We have described in this chapter our current results for map
separates, based on the RGB clustering
algorithm. This method results in a comparable or somewhat lower
quality separation then the backpropagation algorithm, but it is faster
by a factor of 100. It is suggested that our RGB clustering algorithm
is in fact essentially equivalent to a backpropagation algorithm. In
the neural network jargon, we can say that we have found the analytic
representation for the bulk of the hidden unit layer which results in
dramatic performance improvement. This representation can be thought
of numerically as a pixelregion look-up table or
geometrically as a set of polyhedral regions covering the RGB cube.
Further quality improvement of our results will be achieved soon by
refining our software tools and by coupling the RGB clustering with the
zero-crossing-based segmentation and edge detection algorithms. Zero crossing techniques provide in turn a
natural algorithmic connectivity for our intended collaboration with
Coherent Research, Inc. on high-level vision and AI/expert systems
techniques for map understanding.