O
in	O
1979	O
and	O
1981	O
and	O
was	O
later	O
reinvented	O
by	O
Osher	O
and	O
Sethian	O
in	O
1988	O
.	O
</s>
<s>
The	O
level-set	B-Algorithm
method	I-Algorithm
affords	O
numerous	O
advantages	O
:	O
it	O
is	O
implicit	O
,	O
is	O
parameter-free	O
,	O
provides	O
a	O
direct	O
way	O
to	O
estimate	O
the	O
geometric	O
properties	O
of	O
the	O
evolving	O
structure	O
,	O
allows	O
for	O
change	O
of	O
topology	O
,	O
and	O
is	O
intrinsic	O
.	O
</s>
<s>
One	O
can	O
conclude	O
that	O
it	O
is	O
a	O
very	O
convenient	O
framework	O
for	O
addressing	O
numerous	O
applications	O
of	O
computer	B-Application
vision	I-Application
and	O
medical	B-Application
image	I-Application
analysis	O
.	O
</s>
<s>
Research	O
into	O
various	O
level-set	B-Algorithm
data	I-Algorithm
structures	I-Algorithm
has	O
led	O
to	O
very	O
efficient	O
implementations	O
of	O
this	O
method	O
.	O
</s>
<s>
The	O
fast	B-Algorithm
marching	I-Algorithm
method	I-Algorithm
has	O
been	O
used	O
in	O
image	B-Algorithm
segmentation	I-Algorithm
,	O
and	O
this	O
model	O
has	O
been	O
improved	O
(	O
permitting	O
both	O
positive	O
and	O
negative	O
propagation	O
speeds	O
)	O
in	O
an	O
approach	O
called	O
the	O
generalized	O
fast	B-Algorithm
marching	I-Algorithm
method	I-Algorithm
.	O
</s>
<s>
The	O
jump	O
set	O
of	O
defines	O
a	O
segmentation	B-Algorithm
.	O
</s>
<s>
The	O
functional	O
value	O
is	O
the	O
sum	O
of	O
the	O
total	O
length	O
of	O
the	O
segmentation	B-Algorithm
curve	O
,	O
the	O
smoothness	O
of	O
the	O
approximation	O
,	O
and	O
its	O
distance	O
to	O
the	O
original	O
image	O
.	O
</s>
<s>
The	O
Potts	O
model	O
is	O
often	O
called	O
piecewise	O
constant	O
Mumford-Shah	B-Algorithm
model	I-Algorithm
as	O
it	O
can	O
be	O
seen	O
as	O
the	O
degenerate	O
case	O
.	O
</s>
<s>
Classical	O
algorithms	O
are	O
graduated	B-Algorithm
non-convexity	I-Algorithm
and	O
Ambrosio-Tortorelli	B-Algorithm
approximation	I-Algorithm
.	O
</s>
<s>
Graph	B-Application
partitioning	O
methods	O
are	O
an	O
effective	O
tools	O
for	O
image	B-Algorithm
segmentation	I-Algorithm
since	O
they	O
model	O
the	O
impact	O
of	O
pixel	B-Algorithm
neighborhoods	O
on	O
a	O
given	O
cluster	O
of	O
pixels	B-Algorithm
or	O
pixel	B-Algorithm
,	O
under	O
the	O
assumption	O
of	O
homogeneity	O
in	O
images	O
.	O
</s>
<s>
In	O
these	O
methods	O
,	O
the	O
image	O
is	O
modeled	O
as	O
a	O
weighted	O
,	O
undirected	O
graph	B-Application
.	O
</s>
<s>
Usually	O
a	O
pixel	B-Algorithm
or	O
a	O
group	O
of	O
pixels	B-Algorithm
are	O
associated	O
with	O
nodes	O
and	O
edge	O
weights	O
define	O
the	O
(	O
dis	O
)	O
similarity	O
between	O
the	O
neighborhood	O
pixels	B-Algorithm
.	O
</s>
<s>
The	O
graph	B-Application
(	O
image	O
)	O
is	O
then	O
partitioned	O
according	O
to	O
a	O
criterion	O
designed	O
to	O
model	O
"	O
good	O
"	O
clusters	B-Algorithm
.	O
</s>
<s>
Each	O
partition	O
of	O
the	O
nodes	O
(	O
pixels	B-Algorithm
)	O
output	O
from	O
these	O
algorithms	O
are	O
considered	O
an	O
object	O
segment	O
in	O
the	O
image	O
;	O
see	O
Segmentation-based	B-Algorithm
object	I-Algorithm
categorization	I-Algorithm
.	O
</s>
<s>
Some	O
popular	O
algorithms	O
of	O
this	O
category	O
are	O
normalized	O
cuts	O
,	O
random	B-Algorithm
walker	I-Algorithm
,	O
minimum	O
cut	O
,	O
isoperimetric	O
partitioning	O
,	O
minimum	B-Algorithm
spanning	I-Algorithm
tree-based	I-Algorithm
segmentation	I-Algorithm
,	O
and	O
segmentation-based	B-Algorithm
object	I-Algorithm
categorization	I-Algorithm
.	O
</s>
<s>
Their	O
strong	O
mathematical	O
foundation	O
and	O
ability	O
to	O
provide	O
a	O
global	O
optimum	O
even	O
when	O
defined	O
on	O
local	O
features	O
proved	O
to	O
be	O
the	O
foundation	O
for	O
novel	O
research	O
in	O
the	O
domain	O
of	O
image	O
analysis	O
,	O
de-noising	O
and	O
segmentation	B-Algorithm
.	O
</s>
<s>
The	O
criterion	O
for	O
image	B-Algorithm
segmentation	I-Algorithm
using	O
MRFs	O
is	O
restated	O
as	O
finding	O
the	O
labelling	O
scheme	O
which	O
has	O
maximum	O
probability	O
for	O
a	O
given	O
set	O
of	O
features	O
.	O
</s>
<s>
The	O
broad	O
categories	O
of	O
image	B-Algorithm
segmentation	I-Algorithm
using	O
MRFs	O
are	O
supervised	O
and	O
unsupervised	O
segmentation	B-Algorithm
.	O
</s>
<s>
In	O
terms	O
of	O
image	B-Algorithm
segmentation	I-Algorithm
,	O
the	O
function	O
that	O
MRFs	O
seek	O
to	O
maximize	O
is	O
the	O
probability	O
of	O
identifying	O
a	O
labelling	O
scheme	O
given	O
a	O
particular	O
set	O
of	O
features	O
are	O
detected	O
in	O
the	O
image	O
.	O
</s>
<s>
This	O
is	O
a	O
restatement	O
of	O
the	O
maximum	B-General_Concept
a	I-General_Concept
posteriori	I-General_Concept
estimation	I-General_Concept
method	O
.	O
</s>
<s>
The	O
generic	O
algorithm	O
for	O
image	B-Algorithm
segmentation	I-Algorithm
using	O
MAP	O
is	O
given	O
below	O
:	O
</s>
<s>
The	O
common	O
trait	O
of	O
cost	O
functions	O
is	O
to	O
penalize	O
change	O
in	O
pixel	B-Algorithm
value	O
as	O
well	O
as	O
difference	O
in	O
pixel	B-Algorithm
label	O
when	O
compared	O
to	O
labels	O
of	O
neighboring	O
pixels	B-Algorithm
.	O
</s>
<s>
The	O
iterated	B-Algorithm
conditional	I-Algorithm
modes	I-Algorithm
(	O
ICM	O
)	O
algorithm	O
tries	O
to	O
reconstruct	O
the	O
ideal	O
labeling	O
scheme	O
by	O
changing	O
the	O
values	O
of	O
each	O
pixel	B-Algorithm
over	O
each	O
iteration	B-Algorithm
and	O
evaluating	O
the	O
energy	O
of	O
the	O
new	O
labeling	O
scheme	O
using	O
the	O
cost	O
function	O
given	O
below	O
,	O
</s>
<s>
neighboring	O
pixels	B-Algorithm
and	O
chosen	O
pixel	B-Algorithm
.	O
</s>
<s>
Here	O
is	O
neighborhood	O
of	O
pixel	B-Algorithm
i	O
and	O
is	O
the	O
Kronecker	O
delta	O
function	O
.	O
</s>
<s>
Derived	O
as	O
an	O
analogue	O
of	O
annealing	O
in	O
metallurgy	O
,	O
simulated	B-Algorithm
annealing	I-Algorithm
(	O
SA	O
)	O
uses	O
change	O
in	O
pixel	B-Algorithm
label	O
over	O
iterations	B-Algorithm
and	O
estimates	O
the	O
difference	O
in	O
energy	O
of	O
each	O
newly	O
formed	O
graph	B-Application
to	O
the	O
initial	O
data	O
.	O
</s>
<s>
If	O
the	O
newly	O
formed	O
graph	B-Application
is	O
more	O
profitable	O
,	O
in	O
terms	O
of	O
low	O
energy	O
cost	O
,	O
given	O
by	O
:	O
</s>
<s>
the	O
algorithm	O
selects	O
the	O
newly	O
formed	O
graph	B-Application
.	O
</s>
<s>
Simulated	B-Algorithm
annealing	I-Algorithm
requires	O
the	O
input	O
of	O
temperature	O
schedules	O
which	O
directly	O
affects	O
the	O
speed	O
of	O
convergence	O
of	O
the	O
system	O
,	O
as	O
well	O
as	O
energy	O
threshold	O
for	O
minimization	O
to	O
occur	O
.	O
</s>
<s>
They	O
include	O
Maximization	O
of	O
Posterior	O
Marginal	O
,	O
Multi-scale	O
MAP	B-General_Concept
estimation	I-General_Concept
,	O
Multiple	O
Resolution	O
segmentation	B-Algorithm
and	O
more	O
.	O
</s>
<s>
Apart	O
from	O
likelihood	O
estimates	O
,	O
graph-cut	O
using	O
maximum	O
flow	O
and	O
other	O
highly	O
constrained	O
graph	B-Application
based	O
methods	O
exist	O
for	O
solving	O
MRFs	O
.	O
</s>
<s>
The	O
expectation	B-Algorithm
–	I-Algorithm
maximization	I-Algorithm
algorithm	I-Algorithm
is	O
utilized	O
to	O
iteratively	O
estimate	O
the	O
a	O
posterior	O
probabilities	O
and	O
distributions	O
of	O
labeling	O
when	O
no	O
training	O
data	O
is	O
available	O
and	O
no	O
estimate	O
of	O
segmentation	B-Algorithm
model	O
can	O
be	O
formed	O
.	O
</s>
<s>
A	O
general	O
approach	O
is	O
to	O
use	O
histograms	B-Algorithm
to	O
represent	O
the	O
features	O
of	O
an	O
image	O
and	O
proceed	O
as	O
outlined	O
briefly	O
in	O
this	O
three-step	O
algorithm	O
:	O
</s>
<s>
E	O
step	O
:	O
Estimate	O
class	O
statistics	O
based	O
on	O
the	O
random	O
segmentation	B-Algorithm
model	O
defined	O
.	O
</s>
<s>
Exact	O
MAP	B-General_Concept
estimates	I-General_Concept
cannot	O
be	O
easily	O
computed	O
.	O
</s>
<s>
Approximate	O
MAP	B-General_Concept
estimates	I-General_Concept
are	O
computationally	O
expensive	O
to	O
calculate	O
.	O
</s>
<s>
Based	O
on	O
method	O
of	O
optimization	O
,	O
segmentation	B-Algorithm
may	O
cluster	O
to	O
local	O
minima	O
.	O
</s>
<s>
The	O
watershed	B-Algorithm
transformation	I-Algorithm
considers	O
the	O
gradient	O
magnitude	O
of	O
an	O
image	O
as	O
a	O
topographic	O
surface	O
.	O
</s>
<s>
Pixels	B-Algorithm
having	O
the	O
highest	O
gradient	O
magnitude	O
intensities	O
(	O
GMIs	O
)	O
correspond	O
to	O
watershed	B-Algorithm
lines	O
,	O
which	O
represent	O
the	O
region	O
boundaries	B-Algorithm
.	O
</s>
<s>
Water	O
placed	O
on	O
any	O
pixel	B-Algorithm
enclosed	O
by	O
a	O
common	O
watershed	B-Algorithm
line	O
flows	O
downhill	O
to	O
a	O
common	O
local	O
intensity	O
minimum	O
(	O
LIM	O
)	O
.	O
</s>
<s>
Pixels	B-Algorithm
draining	O
to	O
a	O
common	O
minimum	O
form	O
a	O
catch	O
basin	O
,	O
which	O
represents	O
a	O
segment	O
.	O
</s>
<s>
Other	O
important	O
methods	O
in	O
the	O
literature	O
for	O
model-based	O
segmentation	B-Algorithm
include	O
active	B-General_Concept
shape	I-General_Concept
models	I-General_Concept
and	O
active	B-General_Concept
appearance	I-General_Concept
models	I-General_Concept
.	O
</s>
<s>
Image	B-Algorithm
segmentations	I-Algorithm
are	O
computed	O
at	O
multiple	O
scales	O
in	O
scale	B-Algorithm
space	I-Algorithm
and	O
sometimes	O
propagated	O
from	O
coarse	O
to	O
fine	O
scales	O
;	O
see	O
scale-space	B-Algorithm
segmentation	I-Algorithm
.	O
</s>
<s>
Segmentation	B-Algorithm
criteria	O
can	O
be	O
arbitrarily	O
complex	O
and	O
may	O
take	O
into	O
account	O
global	O
as	O
well	O
as	O
local	O
criteria	O
.	O
</s>
<s>
Witkin	O
's	O
seminal	O
work	O
in	O
scale	B-Algorithm
space	I-Algorithm
included	O
the	O
notion	O
that	O
a	O
one-dimensional	O
signal	O
could	O
be	O
unambiguously	O
segmented	O
into	O
regions	O
,	O
with	O
one	O
scale	O
parameter	O
controlling	O
the	O
scale	O
of	O
segmentation	B-Algorithm
.	O
</s>
<s>
There	O
have	O
been	O
numerous	O
research	O
works	O
in	O
this	O
area	O
,	O
out	O
of	O
which	O
a	O
few	O
have	O
now	O
reached	O
a	O
state	O
where	O
they	O
can	O
be	O
applied	O
either	O
with	O
interactive	O
manual	O
intervention	O
(	O
usually	O
with	O
application	O
to	O
medical	B-Application
imaging	I-Application
)	O
or	O
fully	O
automatically	O
.	O
</s>
<s>
Nevertheless	O
,	O
this	O
general	O
idea	O
has	O
inspired	O
several	O
other	O
authors	O
to	O
investigate	O
coarse-to-fine	O
schemes	O
for	O
image	B-Algorithm
segmentation	I-Algorithm
.	O
</s>
<s>
Lindeberg	O
studied	O
the	O
problem	O
of	O
linking	O
local	O
extrema	O
and	O
saddle	O
points	O
over	O
scales	O
,	O
and	O
proposed	O
an	O
image	O
representation	O
called	O
the	O
scale-space	B-Algorithm
primal	O
sketch	O
which	O
makes	O
explicit	O
the	O
relations	O
between	O
structures	O
at	O
different	O
scales	O
,	O
and	O
also	O
makes	O
explicit	O
which	O
image	O
features	O
are	O
stable	O
over	O
large	O
ranges	O
of	O
scale	O
including	O
locally	O
appropriate	O
scales	O
for	O
those	O
.	O
</s>
<s>
Bergholm	O
proposed	O
to	O
detect	O
edges	O
at	O
coarse	O
scales	O
in	O
scale-space	B-Algorithm
and	O
then	O
trace	O
them	O
back	O
to	O
finer	O
scales	O
with	O
manual	O
choice	O
of	O
both	O
the	O
coarse	O
detection	O
scale	O
and	O
the	O
fine	O
localization	O
scale	O
.	O
</s>
<s>
Gauch	O
and	O
Pizer	O
studied	O
the	O
complementary	O
problem	O
of	O
ridges	O
and	O
valleys	O
at	O
multiple	O
scales	O
and	O
developed	O
a	O
tool	O
for	O
interactive	O
image	B-Algorithm
segmentation	I-Algorithm
based	O
on	O
multi-scale	O
watersheds	B-Algorithm
.	O
</s>
<s>
The	O
use	O
of	O
multi-scale	O
watershed	B-Algorithm
with	O
application	O
to	O
the	O
gradient	O
map	O
has	O
also	O
been	O
investigated	O
by	O
Olsen	O
and	O
Nielsen	O
and	O
been	O
carried	O
over	O
to	O
clinical	O
use	O
by	O
Dam	O
.	O
</s>
<s>
A	O
fully	O
automatic	O
brain	O
segmentation	B-Algorithm
algorithm	O
based	O
on	O
closely	O
related	O
ideas	O
of	O
multi-scale	O
watersheds	B-Algorithm
has	O
been	O
presented	O
by	O
Undeman	O
and	O
Lindeberg	O
and	O
been	O
extensively	O
tested	O
in	O
brain	O
databases	O
.	O
</s>
<s>
These	O
ideas	O
for	O
multi-scale	O
image	B-Algorithm
segmentation	I-Algorithm
by	O
linking	O
image	O
structures	O
over	O
scales	O
have	O
also	O
been	O
picked	O
up	O
by	O
Florack	O
and	O
Kuijper	O
.	O
</s>
<s>
Bijaoui	O
and	O
Rué	O
associate	O
structures	O
detected	O
in	O
scale-space	B-Algorithm
above	O
a	O
minimum	O
noise	O
threshold	O
into	O
an	O
object	O
tree	O
which	O
spans	O
multiple	O
scales	O
and	O
corresponds	O
to	O
a	O
kind	O
of	O
feature	O
in	O
the	O
original	O
signal	O
.	O
</s>
<s>
Extracted	O
features	O
are	O
accurately	O
reconstructed	O
using	O
an	O
iterative	B-Algorithm
conjugate	O
gradient	O
matrix	O
method	O
.	O
</s>
<s>
In	O
one	O
kind	O
of	O
segmentation	B-Algorithm
,	O
the	O
user	O
outlines	O
the	O
region	O
of	O
interest	O
with	O
the	O
mouse	O
clicks	O
and	O
algorithms	O
are	O
applied	O
so	O
that	O
the	O
path	O
that	O
best	O
fits	O
the	O
edge	O
of	O
the	O
image	O
is	O
shown	O
.	O
</s>
<s>
Techniques	O
like	O
SIOX	B-Algorithm
,	O
Livewire	B-Algorithm
,	O
Intelligent	O
Scissors	O
or	O
IT-SNAPS	O
are	O
used	O
in	O
this	O
kind	O
of	O
segmentation	B-Algorithm
.	O
</s>
<s>
In	O
an	O
alternative	O
kind	O
of	O
semi-automatic	O
segmentation	B-Algorithm
,	O
the	O
algorithms	O
return	O
a	O
spatial-taxon	O
(	O
i.e.	O
</s>
<s>
Most	O
of	O
the	O
aforementioned	O
segmentation	B-Algorithm
methods	O
are	O
based	O
only	O
on	O
color	B-Language
information	O
of	O
pixels	B-Algorithm
in	O
the	O
image	O
.	O
</s>
<s>
Humans	O
use	O
much	O
more	O
knowledge	O
when	O
performing	O
image	B-Algorithm
segmentation	I-Algorithm
,	O
but	O
implementing	O
this	O
knowledge	O
would	O
cost	O
considerable	O
human	O
engineering	O
and	O
computational	O
time	O
,	O
and	O
would	O
require	O
a	O
huge	O
domain	O
knowledge	O
database	O
which	O
does	O
not	O
currently	O
exist	O
.	O
</s>
<s>
Trainable	O
segmentation	B-Algorithm
methods	O
,	O
such	O
as	O
neural	B-Architecture
network	I-Architecture
segmentation	B-Algorithm
,	O
overcome	O
these	O
issues	O
by	O
modeling	O
the	O
domain	O
knowledge	O
from	O
a	O
dataset	O
of	O
labeled	O
pixels	B-Algorithm
.	O
</s>
<s>
An	O
image	B-Algorithm
segmentation	I-Algorithm
neural	B-Architecture
network	I-Architecture
can	O
process	O
small	O
areas	O
of	O
an	O
image	O
to	O
extract	O
simple	O
features	O
such	O
as	O
edges	O
.	O
</s>
<s>
Another	O
neural	B-Architecture
network	I-Architecture
,	O
or	O
any	O
decision-making	O
mechanism	O
,	O
can	O
then	O
combine	O
these	O
features	O
to	O
label	O
the	O
areas	O
of	O
an	O
image	O
accordingly	O
.	O
</s>
<s>
A	O
type	O
of	O
network	O
designed	O
this	O
way	O
is	O
the	O
Kohonen	B-Algorithm
map	I-Algorithm
.	O
</s>
<s>
Pulse-coupled	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
(	O
PCNNs	B-Algorithm
)	O
are	O
neural	O
models	O
proposed	O
by	O
modeling	O
a	O
cat	O
’s	O
visual	O
cortex	O
and	O
developed	O
for	O
high-performance	O
biomimetic	O
image	B-Algorithm
processing	I-Algorithm
.	O
</s>
<s>
The	O
Eckhorn	O
model	O
provided	O
a	O
simple	O
and	O
effective	O
tool	O
for	O
studying	O
the	O
visual	O
cortex	O
of	O
small	O
mammals	O
,	O
and	O
was	O
soon	O
recognized	O
as	O
having	O
significant	O
application	O
potential	O
in	O
image	B-Algorithm
processing	I-Algorithm
.	O
</s>
<s>
In	O
1994	O
,	O
the	O
Eckhorn	O
model	O
was	O
adapted	O
to	O
be	O
an	O
image	B-Algorithm
processing	I-Algorithm
algorithm	O
by	O
John	O
L	O
.	O
Johnson	O
,	O
who	O
termed	O
this	O
algorithm	O
Pulse-Coupled	O
Neural	B-Architecture
Network	I-Architecture
.	O
</s>
<s>
Over	O
the	O
past	O
decade	O
,	O
PCNNs	B-Algorithm
have	O
been	O
utilized	O
for	O
a	O
variety	O
of	O
image	B-Algorithm
processing	I-Algorithm
applications	O
,	O
including	O
:	O
image	B-Algorithm
segmentation	I-Algorithm
,	O
feature	O
generation	O
,	O
face	O
extraction	O
,	O
motion	O
detection	O
,	O
region	B-Algorithm
growing	I-Algorithm
,	O
noise	O
reduction	O
,	O
and	O
so	O
on	O
.	O
</s>
<s>
A	O
PCNN	B-Algorithm
is	O
a	O
two-dimensional	O
neural	B-Architecture
network	I-Architecture
.	O
</s>
<s>
Each	O
neuron	O
in	O
the	O
network	O
corresponds	O
to	O
one	O
pixel	B-Algorithm
in	O
an	O
input	O
image	O
,	O
receiving	O
its	O
corresponding	O
pixel	B-Algorithm
’s	O
color	B-Language
information	O
(	O
e.g.	O
</s>
<s>
Through	O
iterative	B-Algorithm
computation	O
,	O
PCNN	B-Algorithm
neurons	O
produce	O
temporal	O
series	O
of	O
pulse	O
outputs	O
.	O
</s>
<s>
The	O
temporal	O
series	O
of	O
pulse	O
outputs	O
contain	O
information	O
of	O
input	O
images	O
and	O
can	O
be	O
utilized	O
for	O
various	O
image	B-Algorithm
processing	I-Algorithm
applications	O
,	O
such	O
as	O
image	B-Algorithm
segmentation	I-Algorithm
and	O
feature	O
generation	O
.	O
</s>
<s>
Compared	O
with	O
conventional	O
image	B-Algorithm
processing	I-Algorithm
means	O
,	O
PCNNs	B-Algorithm
have	O
several	O
significant	O
merits	O
,	O
including	O
robustness	O
against	O
noise	O
,	O
independence	O
of	O
geometric	O
variations	O
in	O
input	O
patterns	O
,	O
capability	O
of	O
bridging	O
minor	O
intensity	O
variations	O
in	O
input	O
patterns	O
,	O
etc	O
.	O
</s>
<s>
U-Net	B-Application
is	O
a	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
which	O
takes	O
as	O
input	O
an	O
image	O
and	O
outputs	O
a	O
label	O
for	O
each	O
pixel	B-Algorithm
.	O
</s>
<s>
U-Net	B-Application
initially	O
was	O
developed	O
to	O
detect	O
cell	O
boundaries	B-Algorithm
in	O
biomedical	O
images	O
.	O
</s>
<s>
U-Net	B-Application
follows	O
classical	O
autoencoder	B-Algorithm
architecture	O
,	O
as	O
such	O
it	O
contains	O
two	O
sub-structures	O
.	O
</s>
<s>
In	O
addition	O
to	O
pixel-level	O
semantic	B-Algorithm
segmentation	I-Algorithm
tasks	O
which	O
assign	O
a	O
given	O
category	O
to	O
each	O
pixel	B-Algorithm
,	O
modern	O
segmentation	B-Algorithm
applications	O
include	O
instance-level	O
semantic	B-Algorithm
segmentation	I-Algorithm
tasks	O
in	O
which	O
each	O
individual	O
in	O
a	O
given	O
category	O
must	O
be	O
uniquely	O
identified	O
,	O
as	O
well	O
as	O
panoptic	O
segmentation	B-Algorithm
tasks	O
which	O
combines	O
these	O
two	O
tasks	O
to	O
provide	O
a	O
more	O
complete	O
scene	O
segmentation	B-Algorithm
.	O
</s>
<s>
The	O
task	O
of	O
simultaneously	O
segmenting	O
scenes	O
from	O
related	O
images	O
or	O
video	O
frames	O
is	O
termed	O
co-segmentation	B-Algorithm
,	O
which	O
is	O
typically	O
used	O
in	O
human	B-Application
action	I-Application
localization	I-Application
.	O
</s>
<s>
Unlike	O
conventional	O
bounding	O
box-based	O
object	B-General_Concept
detection	I-General_Concept
,	O
human	B-Application
action	I-Application
localization	I-Application
methods	O
provide	O
finer-grained	O
results	O
,	O
typically	O
per-image	O
segmentation	O
masks	O
delineating	O
the	O
human	O
object	O
of	O
interest	O
and	O
its	O
action	O
category	O
(	O
e.g.	O
,	O
Segment-Tube	O
)	O
.	O
</s>
<s>
Techniques	O
such	O
as	O
dynamic	O
Markov	O
Networks	O
,	O
CNN	B-Architecture
and	O
LSTM	B-Algorithm
are	O
often	O
employed	O
to	O
exploit	O
the	O
inter-frame	O
correlations	O
.	O
</s>
<s>
There	O
are	O
many	O
other	O
methods	O
of	O
segmentation	B-Algorithm
like	O
multispectral	B-Algorithm
segmentation	I-Algorithm
or	O
connectivity-based	O
segmentation	B-Algorithm
based	O
on	O
DTI	O
images	O
.	O
</s>
