COL 10(Suppl.), S11009(2012) CHINESE OPTICS LETTERS June 30, 2012
Automatic splicing algorithm for building super viewing
field from disordered image sequence
Cong Chen (
hhh
)
1,2
and Guixi Liu (
444
BBB
UUU
)
1∗
1
School of Mechano-Electronic Engineering, Xidian University, Xi’an 710071, China
2
Chinese Aeronautical Radio Electronics Research Institute, Shanghai 200233, China
∗
Corresponding author: gxliu@xidian.edu.cn
Received December 27, 2011; accepted February 22, 2012; posted online June 20, 2012
An effective image splicing algorithm based on phase correlation and speeded-UP robust features (SURF)
operator is proposed which can sort the disordered sequence and stitch them into a super viewing field
image without any human intervention. Phase correlation in frequency domain is used for images sorting
and region of interest (ROI) estimation, and guiding features extracting and matching in spatial domain
by SURF operator and bidirectional best bin first (BBF) strategy. The experimental results demonstrate
that this algorithm not only can deal with the input images with translation, rotation and scale changes,
but also outperforms the pre-existing methods on the aspect of repeatability, efficiency and accuracy.
OCIS codes: 100.2000, 110.3010, 110.6980, 350.5730.
doi: 10.3788/COL201210.S11009.
Image splicing refers to stitching a series of partially over-
lapped images and blending them to create a large, seam-
less and high-resolution one. The automatic construction
of super viewing field ima ge from a disordered image se-
quence invo lves two crucial steps: sequence sorting and
image stitching. The present splicing algorithms usu-
ally need artificial intervention or have restrictions on
the source images
[1,2]
.
Image stitching methods fall into the fo llowing cat-
egories: graph-theoretic methods , frequency domain
methods , and feature-based methods
[3]
. However, each
of them has its own characteristic and limitation. A va-
riety of local invariant descriptors recently have made
remarkable progresses. Scale-invariant feature transform
(SIFT)
[4]
is invariant to image scaling and rotation and
partially invariant to illumination and viewpoint changes.
However, SIFT has a huge computational burden and
cannot meet the real-time requirements in some appli-
cations. In 2006, speeded-up robust features (SURF)
algorithm
[5]
is proposed, which approximates or even out-
performs SIFT with res pect to repeatability, distinctive-
ness, and robustness, and yet can be evaluated at a very
low computational co st.
In this letter, an automatic super viewing field image
splicing algorithm based on phase correlation and SURF
operator is presented, which improves the efficiency of
features detecting and accuracy of features matching ob-
viously and obtains a satisfactory super viewing field
splicing image without any manual interference.
The automatic sorting scheme is put forward in fre-
quency domain based on phase correlation technique
[6]
which has high accur acy at the image translation and the
correla tio n output shows a highly peaked shape. Phase
correla tio n is based on the Fourier shift pr operty, which
states that a shift in the coordinate frames of two func-
tions is transformed in the Fourier domain as linear phase
differences. Let f
1
(x, y) and f
2
(x, y) be the two images
that differ only by a displace ment (∆x, ∆y) i.e.,
f
1
(x, y) = f
2
(x − ∆x, y − ∆y). (1)
An unordered sequence with N images would be sorted
automatically by the following steps:
Step 1: compute the normalize d cross-power spectrum
for each image with another, and then work out the prin-
ciple peak value a nd translatio n parameters.
Step 2: pick out the image pair with the maximum cor-
relation degree and they can be identified as the adjacent
images in the sequence. According to the sign of horiz on-
tal displace ment (∆x) to rank the neighboring images (f
1
and f
2
is supposed as Eq. (1)):
a) ∆x > 0, f
1
is on the left of f
2
;
b) ∆x < 0, f
1
is on the right of f
2
.
In addition, if ∆x is over half of the image’s width, it
should be subtra c ted from the image width to obtain the
real displacement. Up to now, N ranked image pairs are
obtained and partly of them might be same.
Step 3: make a serial connection among the N ranking
results:
a) The N image pairs can be chained with each other,
then they join tog e ther to form the final order.
b) Not all of them can be chained. Examine all the
possible per mutations at the disconnected joint and con-
sider the principle peak value to find the true joint. Then
connect them to get the correct order.
SURF operator always extracts abundant and inten-
sive feature points in the ar ea where texture features are
mass. To avoid that, we use SURF to detect interest
points only in region of interest (ROI), and adopt the
bidirectional best bin first (BBF)
[7]
matching strategy.
Hence, the algorithm could dramatically cut down the
number of extracted features and speed up the matching
process, and certainly will greatly reduce the c omputa-
tional burden in imag e stitching.
Let F
1
=
f
1
1
, f
1
2
, · · · , f
1
M
and F
2
=
f
2
1
, f
2
2
, · · · , f
2
N
represent the SURF feature vectors extracted from the
ROIs of image 1 and image 2, respectively. f
NN
and f
SN
are used to denote the nearest neighbor and next nearest
neighbor, δ is defined as a threshold, and the bidirec-
tional matching method is as follows:
a) Forward matching: F
1
→ F
2
. For a n arbitrary fea-
ture f
1
K
from F
1
, find its f
NN
and f
SN
in F
2
. f
1
K
and
f
NN
are matching succe ssful if
f
NN
f
SN
< δ. Then re peat the
1671-7694/2012/S11009(4) S11009-1
c
2012 Chinese Optics Letters
资源评论
