Professional Documents
Culture Documents
An Iterative Dynamic Programming Approach to 2-D Phase Unwrapping
An Iterative Dynamic Programming Approach to 2-D Phase Unwrapping
2-DPHASE UNWRAPPING
Lei fink, BrendanJ. FEY, RalfKoetter', David C. Munson, Jr?
1. Dept. of ElecUical & Computer Engineering. University of Illinois at Urbana-Champaign. Urbana, IL 61801. USA
2. Dept. of Electrical & Computer Engineering,University of Toronto. Toronto, Ontario M5S 3G4. Canada
ABSTRACT flow [7] methods. The minimum norm method defines and
minimizes a norm that penalizes nonsmooth solutions. For
In this paper, we consider a novel Bayesian approach an appropriate 12-norm, the minimization can be efficiently
to 2-D phase unwrapping. The phase is unwrapped accord- computed by FFT, DCT or multigrid methods. However, the
ing to a maximum a posteriori (MAP)rule, where the esti-
least-squares solution can deviate substantially from the true
mate is made through a form of 2-D dynamic programming. phase if noise is large or data is discontinuous [SI. In the
The approach uses structured iterated conditional modes to branch cut method, the path-independent regions are identi-
achieve good performance without examining a large num-
fied and unwrapping is performed within these regions. The
ber of states in the dynamic system. We analyze the perfor- solution is robust to sheared or noisy phase data in general.
mance of the approach by transforming the problem to one
The network flow method converts the phase unwrapping
of decoding a convolutional code. An example with seven
problem to a discrete optimization problem using network
states in the dynamic program is given. We derive an ap- flow theory. Here, the goal of optimization is to minimize
proximate upper bound for probability of pixel error based
overall discontinuity. An appropriate choice of discontinu-
on a Gaussian Markov random field model. Monte Carlo
ity function is important.
simulation results show that the bound offers a good ap-
We proposed a new iterative dynamic programming ap-
proximation to the probability of error. A comparison with
proach to 2-D phase unwrapping in [9]. We take a Bayesian
other phase unwrapping techniques on a real data set sug-
point of view. The 2-D optimization is approximated by an
gests that the new approach is superior.
iterative set of 1-D dynamic programs, typically along rows
and columns of the image. This paper provides an analysis
1. INTRODUCTION of this new approach. In particular, we derive an approxi-
mate upper hound for the probability of pixel error.
Interferometric processing has been widely used for terrain This paper is organized as follows. Section 2 defines
height estimation in synthetic aperture radar (SAR) Ill, wa- a mathematical model for phase unwrapping. The iterative
terlfat decomposition in magnetic resonance imaging (MRI) dynamic programming approach is introduced in Section 3.
[2], and measurement of wavefront distortion in adaptive In Section 4, the performance of the algorithm is analyzed.
optics [3]. In these applications. the phase in a complex In Section 5. experimental results are presented. Finally,
number C = A ( z ,y) exp[iZr$(z, y)] is related to a phys- conclusions are given in Section 6.
ical quantity of interest. In order to extract phase informa-
tion. 2m,!~= tan-'(Im[C]/Re[C]) is computed. This prin-
ciple phase value differs from the true phase 2 4 by an un- 2. MATHEMATICALMODEL
known integer multiple of 271. which makes phase unwrap-
ping necessary to retrieve the physical information from the Suppose the unwrapped phase can be modeled as a first-
measurement. Two-dimensional phase unwrapping is es- order Gaussian Markov random field (GMRF) [lo]:
pecially difficult compared to one-dimensional unwrapping
due to the path-independent requirement. which is math-
ematically equivalent to the imposition of zero-curl con-
straint 141.
Much research has been done on 2-D phase unwrap-
ping. Approaches mainly can he divided into three cat-
egories: minimum norm 151, branch-cut 161, and network where p(r$',') is the distribution of the reference pixel (1,l)
This work was supported under NSF gant CCR01-05719. which offsets all other phases by The normalized phase
L = argmpp(*lE)p(E) d s ( i , j ) = min[d,,(i,j - 1) + A d 8 , , 8 ( i , j ) ] ,
8'
(4)
111 - 830
rf2
Fig. 2. Definition of c
k, 4 CIc2 cl s lqk,
our derivation. we use p = 0 and assume perfect feedback, Fig. 3. Trellis of the binary linear program
which means no errors from either previous rows or itera-
tions affect the current decision. Figure 2 uses a window
of size 2 as an example. We can easily generalize to larger model satisfying the assumptions can be used. In the case
windows. Define c3, c2, c1 and c,, as of a GMRF, Pd can be approximated by
J;i
pd 2Q(s), (14)
I11 - 831
7. REFERENCES
I11 - 832