A Fast Modified Signed Discrete Cosine Transform

For Image Compression

Reem T. Haweel Wail S. El-Kilani Hassan H. Ramadan
Basic Science Department Computer Systems Department Basic Science Department
Ain-Shams University, Faculty of Ain-Shams University, Faculty of Ain-Shams University, Faculty of
Computer Science Computer Science Computer Science
Cairo, Egypt Cairo, Egypt Cairo, Egypt
[email protected] [email protected] [email protected]

Abstract- The Discrete Cosine Transform (DCT) is widely to its orthogonal structure, the DCT has good power
used in image compression for its high power compaction compaction properties. The DCT is the best substitute for the
property. The Signed DCT (SDCT) and its modifications
Karhunen Loeve transform, which is considered to be
approximate the DCT and proceed faster. This paper introduces
statistically optimal for power concentration [2]. For this
an efficient and low complexity 8 point transform. The proposed
reason, the DCT is the core of image coding [3] and video
algorithm is derived by applying the signum function operator to
compression techniques such as JPEG, MPEG1, MPEG2,
an existing SDCT modification transform with good power
H.261, and H263 [4].
compaction capabilities. Consequently, the elements of the

proposed transform are only zeroes and ones. No multiplications

In spite of the existence of many fast algorithms [5] which
or shift operations are required. The introduced transform keeps
reduce the total number of operations required to compute
the high power compaction capabilities of its originating
such transforms, multiplication operations may be inevitable.
transform and, in the same time, provides a saving in

computational complexity. A flow diagram is provided for the The DCT matrix, TllCT of order N, is defined as:
fast implementation of the transform. Only 17 additions are

required for both forward and backward

Simulation experiments are provided to justify the efficiency and

improved performance of the proposed

transformations.

transform in image
. ·l�g
�)crCl,J)=
2 n{2j+l)i
-cos:---"---
--'- ----'-
- --
i = O, OSjsN-I

ISisN-I, OSjsN-I
(1)

N 2N
compression compared to other transforms.
To increase the speed of transformation while keeping the
Keywords-Discrete Cosine Transform, Signed Discrete Cosine
compaction properties of the DCT, the Signed Discrete Cosine
Transform, Fast computation, Image compression, and 8x8
transforms. Transform (SDCT) [6] was suggested simply by applying the
signum function operator to the DCT elements in (1). The
I. INTRODUCTION
SDCT matrix, Tmcl, is given by
Image compression is essential in many applications such
as real time visual communication, surveillance and medical
imaging. Image sizes are increasing massively and the need to (2)
accelerate the compression standards is becoming more
Where sign {.} is the signum function defined as
demanding. Transform methods using orthogonal kernel
functions are commonly used in image compression.
Discrete Cosine Transform (DCT) [I] incorporates real
The
sign{x} =
j +1
0
if x>O
if x=O
sinusoids and possesses many interesting features. In addition
-1 if x<O
(3)

978-1-4799-6594-6/14/$31.00 ©2014 IEEE 56

 Several advantages are apparent from the proposed SDCT Where,
defined by (1)-(3). All the elements of the transform are 0 or
±1, that is no multiplication operations or transcendental
D] = diag (1,.J2, # 1,.J2, # J
2 , 2, 2, 2 , 2
(6)
expressions are required. The transform order need not be a
[n TJ, Some of the SDCT elements have been changed to
specific integer or a power of 2 and Odd orders are permitted.
±0.5 and 24 elements have been cleared (turned to zero). It has
Moreover, the SDCT maintains the periodicity and spectral
been shown in [12] that the power compaction of the
structure of its originating DCT and, in turn, maintains its
transform, and consequently the compression capabilities, is
good de-correlation and energy compaction characteristics [6].
high. The computation of the transform requires 17 additions
The 8-by-8 SDCT transform matrix is given by [6]. and two shifts. The transform in [[2] is not completely

+1 +1 +1 +1 +1 +1 +1 +1 orthogonal. There are two nonzero off-diagonal elements in

+1 +1 +1 +1 -1 -1 -1 -1 TITI!. However, the effect of these two nonzero elements on

+1 +1 -1 -1 -1 -1 +1 +1 the compression is negligible and the approximation of the

1 +1 -1 -1 -1 +1 +1 +1 -1 transform transpose as the transform inverse can be

TmcT = .J8 +1 -1 -1 +1 +1 -1 -1 +1 performed.

+1 -1 +1 +1 -1 -1 +1 -1 Another transform, T2 with 26 inserted zero elements has

+1 -1 +1 -1 -1 +1 -1 +1 been introduced in [13]. The other elements in this transform
+1 -1 +1 -1 +1 -1 +1 -1
(4) are ± 1. Due to the large number of inserted zeroes, the
computational requirements are [7 additions only.
Unfortunately, the SDCT reverse transformation matrix is
Unfortunately, the power compaction of this transform is
not orthogonal. The computation of the SDCT requires 24
relatively low as will be seen in section IV. Moreover, the
additions [6]. Following the introduction of the SDCT a
transform is not orthogonal as there are 4 nonzero elements in
stream of research papers has followed such as the Bouguezel­
the T2Tt The transform transpose cannot substitute the
Ahmed-Swamy (BAS) series of algorithms [7]-[11]. The
transform inverse which offsets the advantage of the reduced
target is to modify the SDCT to further reduce the
computational requirements.
computational complexity and to achieve orthogonality. The
strategy is to change some of the SDCT matrix elements and Alternate approximation to the DCT has been obtained by
to clear others. scaling then rounding-off the scaled DCT matrix [[4], [[5].
The resulting transform has ± [elements and 16 zeroes. The
An interesting modification to the SDCT is introduced in
transform is orthogonal and the computational requirements
[12] with a transform, Tl
are 22 additions which is relatively high. The rest of the paper
is organized as follows. Section III introduces the proposed
1 0 0 0 0 -1 -1
algorithm. Section IV demonstrates the improved performance
0.5 -0.5 -1 -1 -0.5 0.5 1
of the proposed transform in image compression. Finally,
0 0 -1 0 0 0 0
7:=Jl Section V concludes the work.
] 2fi -1 -1 1 1 -1 -1 1
-1 0 0 0 0 1 -1
0.5 0 0 -0.5 -0.5 0 0 0.5 II. PROPOSED TRANSFORM
0 0 0 -1 0 0 0
(5)
The motivation for the proposed algorithm is to introduce
an efficient algorithm in the sense of a high compaction

57
properties as well as low complexity. Referring to the computational requirements for different 8x8 transforms. The
literature review summarized in the introduction of section I lowest complexity is that for the proposed transform.
before, the algorithm of [12] has an enough high compaction
� -r---r----�--r---�-r- Xo

properties and a comparative low complexity. However, there XI X2

X2--+-�-T��-r-X ----7�--��� 4
are shift operations that can be saved. The proposed X3 --�--�+-��r7--��----�-L- X6
� �
transform, referred to as Tr, results from applying the signum Xs X3
X6 X,
function operator to the transform in [12] given in eq. 5. The X7 XI

signum function retains the overall spectrum properties and

consequently retains the high compaction properties.
Moreover, the extra shift operation in [12] will disappear so
that an enhancement in the computational complexity is Fig.l Flow diagram for the fast implementation of the proposed algorithm

achieved. The diagonal matrix in (6) is modified to normalize and the explanation of It's building blocks

the main diagonal elements in TrTrf to ones for a uniform

computation. That is the proposed transform; Tr and its TABLE 1. COMPUTATIONAL REQUIREMENTS FOR DIFFERENT TRANSFORMS

associated diagonal D are given by

Transform Adds Shifts Mult. Total

TI' = DT (7) BAS [10] 28 0 4 32

SDCT [6] 24 0 0 24
Where,
Transfonns in [14]. [15] 22 0 0 22
I 1 I I 1 I BAS [7] 21 0 0 21
0 0 0 0 -1 -1 BAS [8], BAS [II] 18 2 0 20
1 -I -1 -I -1 1 I
BAS [9] 18 0 0 18
0 0 -1 0 0 1 0 0 (8)
T=sign{I;}= Transfonn in [12] 16 2 0 18
I -1 -I 1 I -1 -1 I
Transform in [13] 17 0 0 17
1 -1 0 0 0 0 1 -1
Proposed transform 17 0 0 17
I 0 0 -1 -I 0 0 I
0 0 0 -1 1 0 0 0

III. IMAGE COMPRESSION USING THE PROPOSED

(
And,
TRANSFORM

J
1 I l l 1 I I I I (9)
D=
.JTT' =diag .j8'2' .j8'.J2' .j8'2'2'.J2 As mentioned earlier, the DCT, SDCT and its
modifications have good power compaction properties. For a
As can be seen in the next section, the proposed transform
typical image the two-dimensional such transforms
really maintains the good power compaction and orthogonality
concentrate most of the visually significant information about
properties of its originating transform. While maintaining the
the image in just a few coefficients. The international standard
same 24 zeroes of (5) the ±O.5 elements have been converted
lossy image compression algorithm produced by the Joint
to ± I. Thus the shift operations required in the transform of
Photographic Experts Group (JPEG) employs these transforms
[12] have been saved in the proposed transform. The fast tree
[4]. Basically, in JPEG [16], the input image is divided into 8-
computation for the proposed transform is shown in Fig. 1.
by-8 (or 16-by-16 blocks), and the two-dimensional transform
Only 17 add operations are required. Table-l illustrates the
is computed for each block. Using the proposed transform, the
forward transform, F for an 8x8 block, X is given by,

58
 Trees Image
55 r-----�----�----�----F=========�
(10) -B-DCT
50 -B-SDCT
-B-Tl [12]
A fraction of the transformed data is then quantified, z
ww 45 ----8-T2 [13]

� --+- TP (Proposed)
coded and transmitted. In the receiver the inverse two- Eo 40
z
dimensional transform for each block is computed and then e; 35

wl'
the blocks are set back together to form the original image. g 30
w
The estimated block at the receiver is given by �
'"
25
E
�
"
20

x = TI' ' FTI' = D'(T' XT)D 0..

(11) 15

10 �____L-____L-____L-____L-____L-__�
A simulation of the basic JPEG technique has been carried 5 10 11
Number of Coefficients Employed in Compression

out employing the two-dimensional DCT, SDCT, the

transform in [12] (TJ), the transform in [13] (Tz) and the Fig 2. PENN using different transforms for 'trees' image

proposed transform (Tp). The block size used to scan the Circuit Image

50 �---'r-----�----�----F=========�
image is 8-by-8. Out of the 64-transform coefficients in each -B-DCT

45 -B-SDCT
-B-Tl [12]
block, the coefficients employed to reconstruct the image have z
� 40 ----8-T2 [13]

� --+- TP (Proposed)
been varied from 6 to 10 of the top left coefficients. Such Eo 35
z
coefficients retain most of the image energy [6]. All other e; 30

wl'
coefficients have been set to zero. The employed images are g 25
w
the 256-by-256 pixel, 8-bit, gray level 'cameraman', 'circuit', �
'"
20
E

and 'trees' images respectively. The percentage error energy � 15

"
0..

10
norm (PEEN) has been computed as a measure of the
performance of each transform. The PEEN is defmed as 10 11
Number of Coefficients Employed in Compression

M-JN-J
Fig 3. PENN using different transforms for 'circuit' image
II [I(m,n) - IR(m,n W (12)
PEEN = o
m� n�
o
_1 xlOO
M N_l Cameraman Image
III2(m,n) 55 r-----�----�----�----F=========�
111=0 n=O -B-DCT
50 -B-SDCT
-B-Tl [12]
z
Where, I and IR are the M-by-N matrices of the original ww 45 ----8-T2 [13]

� --+- TP (Proposed)

and reconstructed images respectively. The PEEN for different Eo 40

z
e; 35
transforms and images is shown in Figures 2, 3 and 4. It is
wl'
g
Tp is
30
clear that the performance of the proposed algorithm w
� 25
very close to its originating transform TJ [12] as expected '"
E
�
"
20
since the signum function retains the spectral characteristics of 0..

the original transform. However, the PEEN of Tp is a little bit

higher than TJ because of the slightly loss of information due Number of Coefficients Employed in Compression

to the employment of the signum function. Both Tp and TJ Fig 4. PENN using different transforms for 'cameraman' image
outperforms both Tz [13] and the SDCT. Certainly, the DCT
It is clear that the transform Tz has a poor power
has the best performance.
compaction as its associated PEEN is unacceptably very large

59
when the number of coefficients used in the reconstruction is and modifY others. There is a tradeoff between the number of
6. This is also clear from Fig. 5 which shows the reconstructed inserted zeroes which reduces the computational complexity
images using different transforms for the 6 coefficient and the power compaction capability of the resulting
reconstruction case, where the bad quality of the reconstructed transform. This paper introduces an efficient transform with
image using Tz is poor. low complexity and high power compaction. The proposed
transform results from applying the signum function operator
on the high performance transform in [12}. The proposed
algorithm has removed the ±O.5 elements in [12J, and
consequently saved the required shift operations by the
transform of [12]. Aflow diagram for the fast computation of
the proposed transform is provided. The computational
complexity is 17 additions only, which is the lowest among this
family of transforms. The performance of the proposed
Original image Reconstructed image using OCT
transform in image compression has been investigated and
compared to other competitive transforms for different image
databases and for different compression ratios. The
percentage error energy norm (PEEN) has been employed as
a measure to the performance of each transform. It has been
shown that the high performance (low PEEN) of the proposed
algorithm is very close to its originating transform [12J and
Recnstructed Imaqe usinq SOCT Reconstructed image using T, [12] both of them outperform the most recent BAS series [7J-[1 1J
and related transform [13J . Although this most recent
transform [13J has 26 inserted zeroes, it has a poor power
compaction as its associated PEEN is seriously large when the
number of coefficients used in the reconstruction is low.

V. REFERENCES

[I] N. Ahemd, T. Natarajian, and K. R. Rao, "Discrete cosine transform,"

Reconstructed i mage using T, [13] Reconstructed image uing Tp (Proposed) IEEE Trans. Acoust, Speech, Signal Processing, vol. ASSP-32, pp.1234-
1245, Dec. 1984.
[2] R. 1. Clark, "Relation between Karhunen-Loeve and cosine transform,"
Fig. 5. Original and reconstructed images for different transforms using 6 Proc. Inst. Elec. Eng., pt. F, vol. 128, no 6, pp. 359-360, 1981.
[3] Abdelhafez W. M. and Mofaddel M. A., "Hyprid Scheme for lifting
coef
f icients based image coding," Seventh International Conference on Computer
Engineering & Systems (ICCES), pp.141-146, 2012.
[4] Lakhani, G., "Modifying JPEG binary arithmetic coder for exploiting
IV. CONCLUSION
interlintra-block and DCT coefficient sign redundancies," IEEE Trans.
Image Processing, vol. 22, issue 4, pp. 1326 -1339, 2013.
The Signed Discrete Cosine Transform (SDCT) is the [5] H. S. Hou, "A fast recursive algorithm for computing the discrete cosine
transform," IEEE Trans. Acoust. Speech, Signal Process., vol. ASSP-35,
origin for fast transforms employed in image compression. no. 10, pp. 1455-1461, Oct. 1987.
Haweel T.1., "A new square wave transform based on the OCT", Signal
Several modified SDCT algorithms have been investigated in
[6]
Processing, 81, pp. 2309-2319, 2001.

the literature to improve the performance of the SDCT The [7] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy, "A multiplication­
free transform for image compression," in 2nd Int. Conf. Signals,
strategy is to convert some of the SDCT elements to zeroes Circuits and Systems, pp. 1-4, Nov. 2008.

60
[8] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy, "Low complexity [13] Nadia Brahimi and Saad Bouguezel, " An efficient fast integer DCT
8X8 transform for image compression," Electron. Lett. vol. 44, pp. transform for images compression with 16 additions only, " IEEE 7th
1249-1250, Sep. 2008. International Workshop on Systems, Signal Processing and Their
[9] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy, "A fast 8x8 Applications (WOSSPA), pp. 71-74,2011.
transform for image compression," in 2009 Int. Conf Microelectronics [14] Renato 1. Cintra and Fabio M. Bayer,"A DCT approximation for image
(ICM), pp. 74-77,Dec. 2009. compression," IEEE Signal Processing Letters,vol. 18,NO. 10,pp. 579-
[10] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy, "A novel transform 582,Oct. 2011.
for image compression," in 53rd IEEE Int. Midwest Sump. Circuits and [15] K. Saraswathy, D. Vaithiyanathan and R. Seshasayanan, "A DCT
Systems (MWSCAS),pp. 509-512,Aug. 2010. approximation with low complexity for image compression," IEEE
[II] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy, "A low complexity International Conference on Communication and Signal Processing, pp.
parametric transform for image compression," in Proc. IEEE Int. Symp. 465-468, Apr. 3-5, 2013.
Circuits and systems, 2011. [16] JPEG Standard (JPEG ISOIIEC 10918-1 ITU-T Recommendation T.81),
[12] Ranjan K. Senapati, Umesh C. Pati and Kamala K. Mahapatra, " A low 1999. Available at: http://www. w3. orgiGraphics/JPEG/itu-t81.
complexity orthogonal 8x8 transform matrix for fast image compression,
" Annual IEEE India Conference (lNDICON),pp. 1-4,2010.

61