Database Reference
In-Depth Information
Fig. 5.1
Landmark recognition process, incorporating saliency weighting scheme and re-ranking
s1, s2, and s3. The s1 is the extraction stage, where feature vectors are extracted over
the image plane. This results in a feature map
M
:
2
, where
n
is the image
space. The s2 is the activation stage, which forms an activation map
A
:
[
n
]
ₒ
R
2
[
n
]
ₒ
R
,
2
where the image is salient will correspond to high
values of activation
A
. The last stage s3 will normalize the activation map.
The GBVS method is applied to the last two stages, for a given feature map
M
.
Specifically, the dissimilarity of
M
such that locations
(
i
,
j
)
∈
[
n
]
(
i
,
j
)
and
M
(
p
,
q
)
is computed by:
(
,
)
M
i
j
d
((
i
,
j
)
||
(
p
,
q
)) =
log
(5.1)
M
(
p
,
q
)
Firstly, the fully-connected directional graph
G
A
is obtained by connecting every
node of the lattice
M
, labelled with two indices
2
, with all other
n
(
i
,
j
)
∈
[
n
]
−
1
nodes. The directed edge node
(
i
,
j
)
to node
(
p
,
q
)
will be assigned a weight:
exp
2
2
−
(
−
)
+(
−
)
i
p
j
q
w
((
i
,
j
)
,
(
p
,
q
)) =
d
((
i
,
j
)
||
(
p
,
q
))
·
(5.2)
2
˃
2
where
is a free parameter. Secondly, the graph
G
A
is converted to a Markov chain,
and the equilibrium distribution of this chain results in an activation measure, i.e.,
the activation map
A
. Thirdly, the map is normalized to generate the final saliency
map
S
of the image.
In the current work, at the final stage, the Sigmoid function is applied to conduct
a mapping function from the activation map to saliency map [
346
],
˃
1
−
exp
(
−
b
·
A
(
i
,
j
))
S
(
i
,
j
)=
a
+(
1
−
a
)
(5.3)
+
(
−
·
(
,
))
1
exp
b
A
i
j
Search WWH ::
Custom Search