ling@ria.ccs.uwo.ca (Charles X. Ling) (02/22/90)
Friday, March 2, 1990. 2:00 pm. Middlesex College, Room 316
Gregory Dudek
Department of Computer Science
University of Toronto
will speak on:
THE DECOMPOSITION AND RECOGNITION
OF CURVED OBJECTS
A B S T R A C T
The recognition of a curved object from its silhouette or bounding
contour is a problem of substantial practical as well as theoretical
interest. The extraction of appropriate subparts with which to
describe objects is a natural approach to this problem. Typically,
one first applies a smoothing operation to deal with the noise in
the input data followed by the extraction of appropriate parts
into which to decompose the curve. Such smoothing, however, is
intrinsically scale-specific and hence is optimal only for certain
types of object structure. This is a major drawback, since the
natural decomposition of an object often consists of components of
very different sizes. Furthermore, different structures may overlap
at a single location, making the choice of a single ``best''
smoothing function difficult to define, even locally.
I introduce a technique that accomplishes the simultaneous smoothing
and decomposition of planar curves. This technique, dubbed
``curvature-tuned smoothing'' provides for rotation and translation
invariant smoothing of planar curves, taking the modelling
primitives into account during the smoothing operation. The
technique is based on a family of regularizing functions tuned to
specific curvatures. As the smoothing is performed, parts are
extracted at multiple scales. Because the part extraction is
explicitly based on a smoothing model, it is able to deal with noisy
data. On the other hand, the smoothing technique is specifically
tuned to the set of models being extracted.
The parts extracted by the process are subsequently usable for
object recognition. These parts correspond to regions of roughly
uniform curvature and constitute a rich description
of the original data. The technique naturally describes some
regions of the data using multiple parts, reflecting the fact that
structures at different scales may co-occur. The extension of this
approach to the description of surface data will also be
discussed.
*************
Coffee and cookies will be served after the colloquium in room 300