IAENG International Journal of Computer Science, 47:4, IJCS_47_4_09

______________________________________________________________________________________

Iterative Method for Image Compression by Using

LSPIA
Lijuan Hu, Yeqing Yi, Chengzhi Liu, and Juncheng Li

Abstract—In this paper, the curve fitting by cubic B-spline Curve fitting plays an important role in computer aided
curves is used to compress gray-level images. The fitting tool geometry design (CAGD), image processing, shape modeling
is the progressive and iterative approximation method for least and data mining. There are some works involving with curve
square fitting (LSPIA), which is used to approximate scanned
image data. Different from the existing methods by using fitting in image compression [4], [5], [6], [7], [8], [9], [11].
piecewise curves, the image data are fitted by a single curve. It is known that the data of a scanned image are very
Hence it can well preserve the relative information between large, it may bring a large amount of computational cost
neighborhood pixels. In particular, to reduce the compression and cause computational instability when fitting by higher
ratio, we further exploit some techniques to save storage order polynomials or non-polynomials. Hence we should
space. Numerical implements show that the proposed method
outperforms the existing methods by using fitting curves. find suitable fitting curves and efficient solvers to fit these
data. Due to these reasons, piecewise curves are suggested in
Index Terms—image compression, LSPIA, cubic B-spline, fitting these scanned data. For example, piecewise Bernstein
Hilbert scan, curve fitting.
polynomials of degree 2 [8], [9]; piecewise quasi cubic ra-
tional Bézier curve [10]; trigonometric Bézier curve [11] and
I. I NTRODUCTION so on. Despite the fact that fitting by piecewise curves has
good fitting performance and is easy to operate, there are also

I N digital image processing, many image files with large

amount of data are generated. These files usually need
to be stored or transmitted. Therefore, an effective image
drawbacks. For instance, the continuity condition between
two adjacent curve segments can not be guaranteed, hence it
may loss the relative information of the neighborhood pixels.
compression method is often required to store and transmit Very often, it is necessary to solve a linear system in
these files. Digital images can be compressed because they traditional curve fitting methods. The system can be solved
often contain redundant and unrelated information. The task by direct solvers [12], e.g., Cholesky factorization, Gauss
of image compression is to eliminate these redundant and elimination with partial pivot and QR factorization. However,
unrelated information. Despite the fact that there exist a large it is known that the solutions solved by direct solvers may be
amount of image compression techniques, further research loss of accuracy and instability when the condition number of
is needed to meet the continuous improvements for more the coefficient matrices are very large [12]. We note in [13]
efficient compression methods. Therefore, the image com- that for large scale problems, the iterative methods would be
pression remains a hot topic in image processing, on which good choices. In 2014, an efficient and robust iterative fitting
various compression methods have been presented, see [1], method, named LSPIA, was proposed by Deng et al. [14]. It
[2], [3]. has the property of shape preserving and can deal with large
Roughly speaking, there exist two kinds of methods for data. Moreover, we can adjust the control points until the
image compression [3]. One is lossless compression, which fitting error satisfies a given tolerance. These elegant merits
only removes the redundant information of the image, such make convenience in using the LSPIA in curve fitting. This
as medical images and fingerprint images, and so on. We naturally arises a question that whether we can apply the
can accurately restore the original image after decompression LSPIA to fit the image data and thus compress the image.
without any distortion [3], e.g., Huffman coding, arithmetic In this paper, we study the application of LSPIA in image
coding. However, it is known that the compression ratio (CR) compression by using cubic B-spline curves.
is limited by the theory of statistical redundancy. Another is
This paper is organized as follows. In section II we review
lossy compression, which can obtain high CR by reducing
the least square fitting method and the methodology of
the quality of images, e.g., singular value decomposition
LSPIA. The image compression algorithm is presented in
[3], transform coding [3], predictive coding [3], curve fitting
Section III. Some image compression examples are given to
[4], [5] and so on. For more detailed research on image
illustrate the effectiveness of the proposed method in Section
compression techniques, we refer the reader to read recent
IV. Some improvement are presented in Section V. A brief
surveys ([1], [2], [3]).
conclusion is drawn in Section VI.
Manuscript received June 21, 2020; revised August 28, 2020. This work
was supported in part by the Natural Science Foundation of Hunan Province
(Grant No. 2017JJ3124 and 2020JJ5267) and the Scientific Research Funds II. C URVE FITTING BY USING CUBIC B- SPLINE
of Hunan Provincial Education Department (Grant No. 18C877).
L. Hu and Y. Yi are with School of Information, Hunan University of A. Least square fitting curve
Humanities, Science and Technology, Loudi 417000, P.R. China, e-mail:
yeqingyi@126.com(Yi) and 1023236794@qq.com(Hu). Let us first recall some details for the conventional curve-
C. Liu and J. Li are with School of Mathematics and Finance, Hunan fitting method in CAGD.
University of Humanities, Science and Technology, Loudi 417000, P.R.
China, e-mail:162101002@csu.edu.cn(Liu, Corresponding Author) and li- Given a data points {Qj }m j=0 to be fitted, and tj be the
juncheng82@126.com(Li). parameters of Qj such that 0 = t0 < t1 < . . . < tm < 1.

Volume 47, Issue 4: December 2020

 IAENG International Journal of Computer Science, 47:4, IJCS_47_4_09
______________________________________________________________________________________

The aim of curve fitting is to find a cubic B-spline curve In curve fitting, we need to measure the fitting error.
n Let {Qj }m
j=0 be the points to be fitted and tj be their
corresponding parameters. Then we use
X
r(t) = Pi Bi,3 (t), (1)
v
i=0 u m
u 1 X
that approximates the points {Qj }m n ε=t kQ − r(tj )k2 . (6)
j=0 best. Here the {Pi }i=0 m + 1 j=0 j
in (1) are the control points to be determined, Bi,3 (t) are
the cubic B-spline basic functions defined at the knot vector to represent the fitting error of the fitting curve r(t).
{0 = u0 = u1 = u2 = u3 < u4 < u5 < . . . < un < un+1 =
un+2 = un+3 = un+4 = 1}, in details B. LSPIA by using cubic B-spline
Given an ordered points set {Qj }mj=0 to be fitted, and tj be

0 if ui ≤ t ≤ ui+1 ;
Bi,0 (t) =
1 otherwise. the parameters of Qj such that 0 = t0 < t1 < . . . < tm < 1.
t − ui ui+p+1 − t (0)
Bi,p (t) = Bi,p−1 (t) + Bi+1,p−1 (t), Firstly, we select {Pi }ni=0 from {Qj }m j=0 as the initial
up+1 − ui ui+p+1 − ui+1 control points and construct the initial approximate fitting
p = 1, 2, 3. curve n
(0)
X
For simplicity, we denote the cubic B-spline basic functions r(0) (t) = Pi Bi (t).
Bi,3 (t) by Bi (t) in this paper. Oftentimes, the number of the i=0
control points is less than that of the points to be fitted, i.e., (0) (0)
Let δj = Qj − r (tj ), j = 0, 1, . . . , m. Then the first
n < m.
adjusting vector for the i-th (i = 0, 1, . . . , n) control point
The main idea of the least square fitting (LSF) method is is given by
to find an optimal control polygon {Pi }ni=0 that minimizes m
(0) (0)
X
the distances between r(t) and {Qj }m j=0 , i.e., ∆i = µ Bi (tj )δj ,
j=0
m
X
min f (P0 , P1 , . . . , Pn ) = min kQj − r(tj )k2 where µ ∈ (0, 2/λ0 ) is a constant, λ0 represents the largest
Pi
j=0 eigenvalue of B T B.
m
X n
X Next, we can generate a new approximate fitting curve
= min kQj − Pi Bi (tj )k2 . n
Pi X (1)
j=0 i=0 r(1) (t) = Pi Bi (t),
(2) i=0
(1) (0) (0)
The norm in (2) is the Euclidean norm. where Pi = Pi + ∆i , i= 0, 1, . . . , n.
The optimal curve r(t) obtained by solving (2) is said to be Suppose that we have obtained (k − 1)-th (k = 1, 2, . . .)
the LSF curve of {Qj }m j=0 . To minimize f (P0 , P1 , . . . , Pn ), curve r(k−1) (t), then the k-th approximate fitting curve can
set the gradient of f to zero, i.e., be generated by
m n n
∂f (k)
X
r(k) (t) = Pi Bi (t), (7)
X X
= −2 Bi (tj )kQj − Pi Bi (tj )k = 0,
∂Pi j=0 i=0 i=0

i = 0, 1, . . . , n. where 
(k) (k−1) (k−1)
 P =P + ∆i ,
 i (k−1) i P

 m
Hence, we have (k−1)
∆i =µ Bi (tj )δj , (8)
n
X 
 j=0
 (k−1)
Qj − Pi Bi (tj ) = 0, j = 0, 1, . . . , m. (3)  δj = Qj − r(k−1) (tj ).
i=0
Therefore, we get a sequence of curves r(k) (t), k =
T T
Let P = [P0 , P1 , . . . , Pn ] and Q = [Q0 , Q1 , . . . , Qm ] . 0, 1, . . .. The initial curve is said to have the LSPIA property
Then the equations (3) can be written in the matrix form if r(k) (t) is convergent. The limit curve of r(k) (t) is the LSF
curve of {Qj }m j=0 . Deng et al. proved that the B-spline curves
BP = Q, (4)
have the LSPIA property [14].
(k) (k)
j=0,1,...,m
where the matrix B = (Bi (tj ))i=0,1,...,n is the so-called Let P(k) = [P0 , P1 , . . . , P(k) T
n ] . Then the euqations (8)
collocation matrix resulting from the cubic B-spline basis. can be written in the matrix form
Therefore the control polygon {Pi }ni=0 can be obtained by P(k+1) = B T Q + (I − µB T B)P(k) . (9)
solving the linear system (4). Since n < m, the system (4)
is over-determined and can be solved by solving the related The LSPIA property means that the sequence of control
(k)
system of normal equations, i.e., polygons {Pi }ni=0 converges to the control polygon of the
LSF curve.
B T BP = B T Q. (5) For LSPIA, the fitting error of the k-th approximate fitting
curve r(k) (t) is given by
As mentioned earlier, the system (5) can be solved by di- v
m
rect solvers or iterative methods. In the following subsection,
u
u 1 X
(k) Qj − r(k) (tj ) 2 .

we will introduce an iterative method for curve fitting with ε = t (10)
m + 1 j=0
clear geometric meaning.

Volume 47, Issue 4: December 2020

 IAENG International Journal of Computer Science, 47:4, IJCS_47_4_09
______________________________________________________________________________________

III. I MAGE COMPRESSION ALGORITHM 220

In this section, we employ the LSPIA for cubic B-spline 200
curves to fit the image data. Curve fitting means to replace
180
the original data points with less control points. Thus, we
can reach the aim of data compression by LSPIA, and we 160

illustrate in Fig. 1 the main process of image compression. 140

120
Fitting the Sample at
Original Hilbert Compressed 100
scanned data the fitting
image scan image
by LSPIA curve 80

60
Fig. 1. The main process of image compression.
40

20

A. Construction of Hilbert curve 2000 4000 6000 8000 10000 12000 14000 16000

To compress a image, one has to convert the 2-dimensional (a) Lena

image data to a 1-dimensional sequence by scanning the 220

image. There are many methods to scan images [9], e.g., 200
raster scan, Z-scan and Hilbert scan, etc. Recently, re-
180
searchers are more likely to use Hilbert scan in the area of
image processing because the Hilbert scan can preserve the 160

relationship of the neighborhoods pixels as much as possible. 140

In [9], Biswas compared the raster scan with Hilbert scan and
120
concluded that Hilbert scanned images can provide better CR
than raster scanned images. Therefore, we use Hilbert scan 100

in this paper. For more details about Hilbert scan, readers 80

can refer to [9]. We show in Fig. 3 two Hilbert curves of
60
Lena and Girl given in Fig. 2.
40

20

0
2000 4000 6000 8000 10000 12000 14000 16000

(b) Girl

Fig. 3. The corresponding Hilbert curves of Lena and Girl in Fig. 2.

For comparison, the accumulated chord parametrization is

also used in numerical examples, and is given by
(a) Lena (b) Girl t0 = 0, tm = 1,
0 Lena0 0 Girl0 . kQj − Qj−1 k (11)
Fig. 2. Original 128 × 128 and tj = tj−1 + , j = 1, 2, . . . , m − 1,
D
Pm
where D = kQj −Qj−1 k is the totally chord length [14].
B. Fitting the Hilbert curve by LSPIA j=0
2) Computation of knot vector: We use the same knot
Suppose that we have obtained the 1-dimensional sequence vector as in [14], i.e.,
{Qj }m j=0 by scanning the image. Then the LSPIA for B-
spline is employed to fit {Qj }m u0 = u1 = u2 = u3 = 0,
j=0 . Before giving the image
compression algorithm, we discuss some implementation uj+3 = (1 − α)ti−1 + αti , j = 1, 2, . . . , n − 3, (12)
details of LSPIA. un+1 = un+2 = un+3 = un+4 = 1,
1) Parametrization: Since the abscissas of Qj (j =
0, 1, . . . m) are distributed uniformly, we can employ the where i = djde, α = jd − i, d = m+1 n−2 and dxe is the largest
uniform parametrization in which we assign a parameter integer not large than x.
tj = j/m for Qj . 3) Selection of initial control points: We note in [14] that
(0)
Remark 3.1: The parametrization of data points is a diffi- the initial points {Pi }ni=0 can be set arbitrary in theory. In
cult hotpot and plays important role in curves representation. this paper, the selection of the initial control points {Pi }ni=0
In general, the accumulated chord parametrization is superior is the same as the method used in [14], i.e.,
to the uniform one in curves representation. While in image (0)
P0 = Q0 , P(0)
n = Qm ,
compression, the uniform parametrization will yield smaller (13)
(0)
CR, which will be discussed later. Pi = Qf (i) , i = 1, 2, . . . , n − 1,

Volume 47, Issue 4: December 2020

 IAENG International Journal of Computer Science, 47:4, IJCS_47_4_09
______________________________________________________________________________________

l m
(m+1)i
where f (i) = n . Finally, we summarize the image compression algorithm
into the following Algorithm 3.2.
Algorithm 3.2: (Image compression algorithm)
C. Compression quality metrics
Input: a gray-level image.
Very often, the peak signal to noise ratio (PSNR) and the Output: a compressed image.
CR are used to measure the quality of the compressed image. 1) Scan the image to obtain the Hilbert sequence
We review the definitions of CR and PSNR [1]. {Qj }m j=0 .
For the original data Qj (j = 0, 1, . . . , m) and the recon- 2) Parameterize Qj with tj and select the initial control
structed data Q̃j = r(k) (tj ), the PSNR is defined as (0)
points {Pi }ni=0 according to (13).
2552 3) Compute the knot vector for cubic B-spline according
PSNR = 10 × log10 , (14) to (12).
MSE
m 
4) Compute the collocation matrix B and the optimal
2
1
Qj − r(k) (tj ) . The PSNR is value of µ.
P
where MSE = m+1
i=0 5) For k = 1, 2, . . . , kmax
usually termed as bit per pixel (bpp). Generally speaking,
the bigger the PSNR, the higher the quality of the image, (a) Update P(k) = µB T Q + (I − µB T B)P(k−1) .
and vice versa. (b) Compute the fitting error ε(k) according to (10).
The CR is defined as (c) If ε(k) ≤ θε(k−1) , break for.
End for.
Bc
CR = , (15) 6) Compute the approximate fitting curve r(k) (t) accord-
Buc
ing to (7).
where Bc represents the number of bits in compressed data, 7) Sample at tj and obtain the recovered data Q̃j =
Buc represents the number of bits in uncompressed data. The r(k) (tj ), j = 0, 1, . . . , m.
CR is termed as bit per pixel (bpp) in image compression. 8) Calculate the PSNR and the CR according to (14) and
Next we discuss the CR of the proposed method. Given an (15), respectively.
image of size M ×M pixels with gray levels {0, 1, . . . , 255}. 9) Display the compressed image.
Since the corresponding Hilbert curve is approximated by
a cubic B-spline curve r(k) (t), which can be stored by IV. I MAGE COMPRESSION EXAMPLES
(k)
saving the control points Pi (i = 0, 1, . . . , n) as well as In this section we employed Algorithm 3.2 to test the
the knot vector ui (i = 0, 1, . . . , n). Besides, having obtained following two well-known images, which are often used to
the fitting curve r(k) (t), we need to save the parameters tj illustrate the effectiveness of image compression methods.
when regenerating the approximate image data by sampling All the numerical experiments were done by Matlab R2012b
at tj , i.e., r(k) (tj )(j = 0, 1, . . . , m). Clearly, it will bloats on a PC with Intel(R) Core(TM) i5-5200U CPU @2.20 GHz
the repository unnecessarily, making it difficult to compress and RAM 6GB.
image efficiently. It should be pointed out that when the Firstly, we select n = 6200 control points when we
uniform parametrization is used, we never need to worry compress the Lena image and select n = 4745 control points
about saving tj . In addition, we can obtain the knot vector when we compress the Girl image. The fitting errors of
ui according to (12). This means that we can save massive the approximate fitting curves r(k) (t) and the PSNR of the
storage space. compressed images are shown in Fig. 4(a) and Fig. 4(b),
According to (7), we remark that the coordinates of the respectively. We observe that the fitting error decreases fast
(k)
control points Pi do not always to be integers. It is well in the first several iterations, and then decreases slowly. Sim-
known that the decimals require larger storage space than ilarly, the PSNR increases fast in the first several iterations,
the integers. To save storage space, we only take the integer and then slows down. Therefore, there is no need to iterate
parts of the coordinates as the control points of the cubic many times because it is time-consuming. Hence we add a
B-spine curve. This can be desirable in image compression terminal condition (16) and we set θ = 0.98 in our tests.
for the following two reasons. On one hand, since Bi (t) ≤ By using two different parametrization methods, we list
1 (t ∈ [0, 1]), the decimal parts of the coordinates have a in Table I the number of iterations required, the CR and
little influence on the results. More exactly, the impact of the the PSNR of the compressed images obtained by Algorithm
ignored decimal parts on the recovered data is no more than 3.2. We denote by UP the uniform parametrization, by ACP
1. On the other hand, the image data are in the gray levels the accumulated chord parametrization, by n the number of
{0, 1, . . . , 255}. We have to round off the decimal parts of the control points of cubic B-spline curve, and by k the iteration
recovered data. The rationality of integration of the control number required, respectively. In our tests, we test different
points is also verified by numerical experiments. n.
From Table I, we observe that the quality of the com-
D. Image compression algorithm pressed images is improved as the number of control points
increases. Consequently, the CR increases. Besides, despite
To ensure the computational efficiency, the iteration (9) the fact that the accumulated chord parametrization can
can also be terminated, if the following stopping criterion provide good results for compression of images. As stated in
ε(k) ≤ θε(k−1) , θ ∈ (0, 1), (16) Section III-C, it is not advisable in practice, because it will
take up lots of storage space to save the knot vectors and
is satisfied. parameters.

Volume 47, Issue 4: December 2020

 IAENG International Journal of Computer Science, 47:4, IJCS_47_4_09
______________________________________________________________________________________

300
Lena
Girl
250 200

200
150
Fitter error

150

100
100

50
50

0
0 5 10 15 20 25 30 35 40 0
2000 4000 6000 8000 10000 12000 14000 16000
Iteration number
(a) Fitting error vs iteration (a) n = 5000
32
220

31 200

30 180

160
29
140
28
PSNR

120
27
100

26 80

60
25
40
24 Lena
Girl 20
23 2000 4000 6000 8000 10000 12000 14000 16000
0 5 10 15 20 25 30 35 40
Iteration number (b) n = 8000
(b) PSNR vs iteration
Fig. 5. Fitting the scanned Hilbert curves of Lena in Fig. 3(a).
Fig. 4. The fitting error and PSNR of the compressed image versus the
iteration.

TABLE I
256 gray images. In Table II, we list the numerical results
N UMERICAL RESULTS OF COMPRESSED 128 × 128 IMAGES obtained by Algorithm 3.2. The numerical results obtained
by Biswas’s methods [9] are also listed for comparison. We
UP ACP can observe that with the same CR, the PSNR of Algorithm
Image n CR
k PSNR k PSNR 3.2 is bigger than those of Biswas’s methods. This means that
4000 0.4883 8 26.3166 8 27.4178
our method performs much better than Biswas’s methods.
5000 0.6104 7 27.2148 7 28.3812 The compressed Lena and Girl images are shown in
Lena
6000 0.7324 9 28.1842 9 29.8114 Fig. 8 and 9, respectively. We can see that the compressed
7000 0.8545 10 29.1810 10 31.4707
8000 0.9766 10 30.0596 10 32.8733
images obtained by Biswas’s methods suffer from visible
9000 1.0986 12 31.1216 12 34.0487 blocking artifacts, while the compressed images obtained by
4000 0.4883 7 29.2893 7 29.9472
Algorithm 3.2 could avoid blocking artifacts satisfactorily.
5000 0.6104 7 30.0564 7 30.7359 Furthermore, it can be found from the detailed figures in
6000 0.7324 9 31.0910 9 32.4577 Fig. 8 and 9 that there exists sawtooth effect at the edges,
Girl
7000 0.8545 10 32.0776 10 34.1655
8000 0.9766 9 33.0083 9 35.7128
no matter whether Algorithm 3.2 or Biswas’s methods are
9000 1.0986 11 34.0185 11 36.6352 used. But the images compressed by Algorithm 3.2 have
the weaker sawtooth effect than those by Biswas’s methods.
These results indicate that the proposed method not only
In Fig. 5 and 6, we show the cubic B-spline curves with reduces blocking artifacts and sawtooth effect but also has
different n when fitting the scanned Hilbert curves in Fig. more compression effect than Biswas’s methods.
3. In Fig. 7, we show the compressed images obtained by
Algorithm 3.2. All the numerical results demonstrate that V. D ISCUSSION AND IMPROVEMENT
the proposed method achieves a good performance in image Oftentimes, compressed images are polluted by many
compression. kinds of noises during the process of image compression,
Secondly, we employ Algorithm 3.2 to compress 256 × especially in lossy compression. Consequently, it will bring

Volume 47, Issue 4: December 2020

 IAENG International Journal of Computer Science, 47:4, IJCS_47_4_09
______________________________________________________________________________________

200

180

160

140

120

100

80

60 (a) n = 5000, Lena (b) n = 8000, Lena

40

20

0
2000 4000 6000 8000 10000 12000 14000 16000
(a) n = 5000

200

180

160

140 (c) n = 5000, Girl (d) n = 8000, Girl

120 Fig. 7. Compressed 128 × 128 images.

100

80

60

40

20

0
2000 4000 6000 8000 10000 12000 14000 16000

(b) n = 8000

Fig. 6. Fitting the scanned Hilbert curves of Girl in Fig. 3(b).

(a) Original 256 × 256 image (b) By Algorithm 1 [9], CR = 1.28

TABLE II
C OMPARISON OF A LGORITHM 3.2 WITH B ISWAS ’ S METHODS [9]

Algorithm 1 [9] Algorithm 2 [9] Algorithm 3.2

Image
CR PSNR CR PSNR CR PSNR
1.44 30.577 1.44 31.222 1.44(11796) 33.641
Lena
1.28 29.757 1.34 30.821 1.28(10485) 33.495
1.09 30.577 1.07 30.436 1.07(8765) 36.010
Girl
0.67 27.692 0.68 28.811 0.67(5488) 32.872

(c) By Algorithm 3.2, CR = 1.28 (d) By Algorithm 3.2, CR = 1.44

unbearable block artifacts, sawtooth effect at the edges and Fig. 8. Comparison of compressed Lena images.
other defects. At this time, we can improve the quality of
the compressed images by employing some pre-processing
or post-processing techniques, such as transform coding [3], VI. C ONCLUSIONS
image filtering [15] and so on. In this paper, we have exploited an image compression
Here we use some image filtering algorithms to reduce the algorithm by using LSPIA due to its efficient and reliable
sawtooth effect at the edges. The median filtering [15] and performance in data fitting. Compared with the other curve
the Gauss filtering [16] algorithms are employed to filtering fitting methods, the proposed method can well preserve the
the compressed image in Fig. 7(c). The filtered images are relative information between neighborhood pixels. Numerical
illustrated in Fig. 10. From these two examples we conclude experiments also show that the proposed method outperforms
that the image filtering algorithms can reduce the sawtooth the similar image compression methods, in terms of the
effect of the compressed images but can not eliminate it CR, the PSNR and the blocking artifacts of the compressed
entirely. images.

Volume 47, Issue 4: December 2020

 IAENG International Journal of Computer Science, 47:4, IJCS_47_4_09
______________________________________________________________________________________

[6] S. M. Othman, A. E. Mohamed, Z. Nossair and M. I. El-Adawy, “Image

compression using polynomial fitting,” International Conference on
Electronics, Communication and Aerospace Technology 2019, 12-14
June, 2019, Coimbatore, India, pp. 344-349.
[7] R. Banerjee and B. S. Das, “An energy efficient image compression
scheme for wireless multimedia sensor network using curve fitting
technique,” Wireless Networks, vol. 25, no. 1, pp. 167-183, 2019.
[8] S. Biswas and S. K. Pal, “Image coding using modified Bézier-Bernstein
approximation,” Information Sciences, vol. 83, no. 3, pp. 175-197, 1995.
[9] S. Biswas, “One-dimensional B-B polynomial and Hilbert scan for
graylevel image coding,” Pattern Recognition, vol. 37, no. 4, pp. 789-
800, 2004.
[10] J. C. Li and D. B. Zhao, “Image compression based on quadratic
(a) Original 256 × 256 image (b) By Algorithm 1 [9], CR = 0.67 rational Bézier curve approximation,” Journal of Huazhong University
of ence and Technology(Natural Science Edition), vol. 40, no. 1, pp.
21-25, 2012.
[11] J. C. Li and D. B. Zhao, “An investigation on image compression using
the trigonometric Bézier curve with a shape parameter,” Mathematical
Problems in Engineering, vol. 2013, pp. 1-8, 2013.
[12] J. Demmel, “Applied numeric linear algebra,” Society for Industrial
and Applied Mathematics, 1997.
[13] Yousef Saad, “Iterative methods for sparse linear systems,” IEEE
Computational Science & Engineering, 1996.
[14] C. Deng and H. Lin, “Progressive and iterative approximation for least
squares B-spline curve and surface fitting,” Computer-Aided Design,
vol. 47, pp. 32-44, 2014.
[15] C. M. Bangug, A. C. Fajardo and R. P. Medina, “A modified
median filtering algorithm (MMFA),” Lecture Notes in Engineering and
(c) By Algorithm 3.2, CR = 0.67 (d) By Algorithm 3.2, CR = 1.07 Computer Science: Proceedings of The International MultiConference
of Engineers and Computer Scientists 2019, IMECS 2019, 13-15 March,
Fig. 9. Comparison of compressed Girl images. 2019, Hong Kong, pp. 377-380.
[16] L. Cadena, F. Cadena and A. Legalov, “Improve Method for Processing
Dental Images with Fast Spatial Filter and Shearlet Transform,” Lecture
Notes in Engineering and Computer Science: Proceedings of The
International MultiConference of Engineers and Computer Scientists
2019, IMECS 2019, 13-15 March, 2019, Hong Kong, pp. 371-376.

(a) Median filter (b) Guass filter

Fig. 10. Post-processing the compressed image in Fig. 8(c) with image
filter.

As a shortcoming of this paper, we use a single curve

to approximate the image data. This technique increases the
quality of the compressed image at the cost of computational
time.

R EFERENCES

[1] J. Uthayakumar, T. Vengattaraman and P. Dhavachelvan, “A survey

on data compression techniques: From the perspective of data quality,
coding schemes, data type and applications,” Journal of King Saud
University Computer & Information Sciences, in press, 2018.
[2] M. A. Rahman and M. Hamada, “Lossless image compression tech-
niques: A state-of-the-art survey,” Symmetry, vol. 11, no. 10, pp. 1274-
1295, 2019.
[3] A. J. Hussain, A. Al-Fayadhb and N. Radi, “Image compression
techniques: A survey in lossless and lossy algorithms,” Neurocomputing,
vol. 300, no. 26, pp. 44-69, 2018.
[4] R. Panda and B. N. Chatterji, “Least squares generalized B-spline signal
and image processing,” Signal Processing, vol. 81, no. 10, pp. 2005-
2017, 2001.
[5] K. S. Sim, F. F. Ting, J. W. Leong, and C. P. Tso, “Signal-to-noise
Ratio estimation for SEM single image using cubic spline interpolation
with linear least square regression,” Engineering Letters, vol. 27, no. 1,
pp. 151-165, 2019.

Volume 47, Issue 4: December 2020