October 2019 Layout Flipping Full - page 735

bank. Therefore, the SSSFncc achieves more comprehensive and
complete terrain shape features compared with HOPCncc and
CSLTP in a region. Besides, why the HOPC ncc gets the minimal
tie points is that the HOPCncc removes image boundaries to
eliminate the effects of noise and illumination.
For computational efficiency, the NCC has the fastest cal-
culation. SSSFncc computes faster than HOPCncc , because SSSFncc
and HOPCncc have to integrate the descriptors first and then
compute the NCC between these descriptors, which is time
consuming. SSSFncc requires less run time compared with the
HOPCncc , mainly because the SSSFncc only calculates the shape
context centering on the feature points. This process is less
time consuming than computing the phase congruency fea-
ture applied to build the HOPC ncc. The results in Figure 13 il-
lustrate that SSSFncc is more time consuming than CSLTP mainly
because SSSFncc is required to extract scene shape information
and calculate the NCC between descriptors, which needs more
run time than calculating self-similarity used to construct
CSLTP. Overall, results illustrate that compared with the NCC,
HOPCncc and CSLTP, the SSSFncc is more robust for multisource
remote sensing registration.
Possible Errors and Model Limitations
Experimental results show that the SSSF descriptor is robust
for multisource remote sensing image registration. Although
some influencing factors, such as the template window sizes
and canny operator thresholds are discussed, some pos-
sible problems may still arise. First, the proposed method is
limited to image information and terrain structure. If the test
data contain less scene shape information, the SSSF descriptor
may achieve poor results in terms of matching accuracy. In
hilly or mountainous regions, the terrain structure informa-
tion of these areas is not obvious, because less contours and
corners will be extracted. In addition, a large surface undu-
lation causes significant local distortion. Second, the factor
most affected by seasonality is vegetation. However, changes
of vegetation usually do not affect the extraction of contours
and corners, and the terrain structure information changes
minimally. Therefore, seasonality has little effect on the
method used in this study. Third, changes in small-scale land
cover affect the extraction of contours and corners. Neverthe-
less, from a global perspective, extracting enough contours to
construct SSSF descriptors and complete image matching is
still possible. Seasonality and land cover usually have little
effect on the proposed method, and the SSSF descriptors can
always be built to match the tie points.
Discussion
This paper proposes a novel descriptor based on the shape
context algorithm, SSSF descriptor. the normalized correlation
coefficient of the SSSF descriptors (SSSFncc) as a similarity mea-
sure is applied to extract tie points with a template match-
ing strategy. The piecewise linear transform model is then
selected to rectify the remote sensing image. Compared with
three state-of-the-art methods, the proposed one achieves the
best matching accuracy because it integrates the terrain shape
feature better and is robust to local rotation.
The canny operator is used to detect scene edges to con-
struct the shape similarity in a local region. Different canny
operator thresholds (i.e., from 0.1 to 0.3) were set to ana-
lyze the sensitivities of SSSFncc with regard to the changes of
canny operator. Experimental results show that the proposed
method obtains an improved matching outcome when the
canny threshold is set to 0.2. The reason are as follows. The
image contours detected by the canny threshold of 0.1 are
too dense and plentiful to reduce the impact of the increase
or decrease and deformation of terrain. By contrast, contours
are not enough to use the canny threshold of 0.3 to extract the
terrain shape features, making it difficult to extract a sufficient
number of tie points.
Although the proposed method has successfully used in
multisource remote sensing image registration, some improve-
ments to SSSFncc should be considered further. The different
number of concentric circle layers and partitions can be set to
analyze the influence of shape context algorithm for the SSSF
descriptor, which changes the structure of the SSSF descrip-
tor. Also, an image enhancement technology could be applied
to enhance the shape and contour feature, which may be of
some help to image matching. Only the SSSF descriptor of the
template center is calculated, but SSSFncc is still a little time
consuming compared with CSLTP since SSSFncc need to sample
a number of contour points. In the future work, this issue
could be resolved via a more efficient sampling method. As
mentioned above, if the images contain less shape structure
information, the SSSFncc may achieve poorer matching accura-
cy, because the SSSFncc depends on the local shape properties
of images.
Conclusions
This study proposes a novel similarity measure named SSS-
Fncc for multisource remote sensing registration. The metric
is quite robust for multimodal remote sensing images with
significant nonlinear grayscale differences. First, the shape
context algorithm is used to extract SSSFs. Then, the scene
shape similarity information is used to construct SSSFncc, and a
template matching strategy is used to extract tie points. SSSFncc
aims to acquire the scene shape similarity between multi-
source images and has been assessed by different categories
of multisource test sets, including high-resolution, medium-
resolution, and map sets. The registration results show that
the SSSFncc is more robust than NCC and HOPCncc similarity
measures for significant nonlinear grayscale differences. In
summary, these registration results illustrate that the pro-
posed method achieves robust registration outcomes.
Acknowledgment
The work presented in this paper is supported by the Fun-
damental Research Funds for the Central Universities
(2018QNA21), and aProject Funded by the Priority Academic
Program Development of Jiangsu Higher Education Institu-
tions.
References
Abdel-Hakim, A. E. and A. A. Farag. CSIFT: A SIFT descriptor with
color invariant characteristics. Pages 1978–1983 in the 2006 IEEE
Computer Society Conference on Proceedings of the Computer
Vision and Pattern Recognition, 2006.
Belongie, S., J. Malik and J. Puzicha. 2001. Shape context: A new
descriptor for shape matching and object recognition. In
Advances in Neural Information Processing Systems 13, edited
by T. K. Leen, T. G. Dietterich and V. Tresp, 831–837.
Brunner, D., G. Lemoine and L. Bruzzone. 2010. Earthquake damage
assessment of buildings using VHR optical and SAR imagery.
IEEE Transactions on Geoscience and Remote Sensing 48
(5):2403–2420.
Chen, S. H., X. R. Li, H. Yang and L. Y. Zhao. 2018. Robust local
feature descriptor for multisource remote sensing image
registration. Journal of Applied Remote Sensing 12.
Cole-Rhodes, A. A., K. L. Johnson, J. LeMoigne and I. Zavorin. 2003.
Multiresolution registration of remote sensing imagery by
optimization of mutual information using a stochastic gradient.
IEEE Transactions on Image Processing 12 (12):1495–1511.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING October 2019 735