Paper Review Zhiqiang 9/21/12
GPU ACCELERATED EDGE-REGION BASED LEVEL SET EVOLUTION CONSTRAINED BY 2D GRAY-SCALE HISTOGRAM
Background– Active contour for image segmentation
Level set distance function Initial contourImage to be segmented
Active contour and it’s level set implementation Pro: Topological changes are handled naturally
Background– Active contour for image segmentation
CV-MODEL (region based active contour model) Its main idea is to consider the information inside the
regions, and not only at their edge. Energy function (or cost function):
2
202
10)( cIcIu
udivu
t
u
)(
220)(
210
21
~
1
,,
CoutsideCinsidedxdyuHcIdxdyuHcI
dxdyuHuccE
where u is the distance function. And C is represented as the zero level set of u. Minimizing with gradient descent flow method
~
E
Research Problem -- weakness of region based model
Example which fall
Example which success
Research Problem -- Advantage of edge based model
g is an edge-stopping function defined as follow:
1
g 21 G 0I
GAC-MODEL (Edge based active contour model)
uguu
ugdiv
t
u
50 100 150 200 250 300
50
100
150
200
250
3000.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Original image I
g(I)
Research Problem -- Advantage of edge based model
failure
success
Claimed Contribution
New model which use simultaneously edge, region and 2D histogram information in order to efficiently segment objects of interest in a given scene
Lattice Boltzmann Method (LBM) is proposed to compute the model in parallel
Edge and region estimate
Edge detector: The diffusivity coefficient g(I) is adapted to the image itself. g(I) is large when is small on intra regions. And g(I) become small when is large near edges.(same with GAC model)
Region detector: Inter-class Variance. (same with CV model)
Speed control
Region selector: Using different evolution speed in various regions based on gradient histogram analysis.
Diffusion equation with a body force:
Experiment results
GPU implementation: Parallel computing toolbox of matlab R2012a and NVIDIA GPU GT 430.(ignore the dates transferring time between CPU and GPU)
Contribution analysis
The proposed model isn’t novel, And segmentation results seem not to be art of the state.
Considering other work that focus on using LBM for active contour model, What’s the unique contribution of this paper?
Algorithm analysis
LBM (step 4) may be very fast on GPU, but the computing of image
features which involve statistical information
would be time consuming.
Which parts of algorithm is computed on GPU?
Computing time for each step?
Questions?