02.12.2005 muscle-wp5&7 meeting priors, syntax and semantics in variational level-set approachs...
Post on 22-Dec-2015
217 views
TRANSCRIPT
02.12.2005 MUSCLE-WP5&7 meeting
Priors, Syntax and Semantics
in Variational Level-Set
Approachs
TAU-VISUAL: Nir Sochen & Nahum Kiryati
Based on works with Tammy Riklin-Raviv
02.12.2005 MUSCLE-WP5&7 meeting
Segmentation as an inference problem
Given: a visual data – an image
Infer: (projection of) the scene structure
02.12.2005 MUSCLE-WP5&7 meeting
Maximum Likelihood
ws= World’s State
Then: find the ws that best
explains the data
Ws=argmax P(data|ws)
Problem: UFO
02.12.2005 MUSCLE-WP5&7 meeting
Bayesian formalism
ws= World’s State
Then: Find a ws that explains
the data and has high
probability
ws=argmax P(ws|data)=
argmax P(data|ws)P(ws)
02.12.2005 MUSCLE-WP5&7 meeting
Image Segmentation: General
Task:
Partition the image to its significant domains
Or
Find the relevant shapes
02.12.2005 MUSCLE-WP5&7 meeting
Segmentation Functional
:f is the input imagewhere
C is the segmenting boundary
u is the average color
fidelity
xxx dcdufdufEC
outC
in22
outside
2
inside
shape evolving-prior)(length
inu
outu
Compatibility
with the prior
Smoothnessand length
02.12.2005 MUSCLE-WP5&7 meeting
Image Segmentation: Assumptions
Assumptions:
There are objects and background .
Object pixels: Normal probability dist .
Background pixels: Another normal dist .
Boundary: Close smooth curve
02.12.2005 MUSCLE-WP5&7 meeting
Image Segmentation: Modelling1
)()()(
))(())(()|(
semanticsPSyntaxPCP
xfPxfPCfP
CC
xout
xin
0
0
: Foreground
: Background
C : Separating curve
02.12.2005 MUSCLE-WP5&7 meeting
Image Segmentation: Modelling2
)},(exp{)(
)}(exp{)(
}2/))((exp{))(( 22//
pC
C
outinoutin
CCdNsemanticsP
ClengthNsyntaxP
uxfNxIP
02.12.2005 MUSCLE-WP5&7 meeting
Image Segmentation: steps
Step 0: InitializeStep 1: Updating the constants u_in, u_out as the meanValues of the respective regions
Step 2: Maximization of the posterior w.r.t. C.
Denote E = -log P then
argmax P(C, u_in, u_out) = argmin E(C, u_in, u_out)
Step 3: while not converge go to step 1
02.12.2005 MUSCLE-WP5&7 meeting
Image Segmentation: math
(no shape prior yet)
0 0
)(())((log))((log
),,(
x
out
x
in
outin
ClengthxfPxfP
uuCE
02.12.2005 MUSCLE-WP5&7 meeting
LS Framework for MS
Let : be a level set function which embeds
the contour: 0 xxC
A piecewise constant segmentation of an input image f can be
obtained by minimizing the functional:
dxHHufHufuuECV
1,, 22 (3)
Where H denotes the Heaviside function:
othewise
H0
01 (4)
Chan & Vese Trans. IP 01
02.12.2005 MUSCLE-WP5&7 meeting
LS Framework for MS (Cont.)
Chan & Vese 01
Let
z
zH arctan2
12
1 be a smooth approximation of the
Heaviside function. It’s derivative z
zHz
takes the form: 22
1
zz
(7)
02.12.2005 MUSCLE-WP5&7 meeting
LS Framework for MS (Cont)
The Euler-Lagrange equation for this functional can be implemented
by the following gradient descent :
22 ufufdiv
t
E
(5)
where the scalars uu , are updated in alternation
with the level set evolution:
dxH
dxHxfu
dxH
dxHxfu
1
1
uu , are the mean gray value in the input image inside
and outside C respectively.
(6)
02.12.2005 MUSCLE-WP5&7 meeting
Shape prior + transformation
pT
priorshapeC
outC
in EcdufdufE )(length2
outside
2
inside
xx
),( pTD priorshapeE
02.12.2005 MUSCLE-WP5&7 meeting
Segmentation with Prior
02.12.2005 MUSCLE-WP5&7 meeting
Summary:
There is much to gain from exploiting
The relation statistics <-> variational
approaches
F I N