Graph-Shifts: Dynamic Hierarchical Energy Minimization for Segmentation and Labeling
Speaker: Jason J. Corso , University at Buffalo SUNY
Date: November 7 2008
Time: 2:00PM to 3:00PM
Location: 32-D507
Host: Polina Golland, CSAIL
Contact: Polina Golland, x38005, polina@csail.mit.edu
Relevant URL: Energy minimization plays a key role in many important computational
imaging problems including segmentation and labeling. Conventional
methods like those based on partial-differential equations (PDE) or
Markov Chain Monte-Carlo sampling are either susceptible to local
minima or computationally expensive. There is great potential in
using hierarchical methods to increase both robustness and
efficiency. In this talk, I will present such a method: the graph-
shifts energy minimization algorithm. This algorithm makes use of a
dynamic hierarchical representation of the image to rapidly and
robustly minimize an energy function. Few constraints are imposed on
the energy function: it simply must be a Gibbs energy with local
potentials; i.e., conditional markov random fields are suitable, as
are hybrid discriminative/generative models. Two graph shift
operations are defined, combined splitting-and-merging and new region
spawning, which make it possible to explore the entire solution
space. The dynamic hierarchical representation allows the algorithm
to deterministically compute and select the optimal change to perform
at each iteration.
I will discuss the application of the graph-shifts algorithm to image
segmentation and labeling. The algorithm has been applied to both
anatomical (sub-cortical structures in the brain) and pathological
(brain tumor and multiple sclerosis) 3D medical segmentation problems
as well as 2D natural image labeling. Probabilistic boosting methods
are used to train the energy terms from a set of labeled training
images. The results are comparable or superior to state-of-the-art
approaches but are obtained considerably faster (by orders of
magnitude). The results also quantitatively demonstrate robustness to
initialization and avoidance of local minima in which conventional
boundary PDE methods fall.
Joint work with Alan Yuille and Zhuowen Tu
See other events that are part of Biomedical Imaging and Analysis 2008/2009
See other events happening in November 2008
