With distributed image retrieval, the difficulties associated with traditional CBIR are
also present, but the distributed image retrieval process is further complicated by such
factors as the context of the individual image collections and the similarity algorithms
used in the image retrieval systems.
One of the main challenges in distributed image retrieval is how to take advantage
of the potential that lies in having several databases available for image retrieval.
Distributed collections can help users gain access to a richer and better source of
information, and the challenge is how to provide a user with the most relevant result set
possible. One goal here is to have one result list regardless of the number of databases
queried, and that the result list has the same quality as would be experienced retrieving
images from a single source.
A major obstacle associated with distributed systems consisting of database systems
from different vendors is that different DBMS’ use different algorithms for both query
processing and result ranking. These database system (DBS) specific algorithms are
most often kept secret in commercial systems and thus are not available to the
developers of systems to be used for distributed image retrieval. The consequence is that
result sets returned from different sources may not be ranked in a compatible way,
resulting in a situation that complicates the process of merging distributed result sets.
Prior work demonstrates that significant improvement is often seen when combining
CBIR algorithms with text retrieval algorithms on an arbitrary collection of result sets
from different information retrieval systems.
The main motivation behind the study was to evaluate if the utilization of existing
DBMS functionality in a new way could contribute in narrowing the Semantic Gap.
The experimental approach was evaluated through the development of a prototype
called Context Aware Image Ranking – CAIRANK.