|
The
evolution of a handwritten word recognition algorithm will be
described. The algorithm was developed mainly under U. S. Postal
Service funding. It is a segmentation-based algorithm that relies
on dynamic programming to match the words against lexicons. Several
algorithms of this type have been developed by various researchers.
However, their performance can vary dramatically. In fact, although
the high-level description of our algorithm hasn't changed for
several years, its recognition performance has increased dramatically.
In this talk, I will first provide an overview of the system.
I will then describe some of the changes that were made to improve
its performance. Some of these changes had to do with the character
recognizer. A result that was surprising to us at first was that
high character recognition rates do not necessarily imply high
word recognition rates. Other changes involved using information
extracted from multiple character segments simultaneously to measure,
within the dynamic programming matcher, the spatial compatibility
of a match to a string. Still more changes had to do with methods
of aggregating information. We found that robust methods derived
from the Choquet integral can achieve better results than traditional
methods. Finally, if time permits, I will discuss the problem
of combining outputs of several word recognition algorithms using
several methods including neural networks, Borda counts, and Choquet
integrals.
|
