Eee507 2DFIRDesign2 PDF

FIR Filter Design Using Transformations
Objectives
! Reduces M-D FIR design to 1-D FIR design
! Efficient realization of designed filters
Motivation
! M-D designs based on L (Chebyshev) norm are difficult
! 1-D designs very well studied and understood
Strategy (M = 2; similar for any M-D design)

! Given ideal 2-D specifications I(1, 2)
" Transform 2-D specs into appropriate 1-D specs by choosing appropriate 1D filter I()
1-D filter called prototype filter.
" Design 1-D FIR filter H() approximating I() using known 1-D techniques
" Transform the designed 1-D prototype FIR filter H() into a 2-D FIR filter
H(1, 2) by using an appropriate 2-D mapping function
F(1, 2)
1-D FIR prototype
Resulting designed
H()
,
)
H(
mapping
1
2
2-D FIR filter
" Want to choose 1-D prototype I() and 2-D mapping function F(1, 2) such
that H(1, 2) ~ I(1, 2).
EEE 507 - Lecture 13
Copyright 2004 by Prof. Lina Karam

Two main issues
! How can we transform a 1-D FIR filter into an 2-D FIR filter ?
! How do we choose the 1-D filter (prototype filter) and mapping function
F(1, 2) ?

How to transform 1-D FIR filter into a 2-D FIR filter?
! Only symmetric 1-D filters can be transformed by the proposed
McClellan Transformation method.
! Background:
" Mcclellan transformation method maps 1-D positive-symmetric FIR filters into
2-D (positive-symmetric) FIR filters.
" Consider a 1-D zero-phase symmetric filter of length 2N+1
H ( ) =
h ( n)e
j n
= h ( 0) +
h ( n )e
n = N
N
n= N
jn
+ h(n)e jn
n =1
= h(0) + h(n)[e jn + e jn ]
N
n =1
= a (n) cos n
where
n =0
h (0), n = 0
a (n ) =
2h ( n ), 1 n N

How to transform 1-D FIR filter into a 2-D FIR filter? (continued)
! Background: Chebyshev Polynomials
" It is well-known that cos n can be expressed as a polynomial of
degree n in the variable cos
cos n = Tn (cos )
where
Tn ( x) = n th Chebyshev polynomial
T0 ( x) = 1
T1 ( x) = x
Tn +1 ( x) = 2 xTn ( x) Tn 1 ( x),
" Thus:
1 x 1
H ( ) = a (n) cos n
n =0
H ( ) = a(n)Tn (cos )
n =0

! Background: Derivation of the Chebyshev Recursion
1
1
cos A cos B = cos( A + B ) + cos( A B )
2
2
" Let A = and B = (n 1) , then
or
2 cos cos(n 1) = cos n + cos(n 2)

cos n = 2 cos cos(n 1) + cos(n 2)
Tn (cos ) = 2 cos Tn 1 (cos ) Tn 2 (cos )
Tn ( x) = 2 xTn 1 ( x) Tn 2 ( x)

! The Transformation:
" McClellan suggested that we can obtain a 2-D zero-phase FIR filter
if we make the substitution
cos
replace by
F (1 , 2 )
Then
N
H ( ) = a(n) cos n
n =0
N
H ( ) = a(n)Tn (cos )
n =0
H(1,2 ) = a(n)Tn[F(1,2 )]
n=0

H ( ) = a(n)Tn (cos )
n =0
N
H(1,2 ) = a(n)Tn[F(1,2 )]
n=0
" The mapping function F(1,2) should be real (zero-phase) and
F (1 , 2 ) 1 or
1 F (1 , 2 ) 1
since F(1,2) has to replace cos() and has to take the same
values of cos() since argument of Chebyshev polynomial
=> H(1,2) takes same exact values of designed 1-D FIR
prototype filter H().
Consequence:
- values of H(1,2) are determined by values of H().
- H(1,2) only takes values of H().

" The mapping function F(1,2) should be chosen to be the frequency
response of a zero-phase FIR filter
" If F(1,2) is real and a(n) real H(1,2) is real zero-phase
" If F(1,2) is (2Q+1) x (2Q+1) H(1,2) is (2NQ+1) x (2NQ+1) because
H(1,2) is a polynomial of degree N in F(1,2) same size as FN
FT
f F
(2Q + 1) (2Q + 1)
f * f F2
(4Q + 1)(4Q + 1)
M
f f K f F N
14243
N
(2 NQ + 1)(2 NQ + 1)
Times
g Tn ( F )
123
(2nQ + 1) (2nQ + 1)
144424443
h H (1 , 2 )
1424
3
( 2 NQ + 1) ( 2 NQ + 1)
an F n +...+ a0
size of support of g
bn F n +...+ b0
" Size of support of h(n1,n2) depends on Q (order of transformation subfilter F)

and on N (order of 1-D prototype filter)

H ( ) = a(n)Tn (cos )
n =0
N
H(1,2 ) = a(n)Tn[F(1,2 )]
n=0
" The mapping function F(1,2) determines the symmetries of the

designed 2D FIR filter H(1,2) => H(1,2) have the same
symmetries of F(1,2).
" Example: For zero-phase, even-symmetric H(1,2), F(1,2)
should be zero-phase and even-symmetric of the form:
Q
F (1 , 2 ) = t qr cos( q1 ) cos( r2 )
q =0 r =0
I
+ sij sin(i1 ) sin( j2 )

i =0 j =0
Note: Order of F(1,2) = max(Q,R,I,J).


H ( ) = a(n)Tn (cos )
n =0
N
H(1,2 ) = a(n)Tn[F(1,2 )]
n=0
" The mapping function F(1,2) determines the shape of the

contours of the designed 2D FIR filter H(1,2) => Contours of
H(1,2) have the same shape as the contours of F(1,2).
" Let F (1 , 2 ) = C , C 1
2
N
H (1 , 2 ) = a (n)Tn [C ] = C
n =0
F = C = const
H (1 , 2 ) = C ' = const
1
" Consequence:
Shape of contours of H(1,2) determined by F(1,2).
Value C determined by {a(n)}, n = 1 to N, which depends on the
designed 1-D prototype H() => prototype H() determines the
values of H(1,2) along contours.
Flexible and Modular:

-
Change F(1,2) to change contour shape

Change {a (n )} (prototype) to change values on contours
Procedure:
1) Design transformation subfilter F(1,2) :
a. to get desired symmetries
b. to produce or approximate desired contours
c. OR, if prototype filter already specified, to map values of prototype to
desired 2-D locations or contours
2) Design prototype to get desired values along contours

Example:
Consider
F (1 , 2 ) = A + B cos(1 ) + C cos(2 ) + D cos(1 2 ) + E cos(1 +
Let A=-B=-C=-1/2, D=E=1/4 (proposed by McClellan)
1
1 1
1
F (1 , 2 ) = + cos(1 ) + cos(2 ) + cos 1 cos 2
2 2
2
2
n2
f (n1 , n2 )
1
8
1
4
1
8
1
4
1
2
1
4
1
8
1
4
1
8
n1
F (1 , 2 ) = F ( 1 , 2 ) = F ( 1 , 2 ) = F (1 , 2 ) (quadrant symmetry)
=> designed H (1 , 2 ) has same symmetry
Example: (continued)
1 1
1
1
+ cos(1 ) + cos(2 ) + cos 1 cos 2
2
2 2
2
Contours of F(1,2):
F (1 , 2 ) =
F (1 , 2 )
Tend to be rectangular
1
Circular around origin
Note:
N
1) F (1 ,0) = cos 1 H (1 ,0) = a(n)Tn (cos 1 ) = H (1 )
n =0
=>Prototype exactly mapped along 1 = 0 axis

N
2) F (0, 2 ) = cos 2 H (0, 2 ) = a(n)Tn (cos 2 ) = H ( 2 )
n =0
=>Prototype exactly mapped along 2 = 0 axis

To get lowpass 2-D filter, use
H ( )
To get bandpass 2-D filter, use

H ( )
L
Mapping F(1, 2)
Prototype Filter H()
Designed 2D Filter H(1, 2)
Mapping F(1, 2)

Desired I(1, 2)

Mapping F(1, 2)


Design Example: We want to design a Fan filter using transformation
2
I (1 , 2 )
Requirement: 1) Use a 1st order transformation F(1,2) function => Q=R=1

2) Use a lowpass 1-D prototype
I()
/2
/2

2
Example: Design of a Fan Filter (continued)
I (1 , 2 )
Choice of Transformation Function

- I (1 , 2 ) quadrant symmetric => F (1 , 2 ) needs to be quadrant symmetric
so, F (1 , 2 ) of the form:
Q
F (1 , 2 ) = t qr cos( q1 ) cos( r2 )
q =0 r = 0
= t00 + t10 cos 1 + t01 cos 2 + t11 cos 1 cos 2

Note: since quadrant symmetric => need to consider only 2
one quadrant
0
F (1 , 2 ) = t00 + t10 cos 1 + t01 cos 2 + t11 cos 1 cos 2

! Determination of parameters:
Typically done by:
1. Exploiting other symmetries
2. Mapping values of prototype appropriately
F (1 , 2 ) = t00 + t10 cos 1 + t01 cos 2 + t11 cos 1 cos 2

2
1. Exploiting Other Symmetries:

I (1 , 2 ) symmetric about line 2 = 1 +
=> Choose F (1 , 2 ) to be symmetric about same line
F (1 , 2 ) = F ( 2 , 1 )
I (1 , 2 )
t00 + t10 cos1 + t01 cos 2 + t11 cos1 cos 2
= t00 t10 cos 2 t01 cos1 + t11 cos 2 cos1
t01 = t10
t10 = t01
F (1 , 2 ) = t00 + t10 (cos 1 cos 2 ) + t11 cos 1 cos 2

3 unknowns => need 3 equations to solve for them

2. Map values of prototype:
Since used 1-D prototype is lowpass, appropriate

mapping is as follows:
1-D point
=0
=
=
2
maps into
2-D point (1,2)
(0, )
( ,0)

( , )
2 2
I (1 , 2 )
Note: To map value at 1-D point to a desired 2-D location (1,2),

simply set F(1,2)=cos().

a) = 0 (1 , 2 ) = (0, )
F (0, ) = cos(0) = 1
t00 + t10 (cos 0 cos ) + t11 cos 0 cos = 1 t00 + t10 (1 + 1) t11 = 1
t00 + 2t10 t11 = 1
(1)
b) =
(1 , 2 ) = ( ,0)
F ( ,0) = cos( ) = 1
t00 + t10 (1 1) + t11 cos cos 0 = 1 t00 2t10 t11 = 1
(2)
(1) ( 2) 4t10 = 2 t10 = 1 / 2
c) = / 2
(1 , 2 ) = ( / 2, / 2)
F ( / 2, / 2) = cos / 2 = 0
t00 = 0
t00 = 0
t11 = 2t10 1 = 0
t10 = 1 / 2
F (1 , 2 ) = 1 / 2 cos 1 1 / 2 cos 2
( using (1) )

F (1 , 2 ) = 1 / 2 cos 1 1 / 2 cos 2
(2 N + 1) = length of h( n)
F (1 , 2 ) = 1 / 2(cos 1 cos 2 )
H ( )
F<0
2 > 1
cos 2 < cos 1
=> F > 0
H (1 , 2 ) = h(0) + 2h(n)Tn ( F (1 , 2 ))
cos() >0
cos() < 0
n =1
= a (n)Tn ( F (1 , 2 ))
n =0
where a (0) = h(0), a(n) = 2h(n), 1 n N

Review of design procedure
N
H ( ) = a(n)Tn (cos )
n =0
N
H(1,2 ) = a(n)Tn[F(1,2 )]
n=0
{a (n )}: from 1-D prototype design

F (1, 2 ) : frequency response of low-order FIR zero-phase filter
! If h(n) is 2N+1 points long and f(n1,n2) is (2Q+1)x(2Q+1), then
h(n1,n2) is (2NQ+1)x(2NQ+1).
! Contour shape and symmetries depends only on F.
! Amplitude associated with a particular contour depends upon the
1-D prototype
Implementation of Transformed FIR Designs

Since the filter is FIR, a direct convolution can be used to implement
the filter
A DFT can also be used
There also exists a special implementation that exploits the design
! Based on the recursion of Chebyshev polynomials
Tn [ F (1 , 2 )] = 2 F (1 , 2 )Tn 1 [ F (1 , 2 )] Tn 2 [ F (1 , 2 )]
T0 [ F (1 , 2 )] = 1
T1[F(1, 2 )] = F(1, 2 )
! Thus, if we have Tn1 [ F ] and Tn2 [ F ], we can get Tn [F ],because
F(1,2) is an FIR filter.
Tn2 [ F ]
Tn1 [ F ]
2 F (1 , 2 )
+ +
Tn [F ]

From the Chebyshev polynomials recursion, we obtain:
Tn2 [ F ]
Tn1 [ F ]
+ +
2 F (1 , 2 )
Tn [F ]
In addition, we can get Tn1 [ F ] and Tn2 [ F ] from the polynomial that
come before and so on until we get back to T1 [ F ] and T0 [ F ] .
-1
F
2F
2F
-1
T0 [ F ]
T1 [ F ]
T2 [ F ]
T3 [ F ]

Finally, the frequency response H(1, 2) of the 2D designed FIR
filter is just a linear combination of the Ti [F ]:
N
H (1 , 2 ) = a (n)Tn [ F (1 , 2 )]
n =0
a(0)= h(0) and a(n) = 2h(n), 1 n N .
where
-1
IN
h(0)
2F
2h(1)
-1
+
2F
2h(3)
2h(2)
out
Note: This structure will work for filters of any dimensionality.

To realize an M-dimensional filter H (1 , 2 ,..., M ),
we simply substitute an M-dimensional mapping function
F (1 , 2 ,..., M )

Consider the first-order transformation:
F (1 , 2 ) = A + B cos 1 + C cos 2 + D cos(1 2 ) + E cos(1 + 2 )
A,
B
,
2
C
,
f (n1 , n2 ) = 2
D ,
2
E
,
2
n1 = n1 = 0
n1 = 1, n2 = 0
n1 = 0, n2 = 1
n1 = 1, n2 = n1
n1 = 1, n2 = n1
f ( n1 , n2 ) of size 3x3
D
2
B
2
E
2
C
2
A
C
2
E
2
B
2
D
2
Direct implementation of F requires 5 multiplications per output and

8 adds per output
! Note: Multiplications reduced by combining terms


Computational Complexity
! To implement each filter F of size (2Q+1)x(2Q+1) requires in general:
( 2Q + 1) 2 Multiplications/Output
( 2Q + 1) 2 1 Additions/ Output
( 2Q + 1) Rows of storage
! For the whole filter H(1, 2), we have N filters F:

N ( 2Q + 1) 2 + ( N + 1) Mults/Output
N ( 2Q + 1) 2 + ( N 1) Adds/Output

Eee507 2DFIRDesign2 PDF

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

Eee507 2DFIRDesign2 PDF

Uploaded by

Copyright:

Available Formats

FIR Filter Design Using Transformations

Strategy (M = 2; similar for any M-D design)

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

EEE 507 - Lecture 13

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

EEE 507 - Lecture 13

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

2 cos cos(n 1) = cos n + cos(n 2)

Tn (cos ) = 2 cos Tn 1 (cos ) Tn 2 (cos )

EEE 507 - Lecture 13

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

EEE 507 - Lecture 13

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

" The mapping function F(1,2) should be real (zero-phase) and

EEE 507 - Lecture 13

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

" Size of support of h(n1,n2) depends on Q (order of transformation subfilter F)

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

" The mapping function F(1,2) determines the symmetries of the

+ sij sin(i1 ) sin( j2 )

Note: Order of F(1,2) = max(Q,R,I,J).

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

" The mapping function F(1,2) determines the shape of the

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

Flexible and Modular:

Change F(1,2) to change contour shape

2) Design prototype to get desired values along contours

EEE 507 - Lecture 13

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

F (1 , 2 ) = A + B cos(1 ) + C cos(2 ) + D cos(1 2 ) + E cos(1 +

Let A=-B=-C=-1/2, D=E=1/4 (proposed by McClellan)

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

=>Prototype exactly mapped along 1 = 0 axis

=>Prototype exactly mapped along 2 = 0 axis

EEE 507 - Lecture 13

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

To get bandpass 2-D filter, use

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

Prototype Filter H()

Designed 2D Filter H(1, 2)

EEE 507 - Lecture 13

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations

Prototype Filter H()

Designed 2D Filter H(1, 2)

EEE 507 - Lecture 13

Copyright 2004 by Prof. Lina Karam

FIR Filter Design Using Transformations