J Chongqing Univ.-Eng. Ed.
Computer Science & Application
Vol. 4 No. 2
June 2005
Article ID: 1671-8224(2005)02- 0084-06
Comparison of fast discrete wavelet transform algorithms
*
MENG Shu-ping
a
, TIAN Feng-chun, XU Xin
College of Communication Engineering, Chongqing University, Chongqing 400030, P.R. China
Received 24 February 2005; revised 10 April 2005
Abstract: This paper presents an analysis on and experimental comparison of several typical fast algorithms for discrete wavelet
transform (DWT) and their implementation in image compression, particularly the Mallat algorithm, FFT-based algorithm, Short-
length based algorithm and Lifting algorithm. The principles, structures and computational complexity of these algorithms are
explored in details respectively. The results of the experiments for comparison are consistent to those simulated by MATLAB. It is
found that there are limitations in the implementation of DWT. Some algorithms are workable only for special wavelet transform,
lacking in generality. Above all, the speed of wavelet transform, as the governing element to the speed of image processing, is in
fact the retarding factor for real-time image processing.
Keywords: discrete wavelet transforms (DWT); fast algorithms; computational complexity
CLC number: TN911.72; Document code: A
1 Introduction
1
With the development of the application of wavelet
transform (WT), the speed of wavelet transform as a
hot issue has been put forward. Fast algorithms of
wavelet transform have been studied by many scholars.
The Mallat pyramid algorithm put forward by Mallat,
as the earliest one in this category, greatly reduces the
arithmetic complexity of WT, and plays an important
role in the theoretical research of the theory of wavelets.
Since its coming into being, various fast algorithms of
the WT have been put forward [1-8]. In this study, the
complexity and implementation of several fast
algorithms of WT by software are studied.
2 Concepts of wavelet transform
Wavelet transform includes continuous wavelet
transform (CWT), discrete wavelet transform (DWT)
and wavelet series transform (WST). Suppose that t
denotes the time, ( )t
and
ˆ
()t
ψ
are the mother
wavelet and its dual, respectively, and ( )t
must
satisfy the admission condition. It is defined that
/2
,
() 2 (2 ),
jj
jk
tftk=−
,jk∈Z, and
,,
{()}
kjk
t
ψ
and
,,
ˆ
{()}
kjk
t
ψ
are the wavelet bases of
2
()L R
respectively. Suppose
2
() ( )ft L∈ R
, its CWT is
defined by
a
MENG Shu-ping (
孟书苹
): Female; Born 1981; Postgraduate;
Research interest: optical wavelet transform; E-mail:
mapping81@sina.com.
*
Funded by the Natural Science Foundation of China
(No.60472037).
,
1
ˆ
CWT{ ( ); , } ( ) ( )d
ab
tab ft t t
a
ψ
+∞
−∞
=
∫
1
() ( )d
tb
tt
a
a
ψ
+∞
−∞
−
∫
)
(1)
Its wavelet series transform is
WST{ ( ); , } CWT{ ( ); 2 , 2 }
jj
ft jk ft a b k
−−
===
/2
ˆ
2()(2)d
jj
ttkt
ψ
+∞
−∞
−
∫
, ,jk∈Z (2)
In practice, f(t) often appears in discrete form f (m),
its DWT is defined as
DWT{ ( ); , } WST{ ( ); , }
tjk fmtjk
Δ=
/2
ˆ
2()(2)
jj
m
mmk
ψ
,,mjk∈Z . (3)
3 Fast algorithms for discrete wavelet
transform
Both CWT and WST are continuous transform.
Nevertheless in most cases, signals to be processed are
discrete, and from fast algorithms for DWT, those for
CWT can be derived [2], hence hereinafter only the fast
algorithms of DWT are to be discussed.
DWT is implemented by a group of filters, and the
algorithms can be sorted according to the structure of
the filters into two classes. One class is of “regular”
structure, such as Mallat Straightfoward Filter Bank,
FFT fast filtering, and short-length FIR algorithm. The
other is of “irregular” structure, such as Lifting Scheme
and Integer WT, binary Qudrarature Mirror Filters
(QMF), Classical Lattice Filters and CORDIC Lattice
Filters.