Adaptive Retrieval in a P2P Cloud Datacenter - Multimedia Database Retrieval: Technology and Applications - page 227

8.4.3.3 Advanced Feature Calculation

For extensibility, apart from offline feature calculation, an ideal retrieval system
should also allow new features to be computed on the fly. These new features, which
typically address the specific feature for a query image, are best implemented with
the pseudo RF such for more accuracy in the classification of image relevancy.
The advanced feature descriptors for the distributed retrieval systems can be
generated by the following two approaches: query node pseudo-RF and agent-based
pseudo-RF.

Query Node Pseudo RF

As illustrated in Fig. 8.11 a, the query node, Peer A, makes the initial request to the
destination node, Peer B. The destination nodes will transfer the retrieved images
to the query node. The query node will calculate the advanced feature descriptor on
the fly and perform pseudo RF. A new query message will be generated and sent
from the query node to the destination nodes repeatedly, until a pre-defined number
of iterations is met. This new query message contains updated RBF parameters
(i.e., RBF centers and widths) which are used to perform content search within the
destination nodes.
The query node pseudo-RF approach requires bandwidth for multiple retrieved
data transmission to the query node, and the computation for the query node to
calculate advanced features.

Agent-Based Pseudo RF

A software agent technique [ 225 ] can be applied to offload the bandwidth and
computation cost from the query node. As shown in Fig. 8.11 b, the query node,
Peer A, initiates a software agent to carry the query vector using a standard feature
descriptor, and the algorithm for computing advanced features, to the destination
node, Peer B. Peer B performs the retrieval with pseudo RF using the advanced
features computed on-the-fly. The software agent carries the new query vector,
advanced feature extraction algorithm, and the retrieved data from Peer B, to
the subsequent neighbor node, Peer C. Upon reaching a pre-defined number of
neighbors, the software agent will carry the retrieved data back to the query node.
Offloading computational cost from the query node to destination nodes raises
security concerns, as the flexibility of remote procedure execution opens the door-
way for various malicious attacks. Therefore, authenticating as well as validating
the integrity of the software agent is required.