Quadrature-Mirror Filter Bank

Quadrature-Mirror Filter Bank
In many applications, a discrete-time signal

x[n] is split into a number of subband signals
by
{vk means
[n]} of an analysis filter bank
The subband signals are then processed
Finally, the processed subband signals are
combined by a synthesis filter bank resulting
in an output signal y[n]
1
Copyright S. K. Mitra

If the subband signals {vk [n]}are
bandlimited to frequency ranges much
smaller than that of the original input signal
x[n], they can be down-sampled before
processing
Because of the lower sampling rate, the
processing of the down-sampled signals can
be carried out more efficiently
2

After processing, these signals are then upsampled before being combined by the
synthesis filter bank into a higher-rate
signal
The combined structure is called a
quadrature-mirror filter (QMF) bank

If the down-sampling and up-sampling
factors are equal to or greater than the
number of bands of the filter bank, then the
output y[n] can be made to retain some or
all of the characteristics of the input signal
x[n] by choosing appropriately the filters in
the structure
4

If the up-sampling and down-sampling
factors are equal to the number of bands,
then the structure is called a critically
sampled filter bank
The most common application of this
scheme is in the efficient coding of a signal
x[n]
5
Two-Channel QMF Bank

Figure below shows the basic two-channel
QMF bank-based subband codec
(coder/decoder)

The analysis filters H 0 (z ) and H1(z ) have
typically a lowpass and highpass frequency
responses, respectively, with a cutoff at /2

Each down-sampled subband signal is
encoded by exploiting the special spectral
properties of the signal, such as energy
levels and perceptual importance
It follows from the figure that the sampling
rates of the output y[n] and the input x[n]
are the same
8

The analysis and the synthesis filters are
chosen so as to ensure that the reconstructed
output y[n] is a reasonably close replica of
the input x[n]
Moreover, they are also designed to provide
good frequency selectivity in order to
ensure that the sum of the power of the
subband signals is reasonably close to the
input signal power
9

In practice, various errors are generated in
this scheme
In addition to the coding error and errors
caused by transmission of the coded signals
through the channel, the QMF bank itself
introduces several errors due to the sampling
rate alterations and imperfect filters
We ignore the coding and the channel errors
10

We investigate only the errors caused by the
sampling rate alterations and their effects on
the performance of the system
To this end, we consider the QMF bank
structure without the coders and the
decoders as shown below
11
Analysis of the Two-Channel

QMF Bank
Making use of the input-output relations of
the down-sampler and the up-sampler in the
z-domain we arrive at
Vk ( z ) H k ( z ) X ( z ),
1
1/ 2
1/ 2
U k ( z ) {Vk ( z ) Vk ( z )},
2
^
2
Vk ( z ) U k ( z )
12
k = 0, 1

QMF Bank
From the first and the last equations we
obtain after some algebra
^
1
Vk ( z ) {Vk ( z ) Vk ( z )}
2
1 {H k ( z ) X ( z ) H k ( z ) X ( z )}
2
13
The reconstructed output of the filter bank

is given by
^
^
Y ( z ) G0 ( z )V0 ( z ) G1( z )V1( z )

QMF Bank
From the two equations of the previous
slide we arrive at
1
Y ( z ) {H 0 ( z )G0 ( z ) H1( z )G1( z )}X ( z )
2
1
{H 0 ( z )G0 ( z ) H1( z )G1( z )}X ( z )
2
The second term in the above equation is

due to the aliasing caused by sampling rate
alteration
14

QMF Bank
The input-output equation of the filter bank
can be compactly written as
Y ( z ) T ( z ) X ( z ) A( z ) X ( z )
where T(z), called the distortion transfer
function,
function is given by
1
T ( z ) {H 0 ( z )G0 ( z ) H1( z )G1( z )}
2
15
and
1
A( z ) {H 0 ( z )G0 ( z ) H1( z )G1( z )}
2
Alias-Free Filter Bank

Since the up-sampler and the down-sampler
are linear time-varying components, in
general, the 2-channel QMF structure is a
linear time-varying system
It can be shown that the 2-channel QMF
structure has a period of 2
However, it is possible to choose the
analysis and synthesis filters such that the
aliasing effect is canceled resulting in a
time-invariant operation
16

To cancel aliasing we need to ensure that
A(z) = 0, i.e.,
H 0 ( z )G0 ( z ) H1( z )G1( z ) 0
For aliasing cancellation we can choose
G0 ( z )
H1 ( z )
G1 ( z )
H 0 ( z )
This yields
G0 ( z ) C ( z ) H1( z ), G1( z ) C ( z ) H 0 ( z ),
where C(z) is an arbitrary rational function
17

If the above relations hold, then the QMF
system is time-invariant with an inputoutput relation given by
Y(z) = T(z)X(z)
where
T ( z ) 1 {H 0 ( z ) H1( z ) H1( z ) H 0 ( z )}
2
On the unit circle, we have

j
j
j
j
j ( )
j
Y (e ) T (e ) X (e ) |T (e ) | e
X (e )
18

If T(z) is an allpass function, i.e., | T (e
with d 0 then
19
)| d
|Y (e j )| d | X (e j )|
indicating that the output of the QMF bank
has the same magnitude response as that of
the input (scaled by d) but exhibits phase
distortion
The filter bank is said to be magnitude
preserving

If T(z) has linear phase, i.e.,
arg{T (e
)} ( )
then
j
j
arg{Y (e )} arg{X (e )}
The filter bank is said to be phasepreserving but exhibits magnitude distortion
20

If an alias-free filter bank has no magnitude
and phase distortion, then it is called a
perfect reconstruction (PR) QMF bank
In such a case, T ( z ) dz resulting in
Y ( z ) dz X ( z )
In the time-domain, the input-output
relation for all possible inputs is given by
y[n] d x[n ]
21

Thus, for a perfect reconstruction QMF
bank, the output is a scaled, delayed replica
of the input
Example - Consider the system shown below
x[n ]
2
22
2
2
y[n ]

Comparing this structure with the general
QMF bank structure we conclude that here
we have
1
1
H 0 ( z ) 1, H1( z ) z , G0 ( z ) z , G1( z ) 1
Substituting these values in the expressions
for T(z) and A(z) we get
T ( z ) 1 ( z 1 z 1 ) z 1
A( z )
23
2
1
1 1
(z z ) 0
2

Thus the simple multirate structure is an
alias-free perfect reconstruction filter bank
However, the filters in the bank do not
provide any frequency selectivity
24
An Alias-Free Realization
A very simple alias-free 2-channel QMF bank
is obtained when
H1( z ) H 0 ( z )
The above condition, in the case of a real
coefficient filter, implies
j
j ( )
| H1(e )| | H 0 (e
)|
indicating that if H 0 (z ) is a lowpass filter,
then H1(z ) is a highpass filter, and vice versa
25
The relation | H1(e j )| | H 0 (e j ( ) )|
j
|
H
(
e
)| is a mirror-image
indicates that 1
j
|
H
(
e
)| with respect to /2, the
of 0
quadrature frequency
This has given rise to the name quadraturemirror filter bank
26
Substituting H1( z ) H 0 ( zin)
G0 ( z ) C ( z ) H1( z ), G1( z ) C ( z ) H 0 ( z ),
with C(z) = 1 we get
G0 ( z ) H1( z ), G1( z ) H1( z ) H 0 ( z )
The above equations imply that the two

analysis filters and the two synthesis filters are
essentially determined from one transfer
function
27
H 0 (z )
Moreover, if H 0 (z ) is a lowpass filter, then
G0 (z ) is also a lowpass filter and G1(z ) is a
highpass filter
The distortion function in this case reduces
to
T ( z ) 1 {H 02 ( z ) H12 ( z )} 1 {H 02 ( z ) H 02 ( z )}
2
28
A computationally efficient realization of
the above QMF bank is obtained by
realizing the analysis and synthesis filters in
polyphase form
Let the 2-band Type 1 polyphase
representation of H 0 (z ) be given by
2
1
2
H 0 ( z ) E0 ( z ) z E1( z )
29
Then from the relation H1( z ) H 0 ( z ) it
follows that
2
1
2
H1( z ) E0 ( z ) z E1( z )
Combining the last two equations in a
matrix form we get
2
H
(
z
)
E
(
z
)
0 1 1
0
H 0 ( z ) 1 1 z 1E ( z 2 )
1
30
Likewise, the synthesis filters can be
expressed in a matrix form as
1
2
2 1 1
G0 ( z ) G1( z ) z E1( z ) E0 ( z ) 1 1
Making use of the last two equations we

can redraw the two-channel QMF bank as
shown below
31
Making use of the cascade equivalences, the
above structure can be further simplified as
shown below
32
Substituting the polyphase representations
of the analysis filters we arrive at the
expression for the distortion function T(z) in
terms of the polyphase components as
1
T ( z ) 2 z E0 ( z ) E1( z )
33
1
Example - Let H 0 ( z ) 1 z
Its polyphase components are
2
2
E0 ( z ) 1, E1( z ) 1
Hence
2
1
2
1
H1( z ) H 0 ( z ) E0 ( z ) z E1( z ) 1 z
Likewise
1
2
2
1
G0 ( z ) z E1( z ) E0 ( z ) 1 z
1
G1( z ) z E1( z ) E0 ( z ) 1 z
34
The distortion transfer function for this
realization is thus given by
1
T ( z ) 2 z E0 ( z ) E1( z ) 2 z
35
The resulting structure is a perfect

reconstruction QMF bank
Alias-Free FIR QMF Bank

If in the above alias-free QMF bank H 0 ( z )
is a linear-phase FIR filter, then its polyphase
components
and E0 ( z ,) are also
E1 ( linearz)
phase FIR transfer functions
In this case, T ( z ) 2 z 1E0 ( z 2 ) E1 ( z 2 )
exhibits a linear-phase characteristic
As a result, the corresponding 2-channel
QMF bank has no phase distortion
36
37
However, in general |T (e j ) | is not a

constant, and as a result, the QMF bank
exhibits magnitude distortion
We next outline a method to minimize the
residual amplitude distortion
Let H 0 ( z ) be a length-N real-coefficient
linear-phase FIR transfer function:
N 1
H 0 ( z ) n 0 h0 [n] z n

Note: H (can
0 z ) either be a Type 1 or a Type 2 linearphase FIR transfer function since it has to be a lowpass
filter
Then
satisfy the condition
h0[n]
h0[ N n]
In this case wehcan
0[ n]write
~ function, a real
j
N
/
2
In the above
is
the
amplitude
H 0 (e ) e
H 0 ()
function of
~
H 0 ()
38

The frequency response of the distortion
transfer function can now be written as
jN
e
T ( e j )
{| H 0 (e j ) |2 (1) N | H 0 (e j ( ) ) |2 }
2
From the above, it can be seen that if N is

even, then T (e j ) 0at = /2, implying
severe amplitude distortion at the output of
the filter bank
39
40
N must be odd, in which case we have

jN
e
T ( e j )
{| H 0 (e j ) |2 | H 0 (e j ( ) ) |2 }
2
jN
e
{| H 0 (e j ) |2 | H1 (e j ) |2 }
2
It follows from the above that the FIR 2channel QMF bank will be of perfect
reconstruction type if
| H 0 (e j ) |2 | H1 (e j ) |2 1

Now, the 2-channel QMF bank with linear-phase
filters has no phase distortion, but will always
exhibit amplitude distortion unless
is a
constant for all
j
|T (e )|
If
is a very good lowpass filter with
H0 ( z)
in the passband and
in thej
j
| H 0 (e )| 0
| H 0 (e )then
|1
stopband,
is a very good highpass
H1 ( z ) with the
filter with its passband coinciding
stopband of
, and vice-versa
H0 ( z)
41

j
As a result, |T (e )| 1/ 2 in the passbands

of H 0 ( z ) and H1 ( z )
Amplitude distortion occurs primarily

in the transition band of these filters
Degree of distortion determined by the
amount of overlap between their squaredmagnitude responses
42

This distortion can be minimized by
controlling the overlap, which in turn can be
controlled by appropriately choosing the
passband edge of H 0 ( z )
One way to minimize the amplitude
distortion is to iteratively adjust the filter
coefficients h0[n] of H 0 ( z )on a computer
such that
43
| H 0 (e j ) |2 | H1(e j ) |2 1
is satisfied for all values of

To this end, the objective function to be
minimized can be chosen as a linear
combination of two functions:
(1) stopband attenuation of H 0 ( z ) , and
(2) sum of squared magnitude responses of
H 0 ( z ) and H1(z )
44

One such objective function is given by
1 (1 ) 2
where
and
H (e
j 2
) d
2 1 H 0 (e
and 0 <
0 < 1, and
j 2
2
2
) d
>0
for some small
) H1(e
s
2
45

Since |T (e j )|is symmetric with respect to
/2, the second integral in the objective
function can be replaced with
2 2
/ 2
1 H
0
2
j
(e )
H1
2
2
(e )
d
After has been made very small by the

optimization procedure, both and
1
2
will also be very small
46

Using this approach, Johnston has designed
a large class of linear-phase FIR filters
meeting a variety of specifications and has
tabulated their impulse response
coefficients
Program 10_9 can be used to verify the
performance of Johnstons filters
47

Example - The gain responses of the length12 linear-phase FIR lowpass filter 12B and
its power-complementary highpass filter
obtained using Program 10_9 are shown
below
20
Gain,dB
H (z)
H (z)
20
40
60
48
80
0
0.2
0.4
0.6
0.8
1

The program then computes the amplitude
distortion | H 0 (e j )|2 | H1 (e j )|2 in dB as
shown below
AmplitudedistortionindB
0.03
0.02
0.01
0
0.01
0.02
49
0.2
0.4
0.6
0.8
1

From the gain response plot it can be seen
that the stopband edge s of the lowpass
filter12B is about 0.71, which corresponds
to a transition bandwidth of
(s 0.5) / 2 0.105
The minimum stopband attenuation is
approximately 34 dB
50

The amplitude distortion function is very
close to 0 dB in both the passbands and the
stopbands of the two filters, with a peak
value of 0.02 dB
51
Alias-Free IIR QMF Bank

Under the alias-free conditions of
G0 ( z ) H1 ( z ), G1 ( z ) H 0 ( z )
and the relation
, the distortion
H ( z ) H 0 ( z )
function T(z) is given 1by
1
) 2 zfunction,
E0 ( z then
) E1 (its
z magnitude
)
If T(z) isTan( zallpass
response is a constant, and as a result its
corresponding QMF bank has no magnitude
distortion
52

Let the polyphase components E0 ( z ) and
E1 ( z ) H 0of
( z)
be expressed as
E0 ( z ) 1A 0( z ), E1 ( z ) 1A 1( z )
2
2
with A 0( z ) and A 1( z ) being stable allpass
functions
Thus, H 0 ( z ) 1 A 0 ( z 2 ) z 1A 1 ( z 2 )
2
H1 ( z ) 1 A 0 ( z 2 ) z 1A 1 ( z 2 )
2
53

In matrix form, the analysis filters can be
expressed as
2)
H
(
z
)
A
(
z
0 1 1 1
0
H 0 ( z ) 2 1 1 z 1A ( z 2 )
1
The corresponding synthesis filters in
matrix form are given by
1
1
1
2
2
G0 ( z ) G1( z ) 2 z A 1( z ) A 0 ( z ) 1 1
54

Thus, the synthesis filters are given by
G0 ( z ) 1 [A 0( z 2 ) z 1A 1( z 2 )] H 0 ( z )
G1 ( z )
2
1
[ A 0( z 2 ) z 1A 1( z 2 )] H1 ( z )
2
The realization of the magnitude-preserving

2-channel QMF bank is shown below
55

From
H 0 ( z ) 1 A 0 ( z 2 ) z 1A 1 ( z 2 )
it can be seen that the lowpass transfer

function H 0 (z ) has a polyphase-like
decomposition, except here the polyphase
components are stable allpass transfer
functions
56

It has been shown earlier that a bounded-real
(BR) transfer function
of0 (odd
H0 ( z) P
z ) / D( z )
order, with no common factors between its
numerator and denominator, can be expressed in
the form
if it satisfies the 2power-symmetry condition

H 0 ( z ) 1 A 0 ( z 2 ) z 1A 1 ( z 2 )
1
H 0 (azsymmetric
) H 0 ( z ) numerator
H 0 ( z ) H 0 ( z ) 1
and has
P0 (z )
57

It has also been shown that any odd-order
elliptic lowpass half-band filter H 0 (z ) with
a frequency response specification given by
j
1 p | H (e )| 1, for 0 p
j
| H (e )| s , for s
p s
and satisfying
the
conditions
2
and s 4 p (1 p ) can always be
expressed in the form
H 0 ( z ) 1 A 0 ( z 2 ) z 1A 1 ( z 2 )
58

The poles of the elliptic filter satisfying the
two conditions on bandedges and ripples lie
on the imaginary axis
Using the pole-interlacing property
discussed earlier, on can readily identify the
expressions for the two allpass transfer
functions A 0(z ) and A 1(z )
59

Example - The frequency response
specifications of a real-coefficient lowpass
half-band filter are given by: p 0.4 ,
s 0.6 , and s 0.0155
2
From s 4 p (1 p ) we get p 0.00012013
In dB, the passband and stopband ripples are
Rp=0.0010435178 and Rs=36.193366
60

Using the M-file ellipord we determine
the minimum order of the elliptic lowpass
filter to be 5
Next, using the M-file ellip the transfer
function of the lowpass filter is determined
whose gain response is shown below
Realhalfbandfilter
Gain,dB
61
20
40
60
0
0.2
0.4
0.6
0.8
1

The poles obtained using the function
tf2zp are at z = 0, z = j 0.486625263 ,
and z = j 0.486625263
The pole-zero plot obtained using zplane
is shown below
ImaginaryPart
1
0.5
0
0.5
1
62
0
RealPart

Using the pole-interlacing property we
arrive at the transfer functions of the two
allpass filters as given below:
2
z 0.2368041466
2
A 0( z )
2
1 0.2368041466 z
A 1( z 2 )
63
z 2 0.7149039978
1 0.7149039978 z 2

Quadrature-Mirror Filter Bank

Uploaded by

Copyright:

Available Formats

Quadrature-Mirror Filter Bank

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Quadrature-Mirror Filter Bank

Uploaded by

Copyright:

Available Formats

Quadrature-Mirror Filter Bank

In many applications, a discrete-time signal