Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

Singular Value Decomposition Based Image Matching

Feng Zhao Wen Gao


Institute of Computing Technology, Chinese Academy of Sciences,
Beijing 100080, China
Graduate School of the Chinese Academy of Sciences,
Beijing 100039, China
{fzhao, wgao}@jdl.ac.cn

Abstract In the recent literature some effective strategies for


interest point correspondence between two images
This paper presents a simple and effective method have emerged. A robust technique for matching two
for matching two uncalibrated images. Corner points uncalibrated images has been proposed by Zhang et al.
are firstly extracted as interest points in the two [1], which finds initial matches using correlation and
images. Each interest point is assigned one dominant relaxation methods followed by the LMedS technique
orientation. The initial set of point matches is then to discard false matches. Pilu [2] used a direct method
obtained by singular value decomposition of a cor- based on singular value decomposition for feature cor-
respondence strength matrix. A new expression of this respondence, which was originally proposed by Scott
matrix is introduced to handle more complicated and Longuet-Higgins [7]. The method first sets up a
imaging conditions. Each element of this matrix is the correlation-weighted proximity matrix and then per-
similarity measure between two interest points. The forms singular value decomposition calculation on the
new similarity measure is invariant to image rotation matrix to get initial matches.
by taking into account the dominant orientation of the The initial set of interest point matches usually con-
two interest points. The epipolar geometry constraint is tains some false matches due to the bad locations of
finally imposed to reject the false matches. Experi- corners and the improper matches in the establishment
mental results on real images show this approach to be of correspondence. Constraints like epipolar geometry
effective for general image matching. or disparity gradient limit can be used to reject the
false matches [1, 8, 9].
1. Introduction This paper presents a simple and effective method
for image matching. Interest points are firstly extracted
Matching two uncalibrated images is a classic problem in both images with slightly modified Harris corner
in computer vision. The topic has been researched for detector. In order to achieve invariance to image
several decades. One effective strategy is using interest rotation, each interest point is assigned one dominant
point matching technology. The interest point match- orientation. The algorithm for matching interest points
ing technology first extracts points of interest such as is based on singular value decomposition. The initial
corners in the two images and then uses robust set of point matches is directly obtained by singular
matching technology to establish point correspon- value decomposition of a correspondence strength ma-
dences between the two images [1, 2, 3, 4]. trix. A new expression of the correspondence strength
Corners are considered as good candidates for inter- matrix is introduced to handle more complicated imag-
est points in many computer vision applications such ing conditions. This new definition of the matrix is
as motion correspondence, object tracking and stereo based on the normalized cross-correlation and domi-
matching, etc. Using corner points as interest points nant orientation of each interest point. Last, the epi-
has been proved to be effective in image matching [1, polar geometry constraint is imposed to reject the false
2]. Among the most popular corner detectors, Harris matches. Experimental results on real images show the
corner detector [5] is known to be robust against rota- method is effective for matching two uncalibrated
tion and illumination changes. And its results have images under different imaging conditions.
high repeatability under different imaging conditions, The remainder of this paper is organized as follows.
which is of great advantage to image matching task [6]. Section 2 describes extracting interest points with

Proceedings of the Third International Conference on Image and Graphics (ICIG’04)


0-7695-2244-0/04 $20.00 © 2004 IEEE
dominant orientation. Section 3 introduces the match- A threshold tf is given and a point is considered as a
ing approach based on singular value decomposition corner if CF > tf. In our implementation, tf is set to 1%
and presents the new expression of the correspondence of the maximum observed CF.
strength matrix in detail. Section 4 describes rejecting In order to achieve invariance to rotation, each
the false matches by imposing epipolar geometry con- interest point is assigned one dominant orientation. The
straint. Section 5 presents some experimental results on histogram based approach for orientation assignment
real images. Section 6 concludes the paper. [11] is adopted. An orientation histogram with 36 bins
covering the range of 360 degrees is used to accumu-
2. Detecting interest points with dominant late the local gradient orientations within a region
orientation around an interest point. The gradient orientation of
each sample in the region is weighted by its gradient
Corners are highly informative image locations. They magnitude and by a Gaussian window. Then a histo-
are useful features for many computer vision appli- gram smoothing is performed by iterative local aver-
cations. Corners can be automatically detected without aging of every 3 consecutive bins in a cyclical fashion.
any prior knowledge and they are stable features for The orientation corresponding to the largest bin in the
image matching task. Among the most popular corner smoothed histogram is selected to be the dominant
detectors, Harris corner detector [5] is known to be orientation of the interest point.
robust against rotation and illumination changes.
Moreover, the results of Harris detector have high re- 3. Matching based on singular value
peatability under different imaging conditions, which is decomposition
very important for image matching.
Harris detector is based on the auto-correlation ma- The problem of matching interest points in two images
trix, which is built as follows: is fundamental in computer vision. A direct method for
establishing feature correspondences between two im-
2
x2  y2 ª I x IxIy º ages based on singular value decomposition has been
M exp …« 2 »
, (1) proposed in Scott and Longuet-Higgins [7] and further
2V 2 «¬ I x I y I y »¼ developed in Pilu [2]. The basic idea of this method is
obtaining feature correspondences with singular value
where … is the convolution operation. Ix and Iy indicate decomposition of a correspondence strength matrix. In
the x and y directional derivative respectively. The the following, the algorithm will be briefly described.
auto-correlation matrix performs a smoothing opera- Let I1 and I2 be two images. I1 contains m features
tion on the products of the first derivatives by con- I1i (i = 1…m) and I2 contains n features I2j (j = 1…n).
volving with a Gaussian window. Two sufficient large To get one-to-one feature correspondences between the
eigenvalues of M indicate the presence of a corner. two images, a proximity matrix G  Mm,n is set up first:
To avoid the explicit eigenvalues decomposition of M, 2 2
 r ij / 2V
a corner response function is defined as follows: G ij e , (4)
CH det( M )  k trace( M ) 2 , (2) where rij = ||I1i-I2j|| is the Euclidean distance between
the two features if they are regarded as being on the
where k is usually set to 0.04. Then a threshold th can
same plane. The degree of interaction between the two
be used to select corner points. If CH > th then the point
sets of features is controlled by the parameter V .
is identified as a corner.
The next step of the algorithm is to perform the
In the original Harris detector, Ix and Iy are com-
singular value decomposition of G: G=UDVT. Then a
puted by convolution with the mask [-1 0 1] and the
new matrix E is converted from the diagonal matrix D
corner response function requires setting a parameter k.
by replacing every diagonal element of D with 1. Com-
In our implementation, the mask [-2 –1 0 1 2] is used
puting the product: H=UEVT will get the new matrix H.
to compute the first derivatives. And a corner meas-
The matrix H  Mm,n has the same shape as the prox-
ure function without additional parameter [10] is
imity matrix G. Two features I1i and I2j are pairing up if
adopted. The measure function is defined by the
Hij is both the greatest element in its row and the
following expression:
greatest element in its column. With selecting all such
det( M ) elements in H, the feature correspondences between
CF . (3) the two images can be established.
trace(M )

Proceedings of the Third International Conference on Image and Graphics (ICIG’04)


0-7695-2244-0/04 $20.00 © 2004 IEEE
Pilu [2] uses normalized cross-correlation score as By using the new correspondence strength matrix,
local measurement to quantify feature similarity. The our approach can handle more complicated imaging
elements of G in Pilu’s method is defined as follows: conditions, especially image rotation with large angle.
Good results have been obtained in the experiments on
G
C ij  1
e
 r ij 2
/ 2V 2
, (5) real images under different imaging conditions. Some
ij
2 of the results are demonstrated in Section 5.
where Cij is the normalized cross-correlation score be-
tween the two features. 4. Rejecting false matches
In our method, a new expression of the correspon-
dence strength matrix G is introduced in order to get The initial set of interest point matches usually con-
more reliable initial matches under different imaging tains false matches. Constraints like epipolar geometry
conditions. The new definition of G is given as follows: can be used to reject the false matches. The epipolar
geometry constraint can be described as follows.
G ij C ijr (1  C ijr  H ) , (6) Suppose F is the fundamental matrix. Point p (x, y)
in the image can be represented as: ~p [ x , y ,1 ] T .
where C ijr is the score of normalized cross-correlation For a point match (p, q), the epipolar line of point p in
with rotation based on the dominant orientation of two the first image is defined as: l p F ~p . If the match is
interest points. To prevent the division overflow pro- perfect, point q in the second image should lie on the
blem, a small positive number H is added to the epipolar line lp exactly. The distance dq of point q to
denominator (in our experiments it is 0.001). C ijr is the epipolar line lp is calculated by
defined as follows.
Let m1=I1 (xi, yi) be the i-th interest point with q~ T F ~
p ,
dq (9)
dominant orientation T 1 in the first image I1 and (F~
p )1  ( F ~
2
p)2
2

m2=I2 (xj, yj) be the j-th interest point with dominant


orientation T 2 in the second image I2. Let T T 2  T 1 where ( F ~p ) i is the i-th component of vector F ~p . The
be the rotation angle. W1 and W2 are two windows of distance dp of point p to the epipolar line lq is calcu-
size (2w+1) u (2w+1) centered on each point. W1 is lated similarly. Then a threshold te can be used to find
rotated by | T | around m1 (the direction of rotation is the good matches. A point match is identified as a
counterclockwise if T > 0, and otherwise is clockwise). good match if max (dp, dq) d te. In our implementation,
W1 and W2 can be represented as two (2w+1) u (2w+1) the epipolar geometry constraint is imposed based on
arrays of pixel intensities A and B. RANSAC [12]. The method is described as follows.
First, eight matches are randomly chosen from the ini-
Auv I1( xi  u cosT  v sinT , yi  v cosT  u sinT ) , tial set of interest point matches and the fundamental
(7) matrix F is calculated from them. Then we can find all
Buv I 2 ( x j  u , y j  v) , good matches that are consistent with the epipolar geo-
where u, v  [-w, w]. Auv is calculated using bilinear metry constraint according to this fundamental matrix.
interpolation. C ijr is then defined as: The threshold te is 1 in our system. Repeat the above
steps to get the largest set of good matches.
w w

¦ ¦>A uv  A @˜ >Buv  B @
(8) 5. Experimental results
C r u w v w ,
ij
(2w  1)(2w  1)V ( A)V ( B)
Experiments on real images of various content have
where A ( B ) is the average and V ( A ) ( V ( B ) ) is
been performed. Some of the results are reported here.
the standard deviation of all the elements in A ( B ) . These images are under different imaging conditions
The value of w is 5 in our system. such as rotation, translation, scale changes and the
C ijr is the similarity measure between two interest combination of them. All the results here are obtained
points. By taking into account the dominant orientation, with the same parameter setting as mentioned in the
C ijr is invariant to image rotation. C ijr decreases forgoing sections. Image pairs Mars, East_south and
monotonically from 1 to 0 with the increase of differ- East_park are from INRIA (http://lear.inrialpes.fr/peo-
ence between two interest points. Comparing with the ple/Mikolajczyk/Database/index.html) and image pair
foregoing two definitions of G, the new definition is Car is from our lab. These results (Fig 1. – Fig 4.) are
robust against image rotation and it doesn’t require final results after applying RANSAC on the initial
setting additional parameter V . matches.

Proceedings of the Third International Conference on Image and Graphics (ICIG’04)


0-7695-2244-0/04 $20.00 © 2004 IEEE
6. Conclusions [9] S.B. Pollard, J.E.W. Mayhew, and J.P. Frisby. PMF: A
stereo correspondence algorithm using a disparity gra-
dient limit. Perception, Vol. 14, pp. 449-470, 1985.
We have presented a simple and effective method for [10] W. Faustner. A feature based correspondence algorithm
matching two uncalibrated images. The method ex- for image matching. International Archives of Photo-
ploits singular value decomposition, which is one of grammetry and Remote Sensing, Vol. 26, pp. 150-166,
the stablest numerical matrix operations. The kernel of 1986.
this method is establishing interest point correspon- [11] David G. Lowe. Object recognition from local scale-
dences by singular value decomposition of a well- invariant features. International Conference on Com-
designed correspondence strength matrix. A new ex- puter Vision, pp. 1150-1157, 1999.
pression of the matrix is introduced to handle more [12] M. A. Fischler and R. C. Bolles. Random Sample Con-
complicated imaging conditions. This new definition sensus: A Paradigm for Model Fitting with Appli-
cations to Image Analysis and Automated Cartography.
adopts the similarity measure that is based on the Communications of the ACM, Vol. 24, pp. 381-395,
normalized cross-correlation and dominant orientation 1981.
of each interest point. One characteristic of this algo-
rithm is that it does not require a correlation threshold
for a candidate point match to be accepted. This con-
tributes to the simplicity of the method. Experimental
results on real images show the method is effective for
matching two uncalibrated images under different
imaging conditions.

Acknowledgments
This work is supported by National Hi-Tech Develop-
ment Programs of China under grant No. 2003AA Figure 1. Matching result for image pair Mars
142140. with large-angle rotation.

References
[1] Z. Zhang, R. Deriche, O. Faugeras, and Q.-T. Luong. A
Robust Technique for Matching Two Uncalibrated Im-
ages Through the Recovery of the Unknown Epipolar
Geometry. Artificial Intelligence Journal, Vol. 78, pp.
87-119, 1995.
[2] M. Pilu. A direct method for stereo correspondence Figure 2. Matching result for image pair
based on singular value decomposition. IEEE CVPR, pp. East_south with rotation and scale changes.
261-266, 1997.
[3] A. Baumberg. Reliable feature matching across widely
separated views. IEEE CVPR, pp. 774-781, 2000.
[4] D. Tell and S. Carlsson. Combining Appearance and
Topology for Wide Baseline Matching. 7th European
Conference on Computer Vision (ECCV), pp. 68-81,
2002.
[5] C. Harris and M. Stephens. A combined corner and
edge detector. Fourth Alvey Vision Conference, pp.
148-151, 1988. Figure 3. Matching result for image pair
[6] C. Schmid, R. Mohr, and C. Bauckhage. Evaluation of East_park with rotation and scale changes.
Interest Point Detectors. International Journal of Com-
puter Vision, Vol. 37, pp. 151-172, 2000.
[7] G. Scott and H. Longuet-Higgins. An algorithm for
associating the features of two patterns. Proceedings of
the Royal Statistical Society of London, Vol. B244, pp.
21-26, 1991.
[8] G. Xu, E. Nishimura, and S. Tsuji. Image correspon-
dence and segmentation by epipolar lines: Theory,
algorithm and applications. Technical report, Dept. of Figure 4. Matching result for image pair Car
Systems Engineering, Osaka University, Japan, 1993. with translation and rotation.

Proceedings of the Third International Conference on Image and Graphics (ICIG’04)


0-7695-2244-0/04 $20.00 © 2004 IEEE

You might also like