Database Reference
In-Depth Information
A
put
call first uses Chord to map a node using the lookup service, and does a
put
in its local index storage using the cluster ID as the key.
In Fig.
8.2
b, users search for image files using cluster IDs extracted from a query
image, where one query image may be characterized by more than one cluster ID.
A set of cluster IDs are obtained by self-organized mapping of the query feature
vector. A
get
call uses Chord to map the Cluster ID to a node, and does a
get
in
its local database using the cluster ID as the key. The image list (i.e., index entries)
for each cluster ID is then fetched from the network and the system aggregates the
results. The query node sorts the results list using a similarity measurement between
the query vector and the feature vectors list, and returns the top matching images to
the user. Furthermore, the adaptive retrieval method is used for relevance feedback
of the initial retrieval set and improves the ranking result.
8.3.2
Indexing of Nodes and Data Items on the Distributed
Hash Table
A DHT distributes data among a set of nodes to allow each peer to become
responsible for a range of data items. Each peer has partial knowledge about the
network, and this knowledge will allow the network to route the queries about the
data items to the responsible nodes using efficient and scalable procedures. The
DHT utilizes a circular ring for indexing of nodes and data items. Each data item or
a peer is mapped to a point in a large address space of size 2
m
, where
m
160, i.e.,
m
-bit integer. A node on the network is indexed by a node identification (ID),
=
node ID
=
hash
(
peer IP address
)
(8.1)
by using a mathematical hash function. A data object to be shared is also hashed to
an
m
-bit integer,
key
=
hash
(
Object ID
)
(8.2)
This object is related to the pair (key, data), where data is the data itself or a reference
to the data object.
In the CSS, as illustrated in Fig.
8.2
, when a node wants to share an image, the
feature vector of the image is first extracted. Then, the SOTM is applied to map
the feature vector to the cluster ID. The mapping is denoted by:
d
Z
,fromthe
real-
d
dimensional vector to an integer. Then the cluster ID is hashed and an index
message is sent to the node responsible for the ID through the DHT layer.
Let
R
ₒ
d
be the set of SOTM weight vectors
w
i
previously generated. For an input feature vector
f
v
∈
R
C
=
{
w
i
|
w
i
∈
R
,
i
=
1
,
2
,...,
C
}
d
, its corresponding cluster
ID is constructed by:
Search WWH ::
Custom Search