IEEE Geoscience and Remote Sensing Magazine - March 2021 - 161

TABLE 17. SUMMARY OF THE TYPICAL LOCAL DESCRIPTORS.
LOCAL
DESCRIPTORS GRID

SECTORS

DIMENSIONS

SCALE
SPACE

FEATURE

INVARIANCE

GRADIENT

ENVIRONMENT TYPE

PERFORMANCE

SIFT

Square 4 × 4

128

Gaussian
pyramid

Histograms Scale/rotation Differencing Additive

Sparse Moderate

GLOH

Circle

8+8+1

136

Gaussian
pyramid

Histograms Scale/rotation Differencing Additive

Sparse Moderate

AB-SIFT

Circle

Adaptive

Gaussian
pyramid

Histograms Scale/rotation Differencing Additive

Sparse Moderate

SURF

Square 4 × 4

Box filters Haar
wavelet

DAISY

Circle

200

Orientation Rotation
maps

Convolution Both

Dense Better

HOG

Square -

Histograms ×

Differencing Additive

Dense Poorest

WLD

Square -

Binary
pattern

Dense

Additive

Dense Poorest

SAR-SIFT

Square 4 × 4

128

Gaussian
pyramid

Histograms Scale/rotation Log-ratio

Multiplicative

Sparse Best

SAR-HOG

Square -

Histograms ×

Log-ratio

Multiplicative

Dense Moderate

SAR-WLD

Square -

Binary
pattern

Diff-ratio

Multiplicative

Dense Moderate

8+8+8
+1

poorer than in the previous experiments due to the lower
resolutions and the rough image quality. As can be observed
from Figure 32, the signatures of the ships are changed drastically. The radar shadows are even lost because the resolutions
along the range and the cross range are so poor. The best performance obtained using the holistic features is 0.7649 for
GeF. It is still 2.77% and 3.63% poorer than for keypointbased local features with SIFT and SAR-SIFT, respectively.
The recognition rate for SAR-SIFT is slightly better than for
SIFT, which conforms to the previous experiments.
SUMMARY
We end this section with a summary table, in which the
typical local descriptors are compared from several perspectives. The details can be found Table 17.

Integral
image

Additive

Sparse Poorest

TABLE 18. TYPICAL METRICS TO MEASURE THE
SIMILARITY BETWEEN A PAIR OF HISTOGRAMS.
Correlation

Chi-Square
Intersection
Bhattacharyya

d (H 1, H 2) =

/ i (H 1 (i ) - Hr 1) (H 2 (i ) - Hr 2) a
/ i (H 1 (i ) - Hr 1) 2 / j (H 2 ( j ) - Hr 2) 2

d (H 1, H 2) = / i

(H 1 (i) - H 2 (i)) 2
H 1 (i) + H 2 (i)

d (H 1, H 2) = / i min (H 1 (i), H 2 (i))
d (H 1, H 2) =

1
/ H 1 (i) H 2 (i)
r 1H
r 2 N2 i
H

r j = 1 / H i ( j)
H
N i

◗◗ A keypoint-based local descriptor provides great poten-

CONCLUSION
This article reports a keypoint-based local descriptor for target recognition in SAR images. Two schemes are proposed
to achieve target classification. The representative keypoint
detectors and local descriptors are verified and compared.
In addition, the results have benchmarked evaluation protocols for subsequent research. Multiple comparative studies are performed to compare keypoint-based local descriptors with global features. From the experimental results, we
can reach some conclusions.
◗◗ A keypoint-based local descriptor could be fully tuned
to target recognition in a SAR image.
◗◗ It is important to configure the related factors appropriately.
◗◗ The proposed strategy could achieve comparable or even
better performance than state-of-the-art methods.
MARCH 2021

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

tial for target recognition under the EOCs.
◗◗ The refined image gradient created by considering the
coherent imaging mechanism could achieve significant
improvement in recognition performance.
However, some advanced studies should be pursued further. This article builds sparse signal modeling via a light
dictionary. It is composed of the local descriptors close to
the query. The remaining local features resulting from the
training images are ignored. The proposed trick is, therefore, computationally attractive and memory efficient. A
simple linear cross-correlation response is employed to
measure the (dis)similarity between the local features of
the probe and the gallery. An important research direction in the future is how to define a more suitable metric
to quantify the similarity. The related study on the measurement of similarity has been noted in [83] and [126] and
161

IEEE Geoscience and Remote Sensing Magazine - March 2021

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