Graphics Reference
In-Depth Information
nn
nn
nearest neighbor
and second nearest neighbor
to all other images,
1
2
d
d
d
1
/
d
and their Euclidean distance
and
. If the value of
is less than
2
2
nn
is considered as the best match point;
(e)
Once all image feature points have been traversed, it is also need to validate the
feature matching results. Assume the feature point
0.6,
1
n
I
of image
has matching
J
n
index
with image
, then checks whether the matching index of the feature
ij
J
I
n
n
;
(f)
Using RANSAC algorithm [15] to estimate the motion model between adjacent
images, and exclude the outliers.
point
of image
with image
equals to
ij
4
Multi-focus Image Fusion
Through the above registration process, a series of images with good alignment can
be obtained, and these aligned images are the input of the multi-focus fusion algo-
rithm. Multi-focus fusion needs to compute the desired image by keeping the “best”
parts in the multi-focus image sequence. This process is guided by the quality meas-
ure, which will be consolidated into a scalar-valued weight map. It is useful to think
of the multi-focus image sequence as a stack of images. The final image is then ob-
tained by collapsing the image stack using weighted blending strategy. In this step, it
is assumed that images are perfectly aligned. Otherwise, the input sequence should be
aligned first as in part 3.
One of the most effective and canonical method used to describe image with multi-
resolution is the image pyramid proposed by Burt and Adelson [16]. The essential
idea of image pyramid is to decompose the source image into different spatial resolu-
tions through some mathematical operations. The most commonly used two image
pyramids representation are Gaussian pyramid and Laplacian pyramid, and Laplacian
pyramid can be derived from the Gaussian pyramid, which is a multi-scale representa-
tion obtained through a recursive low-pass filtering and down-sampling operations.
For example, the finest level
G
of Gaussian pyramid is equal to the original image,
G
G
and the next level
and so on. Both
sampling density and resolution are decreased from level to level of the pyramid, and
this local averaging process which generates each pyramid level from its predecessor
is called REDUCE:
is obtained by recursively down-sampling
G
=
+
REDUCE
G
(
)
(1)
l
1
l
2
2
−
G
=
w
(
m
,
n
)
G
(
2
i
+
m
,
j
+
n
)
(2)
l
+
1
l
mn
=
2
=
−
2
Search WWH ::
Custom Search