Operator is a comer detector since it defines interest points as
points where there is a large intensity variation in every
direction. Directional variance is measured over small square
windows. Sums of squares of differences of pixels adjacent in
each of four directions (horizontal, vertical and two diagonals)
over each window are calculated, and the window's interest
measure is the minimum of these four sums. Features are
chosen where the interest measure has local maxima. The
feature is conceptually the point at the center of the window
with this locally maximal value [7].
2.2 Feature point matching
In this step, the correspondence between the feature point
detected in the sensed image and those points in the reference
image is established. This step is the key of the image
registration. With many highly matching precise control points,
we can calculate the relation of the two coordinates system of
the images on the model we proposed with what the
registration will be done. There five key step in this part.
2.2.1 General match: For enhance the matching precise and
efficiency. We should estimate the approximate space of each
feature point. As you know, search the corresponding point all
over the slave image is impossible. So with general match, we
can limit the search area in a small window. In this paper, we
choose 1~3 artificial tie points with which we can obtain the
parallel move, rotate and zoom of the two images. If the tie
points are more than three, we can use affine model. With the
parameters we got, we can cast the feature points on the slave
images. So we just need to search the corresponding point
round the points which been cast.
J(S a -S i t /N)(S,,-S 2 t -/N)
(1)
in which
m n
Sg' ~ ^ i ^ 1 & i+rj+c
i=l y=l
Sgg' ~ ^ 1 ^ 1 &i,j * & ,+r J+ c
/=1 7=1
m n
S g'g' = S' 1 ' l y.s' 2 ‘+r,j+c
<■=1 7=1
m n
<=1 7=1
m n
s gg = X2 g2<j
1=1 7=1
m,n is the template line length and row length. N=m*n, This
statistical measure has the property that it measures correlation
on an absolute scale ranging from [-1,1]. The images which are
more similar with each other will have a large correlation.
B . Mutual Information
2.2.2 Pyramid-layered template matching technique:
Though using general matching, there are still some distance
from the correct corresponding point. If using large range
search method, not only the efficiency would be very low, but
also very hard to get good match result. We solved this math
problem perfectly by using the pyramid-layered template
matching technique. The pyramid-layered template matching
method is a match method that is related from top and bottom.
This correlation method is a strategy that is from coarse to
precise. It gained high match reliability by using the correlation
from the top of the pyramid-image. At the same time it gained
high precision by the correlation from the bottom of the
pyramid-layer.
The number of the pyramid-layer is determined by the deviation
between the initial predicting pixel position and the actual pixel
position. Where, the bottom of the pyramid-image is the
original aviation image. The top of the pyramid-image is
formed by put together 3X3 pixels from the bottom of the
pyramid-image. By using this sequence image correlation
method, we could calculate the correct point position form
coarse to precise.
2.2.3 Similarity metrics: In the points matching process,
one of the most important things is to choose a similarity
metrics. Correlation and Mutual Information are the most
widely used.
A. Correlation
Correlation can be used in multi-source remote sensing images
[11] which defined as
The strength of the mutual information similarity measure lies
in tile fact that no assumptions are made regarding the nature of
the relation between the images intensities in both modalities,
except that such a relationship exists. Therefore, the MI
criterion is very general and has been used in many different
image registration problems [3-6].
Unlike the metrics of the mean-square difference and cross
correlation coefficient, mutual information based algorithms do
not assume any functional relationship between the intensities
of the images to be registered: thus, they are more general,
effective and robust to handle the images when intensities of
one image have a nonlinear relationship with those of the other
image.
Mutual information has its roots in information theory. The
mutual information (MI) of two random variables A and B is
defined by
P AB (a,b)
P A {a).P B {b)
(2)
Where and P ii( b )
are the marginal probability mass
functions, and A.B( a > ) j s the j 0 i n t probability mass
function. MI measures the degree of dependence of A and B by
measuring the distance between the joint
distribution
P a.b (°> b )
and the distribution associated with the