Welcome to Scribd, the world's digital library. Read, publish, and share books and documents. See more
Download
Standard view
Full view
of .
Save to My Library
Look up keyword
Like this
1Activity
0 of .
Results for:
No results containing your search query
P. 1
Gradient Based Optimization of an Emst Image

Gradient Based Optimization of an Emst Image

Ratings: (0)|Views: 17 |Likes:
Published by Shanti Prasad
proj
proj

More info:

Published by: Shanti Prasad on Jan 08, 2010
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

01/08/2010

pdf

text

original

 
GRADIENT BASED OPTIMIZATION OF AN EMST IMAGE REGISTRATION FUNCTION
 Mert R. Sabuncu, Peter J. Ramadge
Princeton UniversityDepartment of Electrical EngineeringPrinceton, NJ 08544
ABSTRACT
This paper examines the problem of registering images us-ing an information theoretic metric (e.g., entropy) estimated using a Euclidean Minimum Spanning Tree (EMST). Theobjective is to find an extremum of the metric with respect to a vector of free parameters. One of the major difficulties posed by such graph theoretic metrics is concurrently ob-taining gradient information when the metric is computed.Obtaining the gradient is a first step in efficiently optimizingthe metric. Our main contribution is to show how to obtaina gradient-based descent direction from the EMST metric.We also indicate how this can be used for optimizing im-age registration over a vector set of parameters and providesome preliminary experimental results.
1. INTRODUCTION
Registration or alignment of images is a fundamental prob-lem in image processing. The ultimate goal is usually tocompare or fuse the information contained in images cap-tured at different times, by different sensors (multi-modal)or from different viewpoints. Typical applications includemedical/biologicalimageanalysis, computervisionandmil-itary applications.Recently, information theoretic registration metrics, e.g.mutual information of pixel intensities, have been proposedand extensively studied [10], [8]. These metrics are well-suited to multi-modal signal registration problems. Mostcommonly, the algorithms rely on histograms to computethe information theoretic metric. An alternative approachthat has produced promising results is based on entropicspanning graphs [3]. However, this approach has a highcomputational cost and it has not been clear how to con-currently obtain derivative information. Hence no efficientvector optimization method has been reported based on thistype of entropy metric. This paper focuses on this problemfor the special case of Euclidean Minimum Spanning Trees(EMST) and indicates how to select a gradient-based de-scent direction for the EMST metric. This leads to a more
Research partially supported by
S
iemens
C
orporate
R
esearch, Prince-ton N.J.
efficient vector optimization algorithm. Although we re-strict attention to the EMST metric, the basic principles of our approach should extend to other entropic graph func-tions.The remainder of the paper is organized as follows. Sec-tion 2 provides some background on graph theoretic entropyestimators and Section 3 defines the EMST registration met-ric. Section 4 formulates the underlying optimization prob-lem and subsection 4.1 proposes a gradient-based solution.Details of the proposed method applied to image registra-tion is provided in Section 5 and some preliminary experi-mental results are presented in subsection 5.1.
2. ENTROPIC SPANNING GRAPHS
It is possible to estimate entropic measures of a probabilitydensity based on computing specific graphs (e.g. a mini-mum spanning tree) on independent samples from the den-sity [9]. In [11], Yukich and Redmond provide a generalframework to obtain convergence results of some Euclid-ean length functionals for specific graphs. Based on thisframework, Hero et al. [2] present the following result toestimate the Renyi entropy
α
of an underlying p.d.f: Let
n
=
{
z
1
,...,z
n
}
be
n
iidsamplesdrawnfromaLebesguedensity
,
G
(
E,
n
)
be a graph with edges weighted by theEuclidean distance between their endpoints and
G
(
n
)
de-noteafamilyofgraphswithaspecifictopologicalconstraint(e.g. spanning trees, k-neighbor graphs, etc.). Define thegraph weight function
γ
(
G
) =
e
E
|
e
|
γ
and the mini-mum graph weight (MGW) function
γ
(
n
) = min
G
∈G
(
Z
n
)
γ
(
G
)
,
Then:
lim
n
→∞
log(
γ
(
n
)
n
α
) =
α
(
) +
c
a
.
s
.,
(1)where
c
is a constant that depends only on
G
(
n
)
, i.e., thetopological constraint. Thus the MGW function can be usedas an asymptotically unbiased and strongly consistent es-timator of the
α
-Renyi entropy. Compared to the popular
 
histogram based “plug-in” estimation techniques [1], thismethod has a faster convergence rate when
is a smoothfunction [2]. In the remainder of the paper, we restrict at-tention to graphs
G
that are spanning trees. In this case, theMGW function is equal to the EMST weight.
3. THE EMST REGISTRATION METRIC
Inspiredbyinformationtheoreticregistrationtechniques[8],and result (1), entropic graphs have recently been usefullyemployed for multi-modal registration [3],[7] and [4]. Thebasic framework of this approach is as follows. For im-ages,
i
(
x,y
)
,
i
= 1
,
2
, and
d
a positive even integer, de-fine
i
(
x,y
) =
(
i
(
x
,y
) : (
x
,y
)
(
x,y
))
R
d/
2
where
(
x,y
)
R
2
is called the “neighborhood” of 
(
x,y
)
and
is the “featurefunction. A trivial but useful exam-ple is
i
=
i
(
x,y
)
where
is the identity function and
(
x,y
) = (
x,y
)
. In this case pixel intensity values arethe “features”. Information theoretic registration methodsare based on treating samples of the joint signal
(
1
,
2
)
R
d
as realizations of a random variable. Intuition suggeststhat samples from
1
and
2
should become “dependent” asthe two images are correctly aligned. This might be mea-sured, for example, by mutual information or joint entropyor, based on (1), by an entropic spanning graph [4], [7].In particular, for a fixed feature function
and constant
α
,we define the EMST registration function to be the EMSTweight
γ
(
n
)
, where
n
is a set of 
n
samples of 
(
1
,
2
)
.Research indicates that using higher dimensional features,i.e. not only pixel intensity values, makes the approachmore robust and accurate by incorporating spatial informa-tion. In contrast to the “plug-in” entropy estimators, theEMST registration function can be employed when high di-mensional features are used [7]. To date, the two major dis-advantages of using the EMST metric are the computationalbottleneck of computing the metric and it was not clear howgradient information could be obtained concurrently fromthe computation. We address the second of these issues be-low.
4. OPTIMIZING THE EMST METRIC
Image registration problems typically consist of three ma- jor components. The
transformation space
determines theallowed spatial transformation applied to the images. Thiscomponent is highly application dependent. Examples arerigid-body, affine and deformable transformations. The
reg-istration function
quantifies the similarity between two im-ages under a given transformation. Some examples are mu-tual information, correlation ratio, the EMST function, etc.The
optimization method 
searches for the optimum transfor-mation that maximizes the similarity between the images.Consider the problem of aligning images
i
(
x,y
)
,
i
=1
,
2
, using the EMST metric with fixed sampling locations
R
2
and feature function
(
.
)
. Let the spatial transfor-mation
:
R
2
be parameterized by
m
parameters,i.e.
t
where
t
= (
t
1
,...,t
m
)
. For example, a rigid-bodytransformation
R
t
has three parameters: two shifts
(
t
x
,t
y
)
and a rotation
θ
. It can be expressed as:
R
t
(
x,y
) =
cos
θ
sin
θ
sin
θ
cos
θ
xy
+
t
x
t
y
.
(2)Let
(
t
)
denote the EMST weight of 
(
t
) =
{
(
1
(
x,y
)
,
2
(
t
(
x,y
))):
(
x,y
)
}
.
Note that in the rest of the paper we will assume
γ 
= 1
.The choice of a different value for
γ 
may change the regis-tration result, but the methods discussed below extend in astraightforward way to these cases. In this framework, theregistration problem boils down to the following optimiza-tion problem:
t
opt
= argmin
t
(
t
)
.
Typically this problem is non-convex. Thus finding theglobal optimum is a difficult task. For illustration, Figure 1shows the profile of the EMST registration function withrespect to the translation along the x-axis for an image pair.
-10-505105.566.577.588.599.510
Translation along the x-axis (in pixels)
   E   M   S   T  m  e   t  r   i  c
Fig. 1
. Profile of the EMST registration function
4.1. Obtaining a Descent Direction
Descent optimization methods use a computed search di-rection that guarantees a local decrease in the value of theoptimization function. Using the gradient or an estimateof the gradient is a common way to determine this direc-tion. These methods usually assure convergence to a localextremum of the optimization function. To obtain a globalextremum, multiple-start and multi-resolution approaches,[6], can be used to augment the basic method in an attemptto avoid getting stuck in an undesired local extremum.The difficulty, however, in employing a gradient basedapproach for the EMST metric is that, in general, the EMSTmetric is not differentiable. For example, consider the ver-texset
=
{
v
1
,v
2
,v
3
}
withedgesandparameterizedweights:
 
|
e
12
|
= 1
2
t
,
|
e
23
|
=
t
+1
and
|
e
13
|
=
2+3
t
. It’s easyto show that at
t
= 0
, the MST consists of 
e
23
and
e
13
,whereas at
t
= 0
+
,
e
12
and
e
13
belong to the MST. Thus
dW 
(0
)
/dt
= 4
and
dW 
(0
+
)
/dt
= 1
. Since the left andright derivatives are not equal, the derivative of the EMSTweight does not exist at
t
= 0
.Our main result, Theorem 1, resolves the above issue byproviding a means to quickly determine a descent directionfor the EMST metric. Let
t
0
denote the edges that belongto an EMST of 
(
t
0
)
. Let
u
be an
m
-dimensional unitvector and
u
denote the directional derivative.
Theorem 1
If 
e
E
t0
u
|
e
(
t
0
)
|
<
0
 , then
ǫ >
0
suchthat 
(
t
0
+
h
u
)
(
t
0
)
 for all
0
h
ǫ
.
Based on the previous theorem, it’s easy to show that adescent direction for a EMST is
u
d
=
e
E
t0
∇|
e
(
t
0
)
|
,
(3)Hence, computation of a descent direction simply involvestheadditionalcomputationofthegradientoftheedgeweights.This requires a negligible amount of additional computationcompared to the computation of the actual EMST weightand it can be done concurrently with that computation.
5. APPLICATION IN IMAGE REGISTRATION
As a concrete application of the above ideas, consider theproblem of rigidly registering two images
1
and
2
usingpixel intensity values as features. For a given set of transfor-mation parameters
t
0
= (
t
x
0
,t
y
0
,θ
0
)
and fixed set of sam-plinglocations
R
2
theregistrationmetricisequaltothetotalweightoftheEMSTof 
{
(
x,y
) = (
1
(
x,y
)
,
t
0
2
(
x,y
)) :
(
x,y
)
}
where
t
0
2
(
x,y
) =
2
(
R
t
0
(
x,y
))
. Let
e
=(
(
x
1
,y
1
)
,
(
x
2
,y
2
))
be an edge in the computed EMST.Define
∇|
e
|
=
∂ 
|
e
|
/∂t
x
;
∂ 
|
e
|
/∂t
y
;
∂ 
|
e
|
/∂θ
and
t
0
2
(
x,y
) = (
∂I 
t
0
2
(
x,y
)
/∂x
;
∂I 
t
0
2
(
x,y
)
/∂y
)
.
Then it’s easy to show that:
∇|
e
(
t
0
)
|
= 1
/
|
e
(
t
0
)
|
(
t
0
2
(
x
1
,y
1
)
t
0
2
(
x
2
,y
2
))
(
t
0
2
(
x
1
,y
1
)
t
0
2
(
x
2
,y
2
))
R,
where
R
=
1 0
y
1
y
2
0 1
x
2
x
1
. Plugging this in Equa-tion 3 gives us a descent direction for each iteration of theformulated optimization problem.We have implemented an iterative optimization schemefor the EMST image registration metric based on the above.In our implementation we employ a “pyramid approach” toavoid “getting trapped” in local minima. Thus the algorithmstarts with registering coarse (low resolution) representa-tions of the images. We use Gaussian blurring and uniformsubsampling to obtain these low resolution images. Thealignment results obtained from the coarse level are thenused to initialize the registration at the next level. The algo-rithm terminates once full resolution is reached.For a given set of transformation parameters
t
0
we cancomputeadescentdirection
u
d
byconsideringonlytheedgesthat belong to the EMST
(
t
0
)
. For any other
t
, the EMSTmetric by definition satisfies
(
t
)
e
E
(
t
0
)
|
e
(
t
)
|
andmovinginadirectionthatdecreasesthetotalweightof 
(
t
0
)
is equivalent to decreasing an upper bound on the EMSTfunction. Hence, in our approach we propose to “updatethe EMST every
k
steps instead of recomputing it at everyiteration. This accelerates the registration algorithm by re-ducing the total number of times we run Kruskal’s MSTalgorithm [5]. The following is the pseudo-code for the im-plementation:
0
Initialize transformation parameters with zero:
t
0
=
0
1
Starting from the lowest resolution, at each level of theimage pyramid:
Initialize transformation parameters:
t
=
t
0
.
For
iter
= 1
...maxIter
if (
iter
%
k
== 1
)run Kruskal on samplesfrom
(
1
,
t
2
)
.
Compute
u
d
.
Update transformation parameters:
t
=
t
+
stepSize
u
d
.
t
0
=
t
.
2
Warp floating image with
t
.
5.1. Experimental results
We present some preliminary experimental results obtainedusing the pair of images
1
shown in Figure 2. Since both im-ages are artificially generated ground truth for the alignmentis known. Prior to the experiment, both images were cor-rupted with i.i.d Gaussian noise with a variance
0
.
1
timesthe maximum value of the original image and one of theimages was transformed using a rigid body transformationwith known parameters (
t
x
,t
y
,θ
). Table 1 provides the re-sults obtained using the proposed algorithm with three lev-els in the image pyramid. The RMS error between the cor-rect and recovered values and the average recovered trans-formationparametersarelistedforthreedifferentcases. These
1
Images obtained from http://www.bic.mni.mcgill.ca/brainweb/ 

You're Reading a Free Preview

Download
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->