Speaker, Murray Loew
         Electrical and Computer Engineering
         George Washington University

Date: 7/12
Time: 4-5 pm
Place: Large Radiology Conference Center (bldg. 10, room 1C520)
Title: Task-Dependent Feature Extraction in Image Computing

Abstract:

Whether the application is recognition, classification, segmentation,
or registration, feature extraction is usually an important step in
image analysis and understanding.  The need for feature extraction
arises from the quest for invariance: for insensitivity to changes in
scale, noise, orientation, and image modality, among other image
properties.  The expectation is that a set of features will exhibit
greater invariance overall than any single measure.  The kinds of
features that are extracted will depend also on the task and so it
will be necessary to overlay knowledge of the application area onto
the problem.

This talk reviews methods that arise from first principles and those
with a more heuristic basis -- including statistics, signal
processing, fractals, wavelets, shape and connectivity analysis, and
texture -- and examines the use of prior information about the problem
domain.  The question of feature selection and dimensionality
reduction is addressed in terms of intrinsic separability, using new
theoretical results.

Definition of the task is important because it will determine the
kinds of data that are used for design - and, at least as
consequential, for evaluation -- of the features.  Validation of the
effectiveness of features (and careful definition of the domain in
which they are effective) is an essential process in practical
problems.  Computational cost is often considered as well, especially
when extensions are made to 3- and 4-dimensional images.

Examples are given from problems in medical imaging, where better
features can reduce the cost and risk of imaging, where better data
bases will increase confidence in the clinical value of the analyses,
and where increased interest in standards will make for real
reproducibility of the methods and outcomes.