IEEE Geoscience and Remote Sensing Magazine - March 2021 - 155

In summary, the performance obtained using GLOH
is always better than that obtained with SIFT and slightly
poorer than that obtained with AB-SIFT, because GLOH
and AB-SIFT develop a more refined division of location
via log-polar sectors. DAISY usually produces the best performance. The combination of the FAST keypoint detector
and DAISY local descriptor is the preferable choice in terms
of recognition performance.
CLASSIFICATION
This article presents two schemes to achieve target recognition
via keypoint features. We study the related factors on recognition performance.
SPARSE REPRESENTATION VIA A LEARNED DICTIONARY
Our first solution builds multiple linear regression models
via a light dictionary in which only the Lth most related
local features are chosen. We then study the recognition

Recognition Accuracy

into 4 × 4 square grids and the gradient angle into eight orientations, resulting in a 128D descriptor. GLOH assigns a
log-polar location grid with three bins in the radial direction
and eight bins in the angular direction, while the central
circle is viewed as a single bin.
The gradient orientations are further quantized into
eight bins, generating a 136D feature. SURF is built by
four Haar-like wavelet responses that are extracted in 4 × 4
square grids, generating a 64D representation. The recognition performance is highly relevant to the keypoint detector. For the DoG detector, the performance obtained using
DAISY is poorer than that obtained with SIFT and GLOH.
For the remaining detectors, DAISY always achieves the best
performance. It produces a gain greater than 3-8% over that
of SIFT and GLOH. The recognition accuracy for SURF is
consistently much poorer. It is 11.63%, 12.40%, and 15.81%
lower than SIFT, GLOH, and DAISY, respectively, for the
Hessian detector. This result can be attributed to the fashion
of keypoint detection and feature representation.
To boost computational efficiency, SURF circumvents
image convolutions by means of integral images. Figure 22
shows a pair of detection maps generated by " FAST-Hessian " and Hessian detectors. The keypoints detected by
" FAST-Hessian " are irregularly scattered in the whole image
and, hence, contain less discriminative information. Contrarily, the interest points produced by the Hessian detector
are mainly concentrated in the target imaging region. These
points reflect the local scattering center and, hence, have
powerful discrimination ability. Moreover, SURF represents the local pattern by a 64D Haar-like wavelet response,
h x, h y, ; h x ;, ; h y ;, while the dimensions of SIFT, GLOH, and
DAISY are 128D, 136D, and 200D. The experimental results
also prove that the approximation of convolution with an
integral image is ineffective against the statistical specificity
of multiplicative noise.

0.92
0.9
0.88
0.86
0.84
0.82
0.8
0.78
0.76
0.74
0.72

Harris−
Laplace

DoG
SURF
AB-SIFT

Hessian Hessian−
Detectors Laplace
SIFT
DAISY

FAST

GLOH

FIGURE 21. The recognition performance obtained using various

descriptors.

3
4

C1 C2

C4
1

Radius
6

8
7

(a)

(b)

FIGURE 22. A comparison of (a) " FAST-Hessian " and (b) Hessian
FIGURE 20. An illustration of DAISY; C 1, C 2, C 3, and C 4 denote the

inner circle and three outer circular regions. The overlapping skill
smooths the transitions between regions, achieving a degree of
rotational robustness.
MARCH 2021

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

operators on a detection map. The interest points detected do not
reflect the target scattering signatures and, hence, have little representation power. The conclusion can be obtained from Figure 21, in
which SURF always produces the poorest performance.
155

IEEE Geoscience and Remote Sensing Magazine - March 2021

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