Contact Information

VISLab
Winston Chung Hall Room 216
University of California, Riverside
900 University Avenue
Riverside, CA 92521-0425

Tel: (951)-827-3954

UCR Collaborators:
CSE
ECE
ME
STAT
PSYC
ENTM
BIOL
BPSC
ECON
MATH
BIOENG
MGNT

Other Collaborators:
Keio University

Other Activities:
IEEE Biometrics Workshop 2019
IEEE Biometrics Workshop 2018
Worshop on DVSN 2009
Multibiometrics Book

Webmaster Contact Information:
Alex Shin
wshin@ece.ucr.edu

Last updated: July 1, 2017

Learning-Integrated Interactive Segmentation and
Classification of Synthetic Aperature Radar Imagery

Presented by: Stephanie Fonder

ABSTRACT: Segmentation is a low-level task that is a first step to many computer vision problems. Although image segmentation algorithms exist which perform reasonably on a limited data set, a general solution for the image segmentation problem has not been developed. We present an approach to image segmentation, in which user selected sets of examples and counter-examples supply information about the specific segmentation problem at hand.
In this approach, image segmentation is guided by a genetic algorithm which learns the appropriate subset and spatial combination of a collection of discriminating functions, associated with image features. The genetic algorithm encodes discriminating functions into a functional template representation, which can be applied to the input image to produce a candidate segmentation. The quality of each candidate segmentation is evaluated within the genetic algorithm, by a comparison to two physics-based segmentations. Through the process of segmentation, evaluation, and recombination, the genetic algorithm non-exhaustively optimizes functional template design. The contributions of this thesis include: genetic learning of functional template design, physics-based segmentation evaluation, novel crossover operator and fitness function, as well as a system prototype and experiments on synthetic and SAR imagery.

Experimental results demonstrate that evolved templates select meaningful features, which complement each other for improved segmentation quality over any single feature. Evolved segmentations consistently outperform segmentations derived from the Bayesian best single feature and typically perform at least as well, if not better than segmentations derived from the actual best single feature defaults.