Total Variation and Geometric Regularization For Inverse Problems

Total Variation and Geometric Regularization for
Inverse Problems
Regularization in Statistics
September 7-11, 2003
BIRS, Banff, Canada
Tony Chan
Department of Mathematics, UCLA
Outline
TV & Geometric Regularization (related concepts)

PDE and Functional/Analytic based
Geometric Regularization via Level Sets Techniques
Applications (this talk):
Image restoration
Image segmentation
Elliptic Inverse problems
Medical tomography: PET, EIT
Regularization: Analytical vs Statistical

Analytical:
Controls smoothness of continuous functions
Function spaces (e.g. Sobolov, Besov, BV)
Variational models -> PDE algorithms
Statistical:
Data driven priors
Stochastic/probabilistic frameworks
Variational models -> EM, Monte Carlo
Taking the Best from Each?

Concepts are fundamentally related:
e.g. Brownian motion Diffusion Equation
Statistical frameworks advantages:

General models
Adapt to specific data
Analytical frameworks advantages:

Direct control on smoothness/discontinuities, geometry
Fast algorithms when applicable
Total Variation Regularization

TV (u ) | u | dx
Measures variation of u, w/o penalizing discontinuities.

|.| similar to Huber function in robust statistics.
1D: If u is monotonic in [a,b], then TV(u) = |u(b) u(a)|,
regardless of whether u is discontinuous or not.
nD: If u(D) = char fcn of D, then TV(u) = surface area of D.
(Coarea formula)
Rn
f | u | dx (
f ds)dr
{u r }
Thus TV controls both size of jumps and geometry of boundaries.

Extensions to vector-valued functions
Color TV: Blomgren-C 98; Ringach-Sapiro, Kimmel-Sochen
The Image Restoration Problem

A given
Observed image
Related to
True Image
Through
Blur
And
Noise
Initial
z K u n
Blur
Blur+Noise
Inverse Problem: restore u, given K and statistics for n.

Keeping edges sharp and in the correct location is a key problem !
Total Variation Restoration

TV (u ) | u | dx
Regularization:
Variational Model:
1
2
min f (u ) TV (u ) || Ku z ||
u
2
* First proposed by Rudin-Osher-Fatemi 92.

* Allows for edge capturing (discontinuities along curves).
* TVD schemes popular for shock capturing.
Gradient flow:
u
ut g (u )
( K Ku K * z )
| u |
anisotropic diffusion
u
0
n
data fidelity
Comparison of different methods for signal denoising & reconstruction
Image Inpainting (Masnou-Morel; Sapiro et al 99)

Disocclusion
Graffiti Removal
Unified TV Restoration & Inpainting model

(C- J. Shen 2000)
J [u ] | u | dxdy | u u 0 |2 dxdy,
2E
E D
u
e (u u 0 ) 0,

| u |
z E;
0 , z D.
TV Inpaintings: disocclusion
Examples of TV Inpaintings
Where is the Inpainting Region?
TV Zoom-in
Inpaint Region: high-res points that are not low-res
Edge Inpainting
edge tube T
No extra data are

needed. Just inpaint!
Inpaint region: points away from Edge Tubes
Extensions
Color (S.H. Kang thesis 02)
Eulers Elastica Inpainting (C-Kang-Shen 01)
Minimizing TV + Boundary Curvature
Mumford-Shah Inpainting (Esedoglu-Shen 01)

Minimizing boundary + interior smoothness:
min
u ,S
( | u | dS
2
Geometric Regularization
Minimizing surface area of boundaries and/or
volume of objects
Well-studied in differential geometry: curvaturedriven flows
Crucial: representation of surface & volume
Need to allow merging and pinching-off of
surfaces
Powerful technique: level set methodology
(Osher/Sethian 86)
Level Set Representation (S. Osher

- J. Sethian 87)
n
C ( x, y ) | ( x, y ) 0
Inside C
Outside C
Outside C
C= boundary of an open domain

Normal n
, Curvature K div
| |
| |
Example: mean curvature motion
* Allows automatic topology changes, cusps, merging and breaking.
Originally developed for tracking fluid interfaces.
Application: active contour

giving an image u0 :
evolve a curve C to detect objects in u0
the curve has to stop on the boundaries of the objects
Initial Curve
Evolutions
Detected Objects
Basic idea in classical active contours

Curve evolution and deformation (internal forces):
Min Length(C)+Area(inside(C))
Boundary detection: stopping edge-function (external forces)
g 0, g , lim g (t ) 0
t
Example: g (| u |)
0
1
p
1 | G u 0 |
Snake model (Kass, Witkin, Terzopoulos 88)

1
inf F (C ) | C ' ( s ) | ds g (| I (C ( s )) |)ds

2
Geodesic model (Caselles, Kimmel, Sapiro 95)

1
inf F (C ) 2 | C ' ( s ) | g (| I (C ( s )) |)ds

C
Limitations
- detects only objects with sharp edges defined by gradients
- the curve can pass through the edge
- smoothing may miss edges in presence of noise
- not all can handle automatic change of topology
Examples
A fitting term without edges

|
u
c
|
dxdy
0
1
inside ( C )
where
Fit > 0
|
u
c
|
dxdy
0
2
outside ( C )
c1 average(u0 ) inside C
c2 average(u0 ) outside C
Fit > 0
Fit > 0
Fit ~ 0
Minimize: (Fitting +Regularization)

Fitting not depending on gradient
detects contours without gradient
An active contour model without edges

(C. + Vese 98)
Fitting + Regularization terms (length, area)
inf F (c1 , c2 , C ) | C | Area (inside(C ))
c1 ,c2 ,C
inside ( C )
| u0 c1 | dxdy
2
|
u
c
|
dxdy
0
2
outside ( C )
C = boundary of an open and bounded domain

|C| = the length of the boundary-curve C
Mumford-Shah Segmentation 89
min
u ,S
( | u | (u u0 ) ) dS
2
S=edges
MS reg: min boundary + interior smoothness

CV model = p.w. constant MS
Variational Formulations and Level Sets

(Following Zhao, Chan, Merriman and Osher 96)
C ( x, y ) : ( x, y ) 0
Length | C | | H ( ) |
The Heaviside function
1, if 0
H ( )
0, if 0
Area (inside(C )) H ( )dxdy
u( x, y ) c1H ( ( x, y )) c2 (1 H ( ( x, y )))
The level set formulation of the active contour model
inf F ( c1 , c2 , )
c1 ,c2 ,
F (c1 , c2 , ) | H ( ) | H ( )dxdy
| u0 ( x, y ) c1 |2 H ( )dxdy | u0 ( x, y ) c2 |2 (1 H ( ))dxdy
The Euler-Lagrange equations

inf F (c1 , c2 , )
c1 ,c2 ,
Using smooth approximations for

the Heaviside and Delta functions
Equations for c1 and c2

c1 ( )
H ( )dxdy
H ( )dxdy
u (1 H ( ))dxdy
0
, c2 ( )
(1 H ( ))dxdy
Equation for (t , x, y )
2
2
(u0 c1 ) (u0 c2 )
( ) div
t
| |
(0, x, y ) 0 ( x, y )
Experimental Results
Evolution of C Averages (c1 , c2 )
Advantages
Automatically detects
interior contours!
Works very well for
concave objects
Robust w.r.t. noise
Detects blurred contours
The initial curve can be
placed anywhere!
Allows for automatical
change of topolgy
A plane in a noisy environment
Europe nightlights
Multiphase level set representations and partitions

allows for triple junctions, with no vacuum and no
n
overlap of
phases
( ,..., ) 2 phases
4-phase segmentation
2 level set functions
Curves :
{1 0} { 2 0}
Curves :
{ 0}
2-phase segmentation
1 level set function
1 0
1 0
2 0
2 0
1 0
2 0
1 0
2 0
Example: two level set functions and four phases

2 level set functions (1 , 2 ) 4 phases or segments :
1 0, 2 0,
1 0, 2 0,
(1 , 2 )
1 0, 2 0,
1 0, 2 0
The level set functions
c (c11 , c10 , c01 , c00 ) Constant vector

u c11 H (1 ) H ( 2 ) c10 H (1 )(1 H ( 2 ))
c01 (1 H (1 )) H ( 2 ) c00 (1 H (1 ))(1 H ( 2 ))
Energy
Inf F (c, ) | u0 c11 |2 H (1 ) H ( 2 )dxdy | u0 c10 |2 H (1 )(1 H ( 2 ))dxdy
( c , )
| u0 c01 |2 (1 H (1 )) H ( 2 )dxdy | u0 c00 |2 (1 H (1 ))(1 H ( 2 ))dxdy
| H (1 ) | | H ( 2 ) |
An MRI brain image
Phase 11
Phase 10
Phase 01
mean(11)=45
mean(10)=159
mean(01)=9
Phase 00
mean(00)=103
min
u ,S
( | u |2 (u f ) 2 ) dS
S
References for PDE & Level Sets in Imaging

* IEEE Tran.
Image Proc. 3/98, Special Issue on PDE Imaging
* J. Weickert 98: Anisotropic Diffusion in Image Processing

* G. Sapiro 01: Geometric PDEs in Image Processing
Aubert-Kornprost 02: Mathematical Aspects of Imaging Processing
Osher & Fedkiw 02: Bible on Level Sets
Chan, Shen & Vese Jan 03, Notices of AMS

Total Variation and Geometric Regularization For Inverse Problems

Uploaded by

Copyright:

Available Formats

Total Variation and Geometric Regularization For Inverse Problems

Uploaded by

Document Information

Original Description:

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

Total Variation and Geometric Regularization For Inverse Problems

Uploaded by

Copyright:

Available Formats

Total Variation and Geometric Regularization for

TV & Geometric Regularization (related concepts)

Regularization: Analytical vs Statistical

Taking the Best from Each?

Statistical frameworks advantages:

Analytical frameworks advantages:

Total Variation Regularization

Measures variation of u, w/o penalizing discontinuities.

Thus TV controls both size of jumps and geometry of boundaries.

The Image Restoration Problem

Inverse Problem: restore u, given K and statistics for n.

Total Variation Restoration

* First proposed by Rudin-Osher-Fatemi 92.

Comparison of different methods for signal denoising & reconstruction

Image Inpainting (Masnou-Morel; Sapiro et al 99)

Unified TV Restoration & Inpainting model

Where is the Inpainting Region?

Inpaint Region: high-res points that are not low-res

No extra data are

Inpaint region: points away from Edge Tubes

Mumford-Shah Inpainting (Esedoglu-Shen 01)

Level Set Representation (S. Osher

C= boundary of an open domain

Application: active contour

Basic idea in classical active contours

Snake model (Kass, Witkin, Terzopoulos 88)

inf F (C ) | C ' ( s ) | ds g (| I (C ( s )) |)ds

Geodesic model (Caselles, Kimmel, Sapiro 95)

inf F (C ) 2 | C ' ( s ) | g (| I (C ( s )) |)ds

A fitting term without edges

Minimize: (Fitting +Regularization)

detects contours without gradient

An active contour model without edges

Fitting + Regularization terms (length, area)

inf F (c1 , c2 , C ) | C | Area (inside(C ))

C = boundary of an open and bounded domain

MS reg: min boundary + interior smoothness

Variational Formulations and Level Sets

The Heaviside function

Area (inside(C )) H ( )dxdy

The Euler-Lagrange equations

Using smooth approximations for

Equations for c1 and c2

A plane in a noisy environment

Multiphase level set representations and partitions

Example: two level set functions and four phases

The level set functions

c (c11 , c10 , c01 , c00 ) Constant vector

| u0 c01 |2 (1 H (1 )) H ( 2 )dxdy | u0 c00 |2 (1 H (1 ))(1 H ( 2 ))dxdy

An MRI brain image

References for PDE & Level Sets in Imaging

Image Proc. 3/98, Special Issue on PDE Imaging

* J. Weickert 98: Anisotropic Diffusion in Image Processing

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.