IEEE Geoscience and Remote Sensing Magazine - December 2021 - 129

FEATURE REGION
Feature region is a general term for all closed boundary regions
of appropriate size, e.g., lakes [125], forests [126], buildings
[113], urban areas [127], and so on. Before the robust feature
point extraction approach was developed, the feature region
was used to indirectly extract feature points. Regions with high
contrast were extracted by filtering [128] and image segmentation
[129] and described with moment-invariant descriptors
[130], [131]. They are often abstracted by their centers of
gravity [128], [132]-[135], which are invariant with respect to
rotation, scaling, and skewing and are stable under random
noise and gray-level variation [16]. Compared with feature
points and lines, the extraction and description of feature regions
were relatively early foci of research, and they have been
used less for recent feature-based registration.
FEATURE MATCHING AND MISMATCHED
FEATURE ELIMINATION
The correspondence relationship between reference and
sensed images can be established based on detected feature
points, lines, and regions, exploiting various descriptors of
features [16], [136], [137]. Mismatched features are an inevitable
byproduct of general feature matching, the elimination
of which purifies correspondences for generating transformation
models that are as accurate as possible. A pair of features
with similar attributes is considered a selectable matching despite
radiometric differences, noise, image distortion, and so
forth. Under the circumstances, a robust matching measurement
is essential. Feature matching approaches can be generally
classified into two categories, namely, feature similarity
and spatial relations.
FEATURE SIMILARITY
The constructed feature descriptors are used to establish the
correspondence between extracted features in the reference
and sensed images through feature similarity comparison.
Feature similarity is conducted in the feature space by using
the Euclidean distance ratio between the first and second
nearest neighbors [92]. For efficiency, the k-dimensional tree
and the best-bin-first algorithms are employed for feature
similarity determination [93], [138]. The clustering technique
[140], chamfer matching [141], and PC models are frequently
used matching approaches, and they are invariant under intensity
changes during matching [1].
SPATIAL RELATIONS
Aimed at tie point matching in poor textural regions, approaches
based on spatial relations have been developed.
Representative of these, graph-based feature points matching
considers feature points as graph nodes. Feature matching is
then transformed into a node-correspondence problem and
solved by graph matching [125], [142]. Graph matching is
applied to image feature correspondences, although it is not
affine invariant [143]. By finding a consensus nearest-neighbor
graph from candidate matches, a graph-transformation
matching approach is developed [144]. Targeting the problem
DECEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
in [143], a similar graph matching for tie point matching in
poor textural images is proposed [101]. Furthermore, Xiong
and Zhang introduced a novel interest point matching for
high-resolution satellite images [145]. For this, the relative
position and angle are used to reduce ambiguity and to avoid
false matching, as the approach is suitable for image shifting
and rotation. Affine and large-scale transformations are not
considered [144].
MISMATCHED FEATURE ELIMINATION
Although the extracted features in a reference image have
been matched with the corresponding ones in the sensed
image via the aforementioned approach, some mismatched
feature points are inevitable, further affecting the transformation
model estimation [32], [76]. Therefore, eliminating
mismatched features with a specified approach is necessary
[146], [147]. Generally, based on the initial matching result,
random sample consensus (RANSAC) is used to remove a
mismatched point. This method randomly selects a sample
from the consensus set in each iteration and finds the largest
consensus set to calculate the final model parameters [33],
[148]. RANSAC performs well and robustly when there are no
more than 50% outliers [144], [149], [150]. Combining the
local structure with global information, a restricted spatial order
constraints algorithm is developed to find exact matched
feature points in reference and sensed images [144].
Based on the affine-invariance property of the triangle-area
representation (TAR), a robust sample consensus judging algorithm
is proposed to efficiently identify bad samples and
ensure accuracy with a light computational load [151]. For
images with simple patterns, large affine transformations, and
low overlapping areas, a mismatch- removal principle based
on the TAR value of the k-nearest neighbors is proposed and
referred to as k-nearest neighbors-TAR [149]. Furthermore,
an improved RANSAC approach called fast sample consensus
is developed to obtain correct matching in a few iterations
[150], [152]. Thus, most of the reserved feature points in
the reference image accurately correspond to the specified
feature points in the sensed image, as the feature points
connected by the yellow lines in Figure 12 will add precision
to the transformation model estimation in the following
step. The geometrical feature-based approach abstracts
an original remote sensing image with distinct features
instead of its intensity information, which is efficient and
can easily process large rotations, translations, and scale
differences between reference and sensed images. However,
position errors in the automatically extracted features are
inevitable, and a few mismatched features cannot be eliminated.
This leads to a relatively low registration precision
compared with the intensity-based approach.
NOVEL FEATURE-BASED REGISTRATION
BY DEEP LEARNING
Deep learning provides a new concept for remote sensing
image registration. It essentially refers to image registration
based on advanced feature extraction [153]. Deep learning
129
IEEE Geoscience and Remote Sensing Magazine - December 2021

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