Sunteți pe pagina 1din 16

Formulating Compressed Sensing

Applications
Conclusions

Image reconstruction using Compressed


Sensing
Anna Scaife
Yves Wiaux, Laurent Jacques, Gilles Puy & Pierre Vandergheynst

CALIM09 31/03/09

Anna Scaife

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

Outline

Formulating Compressed Sensing

Applications

Conclusions

Anna Scaife

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

Deconvolution
CLEAN
Local iterative deconvolution
Matching Pursuit
Implicitly implies sparsity
MEM
Global minimization problem
Assumes an entropic prior
Both methods are flexible enough to consider a variety of bases
(Dirac, wavelet etc)

Anna Scaife

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

Sparsity

The main premise of CS is that although our signal is not


necessarily sparse in real space or Fourier space, it is
sparse or compressible in some basis.
If we consider a real signal x = {xi }1iN
and define a real basis = {i }1iN;1T
Then we can say that the decomposition = { }1T .
x =
...is spare or compressible if it contains only K << N
non-zero or significant co-efficients.

Anna Scaife

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

If we then probe the signal using m real linear


measurements (visibilities) y = {yr }1r m in some sensing
basis = {ri }1r m;1iN
and these measurements are possibly affected by some
independent and identically distributed noise:
n = {nr }1r m , so that:
y = + n, where = RmT

Anna Scaife

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

Restricted Isometry

Defining the `p norm, ||u||p =

P

Q
p
l=1 |ul |

1/p

By definition the matrix satisfies a RIP of order K if there


exists a constant K < 1 such that
The RIP
(1 K )||K ||22 || K ||22 (1 + K )||K ||22

Anna Scaife

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

Satisfying the RIP


The incoherence of the sensing matrix with the sparsity
basis will satisfy the RIP if the number of measurements
(m) is large enough relative to the sparsity K .
For radio interferometry the RIP is satisfied if
K

Cm
.
2 ln4 N

is the mutual coherence of the elements of the Fourier


basis and the elements of the sparsity basis:
=

N max |he |e0 i|.

Anna Scaife

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

Basis Pursuit
the `1 norm of the vector is simply the sum of the
absolute values of the vector
PT components:
||||1 = =1 | |.

Laplacian Prior

The Optimization Problem


min ||0 ||1 s.t. ||y 0 ||2 
Anna Scaife

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

Recovery
If the solution of the BP is then the corresponding recovered
signal is
x = .
Solutions
CS shows
that if satisfies a RIP of order 2K with
2K < 2 1 then the solution x provides an accurate
reconstruction.
It can be said to be optimal in the sense that exactly sparse
signals in the absense of noise are recovered exactly.
In the presence of noise very strong stability results are
obtained.

Anna Scaife

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

Interferometer Data

We consider 5 different sets of coverage in uv with different


% coverage of the Fourier plane
Two examples:
1
2

a field filled with multi-variate compact sources


a CMB cosmic string signal simulation
0

SNR(s,s ) = 20 log10

(ss
(s)

Anna Scaife

0)

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

Cosmic Strings
Topological defects in the CMB.
CDM cosmology.
String signal is well modelled by GGDs in wavelet space
GGDs
 h i 
v
j
j ( ) exp u
j
s-norm
() exp
s
P ||||
vj
|
||||s | u
j

Anna Scaife

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

The three methods

BP
min ||0 ||1 s.t. ||y 0 ||2 
BP+
ri x 0 and x 0 0.
min ||x 0 ||1 s.t. y =
SBP
min ||0 ||s s.t. ||y Wcmb ri s 0 ||2 

Anna Scaife

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

Compact Sources

Anna Scaife

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

Cosmic Strings

Anna Scaife

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

SNR

Anna Scaife

Image reconstruction using Compressed Sensing

Formulating Compressed Sensing


Applications
Conclusions

Conclusions

A new framework for image reconstruction in interferometry


Simple BP provides the same image fidelity as CLEAN
BP is more rapid than CLEAN in terms of no. iterations and
computation time
Prior statistical knowledge of the signal can greatly
improve the reconstruction

Anna Scaife

Image reconstruction using Compressed Sensing