IEEE Geoscience and Remote Sensing Magazine - December 2021 - 128

especially for large remote sensing images. At present, many
methods have been proposed to automatically acquire representative
features. Common geometrical features, including
salient points (line intersections, corners, points on curves
with high curvature, and road crossings) [73], [74], polylines
(roads, contours, and edges) [41], [75], and polygons (closed
boundary regions and lakes) [76], are selected by the specified
approach. As shown in Figure 11, the yellow points, line
segments, and regions are detected to abstractly describe the
original image.
FEATURE POINTS
The local points at which the gray value varies dramatically in
all directions are feature points, including corner points, inflection
points, and T-intersection points. Many attempts have
been made to extract them in computer vision, inspiring the
development of feature point extraction in remote sensing. The
first corner detection approach was proposed by Moravec in
1977 [77]. This algorithm has fast computation but is sensitive
to noise and vulnerable to image rotation, leading to its rare
use in the remote sensing field. The Harris corner detector was
proposed in 1988 [78]. This algorithm is invariant under grayscale
and rotational changes. It and improved Harris algorithms
are applied to remote sensing image processing [38], [74], [79],
[80], mainly with respect to multiscale corner detection.
Smith and Brady presented the smallest unvalued segment
assimilating nucleus operator [81], which is insensitive to local
noise and has high anti-interference ability [82]. However,
it is not widely used in remote sensing image registration [83],
whereas the SIFT algorithm is [45], [58], [74], [84]-[90]. The
SIFT was developed by Lowe [92] and is invariant under rotation,
scale, and translational changes [93]. It has been followed
by many improved versions, such as principal component
analysis SIFT [94], scale-restriction (SR) SIFT [36], [95], affine
SIFT [96], and uniform robust SIFT [97], [98]. Moreover, the
speeded-up robust features (SURF) [99] algorithm was proposed,
by Bay et al. to overcome the time-consuming nature
of the SIFT for large-scale remote sensing images [100]-[102].
SURF applies an integral image to compute image derivations
and quantifies the gradient orientations in a small number of
histogram bins [103]. Additionally, the features from accelerated
segment test (FAST) [104]; binary, robust, independent
elementary features (BRIEF) [105]; oriented FAST and rotated
BRIEF [106], [107]; Kaze [108]; and accelerated Kaze [109] algorithms
are fast tools for descriptor construction but are less
widely utilized in remote sensing.
In addition, a novel key point detector combining corners
and blobs for remote sensing image registration is under development
to increase the number of correctly matched features
[110]. Recently, looking at intensity differences in multimodal
remote sensing images, robust and novel feature descriptors
have been adopted to depict detected feature points; these include
the LSS descriptor, which accommodates effects such
as nonlinear intensity differences [36]; the histogram of oriented
PC, based on structural similarity measures [57]; and
maximally stable PC, representing a novel affine and contrastinvariant
descriptor [111]. All these coincidentally absorb PC
information. PC is similar to the image gradient, presenting
structural information with resistance to variations in illumination
[112]. Therefore, the use of phase consistency information
is a trend in the construction of robust feature descriptors
for multimodal remote sensing image.
FEATURE LINES
Feature Point
Feature Line
Feature Region
Feature Matching
Transformation Model
Construction
Mismatched Features Elimination
Coordinates
Transformation/Resampling
FIGURE 11. The geometrical feature-based registration algorithm.
128
A feature line is also known as a line feature; it is the generalization
of feature points, such as general line segments [113],
object contours [75], roads, coast lines [114], and rivers [115].
Given that feature lines have more attributes than feature
points as control features [116], they
have been gradually developed for use
in image registration [117] as well as
remote sensing image registration
[116], [118], [119]. Standard edge detection,
as with the Canny detector
[120], [121] and detectors based on
the Laplacian of Gaussian [122] are
conventional feature line detection
approaches [16]. Recently, some excellent
detectors generating precise
and robust line segments have been
proposed [123], [124], and they are
suitable for line detection in remote
sensing images. Feature lines are comparatively
less utilized in the remote
sensing field than are feature points
because matching them is an obstacle.
They are often abstracted from
corners, midpoints, and endpoints
as final features [16], thereby losing
their geometric value.
Aligned
Image
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2021
IEEE Geoscience and Remote Sensing Magazine - December 2021

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