Signal Processing 99 (2014) 201–214

Contents lists available at ScienceDirect

Signal Processing
journal homepage: www.elsevier.com/locate/sigpro

Low-complexity 8-point DCT approximations

based on integer functions
R.J. Cintra a,b,n, F.M. Bayer c, C.J. Tablada a
a
Signal Processing Group, Departamento de Estatística, Universidade Federal de Pernambuco, Brazil
b
Department of Electrical and Computer Engineering, University of Akron, OH, United States
c
Departamento de Estatística and LACESM, Universidade Federal de Santa Maria, Brazil

a r t i c l e i n f o abstract

Article history: The discrete cosine transform (DCT) is a central mathematical operation in several digital
Received 22 October 2013 signal processing methods and image/video standards. In this paper, we propose a
Received in revised form collection of twelve approximations for the 8-point DCT based on integer functions.
21 December 2013
Considered functions include: the floor, ceiling, truncation, and rounding-off functions.
Accepted 24 December 2013
Sought approximations are required to meet the following specific criteria: (i) very low
Available online 1 January 2014
arithmetic complexity, (ii) orthogonality or quasi-orthogonality, and (iii) low-complexity
Keywords: inversion. By varying a scaling parameter, approximations could be systematically
Approximate DCT obtained and several existing approximations were identified as particular cases of the
Image compression
proposed methodology. Particular cases include the signed DCT and the rounded DCT.
Integer functions
Four new quasi-orthogonal approximations were introduced and their practical relevance
Low-complexity algorithms
was demonstrated. All approximations were given fast algorithms based on matrix
factorization methods. Proposed approximations are multiplierless; their computation
requires only additions and bit-shifting operations. Additive complexity ranged from 18 to
24 additions. Obtained approximations were compared with the exact DCT and assessed
in the context of JPEG-like image compression. As quality assessment measures, we
considered the peak signal-to-noise ratio and the structural similarity index. Because its
low-complexity and good performance properties, the proposed approximations are
suitable for hardware implementation in dedicated architectures.
& 2013 Elsevier B.V. All rights reserved.

1. Introduction desirable because the DCT is equipped with fast algorithms

The discrete cosine transform (DCT) is widely regarded
as a key operation in digital signal processing [1,2]. In fact,
the DCT is asymptotically equivalent to the Karhunen–Loève
transform (KLT) [3–5], which is an optimal tool in terms of
decorrelation and energy compaction properties [1–3,6–9].
The KLT has been considered in several contemporary
applications [10–13]. In [14], a distributed version of the
KLT, which is suitable for sensor networks data, was
introduced. However, when high correlated first-order Mar-
kov signals are considered [1,2]—such as natural images [8]
—the DCT can closely emulate the KLT. This approximation is
n
Corresponding author at: Signal Processing Group, Departamento
de Estatística, Universidade Federal de Pernambuco, Brazil.
E-mail addresses: rjdsc@stat.ufpe.org (R.J. Cintra),
bayer@ufsm.br (F.M. Bayer).
and is adequate for hardware implementation.
The DCT has been considered and effectively adopted in
a number of methods for image and video coding [15]. As a
result, the DCT is the central mathematical step for the
following standards: JPEG [16,17], MPEG-1 [18], MPEG-2
[19], H.261 [20], H.263 [21], H.264 [22–24], and the recent
HEVC [25–27]. All the above standards consider the
particular 8-point DCT.
Thus, developing fast algorithms for the efficient eva-
luation of the 8-point DCT is a main task in the circuits,
systems, and signal processing communities. Archived
literature contains a multitude of DCT fast algorithms for
this particular blocklength [28,29]. Remarkably extensive
reports have been generated amalgamating scattered
methods for the 8-point DCT [1,2]. Among the most
popular techniques, we mention the following algorithms:

0165-1684/$ - see front matter & 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.sigpro.2013.12.027
202 R.J. Cintra et al. / Signal Processing 99 (2014) 201–214

Wang factorization [30], Lee DCT for power-of-two block- 2. Exact and approximate DCT
lengths [31], Arai DCT scheme [32], Loeffler algorithm [33],
Vetterli–Nussbaumer algorithm [28], Hou algorithm [29], 2.1. Mathematical preliminaries
and Feig–Winograd factorization [34]. All these methods
are classical results in the field and have been considered The N-point DCT is algebraically represented by the N
N
for practical applications [18,35,36]. For instance, the Arai transformation matrix CN whose elements are given by [1,2]
DCT scheme was employed in a number of recent hard-  
1 πðm  1Þð2n  1Þ
ware implementations of the DCT [37–39]. cm;n ¼ pffiffiffiffi βm  1 cos ;
N 2N
Naturally, DCT fast algorithms that result in major
pffiffiffi
computational savings compared to direct computation where m; n ¼ 1; 2; …; N, β0 ¼ 1, and βk ¼ 2, for k a0.
of the DCT were already developed few decades ago. In Let x ¼ ½x0 x1 ⋯ xN  1  > be an input vector, where the
fact, the intense research in the field has led to methods superscript > denotes the transposition operation. The one-
that are in the vicinity of the theoretical complexity of dimensional (1-D) DCT transform of x is the N-point vector
exact DCT [8,32,33,40,41]. Thus, the computation of the X ¼ ½X 0 X 1 ⋯ X N  1  > given by X ¼ CN  x. Because CN is an
exact DCT is a task with little scope for major improve- orthogonal matrix, the inverse transformation can be written
ments in terms of minimization of computational com- according to x ¼ CN>  X. The DCT is mathematically relevant
plexity by means of standard methods. because of its relationship with the KLT. Indeed, if the input
On the other hand, DCT approximations—operations that signal x is modeled after a first-order Markov process with
closely emulate the DCT—are mathematical tools that can correlation coefficient ρ, then the DCT matrix converges to the
furnish an alternative venue for DCT evaluation. Effectively, KLT matrix, as ρ converges to one (ρ-1) [2, p. 61].
DCT approximations have already been considered in a Let A and B be square matrices of size N. For two-
number of works [2,9,42–44]. Moreover, although usual dimensional (2-D) signals, we have the following expressions
fast algorithms can reduce the computational complexity that relate the forward and inverse 2-D DCT operations,
significantly, they still need floating-point operations [2]. In respectively:
contrast, approximate DCT methods can be tailored to
require very low arithmetic complexity and simple number B ¼ CN  A  CN> and A ¼ CN>  B  CN : ð1Þ
representation systems. Although the procedures described in this work can
A comprehensive list of approximate methods for the be applied to any blocklength, we focus exclusively on the
DCT is found in [2]. Prominent techniques include the 8-point DCT. Thus, for simplicity, the 8-point DCT matrix is
signed DCT (SDCT) [9], the binDCT [8], the level 1 approx- denoted as C and is given by
imation by Lengwehasatit–Ortega [45], the Bouguezel–
Ahmad–Swamy (BAS) series of algorithms [43,44,46–49], 1
C 9C8 ¼
the DCT round-off approximation [50], the modified DCT 2
2 3
round-off approximation [42], and the multiplier-free DCT γ3 γ3 γ3 γ3 γ3 γ3 γ3 γ3
6γ γ2 γ4 γ6  γ6 γ 4 γ 2  γ0 7
approximation for RF imaging [51]. 6 0 7
6 7
The goal of this paper is two-fold. First, we aim at 6 γ1 γ5  γ5  γ1  γ1 γ 5 γ5 γ1 7
6 7
proposing a systematic procedure for deriving low- 6 7
6 γ2 γ 6  γ0  γ4 γ4 γ0 γ6  γ2 7
complexity approximations for the 8-point DCT. For such, 6
6 γ3
7;
6 γ 3  γ3 γ3 γ3 γ 3 γ 3 γ3 7 7
we consider several types of rounding-off functions applied 6 7
6 γ4 γ 0 γ6 γ2  γ2 γ 6 γ0  γ4 7
to the scaled exact 8-point DCT matrix. The entries of the 6 7
6γ γ 1 γ1  γ5  γ5 γ1 γ 1 γ5 7
sought approximate DCT matrices are required to possess 4 5 5
null multiplicative complexity; only additions and simple γ6 γ 4 γ2  γ0 γ0 γ 2 γ4  γ6
bit-shifting operations are allowed. Second, we focus on
suggesting practical approximations and assessing them as where γ k ¼ cos ð2πðk þ 1Þ=32Þ, k ¼ 0; 1; …; 6. These quanti-
tools for JPEG-like image compression. ties are algebraic integers explicitly given by [37]
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
The paper unfolds as follows. In Section 2, we describe pffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pffiffiffiffi
the mathematical framework of the DCT and we discuss the 2þ 2þ 2 2þ 2
γ0 ¼  0:9808…; γ 1 ¼  0:9239…;
polar decomposition method for DCT approximation [52]. 2 2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
In Section 3, we propose a systematic method based on pffiffiffiffi pffiffiffi
integer functions for obtaining low-complexity matrices 2þ 2 2 2
γ2 ¼  0:8315…; γ 3 ¼  0:7071…;
useful for generating DCT approximations. The discussed 2 2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
method is based on a computational search over a subset of pffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pffiffiffiffi
2 2 2 2 2
candidate matrices under constraints of low-computational γ4 ¼  0:5556…; γ 5 ¼  0:3827…;
2 2
complexity; and orthogonality or near orthogonality. In
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Section 4, we provide details of the computational search pffiffiffiffi
2 2þ 2
and list the obtained approximations. Section 5 presents γ6 ¼  0:1951…
fast algorithms for the obtained approximations. In Section 6, 2
the resulting approximations are subject to performance Here, we adopt the following terminology. A matrix A is
assessment in the context of image compression using image said to be orthogonal if A  A > is a diagonal matrix. In
quality measures as figures of merit. In Section 7, we state particular, if A  A > is the identity matrix, then A is said to
concluding remarks. be orthonormal.
 R.J. Cintra et al. / Signal Processing 99 (2014) 201–214
2.2. DCT approximations
Generally, a DCT approximation is a transformation C^
that—according to some specified metric—behaves simi-
larly to the exact DCT matrix C. An approximation matrix C^
is usually based on a transformation matrix T of low
computational complexity. Indeed, matrix T is the key
component of a given DCT approximation.
Often the elements of the transformation matrix T
possess null multiplicative complexity. For instance, this
property can be satisfied by restricting the entries of T to
the set of powers of two f0; 7 1; 72; 74; 7 8; …g. In fact,
multiplications by such elements are trivial and require
only bit-shifting operations.
Approximations for the DCT can be classified into two
categories depending on whether C^ is orthonormal or not.
We start discussing the orthogonal case. In principle, given
a low-complexity matrix T it is possible to derive an
orthonormal matrix C^ based on T by means of the polar
decomposition [52,53]. Indeed, if T is a full rank real matrix,
then the following factorization is uniquely determined:
C^ ¼ S  T; ð2Þ
where S is a symmetric positive definite matrix
[53, p. 348]. Matrix S is explicitly related to T according to
the following relation:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
S ¼ ðT  T > Þ  1 ;
pffi
where  denotes the matrix square root operation [54,55].
Being orthonormal, such kind of approximation satisfies
1 > formation matrices have low computational complexity.
C^ ¼ C^ . Therefore, we have that
1
C^ ¼ T>  S> :
1
As a consequence, the inverse transformation C^ inherits
the same computational complexity of the forward
transformation.
From the computational point of view, it is desirable
that S be a diagonal matrix. In this case, the computational
complexity of C^ is the same as that of T, except for the
scale factors in the diagonal matrix S. Moreover, depend-
ing on the considered application, even the constants in S
can be disregarded in terms of computational complexity
assessment. This occurs when the involved constants are
trivial multiplicands, such as the powers of two. Another
more practical possibility for neglecting the complexity of S
arises when it can be absorbed into other sections of a larger
procedure. This is the case in JPEG-like compression, where
the quantization step is present [16]. Thus, matrix S can be
incorporated into the quantization matrix [42,44–47,50,56].
In terms of the inverse transformation, it is also beneficial
1
that S is diagonal, because the complexity of C^ becomes
> [57], we adopt C ¼ C0 [ f 7 3g as the domain set of the
essentially that of T .
In order that S be a diagonal matrix, it is sufficient that
T satisfies the orthogonality condition:
T  T > ¼ D; ð3Þ
where D is a diagonal matrix [52].
Now we address the non-orthogonal case. If (3) is not
satisfied, then S is not diagonal and the advantageous
203
properties of the resulting DCT approximation are in
principle lost. In this case, the off-diagonal elements
contribute to a computational complexity increase and
the absorption of matrix S in JPEG-like schemes cannot be
easily done. However, at the expense of not providing an
orthonormal approximation, one may consider approxi-
mating S itself by replacing the off-diagonal elements of D
by zeros. Thus, the resulting matrix S^ is given by
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
S^ ¼ ½diagðT  T > Þ  1 ;
where diagðÞ returns a diagonal matrix with the diagonal
elements of its matrix argument. Thus, the non-orthogonal
approximation is furnished by
C~ ¼ S^  T:
Matrix C~ can be a meaningful approximation if S^ is, in some
sense, close to S; or, alternatively, if T is almost orthogonal.
From the algorithm designing perspective, proposing
non-orthogonal approximations may be a less demanding
task, since (3) is not required to be satisfied. However,
since C~ is not orthonormal, the inverse transformation
must be cautiously examined. Indeed, the inverse trans-
formation does not employ directly the low-complexity
matrix T and is given by
1 1
C~ ¼ T  1  S^ :
Even if T is a low-complexity matrix, it is not guaranteed
that T  1 also possesses low computational complexity
figures. Nevertheless, it is possible to obtain non-orthogo-
nal approximations whose both direct and inverse trans-
This situation is illustrated by two prominent approxima-
tions: the SDCT [9] and the BAS approximation described
in [46].
3. Scaling and integer mapping
Approximations archived in the literature often possess
transformation matrices with entries defined on the set
C0 ¼ f0; 71; 7 2g [9,42,44,49,50]. Thus such transforma-
tions possess null multiplicative complexity, because the
required arithmetic operations can be implemented exclu-
sively by means of additions and bit-shifting operations.
However, in [57,58], an image compression scheme based
on the Tchebichef transform was advanced for image
compression. This particular method employs a discrete
transformation referred to as the discrete Tchebichef
transform (DTT) [57]. The implied DTT low-complexity
matrix possesses entries defined on C0 but also considers
the elements 73. Multiplications by constants 7 3 can be
implemented by means of one addition and one bit-
shifting operation (3  x ¼ 2  x þ x). Thus, as suggested in
entries for the sought DCT approximations. Nevertheless,
we emphasize that approximations with entries 73 are
expected to possess a higher computational complexity.
In [9] Haweel introduced a simple approach for design-
ing a DCT approximation. The DCT approximation termed
as SDCT was defined as follows [9]:
signðCÞ;
204 R.J. Cintra et al. / Signal Processing 99 (2014) 201–214

where signðÞ is the signum function applied to each entry 8 

>
> 1 2x  1
of C and is given by >
< x þ if A Z;
2 4
8 roundODD ðxÞ ¼  
< þ1 if x 4 0;
> >
>
> x
1
otherwise:
: 2
signðxÞ ¼ 0 if x ¼ 0;
>
:
1 if x o 0: The round-half-away-from-zero function is the imple-
mentation employed in the round function in Matlab/Octave.
The SDCT can be regarded as a seminal work in the field of The international technical standard ISO/IEC/IEEE 60559:2011
DCT approximations. recommends roundEVEN ðÞ as the nearest integer function of
Additionally, in [50,52,59] a low-complexity DCT choice [64]. This latter implementation is adopted in the
approximation was proposed based on the following scientific computation software Mathematica [65].
matrix:
roundð2  CÞ; 4. Computational search
where roundðÞ is the entrywise rounding-off function as
implemented in Matlab/Octave programming languages 4.1. Problem setup
[55,60]. In this work, our goal is to expand and generalize
the methods employed to derive these two approxima- In this section, we exhaustively compute (4) for judi-
tions above. ciously chosen values of α such that the following condi-
As a venue to design DCT approximations, we consider tions about T ¼ intðα  CÞ are satisfied:
integer functions [61, Chapter 3]. An integer function is
simply a function whose values are integers. We aim at (a) matrix T must possess its elements defined on C;
mapping the exact entries of the DCT matrix into integer (b) T  T > must be a diagonal matrix or must exhibit a
quantities. The resulting matrix is sought to approximate small deviation from diagonality in the sense
the DCT. For such end, we adopt the following general described in [66];
mapping: (c) if T is not orthogonal (cf. (3)), but T  T > is approxi-
mately a diagonal matrix, then the inverse matrix T  1
R⟶M8 ðZÞ must possess low-complexity with its elements defined
α⟼intðα  CÞ; ð4Þ on C.
where M8 ðZÞ is the space of 8
8 matrices over the set of
integers Z and intðÞ is a prototype integer function [61, p. 67]. Condition (a) ensures that the forward transformation
Function intðÞ operates entrywise over its matrix argument. is a low-complexity operation. Therefore, in terms of
Parameter α is termed as the expansion factor and scales the implementation, it may require simple hardware struc-
exact DCT matrix allowing a wide range of possible integer tures. If (3) is satisfied, then the inverse transformation is
mappings [62]. guaranteed to have low computational complexity. This is
Particular examples of integer functions are the floor, because T becomes the transpose of itself, apart from the
ceiling, truncation (round towards zero), and round-away- multiplication by a diagonal matrix. On the other hand, if
from-zero functions. These functions are defined, respec- (3) is not satisfied, then resulting matrices may not be
tively, as follows: useful in contexts that depend on orthogonalization.
Nevertheless, if T  T > is ‘almost’ diagonal, then T can be
floorðxÞ ¼ ⌊xc ¼ maxfm A Z∣m rxg;
of interest. In that case, one may explicitly check whether
ceilðxÞ ¼ ⌈x⌉ ¼ minfn A Z∣n Zxg;
T  1 has low complexity. Thus, Condition (c) is considered.
truncðxÞ ¼ signðxÞ  ⌊jxjc; To quantify the deviation from diagonality, as required
roundAFZ ðxÞ ¼ signðxÞ  ⌈jxj⌉; in Condition (b), we adopt the following measure, called
where j  j returns the absolute value of its argument. deviation from diagonality.
Another particularly useful integer function is the round Definition 1. Let A be a square matrix. Then its deviation
to nearest integer function [63, p. 73]. This function possesses from diagonality is given by [66]
various definitions depending on its behavior for input
arguments whose fractional part is exactly 1=2. Thus, relevant J diagðAÞ J F
δð A Þ ¼ 1  ;
rounding-off methods are the round-half-up, round-half- JAJF
down, round-half-away-from-zero, round-half-towards-zero, where J  J F denotes the Frobenius norm for matrices [53].
round-half-to-even, and round-half-to-odd functions. These
different nearest integer functions are, respectively, given by As a design criterion, we adopt the deviation from
diagonality exhibited by the SDCT as the maximum devia-
roundHU ðxÞ ¼ ⌊x þ 12 c; roundHD ðxÞ ¼ ⌈x  12 ⌉; tion acceptable for non-orthogonalpffiffiffi approximations. Such
 
roundHAFZ ðxÞ ¼ signðxÞ  ⌊x þ 12 c; threshold value is equal to 1  2= 5. The SDCT was chosen
  as a reference transformation because (i) it has proven
roundHTZ ðxÞ ¼ signðxÞ  ⌈x  12 ⌉;
8  good properties [9] and (ii) it is widely employed in
> x1
> 2x  1 performance comparisons [42,43,46–50]. Thus, according
>
< if A Z;
2 4 to this criterion, Condition (b) becomes:
roundEVEN ðxÞ ¼  
>
> 1
>
: xþ2 otherwise; 
(b) δ T  T > r 1  p2ffiffi5  0:1056.
 R.J. Cintra et al. / Signal Processing 99 (2014) 201–214 205

In order that the entries of intðα  CÞ are defined on C,
we must restrict the range of α. We notice that the largest
element of C is γ 0 =2. Thus, it is sufficient to solve the
following inequality for α: 0 r intðα  γ 0 =2Þ r 3. For the
ceiling, floor, truncation, round-away-from-zero, and all
nearest integer functions, we have the following ranges
of α: ½0; 6=γ 0 , ½2=γ 0 ; 8=γ 0 , ½2=γ 0 ; 8=γ 0 , ½0; 6=γ 0 , and
½1=γ 0 ; 7=γ 0 , respectively. For computational purposes, we
employed 106 uniformly spaced values of α (cf. (4)) over
each considered interval.

Table 1
Ceiling function.

Approximation Transformation matrix Range of α Orthogonal?

2 3
T~ 0 1 1 1 1 1 1 1 1 ð0; 2=γ 4 Þ No
61 1 1 1 0 0 0 07
6 7
6 7
61 1 0 0 0 0 1 17
6 7
61 0 0 0 1 1 1 07
6 7
6 7
61 0 0 1 1 0 0 17
6 7
61 0 1 1 0 0 1 07
6 7
6 7
41 0 1 0 0 1 0 15
1 0 1 0 1 0 1 0

Table 2
Truncation function.

Approximation Transformation matrix Range of α Orthogonal?

2 3
T0 [50] 1 1 1 1 1 1 1 1 ð2=γ 4 ; 4=γ 0 Þ Yes
61 1 1 0 0 1 1 1 7
6 7
6 7
61 0 0 1 1 0 0 1 7
6 7
61 0 1 1 1 1 0 1 7
6 7
6 7
61 1 1 1 1 1 1 1 7
6 7
61 1 0 1 1 0 1 1 7
6 7
6 7
40 1 1 0 0 1 1 0 5
0 1 1 1 1 1 1 0
2 3
T1 1 1 1 1 1 1 1 1 ð4=γ 0 ; 4=γ 1 Þ Yes
62 1 1 0 0 1 1 2 7
6 7
6 7
60 1 1 0 0 1 1 0 7
6 7
61 0 2 1 1 2 0 1 7
6 7
6 7
61 1 1 1 1 1 1 1 7
6 7
61 2 0 1 1 0 2 1 7
6 7
6 7
41 0 0 1 1 0 0 1 5
0 1 1 2 2 1 1 0
2 3
T2 1 1 1 1 1 1 1 1 ð4=γ 1 ; 4=γ 2 Þ Yes
62 1 1 0 0 1 1 2 7
6 7
6 7
62 0 0 2 2 0 0 2 7
6 7
61 0 2 1 1 2 0 1 7
6 7
6 7
61 1 1 1 1 1 1 1 7
6 7
61 2 0 1 1 0 2 1 7
6 7
6 7
40 2 2 0 0 2 2 0 5
0 1 1 2 2 1 1 0
2 3
T3 2 2 2 2 2 2 2 2 ð4=γ 4 ; 6=γ 2 Þ Yes
63 2 2 0 0 2 2 3 7
6 7
6 7
63 1 1 3 3 1 1 3 7
6 7
62 0 3 2 2 3 0 2 7
6 7
6 7
62 2 2 2 2 2 2 2 7
6 7
62 3 0 2 2 0 3 2 7
6 7
6 7
41 3 3 1 1 3 3 1 5
0 2 2 3 3 2 2 0
2 3
T~ 1 1 1 1 1 1 1 1 1 ð2=γ 3 ; 2=γ 4 Þ No
61 1 0 0 0 0 1 1 7
6 7
6 7
61 0 0 1 1 0 0 1 7
6 7
61 0 1 0 0 1 0 1 7
6 7
6 7
61 1 1 1 1 1 1 1 7
6 7
60 1 0 1 1 0 1 0 7
6 7
6 7
40 1 1 0 0 1 1 0 5
0 0 1 1 1 1 0 0
206 R.J. Cintra et al. / Signal Processing 99 (2014) 201–214

4.2. Obtained approximations respectively. Moreover, from Table 3, we have that

2 3
2 0 0 0 0 0 0 0
In terms of Condition (a), the ceiling function could 60 1
6 0 0 0 0 0 07 7
supply only one low-complexity candidate matrix for DCT 6 7
approximation as displayed in Table 1. The specific range 60 0 1 0 0 0 0 07
6 7
of α for which the obtained matrix arises is also shown in 60 0 0 1 0 0 0 07
~ 6 7 ~
T4 ¼ 6 7  T 3: ð5Þ
Table 1. Henceforth, we adopt the following notation: 60 0 0 0 2 0 0 07
6 7
orthogonal matrices are denoted as Tk , k ¼ 0; 1; 2; …; and 60 0 0 0 0 1 0 07
6 7
non-orthogonal ones are referred to as T~ k , k ¼ 0; 1; 2; … . 6
40 0 0 0 0 0 1 05
7
However, notice that this particular matrix (Table 1) is
0 0 0 0 0 0 0 1
non-orthogonal and its deviation from diagonality is
exceedingly high. In fact, we have that δðT~ 0 Þ  0:4548. As a consequence, by means of (2), both T~ 3 and T~ 4 lead to
For such reason, this approximation will not be further the same DCT approximation.
considered in this paper. Hereafter, we only list matrices The nearest integer function may result in different
that satisfy all prescribed conditions. approximations when the choice of α results in a matrix
The floor function could not furnish any matrix under α  C whose entries are possibly half-integers. The values of
the prescribed requirements. On the other hand, when α that effect half-integers are of the form l=γ k , k ¼ 0; 1; …; 6,
considering the truncation function, five matrices were where l A Z. Apart from these critical points, the different
obtained, being listed in Table 2. Both orthogonal and non- types of nearest integer functions behave identically. By
orthogonal matrices were found. Similarly, for the round- examining these boundary cases, we could establish inter-
away-from-zero function, a set of four distinct matrices vals for which each of the discussed nearest integer
was derived. These matrices are shown in Table 3. From functions results in meaningful DCT approximations.
Tables 2 and 3, we notice that matrices T0 and T~ 2 coincide Table 4 brings the obtained low-complexity matrices
with the rounded DCT reported in [50] and the SDCT [9], derived from the round-half-up and round-half-down

Table 3
Round-away-from-zero function.

Approximation Transformation matrix Range of α Orthogonal?

2 3
T4 1 1 1 1 1 1 1 1 1=γ 0 Yes
61 1 1 0 0 1 1 1 7
6 7
6 7
61 1 1 1 1 1 1 1 7
6 7
61 0 1 1 1 1 0 1 7
6 7
6 7
61 1 1 1 1 1 1 1 7
6 7
61 1 0 1 1 0 1 1 7
6 7
6 7
41 1 1 1 1 1 1 1 5
0 1 1 1 1 1 1 0
2 3
T~ 2 [9] 1 1 1 1 1 1 1 1 ð0; 2=γ 0 Þ No
61 1 1 1 1 1 1 1 7
6 7
6 7
61 1 1 1 1 1 1 1 7
6 7
61 1 1 1 1 1 1 1 7
6 7
6 7
61 1 1 1 1 1 1 1 7
6 7
61 1 1 1 1 1 1 1 7
6 7
6 7
41 1 1 1 1 1 1 1 5
1 1 1 1 1 1 1 1
2 3
T~ 3 1 1 1 1 1 1 1 1 ð2=γ 2 ; 2=γ 3  No
62 2 1 1 1 1 2 2 7
6 7
6 7
62 1 1 2 2 1 1 2 7
6 7
62 1 2 1 1 2 1 2 7
6 7
6 7
61 1 1 1 1 1 1 1 7
6 7
61 2 1 2 2 1 2 1 7
6 7
6 7
41 2 2 1 1 2 2 1 5
1 1 2 2 2 2 1 1
2 3
T~ 4 2 2 2 2 2 2 2 2 ð2=γ 3 ; 2=γ 4 Þ No
62 2 1 1 1 1 2 2 7
6 7
6 7
62 1 1 2 2 1 1 2 7
6 7
62 1 2 1 1 2 1 2 7
6 7
6 7
62 2 2 2 2 2 2 2 7
6 7
61 2 1 2 2 1 2 1 7
6 7
6 7
41 2 2 1 1 2 2 1 5
1 1 2 2 2 2 1 1
 R.J. Cintra et al. / Signal Processing 99 (2014) 201–214 207

Table 4
Round-half-up and round-half-down functions.

Approximation Transformation matrix Range of α Orthogonal?

T0 [50] See Table 2 ð1=γ 4 ; 1=γ 5 Þ Yes

T4 See Table 3 ð1=γ 5 ; 3=γ 0 Þ Yes
2 3
T5 1 1 1 1 1 1 1 1 ð3=γ 0 ; 3=γ 1 Þ Yes
62 1 1 0 0 1 1 2 7
6 7
6 7
61 1 1 1 1 1 1 1 7
6 7
61 0 2 1 1 2 0 1 7
6 7
6 7
61 1 1 1 1 1 1 1 7
6 7
6 2 1 1 7
61 0 1 0 2 7
6 7
41 1 1 1 1 1 1 1 5
0 1 1 2 2 1 1 0
2 3
T6 [51] 1 1 1 1 1 1 1 1 ð3=γ 1 ; 3=γ 2 Þ Yes
62 1 1 0 0 1 1 2 7
6 7
6 7
62 1 1 2 2 1 1 2 7
6 7
61 0 2 1 1 2 0 1 7
6 7
6 7
61 1 1 1 1 1 1 1 7
6 7
6 2 1 7
1 7
61 0 1 0 2
6 7
41 2 2 1 1 2 2 1 5
0 1 1 2 2 1 1 0
2 3
T7 2 2 2 2 2 2 2 2 ð1=γ 6 ; 3=γ 4 Þ Yes
63 2 1 1 1 1 2 3 7
6 7
6 7
62 1 1 2 2 1 1 2 7
6 7
62 1 3 1 1 3 1 2 7
6 7
6 7
62 2 2 2 2 2 2 2 7
6 7
61 3 1 2 2 1 3 1 7
6 7
6 7
41 2 2 1 1 2 2 1 5
1 1 2 3 3 2 1 1
T~ 1 See Table 2 ð1=γ 3 ; 1=γ 4 Þ No

Table 5 Table 6
Nearest integer functions. Orthogonal approximations.

Function Range of α Transformation matrix Approximation T Diagonal elements of T  T >

roundHAFZ ðÞ ½1=γ 4 ; 1=γ 5 Þ T0 T0 [50] ½8 6 4 6 8 6 4 6

½1=γ 5 ; 3=γ 0 Þ T4 T1 ½8 12 4 12 8 12 4 12
ð3=γ 0 ; 3=γ 1 Þ T5 T2 ½8 12 16 12 8 12 16 12
ð3=γ 1 ; 3=γ 2 Þ T6 T3 ½32 34 40 34 32 34 40 34
½1=γ 6 ; 3=γ 4 Þ T7 T4 ½8 6 8 6 8 6 8 6
½1=γ 3 ; 1=γ 4 Þ T~ 1 T5 ½8 12 8 12 8 12 8 12
T6 [51] ½8 12 20 12 8 12 20 12
roundHTZ ðÞ ð1=γ 4 ; 1=γ 5 Þ T0
T7 ½32 30 20 30 32 30 20 30
½1=γ 5 ; 3=γ 0 Þ T4
½3=γ 0 ; 3=γ 1 Þ T5
½3=γ 1 ; 3=γ 2 Þ T6
ð1=γ 6 ; 3=γ 4 Þ T7
½1=γ 3 ; 1=γ 4 Þ T~ 1 apart from the above listed ones. In Table 5, we show the
roundEVEN ðÞ ð1=γ 4 ; 1=γ 5 Þ T0 resulting matrices.
½1=γ 5 ; 3=γ 0 Þ T4 Considering orthogonal
pffiffiffiffi matrices, as shown in (2), the
ð3=γ 0 ; 3=γ 1 Þ T5 diagonal matrix S ¼ D is required to orthonormalize T.
ð3=γ 1 ; 3=γ 2 Þ T6
In Table 6, the required diagonal elements are listed. As
½1=γ 6 ; 3=γ 4 Þ T7
described in [42,44–47,50,56], in the context of JPEG-like
½1=γ 3 ; 1=γ 4 Þ T~ 1
image compression, these diagonal matrices represent no
roundODD ðÞ ½1=γ 4 ; 1=γ 5 Þ T0 additional arithmetic complexity. This is because they can
½1=γ 5 ; 3=γ 0 Þ T4
½3=γ 0 ; 3=γ 1 Þ T5
be absorbed into the image quantization step [16].
½3=γ 1 ; 3=γ 2 Þ T6 Regarding the obtained non-orthogonal approxima-
½1=γ 6 ; 3=γ 4 Þ T7 tions, in Table 7, we show the deviation from diagonality
½1=γ 3 ; 1=γ 4 Þ T~ 1 values for the associate diagonal matrix D as in (3). The
proposed matrices T~ 3 and T~ 4 showed significantly lower
deviation from orthogonality when compared to the well-
functions. Among the resulting matrices, we identify T6 as known SDCT.
the approximation described in [51]. The remaining near- We explicitly computed the inverse of the non-
est integer functions could not supply any distinct matrix orthogonal matrices, confirming their low-complexity
208 R.J. Cintra et al. / Signal Processing 99 (2014) 201–214

2 3
Table 7 1 2 1 2 1 0 1 0
Deviation from diagonality measure. 61 2 1 0 1 2 1 0 7
6 7
6 7
Approximation T δðT  T Þ> 61 0 1 2 1 0 1 2 7
6 7
16
61 0 1 0 1 2 1 2 7
7
T~ 0 0.4548 ¼ 6 7;
861 0 1 0 1 2 1 2 7
T~ 1 0.0646 6 7
61 0 1 2 1 0 1 2 7
T~ 2 [9] 0.1056 6 7
6 7
T~ 3 0.0063 41 2 1 0 1 2 1 0 5
T~ 4 0.0036 1 2 1 2 1 0 1 0

2 3
characteristic. Indeed, we have the following inverse 1 3 2 3 1 1 1 1
61 3 1 1 1 3 2 1 7
matrices: 6 7
6 7
2 3 61 1 1 3 1 1 2 3 7
1 1 1 1 1 1 0 1 6 7
61 61 2 1 1 3 7
6 1 0  1 1 1 1 1 7
7 ~ 1 6
T3 ¼ 6
1 1 3 7
7
6 7 61 1 2 1 1 3 1 3 7
61 1 0 1 1 1 1 1 7 6 7
6 7 61 1 1 1 1 3 7
61 1 1 1 1 1 0 1 7 6 3 2 7
~ 1 6 7 6 7
T1 ¼ 6 7 41 3 1 1 1 3 2 1 5
6 1 1 1 1 1 1 0 1 7
6 7
6 1 1 0 1 1 1 1 1 7 1 3 2 3 1 1 1 1
6 7
6 7
4 1 1 0 1 1 1 1 1 5 2 3
1 1 1 1 1 1 0 1 1=8 0 0 0 0 0 0 0
2 3 6 7
1=8 0 0 0 0 0 0 0 6 0 1=28 0 0 0 0 0 0 7
6 7
6 0 6 7
6 1=4 0 0 0 0 0 0 7
7 6
6
0 0 1=20 0 0 0 0 0 7
7
6 7 6 0 0 0 1=28 0 0 0 0 7
6 0 0 1=4 0 0 0 0 0 7
6 7
6
6
7;
7
6 7 6 0 0 0 0 1=8 0 0 0 7
6 0 0 0 1=4 0 0 0 0 7 6 7

6
6 0
7 6 0 0 0 0 0 1=28 0 0 7
6 0 0 0 1=8 0 0 0 7
7 6 7
6 7 6 0 0 0 0 0 0 1=20 0 7
6 0 0 0 0 0 1=4 0 0 7 4 5
6 7 0 0 0 0 0 0 0 1=28
6 0 0 0 0 0 0 1=4 0 7
4 5
0 0 0 0 0 0 0 1=4
2 3 2 3
1 2 2 2 1 2 0 2 1 3 2 3 1 1 1 1
61 61 1 1 3 2 1 7
6 2 0 2 1 2 2 2 7
7 6 3 1 7
6 7 6 7
61 2 0 2 1 2 2 2 7 61 1 1 3 1 1 2 3 7
6 7 6 7
61 2 1 1 3 7
16
61 2 2 2 1 2 0 2 7
7
1 6
T~ 4 ¼ 6
1 1 3 7
7
¼ 6 7; 61 1 2 1 1 3 1 3 7
861 2 2 2 1 2 0 2 7 6 7
6 7 61
61 2 0 2 1 2 2 2 7 6 1 1 3 1 1 2 3 7
7
6 7 6 7
6 7 41 3 1 1 1 3 2 1 5
41 2 0 2 1 2 2 2 5
1 2 2 2 1 2 0 2 1 3 2 3 1 1 1 1

2 3 2 3
1 1 1 1 1 0 1 0 1=16 0 0 0 0 0 0 0
61 1 1 0 1 1 1 0 7 6 0 7
6 7 6 1=28 0 0 0 0 0 0 7
6 7 6 7
61 0 1 1 1 0 1 1 7 6 0 0 1=20 0 0 0 0 0 7
6 7 6 7
61 6 7
~ 1 6 0 1 0 1 1 1 1 7
7 6 0 0 0 1=28 0 0 0 0 7
T2 ¼ 6 7
6
6 0
7
7
61 0 1 0 1 1 1 1 7 6 0 0 0 1=16 0 0 0 7
6 7 6 7
61 0 1 1 1 0 1 1 7 6 0 0 0 0 0 1=28 0 0 7
6 7 6 7
6 7 6 0 7
41 1 1 0 1 1 1 0 5 4 0 0 0 0 0 1=20 0 5
1 1 1 1 1 0 1 0 0 0 0 0 0 0 0 1=28
2 3
1=8 0 0 0 0 0 0 0 2 3
1=2 0 0 0 0 0 0 0
6 0 7
6 1=4 0 0 0 0 0 0 7 6 0 1 0 0 0 0 0 07
6 7 6 7
6 0 0 1=8 0 0 0 0 0 7 6 7
6 7 6 0 0 1 0 0 0 0 07
6 7 6 7
6 0 0 0 1=4 0 0 0 0 7 6 0 0 0 1 0 0 0 07

6
6 0
7
7 ~  16
¼ T3 6
7
7:
6 0 0 0 1=8 0 0 0 7 6 0 0 0 0 1=2 0 0 07
6 7 6 7
6 0 0 0 0 0 1=4 0 0 7 6 0 0 0 0 0 1 0 07
6 7 6 7
6 0 0 0 0 0 0 1=8 0 7 6 7
4 5 4 0 0 0 0 0 0 1 05
0 0 0 0 0 0 0 1=4 0 0 0 0 0 0 0 1
 R.J. Cintra et al. / Signal Processing 99 (2014) 201–214 209

Table 8 5. Fast algorithm

Constants required for the fast algorithm.
Considering usual decimation-based techniques and
Approximation m0 m1 m2 m3 m4 m5 m6
matrix factorization [67], fast algorithms for the obtained
T0 [50] 1 1 1 1 1 0 0 transformations could be derived. All discussed matrices
T1 2 0 1 1 1 1 0 share the same factorization structure described below:
T2 2 2 1 1 1 0 0
T3 3 3 2 2 2 1 0
T ¼ P  K  B1  B2  B3 ;
T4 1 1 1 1 1 1 0
T5 2 1 1 1 1 1 0
T6 [51] 2 2 1 1 1 1 0 where P is a permutation matrix, K is a multiplicative
T7 3 2 2 2 1 1 1 matrix, and B1 , B2 , and B3 are additive matrices. These
T~ 1 1 1 1 1 0 0 0 matrices are given by
T~ 2 [9] 1 1 1 1 1 1 1
T~ 3 2 2 2 1 1 1 1 2 3
1 0 0 0 0 0 0 0
T~ 4 2 2 2 2 1 1 1 60
6 0 0 0 1 0 0 0 77
6 7
60 0 1 0 0 0 0 0 7
6 7
60 0 0 0 0 1 0 0 7
6 7
P¼6 7;
60 1 0 0 0 0 0 0 7
6 7
60 0 0 0 0 0 0 1 7
6 7
6 7
40 0 0 1 0 0 0 0 5
0 0 0 0 0 0 1 0

Fig. 1. General signal flow graph for proposed transformations. Input

data xn, n ¼ 0; 1; …; 7, relates to output Xk, k ¼ 0; 1; …; 7, according to
X ¼ T  x. Dashed arrows represent multiplications by  1. (a) Full dia-
gram, (b) Block A.

Table 9
Arithmetic complexity of the obtained approximations.

Approximation Multiplications Additions Bit-shifts

T0 [50] 0 22 0
T1 0 22 4
T2 0 22 6
T3 0 30 16
T4 0 24 0
T5 0 24 4
T6 [51] 0 24 6
T7 0 32 12
T~ 1 0 18 0
T~ 2 [9] 0 28 0
T~ 3 0 28 10 Fig. 2. Quality measures of the considered approximations for several
T~ 4 0 28 12 values of r according to the following figures of merit: (a) average PSNR,
and (b) average PSNR absolute percentage error relative to the DCT.
210 R.J. Cintra et al. / Signal Processing 99 (2014) 201–214

2 3 2 3
m3 0 0 0 0 0 0 0 1 0 0 1 0 0 0 0
6 7 60 07
6 0 m3 0 0 0 0 0 0 7 6 1 1 0 0 0 0 7
6 7 6 7
6 0 0 m5 m1 0 0 0 0 7 61 0 0 1 0 0 0 07
6 7 6 7
6 7 60
6 0 0  m1 m5 0 0 0 0 7 6 1 1 0 0 0 0 07
7
K¼6
6
7; B2 ¼ 6 7;
6 0 0 0 0 m4  m6 m2 m0 77 60
6
0 0 0 1 0 0 07
7
6 7 60
6 0 0 0 0  m0 m4  m6 m2 7 6 0 0 0 0 1 0 07
7
6 7 6 7
6 0 0 0 0  m2  m0 m4 m6 7 40 0 0 0 0 0 1 05
4 5
0 0 0 0 m6  m2  m0 m4 0 0 0 0 0 0 0 1
2 3
1 0 0 0 0 0 0 1
2 3 60 1 0 0 0 0 1 0 7
1 1 0 0 0 0 0 0 6 7
6 7
61 60 0 1 0 0 1 0 0 7
6 1 0 0 0 0 0 07
7 6 7
6 7 60 0 0 1 1 0 0 0 7
60 0 0 1 0 0 0 07 6 7
6 7 B3 ¼ 6 7;
60 61 0 0 0 0 0 0 1 7
6 0 1 0 0 0 0 07
7 6 7
B1 ¼ 6 7; 60 1 0 0 0 0 1 0 7
60 0 0 0 0 0 1 07 6 7
6 7 6 7
60 40 0 1 0 0 1 0 0 5
6 0 0 0 0 0 0 17
7
6 7 0 0 0 1 1 0 0 0
40 0 0 0 0 1 0 05
0 0 0 0 1 0 0 0 where constants mi, i ¼ 0; 1; …; 6, depend on the particular
choice of transformation matrix T. In Table 8, these con-
stants are listed for each of the discussed transformations.
As a consequence, all transformations also share the same

Fig. 3. Quality measures of the considered approximations for several Fig. 4. Quality measures of the considered approximations for several
values of r according to the following figures of merit: (a) average SSIM, values of r according to the following figures of merit: (a) average SR-SIM,
and (b) average SSIM absolute percentage error relative to the DCT. and (b) average SR-SIM absolute percentage error relative to the DCT.
 R.J. Cintra et al. / Signal Processing 99 (2014) 201–214
fast algorithm and signal flow structure, as presented in
Fig. 1. For each transform, we could assess the arithmetic
complexity, as measured by multiplication, addition, and
bit-shifting operation counts. A multiplication by 3 was
counted as one addition and one bit-shift operation. Results
are shown in Table 9. All proposed algorithms are multi-
plierless; requiring only additions and bit-shifts operations.
As shown in (5), transforms T~ 3 and T~ 4 lead to same
approximations. Since T~ 4 requires more arithmetic opera-
tions than T~ 3 does, we do not consider T~ 4 for further
analysis.
6. Image compression
6.1. JPEG-like compression
Discussed transformations were considered as tools for
JPEG-like image compression. Adopting the computational
experiment described in [42,50,56], we employed 45
512
512 8-bit images obtained from a public image bank
[68]. All images were subdivided into 8
8 blocks and
were submitted to a 2-D transformation similar to (1),
where the exact DCT matrix was replaced with a selected
DCT approximation. The resulting 64 coefficients in the
transform domain were ordered in the standard zigzag
Fig. 5. Compressed ‘boat’ image using (a) T0 [50], (b) T4 , (c) T~ 1 , (d) T~ 2 [9], (e) T~ 3 , and (f) DCT, for r¼ 10.
211
sequence [16]. Only the r initial coefficients in each block
were retained; the remaining coefficients being discarded.
We adopted 1 rr r 45.
Subsequently, the inverse 2-D transform was applied
and the compressed images were obtained. Original and
compressed images were then evaluated for image degra-
dation. As quality assessment measures, we considered
(i) the peak signal-to-noise ratio (PSNR) [69], (ii) the
structural similarity index (SSIM) [70], and (iii) the spectral
residual based similarity index (SR-SIM) [71]. For each
value of r, average image quality measures based on the
45 images were considered. As opposed to analyzing
particular images as in [43,44,46–48], by taking average
measurements, the suggested approach is less prone to
variance effects and fortuitous data. Therefore, the pro-
posed methodology is more robust [50,72].
6.2. Results and discussion
Figs. 2(a), 3(a), and 4(a) show the obtained plots based on
the selected quality assessment measures. In order to enhance
visualization of the results, we considered the absolute
percentage error (APE) relative to the DCT as shown in
Figs. 2(b), 3(b), and 4(b). The APE measures simply represent
relative errors in percentage. For instance, considering the
212 R.J. Cintra et al. / Signal Processing 99 (2014) 201–214
Fig. 6. Compressed ‘Lena’ image using (a) T0 [50], (b) T4 , (c) T~ 1 , (d) T~ 2 [9], (e) T~ 3 , and (f) DCT, for r ¼25.
PSNR, the APE is calculated according to the following
expression:
 
PSNRC PSNRT 
APEðPSNRÞ ¼  ;
 values of r. In terms of the approximations with 24 addi-
PSNR C
where PSNRC and PSNRT are the values of PSNR considering
the exact DCT and a given approximation T, respectively. The
values of APEðSSIMÞ and APEðSR  SIMÞ are calculated in a
similar manner.
Approximation T7 outperformed all other approxima-
tions in terms of PSNR. This is partially expected because,
by using all possible elements in C, it may potentially
better approximate the actual DCT vector basis. Never-
theless, it should be noticed that T7 also possesses the
highest computational cost among the examined transfor-
mations. On the other hand, the non-orthogonal approx-
imation T~ 3 outperformed all other approximations in
terms of SSIM and SR-SIM in high compression scenarios
(r r 9). Approximation T~ 1 has the lowest computational
complexity, requiring only 18 additions. The orthogonal
approximation T3 showed comparable performance to the
non-orthogonal approximation T~ 3 . However, the non-
orthogonal approximation is less computationally expen-
sive requiring 28 additions and 10 bit-shifts; whereas
T3 requires 30 additions and 16 bit-shifting operations.
Comparing the approximations with 22 additions, we have
that T0 could outperform approximations T1 and T2 in terms
of PSNR, SSIM, and SR-SIM measures for all considered
tions, we have that T4 showed better behavior than approx-
imations T5 and T6 according to the considered measures.
Moreover, in comparison, T4 has the advantage of requiring
no bit-shifting operations. Focusing on the non-orthogonal
transforms, T~ 3 presented the best performance in terms of
PSNR, SSIM, and SR-SIM measures. However, T~ 1 and SDCT
showed lower computational complexities.
The preceding discussion permits us to identify the
approximations with better performance and complexity
trade-off. Thus, we separate the following approximations:
T0 , T4 , T~ 1 , and T~ 3 . Considering this restricted set of
transformations, we processed two particular images for
qualitative analysis. The SDCT and DCT were also consid-
ered for comparison purposes. Figs. 5 and 6 show ‘boat’
and ‘Lena’ images after being submitted to the JPEG-like
compression experiment for r ¼10 and r ¼25, respectively.
We also included the associate difference images with
respect to the original image in order to qualitatively
illustrate the small amount of error effected by the
proposed approximations. Measurements for PSNR, SSIM,
and SR-SIM are also included.
 R.J. Cintra et al. / Signal Processing 99 (2014) 201–214
7. Conclusion
This paper introduces a collection of DCT approxima-
tions derived from the application of common integer
functions to the exact DCT. The proposed mathematical
formalism could encompass—as particular cases—several
transforms already archived in the literature. In particular,
the well-known SDCT was derived as a result of the
suggested systematic procedure. All proposed transforms
were given fast algorithms that have the same structure.
This suggests a common algebraic framework shared
among all discussed approximations. Only additions and
simple bit-shifting operations were necessary for their
evaluation. The low-complexity property of the obtained
approximations makes them suitable for hardware imple-
mentation in dedicated architecture employing fixed-point
arithmetic. The proposed approximations were assessed in
terms of computational complexity and performance in
JPEG-like compression; exhibiting a good balance between
cost and performance.
Acknowledgments
This research was partially supported by CNPq, CAPES,
FACEPE, and FAPERGS.
References
[1] K.R. Rao, P. Yip, Discrete Cosine Transform: Algorithms, Advantages,
Applications, Academic Press, San Diego, CA, 1990.
[2] V. Britanak, P. Yip, K.R. Rao, Discrete Cosine and Sine Transforms,
Academic Press, Boston, MA, 2007.
[3] N. Ahmed, T. Natarajan, K.R. Rao, Discrete cosine transform, IEEE
Trans. Comput. C-23 (1) (1974) 90–93.
[4] J.L. Doob, Stochastic Processes, first edition 1953, John Wiley & Sons,
New York, 1990.
[5] I.I. Gikhman, A.V. Skorohod, The Theory of Stochastic Processes,
Springer-Verlag, New York, 1974.
[6] K. Karhunen, Über lineare Methoden in der Wahrscheinlichkeits-
rechnung, Annales Academiae scientiarum Fennicae: Mathematica-
Physica, Universitat Helsinki, 1947.
[7] R.J. Clarke, Relation between the Karhunen–Loève and cosine trans-
forms, IEEE Proc. F Commun. Radar Signal Process. 128 (6) (1981)
359–360.
[8] J. Liang, T.D. Tran, Fast multiplierless approximation of the DCT with
the lifting scheme, IEEE Trans. Signal Process. 49 (2001) 3032–3044.
[9] T.I. Haweel, A new square wave transform based on the DCT, Signal
Process. 82 (2001) 2309–2319.
[10] F. Gianfelici, RBF-based technique for statistical demodulation of
pathological tremor, IEEE Trans. Neural Netw. Learn. Syst. 24 (10)
(2013) 1565–1574.
[11] F. Gianfelici, D. Farina, An effective classification framework for
brain–computer interfacing based on a combinatoric setting, IEEE
Trans. Signal Process. 60 (3) (2012) 1446–1459.
[12] C. Turchetti, G. Biagetti, F. Gianfelici, P. Crippa, Nonlinear system
identification: an effective framework based on the Karhunen–Loève
transform, IEEE Trans. Signal Process. 57 (2) (2009) 536–550.
[13] F. Gianfelici, C. Turchetti, P. Crippa, A non-probabilistic recognizer of
stochastic signals based on KLT, Signal Process. 89 (4) (2009)
422–437.
[14] M. Gastpar, P.-L. Dragotti, M. Vetterli, The distributed Karhunen–
Loève transform, IEEE Trans. Inf. Theory 52 (12) (2006) 5177–5196.
[15] V. Bhaskaran, K. Konstantinides, Image and Video Compression
Standards, Kluwer Academic Publishers, Boston, 1997.
[16] G.K. Wallace, The JPEG still picture compression standard, IEEE
Trans. Consum. Electron. 38 (1) (1992) xviii–xxxiv.
[17] W.B. Pennebaker, J.L. Mitchell, JPEG Still Image Data Compression
Standard, Van Nostrand Reinhold, New York, NY, 1992.
213
[18] N. Roma, L. Sousa, Efficient hybrid DCT-domain algorithm for video
spatial downscaling, EURASIP J. Adv. Signal Process. 2007 (2) (2007)
1–16.
[19] International Organisation for Standardisation, Generic Coding of
Moving Pictures and Associated Audio Information – Part 2: Video,
ISO/IEC JTC1/SC29/WG11 – Coding of Moving Pictures and Audio,
ISO, 1994.
[20] International Telecommunication Union, ITU-T Recommendation
H.261 Version 1: Video Codec for Audiovisual Services at p
64
kbits, Technical Report, ITU-T, 1990.
[21] International Telecommunication Union, ITU-T Recommendation
H.263 Version 1: Video Coding for Low Bit Rate Communication,
Technical Report, ITU-T, 1995.
[22] T. Wiegand, G.J. Sullivan, G. Bjontegaard, A. Luthra, Overview of the
H.264/AVC video coding standard, IEEE Trans. Circuits Syst. Video
Technol. 13 (7) (2003) 560–576.
[23] Joint Video Team, Recommendation H.264 and ISO/IEC 14 496–10
AVC: Draft ITU-T Recommendation and Final Draft International
Standard of Joint Video Specification, Technical Report, ITU-T, 2003.
[24] A. Luthra, G.J. Sullivan, T. Wiegand, Introduction to the special issue
on the H.264/AVC video coding standard, IEEE Trans. Circuits Syst.
Video Technol. 13 (7) (2003) 557–559.
[25] M.T. Pourazad, C. Doutre, M. Azimi, P. Nasiopoulos, HEVC: the new
gold standard for video compression: How does HEVC compare with
H.264/AVC? IEEE Consum. Electron. Mag. 1 (3) (2012) 36–46.
[26] F. Bossen, B. Bross, K. Suhring, D. Flynn, HEVC complexity and
implementation analysis, IEEE Trans. Circuits Syst. Video Technol.
22 (12) (2012) 1685–1696.
[27] G.J. Sullivan, J. Ohm, W.-J. Han, T. Wiegand, Overview of the high
efficiency video coding (HEVC) standard, IEEE Trans. Circuits Syst.
Video Technol. 22 (12) (2012) 1649–1668.
[28] M. Vetterli, H. Nussbaumer, Simple FFT and DCT algorithms with
reduced number of operations, Signal Process. 6 (1984) 267–278.
[29] H.S. Hou, A fast recursive algorithm for computing the discrete
cosine transform, IEEE Trans. Acoust. Signal Speech Process. 6 (10)
(1987) 1455–1461.
[30] Z. Wang, Fast algorithms for the discrete W transform and for the
discrete Fourier transform, IEEE Trans. Acoust. Speech Signal Pro-
cess. ASSP-32 (1984) 803–816.
[31] B.G. Lee, A new algorithm for computing the discrete cosine trans-
form, IEEE Trans. Acoust. Speech Signal Process. ASSP-32 (1984)
1243–1245.
[32] Y. Arai, T. Agui, M. Nakajima, A fast DCT-SQ scheme for images,
Trans. IEICE E-71 (11) (1988) 1095–1097.
[33] C. Loeffler, A. Ligtenberg, G. Moschytz, Practical fast 1D DCT algo-
rithms with 11 multiplications, in: Proceedings of the International
Conference on Acoustics, Speech, and Signal Processing, 1989,
pp. 988–991.
[34] E. Feig, S. Winograd, Fast algorithms for the discrete cosine trans-
form, IEEE Trans. Signal Process. 40 (9) (1992) 2174–2193.
[35] B. Vasudev, N. Merhav, DCT mode conversions for field/frame coded
MPEG video, in: IEEE Second Workshop on Multimedia Signal
Processing, 1998, pp. 605–610.
[36] M.C. Lin, L.R. Dung, P.K. Weng, An ultra-low-power image compres-
sor for capsule endoscope, BioMed. Eng. OnLine 5 (1) (2006) 1–8.
[37] H.L.P.A. Madanayake, R.J. Cintra, D. Onen, V.S. Dimitrov, L.T. Bruton,
Algebraic integer based 8
8 2-D DCT architecture for digital video
processing, in: IEEE International Symposium on Circuits and
Systems (ISCAS), 2011, pp. 1247–1250.
[38] N. Rajapaksha, A. Edirisuriya, A. Madanayake, R.J. Cintra, D. Onen,
I. Amer, V.S. Dimitrov, Asynchronous realization of algebraic integer-
based 2D DCT using Achronix Speedster SPD60 FPGA, J. Electr.
Comput. Eng. 2013 (2013) 1–9.
[39] A. Edirisuriya, A. Madanayake, V. Dimitrov, R.J. Cintra, J. Adikari, VLSI
architecture for 8-point AI-based Arai DCT having low area-time
complexity and power at improved accuracy, J. Low Power Electron.
Appl. 2 (2) (2012) 127–142.
[40] A. Edirisuriya, A. Madanayake, R.J. Cintra, V.S. Dimitrov, N.
T. Rajapaksha, A single-channel architecture for algebraic integer
based 8
8 2-D DCT computation, IEEE Trans. Circuits Syst. Video
Technol. 23 (12) (2013) 2083–2089.
[41] M.T. Heideman, C.S. Burrus, Multiplicative Complexity, Convolution,
and The DFT, Signal Processing and Digital Filtering, Springer-Verlag,
New York, NY, 1988.
[42] F.M. Bayer, R.J. Cintra, DCT-like transform for image compression
requires 14 additions only, Electron. Lett. 48 (15) (2012) 919–921.
[43] S. Bouguezel, M.O. Ahmad, M.N.S. Swamy, A multiplication-free
transform for image compression, in: 2nd International Conference
on Signals, Circuits and Systems, 2008, pp. 1–4.
214 R.J. Cintra et al. / Signal Processing 99 (2014) 201–214
[44] S. Bouguezel, M.O. Ahmad, M.N.S. Swamy, A low-complexity para-
metric transform for image compression, in: IEEE International
Symposium on Circuits and Systems (ISCAS), 2011.
[45] K. Lengwehasatit, A. Ortega, Scalable variable complexity approx-
imate forward DCT, IEEE Trans. Circuits Syst. Video Technol. 14 (11)
(2004) 1236–1248.
[46] S. Bouguezel, M.O. Ahmad, M.N.S. Swamy, Low-complexity 8
8
transform for image compression, Electron. Lett. 44 (21) (2008)
1249–1250.
[47] S. Bouguezel, M.O. Ahmad, M.N.S. Swamy, A fast 8
8 transform for
image compression, in: International Conference on Microelectro-
nics (ICM), 2009, pp. 74–77.
[48] S. Bouguezel, M.O. Ahmad, M.N.S. Swamy, A novel transform for
image compression, in: 53rd IEEE International Midwest Sympo-
sium on Circuits and Systems (MWSCAS), 2010, pp. 509–512.
[49] S. Bouguezel, M.O. Ahmad, M.N.S. Swamy, Binary discrete cosine and
Hartley transforms, IEEE Trans. Circuits Syst. I: Regul. Pap. 60 (4)
(2013) 989–1002.
[50] R.J. Cintra, F.M. Bayer, A DCT approximation for image compression,
IEEE Signal Process. Lett. 18 (10) (2011) 579–582.
[51] U.S. Potluri, A. Madanayake, R.J. Cintra, F.M. Bayer, N. Rajapaksha,
Multiplier-free DCT approximations for RF multi-beam digital
aperture-array space imaging and directional sensing, Meas. Sci.
Technol. 23 (11) (2012) 114003.
[52] R.J. Cintra, An integer approximation method for discrete sinusoidal
transforms, J. Circuits Syst. Signal Process. 30 (6) (2011) 1481–1501.
[53] G.A.F. Seber, A Matrix Handbook for Statisticians, John Wiley & Sons,
Inc., Hoboken, NJ, 2008.
[54] N.J. Higham, Computing real square roots of a real matrix, Linear
Algebra Appl. 88/89 (1987) 405–430.
[55] MATLAB, version 8.1 (R2013a) Documentation, The MathWorks Inc.,
Natick, MA, 2013.
[56] F.M. Bayer, R.J. Cintra, A. Edirisuriya, A. Madanayake, A digital
hardware fast algorithm and FPGA-based prototype for a novel 16-
point approximate DCT for image compression applications, Measur.
Sci. Technol. 23 (8) (2012) 114010.
[57] S. Ishwar, P.K. Meher, M.N.S. Swamy, Discrete Tchebichef transform—
A fast 4
4 algorithm and its application in image/video compres-
sion, in: IEEE International Symposium on Circuits and Systems
(ISCAS), 2008, pp. 260–263.
[58] K. Nakagaki, R. Mukundan, A fast 4
4 forward discrete Tchebichef
transform algorithm, IEEE Signal Process. Lett. 14 (10) (2007)
684–687.
[59] F.M. Bayer, R.J. Cintra, Image compression via a fast DCT approxima-
tion, IEEE Lat. Am. Trans. 8 (6) (2010) 708–713.
[60] J.W. Eaton, D. Bateman, S. Hauberg, GNU Octave Manual Version 3,
Network Theory Limited, 2008.
[61] R.L. Graham, D.E. Knuth, O. Patashnik, Concrete Mathematics, 2nd
edition, Addison-Wesley, Upper Saddle River, NJ, 2008.
[62] G. Plonka, A global method for invertible integer DCT and integer
wavelet algorithms, Appl. Comput. Harmon. Anal. 16 (2004) 90–110.
[63] K. Oldham, J. Myland, J. Spanier, An Atlas of Functions, 2nd edition,
Springer, 2008.
[64] International Organization for Standardization, ISO/IEC/IEEE 60559:2011,
2011.
[65] Wolfram Research, Round—nearest integer function 〈http://func
tions.wolfram.com/IntegerFunctions/Round/27/01/01/01/〉, Septem-
ber 2013.
[66] B.N. Flury, W. Gautschi, An algorithm for simultaneous orthogonal
transformation of several positive definite symmetric matrices to
nearly diagonal form, SIAM J. Sci. Stat. Comput. 7 (1) (1986) 169–184.
[67] R.E. Blahut, Fast Algorithms for Signal Processing, Cambridge Uni-
versity Press, Cambridge, UK, 2010.
[68] The USC-SIPI Image Database 〈http://sipi.usc.edu/database/〉, Uni-
versity of Southern California, Signal and Image Processing Institute,
2011.
[69] Q. Huynh-Thu, M. Ghanbari, Scope of validity of PSNR in image/
video quality assessment, Electron. Lett. 44 (13) (2008) 800–801.
[70] Z. Wang, A.C. Bovik, H.R. Sheikh, E.P. Simoncelli, Image quality
assessment: from error visibility to structural similarity, IEEE Trans.
Image Process. 13 (4) (2004) 600–612.
[71] L. Zhang, H. Li, SR-SIM: a fast and high performance IQA index based
on spectral residual, in: 2012 19th IEEE International Conference on
Image Processing (ICIP), 2012, pp. 1473–1476. http://dx.doi.org/10.
1109/ICIP.2012.6467149.
[72] S.M. Kay, Fundamentals of Statistical Signal Processing, Volume I:
Estimation Theory, Prentice Hall Signal Processing Series, vol. 1,
Prentice Hall, Upper Saddle River, NJ, 1993.