IEEE Geoscience and Remote Sensing Magazine - December 2021 - 131

Gong et al. employed the robustness of the SIFT and the accuracy
of MI, proposing a novel coarse-to-fine registration
framework aimed at registering optical and SAR remote sensing
images [90]. For multisensor SAR image registration, Suri
et al. proposed a multistage registration strategy. The rough
parameters of the transformation model are estimated by MI,
and this model is introduced during the SIFT matching phase
to increase the number of tie points [175]. Under the SIFT
and MI combinations, Heo et al. introduced a stereo matching
method that produces accurate depth maps [176]. All
these approaches can be considered coarse-to-fine-processing
chains. The basic idea is to improve the result of the featurebased
approach by adopting an optimization process from an
area-based technique [90], [171].
The combined methods integrate the robustness of the
feature-based algorithm with the accuracy of the area-based
approach. They are relatively few compared with individual
methods, but their combination will be the focus in the near
future, from our point of view. To deal with the possible accumulation
of errors, bundle block adjustment is usually needed
[178], [179] to register sequential images. Moreover, the
integration of different geometric feature-based approaches is
being developed, as well, for ever-increasing transformation
model estimation accuracy, generating precise registration results
to the greatest extent possible.
INTEGRATION OF TWO GEOMETRIC
FEATURE-BASED APPROACHES
In addition to combinations of feature- and area-based techniques,
the integration of two geometric feature-based approaches
is a developing trend for high-precision registration.
In particular, the feature points extracted by different methods
are used to register images in two stages. Yu et al. proposed
to extract feature points using the SIFT for the preregistration
of Satellite Pour l'Observation de la Terre-5/Thematic Mapper/
Quickbird images from different sensors [74]. In the fine registration
stage, the Harris algorithm for corner point detection is
enforced to detect the distinct corner, and the extracted point
is matched by the NCC algorithm. Similarly, Lee used SURF
to extract the feature point of a low-resolution image after
Harr wavelet transformation, which is defined as rough registration
[180]. Fine registration is the same as the approach
proposed by Yu et al. Recently, Ye et al. utilized SR-SIFT to
extract the feature point in the preregistration stage for distinct
translation, rotation, and scale difference elimination.
To further optimize registration, the Harris algorithm was
employed to detect feature points in the reference and prealigned
images and describe them by LSS for matching [36].
To register large, high-resolution remote sensing images, a
coarse-to-fine strategy combining the Harris-Laplace detector
with the SIFT descriptor has been proposed. After rough
registration, a large image is divided into small, processable
blocks for fine alignment [181]. Additionally, in a new twostep
registration, the approximate spatial relationship is calculated
with the deep features using a convolutional NN in the
first step. Then, the previous result is adjusted based on the
DECEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
extracted local features [182]. Another technique combines
feature point and feature line methods for the registration
of images covering low-texture scenes in the computer vision
field [183]. Since low- and repeated-texture regions are
common in remote sensing images, feature lines can be employed
to supplement the number of feature points. Therefore,
beside the combination of two geometric feature-based
methods, the integration of different geometrical features has
great potential for the high-precision alignment of remote
sensing images [22].
Since combination schemes integrate the advantages of
two or more registration approaches, they offer remarkable
precision. Moreover, in general, preregistration provides a
rough result that approximates the final alignment. With finetuning
in the optimized registration stage, a high-precision
registration result is finally acquired. This algorithm is suitable
for remote sensing image registration with large spatial
position differences. It is as time-consuming as two or more
alignment strategies.
SOFTWARE-BASED REGISTRATION
Most reviews emphasize the ever-increasing number of image
registration approaches that are improved on the basis
of existing methods for registering larger and more complicated
images [16], [184]. Few studies have evaluated the
performance of software-embedded image registration modules
and the packages/tools for image geometric registration
[185]. Thus, in this section, we present some examples.
The Earth Resources Data Analysis System (ERDAS), ENVI,
PCI Geomatica, ER Mapper, and Arc Geographic Information
System (GIS) are well-known software packages for remote
sensing image processing that include registration modules.
ER Mapper was acquired by ERDAS a few years ago. They
integrate conventional manual and automatic registration
programs. Concretely, ENVI could register two remote sensing
images or align one image with a map covering the same
scene. A user can extract tie points by observing similar objects
lying on two images, such as corners of buildings, road intersections,
inflection points of rivers, and so on. With a uniform
point distribution, the parameters of a specified transformation
model can be estimated. There are some general geometric
mapping functions, including affine, polynomial, and
triangulation transformation models. Geometric mapping is
generally conducted by an expert and is time-consuming and
tedious. It is difficult to avoid subjective factors while extracting
tie points, especially when registering WFV images that
require more time than general image registration. To liberate
the productive forces and improve the registration efficiency,
the automatic alignment technique is also put into ENVI.
We should point out the reference and sensed images, respectively.
After setting the area-based matching parameters,
the tie point for transformation model construction is automatically
extracted; soon, the aligned image is obtained. Neither
the manually extracted tie point nor the automatically acquired
point in ENVI is sufficiently accurate. For example, the
coordinates of the extracted feature point are (. ,),
157 05 171
131
IEEE Geoscience and Remote Sensing Magazine - December 2021

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