Fast Methods to Find Optimal Shapes in Images
Gunay Dogan Mathematical and Computational Sciences Division, NIST
Tuesday, April 21, 2009 15:00-16:00, Building 101, Lecture Room C Gaithersburg Tuesday, April 21, 2009 13:00-14:00, Room 5000 Boulder
Abstract:
Through the last two decades, the field of image processing has seen
a significant increase in the use of energy minimization methods. An
important class of such problems falls under the task of image
segmentation, that is, finding distinct objects or regions in given
images. Applications range from detection of people in surveillance
images to 3d visualization of human organs from medical scans. These
methods are usually based on partial differential equation
formulations and involve iterative update of a candidate solution with
respect to an assigned energy, in order to achieve a "best" solution.
In this talk, we approach these problems from a shape optimization
point of view and introduce a novel iterative method to find the
minimum energy shapes in given images. In our method, we model the
geometry explicitly, respecting the continuous structure of the
problem, and discretize the resulting formulation using the finite
element method. A distinct feature of our method is that it allows us to
apply specially-designed gradient descent schemes tuned for improved
performance. We demonstrate the power of this new method with several
examples in 2d and 3d.
Speaker Bio:
Gunay Dogan received his PhD degree in Applied Math and Scientific Computing
from the University of Maryland, College Park in 2006. His PhD advisor was Ricardo H. Nochetto. Following his PhD, he worked as a postdoctoral researcher with George Biros at the University of Pennsylvania for two years. Currently he is a guest researcher in the Mathematical and Computational Sciences Division at NIST. Gunay Dogan's scientific contributions have been in the areas of shape optimization, image segmentation and numerical methods for inverse problems. His scientific interests are broadly in image processing, scientific computing and inverse problems, with emphasis on large-scale optimization and numerical solutions of partial differential equations."
Presentation Slides: PDF
Contact: S. LangerNote: Visitors from outside NIST must contact
Robin Bickel; (301) 975-3668;
at least 24 hours in advance.
|