XIX-B3, 2012
response spread,
'ecome penalized,
ess of the cut-off.
to calculate this
tor.
act the common
r images, such as
the two image
1g and, therefore,
rast to commonly
s, point features),
s within the local
left to right: the
ote that the LWA
n the PC map,
; the sky, sea, and
LARITY
e independent of
n ofthe LAP and
it: CSM, which is
tion than those
] is therefore able
mage matching.
mmon amplitude
1sor images, and
commonly used
etic analysis and
VA is an image
his problem, we
lation (ZNCC) as
contrast invariant
'esponding LWA
hin the template
(u, v) in 84 3
window, where
sed as follows:
YYX(AGsiy*D-7) (a, u+iv+ )-8,)
Xen) EX ene n-i, y
ZNCC(u,v) =
(11)
Because of the normalization, ZNCC is invariant to image
contrast linear changes, which can be used to compensate the
disadvantage of the LWA.
As the LAP represents the phase of local frequency vectors, it is
an image contrast invariant variation. Therefore, for efficiency,
we present an Extended Mean Absolute Difference (EMAD) as
the similarity measure rather than ZNCC. If we define f, and g,
as the LAP pair, the definition of EMAD is given by:
EMAD(u,v) = YX255- (Us Gy 7g. Qi iv 7) do
If we define y... as the maximum value of ZNCC matrix and
Max, 98 the maximum value of EMAD matrix respectively,
then the newly presented CSM can be expressed as follows:
CEMOLY)S ZNCC(u,v)+1 X EMAD(u,v) (13)
Mayo tl +e Max, +2
A very small positive constant € is added to the denominator in
case of a small Max jvc and/or Max mp" From Equ.13, we can
see the value range of EMAD and ZNCC are both normalized,
therefore, they have the same value range. The maximum value
of the ZNCC component, ZNCC(u,v)+1 is 1, and the same
Max. trt €
applies to the EMAD component, EMAD(u,v) . Therefore,
Max, ,, + ©
the CSM is able to combine the LAP and LWA information
with equal weight, and make full use of them.
4. LOCAL BEST MATCHING POINT DETECTION
In this work, the goal of local best matching point detection is
to determine the template which has the highest matching
accuracy within a certain image region. The centre of the
template is named as Local Best Matching Point. In order to
find this template, we must clarify what the feature of the
template is. If the template centered on a point is shifted, the
texture within the template obviously changes, and then we can
know this template is unique, and is also suitable for image
matching. Therefore, the local best matching point can be
detected using the self-similarity measurement. We first need to
evaluate the suitability measurement of each point surrounding
the target point and then choose the point with the highest
suitability measurement as the local best feature point. The
detailed algorithm proceeds as follows:
(1) Pick a point from the region centered on the target point,
and then calculate the suitability of the selected point. The
definition of suitability can be expressed as follows:
First, as shown in Figure 3, pick another eight points which are
centered on the selected point, and equally spaced on a circle of
radius, 1; Second, if we define the template centered on the
selected point as the centre template, and the template centered
on the other eight points as the neighboring templates, we can
choose the centre template as the reference template, and
calculate its self-similarity measurement with neighboring
templates. In this work, we use ZNCC as the self-similarity
measurement. If we define ZNCC, as the self-similarity
measurement of the neighboring template, then the suitability
measurement of the template, S , can be defined as:
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
S «1- Max(ZNCC, ) (14)
where Max(ZNCC, ) is the maximum self-similarity value
of ZNCC, .
(2) Successively pick a point from the region, which are
centered on the target point with a circle of radius, R, as shown
in Figure 3. Similar to step 1, get the suitability measurement
for all these selected points.
(3) Find the point with the highest suitability measurement, and
identify this point as the local best matching point.
(4) Conduct the image matching using the template centered on
the local best matching point. After matching, based on the
geometric transformation between the reference image and the
searching image, calculate the corresponding point of the target
point.
# PE hey t ym muy :
e mi Pd i i
# * ag + #
i >
7 E^ # * » %
* % f * i En %
$9 5 * *
2 Target 4 T
# $ E on 4
# R i ’ , #
3 * # * É *
% 2 % 3
3 % $ $ A £
# : + e
%, E P . P
d : . X M dé
* Neighboring # a 4 1
d *
“i ab
M d > qe
ues 5
Figure 3. Local Best Matching Point Detection
5. EXPERIMENTS
5.1 Experiments Using Real Images
We evaluate the performance of the proposed method using
some real images, which include: a pair of infrared and visible
images and a pair of SAR and visible images. We compare the
matching results obtained from the proposed algorithm with
those from three existing state-of-the-art methods based on
Local Frequency Response Vectors (LFRV) [9], Phase
Congruence (PC) [10], and Four Directional-Derivative-Energy
Image (FDDEI) [12]. As shown in Figure 4, many target points
are first selected from the reference image (left), and the interval
of the target points is 20 pixels. The four different image
matching approaches then conduct on the searching images
(right) to search the corresponding points. The size of the
template is 101(pixel) X 101(pixel), and the size of the
searching region is 201(pixel) X 201 (pixel). If the distance from
a matching result to its corresponding truth-value is less than
1.5 pixels, we identify this matching result as correct. The
Correct Rate obtained from four different methods are shown in
Figure 5.
The experiments using real images show that our new method is
effective for matching multi-sensor and multi-temporal images
which cannot be effectively handled by the traditional methods.
From Figure 5, we can see the average accuracy rate of our new
method is much higher than other methods. Moreover, when
matching the SAR and Visible images pair, the performances of
the three traditional methods reduce dramatically. However, our
new method is still able to robustly handle the image pair.
