Digital Signal Processing Reference
In-Depth Information
constructing the probabilistic latent semantic analysis (PLSA) model, and introduces
an unsupervised learning classification approach to wireless capsule endoscopy video
segmentation. A novel method [9] is based on BFSIFT (bilateral filter SIFT) to find
feature matches for synthetic aperture radar image registration. New AIR methods
[10] is based on the combination of image segmentation and SIFT, it is complemented
by a robust procedure of outlier removal and directly applied to remote sensing
images.
In this paper, we propose an image matching algorithm based on the 2DPCA-SIFT,
which uses 2DPCA (Two-dimensional PCA) [11] to describe SIFT feature vector, it
retains the integrity of the image of the two-dimensional spatial structure information.
2
Instruction of SIFT
SIFT algorithm is one of the most representative methods to extract local invariant
features. The algorithm mainly contains the following four steps: 1) Construct scale
space. 2) Accurately locate the keypoint. 3) Assign orientation for each keypoint. 4)
Generate keypoint descriptor.
2.1
Construct Scale Space
The scale space of an image is defined as a function,
Lxy
(, , )
σ
, that is produced
from the convolution of a variable-scale Gaussian,
Gxy
( ,
,
σ
)
, with an input image,
Ixy
:
(, )
Lxy Gxy Ixy
(,,)
σ
=
(,,) (,)
σ
∗
(1)
1
2
2
−+
(
2
y
)
2
σ
x
Gxy
(, , )
σ
=
e
(2)
2
2
π
σ
Where ( ,
can be computed
from the difference of two nearby scales separated by a constant multiplicative factor
k
:
xy
is coordinate,
)
σ
is scale coordinate.
Dxy
( ,
,
σ
)
Dxy Gxyk Gxy
(,, ) ( (,,
σ
=
σ
)
−
(,, )
σ
∗
Ixy Lxyk Lxy
(,)
=
(,,
σ
)
−
(,, )
σ
(3)
k
4
σ
k
3
σ
k
4
k
3
σ
k
2
σ
k
2
σ
k
σ
σ
σ
Fig. 1.
Image pyramid and extrema detection
