The aforementioned analysis is performed using the auto-
matic evaluation method. To rigorously evaluate the matching
performance, we also manually select samples and determine
the correct pairs of matches for the evaluation, and the results of
manual evaluation is shown in Figure 11. It can be seen that un-
der manual evaluation, the proposed method still performs best
in all image pairs, with the largest improvement being more than
30 percent in
MP
(i.e., on dataset 2), demonstrating the effective-
ness of our proposed method. The correspondences matched by
the proposed matching method are displayed in Figure 12.
Conclusions and Future Work
In this paper, we propose a reliable matching method for
remote sensing images containing highly repetitive patterns.
The proposed method integrates a novel
LDF
detector and
an effective match searching strategy. First, the
LDF
detec-
tor is designed to detect distinctive features from the source
image and target image. Both the distinctiveness of pixel
and the distinctiveness of support region are evaluated in
the feature response computation, which improves matching
accuracy and robustness. The region distinctiveness factor in
the response function effectively
eliminates non-distinctive points
in the feature detection step.
Second, seed points based on
LDF
s
are selected and matched. A novel
matching reliability indicator de-
scribing the confidence of the seed
matches is proposed to find reliable
seed matches. Third, based on our
method, a
FIPS
based search strat-
egy is proposed to ensure matching
performance while improving the
time efficiency. The experimental
results in this paper demonstrate
the effectiveness of the proposed
method. Each of its components
(i.e.,
LDF
detector and the match-
ing method) performs the best
among the testing methods derived
from the state-of-the-art methods,
with over 90 percent of matching
precision and the largest improve-
ment reaches 30 percent over the
comparative methods. However,
one limitation of our method is that
it relies on the use the geo-referenc-
ing information or platform flight
parameters to compute a coarse
overlap region of the input images
at the beginning of the method.
Also, some of the parameters/
thresholds may need to be tuned if
the dataset is significantly different
than normal remote sensing images
(e.g., terrestrial images, or images
acquired under suboptimal condi-
tions such as platform stabilities).
Therefore, in our future work we
tend to use automatic referencing
methods and adaptive parameter
tuning to improve the general ap-
plicability of the proposed method.
Acknowledgments
This research was supported by the
National Natural Science Founda-
tion of China (No. 41631174, No.
41501492 and No. 41601476),
the National Key Research and
Figure 12. Matching results of the proposed method. (a) to (d) shows the matching
result on Dataset 1-4, respectively. The enlarged sub-images show that most of the
matches are located in the correct positions exactly.
Figure 11. Manual evaluation results.
522
August 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
