Friday, February 20, 1998
3:00 - 4:00 pm
1001 EECS
abstract-
Desktop imaging systems have become ubiquitous. There are many
interesting
signal processing problems associated with the task of
rendering high
quality images with these systems. After a brief overview of
this area, I
will focus on two such problems in particular. The first is
the efficient
implementation of the highly nonlinear transformations
between a standard
color space such as CIE L*a*b* and the native space of a
color printer.
This is a problem in multidimensional approximation theory.
Our method
which we call sequential linear interpolation is mediated by
the need for
extremely fast search and computation, as well as the desire
to minimize
the number of grid points in the lookup tables. The approach
is inspired
by the asymptotic theory of vector quantization. We
previously developed
design rules that are optimal under a mean-squared error
criterion. I will
describe our more recent results that are optimal in a
minimax sense. The
second problem that I will discuss is that of digital
halftoning, which
determines the actual bit map sent to the printer. Our
approach is based
on a heuristic search for the bit map that achieves a local
minimum of a
cost function based on a simple model for the human visual
system. This
method yields halftone images of excellent quality. Despite
the apparent
complexity of the algorithm, I will show that it can be
implemented with
urprising efficiency. In addition, I will briefly describe
derivative
algorithms that are still more efficient, requiring only a
few operations
per pixel.
Biosketch -
Please refer to the homepage
found through the link shown above.