Sunteți pe pagina 1din 45

Decimationand

Interpolation

DecimationandInterpolationhave
alreadybeendefinedinthefirstchapter.
Thischapterdealswiththecombined
usageofdownsamplerswithanti
aliasingfilters,andofupsamplerswith
antiimagingfilters.Theaimisto
developwaystoefficientlydesign
decimatorsandinterpolators.

4.1 Decimation with


Transversal Filters
4.1.1 Convolution with Subsequent
Downsampling
In section 1.2.3, it was shown that a
decimator generally consists of an antialiasing filter h(n) followed by a
downsampler M, as shown in Fig.4.1.

u(n)

h(n)

x(n)

y(n)

Figure 4.1 Decimator consisting


of an anti-aliasing filter h(n) and
a downsampler M

Assuming the use of an FIR filter with N


coefficients, the filtering can be described by
the convolution
N -1

x(n) h(n) * u(n) h(k) u(n - k)

(4.1)

k 0

and the downsampling as


y(m) x(m M)

(4.2)

Decimation can thus be described as convolution followed


by downsampling :
N -1

y(m) h(k) u(mM - k)


k 0

(4.3)

4.1.2
Efficient Transversal
Structure

a)

b)

u(n)
z-1
z-1

h(0)
h(1)
h(N-1)

y(m)

u(n)
z-1
z-1

M
M

h(0)
h(1)
h(N-1)

Figure 4.3 Decimator: original structure with


a transversal filter followed by a downsampler
(a), and an efficient structure (b).

y(m)

4.2 Interpolation with


Transversal Filters
4.2.1 Upsampling with
Subsequent Convolution
In section 1.3.3 it was shown that in general an
interpolator consists of an upsampler L followed by
an anti-imaging filter g(n) as shown in Fig. 4.4.

y(m)

v(n)
L

g(n)

x(n)

Figure 4.4 Interpolator made up of an


upsamler L and an anti-imaging filter g(n)

The upsampling described in (1.82) produces the


interim signal v(n) from the input signal y(m).
The output signal x(n) is obtained by convolving this
with the impulse response g(n), which is assumed in
what follows to be finite, with N coefficients:

N -1

x (n )

v(n-k) g(k)
k 0

(4.4)

4.2.2
Efficient Transversal
Structure

4.3 Decimation with


Polyphase Filters
No Improvement
Used in filter banks

4.3.1 The Principle of


Polyphase Decimators
Let us consider, in the Fig. 4.2, only those multiplications
between the impulse response and the input signal, which
contribute to the output values y(0), y(1) .....
Then we recognize that certain input values, such as u(3),
u(0), u(-3), u(-6). are combined only with a few of the
impulse response coefficients, in this case
with h(0), h(3), h(6), h(9)....
The input values u(2), u(-1), u(-4), u(-7) are combined
only with the coefficients h(1), h(4), h(7), h(10)..., and the
input values u(1), u(-2), u(-5), u(-8) only with
h(2), h(5), h(8), h(11)...

The convolution shown in Figs 4.2a,b and c can


thus be considered as a combination of three
independent convolutions. The three convolution
products are added at the end, giving the output
sequence y(0),y(1)...
Fig 4.7 shows the three independent convolution
processes. For example, Fig.4.7a shows how the
output subsequence y0(m) is obtained from the
impulse response component h(3n) and the input
subsequence u(3n). Similarly, Fig. 4.7b shows the
formation of the output subsequence y 1(m) from
h(3n+1) and u(3n-1). The total output sequence
y(m) is then the sum of the subsequences y 0(m) ,
y (m) and y (m).

In this method, no unnecessary


calculations are performed. The number
of filter operations is the same as with the
transversal filter structure in Fig. 4.3b.
Since the subsequences used are basically
polyphase signals, the process is called
polyphase filtering. In the following
section, this will be formally derived.

4.3.2
Representation of Polyphase
Decimators in the Time Domain
Beginning with the convolution
relationship (4.3), and using the
substitution
k = rM+

(4.5)

The decimated signal can be expressed as


y( m )

h(k) u(mM - k)

k -
M 1

h(rM ) u([m - r] M - ).
0 r -

(4.6)

To simplify (4.6), the abbreviations


h(r)

= h(rM+ )

and
u(r) = u(rM - )

(4.7)
(4.8)

are used for the polyphase components


h(n) and u(n), see Fig. 1.16. Substituting
(4.7) and (4.8) into (4.6) gives
y(m)

M 1

h (r) u (m - r)

0 r -
M -1

h (m) u

( m)

(4.9)

4.3.3
Representation with an
Input Commutator

4.3.4 Memory-Saving
Polyphase Filters
If all the subfilters in the polyphase configurations
of Figs. 4.8 to 4.10 are implemented in the direct
form of Fig. 2.8, the resulting system requires less
memory. Fig. 4.12 shows the polyphase
configuration from Fig 4.8, with an input
commutator and transposed subfilters.

4.3.5 Representation of
Polyphase Decimation Using the
Z-transform
U(z)

H(z)

X(z)

Y(z)

Figure 4.15 General form of a


decimator with z-transformed signals

The structure to be derived is based


upon the polyphase representation of
the z-transform in (1.10) and (1.11).
From these, it follows that the
transfer function H(z) of the FIR
filter in Fig. 4.15 can be written as
M -1

H(z)

z - H ( z M )

(4.10)

The subfilters H0(z)...... HM-1(z)


are likewise FIR filters, and when
combined in the right phase
sequence produce the original
filter H(z). It follows that the
general structure in Fig. 4.15 can
be replaced by the polyphase
structure in Fig. 4.16a.

The first and the thrid identites in Figs.1.23


and 1.25 can be used to derive the version in
Fig. 4.16b, in which both the number of the
filter operations and the amount of memory
required are reduced by a factor of M. This
structure corresponds to the one in Fig. 4.9,
which was derived in the time domain. In
particular, we have
H(z)
h (m), =0,1,2...M-1, (4.11)
for the filters of both structures.

Figure 4.16 Original form of a


polyphase decimator (a), and a
more efficient and memory
saving version(b)

a)
U(z)
H0(zM)
z-1

H1(zM)

z-1
H2(zM)
z-1
HM-1(zM)

Y(z)

b)
U(z)

z-1

Y(z)

H0(z)

H1(z)

H2(z)

HM-1(z)

z-1
z-1

4.4 Interpolation with


Polyphase Filters
Interpolators are the duals of decimators. The
respective signal flow graphs can be derived from
each other - the directions of all signals are here
reversed. Downsamplers and upsamplers
interchanged and inputs and the outputs
swapped. Because of this, the intrepolator
structures derived below are of a very similar
form to the decimator structures in the previous
section.

4.4.1 Principle
of Polyphase
Interpolators

4.4.2 Polyphase Interpolation in


the Time Domain
Consider the convolution
relationship in (4.4). The substitution
k = rL +
(4.12)
yields the following expression for
the interpolated signal :

N -1

x (n )

v(n k) g(k)
k 0
L 1 R -1

v(n rL ) g(rL ).

(4.13)

0 r 0

This assumes that the number N of


coefficients of the FIR filter g(n) is a multiple
of the interpolation factor L: N=RL.
From (1.82), we have the following:

y(m - r)
v(n rL )
0

for n mL
otherwise.

(4.14)

For each index n of the output sequence


x(n) there is exactly one index = n mod L
which makes a nonzero contribution in
(4.13). Using (4.14), this contribution is

R -1

(m)

v(n rL - ) g(rL )

r 0
R 1

y(m r ) g
r 0

y( m ) *

| n mod L

(r )

( m)

g (m) g (mL )

(4.15)

(4.16)

By upsampling with a factor L and introducing a delay


of clock periods (at the higher sampling rate), the
polyphase components:

( p)

x ( m)

( n)

for n mL
otherwise.

(4.17)

are obtained from the components x(m). These


are then summed to give the output signal x(n).
This process is shown in Fig. 4.18 for the
earlier example.

S-ar putea să vă placă și