IET Radar, Sonar & Navigation
Research Article
Improved method for SAR image registration
based on scale invariant feature transform
ISSN 1751-8784
Received on 23rd May 2016
Accepted on 20th June 2016
doi: 10.1049/iet-rsn.2016.0261
www.ietdl.org
Deyun Zhou
1
, Lina Zeng
1
, Junli Liang
1
, Kun Zhang
1
1
School of Electronics and Information, Chang'an Campus of Northwest Polytechnical University, Xi'an Shaanxi 710072, People's Republic of
China
E-mail: zenglina@mail.nwpu.edu.cn
Abstract: Scale invariant feature transform (SIFT) is one of the most common registration algorithms for synthetic aperture
radar (SAR) images. However, the occurrence of speckle noise and geometric distortion within SAR images usually leads to
limited effectiveness, challenging the stability of SIFT and its variants in real actual applications. In this study, significant
improvements for SAR image registration with SIFT are made, which lie mainly in two aspects. First, a scheme is developed to
enhance the description of keypoints with improved dominant orientation assignment and support region. Second, an optimised
matching method for further enhancing the matching performance is developed to reduce the mutual interference among the
keypoints with similar location and dominant orientations. Extensive experiments confirm the effectiveness of the proposed
algorithm for SAR images.
1 Introduction
Image registration is a key step for synthetic aperture radar (SAR)
applications, such as navigation, change detection, and hitting
effect evaluation [1–3]. In SAR image registration, feature-based
algorithms present great advantages over those based on intensity
or area feature [4–6]. Among them, scale invariant feature
transform (SIFT) [7] is one of the most widely applicable
algorithms owing to its high effectiveness in describing local
features of SAR images.
However, in application of SAR image registration, some
important characteristics of SIFT are degenerated, due to the fact
that SIFT is easily influenced by the speckle noise and geometric
distortion, which results from the scattering structure of radar with
more complex mechanism. In current literature, many
modifications of SIFT have been developed to improve its
performance in SAR image registration. Wang et al. [8] adopted a
novel bilateral filter SIFT (BF-SIFT) method with the scale space
built by a bilateral filter. It enhanced the applicability to SAR
images with different aspects and wavebands. Wang et al. [9]
proposed the adapted anisotropic Gaussian SIFT (AAG-SIFT)
matching strategy for SAR image registration by applying an
adapted anisotropic Gaussian filter, which was proved to be robust
to noise and able to retain more details. Dellinger et al. [10]
proposed a SIFT-like algorithm for SAR image registration,
namely SAR-SIFT. In this algorithm, a multi-scale SAR-Harris
matrix based on the multi-scale Harris detector was used to offer
stable keypoints in SAR images, and thus robust gradient
orientations calculation by gradient by ratio was applied to obtain
more distinctive feature descriptor. Wang et al. [11] further
improved the SAR-SIFT algorithm with uniformly distributed
features on the basis of a Voronoi diagram to select stable
keypoints and extract optimal features. In the aforementioned
algorithms, the improvements lie mainly in two aspects. First, more
adaptive scale spaces were put into use to eliminate the instability
of Gaussian scale space existing in the original SIFT algorithm.
Second, more robust features were extracted with the reliable
orientation and description for keypoints. Although the above
extensions of SIFT algorithm were improved considerably with
more robust and reliable performance, there is still some room for
improvement. Due to the interference of speckle noise and the low-
contrast characteristics of SAR images, it was still unstable in the
dominant orientations (DOs) assignment. As a result, lots of
keypoints were invalidly described in the SAR image registration.
The phenomenon was more serious when the SAR image pairs are
imaged with an angle change. Though Fan et al. [12] and Suri et al.
[13] suggested suppressing the orientation assignment for accuracy
and efficiency, the improvement is limited. Hence, the correct
matching rate is very low in SAR image registration. On the other
hand, the mutual interference among keypoints reduces the
matching performance with the current matching methods.
In this paper, an enhanced SAR image registration strategy on
the basis of SIFT is introduced with two improvements. First, an
improved orientation assignment and descriptor calculation to
achieve robust feature descriptors are adopted for SAR images.
Second, the matching method for eliminating the mutual
interference among the keypoints with similar location and DOs is
optimised by calculating the spatial relationship. Experiments
demonstrate that the proposed strategy can reduce the interference
of speckle noise and geometry distortions for SAR image
matching. The main contributions of this paper can be summarised
as follows.
i. This work can increase the correct matches with the same
quantities of keypoints and has the invariance to scale and
rotation, which is meaningful for more matches required
especially in the area with fewer keypoints detected.
ii. This work can eliminate the mutual interference among the
keypoints with similar location and DOs in the matching step,
which further reduces leakage matches.
iii. The proposed methods can be seamlessly embedded into the
current SIFT-like registration algorithms, ensuring a more
generalised performance.
2 Effective feature extraction and eliminating
interference nearest neighbour distance ratio (EI-
NNDR) matching method
Conventional SIFT algorithm usually consists of the following core
stages. The scale space is built to identify the extrema as potential
keypoints. At each candidate location, a detailed model is fit to
determine location and scale, and then the keypoints are selected
on the measures of their stability. Afterwards, a series of operations
are done on the keypoints. First, one or more orientations are
assigned. Then one single scale-dependent support region around
each keypoint is used for descriptor construction. The stability of
DOs determines the robustness of feature description, which affects
the matching results. The correct matching rate is low by the SIFT
algorithm. Only a few detected keypoints can obtain correct
IET Radar Sonar Navig.
© The Institution of Engineering and Technology 2016
1