Escolar Documentos
Profissional Documentos
Cultura Documentos
Abstract: Image registration has great significance in segmentation. In Image analysis technique, the same input
medicine. Hence numerous techniques were developed. This gives out fairly detail description of the scene whose image
paper introduced a method for medical image registration using is being considered. Hence the image analysis algorithms
Fast Walsh-Hadamard transform and measured their perform registration as a part of it towards producing the
performance using correlation coefficient and time taken for description. Also in single subject analysis, the statistical
registration. This algorithm can be used to register images of analysis is done either before or after registration. But in
the same or different modalities. Each image bit is lengthened
group analyses, the statistical analysis is done after
in terms of Fast Walsh-Hadamard basis functions. Each basis
function is a concept of determining various aspects of local
registration. Though registration transforms the different
structure, e.g., horizontal edge, corner, etc. These coefficients sets of data into one co-ordinate system, it’s required for
are normalized and used as numerals in a chosen number comparison or integration.
system which allows one to form a distinct number for each type Generally registration is the most difficult tasks among
of local structure. The research outcomes show that Fast Walsh all in image processing. It is because aligning images to
-Hadamard transform accomplished better results than the overlap the common features and differences if any are to be
conventional Walsh transform in time domain. In addition Fast emphasized for immediate visibility to the naked eye.
Walsh-Hadamard transform established medical image There is no general registration [1-17] algorithm that
registration consuming less time and more correlation work reasonably well for all images. A suitable registration
coefficient. Since in Medical Images information is very algorithm for the particular problem must be chosen or
important than time, hence Correlation Coefficient is used as a
developed, as they are adhoc in nature. The algorithms can
measure. It proves that Fast Walsh Hadamard Transform is
be incorporated explicitly or implicitly or even in the form
better than Walsh Transform in terms of both the measures.
of various parameters. This step determines the success or
Keywords: Walsh Transform, Fast Walsh-Hadamard
failure of image analysis. The method generally involves
Transform, Local Structure, Medical Image Registration, determining a number of corresponding control points in the
Normalization. images and, from the correspondences, determining a
transformation function that will determine the
1. Introduction correspondence between the remaining points in the images.
Digital image processing is building up the basic This technique may be classified based on four different
machine that could achieve the visual functions of all, that aspects given as follows: (i) the feature selection (extracting
is, it is an enhancement by improving image quality by features from an image) using their similarity measures and
filtering the noise and restoration of images by performing a correspondence basis, (ii) the transformation function, (iii)
compression to save storage area and channel capacity the optimization procedure, and (iv) the model for
during transmission. It is a rapidly growing field with processing by interpolation.
emergent applications in many areas of science and Amongst the numerous algorithms developed for image
engineering. The main principle of registration is to registration so far, methods based on image intensity values
combine the sets of data with the deviations if any or with are particularly excellent as they are simple to automate as
their similarities into a single data. These sets of data are solutions to optimization problems. Pure translations, for
obtained by sampling the same scene or object at different example, can be calculated competently, and universally, as
times or from different perspectives, in different co-ordinate the maxima of the cross correlation function between two
systems. The purpose of registration is to envision a single images [11] [15] [17]. Additional commands such as
data merged with all the details about these sets of data rotations, combined with scaling, shears, give rise to
acquired at different times or perspectives or co-ordinate nonlinear functions which must be resolved using iterative
systems. Such data is very vital in medicine for doctors to nonlinear optimization methods [11].
prepare for surgery. The most familiar and significant In the medical imaging field, image registration is
classes of image analysis algorithm with medical regularly used to combine the complementary and
applications [1,3] are image registration and image synergistic information of images attained from different
106 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 10, 2010
modalities. A widespread problem when registering image toward the effectiveness of the occurrence of the analogous
data is that one does not have direct access to the density structure at the particular position. If these coefficients are
functions of the image intensities. They must be estimated normalized by the dc coefficient of the expansion, i.e., the
from the image data. A variety of image registration local average gray value of the image, then they measure
techniques have been used for successfully registering purely the local structure independent of modality. Walsh
images that are unoccluded. This is generally practiced with basis functions correspond to local structure, in the form of
the use of Parzen windows or normalized frequency positive or negative going horizontal or vertical edge, corner
histograms [12]. of a certain type, etc. In addition, registration schemes based
The work proposed in this paper uses Fast Walsh- on wavelet coefficient matching do not present a general
Hadamard Transform (FWHT) [18, 19] for image mechanism of combining the matching results across
registration. The coefficients obtained are normalized to different scales.
determine a unique number which in turn represents the Suppose if there are two images I1 and I2, in which I1 is
digits in a particular range. The experiments conducted on assumed to represent the reference image whereas I2
clinical images show that proposed algorithm performed represents an image that has to be deformed to match I1.
well than the conventional Walsh Transform(WT) method First, we consider around each pixel, excluding border
in medical image registration. In addition, this paper pixels, a 3X3 neighborhood and compute from it, the nine
provides a comparative analysis of Fast Walsh-Hadamard Walsh coefficients (3X3 WT of a 3X3 image patch). If ‘f’ is
transform and Walsh transform in Medical image the input image, the matrix of coefficients ‘g’ computed for
registration in terms of correlation coefficient and time it using equation (1),
taken for registration. g = (W −1 ) T . fW −1 (1)
The remainder of the paper is ordered as follows.
Matrix contains the coefficients of the expansion of the
Section 2 provides an overview on the related work for
image, in terms of the basis images formed by taking the
image registration. Section 3 explains Walsh transform in
vector outer products of the rows of matrix W and its inverse
image registration. Section 4 describes the proposed
W-1. These basis images are shown in Fig. 1(a). These
approach for image registration using Fast Walsh-Hadamard
coefficients are denoted by a00, a01, a02, a10, a11, a12, a20, a21,
Transform. Section 5 illustrates the experimental results to
a22, and in a matrix form as shown in Fig. 1(b). These
prove the efficiency of the proposed approach in image
coefficients take the value in the range [0, 9]. Moreover
registration in terms of correlation coefficient and time
normalization given by equation (2) makes the method
taken for registration and Section 6 concludes the paper
robust to global levels of change of illumination.
with a discussion.
αij = aij / a00 (2)
2. Related Works However, the information having dense features and
rigid body transformation allows for plenty of redundancy in
This section of paper provides a quick look on the the system and makes it robust to noise and bad matches of
relevant research work in image registration. individual pixels which effectively represent lack of local
An automatic scheme using global optimization information. One may construct a unique number out of
technique for retinal image registration was put forth by eight numbers if one uses these numbers as the digits of the
Matsopoulos et al. in [1]. A robust approach that estimates unique number. The number of levels depends on the
the affine transformation parameters necessary to register number system adopted. If one decide to stick with the
any two digital images misaligned due to rotation, scale, decimal system, and then the normalized coefficients are
shear, and translation was proposed by Wolberg and Zokai quantized so that they take integer values in the range [0, 9].
in [2]. Zhu described an approach by cross-entropy
optimization in [3]. Jan Kybic and Michael Unser together
put forth an approach for fast elastic multidimensional
intensity-based image registration with a parametric model
of the deformation in [4]. Bentoutou et al. in [5] offered an
automatic image registration for applications in remote
sensing. A novel approach that addresses the range image (a). WTs basis images for a 3X3 images
registration problem for views having low overlap and
which may include substantial noise for image registration a0 a0 a0
was described by Silva et al. in [6]. Matungka et al. 0 1 2
proposed an approach that involved Adaptive Polar a1 a1 a1
Transform (APT) for Image registration in [7, 10]. A 0 1 2
feature-based, fully non supervised methodology dedicated a2 a2 a2
to the fast registration of medical images was described by
0 1 2
Khaissidi et al. in [8]. Wei Pan et al. in [9] proposed a (b). Nine coefficients in matrix form
technique for image registration using Fractional Fourier Figure 1. Walsh Transformation
Transform (FFT).
3. Walsh Transform In Figure 1(a) the coefficients along the first row and the
first column are of equal importance, as they measure the
The Walsh, Haar [13], etc are examples of orthogonal presence of a vertical or a horizontal edge, respectively. The
transforms. The coefficients of such an extension point
(IJCNS) International Journal of Computer and Network Security, 107
Vol. 2, No. 10, 2010
remaining four coefficients measure the presence of anumber. This unique number can then be used as feature for
corner. image registration. The implementation of FWHT readily
The following ordering of coefficients are used in images, reduces the time consumption for medical image registration
Ordering IA α01, α 10, α 20, α 02, α 11, α 21, α 12, α 22 when comparing the same with conventional WT technique for
Ordering IB α 10, α 01, α 02, α 20, α 11, α 12, α 21, α 22 image registration.
Ordering IIA α 22, α 21, α 12, α 11, α 02, α 20, α 10, α 01
Ordering IIB α 22, α 12, α 21, α 11, α 20, α 02, α 01, α 10 5. Experimental Result
The performance of the FWHT is better than the WT in terms A series of experiments is performed using medical
of Correlation Coefficient(CC) and Time taken for registration images. The tests are performed using different images of
of mono and multimodal image registration. different sizes. A set of CT and MR medical images which
depict the head of the same patient is considered. The
The drawbacks of WT were found to be as follows: original size of these images is given as pixels. In order to
(i) Errors are present in the final registration stage. remove the background parts and the head outline, the
(ii) Consumes more CPU time for registration as lot of original images are cropped, creating sub-images of
calculations are involved. This is the major drawback when different dimension pixels. The algorithms are evaluated by
compared to FWHT. determining the CC and Time taken for registration.
The CC is a statistical measure of how well trends in
The advantages of FWHT were found to be as follows: the predicted values follow trends in past actual values. It is
(i) More reliable as produces least error results in the final a measure of how well the predicted values from a forecast
registration stage when compared to WT. model “fit” with the real-life-data. As the strength of the
(ii)Consumes very less CPU time for registration as calculations relationship between the predicted values and actual values
are performed using divide and conquer method. This is the increases, so does the CC. Thus higher the CC the better it
major advantage when compared to WT. is.
(iii)The performance of the FWHT is better than the WT in Correlation Coefficient
terms of CC for monomodal and multimodal image registration.
The Computed Tomography (CT) or Magnetic Resonance
(MR) images represent monomodal images and CT-CT, CT-MR,
and MR-MR images represent real multimodal pairs and the
successful registration shows that FWHT produce accurate
results while the measures correlation coefficient and time taken
for registration are considered in a diversity of image sets,
without any tuning in the preprocessing step.
4. Proposed Approach
0.5
0.4
FWHT
(iii) Sagittal 500 x 500 - 72k - jpg MRI & Sagittal 840 x 754 -
0.3
0.2 69k - jpg CT - WT
0.1
0
1 2 3 4 5
Ima ges
Table 2.Represents Time consumption for Image Registration a) Registered Image obtained using WT
using WT & FWHT
S X in Y in Angle in Elapsed Elapsed
No mm mm degrees Time in Time in
seconds seconds
for WT for
FWHT b) Difference in images obtained using WT
1 4 -10 9 138.297 6.829 Figure 8. Images obtained using WT
2 -12 -7 13 135.922 7.328
3 5 -7 5 133.406 6.328 (iv) Sagittal 500 x 500 - 72k - jpg MRI & Sagittal 840 x 754 -
4 -14 -15 2 136.000 6.750
69k - jpg CT - FWHT
5 -8 -7 1 141.125 6.125
160
140
120
Time in Secs
100
WT
80
60
FW HT
a) Registered Image obtained using FWHT
40
20
0
1 2 3 4 5
I ma ge s
b) Difference in images obtained using FWHT b) Difference in images obtained using FWHT
Figure 17. Images obtained using FWHT
110 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 10, 2010
Analysis and Machine Intelligence, vol. 31, no. 3, pp. R. Neelaveni is presently working as a
400-413, 2009. Assistant Professor, Department of EEE,
[10] R. Matungka, Y. F. Zheng, and R. L. Ewing, “Image PSG College of Technology, Coimbatore.
registration using Adaptive Polar Transform,” IEEE She has a Bachelor’s degree in ECE, a
Master’s degree in Applied Electronics and
Transactions on Image Processing, vol. 18, no. 10, pp.
PhD in Biomedical Instrumentation. She has
2340-2354, 2009. 19 years of teaching experience and has
[11] Jr. Dennis M. Healy, and Gustavo K. Rohde, “Fast guided many UG and PG projects. Her
Global Image Registration using Random Projections,” research and teaching interests includes
4th IEEE International Symposium on Biomedical Applied Electronics, Analog VLSI, Computer Networks, and
Imaging: From Nano to Macro, ISBI 2007, pp. 476- Biomedical Engineering. She is a Life member of Indian Society
479, 2007. for Technical Education (ISTE). She has published several
[12] C. Fookes and A. Maeder, “Quadrature-Based Image research papers in national and international Journals and
Registration Method using Mutual Information,” IEEE Conferences.
International Symposium on Biomedical Imaging: Nano
to Macro, vol. 1, pp. 728-731, 2004.
[13] M. Petrou and P. Bosdogianni, Image Processing—The
Fundamentals. New York: Wiley, 1999.
[14] Pere Marti-Puig, “A Family of Fast Walsh Hadamard
Algorithms With Identical Sparse Matrix
Factorization,” IEEE Transactions on Signal Processing
Letters, vol. 13, no. 11, pp. 672-675, 2006.
[15] J. L. Moigne, W. J. Campbell, and R. F. Cromp, “An
automated parallel image registration technique based
on correlation of wavelet features,” IEEE Trans. Geosci.
Remote Sens., vol. 40, no. 8, pp. 1849–1864, Aug.
2002.
[16] J. P. W. Pluim, J. A. Maintz, and M. A. Viergever,
“Image registration by maximization of combined
mutual information and gradient information,” IEEE
Trans. Med. Imag., vol. 19, no. 8, pp. 899–814, Aug.
2000.
[17] Z. Zhang, J. Zhang, M. Liao, and L. Zhang, “Automatic
registration of multi-source imagery based on global
image matching,” Photogramm. Eng. Remote Sens.,
vol. 66, no. 5, pp. 625–629, May 2000.
[18] M. Bossert, E. M. Gabidulin, and P. Lusina, “Space-
time codes based on Hadamard matrices proceedings,”
in Proc. IEEE Int. Symp. Information Theory, Jun. 25–
30, 2000, p. 283.
[19] L. Ping, W. K. Leung, and K. Y. Wu, “Low-rate turbo-
Hadamard codes,” IEEE Trans. Inf. Theory, vol. 49, no.
12, pp. 3213–3224, Dec. 2003.
Author’s Profile
D.Sasikala is presently working as Assistant
Professor, Department of CSE, Bannari
Amman Institute of Technology,
Sathyamangalam. She received B.E.( CSE)
from Coimbatore Institute of Technology,
Coimbatore and M.E. (CSE) from
Manonmaniam Sundaranar University,
Tirunelveli. She is now pursuing Phd in Image
Processing. She has 11.5 years of teaching experience and has
guided several UG and PG projects. She is a life member of ISTE.
Her areas of interests are Image Processing, System Software,
Artificial Intelligence, Compiler Design.