Image Segmentation Using A Generalized Fast Marching Method

Image Segmentation using a Generalized Fast Marching
Method
Nicolas Forcadel (ENPC-CERMICS, Paris), Carole Le Guyader (INSA

Rennes), Christian Gout (INRIA, France)
Multivariate Approximation: Theory and Applications
April 26 - May 01, 2007
N. Forcadel C. Le Guyader, C. Gout Image Segmentation

Plan
1 Fast Marching Method
2 Generalized Fast Marching Method
3 Image segmentation
4 Numerical results

Plan
4 Numerical results

Plan
4 Numerical results

Plan
4 Numerical results

Fast Marching Method
Generalized Fast Marching
Image segmentation
Numerical results
Plan
4 Numerical results

Image segmentation
Numerical results
Evolution of a front
c ∆ t n Γt

Ωt

Γt+ ∆t

Γt

dΓt
= c nΓt
dt

Image segmentation
Numerical results
Mathematical difficulty
How to define the evolution with change of topology?

Image segmentation
Numerical results
Mathematical difficulty
Change of topology = frightening

Image segmentation
Numerical results
Level set Method
Level set formulation:
Γt = {x, u(x, t) = 0}.
Formally :
∂Γt
u(Γt , t) = 0 ⇒ ut + ∇u · = 0.
∂t
∂Γt ∇u
with = c(x)~n, ~n = − .
∂t |∇u|
Level set equation for moving fronts:
ut = c(x)|∇u|.

Image segmentation
Numerical results
Level set Method
Γt = {x, u(x, t) = 0}.
Formally :
∂Γt
u(Γt , t) = 0 ⇒ ut + ∇u · = 0.
∂t
∂Γt ∇u
with = c(x)~n, ~n = − .
∂t |∇u|
ut = c(x)|∇u|.

Image segmentation
Numerical results
Level set Method
Γt = {x, u(x, t) = 0}.
Formally :
∂Γt
u(Γt , t) = 0 ⇒ ut + ∇u · = 0.
∂t
∂Γt ∇u
with = c(x)~n, ~n = − .
∂t |∇u|
ut = c(x)|∇u|.

Image segmentation
Numerical results
Level set Method
Γt = {x, u(x, t) = 0}.
Formally :
∂Γt
u(Γt , t) = 0 ⇒ ut + ∇u · = 0.
∂t
∂Γt ∇u
with = c(x)~n, ~n = − .
∂t |∇u|
ut = c(x)|∇u|.

Image segmentation
Numerical results
Stationary approach
1
Setting u(x, t) = t − T (x) ⇒ |∇T (x)| = .
c(x)
T (x) = time when the front cross the point x.

Image segmentation
Numerical results
Stationary approach
1
Setting u(x, t) = t − T (x) ⇒ |∇T (x)| = .
c(x)
T (x) = time when the front cross the point x.

Image segmentation
Numerical results
Discretization
∆x
I
G D
∆x
B
max(TI − TG , TI − TD , 0)2 2
∆x
+ = (1)
cI
max(TI − TH , TI − TB , 0)2
Image segmentation
Numerical results
Introduction of the Narrow Band
Processing the nodes in a special ordering, one can compute the

solution in just one iteration.
This special ordering corresponds to the increasing values of T
Idea : Compute the time only near the front.

Image segmentation
Numerical results


Image segmentation
Numerical results

Far
Accepted
Narrow Band
Image segmentation
Numerical results
Far
Accepted
Narrow Band
Computation of TI on NB
Call tn the minimal time TI on the NB
and accept at the time tn the minimizing points I
The new NB is defined as the boundary of the new accepted region
Image segmentation
Numerical results
Far
Accepted
Narrow Band
Image segmentation
Numerical results
Far
Accepted
Narrow Band
Image segmentation
Numerical results
Far
Accepted
Narrow Band
Image segmentation
Numerical results
Far
Accepted
Narrow Band
Image segmentation
Numerical results
Far
Accepted
Narrow Band
Image segmentation
Numerical results
References
[Sethian]: http://math.berkeley.edu/∼sethian/
[Vladimirsky] : case c = c(x, t) > 0
[Cristiani, Falcone] : Proof of convergence

Image segmentation
Numerical results
Plan
4 Numerical results

Image segmentation
Numerical results
Dependence in time of the velocity
The time is given implicitly by the algorithm!
tn = min TI
I∈N B
Two difficulties:
tn cans be smaller than tn−1 :
tn := tn−1
tn can be very large:
⇒ Introduction of a time step ∆t :
if tn ≥ tn−1 + ∆t ⇒ tn := tn−1 + ∆t

Image segmentation
Numerical results
tn = min TI
I∈N B
Two difficulties:
tn := tn−1
if tn ≥ tn−1 + ∆t ⇒ tn := tn−1 + ∆t

Image segmentation
Numerical results
tn = min TI
I∈N B
Two difficulties:
tn := tn−1
if tn ≥ tn−1 + ∆t ⇒ tn := tn−1 + ∆t

Image segmentation
Numerical results
tn = min TI
I∈N B
Two difficulties:
tn := tn−1
if tn ≥ tn−1 + ∆t ⇒ tn := tn−1 + ∆t

Image segmentation
Numerical results
Change of the sign of the velocity

c=0
c>0 c<0
θ = −1
θ=1
Narrow Band
Introduction of a field θIn ∈ {−1, 1}

Regularisation of the velocity in space
Double Narrow band
The time is defined in each side of the front
Image segmentation
Numerical results

c=0
c>0 c<0
θ = −1
θ=1
Narrow Band

Double Narrow band
Image segmentation
Numerical results

c=0
c>0 c<0
θ = −1
θ=1
Narrow Band

Double Narrow band
Image segmentation
Numerical results

c=0
c>0 c<0
θ = −1
θ=1
Narrow Band

Double Narrow band
Image segmentation
Numerical results
Exemple
c=0 c=0
c>0 c<0 c>0 c<0
θ = −1 θ = −1
θ=1 θ=1
Narrow Band Accepted Points

Image segmentation
Numerical results
Reference
1 E. Carlini, M. Falcone, N. Forcadel, R. Monneau, Convergence of a

Generalized Fast Marching Method for a non-convex eikonal equation
(submitted, http://cermics.enpc.fr/ forcadel/publications.html).
2 E. Carlini, E. Cristiani, N. Forcadel, A non-monotone Fast Marching
scheme for a Hamilton-Jacobi equation modeling dislocation dynamics,
ENUMATH 2005.

Image segmentation
Numerical results
A straight line
c(x, t) = x1

Image segmentation
Numerical results
A circle
c(x, t) = −x1

Image segmentation
Numerical results
Two circles
c(x, t) = 1 − t

Image segmentation
Numerical results
Plan
4 Numerical results

Image segmentation
Numerical results
Image Segmentation with classical FMM
Given an image I, define the velocity as
c(x) = g(|∇I(x)|)
where g is an edge-detector function satisfying

g : [0, +∞[→ [0, +∞[, g(0) = 1, g strictly decreasing and
limr→∞ g(r) = 0.
Very efficient method
But problem of the sign of the velocity implies constaints on initial data.

Image segmentation
Numerical results
c(x) = g(|∇I(x)|)

limr→∞ g(r) = 0.

Image segmentation
Numerical results
c(x) = g(|∇I(x)|)

limr→∞ g(r) = 0.

Image segmentation
Numerical results
Image Segmentation with generalized FMM
The gradient field of −∇g(|∇(I)|) give the direction of the edges:
The vectors point towards the middle of the boundaries.

Image segmentation
Numerical results
The gradient field of −∇g(|∇(I)|) give the direction of the edges:
The vectors point towards the middle of the boundaries.

Image segmentation
Numerical results
Define the velocity as

R ||r−x||2
Q(x) h−∇g(|∇I(r)|), ~
n (x, t)i exp − 2σ 2
dr
c(x, t) = g(|∇I(x)|) R
||r−x||2
Q(x) h−∇g(|∇I(r)|), ~n(x, t)i exp − 2σ2 dr


Image segmentation
Numerical results
Plan
4 Numerical results

Image segmentation
Numerical results
Geophysics data
20 20
40 40
60 60
80 80
100 100
120 120
140 140
−20 0 20 40 60 80 100 120 140 −20 0 20 40 60 80 100 120 140
Figure: Initial data Figure: Final result

Image segmentation
Numerical results
Medical data

Image segmentation
Numerical results
Medical data
20 20
40 40
60 60
80 80
100 100
120 120
140 140
−20 0 20 40 60 80 100 120 140 160 −20 0 20 40 60 80 100 120 140 160
Figure: Initial data Figure: Final result

Image segmentation
Numerical results
From the initial condition to the final result

Image Segmentation Using A Generalized Fast Marching Method

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Image Segmentation Using A Generalized Fast Marching Method

Uploaded by

Copyright:

Available Formats

Image Segmentation using a Generalized Fast Marching

Nicolas Forcadel (ENPC-CERMICS, Paris), Carole Le Guyader (INSA

Multivariate Approximation: Theory and Applications

April 26 - May 01, 2007

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

1 Fast Marching Method

2 Generalized Fast Marching Method

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

1 Fast Marching Method

2 Generalized Fast Marching Method

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

1 Fast Marching Method

2 Generalized Fast Marching Method

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

1 Fast Marching Method

2 Generalized Fast Marching Method

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

1 Fast Marching Method

2 Generalized Fast Marching Method

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

How to define the evolution with change of topology?

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

Change of topology = frightening

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

Level set Method

Level set formulation:

Γt = {x, u(x, t) = 0}.

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

Level set Method

Level set formulation:

Γt = {x, u(x, t) = 0}.

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

Level set Method

Level set formulation:

Γt = {x, u(x, t) = 0}.

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

Level set Method

Level set formulation:

Γt = {x, u(x, t) = 0}.

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

Introduction of the Narrow Band

Processing the nodes in a special ordering, one can compute the

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

Introduction of the Narrow Band

Processing the nodes in a special ordering, one can compute the

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

Introduction of the Narrow Band

Processing the nodes in a special ordering, one can compute the

Introduction of the Narrow Band

Introduction of the Narrow Band

Introduction of the Narrow Band

Introduction of the Narrow Band

Introduction of the Narrow Band

Introduction of the Narrow Band

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

1 Fast Marching Method

2 Generalized Fast Marching Method

N. Forcadel C. Le Guyader, C. Gout Image Segmentation

Dependence in time of the velocity

The time is given implicitly by the algorithm!