⎡ ⎤

DCT-like transform for image compression 1 0 0 1 0 0 0 0

requires 14 additions only ⎢0 1 1 0 0 0 0 0⎥
⎢ ⎥
⎢ ⎥
⎢0 1 −1 0 0 0 0 0⎥
F.M. Bayer and R.J. Cintra ⎢ ⎥
⎢1 −1 0⎥
⎢ 0 0 0 0 0 ⎥
A2 = ⎢ ⎥
⎢0 0 0 0 −1 0 0 0⎥
A low-complexity 8-point orthogonal approximate discrete cosine ⎢ ⎥
⎢0 −1 0⎥
transform (DCT) is introduced. The proposed transform requires no ⎢ 0 0 0 0 0 ⎥
multiplications or bit-shift operations. The derived fast algorithm ⎢ ⎥
⎣0 0 0 0 0 0 −1 0⎦
requires only 14 additions, less than any existing DCT approximation.
Moreover, in several image compression scenarios, the proposed trans- 0 0 0 0 0 0 0 1
form could outperform the well-known signed DCT, as well as state-of- ⎡ ⎤
the-art algorithms.
1 1 0 0 0 0 0 0
⎢1 −1 0 0 0 0 0 0⎥
⎢ ⎥
⎢ ⎥
Introduction: The discrete cosine transform (DCT) is an essential tool ⎢0 0 −1 0 0 0 0 0⎥
⎢ ⎥
in digital signal processing (DSP). In recent years, signal processing lit- ⎢0 0⎥
⎢ 0 0 1 0 0 0 ⎥
erature has been populated with low-complexity methods for the effi- A3 = ⎢ ⎥
⎢0 0 0 0 1 0 0 0⎥
cient computation of the 8-point DCT [1]. Prominent approximation- ⎢ ⎥
⎢0 0⎥
based techniques include the signed DCT (SDCT) [2], the level 1 ⎢ 0 0 0 0 1 0 ⎥
⎢ ⎥
approximation by Lengwehasatit-Ortega [3], the Bouguezel-Ahmad- ⎣0 0 0 0 0 0 1 0⎦
Swamy (BAS) series of algorithms [4– 6], and the DCT round-off 0 0 0 0 0 0 0 1
approximation [7]. ⎡ ⎤
In general, the transformation matrix entries required by approximate 1 0 0 0 0 0 0 0
DCT methods are only {0, +1 /2, +1, +2}. This implies null multi- ⎢0 0 0 0 0 0 0 1⎥
⎢ ⎥
⎢ ⎥
plicative complexity, because the involved operations can be ⎢0 0 0 1 0 0 0 0⎥
⎢ ⎥
implemented exclusively by means of additions and bit-shift operations. ⎢0 0⎥
⎢ 0 0 0 0 1 0 ⎥
In this Letter, we introduce a low-complexity DCT approximation that P=⎢ ⎥
⎢0 1 0 0 0 0 0 0⎥
required only 14 additions. The proposed algorithm attains the lowest ⎢ ⎥
⎢0 0⎥
computational complexity among available methods found in the litera- ⎢ 0 0 0 0 0 1 ⎥
⎢ ⎥
ture. At the same time, the proposed transform could outperform state- ⎣0 0 1 0 0 0 0 0⎦
of-the-art approximations.
0 0 0 0 1 0 0 0

Proposed transform: The proposed approximation is based on the

approximate DCT introduced in [7], hereafter referred to as the CB- The signal flow graph for T is shown in Fig. 1.
2011 matrix. After judiciously replacing elements of the CB-2011
matrix with zeros, we obtained the following matrix: x0 X0

⎡ ⎤ x1 X4
1 1 1 1 1 1 1 1 x2
⎢1 0 0 0 0 0 0 −1 ⎥
X6
⎢ ⎥
⎢1 0 0 −1 −1 0 0 1 ⎥
x3 X2
⎢ ⎥
⎢0 0 −1 0 0 1 0 0 ⎥ x4 X7
T=⎢
⎢1
⎥
⎢ −1 −1 1 1 −1 −1 1 ⎥⎥
x5 X3
⎢0 −1 0 0 0 0 1 0 ⎥
⎢ ⎥ x6 X5
⎣0 −1 1 0 0 1 −1 0 ⎦ x7 X1
0 0 0 −1 1 0 0 0
Fig. 1 Signal flow graph for T
Input data xn , n ¼ 0, 1, . . . , 7, relates to output Xk , k ¼ 0, 1, . . . , 7, according to
The above matrix furnishes the approximate DCT expressed by: Ĉ ¼ D . T, X ¼ T . x. Dashed arrows represent multiplication by 21
1 1 1 1 1 1 1 1
where D ¼ diag √ , √ , , √ , √ , √ , , √ . Arithmetic complexity assessment and comparisons with state-of-the-
8 2 2 2 8 2 2 2
The entries of T are {0, +1}. This is an attestation of its null multi- art DCT approximations are shown in Table 1. Demanding only 14
plicative complexity. Moreover, bit-shift operations are fully absent. additions, the proposed transform Ĉ possesses 22.2, 30.0, and 41.7%
Not only does Ĉ inherit the low computational complexity of T, but it lower arithmetic costs than the BAS-2009 transform [5], the
is also orthogonal. In terms of complexity assessment, matrix D may BAS-2011 transform [6], and the SDCT, respectively. Notice that
not introduce any computational overhead [3– 7]. In image compression, the BAS-2011 transform is the most recent algorithm in the BAS series.
the DCT operation is a pre-processing step for subsequent coefficient
quantisation. In this context, matrix D, in the form of D2, can be Table 1: Arithmetic complexity analysis
merged into the quantisation matrix. Moreover, all elements of D2 are
Method Additions Multiplications Shifts Total
negative powers of two {1/2,1/4,1/8}. Therefore, any implementation
Proposed transform 14 0 0 14
of the quantisation step for the exact DCT can be easily adapted to
SDCT [2] 24 0 0 24
the proposed method by adequately bit-shifting the elements of the
Level 1 approximation [3] 24 0 2 26
quantisation matrix.
A fast algorithm based on sparse matrix factorisation leads to T ¼ P . BAS-2008 transform [4] 18 0 2 20
A3 . A2 . A1 , where: BAS-2009 transform [5] 18 0 0 18
BAS-2011 transform [6] 18 0 2 20
⎡ ⎤
1 0 0 0 0 0 0 1 CB-2011 transform [7] 22 0 0 22
⎢0 1 0 0 0 0 1 0 ⎥
⎢ ⎥
⎢ ⎥ The DSP literature contains the DCT approximation described in [8],
⎢0 0 1 0 0 1 0 0 ⎥
⎢ ⎥ which is claimed to require 16 additions. However, we could not repro-
⎢0 0 ⎥
⎢ 0 0 1 1 0 0 ⎥ duce the performance results shown in [8]. Indeed, contrary to [8], such
A1 = ⎢ ⎥
⎢0 0 0 1 −1 0 0 0 ⎥ approximation could not be verified to be orthogonal. Thus, we could
⎢ ⎥
⎢0 −1 0 ⎥ not consider [8] for any meaningful comparison.
⎢ 0 1 0 0 0 ⎥
⎢ ⎥
⎣0 1 0 0 0 0 −1 0 ⎦
Image compression: To assess the performance of the proposed trans-
1 0 0 0 0 0 0 −1 form for image compression, we used the methodology described in

ELECTRONICS LETTERS 19th July 2012 Vol. 48 No. 15

[2] and supported by [3 – 7]. A set of 45 512 × 512 8-bit greyscale described in [2– 6], we adopted the average quality measure from all
images obtained from a standard public image bank [9] was considered. considered images instead of the results obtained from particular
We implemented the JPEG image compression technique for the 8 × 8 images. Thus, our analysis is more robust; being less prone to variance
matrix case. Each image was divided into 8 × 8 sub-blocks, which were effects and fortuitous input data. Among available algorithms, we separ-
submitted to the two-dimensional transforms. This computation fur- ate the SDCT [2] and BAS-2011 [7] for comparison. The SDCT is a
nished 64 coefficients in the approximate transform domain for each classic reference in the field [10] and the BAS-2011 transform is the
sub-block. According to the standard zigzag sequence only the r most recent method in the BAS series of algorithms. The parametric
initial coefficients in each block were employed to reconstruct the transform BAS-2011 was considered with parameter a ¼ 0.5 [6].
image [7]. All the remaining coefficients were set to zero. We adopted Fig. 2a shows the resulting PSNR measures. The proposed approxi-
2 ≤ r ≤ 45, which corresponds to compression ratios between 96.875 mation Ĉ has comparable performance to the SDCT at high compression
and 29.690%, respectively. The inverse procedure was then applied to rates and could indeed outperform it at low compression rates. At the
reconstruct the processed data and image degradation was assessed. mid-range compression ratios (20 , r ≤ 35), the proposed matrix out-
performed both the BAS-2011 transform and the SDCT. This similar
result could be achieved despite requiring only 70.0 and 58.3% of the
DCT
SDCT arithmetic cost of the BAS-2011 transform and the SDCT, respectively.
BAS−2011
35 proposed transform Fig. 2b depicts the absolute percentage error (APE) relative to the exact
average PSNR, dB

DCT for the average MSE. According to this metric, the proposed
approximation led to a better performance at compression ratios
30
ranging from 90.625% (r ¼ 6) to 45.310% (r ¼ 35), in which
popular compression ratios are included.
25
Fig. 3 shows a qualitative comparison including the DCT, the pro-
posed transform, the BAS-2011 transform, and the SDCT. A 60.937%
compression (r ¼ 25) was applied to the standard Lena image. The pro-
a posed transform offered results that are comparable to those furnished by
2.5 the exact DCT.
SDCT
BAS−2011
2.0 proposed transform
Conclusions: This Letter introduces an 8-point transform suitable for
image compression. The proposed transform requires only 14 additions
APE (MSE)

1.5
and has comparable or better image compression performance than the
1.0 classic SDCT and the state-of-the-art BAS-2011 transform.

0.5 Acknowledgements: This work was supported by CNPq and FACEPE.

0
5 10 15 20 25 30 35 40 45 # The Institution of Engineering and Technology 2012
r
b 4 April 2012
doi: 10.1049/el.2012.1148
Fig. 2 Quality measures for several compression ratios F.M. Bayer (Departamento de Estatística, Universidade Federal de
a Average PSNR Santa Maria, RS, Brazil)
b Average MSE absolute percentage error relative to DCT
R.J. Cintra (Universidade Federal de Pernambuco, PE, Brazil)
E-mail: [email protected]
R.J. Cintra: Also with the University of Akron, OH, USA

References
1 Lecuire, V., Makkaoui, L., and Moureaux, J.-M.: ‘Fast zonal DCT for
energy conservation in wireless image sensor networks’, Electron.
Lett., 2012, 48, (2)
2 Haweel, T.I.: ‘A new square wave transform based on the DCT’, Signal
Process., 2001, 82, pp. 2309–2319
3 Lengwehasatit, K., and Ortega, A.: ‘Scalable variable complexity
DCT (PSNR=37.21) proposed (PSNR=31.44)
approximate forward DCT’, IEEE Trans. Circuits Syst. Video
a b
Technol., 2004, 14, (11), pp. 1236–1248
4 Bouguezel, S., Ahmad, M.O., and Swamy, M.N.S.: ‘Low-complexity
8 × 8 transform for image compression’, Electron. Lett., 2008, 24,
(21), pp. 1249– 1250
5 Bouguezel, S., Ahmad, M.O., and Swamy, M.N.S.: ‘A fast 8 × 8
transform for image compression’. Int. Conf. Microelectronics,
Marrakech, Morocco, 2009, pp. 74–77
6 Bouguezel, S., Ahmad, M.O., and Swamy, M.N.S.: ‘A low-complexity
parametric transform for image compression’. Proc. 2011 IEEE Int.
Symp. Circuits and Systems, Rio de Janeiro., Brazil, 2011
7 Cintra, R.J., and Bayer, F.M.: ‘A DCT approximation for image
compression’, IEEE Signal Process. Lett., 2011, 18, (10), pp. 579–582
BAS-2011 (PSNR=31.33) SDCT (PSNR=31.25) 8 Brahimi, N., and Bouguezel, S.: ‘An efficient fast integer DCT
c d transform for images compression with 16 additions only’. 7th Int.
Workshop on Systems, Signal Processing and their Applications,
Fig. 3 Compressed Lena image using DCT (Fig. 3a), proposed transform Tipaza, Algeria, 2011, pp. 71–74
(Fig. 3b), BAS-2011 transform (Fig. 3c), and SDCT (Fig. 3d), for r ¼ 25 9 USC-SIPI Image Database, University of Southern California, Signal
and Image Processing Institute, http://sipi.use.edu/database/
For the sake of image compression performance assessment, the peak 10 Britanak, V., Yip, P., and Rao, K.R.: ‘Discrete cosine and sine
signal-to-noise ratio (PSNR) and mean square error (MSE) were utilised transforms’ (Academic Press, 2007)
as figures of merit. However, in contrast with the numerical experiments

ELECTRONICS LETTERS 19th July 2012 Vol. 48 No. 15