You are on page 1of 19

Mathematical Image Segmentation:

Variational and Level Set Techniques.


Adrin Galdrn
1,2
, David Pardo
2,3
, Artzai Picn
1
1
Tecnalia Research & Innovation, Parque Tecnolgico de Bizkaia, Zamudio, Spain.
2
Basque Country University (UPV/EHU), Bilbao, Spain.
3
IkerBasque Foundation, Bilbao, Spain.
16 de septiembre de 2011
Tecnalia + UPV + Image Processing
Image Processing
Great Variety of Tasks
Image Filtering and
Enhancement
Image Features
Extraction
Segmentation
Registration
Classication
Recognition
Mathematical Image Segmentation
Many types of images
Light Intensity Images (LI), Color Images (RGB, RGBA, CIELAB), Magnetic
Resonance Images (MRI), X-Ray Images, Ultra-Sound Images, Diffusion
Tensor Images (DTI)
Biomedical Images extracted form the Bioimage Suite of image processing developed at Yale University, www.bioimagesuite.org
Different Approaches
An important decision:
Continuous Vs Discrete
Classical algorithms are simpler, but have some heavy disadvantages.
Algorithm complexity is greatly
increased with the complexity
of the application.
Not able to evolve to adapt
new image data or conditions.
Difcult to include prior or
acquired knowledge not related
with image data.
Must be adapted for use in
other applications.
PDE-based Approach
Variational Techniques
Consider the image as the graph of a function I : R
2
/R
3
X.
Dene an Energy modeling the problem, and nd the
minimum of that functional. For that, differentiate the functional,
nd Euler-Lagrange equations, and solve numerically.
PDEs are not only used for
Segmentation
Noise Removal, Image
Restoration, Image
Classication,. . . .
C. Ballester, M. Bertalmio, V. Caselles, G. Sapiro, and J. Verdera,
IMA Report 2000, IEEE Trans. IP 200
Most Classical Model: Snakes
The idea
Put a closed parametric curve c on the image, and make it
evolve in time, to get close to the contour of the object we want
to segment.
Minimize
J(c) =
_
b
a
|c

(q)|
2
dq+
_
b
a
|c

(q)|
2
dq+
_
b
a
1
1 +|I(c(q))|
2
dq
Snakes: Details
J(c;

c) = . . . =
_
b
a
_
c

+ c
iv
+ E
I
_

c
Euler-Lagrange for Snakes
c

(q) + c
iv
(q) +E
I
(c(q)) = 0
c(a) = c(b), c

(a) = c

(b)
_
Gradient Flow for Snakes
c
t
= c

(t , q) + c
iv
(t , q) +E
I
(c(t , q))
c(0, q) = c(q), c(t , a) = c(t , b)
c

(t , 0) = c

0
(t ) c

(t , a) = c

(t , b)
_

_
(where c
0
is the initial curve, supposed to be enclosing the object to be detected.)
Pros, Cons, Limitations
Snakes Model
Pros
Low complexity.
Easy to introduce prior
knowledge.
Well established
technique, lots of
publications.
User interactivity
Cons
It depends on the choice of
the parametrization.
The sampling affects the
nal result
Very sensitive to initial
conditions.
Unable to perform
topological changes.
Level Set Method
What is the Level Set Method?
An implicit data representation of a hypersurface, a
set of PDEs that govern how the surface moves, and
the corresponding numerical methods for
implementing this on computers. Stanley Osher
Many Different Applications
Geometry, Fluid Dynamics,
Optimal Design, Seismic
Analysis, Robotics,
Semiconductors, Optimal
Control, Image
Processing. . .
Level set uid simulation, R. Fedkiw et al.
(http://physbam.stanford.edu/ fedkiw/)
Level Set Method
In the most simple case, represent a plane curve as the set of
zeros of : R
2
R.
An initial curve c(p, t ) = (x(p, t ), y(p, t )), parameterized by p, is
moving across the image following a velocity eld

V : R
2
R
2
. Parameter t controls the temporal evolution.
Level Set Method
tracks c(t ) through its 0-level set (c(t ), t ) = 0, so:

t
(x(t ), y(t ), t ) +

x
x
t
+

y
y
t
=
t
+
_
x
t
,
y
t
_
. .
c

= 0
The time derivative c

is the velocity at which it is evolving, so:

t
+

V = 0.
Decomposing the velocity eld in its normal and tangential
components,

V = V
n
n + V
t
t, and using n = /||,

t
+ V
n
|| = 0
Level Set Method: How does it Apply to Image
Segmentation?
We build an Integral Cost Function J : C R that depends on the
image, and pushes the curve towards high gradients, or region
dissimilarities.
Dictionary
Heaviside Function, H() =
_
0, if 0
1, if < 0
Normal,

n =

||
Curvature, = div
_

||
_
Level Set Method: How does it Apply to Image
Segmentation?
Geodesic Active Regions
Extending Snakes, we can arrive to
min
C(q)
_
g(|I(C(q))|) |C

(q)|dq
where g(t ) = 1/(1 + t
2
)
Level Set Version

t
= g(I) || +g(I)
Level Set Method: How does it Apply to Image
Segmentation?
Active Contours without edges
It is easy to introduce simple global statistics,
E(c
1
, c
2
) =
_

(c
1
)
2
H()+(c
2
)
2
(1H())+|H()|dxdy
where c
1
is the mean inside , c
2
is the mean outside , and
_

|H()| is the length of the zero level set of .


We can make the energy function dependant on more complex
probabilistic and statistical tools (prior knowledge, histogram,
texture, 3D shape, location of landmarks).
Level Set Method: How does it Apply to Image
Segmentation?
Level Set Segmentation
Level Set Method: Advantages
Advantages of the Implicit Representation
The Level Set Method does not
suffer of
Parametrization-Dependency
problems.
It is able to account for
Changes of the Topology
automatically: the embedding
function can vary smoothly,
while its zero level set inherits
non-continuous changes.
The resulting law of evolution
for a surface becomes a
Hamilton-Jacobi PDE, a
classical equation that has
been intensively studied. The
Level Set Method allows us to
introduce the machinery of the
Numerical Analysis, both for its
theoretical study and for the
computational solution.
Straightforward Generalization
to higher dimensions, since the
formulas for curvature and
normal vector are not
dimension-dependent.
Conclusions
Mathematical and statistical modeling techniques are essential
to deal with complex applications in Image Processing and
Computer Vision.
The Level Set Method is a powerful tool that has been applied in
different scientic elds.
Due to its numerous advantages over traditional methods, it is
also being applied by the Image Processing community.
The goal is to be able to combine eperience and expert
knowledge in statistical tools with advanced numerical analysis
techniques to develop Image Segmentation Algorithms that rely
on the Level Set Framework.
Mathematical Image Segmentation:
Variational and Level Set Techniques.
Adrin Galdrn
1,2
, David Pardo
2,3
, Artzai Picn
1
1
Tecnalia Research & Innovation, Parque Tecnolgico de Bizkaia, Zamudio, Spain.
2
Basque Country University (UPV/EHU), Bilbao, Spain.
3
IkerBasque Foundation, Bilbao, Spain.
16 de septiembre de 2011

You might also like