Está en la página 1de 4

A visual comparision of Rigid-Transformation

performance on Images
Jigar Patel
Dept. of Electrical and Computer Engineering
Florida State University (FSU)
Tallahassee, FL, USA
jp09x@my.fsu.edu

AbstractImage registration is a fundamental task in image


processing used to match/align two or more images that are obtained
at different times, from different sensors, from different viewpoints,
or different modalities. Image registration is of importance in
medical image analysis as a preprocessing tool that makes fusing
images together meaningful by accurately and correctly mapping
two or more images together. This paper investigates the
performance of Iterative Closest Point (ICP) rigid-transformation on
monomodal images. The goal is to gain insight on the effectiveness
of the technique when applied to Images that have different
histogram distributions.
KeywordsImage registration, Medical Image Analysis, PointBased Methods, Iterative Closest Point

I. INTRODUCTION
Image registration, also known as image fusion,
matching or warping, can be defined as the process of aligning
two or more images together. The goal is to find an optimal
transformation that aligns the structures in the input images.
Image acquisition methods include but are not limited to images
from different times, sensors, viewpoints, or modalities. Once
the images are acquired the images are then aligned/registered
by setting one image as the reference image and setting all
subsequent images as the sensed images. Registration is
especially important in Medical Image Analysis as
doctors/radiologists usually want to analyze a patients scans
from multiple imaging equipment. Imaging in the medical field
is done on patients using a wide variety of equipment such as
computer tomography (CT), X-Ray, positron emission
tomography (PET), magnetic resonance imaging (MRI) to
name a few.
The scope of this paper is to do a simple comparison
of how ICP performs on different images. Points are only one
approach to feature selection, other techniques that are not
discussed are lines, regions, and templates. Points are more
desirable compared to the other techniques. It is not in the scope
of this paper to propose a novel approach or modification to the
existing algorithms.
The paper is organized as follows, Section II
introduces the ICP algorithm in detail followed by section III
which provides pseudo-implementations of the algorithm. In
section IV experiment results are discussed followed by the
conclusion in section V.

II. ITERATIVELY CLOSEST POINT ALGORITHM


A point is only considered reliable if it lies in a clearly
discernible feature, if there exists such a point or set of points,
they are called fiducial features.
Fiducial localization is a precise point within a feature. The
fiducial localization process can be based on interactive visual
identification of anatomical landmarks, or it can be done using
external markers attached to the subject. Marker-based
registration proves to be more advantageous since the process
can be more precisely automated eliminating interaction for the
purposes of registration. Examples of markers can be seen in
Figure 1. Such points that can be used as a reference for
matching the images are known as fiducial points or fiducials.

Figure 1: (a) Transaxial CT with marker (b) Transaxial MR with markers [1]

Another approach to registration is an automated approach


based on the internal markers that can be estimated from the
images. This approach is of more interest and poses more
challenges since there is not one size fits all solution due to
the wide range of image acquisition technologies as well as the
wide range of images that can be acquired from these
techniques.
Rigid transformations are also known as linear
transformations and include rotation, translation and in some
cases mirrored transformations. Rigid transformations are
global in nature and cannot model the local geometric
correspondence between images. Scaling and sheering are not
part of the rigid transformation.
ICP is one such transformation which was introduced by
Besl and McKay. ICP is a rigid-transformation technique that is
widely used. It works on the principle of having some
knowledge of the correspondence between non-collinear points
of the two images (initial guess). If at least three set of noncollinear points are known from the images then the remaining

correspondence points can be known by aligning the three points


using some sort of transformation matrix. This requires that the
correspondence of the initial points is a close match in both
images. Errors in these points will result to a mismatch in the
registered images. The algorithm performs rigid transformation
in an iterative process by assuming that every point in the
reference image corresponds to a point in the sensed image, from
this point on, the paper will use the notation:
I1 for the reference image
I2 for the sensed image
unless otherwise specified. More specifically for every point in
I1 there is a point in I2 that is close to it. Once these set of points
are found a least square rigid transformation is found to map the
point of I2 to the points in I1. The algorithm works best if the
initial guess is relatively close to the actual value. Section III
provides the pseudo code for the algorithm and discusses
methods for selecting the initial values for the points.
III. PSEUDO IMPLEMENTATION OF ITERATIVE CLOSEST
POINT ALGORITHM
The pseudo code for the algorithm is shown below.
Step 1 requires selection of three points. These points cannot be
random points since some correspondence between the points
in the two images is required. When the number of
corresponding points are few (fewer than 15 points), all of the
points can be used in determining the parameters of equation 1.
When the number of corresponding points is a very large
number the computational complexity gets intensive, this is
where picking a fewer number of points is useful (in this case
we pick three points). Using points that lie on the convex hull
is one technique of reducing the search space without
immensely reducing the probability of missing the correct
match. Once a sufficiently large number of corresponding
points has been found ICP can be performed to register the
image.
Pseudo-Code for ICP [4]
Let P = points from I1 and
Q = points from I2
Note: P and Q are M-by-2 matrix containing the location of
the corresponding points in the two images.
Step 1: Select 3 points from each set P and Q
Step 2: Determine the a-f parameters using the points from
step 1
= + +
} . 1
= + +
Step 3: For j = 1,,ns
3.1: Substitute qj into Eq. 1and find the
corresponding point qj in the reference image.
3.2: Find the point in the set P that is closest to qj.
Let pi be such a point. If d = |pi qj| < D, save that
entry in a list.
Step 4:
If n is large enough
o Move to step 5
Else

Go back to step two and re-do process or


report failure

Step 5:
Determine the transformation based on the list
created in step 3.2 using least-squares method and
mat set Q to set P and determine the
correspondence again as in step 3.
Step 5 is added into the algorithm to reduce the effect of any
noise that might exist in the image. The value of D in step 3.2
is an arbitrary value which is user defined. This process holds
for both 2-D and 3-D images with a minor change. Instead of
six parameters (a-f) the 3-D version requires computation of 12
parameters (a-l) which are given the equation 2 below.
= + + +
= + + + } . 2
= + + +
Also instead of using three non-collinear point pairs four noncoplanar point pairs are used to calculate the parameters (a-l) of
transformation of equation 2.
IV. EXPERIMENTAL RESULTS
ICP is performed on the following two images shown below
along with the histograms of each image.

Figure 2: Original Images and their histograms: Lena (top) Mammogram (bottom)

The original images were transformed using the following


parameters to acquire the sensed images shown in
Crop
Rotation
Translate

Starting point: (60,90), Window Size: 128 by 96


45 clockwise
32 by 40
Table 1: Transform Parameters to obtain Sensed Images

Figure 3: Sensed Images

Cropped
Image

1
0 0
[0
1 0]
59 89 1

1.011 0.016 0
[ 0.016
1.011 0]
57.528 90.023 1

The matrix above for each image shows the values obtained for the
translation matrix. As seen in row three of each matrix the equation obtained
is fairly close to the actual values from Table 1 where the cropped image
started at point (60,90) and the window size was [128,96] which is the size
of the registered image.

Translation Image

1
0
0
[ 0
1
0]
32 40 1

1
0
0
[ 0
1
0]
32 40 1

The translation performed on the original images was [32,40] which is


exactly what was obtained from the performance of the ICP as seen above
numerically, visually the images are shown below.

Figure 6: Matched Images (top) and the registered images (bottom)


Figure 4: Matched Images (top) and the registered images (bottom)

V. CONCLUSION
Rotation
Image

0.710
0.714 0
[ 0.714
0.710 0]
132.18 129.26 1

0.886 1.120 0
[1.120 0.886 0]
432.71 313.21 1

The matrix above for each image shows the values obtained for the rotation
matrix. This matrix is harder to compare to the actual rotation value from
Table 1 however visually it is noticed that the Lena image is correctly
registered which the Mammogram image is not registered correctly. This
shows that fewer points leads to errors in registration. The reference and
sensed Mammogram images only had three matched points which is not
sufficient to perform the registration correctly in this case.

Rigid transformations works best when the images


have distinct boundaries and corner features. Having more
information about the features of the image lead to a better
registration. A good example can be seen from the rotated
images of Lena and Mammogram images. Since the
Mammogram image does not contain distinct/unique corners it
is harder to select corresponding points in the images making it
harder to register the images.
VI. ACKNOWLEDGEMENTS
All Lena and Mammogram images were provided by
instructor, Dr. V. DeBrunner.
REFERENCES
[1]

Figure 5: Matched Images (top) and the registered images (bottom)

F. J. M., D. L. G. Hill and C. R. Maurer Jr., "Image


Registration, Chapter 8," in Handbook of Medical
Imaging, Vol. 2: Medical Image Processing and Analysis,
Bellingham, WA, SPIE Press, 2000, pp. 447-513.
[2] A. A. Goshtasby, 2-D and 3-D Image Registration for
Medical, Remote Sensing, and Industrial Applications,
Hoboken, New Jersey: Wiley & Sons, Inc., 2005.
[3]
J. Shi and C. Tomasi, "Good Feature to Track," in
Computer Vision and Pattern Recognition, Seattle, WA,
1994.

[4]

[5]

A. Myronenko and X. Song, "Point Set Registration:


Coherent Point Drift," Pattern Analysis and Machine
Intelligence, pp. 2262-2275, 2010.
R. C. Gonzalez and R. E. Woods, Digital Image
Processing, Upper Saddle River, NJ: Pearsons, 2008.

También podría gustarte