Principal
Investigator:
Leon Cooper, PhD,
Professor
Department of Physics
Brown
University
Providence,
RI
Brief
Description:
Automated recognition
systems classify and identify patterns as objects within a library of one- /
two- dimensional objects, such as handwriting, speech and/or visual data. Current recognition systems are
comprised of several steps categorized into two general stages: preprocessing
and postprocessing. Preprocessing
forms an integral part of many existing artificial recognition systems, whereby,
an input signal is converted or transformed into a more suitable form for
further processing. Common steps
performed during preprocessing are normalization, noise reduction (filtering),
and feature extraction, all of which are used to help eliminate noise that
interfere with signal clarity and to effectively eliminate irrelevant
information. There are some
inherent problems that exist during preprocessing, which cannot be resolved,
because of necessary information that is sometimes unavailable at this
level. In summary, current
recognition systems are imperfect in several important ways.
A significantly
improved system over existing state-of-the-art would provide a/an: 1)
representation of an object in terms of its constituent parts; 2)
translationally invariant representation of the object with scale invariant
recognition capability; 3) effective recognition of some patterns that are
partially present/occluded in the input signal; 4) suitable representation for
sequences within the input signal to preserve sequential ordering; 5)
procedure/algorithm based on dynamically determined, context based expectations
for identifying individual features or parts of an object to be recognized. An improved system should be
computationally efficient and capable of implementation in a highly parallelized
configuration, while providing an information processing system that utilizes
the interaction between higher (cognitive) processing level and various lower
level modules. Furthermore, a
mechanism is needed for improving the preprocessing of individual sections of an
input pattern, either by applying one or more preprocessors selected from a set
of several, or by changing the parameters within a single
one.
The innovative
technology is a more accurate recognition system that includes the desired
advantages above; it is comprised of feed forward, feed back multiple neural
networks with context driven recognition.
This novel, robust system can be applied to recognition of a variety of
input signals – [cursive or block]
handwriting, speech, and/or visual data – and hence, is suitable for more rapid
recognition of one-dimensional sequences and two-dimensional image analyses such
as face recognition and/or vehicle identification systems. Broadly, the invention includes 4 parts
that can be embodied in computer software or hardware: binding network,
segmentation network, preprocessor, and a word detector.
Markets for this
technology include, but are not limited to, military field equipment,
automotive, security/surveillance, biomedical prosthetics for vision, and
scientific R&D tools. Examples
of applications for a recognition system are in the manufacture of global
positioning systems (GPS), devices/sensors for military surveillance
applications to identify objects of any type in the field, security
devices/systems to identify appropriate person(s) in residential, commercial,
industrial, or other arenas. Also,
research scientists may find use for such a system in laboratory experimentation
in the fields of computer science, mechanical engineering, robotics, sensors,
and biomimetics, and biomedical
engineering/perception,
among others.
Information:
US patent 6,560,360is
issued (5/06/03)
