Professional Documents
Culture Documents
Zhao-Swamy2013 Article TheAnalysisAndDesignOfTwo-dime
Zhao-Swamy2013 Article TheAnalysisAndDesignOfTwo-dime
Zhao-Swamy2013 Article TheAnalysisAndDesignOfTwo-dime
DOI 10.1007/s11045-011-0165-0
1 Introduction
Sub-band transforms have been successfully used in many areas of signal processing, espe-
cially in the compression of still image and video signals. This technique normally employs
analysis/synthesis multirate digital filter banks for the decomposition and reconstruction of a
signal (Vetterli and Kovacevic 1995; Vaidyanathan 1993). It has been shown that filter banks
are closely connected with the wavelet transform which is a decomposition of a function
(signal) into a set of basis functions consisting of contractions, expansions, and translations
Y. Zhao (B)
The Key Lab of Integrated Microsystems, Shenzhen Graduate School, Peking University, Xili,
518055 Shenzhen, China
e-mail: zhaoyong@szpku.edu.cn
URL: www.szpku.edu.cn
M. N. S. Swamy
Centre for Signal Processing and Communications, Department of Electrical and Computer Engineering,
Concordia University, 1455 de Maisonneuve Blvd. West, Montreal, QC H3G 1M8, Canada
123
200 Multidim Syst Sign Process (2013) 24:199–218
of a mother function, called wavelet (Vetterli and Kovacevic 1995). Lots of wavelet transform
decompositions of functions can be implemented as filter banks, and on the other hand, filter
banks can be used to generate wavelet base if they satisfy some basic conditions such as the
perfect reconstruction and regularity. For two-channel filter banks, no matter whether they
are one-dimensional or two-dimensional, there is not a single wavelet, which can be simulta-
neously symmetric (linear phase), compactly supported (FIR) and orthogonal, except for the
Harr wavelet which is not continuous (Vetterli and Kovacevic 1995). However, linear phase
(symmetry), FIR (compactly-supported), and orthogonality are all often desired in signal
processing (Vaidyanathan 1993; Antonini et al. 1992; Shapiro 1993; Villasenor et al. 1995).
The use of wavelet filter banks for image processing has been studied widely and it
has been shown to have potential for image coding (Antonini et al. 1992; Shapiro 1993;
Villasenor et al. 1995; da Silva and Ghanbari 1996; Sriram and Maicellin 1995). While most
of these developments have dealt with one dimensional signals and the multidimensional
case was handled via the tensor product, some of the more recent efforts have concentrated
on the “true” multidimensional case, from the filter bank aspect as well as the wavelet aspect
(Kovacevic and Vetterli 1997; Bamberger and Smith 1992; Chen and Vaidyanathan 1993;
Cohen and Daubechies 1993; Karlsson and Vetterli 1990; Kovacevic and Vetterli 1992; Tay
and Kingsbury 1993; Venkataraman and Levy 1994; Viscito and Allebach 1991). The main
difference, compared to the one dimensional treatment, is that multidimensional sampling
requires the use of lattices. A lattice is the set of all vectors generated by Dk, k ∈ Z2 , where
D is the matrix characterizing the sampling process and Z is the set of all integers. With non-
rectangular subsampling, more general subband shapes can be achieved. These may exhibit
directional properties. The quincunx subsampling in a 2-channel filter bank is a particularly
important nonrectangular subsampling scheme. It results in ideal subbands that are in a dia-
mond shape. The diamond subband contains significant amounts of horizontal and vertical
frequencies, but rejects most of the diagonal high frequencies. Psychophysics experiments
have shown that the human visual system is less sensitive to diagonal high spatial frequencies.
Regularity is a crucial distinction between an ordinary filter bank and a wavelet filter bank
and it has been shown in Vetterli and Kovacevic (1995), Vaidyanathan (1993), Antonini et
al. (1992), Shapiro (1993), Villasenor et al. (1995), da Silva and Ghanbari (1996) that the
vanishing moments and regularity are relevant to denoising, subband coding, etc, although
it is not necessary to maximize the regularity of a filter bank in all applications (Villasenor
et al. 1995; da Silva and Ghanbari 1996). In Basu (1998), multidimensional filter banks and
wavelets have been studied from a system theoretic perspective. A number of techniques
have been proposed to design quincunx decimation wavelet filter banks. Some of them use
the cascade form of the polyphase matrix to design the filter bank. The disadvantage is that it
is not convenient to impose a high order of zeros in the filter (KV97). The same thing holds
for the technique of using transformation of variables in Tay and Kingsbury (1993). In the
McClellan cases, however, wavelets with compact support are possible, but not orthonormal
ones (Vetterli and Kovacevic 1995). However, one-dimensional nearly-orthogonal symmetric
wavelets can be used to generate two-dimensional nearly-orthogonal nonseparable wavelets
by using McClellan transform.
The study of one-dimensional nearly-orthogonal symmetric compactly-supported wave-
let filter bank has been studied in Zhao and Swamy (2000). In Dumitrescu (2010), a two-
stage constrained optimization in the frequency domain has been used for designing 2D
nearly-orthogonal filter banks; this method is very complicated, although good passband
phase linearity is achieved. In Marshall (1997), a fast algorithm for zero-phase filter bank
and wavelet design has been proposed via decomposition of polyphase matrix. In Stanhill
and Zeevi (1998), two-dimensional orthogonal filter banks and wavelets with two-fold and
123
Multidim Syst Sign Process (2013) 24:199–218 201
four-fold symmetry have been studied, and it shows that orthogonal order-factorable wave-
lets cannot be both continuous and have four-fold symmetry. In the present paper, zeros at
aliasing frequency points are directly imposed on the two-dimensional symmetric filter, and
optimization is performed to satisfy the orthogonality condition as closely as possible. Then,
a nearly-perfect reconstructed (NPR) and nearly-orthogonal (NO) 2D FIR filter bank with
quincunx decimation can be obtained. A corresponding scaling function can be generated.
However, the question arises as to what the corresponding wavelet is, and what its degree
of orthogonality is, since the filter bank is not perfect reconstructed. First, we can construct
a semi-orthogonal 2D wavelet filter bank which is perfect reconstructed. Then, correlation
analysis is carried out on this semi-orthogonal filter bank. Both theoretical and experimental
analyses are given, which show that the semi-orthogonal filter bank is nearly-orthogonal. We
will also show that the semi-orthogonal filter is very close to the nearly-orthogonal FIR filter
bank in the sense that the corresponding filters in the two filter banks have the exact phase
response and almost the same amplitude response. The FIR NPR filter bank is regarded as
an implementation of the corresponding semi-orthogonal filter bank.
It is known that there is a close relationship between filter banks and wavelets. A lot of discrete
wavelet transforms can be implemented as filter banks, and perfect reconstructed filter banks
may generate wavelets, provided that they satisfy some regularity condition. The condition
for a filter bank to be able to generate wavelet bases is very complicated, as indicated in
Chui (1992). However, a necessary condition that a filter can be used to generate a scaling
function is that it must have some zeros at the point z = −1. If this condition is satisfied, the
filter bank may be capable of generating a multiresolution analysis and a wavelet basis. The
one-dimensional case has been studied in Zhao and Swamy (2000).
A quincunx two-channel filter bank is shown in Fig. 1, where G(z 1 , z 2 ), P(z 1 , z 2 ),
H (z 1 , z 2 ), and Q(z 1 , z 2 ) are the analysis low-pass filter, synthesis low-pass filter, analy-
sis high-pass filter and the synthesis high-pass filter, respectively, and D is the quincunx
decimation, which can be characterized by the matrix
1 1
D= (1)
1 −1
123
202 Multidim Syst Sign Process (2013) 24:199–218
where ω = (ω1 , ω2 ) and N(DT ) denotes the set of all integer vectors in DT a, a ∈ [0, 1)2 .
The quincunx expanded version of x(k) is defined as
x(D−1 k), if D−1 k ∈ Z2
y(k) = (3)
0, otherwise
Hence,
Y (ω) = X (DT ω) (4)
From (2) and (4), the output can be obtained as follows
1
y(z 1 , z 2 ) = [P(z 1 , z 2 )G(z 1 , z 2 ) + Q(z 1 , z 2 )H (z 1 , z 2 )]x(z 1 , z 2 )
2
1
+ [P(z 1 , z 2 )G(−z 1 , −z 2 ) + Q(z 1 , z 2 )H (−z 1 , −z 2 )]x(−z 1 , −z 2 ) (5)
2
A necessary condition
√ for a PR filter bank to generate a wavelet basis is that P(−1, −1) = 0
and P(1, 1) = 2, implying that there must be some zeros at the aliasing frequency of
(π, π). If a multiresolution analysis and wavelet basis exist, then a scaling function can be
defined by the product
1
∞
T −i
ϕ (ω) = √ P e j (D ) ω (6)
i=1
2
where ϕ is a scaling function generated by P(z 1 , z 2 ) with the above equation, and ϕ is its
Fourier transform. We have the following dilation equations:
√
ϕ(t) = 2 p(k) ϕ(Dt − k) (7)
k∈Z 2
√
ψ(t) = 2 q(k) ϕ(Dt − k) (8)
k∈Z 2
Since the nearly-orthogonal filter banks are symmetric, we can use McClellan transform to
obtain the nonseparable two-dimensional nearly-orthogonal filter banks, i.e., we use
cos ω1 + cos ω2
cos ω = (11)
2
to transform all the one-dimensional nearly-orthogonal symmetric filters into 2D nonsepara-
ble symmetric near-orthogonal wavelet filter banks with quincunx decimation. Another way
123
Multidim Syst Sign Process (2013) 24:199–218 203
123
204 Multidim Syst Sign Process (2013) 24:199–218
π π 2
C(x) = |P(e jω1 , e jω2 )|2 + |P(e j (ω1 +π ) , e j (ω2 +π ) )|2 − 2 dω1 dω2 (17)
0 0
The integral can be approximated by using the average sum over a two-dimensional sampled
frequency. This technique involves substantial computation for numerical integral in this cost
function. Hence, we propose another cost function by comparing the coefficients on the two
sides of (16).
Let S0 be the matrix of the coefficients of the polynomials in the first portion in (12)
⎛ ⎞
0 0.125 0
S0 = ⎝ 0.125 0.5 0.125 ⎠ (18)
0 0.125 0
and Sn be matrix of the polynomial of Sn (z 1 , z 2 ). Then, the matrix P characterizing the filter
P(z 1 , z 2 ) can be obtained by using the 2D convolution
P = (S0 ∗ · · · ∗S0 ) ∗Sn (19)
m
where the operation ∗ is the matrix convolution operation, since the matrix corresponding to
the product of two polynomials is equal to the convolution of the matrix corresponding to the
two polynomials. Let W be a diagonal matrix of the same of dimension as that of P, with all
its entries being zero except for the diagonal entries which are alternately +1 and −1 with
the central entry being +1. Then the matrix corresponding to P(−z1 , −z2 ) is WPW, and the
matrix P F with a superscript of F is the matrix obtained from P by flipping its rows up-side
down and its columns left-side right; the matrix P F corresponds to P(z−1 −1
1 , z2 ). Then, the
cost function can be defined as follows
L(x) = P∗P F + (WPW)∗(WPW) F − 2I0 (20)
where |X | is the norm of a matrix or a vector X , which is defined here as the largest singular
value of X , or the square root of the largest eigenvalue of the positive-semidefinite matrix
X ∗ X , where X ∗ denotes the conjugate transpose of X , and I0 is a matrix of the size of P∗P F
with all entries being zero, except the central entry being 1. The entries of Sn are dependent
on the coefficients x of the polynomials of Sn (z 1 , z 2 ), and hence L is a cost function of those
coefficients x. This cost function involves only matrix manipulation and the computation is
much less than that of the cost function given by (17).
Usually, the cost function L(x) is not a globally convex nonlinear function. Therefore,
in order to find the global minimum, a proper choice of the initial point is essential. We
can find the corresponding one-dimensional nearly-orthogonal filter, then use the McClellan
transform to get the initial point. The initial point is just an estimate close to the optimal
point and is set to be the starting point for searching the optimal, the cost function could be
substantially more reduced from the initial point.
We then minimize (20) using the Simplex downhill technique. Experiments for m = 1 : 3
and n = 2 : 5 are performed. We can the check the amplitude distortion of the filter bank
derived from P(z 1 , z 2 ), which is:
max |P(e jω1 , e jω2 )|2 + |P(e j (ω1 +π ) , e j (ω2 +π ) )|2 − 2
ω1 ,ω2
The maximum amplitude distortion in decibels is given in Table 1. Usually, if the maximum
distortion is less than −40 decibels, then the distortion is less than 1% (0.01), and the recon-
struction error is acceptable for many applications. However, if it is greater than that, then
123
Multidim Syst Sign Process (2013) 24:199–218 205
1.5
0.5
0
1
0.5 1
0 0.5
0
−0.5 −0.5
−1 −1
there might be a problem. Hence, except for the case of m = 3 and n = 2, for which the
amplitude distortion is too large, the others may be useful for some practical applications.
The amplitude response of the filter with four zeros at the aliasing frequency and two-
ring remainder symmetric polynomial (m = 2, n = 2) is shown in Fig. 2, with normalized
frequency. The amplitude distortion is shown in Fig. 3.
The amplitude distortion of this filter is relatively large compared to that of the other filters
(except when m = 3, n = 2, which is useless). All the other filters in Table 1 have very
small distortions. For larger n, we can obtain a much smaller distortion for orthogonality and
amplitude. Taking m = 2 for example, the deviation reduces from −45.39 dB for n = 2 to
−103.19 dB for n = 5. In applications that require a high precision and there is less emphasis
on computation, we can use a relatively large n to achieve filters that are more orthogonal.
With P(z 1 , z 2 ), a two-dimensional nearly-orthogonal filter bank can be obtained as fol-
lows
Pno (z 1 , z 2 ) = P(z 1 , z 2 ) (21)
G no (z 1 , z 2 ) = P(z 1−1 , z 2−1 ) (22)
Q no (z 1 , z 2 ) = −z 1−1 P(−z 1−1 , −z 2−1 ) (23)
Hno (z 1 , z 2 ) = −z 1 P(−z 1 , −z 2 ) (24)
It is easy to verify that the aliasing is completely cancelled and it is nearly perfect recon-
structed.
123
206 Multidim Syst Sign Process (2013) 24:199–218
−40
−50
−60
dB
−70
−80
−90
−100
1
0.5 1
0 0.5
0
−0.5 −0.5
−1 −1
Consider the convergence of the product of (6) to the Fourier transform of a scaling function.
It is equivalent to consider the convergence of the recursion of the following function.
ϕi (t1 , t2 ) = p(k1 , k2 )ϕi−1 (t1 + t2 − k1 , t1 − t2 − k2 ) (25)
k1 ,k2
Hence we can just examine the convergence of the function ϕi at these grid points only. The
values of ϕi−1 at these grid points are computed and stored for computing ϕi . In this way,
the computation increases only linearly with respect i, thus allowing us to calculate for large
i. For example, consider m = 2, n = 2 (fourth order zeros at the aliasing frequency and
2 ring remainder filter); Fig. 4 shows the process of convergence to the scaling function, in
which ϕi with i = 3, i = 10, i = 20 and i = 50 are displayed. We see that ϕi converges
gradually as i increases. The scaling function for m = n = 2 at the 100th recursion is
shown in Fig. 5. However, for the case of m = 1, n = 2 (with only two zeros at the aliasing
frequency and two ring remainder filter), the recursion of the scaling function of the filter
shows that unfortunately the scaling function does not converge. The maximum of value of
123
Multidim Syst Sign Process (2013) 24:199–218 207
ϕi for i = 3 is 2.0655, for i = 10, it is 6.3837, for i = 50, it becomes 21.7662 and for
i = 1, 000, it hikes to 1.0356e+05. From the iteration process, i.e., when i = 3, 10, 20 and
50, we can see from Fig. 4 that the graph gradually converges; however, as seen from Fig. 5,
the graph becomes a pulse in the center, and gradually goes to ∞ at the center and becomes 0
elsewhere. Figure 5 shows that the maximum value of ϕi increases quickly as i increases, and
hence it does not converge. Therefore, in the two-dimensional case, more zeros are needed
at the aliasing frequency to guarantee the convergence.
One way to obtain a perfect reconstructed filter bank is to construct a semi-orthogonal filter
bank in a manner similar to that in the one dimensional case (Chui 1992). A correlation
polynomial can be defined as
E(z 1 , z 2 ) = r (k1 , k2 )z 1−k1 z 2−k2 (28)
k1 ,k2
where
∞ ∞
r (k1 , k2 ) = φ(t1 , t2 ) φ(t1 + k1 , t2 + k2 ) dt1 dt2 (29)
−∞ −∞
123
208 Multidim Syst Sign Process (2013) 24:199–218
An important relation between E(z 1 , z 2 ) and P(z 1 , z 2 ) can be derived from the above equa-
tion by using (9)
2E(e j (ω1 +ω2 ) , e j (ω1 −ω2 ) ) = 2 |ϕ̂(ω1 + ω2 + 2k1 π, ω1 − ω2 + 2k2 π )|2
k1 ,k2
2π(k1 +k2 ) 2π(k1 −k2 ) 2π(k1 + k2 ) 2π(k1 − k2 ) 2
= |P e j ω1 + 2 , e j ω2 + 2 |2 |ϕ̂ ω1 + , ω2 + |
2 2
k1 ,k2
k1 −k
2 =2n 2
= |P(e j (ω1 +2π n 1 ) , e j (ω2 +2π n 2 ) )|2 |ϕ̂(ω1 + 2π n 1 , ω2 + 2π n 2 )|2
k1 +k2 =2n 1
k1 −k2
=2m 2 +1
∧
+ |P(e j (ω1 +2π m 1 +π ) , e j (ω2 +2π m 2 +π ) )|2 | ϕ (ω1 + 2π m 1 + π, ω2 + 2π m 2 + π )|2
k1 +k2 =2m 1 +1
(35)
Since k1 , k2 can assume any integer value, so can n 1 , n 2 , m 1 , m 2 . Hence, the above summa-
tion becomes
123
Multidim Syst Sign Process (2013) 24:199–218 209
2E(e j (ω1 +ω2 ) , e j (ω1 −ω2 ) ) = |P(e jω1 , e jω2 )|2 |ϕ̂(ω1 + 2n 1 π, ω2 + 2n 2 π)|2
n 1 ,n 2
+ |P(−e jω1
, −e jω2
)|2 |ϕ̂(ω1 + 2m 1 π + π, ω2 + 2m 2 π + π)|2
m 1 ,m 2
= |P(e jω1 , e jω2 )|2 E(e jω1 , e jω2 ) + |P(−e jω1 , −e jω2 )|2 E(−e jω1 , −e jω2 ) (36)
2E(z 1 z 2 , z 1 z 2−1 )
= P(z 1 , z 2 )P(z 1−1 , z 2−1 )E(z 1 , z 2 ) + P(−z 1 , −z 2 )P(−z 1−1 , −z 2−1 )E(−z 1 , −z 2 )
(37)
Using this equation, it is easy to verify that the semi-orthogonal wavelet filter bank is perfect
reconstructed. By using (5), (30–33), we have
1
y(z 1 , z 2 ) = [PS O (z 1 , z 2 )G S O (z 1 , z 2 ) + Q S O (z 1 , z 2 )HS O (z 1 , z 2 )]x(z 1 , z 2 )
2
1
+ [PS O (z 1 , z 2 )G S O (−z 1 , −z 2 ) + Q S O (z 1 , z 2 )HS O (−z 1 , −z 2 )]x(−z 1 , −z 2 )
2
1 P(z 1−1 , z 2−1 )E(z 1 , z 2 )
= P(z 1 , z 2 )
2 E(z 1 z 2 , z 1 z 2−1 )
−z 1 P(−z 1 , −z 2 )
+ (−z 1−1 )E(−z 1 , −z 2 )P(−z 1−1 , −z 2−1 ) x(z 1 , z 2 )
E(z 1 z 2 , z 1 z 2−1 )
1 P(−z 1−1 , −z 2−1 )E(−z 1 , −z 2 )
+ p(z 1 , z 2 )
2 E(z 1 z 2 , z 1 z 2−1 )
z 1 P(z 1 , z 2 )
+ (−z 1−1 )E(−z 1 , −z 2 )P(−z 1−1 , −z 2−1 ) x(−z 1 , −z 2 )
E(z 1 z 2 , z 1 z 2−1 )
1 P(z 1 , z 2 )P(z 1−1 , z 2−1 )E(z 1 , z 2 ) + P(−z 1 , −z 2 )P(−z 1−1 , −z 2−1 )E(−z 1 , −z 2 )
= x(z 1 , z 2 )
2 E(z 1 z 2 , z 1 z 2−1 )
1 P(z 1 , z 2 )P(−z 1−1 , −z 2−1 )E(−z 1 , −z 2 )−P(z 1 , z 2 )E(−z 1 , −z 2 )P(−z 1−1 , −z 2−1 )
+ x(−z 1 , −z 2 )
2 E(z 1 z 2 , z 1 z 2−1 )
Note that the second part of the above equation is equal to zero. From (37), we have:
y(z 1 , z 2 )
1 P(z 1 , z 2 )P(z 1−1 , z 2−1 )E(z 1 , z 2 )+P(−z 1 ,−z 2 )P(−z 1−1 , −z 2−1 )E(−z 1 , −z 2 )
= x(z 1 , z 2 )
2 E(z 1 z 2 , z 1 z 2−1 )
1 2E(z 1 z 2 , z 1 z 2−1 )
= = x(z 1 , z 2 )
2 E(z 1 z 2 , z 1 z 2−1 )
Hence, the semi-orthogonal filter bank is perfect reconstructed. With this PR semi-orthogonal
filter bank, a wavelet function can be obtained by using the two-scale function. The wavelet
corresponding to the scaling function depicted in Fig. 6 is shown in Fig. 7.
123
210 Multidim Syst Sign Process (2013) 24:199–218
First, we consider the correlation of the scaling function. Similar to the one dimensional case,
we have (see Appendix)
ϕ̂(0, 0) = 1 (38)
φ̂(2πk1 , 2πk2 ) = 0, k1 = 0 or k2 = 0 (39)
123
Multidim Syst Sign Process (2013) 24:199–218 211
Hence
E(1, 1) = 1 (40)
From (37) and (40), E(z 1 , z 2 ) can be found. Assume that
N
−j
P(z 1 , z 2 ) = pi j z 1−i z 2 (41)
i, j=−N
2N
−j
B(z 1 , z 2 ) = bi j z 1−i z 2 (42)
i, j=−2N
where B(z 1 , z 2 ) = P(z 1 , z 2 )P(z 1−1 , z 2−1 ), and B = P∗P , where B is the matrix correspond-
ing to B(z 1 , z 2 ). Since the support of φ(t1 , t2 ) is within [−N , N ] × [−N , N ], E(z 1 , z 2 ) can
be written in the following form according to (28)
2N
−j
E(z 1 , z 2 ) = ri j z 1−i z 2 (43)
i, j=−2N
123
212 Multidim Syst Sign Process (2013) 24:199–218
−20
−40
dB
−60
−80
−100
5 8
6
0 4
2
0
−2
−5 −4
−6
−8
we get
Ae = a (54)
and
123
Multidim Syst Sign Process (2013) 24:199–218 213
Define the correlation polynomial of the wavelet in the 2D semi-orthogonal wavelet filter
bank as
E ψ (z 1 , z 2 ) = rψ (k1 , k2 )z 1−k1 z 2−k2 (56)
k1 ,k2
where
∞ ∞
rψ (k1 , k2 ) = ψ(t1 , t2 )ψ(t1 + k1 , t2 + k2 ) dt1 dt2 (57)
−∞ −∞
Using a technique similar to that used in obtaining Eq. (37), and following the method of
derivation of the correlation polynomial of the wavelet in the one-dimensional case, we have
2E ψ (e j (ω1 +ω2 ) , e j (ω1 −ω2 ) ) = 2 |ψ̂(ω1 + ω2 + 2k1 π, ω1 − ω2 + 2k2 π)|2
k1 ,k2
2π(k1 +k2 )
2π(k1 −k2 ) 2π(k1 + k2 ) 2π(k1 − k2 ) 2
= |Q e j ω1 + 2 , e j ω2 + 2 |2 |ϕ̂ ω1 + , ω2 + |
2 2
k1 ,k2
k1 −k
2 =2n 2
= |Q(e j (ω1 +2πn 1 ) , e j (ω2 +2πn 2 ) )|2 |ϕ̂(ω1 + 2πn 1 , ω2 + 2πn 2 )|2
k1 +k2 =2n 1
k1 −k2
=2m 2 +1
+ |Q(e j (ω1 +2πm 1 +π) , e j (ω2 +2πm 2 +π) )|2 |ϕ̂(ω1 +2πm 1 +π, ω2 +2πm 2 + π)|2
k1 +k2 =2m 1 +1
= |Q(e jω1 , e jω2 )|2 |ϕ̂(ω1 + 2πn 1 , ω2 + 2πn 2 )|2
n 1 ,n 2
+ |Q(−e jω1 , −e jω2 )|2 |ϕ̂(ω1 + π + 2πm 1 , ω2 + π + 2πm 2 )|2
m 1 ,m 2
= |Q so (e jω1 , e jω2 )|2 E(e jω1 , e jω2 ) + |Q so (−e jω1 , −e jω2 )|2 E(−e jω1 , −e jω2 ) (58)
E ψ (e jω1 , e jω2 )=E(e j (ω1 +ω2 )/2 , e j (ω1 −ω2 )/2 )E(−e j (ω1 +ω2 )/2 , −e j (ω1 −ω2 )/2 )E(e jω1 , e jω2 )
(59)
where ε is the maximum absolute difference between E(e jω1 , e jω2 ) and 1, which is usually
a very small value for the semi-orthogonal filter banks designed above. Then we have
Similarly, consider
E ϕ,ψ (z 1 , z 2 ) = rϕ,ψ (k1 , k2 )z 1−k1 z 2−k2 (62)
k1 ,k2
123
214 Multidim Syst Sign Process (2013) 24:199–218
From the results in Table 2 we see that E(z 1 , z 2 ) ≈ 1. Therefore the semi-orthogonal filter
bank can be approximately implemented as the corresponding nearly-orthogonal filter bank,
since
Pso (z 1 , z 2 ) = P(z 1 , z 2 ) = Pno (z 1 , z 2 ) (71)
Q so (z 1 , z 2 ) = −z 1−1 E(−z 1 , −z 2 )P(−z 1−1 , −z 2−1 )
≈ −z 1−1 P(−z 1−1 , −z 2−1 ) = Q no (z 1 , z 2 ) (72)
G so (z 1 , z 2 ) = P(z 1−1 , z 2−1 )E(z 1 , z 2 )/E(z 1 z 2 , z 1 z 2−1 )
≈ P(z 1−1 , z 2−1 ) = G no (z 1 , z 2 ) (73)
Hso (z 1 , z 2 ) = −z 1 P(−z 1 , −z 2 )/E(z 1 z 2 , z 1 z 2−1 )
≈ −z 1 P(−z 1 , −z 2 ) = Hno (z 1 , z 2 ) (74)
123
Multidim Syst Sign Process (2013) 24:199–218 215
0
−50
−100
−150
dB
−200
−250
−300
−350
1
0.5 1
0 0.5
0
−0.5 −0.5
−1 −1
Fig. 9 The absolute difference in dB between the analysis lowpass filters of the semi-orthogonal filter bank
and that of the corresponding nearly-orthogonal filter bank, with normalized frequency
There is no phase difference between these filters and their associate approximate filters,
since E(z 1 , z 2 ) has zero phase. There is only a small amplitude difference. For larger n, the
difference will be even much smaller. The difference between the analysis lowpass filters for
the case m = 2, n = 5 is shown in Fig. 9. From these figures, we see that the approximate
filter is a very good approximate implementation of the original filter.
5 Conclusion
123
216 Multidim Syst Sign Process (2013) 24:199–218
perfect reconstructed filter bank. We hereby construct a semi-orthogonal filter bank. Correla-
tion analysis made on the semi-orthogonal filter bank shows that it is nearly-orthogonal. The
integer translates of the semi-orthogonal wavelet at the same scale are nearly-orthogonal to
one another. The integer translates of the wavelet at different scales are completely orthog-
onal to each other. Theoretical analysis and experiments have shown that the filter bank is
very close to the nearly-orthogonal FIR filter bank if the filter bank is very close to orthog-
onal. The semi-orthogonal wavelet filter bank can be implemented with the corresponding
nearly-orthogonal FIR filter bank, in which every filter in the nearly-orthogonal FIR filter
bank is a very good approximation and implementation of the corresponding filter in the
semi-orthogonal filter bank.
Appendix
√
If P(1, 1) = 2, P(−1, −1) = 0 and the product in (9) converges, then (38) and (39) hold.
√
Proof Equation (38) obviously holds since P(1, 1) = 2. Rewrite the product of φ̂(ω) as
1 j Di ω
∞ ∞
1 T −i
ϕ̂(ω) = √ P(e j (D ) ω ) = √ P e 2i
i=1
2 i=1
2
1 j D2n ω 1 j D2n+1
2n+1
∞ ∞
ω
= √ P e 22n √ P e 2
i=1
2 i=0
2
1 ω1 ω2 1 ω1 +ω2 ω −ω
∞ ∞
j j 1 2
= √ P e j 2n , e j 2n √ P e 2n+1 , e 2n+1
i=1
2 i=0
2
where the product has been split into an odd term product and an even term product. Hence
(1). If k1 , k2 are simultaneously odd, the term of n = 1 in the first product will be
P(−1, −1) = 0, hence the total product is equal to zero.
(2). If k1 , k2 are not simultaneously even or odd, then k1 + k2 , k1 − k2 will be both odd,
the term of n = 0 in the second product will be zero.
(3). If k1 , k2 are both even, then it can be written in the following form
Hence, the term of m = n in the second product will be zero. Therefore, for all integer values
of k1 , k2 , (39) holds.
123
Multidim Syst Sign Process (2013) 24:199–218 217
References
Adelson, E. H., Simoncelli, E., & Hingorani, R. (1987). Orthogonal pyramid transform for image coding.
In Proceedings of SPIE (Vol. 845, pp. 50–58). Cambridge, MA.
Antonini, M., Barlaud, M., Mathieu, P., & Daubechies, I. (1992). Image coding using wavelet transform. IEEE
Transactions on Image Processing, 1(2), 205–220.
Bamberger, R., & Smith, M. (1992). A filter bank for the directional decomposition of images: Theory
and design. IEEE Transactions on Signal Processing, 40(4), 882–893.
Basu, S. (1998). Multidimensional filter banks and wavelets, a system theoretic perspective. Journal of
Franklin Institute, 335B(8), 1367–1409.
Chen, T., & Vaidyanathan, P. (1993). Multidimensional multirate filters and filter banks derived from
one-dimensional filters. IEEE Transactions on Signal Processing, 41(5), 1749–1765.
Chui, C. K. (1992). An introduction to wavelets. New York: Academic Press.
Cohen, A., & Daubechies, I. (1993). Non-separable bidimensional wavelet bases. Revista Matemd́ftica
Iberoamericana, 9(1), 51–138.
Crochiere, R. E., & Rabiner, R. L. (1983). Multirate digital signal processing. Englewood Cliffs,
NJ: Prentice-Hall.
da Silva, E. A. B., & Ghanbari, M. (1996). On the performance of linear phase wavelet transforms in
low bit-rate image coding. IEEE Transactions on Image Processing, 5(5), 689–704.
Daubechies, I. (1986). Orthogonal bases of compactly supported wavelets. Communications on Pure and
Applied Mathematics, XLI, 909–996.
Dumitrescu, B. (2010). A moulding technique for the design of 2-D nearly orthogonal filter banks. IEEE
Signal Processing Letters, 17(3).
Johnston, J. D. (1980). A filter family design for use in quadrature mirror filter banks. In Proceedings of
ICASSP (pp. 291–294).
Karlsson, G., & Vetterli, M. (1990). Theory of two- dimensional multirate filter banks. IEEE Transactions
on Acoustic, Speech, Signal Processing, 38(6), 925–937.
Kovacevic, J., & Vetterli, M. (1992). Non-separable multidimensional filter banks and wavelet bases for
R n . IEEE Transactions on Information Theory, 38, 533–555.
Kovacevic, J., & Vetterli, M. (1997). Nonseparable two- and three-dimensional wavelets. IEEE Transactions
on Signal Processing, 43(5), 1269–1273.
Mallat, S. (1989). A theory for multiresolution signal decomposition: The wavelet representation. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 11(7), 674–693.
Marshall, T. G. (1997). Zero-phase filter bank and wavelet code r matrices: Properties, triangular decom-
positions, and a fast algorithm. Multidimensional Systems and Signal Processing, 8(1–2), 71–88.
Shapiro, J. M. (1993). Embedded image coding using zerotrees of wavelet coefficients. IEEE Transactions
on Signal Processing, 41(12), 3445–3462.
Simoncelli, E., & Adelson, E. (1990). Non-separable extensions of quadrature mirror filters to multiple
dimensions. Proceedings of IEEE, 78(4), 652–663.
Sriram, P., & Maicellin, M. W. (1995). Image coding using wavelet transforms and entropy-constrained
trellis-coded quantization. IEEE Transactions on Image Processing, 4(6), 725–733.
Stanhill, D., & Zeevi, Y. Y. (1998). Two-dimensional orthogonal filter banks and wavelets with linear
phase. IEEE Transactions on Signal Processing, 46(1), 183–190.
Strang, G. (1989). Wavelet and dilation equation. SIAM Journal of Mathmatics Analysis, 31, 614–627.
Tay, D., & Kingsbury, N. (1993). Flexible design of multidimensional perfect reconstruction FIR 2-band
filters using transformations of variables. IEEE Transactions on Image Processing, 2, 466–480.
Vaidyanathan, P. P. (1993). Multirate systems and filter banks. Englewood Cliffs, New Jersey: Prentice
Hall PTR.
Venkataraman, S., & Levy, B. (1994). State space representations of 2-D FIR lossless transfer matrices. IEEE
Transactions on Circuits System II: Analog and Digital Signal Processing, 41(2), 117–132.
Vetterli, M., & Kovacevic, J. (1995). Wavelets and subband coding. Englewood Cliffs, New Jersey: Prentice
Hall PTR.
Villasenor, J. D., Belzer, B., & Liao, J. (1995). Wavelet filter evaluation for image compression. IEEE
Transactions on Image Processing, 4(8), 1053–1060.
Viscito, E., & Allebach, J. (1991). The analysis and design of multidimensional FIR perfect recon-
struction filter banks for arbitrary sampling lattice. IEEE Transactions on Circuits System (Video
Technol.), 38, 29–42.
Zhao, Y., & Swamy, M. N. S. (1999). A technique for designing new biorthogonal filters and its application
to image compression. Electronics Letters, 35(18).
123
218 Multidim Syst Sign Process (2013) 24:199–218
Zhao, Y., & Swamy, M. N. S. (2000). New technique for designing nearly-orthogonal wavelet filter banks
with linear phase. IEE Proceeding on Vision, Image and Signal Processing, 147(6), 527–533.
Author Biographies
Yong Zhao graduated from Southeast University, China, in 1991
with Ph.D. awarded. He joined Zhejiang University as a faculty
member in 1991. He went to Concordia University, Canada as a
Post-Doctroal Fellow in 1997. Then he was a senior Audio/Video
compression Engineer for Honeywell in Ottawa, Canada, in 2000.
He is currently with Peking University, Shenzhen Graduate School.
His main research fields are video compression, video analytics, and
signal processing.
123