International Journal of Advancements in Research & Technology, Volume 7, Issue 9, September-2018

ISSN 2278-7763 44

Improved Kuan Filter and Wavelet Based SAR Image Despeck-

ling and Compression
Sikiru Yommy Aiyeola1, Rahmon Ariyo Badru2, Olufunke Janet Alao3, Tofade Christopher Temidayo4, Wasiu
Akande Ahmed5, Alapa Frederick6
1,2,3,4,6
National Space Research and Development Agency (NASRDA), Abuja, Nigeria; 5School of Instrumentation Science and Opto-Electronics
Engineering, Beihang University, Beijing, China
Email: [email protected]

ABSTRACT
Synthetic aperture radar (SAR) image compression is very essential in minimizing storage and transmission costs in bandwidth-
limited channels. In this paper, we present a SAR image compression approach in which speckle noise removal precedes com-
pression. The image compression is improved by first eliminating speckle noise from the image. We start by using the k-nearest
neighbour (K-NN) algorithm to improve the Kuan filter. The improved Kuan filter is then used to remove speckle noise from the
SAR image. The despeckled image is then compressed with the discrete wavelet transform, followed by further encoding of the
wavelet coefficients using the set partitioning in hierarchical trees (SPIHT) scheme with the removal of the slightly complex arith-
metic coding scheme to reduce computational time. The objectives of the present research include; removal of speckle noise from
SAR image, accomplishing efficient SAR image encoding, and exploring enormous advantages of wavelet transform. From the
results obtained the proposed approach favorably compared with the conventional wavelet-SPIHT approach, judging by the val-
ues of the performance evaluation metrics obtained.

IJOART
Keywords : Synthetic Aperture Radar (SAR), SAR image compression, speckle noise, k-nearest neighbour (KNN) algorithm,
Kuan filter, set partitioning in hierarchical Trees (SPIHT), wavelet transform.

1 INTRODUCTION
Synthetic Aperture Radar (SAR) images are an important and edge preservation are simultaneously required [7]. Lee,
means of gathering information about the earth surface by sci- Kuan, Frost, and others are spatial domain filters [8], [9], [10],
entists. However, while the amount of data obtained is increas- [11]. In this paper, we propose a SAR image compression ap-
ing speedily, the speed of transmission to the ground receiving proach in which the despeckling precedes compression. We use
station, and storage capacity, are not increasing rapidly. Hence, the k-nearest neighbour algorithm to significantly improve the
algorithms for encoding and decoding image data that will en- performance of Kuan filter. The use of SPIHT without the arith-
sure excellent compression while at the same time give good metic coding also reduces the computational time of the ap-
quality reconstructed image need to be developed [1]. But in proach. This paper is organized as follows: Section 2 discusses
SAR images, the quality of the reconstructed image is always the objectives of the present research, Section 3 discusses
degraded due to the presence of speckle noise. The speckle phe- speckle noise in SAR images, and Section 4 discusses Kuan fil-
nomenon, is caused by coherent radiation during SAR image ter, In Section 5 the K-Nearest Neighbour Algorithm is dis-
processing, and in order to get high quality reconstructed SAR cussed, while Section 6 discusses Wavelet Transform, SPIHT
image, the image must be despeckled prior to further pro- Algorithm is discussed in Section 7, Section 8 presents the per-
cessing. Researchers in the image processing community have formance Evaluation Metrics, while Sections 9, 10, and 11 re-
embraced the use of wavelet transforms due to their high en- spectively present the Proposed Approach, Experimental Re-
ergy compaction capability and multiresolution analysis for ef- sults and Conclusion.
fectively analysing images [2], [3]. Wavelets are a mathematical
tool suitable for image compression [4]. Wavelet transform al- 2 OBJECTIVES OF THE PRESENT RESEARCH
gorithm creates a pyramid containing the low resolution ap-
(a) Since speckle noise affects the result of any SAR image en-
proximation and the details of the original image at different coding method, it is therefore essential to first deal with this
scales [5]. It decorrelates an image both in space and frequency, phenomenon before the image is encoded.
and has the capability to concentrate the image energy into a (b) To accomplish the goal of perfectly encoding SAR images
few low and high frequency coefficients [6]. Among the well- using two dimensional discrete wavelet transform.
known SAR image despeckling filters are Lee, Kuan, and Frost, (c) The aim is to explore the enormous advantages of wavelet
transform for image encoding. Wavelet-based image compres-
but with these filters however, difficulties arise when denoising

IJOART
Copyright © 2018 SciResPub.
 International Journal of Advancements in Research & Technology, Volume 7, Issue 9, September-2018
ISSN 2278-7763 45

sion schemes have gained popularity due to their ability to re- 𝑊𝑓 = Filter window centered at the pixel with coordinates (𝑟, 𝑐)
duce the blocking artifacts that are commonly experienced in 𝐶𝑊𝑓 = Center pixel in 𝑊𝑓
JPEG-based image compression. Wavelet transform also has the 𝐶2
1− 𝑢⁄ 2
ability to achieve multiresolution analysis, which leads to better 𝐾 = Weighting function, given by: 𝐾 =
𝐶𝑒
energy compaction with good quality reconstructed images. 1+𝐶𝑢2
𝜇𝑊𝑓 = Local mean within the filter window, 𝑊𝑓
𝜎𝑊𝑓 = Standard deviation of 𝑊𝑓
3 SPECKLE NOISE IN SAR IMAGES
A despeckled pixel can be expressed as follows:
When a surface is illuminated by a radar and such surface is
rough compared to the wavelength of the radar, the echoes em- 𝑌(𝑟, 𝑐) = 𝜇𝑊𝑓 + 𝐾(𝐶𝑊𝑓 − 𝜇𝑊𝑓 ) (2)
anating from the surface comprises of waves that are reflected 𝜎𝑊𝑓
from a number of scatterers contained in what is called a reso- 𝐶𝑒 = and the noise variance coefficient is given by:
𝜇𝑊𝑓
lution cell. As a result of the roughness of this surface, the sep-
𝐶𝑢 = 1⁄
aration between the scatterers and receiver differs, hence the re- √𝑁𝐿𝑜𝑜𝑘
ceived echoes are no more coherent in phase, even though they
are still coherent in frequency. If the received waves sum up The NLook value controls the amount of smoothing that can be
constructively, the strength of the received signal is strong, if on applied to an image, by simply varying the noise variance coef-
the other hand, they are out of phase during their reception, the
ficient. The higher the NLook, the more the image details that
received signal’s strength is weak. The formation of SAR image
is due to coherent processing of radar pulses that return in suc- are preserved. On the other hand, by reducing the NLook better
cession at the receiver. This process results in pixels with vary- smoothing is achieved [14].
ing intensities. This variation in intensities is what results in the
granular appearance seen in SAR images, and that is referred to
as speckle noise. This noise modifies the intensities of the im-
5 K-NEAREST NEIGHBOUR ALGORITHM
aged surface [12]. Figure 1 shows a SAR image with speckle It is a non-parametric search algorithm used for classifying objects

IJOART
contamination. and it is based on closest training examples. It is a technique for
locating the points in a group of data that are closest to the object
of interest. If in a given-dimensional space, we represent a group
of 𝑦 reference positions by 𝑅 = {𝑟1 , 𝑟2 , 𝑟3 , … 𝑟𝑦 } and 𝑦 represents a
group of query positions 𝑄 in the same space, then 𝑄 =
{𝑞1 , 𝑞2 , 𝑞3 , … 𝑞𝑦 }. The 𝑘 closest point determination requires
searching for the k nearest points to each position q in the reference
group R. Euclidean and other distance determination methods can
be used for this purpose [15]. Figure 2 shows an example with k=3.
The blue dots represent some reference positions, while the green
cross represents the query position. This algorithm is used in pat-
tern recognition [16], [17]. We can find the Euclidean distance be-
tween the points 𝑅 and 𝑄 from:
Fig. 1. SAR Image with speckle contamination
𝑫𝑬 = √(𝒓𝟏 − 𝒒𝟏 )𝟐 + (𝒓𝟐 − 𝒒𝟐 )𝟐 + ⋯ + (𝒓𝒚 − 𝒒𝒚 )𝟐 (3)
3.1 Speckle Noise Model
The basic steps required in the K-NN algorithm classification
Speckle noise is a characteristic of SAR images. It is a random are illustrated in this paper, [18].
noise which impacts SAR image negatively [13]. Speckle noise
is multiplicative. Degraded SAR image can be modelled by the
equation below:

𝐼𝐷 (𝑚, 𝑛) = 𝐼𝑜 (𝑚, 𝑛) ∗ 𝑁(𝑚, 𝑛) (1)

where 𝐼𝐷 (𝑚, 𝑛) is the degraded image with speckle noise,

𝐼𝑂 (𝑚, 𝑛) is the original image, 𝑁(𝑚, 𝑛) is the speckle noise.

4 KUAN FILTER
Kuan filter is a well-known SAR image despeckling filter that
had been used by the image processing community. To apply Fig. 2. An example of K-Nearest Neighbour (K-NN) Classification
this filter, the following representations are made:

IJOART
Copyright © 2018 SciResPub.
 International Journal of Advancements in Research & Technology, Volume 7, Issue 9, September-2018
ISSN 2278-7763 46

6 WAVELET TRANSFORM D. Step Four

Increment N by one, and go back to the sorting pass (step B).
A wavelet is a “small wave” which has its energy compacted in
time and this provides a means for analyzing non-stationary
and transient phenomena. Wavelet transform has been used in 8 PERFORMANCE EVALUATION METRICS
data compression [19]. A function f(t) can be analyzed, or pro- We use the mean squared error (MSE), the peak signal-to-noise
cessed if it is expressed as a linear decomposition by: ratio (PSNR) and the feature preserving index (FPI) as the per-
formance evaluation metrics.
𝑓(𝑡) = ∑𝑠 𝑎𝑠 𝜓𝑠 (𝑡) (4)
1
𝑀𝑆𝐸 = ∑𝑋−1 𝑌−1 ̂
𝑥=0 ∑𝑦=0(𝐼[𝑥, 𝑦] − 𝐼 [𝑥, 𝑦])
2
(8)
where the integer s is an index of for the summation, and a_s is 𝑋𝑌

the expansion coefficients and ψ𝑠 (𝑡) is the set of real-valued where the original image is 𝐼[𝑥, 𝑦] and 𝐼̂[𝑥, 𝑦] is the decom-
function of t called the expansion set. Wavelets are generated pressed image. The peak signal-to-noise-ratio is given as:
from a mother wavelet by scaling and translation. Two-dimen-
sional representation of wavelets is achieved from the func- 𝑃𝑆𝑁𝑅 = 10𝑙𝑜𝑔10 (2𝑁 − 1)2 /𝑀𝑆𝐸 (9)
tion 𝑓(𝑡) [20].
𝑎⁄ where the pixels are represented using N number of bits,
ψ𝑎,𝑏 (𝑡) = 2 2 𝜓(2𝑎 𝑡 − 𝑏), 𝑎, 𝑏 ∈ 𝑍 (5) while the feature preserving index is given as:

∑𝑛
1 (2.𝑅𝑓 −𝑅𝑓1 −𝑅𝑓2 )
7 SPIHT ALGORITHM 𝐹𝑃𝐼 = ∑𝑛
(10)
1 (2.𝑅−𝑅1 −𝑅2 )
The SPIHT is an important wavelet coefficients coding scheme. where the original pixel value on the linear feature is repre-
It was brought into the image coding community in the year sented by R, while 𝑅1 and 𝑅2 give the values of the surrounding
1996. In this algorithm, the wavelet coefficients obtained from pixels on both sides of the feature. . 𝑅𝑓 ,𝑅𝑓1 and 𝑅𝑓2 respectively
the wavelet transformation of the image are first encoded and give the filtered values of the pixels [22]. The higher the value

IJOART
then transmitted using multiple passes [2]. SPIHT uses most of the FPI, the better a filter to preserves linear features in an
significant bit (msb) approach for sorting the transform coeffi- image [23].
cients from the wavelet decomposition and it employs the list
of significant pixels (LSP), list of insignificant sets (LIS) and list
of insignificant pixels (LIP) in the coding process. 9 PROPOSED APPROACH
We present a method of encoding SAR images which gives a
7.1 SPIHT Coding Steps compressed image of good quality. In this method, attempt is
A. Step One made to remove speckle noise from the SAR image by using an
In each of the multiple passes undergone during the SPIHT en- improved Kuan filter. The KNN classification algorithm is used
coding, only the coefficients that are equal or greater in magni- to improve the performance of this filter for use in removing
tude than the threshold are encoded. The threshold (Thresh) for speckle noise from the image before its compression using the
each pass is computed as follows: discrete wavelet transform. A window size of KxK is chosen
and the method is such that instead of using all the neighbour
𝑇ℎ𝑟𝑒𝑠ℎ = 2𝑃−𝑁 (6) pixels to calculate the filtering parameters, one can select a
number of neighbour pixels to ensure that the best results are
where, N=0, 1,2,3,4 ...,P obtained in terms of performance evaluation metrics values.
N= the pass number, and P is obtained as follows: The number of neighbour pixels specified is k. For this experi-
ment, we used a sliding window size of 3x3, and out of the 9
𝑃 = ⌊𝑙𝑜𝑔2 𝑚𝑎𝑥𝑖𝑚𝑢𝑚(𝑐𝑜𝑒𝑓𝑓(𝑖, 𝑗))⌋ (7) elements, we chose 6 to calculate the filtering parameters. The
𝑐𝑜𝑒𝑓𝑓(𝑖, 𝑗) is the coefficient at coordinate (i,j)in the image. resulting image is then compressed using the 2-D Discrete
Wavelet Transform. This is followed by the use SPIHT algo-
B. Step Two: Sorting Pass rithm to further encode the wavelet coefficients. Figure 3 shows
When N=n, where n is an integer. The pixel that satisfies the an illustration of the despeckling approach using 3x3 sliding
condition 𝑇(𝑛) ≦ |𝑐𝑜𝑒𝑓𝑓(𝑖, 𝑗)| < 2𝑇(𝑛) is regarded to as being window. The blue dots represent the image pixels as an exam-
significant. The significant pixel’s position and sign bit are then ple and the center pixel of the sliding window is the red dot.
encoded.

C. Step Three: Refinement Pass

The third step of the SPIHT coding is the refinement pass. Dur-
ing this pass, the pixels that satisfy the following condition
|𝑐𝑜𝑒𝑓𝑓(𝑖, 𝑗) ≧ 2𝑇(𝑛)| are refined by encoding the 𝑛𝑡ℎ most sig-
nificant bit (that is those pixels that had their coordinates trans-
mitted during the previous sorting pass).

IJOART Fig. 3 Illustration of the center pixel sliding window

Copyright © 2018 SciResPub.
 International Journal of Advancements in Research & Technology, Volume 7, Issue 9, September-2018
ISSN 2278-7763 47

TABLE 1
10 EXPERIMENTAL RESULTS PERFORMANCE EVALUATION RESULTS FOR THE SAR IMAGE FOR
The experiment is carried out on a sample SAR image contam- ONE OF THE DECOMPOSITION LEVELS
inated with speckle noise of variance 0.03. The image is com- Evaluation Metrics Compression without Proposed
pressed with wavelet transform without speckle noise removal Despeckling
which is the conventional way of compressing SAR images. We MSE 0.1028 0.0672
then use our two-stage approach on the image. First, we des- PSNR (dB) 19.7630 23.4559
peckled the image with the improved Kuan filter and second, FPI 0.6812 0.7581
we compressed the image using wavelet transform and further Time (seconds) 5.2458 2.8483
encoded the wavelet coefficients with SPIHT scheme, but with dB = decibel.
the removal of the slightly complex arithmetic coding stage.
Displayed in Figure 4 are three SAR images with (a) being the 0.5
original image, (b) is the image that is encoded without dealing Proposed
0.45 Compress only
with the inherent speckle noise and (c) result of the proposed
approach. The values of the performance evaluation metrics 0.4

clearly show the superiority of our proposed approach. Figure 0.35

5 shows the plots of the performance evaluation metrics against 0.3

wavelet decomposition levels. The values of the performance

MSE
0.25
evaluation metrics are contained in Table 1. The highlighted
values are those obtained for our approach. The values of MSE, 0.2

PSNR FPI, and computational time for the third level of decom- 0.15

position have been tabulated. And for all the levels of decom- 0.1
position, our approach shows better performance.
0.05

IJOART
1 2 3 4 5 6 7
Decomposition Levels

(a)

25

20

Proposed
Compress only
PSNR

15
(a) (b)

10

5
1 2 3 4 5 6 7
Decomposition Levels

(b)

0.9

0.8

Despeckling and Compression

0.7
(c)
0.6
Proposed
Fig. 4 (a) SAR image with speckle of variance 0.03 (b) Image com-
Compress only
pressed only (c) (proposed) 0.5
FPI

0.4

0.3

0.2

0.1

0
1 2 3 4 5 6 7
Decomposition Levels

(c)
IJOART
Copyright © 2018 SciResPub.
 International Journal of Advancements in Research & Technology, Volume 7, Issue 9, September-2018
ISSN 2278-7763 48

6
Video Technology volume 6, pp. 243–250, June 1996.
Proposed
Compress only
[3] J. Shapiro, “Embedded image coding using zerotrees of
5 wavelet coefficients,” IEEE Transactions on Signal Pro-
cessing, volume 41, pp. 3445–3462, December 1993.
4 [4] P. Subashini, M. Krishnaveni and S. K Thakur, “Perfor-
mance Measures of Wavelet Families on SAR Images for
Time (seconds)

3 Compression Hypothesis,” 2010 International Conference

on Computer Design and Applications (ICCDA 2010), vol-
2
ume 1, pp. 377-381, June 25-27, 2010.
[5] S.G. Mallat, “A Theory for Multiresolution Signal Decom-
1
position: The Wavelet Representation,” IEEE Transactions
on Pattern Analysis and Machine Intelligence, volume 2,
0
1 2 3 4 5 6 7 number 7. pp. 674-693, July 1989.
Decomposition Levels
[6] M.B.I Reaz, M. Akter and F. M Yasin, “An efficient image
. (d) coder using modified SPIHT for wireless communica-
tion,” WSEAS Transactions on Communications, volume
Fig. 5 (a) MSE vs decomposition levels (b) PSNR vs decomposi- 5, issue 6, pp. 1146-1150, June 2006.
tion levels (c) FPI vs decomposition levels (d) Computational time
vs decomposition levels [7] J. S Lee, “Refined filtering of image noise using local sta-
tistics,” Computer Graphics and Image Processing, vol-
11 CONCLUSION ume 15, issue 4, pp. 380-389, April 1981.
[8] J. S. Lee, “Digital image enhancement and noise filtering
In this paper, we have introduced a new approach for SAR im-
by use of local statistics,” IEEE Transactions on Pattern

IJOART
age compression where speckle noise is first dealt with, due to
Analysis and Machine Intelligence, volume 2, issue 2, 165–
its adverse effects on the quality of reconstructed image. Our
approach employed the use of k-nearest neighbour algorithm 168 March 1980.
to improve the performance of Kuan filter for better speckle [9] D. T. Kuan, A. A. Sawchuk, T. C. Strand and P. Chavel,
noise removal. The despeckled image is then compressed using “Adaptive noise smoothing filter for images with signal-
the discrete wavelet transform. For the despeckling we used a dependent noise,” IEEE Transactions on Pattern Analysis
window size of 3x3 and the filtering parameters were obtained and Machine Intelligence, volume 7, issue 2, pp. 165–177,
using only 6 out of the 9 elements of the sliding window. We March 1985.
observed that when we attempted to use larger window sizes, [10] V. S. Frost and J. A. Stiles, K. S Shanmugan, and J. Holtz-
blur was introduced into the despeckled image. Usually, when man “A model for radar images and its application to
too large window sizes are used blur is introduced into the des- adaptive digital filtering of multiplicative noise,” IEEE
peckled image. Hence, we had to limit our window size to 3x3. Transactions on Pattern Analysis and Machine Intelli-
In this work, speckle noise was been removed to prepare the gence, volume 4, issue 2, pp. 157-166, March 1982.
SAR image for better compression using discrete wavelet trans- [11] A. Lopes, C.E.S.R, E. Nezry, R. Touzi, and H. Laur, “Max-
form (DWT) and from the values of the evaluation metrics, it imum a posteriori speckle filtering and first order texture
can be concluded that our approach outperformed the conven- model in SAR images,” Proceedings of IEEE Geoscience
tional SAR image compression method using wavelet trans- and Remote Sensing Symposium, IGRASS 1990, pp. 2409-
form. 2412, May 1990.
[12] Y. Li, “Synthetics Aperture Radar Image Compression Us-
ACKNOWLEDGMENT ing Wavelet Transform,” Faculty of Engineering and Ap-
We wish to express our profound appreciation to the manage- plied Sciences, Memorial University of Newfoundland,
ment of the National Space Research and Development Agency pp. 1-125, August, 1977,
(NASRDA), Abuja, Nigeria. [13] K.M.S Raju, M.S Nasir and T.M Devi, “Filtering Tech-
niques to reduce Speckle Noise and Image Quality En-
REFERENCES hancement methods on Satellite Images,” IOSR Journal of
[1] Z. Zeng and I. G. Cumming, “SAR Image Data Compres- Computer Engineering (IOSR-JCE), volume 15,issue 4, pp.
sion Using a Tree-Structured Wavelet Transform,” IEEE 10-15, November – December, 2013.
Transactions on Geoscience and Remote Sensing, volume [14] A. Akl, K. Tabbara, and C. Yaacoub, “An enhanced Kuan
39, number 3, pp. 546-552, March 2001. filter for suboptimal speckle reduction,” 2nd IEEE Interna-
[2] A. Said and W. Pearlman, “A new, fast, and efficient image tional Conference on Advances in Computational Tools
codec based on set partitioning in hierarchical trees for Engineering Applications (ACTEA), pp. 91-95, 2012.
(SPIHT),” IEEE Transactions on Circuits, Systems and [15] V. Garcia, E. Debreuve and M. Barlaud, “Fast k nearest

IJOART
Copyright © 2018 SciResPub.
 International Journal of Advancements in Research & Technology, Volume 7, Issue 9, September-2018
ISSN 2278-7763 49

neighbour search using GPU,” Computer Vision and Pat-

tern Recognition Workshops, 2008. CVPRW'08. IEEE
Computer Society Conference, pp. 1-6, 2008.
[16] Z. Voulgaris, and G. D. Magoulas, “Extensions of the k
nearest neighbour methods for classification problems,”
26th IASTED International Conference on Artificial Intel-
ligence and Applications, CD Proceedings ISBN: 978-0-
88986-710-9, pp. 23-28, Feb. 2008.
[17] F. Moreno-Seco, L. Mico, and J. A Oncina, “Modification
of the LAESA Algorithm for Approximated k-NN Classi-
fication,” Pattern Recognition Letters 24, pp. 47–53, 2003.
[18] Q. Kuang and L. Zhao, “A practical GPU-based KNN al-
gorithm,” ISCSCT, Huangshan, China, pp. 151-155, 2009.
[19] O.O Khalifa, S.H Harding and A.A Hashim, “Compres-
sion using Wavelet Transform,” [J], Signal Processing, An
International Journal, volume 2, issue 5, pp. 17-26.
[20] D. Mistry and A. Banerjee, “Discrete Wavelet Transform
using Matlab,” International Journal of Computer Engi-
neering and Technology (IJCET), volume 4, issue 2, pp.
252-259, 2013.
[21] P. Liu, and G. Li, “An Improved SPIHT Algorithm for Im-
age Compression in Low Bit Rate,” Communications and

IJOART
Network, Scientific Research Journal, pp. 245-248, 2013.
[22] A. Rajamani and V. Krishnaveni, “Performance analysis
survey of various SAR image despeckling techniques,” [J],
International Journal of Computer Applications (0975-
8887), Volume 90, No. 7, pp. 5-13, March, 2014.
[23] S.Y Aiyeola, L. Rongke, O. O Spencer and C. I Ani “SAR
Image Despeckling and Compression Using K-Nearest
Neighbour Based Lee Filter and Wavelet,” 2015 8th Inter-
national Congress on Image and Signal Processing (CISP
2015), 978-1-4673-9098-9/15/$31 ©2015 IEEE, pp. 158-167.

IJOART
Copyright © 2018 SciResPub.