C A B S  C O L L O Q U I U M


The CABS colloquium is organized by the Collective Agent Based  
Systems group (TU-Delft) the Intelligent Systems Group 
(Universiteit van Utrecht) in close cooperation with SIKS.
------------------------------------------------------------------------ ------------------------------------------------------------


Title: M A P R :  M u l t i p l e  A g e n t s  f o r  P a t t e r n  R e c o g n i t i o n



Speaker: dr. Louis Vuurpijl, NICI, KUN


Date:  Friday 14 maart 2003


Time: 14.30 - 16.00 uur



Location: Snijderzaal, Laagbouw 1.01 ITS gebouw, Mekelweg 4 2628 CD  TU-Delft



Abstract
The traditional approach for pattern recognition uses a pipeline of
processing operations: 1) preprocessing, 2) segmentation, 3) feature
extraction and 4) classification. Today, more elaborate and complex
classification schemes, including feedback, multi-stage, hierarchical
and multiple classifiers are being researched. The combination of
classifier hypotheses in the area of pattern recognition is an
example of the more general and fundamental problem of integration of
information from multiple sources. This problem, formerly only known
as the bottom-up vs top-down processing problem, takes place at all
stages of information processing in pattern recognition. Consider,
for example, two character classifiers in an optical character
recognition system, one for the digits, and the other for the letters
in the alphabet. If the letter classifier claims that the letter 'o'
is the best matching pattern, whereas the digit classifier claims
that the digit '0' is the best match, then more information is needed
from an independent source. This source, however, may reside at
several possible abstraction levels. The disambiguation may come from
the signal (the image elements and their detailed shape) itself, from
higher-level current expectancies on possible character classes, or
from sibling classifiers at the same processing level. We propose to
make use of a new and promising paradigm, i.e., the use of so-called
'multiple agents' for pattern recognition.


In this talk, I will discuss the problem of handwritten text
recognition. Results using multiple classifiers combined through
voting mechanisms, weighing schemes and organized through
hierarchical decision trees are presented. I will argue that such
elaborate classification mechanisms are still not enough by showing
various examples where the recognition techniques could benefit from
negotiation strategies well-known from the intelligent agent
communities ...