Database Reference
In-Depth Information
The query peer maintains a table of community neighbors based on past retrieval
results to identify the peers which collect a similar image database.
8.4.2
Query Within the Social Network
Once the social network is identified, subsequent queries will be made to limited
peers within the social network, as illustrated in Fig.
8.10
b. To improve the
communication efficiency, instead of forwarding the request hop-by-hop in the
social network discovery stage, direct communication between the peers is utilized.
The same packet format is used for query and query response within the social
network.
Each peer in the social network collects more than one category of images, with
at least one common category as the requesting peer to satisfy the criteria to be
listed in the social network. Therefore, the same image appearing in multiple peers
is likely to belong to a common category in the social network. Let
Ret
denote
the retrieval result using query image
Q
from peer
P
n
, where
P
n
is the
n
-th member of
social network. Also, let
N
(
Q
,
P
n
)
denote the number of occurrences of each
retrieved image
I
. We can calculate an occurrence distance
D
O
for each retrieved
image, with the normalized value of
N
(
∩
Ret
(
Q
,
P
n
))
. This can be used to adjust
the ranking of images, in addition to the distance
D
F
that is usually calculated by
feature vectors. The integration of
D
O
and
D
F
is done by a weight assignment, with
the weighting factor
w
p
2
p
=[
(
∩
Ret
(
Q
,
P
n
))
w
F
w
O
]
. In this way, the similarity ranking for image
I
denoted by
Rank
(
I
)
, is obtained as:
t
Rank
(
I
)=
w
p
2
p
·
[
D
F
D
O
]
(8.12)
This integration of the two distances is referred to as the occurrence weighting
scheme [
361
].
8.4.3
Pseudo Relevance Feedback in the Distributed
Database System
While pseudo RF reduces the need for user interaction in relevance feedback,
integrating pseudo RF into the distributed retrieval system gives rise to new
challenges for repeated requests to multiple peers, which consumes bandwidth and
computational resources. To address this issue, an incremental searching mechanism
is introduced to reduce the level of transactions between the peers.
Search WWH ::
Custom Search