Date post: | 04-Apr-2018 |
Category: |
Documents |
Upload: | akfaditadikaparira |
View: | 213 times |
Download: | 0 times |
of 12
7/29/2019 Zang 2009
1/12
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 47, NO. 12, DECEMBER 2009 4189
A Novel Interest-Point-Matching Algorithm forHigh-Resolution Satellite Images
Zhen Xiong and Yun Zhang, Member, IEEE
AbstractInterest-point matching is a key technique for imageregistration. It is widely used for 3-D shape reconstruction, changedetection, medical image processing, computerized visioning sys-tems, and pattern recognition. Although numerous algorithmshave been developed for different applications, processing localdistortion inherent in images that are captured from differentviewpoints remains problematic. High-resolution satellite imagesare normally acquired at widely spaced intervals and typicallycontain local distortion due to ground relief variation. Interest-point-matching algorithms can be grouped into two broad cate-gories: area based and feature based. Although each type has itsown particular advantages in specific applications, they all face
the common problem of dealing with ambiguity in smooth (low-texture) areas, such as grass, water, highway surfaces, buildingroofs, etc. In this paper, a new algorithm for interest-point match-ing of high-resolution satellite images is proposed. The concep-tual basis of this algorithm is the detection of super points,those points which have the greatest interest strength (i.e., whichrepresent the most prominent features) and the subsequent con-struction of a control network. Sufficient spatial information isthen available to reduce the ambiguity and avoid false matches.We commence this paper with a brief review of current researchon interest-point matching. We then introduce the proposed al-gorithm in detail and describe experiments with three sets ofhigh-resolution satellite images. The experiment results show thatthe proposed algorithm can successfully process local distortionin high-resolution satellite images and can avoid ambiguity in
matching the smooth areas. It is simple, fast, and accurate.
Index TermsControl network, high-resolution satellite image,interest-point matching, super point.
I. INTRODUCTION
INTEREST-POINT matching refers to the process of match-
ing two sets of features and finding correspondences be-
tween them. Matching interest points (sometimes called feature
points or key points) is a key requirement for image regis-
tration. Image registration is widely used in photogramme-
try, remote sensing, computer vision, pattern recognition, and
medical image processing [6], [31]. Unfortunately, there are
still many challenges with interest-point matching. The maininterest-point-matching algorithms currently in use are area
based or feature based. Neither type of algorithm can avoid
Manuscript received September 17, 2008; revised December 29, 2008 andMarch 18, 2009. First published September 1, 2009; current version publishedNovember 25, 2009. This work was supported in part by the Discovery GrantsProgram of the Natural Sciences and Engineering Research Council of Canadaand in part by the Canada Research Chairs Program.
The authors are with the Department of Geodesy and Geomatics Engineer-ing, University of New Brunswick, Fredericton, NB E3B 5A3, Canada (e-mail:[email protected]; [email protected]).
Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TGRS.2009.2023794
the problem of dealing with ambiguity in smooth (low texture)
areas. Feature-based algorithms face the additional problem of
the effect of outliers (points with no correspondences) on the
results [31].
In this paper, we propose a novel interest-point-matching
algorithm, in which super points, those points which have
the greatest interest strength (i.e., which represent the most
prominent features), are extracted first. A control network is
then constructed using these super points. Next, each remaining
interest point is assigned a unique position with regard to
the closest control network point. Finally, an iterative closestpoint algorithm is applied to search for correspondences based
on the position that has been assigned to each interest point.
After each iteration, the new correspondences are added to the
control network as new leaves. The control network therefore
gradually becomes larger and denser. The iterations continue
until no more correspondences are found. Because every point
is located in a unique position relative to the control network,
this method avoids the problem of how to deal with local
minimums.
The first section of this paper contains a brief review of
previous relevant work by others. In the second section, the new
algorithm is introduced in detail. Next, we present some ex-
periments using high-resolution satellite images. Finally, someconcluding remarks are provided.
II. LITERATURE REVIEW
Interest-point matching is problematic and remains the
subject of much research within the communities of photogram-
metry, remote sensing, computer vision systems, pattern recog-
nition, and medical image processing. Interest-point-matching
algorithms can be grouped into two main categories: area-based
algorithms and feature-based algorithms. In remote sensing,
area-based algorithms are normally suitable for open terrain
areas, but the feature-based approaches can provide more accu-rate results in urban areas. No single technique performs well
in both circumstances [16]. Both algorithms have their own
unique strengths and weaknesses.
Our review of previous research in interest-point matching
revealed that about 90% of the papers are from the fields
of computer vision, pattern recognition, and medical image
processing. Such applications have a number of common char-
acteristics: 1) The images they deal with have no baseline or a
short baseline; 2) the images are normally processed in a short
time; and 3) feature-based algorithms are widely used.
Because of the large number of feature-based algorithms
used in interest-point matching, there are many classification
0196-2892/$26.00 2009 IEEE
7/29/2019 Zang 2009
2/12
4190 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 47, NO. 12, DECEMBER 2009
TABLE ILIMITATIONS OF ARE A-BASED AND FEATURE-BASED ALGORITHMS
Fig. 1. (Above) Original image and (below) corresponding interest strength.The brightness is directly proportional to the interest strength.
methods for describing these algorithms. Normally, feature-
based algorithms can be categorized into rigid and nonrigid(according to the transformation between images), global and
local (according to the image distortions), or corrected and un-
corrected (according to the image variations). In addition, most
of the feature-based algorithms search for correspondences
and also address the refinement of a transformation function.
Therefore, feature-based algorithms can also be grouped into
three additional categories [8]. They either solve the correspon-
dence only, solve the transformation only, or solve both the
correspondence and the transformation.
Although numerous feature-based algorithms have been de-
veloped, there is no general algorithm which is suitable for a
variety of different applications. Every method must take into
account the specific geometric image deformation [31]. Thefirst category of algorithms processes the global distortions. The
Fig. 2. Extracted super points (above: 99 super points in super-point set 1 and111 super points in super-point set 2) and interest points (below: 737 interestpoints in set 1 and 707 interest points in set 2).
iterative closest point algorithm is a classical global algorithm[3], [29]. Because this algorithm requires the assumption that
one surface is a subset of the other, it is only suitable for
global distortion image registration [28]. For medical image
registration and pattern recognition, many rigid global transfor-
mations are used [3], [22], [27]. The B-Spline and Thin Plate
Spline (TPS) deformation model is a common model for global
distortion in medical image registration [5], [18].
The second category of algorithms deals with the local
distortions. For nonrigid local distortions, more complicated
transformations are developed. TPS was proposed initially for
global transformations, but it was improved for smooth local
distortions for medical image registration [1], [8], [13]. Another
common local distortion model is the elastic deformation model[1], [23].
7/29/2019 Zang 2009
3/12
XIONG AND ZHANG: NOVEL INTEREST-POINT-MATCHING ALGORITHM 4191
Fig. 3. Flowchart of super-point-matching procedure.
Some algorithms do not need a transformation function.
In computer vision systems and pattern recognition, feature
descriptors extracted from an images gray values are usu-
ally used [2], [17], [20], [26], [30]. Scale invariant feature
transform is one of the best descriptors for interest-point
matching [21]. In graph-matching algorithms, topological re-
lationship is the key feature and is widely used in pattern
recognition [7], [10][12], [25]. Another idea is to consider
interest-point matching as a classification problem. The fea-
tures from the reference image are used to train the classifier
[4], [19].
Although many of the feature-based algorithms describedearlier are useful in solving problems for specific applications,
they have four common drawbacks: 1) The features cannot be
exactly matched, because of the variations of features between
different images; 2) outliers are difficult to reject [8]; 3) for
local image distortion, high-dimensional nonrigid transforma-
tions are required, and a large number of correspondences are
needed for the refinement of mapping functions [6], but too
many features will make the feature matching more difficult;
and 4) the feature description should fulfill several conditions,
the most important ones being invariance (the descriptions
of the corresponding features from the reference and sensed
image have to be the same), uniqueness (two different features
should have different descriptions), stability (the description ofa feature, which is slightly deformed in an unknown manner,
Fig. 4. Control network constructed with super points. P and P are roots,and the others are leaves. A and A are start points. Sixteen tie points(correspondences) are obtained after super-point matching. denotes anoutlier.
Fig. 5. Relative position and angle assignment and correspondence search.After the root and start points are determined, every point (e.g., C) can beassigned a relative position (R) and angle () (Image 1). The closest candidatein the searching area is the correspondence (Image 2).
should be close to the description of the original feature), andindependence (if the feature description is a vector, its elements
should be functionally independent). Usually, these conditions
cannot be satisfied simultaneously, and it is necessary to find an
appropriate tradeoff [31].
Images in photogrammetry and remote sensing contain local
distortions caused by ground relief variations and differing
imaging viewpoints. Because of their stability and reliability,
area-based methods are usually used in remote sensing for
interest-point matching. Photogrammetric scientists are always
attempting to improve the stability and reliability of interest-
point-matching techniques. Hierarchical matching and relax-
ation algorithms are typical examples of such attempts. Atthe same time, great efforts are also being made to reduce
the search area and increase the matching speed. The use of
epipolar geometry is one of the most important achievements
of such work. Despite the progress that has been made, the
area-based methods still have many drawbacks. The main
limitations can be summarized as follows: 1) The rectangular
image window is only suitable for image distortion caused
by translation (in theory); 2) these methods cannot process
smooth areas (areas without prominent texture); and 3) the
methods are sensitive to image intensity changes which are
caused by noise, varying illumination, and the use of different
sensors [31].
In summary, Table I shows the characteristics and limitationsof area-based and feature-based algorithms.
7/29/2019 Zang 2009
4/12
4192 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 47, NO. 12, DECEMBER 2009
Fig. 6. Result of super-point matchingcontrol networks (41 correspondences).
III. METHODOLOGY
The proposed algorithm first detects and extracts superpoints, which have the greatest interest strength (i.e., those
points which represent the most prominent features). A control
network can then be constructed based on these super points.
This control network, like a sketch, can then control the entire
image, and ambiguities in the smooth areas can be avoided.
Next, every point in the image is assigned a unique position
and angle relative to the closest super point in the control
network. Finally, for interest-point matching, those points with
the smallest position and angle differences are the correspon-
dences. The correspondences are then added to the control
network to construct a bigger and stronger control network. The
process is continued until no more correspondences are found.The algorithm proposed in this paper includes three parts:
1) super-point detection; 2) super-point matching; and
3) interest-point matching.
A. Super-Point Detection
The Harris detector is a well-known interest-point detection
algorithm and was used in this paper to detect and extract the su-
per points and interest points. The Harris algorithm determines
whether a point is a corner based on the Harris matrix A at the
point P(x, y)
A = I2
x IxIy
IxIy
I2y
(1)
where Ix is the first derivative in x-direction and Iy is the first
derivative in y-direction; the angle brackets denote averaging
[summation over the image patch around the point P(x, y)].The interest strength is determined based on the magnitudes
of the eigenvalues (1 and 2) of A. Because the exact
computation of the eigenvalues is computationally expensive,
the following function Mc was suggested by Harris and
Stephens [15] as the interest strength:
Mc = det(A) trace2(A). (2)
The value of has to be determined empirically, and in theliterature, values in the range of 0.040.06 have been reported
Fig. 7. Flowchart of interest-point-matching procedure.
as feasible [24]. IfMc > 0, it is a corner; otherwise, it is not acorner. Obviously, the corner should be the point with the local
maximum value of Mc. By calculating the interest strength
Mc over the whole image, an image which shows the interest
strength can be obtained (Fig. 1). Two thresholds T A and T B
can be set, with TA > TB for the interest-point detection
and super-point detection. The point with an interest strength
greater than the threshold T B and also representing the local
maximum can be extracted as an interest point. If the interest
strength of such a point is greater than the threshold T A and
its interest strength is a local maximum, then a super point isdetected (Fig. 2). After super-point matching, a super-point
7/29/2019 Zang 2009
5/12
XIONG AND ZHANG: NOVEL INTEREST-POINT-MATCHING ALGORITHM 4193
Fig. 8. Subcontrol network. Interest points 17, 18, 19, and 20 are groupedwith their closest control network point 10. A subcontrol network is constructedwith interest points 17, 18, 19, 20, and 10 and node 10s father node P. Thefather node P will be the starting point in the subcontrol network. Interest-point matching is performed between two subcontrol networks whose roots arecorrespondences (tie points).
Fig. 9. Image distance difference caused by ground relief variation.
Fig. 10. Image distance difference. The distance difference changes with theincidence angle and ground slope (assuming that the forward incidence angle1 equals the backward incidence angle 1).
set can be obtained for each image (see Fig. 2). Like most
other interest-point-matching processes, super-point matching
is an exhaustive search process, so the number of super points
should be limited to an acceptable range.
B. Super-Point Matching
The goal of the super-point matching is to find a root from
each super-point set and identify the first group of correspon-
dences (tie points). The super-point matching consists of three
steps (Fig. 3): 1) control network construction; 2) assignment of
relative positions and angles; and 3) correspondence searching.A more detailed description of each step follows.
Fig. 11. Test data 1 from stereo pair of IKONOS images of Penang (fromCRISP).
Fig. 12. Test images from the Penang stereo pair: (a) and (a) are a pair (400by 400 pixels) without rotation, while (b) and (b) are a pair (400 by 400 pixels)with (b) rotated 45.
In Step 1), a super point from each super-point set is ran-
domly selected as a root, and a control network is constructed.
One control network is constructed for each super-point set.
In this control network, there is only one root, whereas other
points are leaves (Fig. 4).
Step 2) includes the following three stages.
1) A leaf from control network 1 is selected randomly as the
starting point. The distance between the starting point and
the root is denoted as S.
2) The corresponding starting point in control network 2 is
determined according to the distance between the root
and the leaf. The leaf point of control network 2 with
the closest distance to S is selected as the correspondingstarting point in control network 2.
7/29/2019 Zang 2009
6/12
4194 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 47, NO. 12, DECEMBER 2009
Fig. 13. Results of interest-point matching corresponding to the image pair in Fig. 12(a) and (a) (410 correspondences) and Fig. 12(b) and (b) (264correspondences).
3) After the two starting points for both control networks
have been determined, the relative positions (distance
between root and leaf) and angles (clockwise from the
starting point) are assigned to every point in both control
networks (Fig. 5).
Correspondence searching commences in Step 3). After each
point in both control networks has been assigned a relative
position and angle, a corresponding point in control network 2may be found for every leaf point in control network 1 ac-
cording to their positions and angles based on the following
function:
correspondence = M in
m1
i=1
n1j=1
abs(P i Pj)
M in
m1
i=1
n1
j=1
abs(Pi Pj)
(3)
7/29/2019 Zang 2009
7/12
XIONG AND ZHANG: NOVEL INTEREST-POINT-MATCHING ALGORITHM 4195
where m and n denote the number of leaves in control network
1 and control network 2, respectively, Pi and Pj are relative
distances between root and leaf in the two control networks,
and Pi and Pj are relative angles between starting point and
leaf in the two control networks.
The closest points with the smallest position differences
and smallest angle differences, where both differences are lessthan their corresponding thresholds, will be selected as tie
points (correspondences). Otherwise, if a point does not have
a correspondence, it is an outlier (Fig. 4). The outlier will be
processed as an interest point in the next iteration.
Every super point can be either the root or the starting point.
After super-point matching, a number of correspondences are
obtained. When the maximum possible number of correspon-
dences is obtained, the corresponding root and starting points
will be the final root and starting points of the super-point
control network.
Only image shift and image rotation are considered when
interest points are matched by determining the root and the
starting point. This is acceptable because for high-resolution
satellite images with narrow fields of view, affine transforma-
tions can accurately simulate the geometric distortion between
two images [14].
The process of super-point matching is an iterative and
exhaustive search process. Every point can be either a root or
a starting point. For example (Fig. 4), there are 20 super points
in super-point set 1 and 21 super points in super-point set 2.
Therefore, there are C120
C121
combinations for root selection,
C119
C120
combinations for starting point selection, and C118
C119
combinations for the correspondence search. Thus, there will
be C120
C121
C119
C120
C118
C119
= 54583200 combinations in total.
Therefore, in order to avoid combination explosion and reducethe matching time, the number of super points should be limited
to an acceptable range.
After super-point matching, a control network, which con-
sists of all the extracted correspondences, is obtained (Fig. 6).
C. Interest-Point Matching
After the super-point matching, two control networks corre-
sponding to the two interest-point sets are obtained (Fig. 6).
Under the control of the super-point network, interest-point
matching becomes simple. Fig. 7 shows a flowchart of the
interest-point-matching process, which includes four steps.First, through a process of K-means clustering, every interest
point can be grouped with the closest node of the control net-
work. For example (Fig. 8), the interest points in the circle are
grouped with the closest control point 10. Then, taking node
10 as the root, together with all the interest points grouped
with it (17, 18, 19, 20), a subcontrol network is constructed. In
this subcontrol network, the father node P of node 10 is the
starting point. Next, every point in this subcontrol network is
assigned a position and angle with respect to node 10 and the
starting point P. In this way, every interest point is assigned
a relative position and angle with respect to its closest control
network point. Finally, interest-point matching is performed
between the two subcontrol networks whose root nodes arecorrespondences.
Fig. 14. Test area 2 from stereo pair of IKONOS images in Hobart (from theUniversity of Melbourne).
Fig. 15. Test images from the Hobart stereo pair: (c) and (c) are a test imagepair (400 by 400 pixels) without rotation, and (d) and (d) are a test image pair(400 by 400 pixels) with (d) rotated 315.
Correspondences are defined as those interest points with the
minimum difference in position and angle. The new correspon-
dences are added to the control network to construct a biggernetwork. This is an iterative process that continues until no
new correspondence is added to the control network. The final
control network is the result of interest-point matching.
D. Threshold Selection
In the process of interest-point matching, it is crucial to set
a suitable threshold for the position and angle differences. In
remote sensing and photogrammetry, the images always contain
complicated local distortions because of the long baselines
(long distance between images) and ground relief variations.
In such a situation, the effective ground distance for different
sensors will vary with changes in ground relief, incidence angle,and sensor position (Fig. 9).
7/29/2019 Zang 2009
8/12
4196 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 47, NO. 12, DECEMBER 2009
Fig. 16. Results of interest-point matching corresponding to the image pair in Fig. 15(c) and (c) (641 correspondences) and Fig. 15(d) and (d) (561correspondences).
For example, a distance S on the ground with a slope is
acquired by two sensors S1 and S2 with incidence angles 1and 2, respectively, (Fig. 9). In this case, the effective distance
for sensor S1 and the effective distance for sensor S2 can be
calculated as follows:
S1e = s cos(1 ) (4)
S2e = s cos(2 + ) (5)
where S1e and S2e are effective distances for sensor S1 and
sensor S2, respectively. 1 and 2 are the incidence angles of
sensor S1 and sensor S2, respectively, is the slope of the
ground, and s is the ground distance.
Therefore, the difference between two effective distances
caused by ground relief variation and incidence angle in such
a case can be defined as follows:
ds = s [cos(1 ) cos(2 + )] (6)
7/29/2019 Zang 2009
9/12
XIONG AND ZHANG: NOVEL INTEREST-POINT-MATCHING ALGORITHM 4197
Fig. 17. Test data 3 from stereo pair of IKONOS images in Penang (fromCRISP).
where ds is the difference between two effective dis-
tances caused by the ground relief variation and incidence
angle.
Obviously, the difference between two effective distances
can vary with ground slope and incidence angle. Fig. 10 shows
the situation.
In fact, satellite elevation and pixel size can also affect
the distance of two effective distances. However, the satellite
elevation affects the effective distance in the form of incidence
angle. As the same, the pixel size changes with the incidence
angle and slope. Therefore, the effective distance could be
affected mainly by incidence angle and slope. That is why we
set a threshold for the correspondence search. That is also a
tolerance for the difference of effective distance.
The difference between two effective distances is propor-
tional to the ground slope and the incidence angle. For animage pair, the incidence angles are constants, so the ground
slope is the only variable. In an image, the slope varies
with the ground relief variation. Therefore, the only way to
limit the distance difference between two images is to limit
the ground distance. A small ground distance will lead to
a small distance difference and vice versa. That is why in
the proposed interest-point-matching algorithm, all interest
points should be grouped with their closest control network
points.
It is important to determine the best way to select the thresh-
old for the distance difference and angle difference. Obviously,
a large threshold will increase the number of false matches, so
in order to reduce false matches, the threshold should be set
as small as possible. However, when the distance difference
between two images is large, a small threshold may mean that
correspondences are overlooked and more iterations may be
needed to find matches. Another concern may be that a small
threshold may lead to false matches and exclude the correct
correspondence. This is possible, but because the interest point
is a local maximum, there is only a small probability that, in
the small search area, there is another local maximum and the
correct one is farther away from the search area. The threshold
can therefore be set by considering the radius of the local
maximum. For example, if the local maximum is contained in
a 5 by 5 pixel window, a threshold of 2 pixels or less can beconsidered as a safe threshold.
Fig. 18. Test area 3 in mountainous area (2000 by 2000 pixels).
IV. EXPERIMENTS
Three sets of high-resolution satellite images were used for
our experiments.
A. Test Data 1
A stereo pair of level 1A IKONOS images acquired on
June 25, 2004 over Penang, Malaysia, was used for this experi-
ment (Fig. 11). The incidence angles are 30 and 3.5.
A rectangular area (400 by 400 pixels) was selected as the
test area. Fig. 12 shows two pairs of images. The pair (a) and
(a) was taken directly from the original images. A second pair
(b) and (b) is composed of (b) which was taken from the
original left image and (b) which was taken from the right
image which has been rotated 45. In this test area, there isa large area of grass which was used to test the algorithms
7/29/2019 Zang 2009
10/12
4198 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 47, NO. 12, DECEMBER 2009
Fig. 19. Result of interest-point matching corresponding to the image pair (e) and (e). There are 5674 correspondences in total.
TABLE IISTATISTICS OF INTEREST-POINT MATCHING FOR FIV E CASES
capability of reducing ambiguity and avoiding false matching
in a smooth area.
Fig. 13 shows the results of interest-point matching cor-
responding to the image pairs in Fig. 12(a) and (a) and
Fig. 12(b) and (b), respectively.
B. Test Data 2
A stereo pair of IKONOS images, which was acquired in
February 2003 in Hobart, Australia, was used for this exper-iment (Fig. 14). The incidence angles are forward 75 and
backward 69 [32].
A rectangular area (400 by 400 pixels) was selected as the
test area. Fig. 15 shows two pairs of images: Pair (c) and (c) is
an image pair taken directly from the original images, while pair
(d) and (d) is another image pair where (d) was taken directly
from the original left image and (d) was taken from the right
image which has been rotated 315. This is an urban area with a
large area of grass where the algorithms capability of reducing
ambiguity and avoiding false matching in smooth areas could
be tested.
Fig. 16 shows the results of interest-point matching corre-
sponding to the image pairs in Fig. 15(c) and (c) and Fig. 15(d)and (d), respectively.
C. Test Data 3
Test area 3 is also from the stereo image pair in Penang.
Because the aforementioned two test areas are relatively flat
and somewhat small, a larger test area from a mountainous area
was selected as test area 3 (Fig. 17) in order to test the algorithm
under a different set of conditions.
A rectangular area (2000 by 2000 pixels) was selected as test
area 3. Fig. 18 shows image pair (e) and (e), taken directly
from the original images. This is a mixed area of mountain and
urban land cover. In this test area, there is a large area of forestwhich was used to test the algorithms capability of reducing
ambiguity and avoiding false matching in a smooth area. The
mountainous area was used to test the algorithms capability of
processing large distortions.
Fig. 19 shows the results of interest-point matching corre-
sponding to the image pair in Fig. 18(e) and (e).
V. EXPERIMENTAL RESULTS
All the experiments showed satisfactory results. Table II
illustrates the statistics of interest-point matching for five image
pairs. We carefully checked every correspondence in each of thetest areas in test data 1 and test data 2 and did not find any false
7/29/2019 Zang 2009
11/12
XIONG AND ZHANG: NOVEL INTEREST-POINT-MATCHING ALGORITHM 4199
matches. Even in the smooth areas (e.g., a large area of grass),
this algorithm avoided false matches efficiently. In addition,
because each interest point is assigned a unique position and
angle with regard to its closest control point, it is only necessary
to search for correspondences within the corresponding
subcontrol network, so the process of interest-point matching
is very fast. By using IBM (processor 1.70 GHz, 1.69 GHz,768 MB of RAM), each experiment took only a few seconds.
The success of this algorithm completely depends on the con-
trol network. On the one hand, the control network incorporates
the spatial information and easily overcomes the problem of
ambiguity in the homogeneous area; on the other hand, once the
first group of correspondences from the super-point matching
is wrong, then all the other correspondences extracted based
on this control network later on will be false. This may be the
main concern for this algorithm. However, for every different
image, the control network of super points is almost always
unique, except that there is not any prominent texture in the
image and the whole image is homogeneous or filled with man-
made texture. Therefore, this algorithm does not work in the
complete homogeneous area, such as the area covered by snow,
water, or sand.
VI. CONCLUSION
We have presented and successfully tested a novel algorithm
for interest-point matching of high-resolution satellite images.
This algorithm has the following characteristics.
1) It can avoid local minimum problems and can process
areas without prominent details because the proposed
algorithm uses spatial information by first constructing a
super-point control network.2) It can remove outliers easily because every interest point
is assigned a unique position and angle with regard to its
closest control point.
3) Because of the super-point control network, the algorithm
does not require an exhaustive search during the interest-
point matching, so it is a simple, fast, and accurate
algorithm.
Of course, like other algorithms, the proposed algorithm can-
not solve every interest-point-matching problem. Because only
shift and rotation are considered in the algorithm, we think
that this algorithm can only be used for high-resolution satellite
images that were captured with a narrow field-of-view cameraor other images that were captured with a short baseline.
ACKNOWLEDGMENT
The funding for this project was provided by the Discov-
ery Grants Program of the Natural Sciences and Engineering
Research Council of Canada (NSERC) and the Canada Re-
search Chairs Program. The authors would like to acknowledge
that Professor Clive Fraser, Department of Geomatics Engi-
neering, University of Melbourne, Australia, provided satellite
images for the tests. Thanks also go to Mr. David C. Whyte,
Department of Environment, NB, Canada, who reviewed the
manuscript of this paper. The authors appreciate the reviewerscomments and suggestions.
REFERENCES
[1] M. Auer, P. Regitnig, and G. A. Holzapfel, An automatic nonrigid reg-istration for stained histological sections, IEEE Trans. Image Process.,vol. 14, no. 4, pp. 475486, Apr. 2005.
[2] S. Belongie, J. Malik, and J. Puzicha, Shape matching and object recog-nition using shape contexts, IEEE Trans. Pattern Anal. Mach. Intell.,vol. 24, no. 4, pp. 509522, Apr. 2002.
[3] P. J. Besl and N. D. McKay, A method for registration of 3-D shapes,IEEE Trans. Pattern Anal. Mach. Intell., vol. 14, no. 2, pp. 239256,Feb. 1992.
[4] A. Boffy, Y. Tsin, and Y. Genc, Real-time feature matchingusing adaptive and spatially distributed classification trees, pre-sented at Brit. Mach. Vis. Conf., Jul. 2006. [Online]. Available:http://www.macs.hw.ac.uk/bmvc2006/papers/397.pdf
[5] F. L. Bookstein, Principal warps: Thin-plate splines and the decomposi-tion of deformations, IEEE Trans. Pattern Anal. Mach. Intell., vol. 11,no. 6, pp. 567585, Jun. 1989.
[6] L. G. Brown, A survey of image registration techniques, ACM Comput.Surv., vol. 24, no. 4, pp. 325376, Dec. 1992.
[7] T. S. Caetano, T. Caelli, and D. A. C. Barone, A comparison of junctiontree and relaxation algorithms for point matching using different distancemetrics, Dept. Comput. Sci., Univ. Alberta, Edmonton, AB, Canada,Tech. Rep. TR 04-04, Feb. 2004.
[8] H. Chui and A. Rangarajan, A new point matching algorithm for non-
rigid registration, Comput. Vis. Image Underst., vol. 89, no. 2/3, pp. 114141, Feb. 2003.
[9] A. Croitoru and V. Tao, An alternative approach to the point corre-spondence problem, presented at ASPRS Annu. Conf., Anchorage, AK,May 2003.
[10] A. D. J. Cross and E. R. Hancock, Graph matching with a dual-stepEM algorithm, IEEE Trans. Pattern Anal. Mach. Intell., vol. 20, no. 11,pp. 12361253, Nov. 1998.
[11] M. F. Demirci, A. Shokoufandeh, S. Dickinson, Y. Keselman, andL. Bretzner, Many-to-many feature matching using spherical coding ofdirected graphs, in Proc. ECCV, T. Pajdla and J. Matas, Eds., 2004,vol. 3021, pp. 322335.
[12] S. Gold and A. Rangarajan, A graduated assignment algorithm forgraph matching, IEEE Trans. Pattern Anal. Mach. Intell., vol. 18, no. 4,pp. 377388, Apr. 1996.
[13] S. Gold, A. Rangarajan, C. P. Lu, S. Pappu, and E. Mjolsness, Newalgorithms for 2D and 3D point matching: Pose estimation and correspon-dence, Pattern Recognit., vol. 31, no. 8, pp. 10191031, Aug. 1998.
[14] A. Habib and R. Ai-Ruzouq, Semi-automatic registration of multi-sourcesatellite imagery with varying geometric resolutions, Photogramm. Eng.
Remote Sens., vol. 71, no. 3, pp. 325332, Mar. 2005.[15] C. Harris and M. Stephens, A combined corner and edge detector, in
Proc. Alvey Vis. Conf., 1988, pp. 147151.[16] Y. C. Hsieh, D. M. McKeown, and F. P. Perlant, Performance evaluation
of scene registration and stereo matching for cartographic feature extrac-tion, IEEE Trans. Pattern Anal. Mach. Intell., vol. 14, no. 2, pp. 214238,Feb. 1992.
[17] A. Kaplan, E. Rivlin, and I. Shimshoni, Robust feature matchingacross widely separated color images, in Proc. 17th ICPR, 2004,pp. 136139.
[18] J. Kybic and M. Unser, Fast parametric elastic image registration, IEEETrans. Image Process., vol. 12, no. 11, pp. 14271442, Nov. 2003.
[19] V. Lepetit, J. Pilet, and P. Fua, Point matching as a classifi-
cation problem for fast and robust object pose estimation, Proc.Comput. Vis. Pattern Recog., 2004, pp. II-244II-250. [Online].Available: http://cvlab.epfl.ch/~vlepetit/papers/lepetit_cvpr04.pdf
[20] V. Lepetit, P. Lagger, and P. Fua, Randomized trees for real-timekeypoint recognition, in Proc. Comput. Vis. Pattern Recog., 2005, vol. 2,pp. 775781.
[21] D. G. Lowe, Distinctive image features from scale-invariant keypoints,Int. J. Comput. Vis., vol. 60, no. 2, pp. 91110, Nov. 2004.
[22] D. M. Mount, N. S. Netanyahu, and J. Le Moigne, Efficient algo-rithms for robust feature matching, in Proc. CESDIS Image RegistrationWorkshop, 1997, pp. 247256.
[23] J. Rexilius, S. K. Warfield, C. R. G. Guttmann, X. Wei, R. Benson,L. Wolfson, M. Shenton, H. Handels, and R. Kikinis,A novel nonrigidregistration algorithm and applications, in Proc. MICCAI, W. Niessenand M. Viergever, Eds., 2001, vol. 2208, pp. 923931.
[24] C. Schmid, R. Mohr, and C. Bauckhage, Evaluation of interest point
detectors, Int. J. Comput. Vis., vol. 37, no. 2, pp. 151172, Jun. 2000.[25] A. Shokoufandeh, Y. Keselman, F. Demirci, D. Macrini, and S. Dickinson,Many-to-many feature matching in object recognition, in Cognitive
7/29/2019 Zang 2009
12/12
4200 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 47, NO. 12, DECEMBER 2009
Vision Systems, vol. 3948, H. I. Christensen and H. H. Nagel, Eds.Berlin, Germany: Springer-Verlag, 2006, pp. 107125.
[26] K. Terasawa, T. Nagasaki, and T. Kawashima, Robust matching methodfor scale and rotation invariant local descriptors and its application toimage indexing, Proc. AIRS, G. G. Lee et al., Eds., 2005, vol. 3689,pp. 601615.
[27] Z. Tu, S. Zheng, and A. Yuille, Shape matching and registration by data-driven EM, Comput. Vis. Image Underst., vol. 109, no. 3, pp. 290304,
Mar. 2008.[28] J. Williams and M. Bennamoun, Simultaneous registration of multiple
corresponding point sets, Comput. Vis. Image Underst., vol. 81, no. 1,pp. 117142, Jan. 2001.
[29] G. Yang, C. V. Steward, M. Sofka, and C. L. Tsai, Registration ofchallenging image pairs: Initialization, estimation, and decision, IEEETrans. Pattern Anal. Mach. Intell., vol. 29, no. 11, pp. 19731989,Nov. 2007.
[30] W. Zhao, Y. Jiang, and C. W. Ngo, Keyframe retrieval by keypoints: Canpoint-to-point matching help?, in Proc. CIVR, H. Sundaram et al., Eds.,2006, vol. 4071, pp. 7281.
[31] B. Zitova and J. Flusser, Image registration methods: A survey, ImageVis. Comput., vol. 21, pp. 9771000, 2003.
[32] C. S. Fraser and H. B. Hanley, Bias-compensated RPCs for sensor ori-entation of high-resolution satellite imagery, Photogramm. Eng. RemoteSens., vol. 71, no. 8, pp. 909915, Aug. 2005.
Zhen Xiong received the A.B. degree in pho-togrammetry and the M.Sc. degree in geomaticsfrom the Wuhan University of Surveying and Map-ping, Wuhan, China, in 1986 and 1993, respectively,the Ph.D. degree in hyperspectral remote sensingfrom the Institute of Remote Sensing Applications,Chinese Academy of Sciences, Beijing, China, in2000, and the second Ph.D. degree in remote sensingfrom the University of New Brunswick, Fredericton,NB, Canada, in 2009.
From 1986 to 1990, he was a Land Surveyor withthe Institute of Surveying and Mapping of Hubei Province, where he completedmore than 20 engineering projects. From 1993 to 1997, he was a Lecturer ofgeomatics with Beijing Jiaotong University, Beijing, China, where he taughtsurveying, railway engineering surveying, construction engineering surveying,
and other geomatics courses. From 2000 to 2005, he was a Research Scientistwith the Centre for Remote Imaging, Sensing and Processing, National Uni-versity of Singapore, Singapore, where he mainly focused on the research ofsatellite photogrammetry and image geometric processing. He is currently aResearch Fellow with the Department of Geodesy and Geomatics Engineering,University of New Brunswick. His research interests include camera sensormodel generation and refinement, feature extraction and feature matching,digital elevation model generation, 3-D building reconstruction, and treescanopy reconstruction from LiDAR data. He is the coauthor of more than 30scientific publications and is a referee of several international journals.
Dr. Xiong is a member of the American Society of Photogrammetryand Remote Sensing (ASPRS). He is a first place co-recipient of 2009John I. Davidson Presidents Award of ASPRS.
Yun Zhang (M02) received the B.Sc. degree in landinformation and mapping from Wuhan University,Wuhan, China, in 1982, the M.S. degree in geogra-phy and remote sensing from the East China NormalUniversity, Shanghai, China, in 1989, and the Ph.D.degree in remote sensing from the Free University ofBerlin, Berlin, Germany, in 1997.
He is currently the Canada Research Chair in
Advanced Geomatics Image Processing with theDepartment of Geodesy and Geomatics Engineer-ing, University of New Brunswick, Fredericton, NB,
Canada. Several new technologies currently developed by Dr. Zhang and hisresearch group are being used by leading organizations globally, includingthe NASA, NOAA, U.S. Geological Survey, U.S. Department of Agriculture,Natural Resources Canada, Department of National Defence (Canada), PCIGeomatics, DigitalGlobe, Google Earth, and many leading universities. He isthe author of more than 30 peer-reviewed journal papers and more than 100refereed conference proceeding papers. With international research experience,he is the holder of two patents and four patent-pending technologies (three withhis student), and the developer of six commercially licensed technologies (onewith his student).
Dr. Zhang was the recipient of the 2005 American Society of Photogram-metry and Remote Sensing (ASPRS) Talbert Abrams Grand Award, and a 1stPlace Co-Recipient of the 2009 ASPRS John I. Davidson Presidents Awardfor Practical Papers. In 2006, his research work was selected as one of nine
successful Canadian technologies for Technology Transfer Works: 100 Casesfrom Research to Realization by the Association of University TechnologyManagers.