View and illumination invariant iterative based image

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
Volume: 02 Issue: 11 | Nov-2013, Available @ http://www.ijret.org 185
VIEW AND ILLUMINATION INVARIANT ITERATIVE BASED IMAGE
MATCHING
Y. Ramesh1
, J. Sofia Priya Dharshini2
1
Student, 2
Associate Prof., Department of ECE, RGMCET, Nandyal, Kurnool Dist., Andhra Pradesh, INDIA
ramesh7792@gmail.com, sofi_prida@yahoo.co.in
Abstract
In this paper, the challenges in local-feature-based image matching are variations of view and illumination. Different methods have
been recently proposed to address these problems by using invariant feature detectors and distinctive descriptors. However, the
matching performance is still unstable and inaccurate, particularly when large variation in view or illumination occurs. In this paper,
we propose a view and illumination invariant image matching method. We iteratively estimate the relationship of the relative view and
illumination of the images, transform the view of one image to the other, and normalize their illumination for accurate matching. The
performance of matching is significantly improved and is not affected by the changes of view and illumination in a valid range. The
proposed method would fail when the initial view and illumination method fails, which gives us a new sight to evaluate the traditional
detectors. We propose two novel indicators for detector evaluation, namely, valid angle and valid illumination, which reflect the
maximum allowable change in view and illumination, respectively.
Keywords-Feature detector evaluation, image matching, Iterative algorithm.
-----------------------------------------------------------------------***-----------------------------------------------------------------------
1. INTRODUCTION
Image matching is a fundamental issue in computer vision. It
has been widely used in tracking, image stitching, 3-D
reconstruction, simultaneous localization and mapping
(SLAM) systems, camera calibration, object classification,
recognition, and so on. Image matching aim to find the
correspondence between two images of the same scene or
objects in different pose, illumination and environment In this
paper, we focus on local feature-based image matching. The
challenges of this work reside in stable and invariant feature
extraction from varying situations and robust matching.
Generally speaking, the framework of a region feature based
image matching consists of three steps. Detecting stable
regions. Interesting points are extracted from images, and the
region of interest is the associated circular (or elliptical) region
around the interesting point. Generally, researchers use corner
(Harris [1], SUSAN [2], CSS [3], etc.) or center of silent
region (SIFT [4], SURF [5], DoH [6], HLSIFD [7], etc.) as the
interesting point since they are stable and easy to locate and
describe. The radius of the region is determined by a priori
setting (Harris corner) or the region scale (scale invariant
features). The total number of features detected is the
minimum number of the features extracted from the matched
images. Describing regions. Color, structure, and texture are
widely used to describe images in the recent literature
Descriptors with edge orientation information (SIFT and
HOG) are also very popular since they are more robust to
scale, blur, and rotation. Matching features. Local features
from two images are first matched when they are the nearest
pair. A handful of distances can be used in practice, such as
L1distance, L2 distance, histogram intersection distance [8],
and earth mover’s distance [9]. If the nearest distance is higher
than k times (k ϵ (0, 1) empirically) of the second nearest
distance, the nearest matching pair will be removed. These are
the very initial matching results.
Fig1 Illustration of the proposed matching algorithm Ir and It
are the images to be matched. Ie is simulated from It by
transformation T. Ir is difficult to match with It for the
difference of view point and illumination, whereas Ie is easier
to match with It since they are closer in the parameter space.
The three parts of the detect–describe–match (DDM)
framework determine the performance of image matching.
The first step is the basis of this framework. Unstable and
variant features increase the difficulties of the next two steps.
Researchers mostly focus on the first step for invariant feature
extraction and have proposed many excellent detectors [1],

__________________________________________________________________________________________
[4], [5], [7]. However, an important experience of a previous
work is that all the aforementioned feature detectors are not
strictly invariant to the changes of view and illumination. For
larger changes, there would be few invariant features that can
be extracted from both images to be matched. This motivates
us to think the essential difference of images with different
view and illumination. Normally, a question need to be
answered: whether an object in two images with different
views and illumination looks like the same one, supposing
there are two images with a large view change, as shown in
Fig. 1. The two top images are the same object in different
views. They are so different in appearance that they can be
considered as two different objects. We do not attempt to find
invariant local feature detectors as in a previous work but
focus on a better framework for image matching.
2. VIEW AND ILLUMINATION INVARIANT
IMAGE MATCHING
2.1. General Definition of Image Matching
Two images of the same object or scene are shown as two
points in parameter space P of the object (scene). Let I be the
original appearance of an object, = ( ( )) be the real
appearance of the object shown from an image, where L
indicates the illumination, and H is the object transformation
factor from a normal pose. Here, we define the parameter
space of a given image I as P = { . } (simply written as
= { . }in the following). Translation L and H is a point in
the parameter space; thus, the observed image is shown as a
point in the parameter space, which is expanded by object I.
Therefore, the purpose of image matching is to find
transformation T between the two points in the parameter
space { . } → { . } or, in other words, | =
( ). . ). The purpose isto find the coordinate
differences between the two points. The norm of this space is
difficult to define since illumination factor L and
transformation H are totally independent and cannot be
combined together. In this paper, we simply use images with
planar objects; therefore, H is the homograph transform
matrix, and L is the histogram matching function that
transforms the histogram of one image to a specific one.
2.2. Proposed Method
Denote the reference image and test image to be matched as
and. Suppose that the true pose transformation matrix from
I to I is and the illumination change function is . The
relationship between !"# is
($) = %( ) = & ( )' = & ( $)' (1)
Where % is the true transformation between !"# , is the
homogeneous co-ordinates, and + = (,, ., 1). If there exists
approximate estimations about illumination and
transformation, the could be transformed to an estimated
image , i.e.,
(/) = ( ) = ( ( $)) (2)
Where H denotes the view point transformation L and denotes
the illumination transformation. If T is not a very rough
estimation between !"# , the estimated image I would be
more similar to 1ℎ!" itself. In other words,
34 56748 17 than to . Thus, the matching between
!"# will be easier, as shown in Fig. 1.
In this way, we propose the following iterative image
matching process:
9X = 9( ;) = <( =( <X>)) ( ; = )
?X = TA(IAB9) = C( CB<( C$D)) (i > 1) (3)
2.3. Algorithm 1: The Proposed Method:
Initial: ; = { =. =} = LM. 1NOP, = ;, QR, QS;
Iterate
3 = 3 + 1;
V413W!18 ?: C. C;
= ? ∘ ;
Y = YZ ∗ Y.
Transform ?B9 17 ? `. (3);
Until C − M < σd, e C − 1NOe < QS7 3 > ".(E is the
unit matrix. f?is a histogram transformation
vector, QR !"# QS are convergence thresholds.)
Return T, H
The algorithm is summarized in Algorithm 1. The final
estimation of the % is
% = ⋯ ∘ h ∘ hB9 ∘ ⋯ ∘ i ∘ 9 (4)
≈ l ∘ lB9 ∘ ⋯ ∘ i ∘ 9 (5)
Where “∘” denotes function composition. Our experiments in
Section IV-B show the convergence of the iteration with SIFT
and the performance with respect to the number of iterations
2.4. Estimate the Parameters H and L
General image-matching methods by local features focus on
the first parameter H since the concerned issue is the space
correspondence between the two images. One of the advantage
of the proposed method is that it also estimates the
illumination change, which makes matching much better when
illumination has changed.

__________________________________________________________________________________________
The purpose of general image-matching methods is to find the
transformation matrix between the reference image and the
test image. These methods are invariant to rotation, scale, and
partially affine changes. The H can be easily estimated by the
general methods without other information. First, we extract
features from the matching images and obtain features
descriptions (which method is used is not important). Then,
we match two features when they are the nearest pair in the
feature space. Here, fi norm is used to calculate the distance
between to the features The RANSAC algorithm is employed
to calculate transformation matrix. The general methods, i.e.,
HarAff, HesAff, SURF, SIFT, and HLSIFD, all can be used as
the feature extraction method. We call them I-HarAff, I-
HesAff, ISURF, ISIFT, and IHLSIFD (“I” indicates
“Iterative”), respectively. Moreover, image matching is
usually used in video sequences.
Fig.2 Illustration of the histogram transformation (a) The original image. (b) Darker image. (c) Transformed image from (b) according
to the histogram of (a). (d) Brighter image. (e) Transformed image from (d) according to the histogram of (a). (f)–(j) The
corresponding histograms of (a)–(e).
We assume that the difference between two consecutive
frames is not large, and the object or the camera smoothly
moves. Thus, the th frame’s transformation Ccan be
approximated by the previous results. Different detectors and
descriptors have been developed to extract illumination
invariant local features. The gradient direction histogram is
normalized to form the descriptors. There is usually a tradeoff
between the distinction and the invariance. If we do not
normalize the descriptors, they will be sensitive to
illumination changes but more distinctive. Computing
detectors and descriptors also cost much time. Conversely, the
detector will be more efficient if we do not require the detector
to be invariant to illumination change. We want to keep both
illumination invariant and descriptor distinctive in our method.
Thus, it is necessary to estimate the illumination change
between the two images.
Estimating the illumination is a challenging issue since the
objects in the images are often accompanied by clutter
background or noise. Benefitting from the estimation of the
transformation matrix, we can warp the test image to another
pose in which the object pose looks similar to that in the
reference image. Accordingly, approximate object
segmentation would be obtained on the simulated image. To
eliminate the occlusion, we only use the matched regions. The
matched regions are the region in the scale of the matched
interesting points. First, we calculate the illumination
histogram of the two images in the matched region. Second,
we fix one image and calculate histogram translation function
L from the other image to the fixed one. Suppose the
histogram of the fixed image is ℎ9 and the histogram of the
other image is ℎi . We calculate the cumulative functions
ofℎ9!"#ℎi − n9!"#ni. Finally, the translation function is
= op
B<
o<. q
Since the cumulative function of gray histogram is always
monotonically increasing, inverse function oB<
always exists.
We transform the histogram of the test image according to the
histogram of the reference image to normalize the illumination
between the pair, as shown in Fig. 2, and the whole procedure
is illustrated in Fig. 3. To sum up, we estimate transformation
matrix H between the matching pairs by feature detector,
estimate the illumination relationship, and change one of the
images according to the color histogram of the other to map
the pose and illumination of the object in one image to the
other.
2.5. Relationship between the Iterative Algorithm
and ASIFT
The proposed iterative method is similar to ASIFT in ASIFT,
the features are not invariant to affine change, but they cover
the whole affine space, as shown in the middle block in Fig. 4.
Every simulation of the reference image is one pose of the
image in the affine space. Therefore, parts of the simulations
of the reference image and the test image should have similar
poses in the affine space theoretically. The simulations of the
reference image and the test image are independently
constructed. No mutual information is used in the simulations

__________________________________________________________________________________________
Table 1: Comparition of Asift and our Method
ASIFT Our method
Simulation to ref image Yes No
Simulaton to test image Yes Yes
Number of simulations Many Few
Number fo features 104
~ 105
103
Pose simulation Yes Yes
Illumination simulation No Yes
NCM High High
RS Vey low High
Affine invariancy Full Partial
Computational cost High low
Rea-time No Yes
Simulating in a high density in the affine space, may supposed
image poses are constructed, and then, they are matched in a
general way. The number of matches increases with the
number of the simulations. ASIFT indeed increases the
invariability of the image-matching method. However, it does
not care what the transformation matrix between the reference
and test images is, by trying many possible transformations
and combining the matches. Thus, ASIFT can be regarded as a
sampling method around the original points in parameter
space, whose properties are shown in the left column of Table
I. Essentially, our method also constructs “simulation.” We
simulate the image not only in the pose but also in
illumination, as shown in the right part of Fig. 4. In addition,
we transform one simulation per iteration, and in most tasks,
two iterations are enough. We will give an experiment to
illustrate this in Section IV-B. Benefiting from few
simulations, the computational cost of our method is very low,
compared with ASIFT, which simulates much more images
than our method.
Fig. 3 Procedure of illumination estimation (a) The test image. Warp (a) by the estimated transformation matrix to generate (b). (c)
Mask with the matched regions labeled as 1, and the unmatched regions labeled as 0. (d) The inner product of (b) and (c). (e) The
reference image. (f) The inner product of (c) and (e). (g) Illumination simulated image from (d) according to the histogram of (f).
Fig. 4 Relationship among the general framework, ASIFT, and the proposed method. (Left block) The general framework, (middle
block) ASIFT, and (right block) ours. The general DDM framework directly estimates the transformation between two images. It is
simple but coarse. ASIFT simulates many poses of the two images to cover the affine space, whereas our method estimates the
transformed pose first and then accurately matches in the projective space.

__________________________________________________________________________________________
A coarse-to-fine scheme can reduce the computational time of
ASIFT to three times of the SIFT, whereas our method only
costs two times. One drawback of the proposed method is that
it does not increase the invariability of the original method.
When the initial method fails in matching images, the
proposed method also fails. One promising method to
overcome this shortage is to combine the proposed method
with the ASIFT, which improves both the invariability and the
accuracy. Furthermore, the histogram matching may amplify
noise that seems to affect the performance. A few more key
points would be extracted after the histogram matching, but
they would not affect the performance too much. We will
show this in Section IV-C. Experimental results show that the
performance of the proposed framework reaches a comparable
level, compared with ASIFT with much fewer features totally
detected, as shown in Table III. Therefore, the RS of our
method is much higher than that of ASIFT. The computational
cost of our method is much lower than that of ASIFT because
much fewer features are required.
Above all, there are some common properties between
iterative SIFT (ISIFT) and ASIFT. Instead of directly
matching the original images, both methods find good
simulations of the original pairs. ASIFT samples the
imaginary images in the whole affine space, whereas our
method directly estimates in the whole parameter space. We
should point out here that these comparisons and the
experiments shown in the following section are all under the
situation that the original method, i.e., SIFT, still works. When
it fails, the proposed method also fails, whereas the ASIFT can
still obtain a valid result.
3. EXPERIMENTAL RESULTS
3.1. Database
In the first experiment, we want to show the performance of
the proposed method. We capture two images with changes
both in illumination and view. This experiment is not used for
comparison, but it only shows the effectiveness of the
proposed method. To evaluate the performance of the
proposed image-matching framework, we do experiments on
the database provided by Mikolajczyk.1 This database
contains eight groups of images with challenging
transformations. Parts of them are shown in Fig. 5. We
compare the proposed method with ASIFT and the usual
DDM framework with the state-of-the-art detectors: SURF,
SIFT. In addition, two evaluations on the detectors through
our strategy are proposed. One of them tests the adaptive
capacity on the view change, and the other tests the capacity
on the illumination change. To finish the two evaluations, we
build two databases. One of them contains 88 frames with
view changes from 0^ to 87^. The other one contains 55
frames with light exposure changes from 40 to 14 (0.1 EV).
The two databases contain continuous transformation frames.
Thus, we can evaluate the view invariant ability of the
detectors at a 1 interval and the illumination change invariant
ability at a step of 0.1 EV. Such databases seldom appear in
the open literature, and they will be currently available on the
Internet.
3.2. Convergence
As we mentioned in Section III-B, the number of iteration is
an important parameter. A question that should be answered is
whether more iterations bring better performance.
Experiments show that, under the proposed framework, our
method converges very fast. Fig. 6 shows an experiment on
matching two images. The reference image is captured from a
frontal view, and the test image is captured from a view angle
of 60 ^,as shown in Fig. 6(a) and (b), respectively. Here, SIFT
is used as the base detector. The RS and NCM of our method
and the DDM framework with SIFT are drawn for
comparison, as shown in Fig. 6(c) and (d).
Table 2: Performance of sift, isift with only pose estimation,
isift both pose and illumination estimation
SIFT ISIFT(H) ISIFT(H&L)
Total detected 436 388 2021
Total matches 50 64 169
NCM 39 57 153
RS(%) 8.95 14.7 7.57
MP(%) 78.0 89.1 90.5
The results show that more iteration does not necessarily
increase the performance significantly, whereas it increases
the computation time linearly. When, n=2 the performance
significantly increases. The NCM increases more than 300
matches from only 12 to 365, and the RS increases from
12.1% to 37.1%. However, as n further increases the
performance little, the NCM only moves around 360, and the
RS moves around 37%. Thus, two iterations are enough in
general situations, and we use n=2 in the following
experiments. Moreover, all the features in this experiment and
the following experiments are described by a SIFT descriptor,
except SURF, which is described by a SURF descriptor.
Following the general evaluation, three criteria are often used
as feature evaluator.
1) NCMs are the number of total correct match pairs.
2) RS is the ratio between the NCM and the minimum of total
number of features detected from the image pair RS
NCM/TOTAL.
3) Matching precision (MP) is the ratio between the NCM and
the number of matches MP NCM/Matches.
3.3. Performance
In this experiment, a brief view of the performance of the
proposed method is given. We use SIFT as the base detector in
this experiment (ISIFT). Two images with both view and
illumination changes are matched here.

__________________________________________________________________________________________
We first match the two images by SIFT, and then, we only
simulate the pose of the left image in our strategy. Finally, we
simulate both pose and illumination. The matching results are
shown in Fig. 7 and Table II. View and illumination changes
both degrade the performance of the general method. SIFT
could achieve 8.95% RS with 39 correct matches. ISIFT, with
the pose estimation only, could achieve 14.7% RS with 57
correct matches. When we estimate the pose and illumination
changes, the number of total detected features rapidly
increases, and the NCM increase to 153. Because histogram
matching amplifies noise in simulation, many fake features are
detected, and the RS is reduced to 7.57%. This experiment is
only a brief view of our strategy, and more experiments will
be presented in the following. We estimate the global
illumination change between the matching pair to increase the
NCMs. The illumination change is usually continuous in the
image. Thus, revising the illumination of part of the image
could benefit to other regions. Our algorithm does not increase
the invariance of the original detector, but it increases the
accuracy, stability, and reliability of the matching results.
When SIFT fails, our method also fails. However, when SIFT
works, but not robust, the proposed method will play an
important role. More matches could not increase the
invariance, but it can increase the accuracy of alignment when
the matching by SIFT is inaccurate.
Fig.5. Four groups of images that we used for comparison [33] Each group contains one or two transformations with six images,
and only parts of them are shown here. (a) Boats (scale rotation). (b) Graf (view). (c) Wall (view). (d) Leuven (illumination)
Fig.6. Experiments of convergence. (a) The reference image. (b) The test image. (c)–(d) The NCM and RS of ISIFT compared with
SIFT.

__________________________________________________________________________________________
Fig.7. Matching results of SIFT and ISIFT. (a) Matching
result of SIFT. (b Matching result of ISIFT with only pose
simulation (H). (c) Result of ISIFT with both pose and
illumination ( H and L) simulation.
In other words, the advantage of the proposed method is that
the performance does not degrade with the increase in the pose
change or transition tilt, which is addressed in the valid range.
Additionally, the local key point location will be more
accurate than that of the original detected point. To
corroborate this point of view, we show an extra experiment in
the following. The first row in Fig. 8 is the matching results of
SIFT, and the second row is the results of ISIFT. Both the
matches and the alignment residual error are shown. From this
experiment, we can find that our algorithm can obtain less
error than SIFT, and the NCM affects the accuracy of
matching very much.
3.4. Comparison
We choose the database provided by Mikolajczyk and
compare them with SURF, SIFT and ASIFT. Four pairs of
images with scale, view, and illumination change are tested, as
shown in Fig. 9. The images on top are the reference image,
and those at the bottom are the test image. Table 3 is a
comparison of this experiment in terms of NCM, RS, and MP.
Our method estimates the pose and illumination of the
matching pairs and simulates the reference image. Therefore,
the simulated image is closer to the original image, which
contains most information of the original image, shortening
the distance of the matching pairs in the parameter space.
First, the NCM of the ISIFT is much higher than that of the
traditional methods. They obtain 584 matches, whereas SURF
obtains 9 matches, and SIFT obtains 46 matches in the Graf
(affine change situation; second row in Fig. 9). SURF and
SIFT obtain 793 and 2837 features, respectively. Thus, the RS
of ISIFT increases to 36.4%, whereas that of SURF and SIFT
is only 1.14% and 1.62%. This implies that the efficacy of IIM
framework is much better than the traditional DDM
framework. We increase about 32 times and 22 times RS in
this view-change experiment. With the significant increasing
performance, we can make the matching more stable and
reliable. Similarly, more correspondences are found in other
experiments, particularly under affine and illumination change
situations. Our method does not significantly increase NCM
under only scale change comparing to SIFT, SURF, and
HLSIFD since they are theoretically scale invariant. The RS
and MP also significantly increase.
However, in extreme situations when SIFT fails in the first
matching, our algorithm also fails. The proposed method can
increase the stability, reliability, and accuracy of the original
detector, but it cannot increase the invariance. A solution is
integrating the proposed method into ASIFT as the second
layer to refine the original matching results. We will show an
experiment in Section IV-E. ASIFT also obtains 105, 465,
556, and 157 matches from Boat, Graf,Wall, and Leuven
matching images (61, 46, 409, and 259 matches are found by
SIFT, respectively). However, these matches are calculated
from 29 985, 45 151, 64 908, and 22 562 extracted features.
Indeed, ASIFT increases the NCM, but they need to extract
much more features from the images, which cost much time in
computation. More detail results are summarized in Table III.
In this paper, we try to link our method with the general
optimization theory. Essentially, the target of image matching
is finding the correspondence. We want to find the
transformation function between the matching pair, which can
minimize the matching error. Thus, we optimize the view
difference and then optimize the illumination. With the two-
step optimization, our method can find more accurate
transformation function. Different from ASIFT, the proposed
method does not increase the invariance of the original
detector, but it increases the stability and reliability.

__________________________________________________________________________________________
Fig.8. Matching error of the SIFT and the proposed method. (a) The matches of SIFT. (b) The residual error of SIFT. (c) The matches
of ISIFT. (d) The residual error of ISIFT.
Fig. 9 Matching results of four groups of images. (Test images from top to bottom) Boat, Graf, Wall, and Leuven. The results of the
correct matches are drawn in blue or white lines.

__________________________________________________________________________________________
Table 3: Comparison of the algorithms on view change pairs
Methods SURF SIFT ASIFT(HR) ISIFT
Boat / Scale Total
Matches
NCM
RS(%)
MP(%)
722
43
8
1.25
18.6
7986
125
61
0.764
48.8
29985
-
105
0.35
-
615
94
79
12.8
84
Graf / Affine Total
Matches
NCM
RS(%)
MP(%)
793
34
9
1.14
26.5
2837
210
46
1.62
21.9
45151
-
465
1.03
-
1605
586
584
36.4
99.7
Wall / Affine Total
Matches
NCM
RS(%)
MP(%)
1730
78
40
2.31
51.3
7094
452
409
5.77
90.5
64908
-
556
0.857
-
5358
834
833
15.5
99.9
Leuven /
Illumination
Total
Matches
NCM
RS(%)
MP(%)
647
172
161
24.9
93.6
999
289
259
25.9
89.6
22562
-
157
6.96
-
1159
379
344
29.7
90.8
3.5. Real-Time Image Matching
An important application of image matching is object
detection and poses estimation in video frame. Suppose that
the camera smoothly moves and the reference image can be
matched with the first frame, the estimation of the
transformation matrix from the reference image to certain
frame in video can be initialized from the matching of the
previous frame. In addition, we match the first frame with the
reference image directly by local-feature-based image-
matching method. We directly use SIFT here. The RS and
NCM of our method and SIFT are shown in Fig. 14, and parts
of the matching results of ISIFT and SIFT are shown in Fig.
13. The RS of our method (ISIFT is used here) stays around
30%, and NCM is always higher than 100 pairs in this
experiment.
The RS of SIFT is running around 7%. Only a small part of
features are useful for the correspondence calculation. The
NCM of SIFT is about 70 matches, which is lower than that of
the proposed method. The mean of the RS and NCM of the
ISFIT and SIFT is, respectively 29.6%, 137, 5.7%, and 66.
Our method accurately calculates matches all through the
video frames, even in large view changes such as frames 750
to 900. To sum up, ISIFT is very accurate and stable in real
applications.
We develop a real-time image-matching system to show the
efficiency. The proposed method could cope with a wide
range of view and illumination changes with stable matches,
as shown in Fig. 15.We compare the real performance of
SURF and SIFT by using them as our basic detector. ISURF is
faster than ISIFT; however, it is not as stable as ISIFT. The
system is implemented on a computer with two dual-core 2.8-
GHz central processing unit, and the processed image size is
640 X 480. The matching could be finished in 80 ms, with
parallel coding in a algorithmic level.
CONCLUSIONS
In this paper, we have proposed a novel image-matching
algorithm based on an iterative framework and two new
indicators for local feature detector, namely, the VA and the
VI. The proposed framework iteratively estimates the relative
pose and illumination relationship between the matching pair
and simulates one of them to the other to degrade the
challenge of matching images in the valid region (VA and VI).
Our algorithm can significantly increase the number of
matching pairs, RS, and matching accuracy when the
transformation is not beyond the valid region. The proposed
method would fail when the initial estimation fails, which is
relative to the ability of the detector. We have proposed two
indicators, i.e., the VA and the VI, according to this
phenomenon to evaluate the detectors, which reflect the
maximal available change in view and illumination,
respectively. Extensive experimental results show that our
method improves the traditional detectors, even in large
variations, and the new indicators are distinctive.

IJRET: International Journal of Research in
__________________________________________________________________________________________
Volume: 02 Issue: 11 | Nov-2013, Available @
REFERENCES
[1]. C. Harris and M. Stephens, “A combined corner and edge
detection,” in Proc. 4th Alvey Vis. Conf., 1988, pp. 147
[2]. S. M. Smith and J. M. Brady, “Susan—A new approach to
low level image processing,” Int. J. Comput. Vis., vol. 23, no.
1, pp. 45–78, May 1997.
[3]. F. Mokhtarian and R. Suomela, “Robust image corner
detection through curvature scale space,” IEEE Trans. Pat
Anal. Mach. Intell., vol. 20, no. 12, pp. 1376
1998.
[4]. D. G. Lowe, “Distinctive image features from scale
invariant keypoints,” Int. J. Comput. Vis., vol. 60, no. 2, pp.
91–110, Nov. 2004.
[5]. H. Bay, A. Ess, T. Tuytelaars, and L. V. Go
up robust features (SURF),” Comput. Vis. Image Understand.,
vol. 110, no. 3, pp. 346–359, Jun. 2008.
[6]. T. Lindeberg, Scale-Space Theory in Computer Vision.
Norwell, MA: Kluwer, 1994.
[7]. Y. Yu, K. Huang, and T. Tan, “A Harris
invariant feature detector,” in Proc. Asian Conf. Comput. Vis.,
2009, pp. 586–595.
[8]. A. Barla, F. Odone, and A. Verri, “Histogram intersection
kernel for image classification,” in Proc. Int. Conf. Image
Process., 2003, vol. 3, p. III–513–16 [Online]. Avai
http://dx.doi.org/10.1109/ICIP. 2003.1247294, vol.2
[9]. Y. Rubner, C. Tomasi, and L. J. Guibas, A Metric for
DistributionsWith Applications to Image Databases.
Washington, DC: IEEE Comput. Soc., 1998, p. 59.
BIOGRAPHIES
Y. Ramesh received B.Tech degree from
GATES Engineering College in the year
2010 and currently pursing M.Tech in
Digital Systems and Computer Electronics
at RGM College of Engineering and
Technology, Nandyal, Kurnool(dist),
Andhra Pradesh. His areas of interest are
Digital Image Processing, Data
Communication and the application of Digital systems.
J. Sofia priya dharshini
B.Tech degree (Electronics &
Communication Engineering) in the year
2005, M.Tech (Digital Systems and
Computer Electronics) in 2009
pursuing Ph.D. (Wireless Communications
and Networking) from JNTU Anantapur,
A.P, India. Presently she is working as
Associate Professor in the dept. of ECE, RGMCET, Nandyal.
She has published two papers in International Journals and
nine papers in National and International Conferences. Her
area of interest includes Wireless communications and
Networks, Mobile Computing and Video and Image
processing.
IJRET: International Journal of Research in Engineering and Technology eISSN: 2319
__________________________________________________________________________________________
2013, Available @ http://www.ijret.org
C. Harris and M. Stephens, “A combined corner and edge
in Proc. 4th Alvey Vis. Conf., 1988, pp. 147–151.
A new approach to
image processing,” Int. J. Comput. Vis., vol. 23, no.
F. Mokhtarian and R. Suomela, “Robust image corner
through curvature scale space,” IEEE Trans. Pattern
Intell., vol. 20, no. 12, pp. 1376–1381, Dec.
D. G. Lowe, “Distinctive image features from scale-
Int. J. Comput. Vis., vol. 60, no. 2, pp.
H. Bay, A. Ess, T. Tuytelaars, and L. V. Gool, “Speeded-
features (SURF),” Comput. Vis. Image Understand.,
Space Theory in Computer Vision.
Y. Yu, K. Huang, and T. Tan, “A Harris-like scale
detector,” in Proc. Asian Conf. Comput. Vis.,
A. Barla, F. Odone, and A. Verri, “Histogram intersection
image classification,” in Proc. Int. Conf. Image
16 [Online]. Available:
2003.1247294, vol.2
Y. Rubner, C. Tomasi, and L. J. Guibas, A Metric for
Applications to Image Databases.
Soc., 1998, p. 59.
received B.Tech degree from
GATES Engineering College in the year
2010 and currently pursing M.Tech in
Digital Systems and Computer Electronics
at RGM College of Engineering and
Technology, Nandyal, Kurnool(dist),
. His areas of interest are
Digital Image Processing, Data
Communication and the application of Digital systems.
J. Sofia priya dharshini has received
B.Tech degree (Electronics &
Communication Engineering) in the year
2005, M.Tech (Digital Systems and
Computer Electronics) in 2009 and also
pursuing Ph.D. (Wireless Communications
and Networking) from JNTU Anantapur,
A.P, India. Presently she is working as
Associate Professor in the dept. of ECE, RGMCET, Nandyal.
She has published two papers in International Journals and
in National and International Conferences. Her
area of interest includes Wireless communications and
Networks, Mobile Computing and Video and Image
eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
194

View and illumination invariant iterative based image

More Related Content

What's hot (17)

Viewers also liked (20)

Similar to View and illumination invariant iterative based image (20)

More from eSAT Publishing House (20)

Recently uploaded (20)

View and illumination invariant iterative based image