1646 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 13, NO. 11, NOVEMBER 2016
A Nonlocal Means for Speckle Reduction of SAR
Image With Multiscale-Fusion-Based
Steerable Kernel Function
Jie Wu, Member, IEEE,FangLiu,Senior Member, IEEE, Hongxia Hao, Member, IEEE, Lingling Li,
Licheng Jiao, Senior Member, IEEE, and Xiangrong Zhang, Member, IEEE
Abstract—For the robustness of a patch-based metric, the non-
local means method is widely applied for speckle reduction of
synthetic aperture radar (SAR) images, where the similarity com-
puted by the patch-based metric is used as weight, and weighted
averaging is used to obtain the true value. However, not knowing
the local spatial property, a fixed kernel (e.g., Gaussian kernel or
uniform kernel) is always used to compute the weight. This is not
good for the preservation of geometrical features (e.g., edges, lines,
and points). In this letter, considering the characteristics of SAR
imagery, a multiscale-fusion-based steerable kernel function was
formed to explore the local spatial property of SAR images. In
addition, by combining the kernel function with a ratio-based sim-
ilarity metric designed with the distribution of the speckle’s ratio,
a new patch-based metric was formed and used with the nonlocal
scheme for speckle reduction. In the experiments, by comparing
with two state-of-the-art methods, a reasonable performance was
obtained by our method, in terms of speckle reduction and detail
preservation.
Index Terms—Multiscale fusion, nonlocal means (NLM),
patch-based similarity, speckle reduction, steerable kernel func-
tion (StKF).
Manuscript received January 20, 2016; revised June 11, 2016; accepted
July 18, 2016. Date of publication September 5, 2016; date of current version
October 12, 2016. This work was partly supported by the National Basic
Research Program of China (973 Program) under Grant 2013CB329402; by the
National Natural Science Foundation of China under Grants 61573267,
61571342, 61572383, and 61501286; by the Program for Cheung Kong Schol-
ars and Innovative Research Team in University under Grant IRT_15R53; by
the Fund for Foreign Scholars in University Research and Teaching Programs
(the 111 Project) under Grant B07048; by the Major Research Plan of the
National Natural Science Foundation of China under Grants 91438201 and
91438103; and by the Fundamental Research Funds for the Central Universities
under Grant GK201603083.
J. Wu is with the School of Computer Science and Technology, Xidian Uni-
versity, Xi’an 710126, China; with the School of Computer Science, Shaanxi
Normal University, Xi’an 710062, China; and also with the Key Laboratory
of Intelligent Perception and Image Understanding of Ministry of Education,
International Research Center for Intelligent Perception and Computation, Joint
International Research Laboratory of Intelligent Perception and Computation,
Xidian University, Xi’an 710071, China (e-mail: 607wujie2005@163.com).
F. Liu and H. Hao are with the School of Computer Science and Technology,
Xidian University, Xi’an 710126, China, and also with the Key Laboratory
of Intelligent Perception and Image Understanding of Ministry of Education,
International Research Center for Intelligent Perception and Computation,
Joint International Research Laboratory of Intelligent Perception and Com-
putation, Xidian University, Xi’an 710071, China (e-mail: F63liu@163.com;
chilamhaohao@163.com).
L. Li, L. Jiao, and X. Zhang are with the Key Laboratory of Intelligent
Perception and Image Understanding of Ministry of Education, International
Research Center for Intelligent Perception and Computation, Joint Interna-
tional Research Laboratory of Intelligent Perception and Computation, Xidian
University, Xi’an 710071, China (e-mail: linglingxidian@gmail.com; lchjiao@
mail.xidian.edu.cn; xrzhang@mail.xidian.edu.cn).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LGRS.2016.2600558
I. INTRODUCTION
O
WING to the operation at any time and under any
weather, synthetic aperture radar (SAR) has become an ir-
replaceable observation tool in many fields, such as forest mon-
itoring, town planning, and disaster assessment. However, due
to the coherent property of the SAR system, granular speckle
is always inherited in a SAR image, which makes the visual
interpretation of a SAR image very difficult. Thus, speckle
reduction is an important phase for the subsequent processing
of a SAR image. For a good speckle reduction method, not only
should the speckle be greatly suppressed but also should the
resolution of the details (e.g., edges, lines, and points) be well
preserved in its results [1].
In the past decades, several well-known adaptive filtering
methods have been proposed for the speckle reduction of SAR
images. In the image domain, a local stationary assumption is
always used within a sliding window for adaptive despeckling,
such as in [2]–[4]. Moreover, given the ability of the wavelet on
the analysis of the nonstationary case, many filtering methods are
designed in the transform domain for speckle reduction [1], [5].
Recently, with the flourish of the nonlocal means (NLM)
method in image filtering, many NLM-based speckle reduction
methods have been proposed [6]–[8]. In [9], for full use of
the spatial constraints in a neighborhood, a block-wised NLM
method was proposed for speckle reduction. Moreover, by
using the patch-based grouping and the collaborated filtering,
a SAR Block Matching and 3-D filter (SAR-BM3D) [10] was
designed for speckle reduction. In addition, the mechanism and
the parameters of a classical NLM filter (e.g., smoothing factor,
sizes of the patch, and the search window) are being studied
[11], [12]. In [13] and [14], by fusing the results obtained
with the patches of different shapes, the true signal was better
restored from additive noise. In [15], a unified nonlocal frame-
work was proposed for the auto-despeckling of SAR images,
where the results obtained with different parameter settings
were adaptively fused for speckle reduction.
From the above, we find that the patch-based metric plays
a central role in NLM methods, and the support of the patch
should be adaptively adjusted for a good preservation of the
details’ resolution. In fact, the determination of a patch’s
support can be considered as the exploration of the spatial
correlation between the neighbor pixel and the central pixel
in the patch. It is well known that, near the edges and lines,
a stronger correlation exists along the local orientation rather
than across it, whereas the correlation is of a rapid decline
away from the center of the point-like feature. Hence, local
1545-598X © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.