Journal of Physics: Conference Series

PAPER • OPEN ACCESS

Towards Better Understanding of SAR Image: Feature Enhancement via

Non-Local and Low-Rank Approach
To cite this article: Xintong Tan et al 2021 J. Phys.: Conf. Ser. 2025 012071

View the article online for updates and enhancements.

This content was downloaded from IP address 181.214.186.171 on 29/09/2021 at 10:42

AICS 2021 IOP Publishing
Journal of Physics: Conference Series 2025 (2021) 012071 doi:10.1088/1742-6596/2025/1/012071

Towards Better Understanding of SAR Image: Feature

Enhancement via Non-Local and Low-Rank Approach

Xintong Tan1, Qi Yu1, Zelong Wang1 and Jubo Zhu1,2

1
College of Liberal Arts and Sciences, National University of Defense Technology,
Changsha, China
2
School of Physics and Astronomy, Sun Yat-sen University, Zhuhai, China
Email: [email protected]

Abstract. Feature enhancement for synthetic aperture radar (SAR) images is of great
significance for their understanding and interpretation. In this work, we aim to address the issues
by introducing the low-rank constraint into non-local means framework, dubbed NL_LR. The
non-local means framework takes advantages of the non-local self-similarity of SAR images,
which makes this approach efficient in noise suppression and preservation of structures and
resolution. When estimating the value of the target pixel, a low-rank matrix can be constructed
with vectorization of similar image patches. By exploiting this low-rank prior of patch matrix
and decomposition of sparse and low-rank matrices, the denoised low-rank patch matrix is more
accurate which will also increase the accuracy of feature enhancement. Afterwards, the
numerical algorithm is designed. Numerical experiments on the real-data of SAR images show
that our novel method can reduce the noise in homogeneous areas especially speckle noise
efficiently, preserve the structural feature, especially edges and textures and improve the
resolution at the same time. Visually, the result of the proposed method is obviously improved.
Keywords. SAR image; feature enhancement; non-local; low-rank.

1. Introduction
Synthetic aperture radar (SAR) is a high-resolution remote sensing technique, and is widely applied in
numerous fields such as environmental monitoring, earth-resource mapping, and military systems [1].
However, due to the errors caused by the imaging system and the signal propagation, the quality of SAR
images decreases and the features degrades. Feature enhancement can highlight these features, which
makes SAR images easier and more accurate to interpret and recognize. Therefore, feature enhancement
for SAR images is of great significance for SAR image applications.
The existing methods for feature enhancement can be divided into several categories, such as the
multi-look approaches which sacrificing the resolution [2, 3], filtering methods based on local statistic
characteristics which cannot protect strong scatters or structural features well [4-12], and the multi-
resolution methods [13]. In addition, the regularization methods [14] and partial differential equation
(PDE) methods [15] are also common. However, the regularization methods are not effective in speckle
suppression and may lead to the loss of weak targets while PDE methods are insufficient in resolution
improvement and target enhancement. Buades proposed the non-local means (NL-means) and the
estimated value of target pixel is a non-local averaging of all similar pixels in the image, which is more
favorable for the textured or periodic case due to the nature of the algorithm [16]. The problem of sparse
and low-rank matrix decomposition was highlighted and intensively studied by [17] and alternating
direction methods to solve the problem were proposed in [18]. Since then, methods based on non-local
Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
AICS 2021 IOP Publishing
Journal of Physics: Conference Series 2025 (2021) 012071 doi:10.1088/1742-6596/2025/1/012071

and low-rank prior are widely used in compressive sensing and image processing. Dong proposed a non-
local low-rank regularization approach and explored its application into compressive sensing of both
photographic and MRI images [19]. Based on edge detection and neural network, Chen proposed a non-
local low-rank matrix completion method and applied the non-local inter-pixel correlation to image
interpolation effectively [20]. Liu proposed a super resolution reconstruction method for single image
based on non-local sparse and low-rank regularization [21]. Wen proposed a joint adaptive patch sparse
and group low-rank model and develop an image restoration framework based on the model [22].
In this paper, we directly implement feature enhancement for SAR images in space domain. The
novel model is established by introducing the low-rank constraint into NL-means framework, dubbed
NL_LR. Figure 1 shows the schematic diagram of the approach. Due to the non-local self-similarity,
there are many neighborhoods similar to neighborhood of the target pixel. With vectorization of these
similar image patches, the row of the patch matrix is the vector of similar image patch and the patch
matrix is low-rank. Assuming the patch matrix is noisy and the noise is sparse, decomposition of sparse
and low-rank matrices is carried out to obtain the low-rank component, namely the denoised patch
matrix. Based on the denoised patch matrix, more accurate weights and values of pixels with similar
neighborhoods will be used to estimate the value of target pixel.

Figure 1. Schematic diagram of the approach.

The paper is structured as follows: In Section 2, we give an overview of weighted NL-means method
and propose the novel SAR image feature enhancement approach based on non-local and low-rank
constraints. Numerical experiments are provided in Section 3 to illustrate the effectiveness of our
proposed method.

2. Feature Enhancement via Non-Local and Low-Rank Approach

In this section, the novel feature enhancement approach for SAR images is proposed based on non-local
and low-rank constraints. Sparse and low-rank matrix decomposition is applied on the patch matrix
constructed by the non-local similar image patches. Finally, we show the algorithm for numerical
solution.

2.1. Weighted Non-local Means

NL-means method takes the non-local averaging of all similar pixels in the image as the value of target
pixel [16]. Namely, the value of target pixel is estimated as the weighted mean of values of pixels whose
neighborhood is similar to neighborhood of the target pixel.

2
AICS 2021 IOP Publishing
Journal of Physics: Conference Series 2025 (2021) 012071 doi:10.1088/1742-6596/2025/1/012071

For a given image g , the estimated value of target pixel xi is computed as a weighted average of all
the pixels in the image
NLg ( xi ) =  w ( xi , x j ) g ( x j ) , (1)
j

( ) (
where g x j is the value of pixel x j in the given image g , w xi , x j is the weight between pixels xi )
and x j to measure the similarity between them, and satisfies the usual conditions 0  w xi , x j  1 and ( )
 w( x , x ) = 1. It shows that the feature-enhanced value at xi is a weighted mean of the values of all
j i j

pixels whose neighborhood is similar to neighborhood of xi . For the convenience of discussion, Ni is

defined as a square neighborhood with its central pixel xi and fixed size, and v (  ) refers to the
vectorization of matrices. So that v ( N i ) represents the intensity gray level vector of neighborhood of
xi . Choose the weighted Gaussian Euclidean distance, namely

d ( xi , x j ) = v ( Ni ) − v ( Ni ) 2,a
2
(2)

where a  0 is the standard deviation of the Gaussian kernel. Its properties are briefly discussed in [16].
Then, the weight can be represented as
v ( Ni ) − v ( Ni )
2

w ( xi , x j ) =
2,a
1 −
h2 (3)
e
Z (i )

where
v ( Ni ) − v ( Ni )
2
2 ,a
−
Z ( i ) = e h2

is a normalizing constant to ensure  w ( x , x ) = 1 , and h

j i j acts as a filtering parameter which controls
the decay of the exponential function. Considering the relationship between the exponential function
and the weight, h controls the decay of the weights as a function of the Euclidean distances. It shows
that the NL-means not only compare the values of the single central pixels but also the values of pixels
in the neighborhood. This means that the geometrical configuration in the whole neighborhood is also
compared to make the method more robust.

2.2. Feature Enhancement via Non-Local and Low-Rank Approach

Since the neighborhoods are similar to neighborhood of xi , the matrix formed by their vectorizations is
low-rank. P ( i ) is the patch matrix, which is a low-rank matrix consisting of the neighborhood of xi
and its similar neighborhoods
 v ( Ni ) 
 
 v ( N j1 ) 
P (i ) =   (4)
 
v ( N ) 
 jn 

3
AICS 2021 IOP Publishing
Journal of Physics: Conference Series 2025 (2021) 012071 doi:10.1088/1742-6596/2025/1/012071

Assuming the noise in P ( i ) is sparse, P ( i ) is consisted of a low-rank matrix and a sparse matrix. The
problem of sparse and low-rank matrix decomposition is discussed in [18]. As a result, P ( i ) is
recoverable for its sparse and low-rank components and they can be recovered by solving the following
convex optimization problem:
min P ' ( i )  +  ε ( i ) l1
P '( i ),ε ( i )
(5)
s.t.P ' ( i ) + ε ( i ) = P ( i )
where P ' ( i ) is the denoised low-rank component of P ( i ) , ε ( i ) is the noise of P ( i ) which is sparse,
and   0 is the regularization parameter which makes a balance between the sparse and low-rank
components,.   refers to the nuclear norm which is the sum of all singular values, and  l1 represents
the l1 norm which is the component-wise sum of absolute values of all entries. The nuclear norm
constrains the low-rank property of P ' ( i ) while the l1 norm constrains the sparsity of ε ( i ) . With less
effect of noise, the value of the target pixel estimated from P ' ( i ) will be more accurate.
To sum up, the feature enhancement method based on non-local and low-rank constraints can be
expressed as
NL _ LRg ( xi ) =  w ' ( xi , x j ) g ' ( x j )
j

min P ' ( i )  +  ε ( i ) (6)

P '( i ) , ε ( i ) l1

s.t.P ' ( i ) + ε ( i ) = P ( i )

( ) is the denoised value of pixel

where g ' x j (
x j , and w ' xi , x j ) is weight between xi and x j computed
based on the denoised patch matrix P ' ( i ) .
The algorithm can be briefly summarized as follows:
Algorithm for feature enhancement
Input g  Rmn
For 𝑖 = 1: m  n
Find the neighborhoods of xi and the similar neighborhoods
Construct P ( i )
Use LRSD for the low-rank component of P ( i ) , namely P ' ( i )
Compute the weights and estimate the value of xi based on P ' ( i )
End
Output g

3. Numerical Experiments
In this section, our novel feature enhancement approach is applied to the RADARSAT-2 data of
Vancouver in April 2008 and data of the Pentagon provided by Sandia National Laboratories, and then
compare the results with the existing methods, such as the sparse regularization method, PDE method,
and the enhanced Wiener filter method (EWF) [23] to evaluate its performance.

3.1. Evaluation of Image Features

In this subsection, four indexes are used to evaluate the performance of different approaches for SAR
image feature enhancement.

4
AICS 2021 IOP Publishing
Journal of Physics: Conference Series 2025 (2021) 012071 doi:10.1088/1742-6596/2025/1/012071

Spatial resolution is the number of pixels which is corresponding to the relative 3-dB mainlobe width.
When the value is smaller, the resolution is higher. Equivalent number of looks (ENL) is an index to
describe the degree of speckle suppression. When ENL is larger, the speckle is reduced more efficiently.
Edge preservation index (EPI) is an index to describe the effect of structure preservation. When EPI is
larger, the structures are preserved better. Target-to-clutter ratio (TCR) is an index to measure the
contrast of targets and clutters. When TCR is larger, the clutter is suppressed more efficiently. Formulas
can be seen in Chapter III of [24] and its references of further details and extensions, which will not be
repeated here.

3.2. Experimental Results and Analysis

The original SAR image dataset 1 is shown in figure 2, from which it can be seen that the speckle noise
and defocus is very heavy. Besides, the feature-enhanced results of the image dataset 1 by different
approaches are also presented in figure 2 with a zoomed-in region shown in figure 3. It shows that all
the methods enhance the features to some extent. The novel method reduces the noise sufficiently,
especially in the homogeneous areas and dark areas, but the targets in the homogeneous areas are well-
preserved and even enhanced. At the same time, the structural feature such as edges and textures are
also well-preserved. In addition, the sidelobe is suppressed and the image resolution is improved to a
certain extent. However, the sparse regularization method sharpens the target at the cost of loss of weak
targets and denoising, and the PDE method cannot preserves the resolution while suppressing the
speckle noise. The EWF method reduces the noise especially in the homogeneous areas, but causes
serious blurring. Table 1 shows the experiment indexes of the original SAR images and the feature-
enhanced ones by different approaches. From the point of view of indexes, all the indexes of our
proposed approach are much better than that of the original. Compared with the original image, the
novel approach increases the spatial resolutions by 10.16% in azimuth and 30.44% in range, ENL by
3.32, EPI by 12.00% and TCR by 0.59dB. It is remarkable that, although our approach does not achieve
the first place according to ENL and EKI, it greatly increases the resolution, EKI and TCR at a small
cost of ENL when compared with the PDE method and the EWF method, and it greatly increases the
resolution, ENL and TCR at the cost of EKI when compared with the regularization method. This is
what we except. That is to say, our proposed approach can efficiently take a compromise among speckle
noise suppression, resolution improvement and structure preservation such as edges and textures.

(a) (b) (c) (d) (e)

Figure 2. Comparison of feature-enhancement images of dataset 1 by different methods. (a) original
image, (b) NL_LR, (c) regularization, (d) PDE, (e) EWF.

(a) (b) (c) (d) (e)

Figure 3. Zoomed-in of feature-enhancement regions in the red block of dataset 1 by different
methods. (a) original image, (b) NL_LR, (c) regularization, (d) PDE, (e) EWF.

5
AICS 2021 IOP Publishing
Journal of Physics: Conference Series 2025 (2021) 012071 doi:10.1088/1742-6596/2025/1/012071

Table 1. Experiment indexes of image 1.

Resolution
ENL EKI TCR (dB)
Azimuth Range
original 2.09 2.80 3.84 1 26.40
NL_LR 1.89 2.01 7.16 1.12 26.99
Reg 2.06 2.69 0.72 1.38 22.48
PDE 2.05 2.47 8.48 0.77 26.71
EWF 2.45 5.29 9.02 0.58 26.90

Figure 4 shows the original SAR image dataset 2, whose quality and features are severely degraded,
and the feature-enhanced results by different methods. Figure 5 illustrates the zoomed-in. Similarly, all
the methods can accomplish the feature enhancement. The regularization method can enhance the strong
scattering targets, while it acts like truncation, which leads to loss of information with low gray values.
The PDE method reduces the noise but it causes image blurring and resolution decrease. The EWF
method can suppress the noise efficiently at a large cost of resolution and the single point targets are
almost lost. Our proposed mothed sufficiently suppresses the noise and improves the resolution, which
highlights the edges of large homogeneous areas with different gray value level and visually make the
edges and textures more obvious. Table 2 shows the experiment indexes of the original SAR images and
the feature-enhanced ones by different approaches. From the point of view of indexes, all the indexes of
our proposed approach are much better than that of the original at an acceptable loss of EKI. Compared
with the original image, our proposed approach increases the spatial resolutions by 21.50% in azimuth
and 4.34% in range, ENL by 2.70 and TCR by 0.57dB while EPI decreases 3.58% which is acceptable.
When compared with the results of the other two methods, it can make a balance among speckle
suppression in the homogeneous areas, edge preservation and even enhancement of strong scatter points.

(a) (b) (c) (d) (e)

Figure 4. Comparison of feature-enhancement images of dataset 2 by different methods. (a) original
image, (b) NL_LR, (c) regularization, (d) PDE, (e) EWF.

(a) (b) (c) (d) (e)

Figure 5. Zoomed-in of feature-enhancement regions in the red block of dataset 2 by different methods.
(a) original image, (b) NL_LR, (c) regularization, (d) PDE, (e) EWF.

6
AICS 2021 IOP Publishing
Journal of Physics: Conference Series 2025 (2021) 012071 doi:10.1088/1742-6596/2025/1/012071

Table 2. Experiment indexes of image 2.

Resolution
ENL EKI TCR (dB)
Azimuth Range
original 1.98 1.27 7.63 1 24.24
NL_LR 1.62 1.21 10.33 0.96 24.81
Reg 1.77 1.15 0.26 1.41 21.52
PDE 1.63 1.23 10.03 0.65 24.39
EWF 2.61 1.79 10.21 0.56 23.00

4. Conclusion
In this paper, we investigate a novel feature enhancement approach for SAR images based on non-local
and low-rank constraints. In our work, the low-rank constraint is introduced into NL-means framework.
The NL-means framework takes advantages of the non-local self-similarity of SAR images, which will
make this feature enhancement approach better in noise suppression and preservation of structures and
resolution. Besides, when estimating the value of the target pixel, a low-rank patch matrix can be
constructed with vectorization of neighborhood of the target pixel and the similar neighborhoods. By
exploiting this low-rank prior and decomposition of sparse and low-rank matrices, the denoised low-
rank patch matrix is more accurate than the original one, so that this will also increase the accuracy of
the novel approach.
The experiment results show that the performance of the novel approach is much better when
compared with the sparse regularization approach and PDE approach. The proposed feature
enhancement approach for SAR images based on non-local and low-rank constraints can suppress the
noise sufficiently in homogeneous areas and preserve the structural feature, especially edges and
textures. What’s more, it can preserve and even improve the resolution at the same time. And visually,
the result of the proposed method is obviously better than results of the other two.
Despite the results above, we will further improve our research in several aspects. Considering the
computational complexity, fast algorithms need to be designed. Based on the NL-means framework, the
parallel implementation of the algorithm can greatly reduce the computational cost. Besides, we will
take advantages of different types of low-rank prior, such as the low-rank tensor consisting of similar
image patches to improve the performance of feature enhancement. What’s more, taking multiple SAR
images of the same area as the original images to be feature-enhanced, and fusing information in
multiple SAR images may further improve the performance.

Acknowledgments
This work was supported by the National Key Research and Development Program of China
(2020YFA071350504), and National Natural Science Foundation of China (No.61671015).

References
[1] Curlander J C and Mcdonough R N 1991 Synthetic Aperture Radar: Systems and Signal
Processing (Wiley).
[2] Moreira A 1991 Improved multilook techniques applied to SAR and SCANSAR imagery IEEE
Transactions on Geoscience & Remote Sensing 29 (4) 529-534.
[3] Bo Z and Tao L 2016 Performance evaluation in subaperture measurement of synthetic aperture
Radar Ship Electronic Engineering 36 46-49.
[4] Lee J S 1986 Speckle suppression and analysis for synthetic aperture radar images Optical
Engineering 25 636-643.
[5] Kuan D, Sawchuk A and Strand T 1987 Adaptive restoration of images with speckle IEEE
Transactions Acoustics Speech & Signal Processing 35 (3) 373-383.

7
AICS 2021 IOP Publishing
Journal of Physics: Conference Series 2025 (2021) 012071 doi:10.1088/1742-6596/2025/1/012071

[6] Frost V S, Stiles J A and Shanmugan K S 1982 A model for radar images and its application to
adaptive digital filtering of multiplicative noise IEEE Transactions on Pattern Analysis &
Machine Intelligence 4 (2) 157-66.
[7] Lopes A and Touzi R 1990 Adaptive speckle filters and scene heterogeneity IEEE Transactions
on Geoscience and Remote Sensing 28 992-1000.
[8] Lopes A, Nezry E, Touzi R and Laur H 1993 Structure detection and statistical adaptive speckle
filtering in SAR images International Journal of Remote Sensing 14 1735-1758.
[9] Cunha A, Zhou J and Do M 2006 The nonsubsampled contourlet transform: Theory, design, and
applications IEEE transactions on image processing 15 3089-3101.
[10] Zhang W, Fang L and Jiao L 2010 SAR image despeckling using edge detection and feature
clustering in bandelet domain IEEE Geoscience & Remote Sensing Letters 7 (1) 131-135.
[11] Starck J, Candès E and Donoho D 2002 The curvelet transform for image denoising IEEE
Transactions on Image Processing 11 (6) 670-84.
[12] Donoho D 1999 Wedgelets: Nearly minimax estimation of edges The Annals of Statistics 27 (3)
859-897.
[13] Na Y Jiao L 2007 Image Fusion Methods Based on Multi-resolition Analysis Theory (Xidian
University Press).
[14] Tikhonov A 1963 Solution of incorrectly formulated problems and the generalization method
Soviet Meth Dokl 4.
[15] Yu Y and Acton S 2002 Speckle reducing anisotropic diffusion IEEE Transactions on Image
Processing 11 (11) 1260-1270.
[16] Buades A, Coll B and Morel J 2005 A non-local algorithm for image denoising 2005 IEEE
Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) pp 60-
65.
[17] Chandrasekaran V, Sanghavi S and Parrilo P 2009 Rank-sparsity incoherence for matrix
decomposition SIAM Journal on Optimization 21 (2) 572-596.
[18] Yuan X and Yang J 2009 Sparse and low-rank matrix decomposition via alternating direction
methods automation and remote control 25 821-837.
[19] Dong W and Shi G 2014 Compressive sensing via nonlocal low-rank regularization IEEE
Transactions on Image Processing 23 (8) 3618-3632.
[20] Chen W, Tian Q and Jin L 2014 Nonlocal low-rank matrix completion for image interpolation
using edge detection and neural network Signal Image & Video Processing 8 (4) 657-663.
[21] Liu C, Fang F and Xu Y 2016 Single Image Super-Resolution Based on Nonlocal Sparse and
Low-Rank Regularization Pacific Rim International Conference on Artificial Intelligence
(PRICAI) pp 251-261.
[22] Wen B, Li Y and Bresler Y 2017 When sparsity meets low-rankness: Transform learning with
non-local low-rank constraint for image restoration 2017 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP) pp 2297-2301.
[23] Kanoun B, Ferraioli G and Pascazio V 2019 Fast GPU-based enhanced wiener filter for
Despeckling SAR data Remote Sensing 11 (12).
[24] Tan X, Qi Y and Zhu J 2016 Feature enhancement for multichannel SAR image: A non-local
vectorial total variation approach 2016 IEEE International Geoscience and Remote Sensing
Symposium (IGARSS) pp 2541-2544.