It is exercises like that I would find most useful and
stimulating. There is a whole body of material that one has - sliders, reproductions in
books, etc. - and perhaps it can suffice to launch one into this area of inquiry which, in
turn, might become an opening into still larger areas.
The second specific issue that I would like to address is
the possibility of using the computer as tool for understanding the micro-measurements and
the system of proportions of forms, figures etc. which artists in the tradition carried in
their heads and worked with. When we identify the family-workshops of painters, as I have
tried to do in my work, we know that individual artists had distinct preferences of their
own, in respect of proportions, stances etc. The case of the four sons and two nephews of
the 18th century Pahari painter Nainsukh, whom we group together as 'The First Generation
after Manaku and Nainsukh' is of great interest. This generation turned out some of the
greatest paintings produced in India, and yet at present we are unable to go beyond a
point in isolating their individual hands. Can, I wonder, the computer assist us in doing
this by carrying this kind of analysis our?
Multimedia and internet technologies offer a great
potential for on-line dissemination of information pertaining to art and culture. With the
growth of such collection across the network, an efficient retrieval mechanism becomes
extremely important. It is more so for audio or video documents which cannot be easily
browsed unlike text or picture documents, and puts more cognitive load on the users. In
this paper, we describe a knowledge based retrieval method for multimedia documents from a
distributed repository.
The problem of retrieval deals with matching documents to
queries based on some similarity measures. A retrieval engine operating on multimedia
document repository should be able to exploit the rich information content of the document
both in textual as well as in non-textual media forms in establishing the similarity
measures. While there has been significant research on content-based retrieval of
multimedia documents, the unique feature of our retrieval method is that it computes the
similarity measure in terms of conceptual entities abstracted over multiple media forms.
We model the retrieval problem as a problem of abduction. Abduction is a reasoning process
for constructing an appropriate explanation for a set of observed patterns. Concepts are
abstract entities which cannot be directly observed in the document. We define a set of
observation mode for a concept in terms of perceptable media objects and their
relationships. The recognition models for the media objects comprises a set of predicates
involving elementary media patterns that are invariant over media specific
transformations. In our abductive model of reasoning we identify the documents which can
account for the expected observation patterns as candidates for retrieval.
We use multiple forms of knowledge at different levels of
abstraction to provide the solution for a retrieval problem. Conceptual knowledge deals
with the concepts and their observation models. Media knowledge encodes the recognition
functions. Classification knowledge provides a set of supportive assumptions for
abduction. Above all, a planning module coordinates the use of knowledge for an optimal
retrieval. This method of partitioning and distributing the required knowledge-base is a
distinguishing feature of our approach.
We present an agent-based architecture to support the
reasoning framework. This architecture enables us to expoit the knowledge base possibly
distributed over a network. It allows dynamic growth of the system by introduction of new
agents and customising user interfaces. Another interesting aspect of our architecture is
introduction of mobile agents which can move across an arbitrary network and perform
requisite functions.
There has been considerable research interest on
application of AI-based techniques for multimedia information retrieval in the recent
past. Our approach is distinct in the following ways:
