September 2019 Full - page 680

terrain; however, there are two difficulties in an aerial im-
age task (Song et al. 2019). One is the multi-temporal image
matching, meaning that radiation information of the pixel can
unexpectedly change greatly due to weather and sun condi-
tions. For example, one image is taken in the daytime and the
other at night. Registration of multi-temporal images is gener-
ally considered difficult because of the scene variation; that
is, illumination variation can occur when images are obtained
at different times (Zhao and Goshtasby 2016). The other dif-
ficulty of point matching in aerial images results from the
viewpoint variation. An unmanned aerial vehicle (UAV) could
always rotate during a flight, so an on-board camera may not
always fly parallel to the ground; that is, viewpoint variation
occurs between adjacent images. The registration could be
more challenging when the images are taken in urban areas
where building exhibit different appearances according to the
angle of the camera. The learned descriptors PPD and TFeat
are used to evaluate the image matching performance between
images under viewpoint and illumination variation along
with the classical handcrafted descriptors SIFT and BRISK.
Close range images: i_leuven and v_graffiti, two image pairs
under illumination and viewpoint change, from the HPatches
data set are used as the experimental data set. Table 6 shows
basic information, including the change type of the two image
pairs and the number of interest points extracted from the two
image pairs using BRISK and SIFT detector.
The number of correspondences after outlier removal using
RANSAC for different combinations of detectors and de-
scriptors is shown in Figure 11. For the image pair v_graffiti
with rotation variation, the learned deep descriptors, includ-
ing TFeat and the proposed PPD, obviously outperform the
traditional handcrafted descriptors, such as SIFT and BRISK.
Figure 8. Feature point matching results of image pair
Pair1 after outliers removal using RANSAC through different
combinations of different detectors and descriptors. The
first row of images use BRISK detector to detect interest
points and (a) BRISK, (b) TFeat, and (c) the proposed PPD to
obtain the descriptors. Analogously, the second to the fifth
rows show the correspondences using the SIFT, ORB, SURF ,
and AKAZE detection algorithms.
Figure 9. Feature point matching results of image pair
Pair2 after outlier removal using RANSAC through different
combinations of different detectors and descriptors.
Figure 10. Feature point matching results of image pair
Pair3 after outlier removal using RANSAC through different
combinations of different detectors and descriptors.
Table 5. The number of correspondences of the three image
pairs using different combinations of detector and descriptors.
PPD is the proposed learned deep descriptor. Bold numbers
indicate that the corresponding descriptors have the top
performance.
Combination Pair1 Pair2 Pair3
BRISK 2658 2047 10 361
BRISK+TFeat 2717 2366 4081
BRISK+PPD 5112 4081 12 477
ORB 3017 1690 12 236
ORB+TFeat 3095 1978 11 301
ORB+PPD 4031 2266 13 819
AKAZE 2030 1727 4432
AKAZE+TFeat 1952 1605 5080
AKAZE+PPD 3051 2522 5856
SIFT 1296 1166 3749
SIFT+TFeat 1493 1205 4377
SIFT+PPD 2274 1943 5721
SURF 1223 1205 4999
SURF+TFeat 734 745 3984
SURF+PPD 1688 1450 5471
680 September 2019 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING