IEEE Geoscience and Remote Sensing Magazine - March 2021 - 160

TABLE 15. COMPARISONS OF DIFFERENT LOCAL DESCRIPTORS
UNDER THE 10-CLASS SCENARIO.
SIFT

GLOH

AB-SIFT

DAISY

SAR-SIFT

DoG

0.8958

0.9075

0.9374

0.9243

0.9573

FAST

0.9015

0.9235

0.9403

0.952

0.9737

significantly different. The correlation filters and keypointbased local descriptors are robust toward spatial translation. The fluctuation in recognition rate is much slighter
for OTSDF, MINACE, NSR, and BoW. This is because the
correlation filters reach the inference according to the correlation response of Fourier-transformed coefficients, while
the Fourier-transformed coefficients are robust to shift
and distortion.
On the contrary, the performances for the remaining references are degraded sharply. Compared to the well-aligned
images, the recognition rates for SVM, SRC, and MSRC are
decreased more than 50% for the 16-pixel-translation images compared to 1.12% for NSR. These results prove that
keypoint-based local descriptors hold great potential for the
real-world scenario of target recognition.

rate for SAR-SIFT is 6.15% and 7.22% better than that for
the prototype of SIFT. This is due to the introduction of a
coherent imaging mechanism during the generation of
local features. Meanwhile, there are some issues worthy
of consideration.
◗◗ As widely reported, the approximation of the image gradient by differencing the adjacent pixels is not effective
in multiplicative noise environments. The refined version specific to the SAR image is, therefore, needed for
the mission of target recognition.
◗◗ The ROA has proven to be successful in edge detection
from a SAR image, yet how to adapt it to local feature
generation should be studied further. Although the
logarithm of the ROA performs well in the preceding
works, the operation is invalid for zero values. It is,
therefore, necessary to explore various alternatives according to the task at hand.

SEA-AREA TARGET RECOGNITION
To further verify the performance of the local feature we
pursue some additional experiments on ship target recognition. They are performed on the OpenSARShip data set,
a publicly released ship album [125] that contains 11,346
chips from 41 Sentinel-1 SAR scene images. Two interferoTEN-CLASS SCENARIO
metric, wide-swath modes are available: single-look comA more challenging task, 10-target recognition, is considered
plex and ground-range detected (GRD). The prototype data
in this section. All of the 10 targets available in the MSTAR
set includes 17 kinds of ships, yet they are sharply imbaldata set are then used to verify the typical local features. The
anced. This article employs found types of targets, cargo,
ground truth photo of the targets and the number of aspect
tankers, dredging, and others taken by the GRD mode. The
view images are shown in Figure 15. Again, 17° depression
resolutions are 20 × 22 m for the GRD imaging mode. Some
target images (the standard) are used to train a 10-target
exemplars are displayed in Figure 32.
classifier by which 15° target images (the variants) are clasIn the OpenSARShip data set, the images differ in size,
sified. The experimental results are shown in Table 15.
ranging from 17 × 17 pixels to 450 × 450 pixels; the baseThe results are similar to those of the previous experilines presented previously are, therefore, not applicable.
ments. The performance obtained using SAR-SIFT is alWe then compare keypoint-based local descriptors with
ways much better than that of the remaining techniques,
the holistic features presented in the prototype study [125].
regardless of the keypoint detection trick. The recognition
The competitors include GeF; Hu moment statistics (HuM);
scattering center features (SCF); principal
axis (PA); PA of the bow, middle, and stern
of the ship (PA-3); local radar cross section
TABLE 16. A COMPARISON OF HOLISTIC FEATURES AND KEYPOINT-BASED LO(RCS); and local RCS of the bow, middle, and
CAL DESCRIPTORS ON RECOGNITION PERFORMANCE.
stern of the ship (RCS-3).
GeF
HuM
SCF
PA
PA-3
RCS
RCS-3
SIFT
SAR-SIFT
The recognition performance is shown in
0.7926 0.8012
0.7649
0.7419
0.7317
0.7518
0.7361
0.7554
0.7521
Table 16, where FVs are used to achieve target recognition. The overall performances are

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

FIGURE 32. Exemplars of (a) cargo, (b) cargo, (c) dredging, (d) fishing, (e) fishing, (f) other, (g) other, and (h) tank. The sets of images are of
(a) 105 × 105, (b) 117 × 117, (c) 45 × 45, (d) 49 × 49, (e) 37 × 37, (f) 101 × 101, (g) 49 × 49, and (h) 57 × 57 pixels in size.

160

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

MARCH 2021
IEEE Geoscience and Remote Sensing Magazine - March 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2021
