Database Reference
In-Depth Information
Peer A. Consequently, Peer A evaluates the retrieval results using pseudo-RF, and
generates a new feature vector (i.e., RBF center). A new query request using the new
feature vector will be sent to Peer B, as well as incrementing the audience to Peer
C. The query request and automated retrieval evaluation process is repeated until a
pre-defined number of peers is reached.
The pseudo-RF is implemented by the SOTM algorithm discussed in Chap.
3
.
SOTM is applied for the pseudo labelling of retrieved samples, and the single-RBF
method is applied for relevance feedback. In the SOTM algorithm, steps 1-5 are
the same as the ones discussed in Sect.
3.2.2
. However, the updating of the winning
node
w
j
∗
in step 3 [cf. Eq. (
3.3
)] is modified to:
w
j
∗
(
t
+
1
)=
w
j
∗
(
t
)+
ʱ
(
t
)
ʲ
(
v
(
t
)) [
v
−
w
j
∗
]
(8.13)
where
is the ranking function which is inversely proportional to the ranking
of the feature vector
v
at iteration
t
. This ranking is obtained by the similarity scores
between the query feature vector and the feature vector
v
. A large value of
ʲ
(
v
,
t
)
))
indicates a high relevance of the feature vector compared with the respective query
feature [
361
]. As a result, the prototype vectors
w
j
are adjusted so that they learn
more from statistically similar inputs and less from statistically irrelevant ones.
The RBF method [i.e. Eqs. (
8.8
)-(
8.9
)] is utilized for nonlinear similarity mea-
surement by the adjustment of RBF centers and widths. Since pseudo-RF may cause
some errors in the pseudo labeling of retrieved samples, the error will propagate
into subsequent modification of the RBF center. To minimize the error preparation,
a bias weighting
ʲ
(
v
(
t
is introduced to the original query vector corresponding to the
initial RBF center
c
ʳ
(
0
)
in Eq. (
8.10
). The new RBF center updating function is
obtained by:
)+
ʱ
N
c
F
−
F
+
+
ʳ
c
(
t
+
1
)=(
1
−
ʳ
)
c
(
0
(
t
)
−
(8.14)
This formula weights the importance of the original query and the mean of the
positive samples, and will replace Eq. (
8.9
) for calculation of the RBF center in the
RF learning.
8.4.3.2
Offline Feature Calculation
Online feature calculation requires high computational resources and results in delay
in content retrieval. Redundant online feature computation can be eliminated by the
following specifications:
Each image stored in the social network is attached with its feature descriptor.
When a peer creates a new image, the feature descriptors will be computed and
attached with the image file before announcing the availability of the new image.
Any image transmission over the social network will be coupled with the
transmission of the image's feature descriptor.
Search WWH ::
Custom Search