(Georgia Institute of Technology, 2002)
Yaner, Patrick W.; Goel, Ashok K.
We consider the problem of associative image retrieval, focusing on
retrieval of 2-D line drawings by example. We represent 2-D line
drawings as semantic networks of spatial elements and relations among
them. We describe a process for retrieving the drawings based on a
structural analogy between the query and the stored images. We then
present several methods of retrieving the drawings: the first family
methods uses logical unification and resolution to accomplish the
matching; the second family of methods heuristically prunes the stored
drawings and then does the resolution and unification on the remaining
drawings; two more methods treat the retrieval problem as a constraint
satisfaction problem and use common CSP techniques for solving it; and
the last two methods combine the heuristic step of the second method
with the CSP technique of the third and fourth. We report on
experimental results that compare the performance of these methods on
computer-based libraries of drawings. A surprising result of our work is
that for the fastest of these methods two stage retrieval appeared to
offer no benefit over one stage retrieval.