1

Design of Double Density Wavelet Filter Banks

ASHOKA JAYAWARDENA and CHRIS RADFORD
School of Mathematical and Computer Sciences,
University of New England,
Armidale, NSW, 2351,
Email:
ashoka,chris@turing.une.edu.au

Abstract— We look at the design of oversampled filter banks


 ¼
 

¼



¼
  




and the resulting framelets. The undecimated wavelet transform 

is known for its shift invariant properties and has applications in

 ½
 

½



½
 
areas such as denoising [7]. The framelets we will design, will have
improved shift invariant properties over decimated wavelet trans-

 ¾
 

¾



¾
 
form. Shift invariance has applications in many areas particularly
denoising [6], [10], [5] and coding and compression [11]. Our con-
Fig. 1. The double density filter bank
tribution here is on filter bank completion. We will develop fac-
torization methods to find wavelet filters from given scaling filters.
Alternatively we can write the above perfect reconstruc-
tion conditions in the polyphase domain. Given the following
We will look at a special class of framelets from a filter bank
polyphase matrices:
perspective, in that we will design double density filter banks
(DDFB’s) as shown in Figure 1. We denote the z-transform of
   
a sequence

 as 
  and its Fourier transform as 

 .  
  
¼¼ 
 
¼½
   ¼¼
  ¼½
 


   
  ½¼
   ½½
 
 and
   
½¼
  ½½
 

Using the basic multirate identities we obtain the following ex-
   
pression for 
 .  ¾¼
  

 
¾½ ¾¼
  ¾½
 
(4)
where 

  is the type 1 analysis polyphase matrix, and

  

 
  is the type 2 synthesis polyphase matrix, we can write
¼
  
   
 
¼ ½ 
   
 
½ ¾ 
  
 
¾
the perfect reconstruction conditions as




   
 
¼
   

   
 

  

 
¼ ½ ½ ¾ ¾
 
(1) 
 ℄ 
    (5)

Now, for the perfect reconstruction, i.e. 
   
 , it

must be necessary that I. N ECESSARY CONDITIONS FOR PERFECT
RECONSTRUCTION

¼
  ¼
   ½
 

   
 
½ ¾ 
   
¾ (2) In the two-band maximally decimated filter banks, for perfect
reconstruction it is necessary that the scaling filters,  ¼
  and
 
 

 , satisfy the bi-orthogonality constraint, 
 
¼ ¼ ¼

¼
  ¼

   ½
 


   
 
½ ¾ 

    (3)
¾ ℄  . Thus most of the designs were dominated to ensure this
 2



  

 


  
bi-orthogonality constraint [2], [1], [3]. In the design of double 
     (8)
density filter banks we no more have the bi-orthogonality con-
The equation 2 can be written in frequency domain as
straint. Thus strictly speaking we do not have bi-orthogonal







or orthogonal double density filter banks. But we will de- 


 

   
 
   
 
    (9)

sign bi-orthogonal-like and orthogonal-like double density fil-

from which we deduce that
ter banks. We use the term orthogonal-like when the analysis



  




  
filters and synthesis filters are related via, 
   

 
    (10)
 

for       . When the filters are not orthogonal-like we

Since at least one of 

 and 

 are nonzero, the co-
say the double density filter bank is bi-orthogonal-like. Thus
efficient matrix of the equations 10 and 7 must be singular.
we can select scaling filters as we like and complete the filter
Thus,
bank such that it satisfy the perfect reconstruction. But vanish- 






 

 







   

  (11)
ing moments for the wavelet filters do impose constraints on the
Now we have the following system of equations in variables
design of scaling filters.



.
 
 and  

II. C ONSTRAINTS ON T HE S CALING F ILTER 







 







   
(12)
We first formalize the low pass nature of scaling filters for 





   
 

    
the double density filter banks.
The coefficient matrix of the system has full rank since
Definition 1: A filter 
  is said to be a possible scaling



   

 


   
filter if 

  . 
     (13)
The following lemma provides the relationship between the ze-
To see this, we assume
ros of the scaling filters at the aliasing frequency  and the first    




 
 
vanishing moments of wavelet filters.  
   


 

 
 
 



 
 


 
(14)
Lemma 1: Let 

  , 
. Then the fol-
lowing are true.


and find a contradiction. Substitute 

   

  and

 
¯ If at least one of  

 and 

 are nonzero, 


   

  

 ¾ in equation 13 to see the contradiction.

, and  

    then  

   and 


  . 


  Since coefficient matrix has full rank, we must have  
  

¯ 


If at least one of 
 and 
 are nonzero, 


   and 

  .

, and  

    then 

   and 

  .

  The third case is easier. Substituting    in equation 6 we





  
¯ 
If 
   
   
  , and   get 


 


  , which implies 

   . Similarly



then 


   , and  

   .



we get 

    after substituting    in equation 6.
Proof: First, we prove the first case and the second fol- Thus unlike in the wavelets which corresponds to the max-
lows similarly. We can write equation 3 in frequency domain imally decimated filter banks, vanishing moments of the high
as pass filters are not guaranteed simply by enforcing zeros at the

aliasing frequency for the scaling filters in general. But vanish-



    

 

  

    

 


     

   
ing moments of the high pass filters are still guaranteed for the
(6)
orthogonal like double density filter banks.
from which we can deduce the following after substituting zero
Corollary 1: Let 
be the scaling filter of a orthogonal like
and  for  respectively.
double density filter bank and   and  are the wavelet filters.



   




   

    (7) Then 

   and 

   if and only if 


   .
 3

Proof: It follows immediately from Lemma 1 since van- Now the result follows after taking the derivatives of the above
ishing moments for the analysis wavelet filter implies vanishing equation in the frequency domain and then substituting   .

moments for the synthesis wavelet filter and vice versa.

III. D ESIGN OF THE S CALING F ILTER

A. Vanishing moments for wavelet filters
The most popular way of achieving smoothness for the scal-
First we will look at necessary conditions which must be sat-
ing filter is to enforce polynomial interpolation property for the
isfied by the scaling filters such that wavelet filters has a given
scaling filter. We will enforce a factor
 
 ¼ for the
number of vanishing moments. The following lemma gives the ´µ


scaling filter 

 , which in turn results in 

    for
minimal number of higher derivatives of scaling filters which
vanish at the aliasing frequency.
   

.
We will discuss a technique due to [4]. In that, Selesnick de-
Theorem 1: Let the analysis wavelet filters has at least 
 
scribes a maximally flat symmetric FIR filter which was origi-
vanishing moments and the synthesis filters has at least   van-

´µ
´µ nally described by Herrman [8]. Let 

  be the product filter
ishing moments. Then 

    and 

    for

´µ
which must satisfy the constraints 


  , 


   for
     where      .
min
 
   , 


´µ

    for    
, and 

  is
Proof: We prove the Theorem by mathematical induction.
symmetric, then the product filter is given by
The case for    is covered by the Lemma 1. Now we assume
that the Lemma is true for   
, i.e. 


´µ

    and
´µ


    for 
  
. By taking the   derivative of 

  
  
¼ 






(18)
equation 6 and substituting    we get


 




´µ 

  


 
(15)
Now orthogonal-like scaling filters can be obtained by spec-
´µ

 tral factorization of the product filter while bi-orthogonal-like
which implies 

   . Similarly, by taking the 

derivative of equation 6 and substituting    we prove that scaling filters can be obtained by polynomial factorization and
´µ

appropriate regrouping of the factors.


   .

The vanishing moments of wavelet filters imposes further

IV. A P OLYPHASE BASED D ESIGN A PPROACH
constraints on the scaling filter. The following Lemma indicates

  As with many problems in signal processing, polyphase con-
that the corresponding product filter 

   

 
of
structions substantially simplify constraints. It is the same with
the scaling filters must have the coiflet like property, i.e. zeros
double density filter banks. Let the analysis and synthesis
of higher derivatives at zero.
   wavelet filters are of the form
Lemma 2: Let analysis wavelet filters   and   has  
 vanishing moments and the synthesis wavelet filters 
and  

 
 


 

 

and  has  and  vanishing moments. Then 
 

 


 

 
(19)

´µ

     
min
     (16)

  



 

 

   

  



 

 

Let the type 1 polyphase matrix of analysis wavelet filters and

Proof: Let  
   
 

   and 
  
 type 2 polyphase matrix of synthesis wavelet filters are given by

 .

   Then the equation 2 can be written as  

  
 

    
 
 



  
 


   
   
    (17)  
 
 4

and   we get


  
  
 

      
   
  (25)



  
 
Many factorization methods of Laurent polynomial matrices
Then from equation 5 it is easy to see that
  are crucially dependent on the determinant of the matrix. For-



 


 

  
  
 




 

  (20) tunately, for a good subclass of double density filter banks, we


 
have a simplified result for the determinant of 
  . We as-
We can obtain a simplified expression for both 

  and sume that the number of vanishing moments of analysis high

 . Note thatthe type 1polyphase vector of

 
  pass filters and synthesis high pass filters are equal, i.e.  


 
   . It can be seen that the determinant of  
 

 
can be written as 
  where 
  is the type 1


 


 

 . Since
 is given by 
   
 



  



polyphase vector of 
  . Thus the type 1 polyphase vector

 and 
 are
 
  is FIR and both 


  can be written in the following form.




 
 
     factors of  
 

 , it must be that





 



 
 


  is a factor of 
 . We have the following Lemma.

 

 
 
 Lemma 3: Let the high pass filters has  number of vanish-
Simmilarly  , have the
ing moments each and the low pass filters, 
and 

    


 
 


 
 
equal regularity,  , and given by


 
 

 

 


  
  
 
 





Thus
  




 
 

 
  

    (21)

(26)
 
 
  Then the determinant of 
  is a real number when,
where 
    . Simmilarly


 
 
 
 
 
  (27)
  


 

   
    (22)

Proof: We have
Let
       det
 
 

  





 
 det

 


     
 


 
 
 



 


 

 



 



 


 


 

 
(28)
  

  
 
 
 Also note that, 


 

 

 
 

 


 
 ℄. Now the largest power of  in
    and hence the largest
(23) det
 
 

  is

  
and 
    
 . Since power of  in det

  is 
 . By symmetry, the
           .
     smallest power of  in det

  is


 

 
 The above lemma covers some useful number of dou-



     (24) ble density filter banks irrespective whether they are bi-

 
 
  
 orthogonal-like or orthogonal-like such as
   




 



 
 
 
 
 
 
 
 
  etc.
 5

A. Factorizations for Orthogonal-like DDFB’s From equations 32 and 31, we get

Note that the polyphase matrix of the orthogonal-like DDFB

is a  lossless system [9]. Selesnick [4] designs the high pass   
  


    (33)

filters indirectly by first designing a lossless system and Thus it is clear that the coefficients of 
 and 
 in

then extracting only the first two rows to form a  lossless 
  
 
  are zero.




system. We provide a more direct method for a class of DDFB’s Remark 1: Note since  

   
 
  ,


by factorizing 
  when det

  is a real number (as in


   

 ℄


 



℄
 (34)
Lemma 3).
Lemma 4: Let is also a possible factorization but it does not give us new filters!
             Assuming each degree reduction step is invertible, repeating


   

  


   the process given in Lemma 4, we could write
 
    
 

and assume that det


  is a nonzero real number. Let


  
  
 
 
   
 


   (35)
  


       

     

 


 
   where    is a constant symmetric matrix (i.e.    
 

   


 ).

  Then    is orthogonally diagonalizable and let
and 

   
  
 and 

  is nonsingular. Then for   
 
 
    . However we require the eigenvalues of
some 
 ,    be positive. Then we could write



   



℄
 
 
 ℄  (29) 
    
   
   


   (36)

where We have one degree of freedom in the factorization as given

             by

         

   
          

    
   
     
 


  
Proof: Since 

 

   , we have
   

(37)


 
   (30) Such parameterization is used by Selesnick [4] to achieve near

shift invariance among the wavelet functions.
    (31)


 Example 1: Consider the (2,1) orthogonal-like DDFB with
     
  


  ½

and 

    and 
 . We get
 
 


  
  
 
    (32)

     which leads to
 ½
From equations 30 and 31 it is clear that
      
   


 

  

    
      


 
   
  
  
 
 

  



    
  

 
 


  



 



  
Now   high pass filters are given by
The synthesis 

 

  
 
  
            
  . This gives 
   
  and







 
  


 


  


  



 
 
  

 
  


   
  
 

  
  
 .
 6

Orthogonal-like double density filter

[4] Selesnick I. The double density dwt. In A. Petrosian and F. G. Meyer,
    




  
  
  

   
   



  


 
  






 editors, Wavelets in Signal and Image Analysis: From Theory to Practice.

    



 





  
  

 


  Kluwer.

         
     
 
 


  [5] Munch N. J. Noise reduction in tight weyl-heisenberg frames. IEEE
    
   
 
 
   


  Trans. on Information Theory, 38(2):608–616, March 1992.

    
   



            maximum phase   
[6] Berkner K. and Wells R. O. Jr. A correlation-dependent model for denois-

 
 
    
        
 

    


ing via nonorthogonal wavelet transforms. Technical Report CML TR98-
  
 

   07, Computational Mathematics Laboratory, Rice University, 1998.

   
    

     
 


[7] Lang M., Guo H., Odegard J. E., Burrus C. S., and Wells R. O. Jr. Noise
         

           minimum phase   
reduction using an undecimated discrete wavelet transform. IEEE Signal


 

    
       
  



 
  Processing Letters, 3(1):10–12, January 1996.
 
    


 



 
 



  

   
      
[8] Herrmann O. On the approximation problem in nonrecursive digital filter


 

 

 
  
      
 

 
 design. IEEE trans. on Circuit Theory, (18):411–413, May 1971.
  
[9] Vaidyanathan P. P. Multirate Systems and Filter Banks. Prentice Hall,
TABLE I
Englewood Cliffs, New Jersey, 1993.
PARAMETRIC D OUBLE D ENSITY F ILTER B ANKS .
[10] Coifman R. R. and Donoho D. L. Translation-invariant de-noising. In
Antoniadis A., editor, Wavelets and Statistics, Lecture Notes. Springer-
Verlag, 1997.
Table I shows parametric orthogonal-like filters for smaller [11] Xiong Z., Orchard M. T., and Zhang Y.-Q. A deblocking algorithm

values of  and  . Clearly Lemma 4 is not always applicable for jpeg compressed images using overcomplete wavelet representations.
IEEE Trans. on Circuits Syst. Video Technol., 7(2):433–437, April 1997.
since there is no guarantee that the degree reduction step is in-
vertible. But so far in all the examples I have computed, I have
not run into this problem. Whether it is that we have just been
lucky or that we haven’t uncovered some of the hidden results,
is not yet clear to me. But when we generalize double density
into multiple density case, we will realize that we have been
lucky in the double density case.

V. C ONCLUSION AND F URTHER R ESEARCH

We have developed a factorization approach to obtain double

density wavelet filters for a special case where the determinant
of transfer polyphase matrix is a real number. We have analyti-
cally obtained number of example filters for with small number
of filter coefficients. However, optimization of such filters was
left undone since it can be done similarly as in Selesnick’s [4].

R EFERENCES
[1] Cohen A., Daubechies I., and Feauveau J.-C. Biorthogonal bases of
compactly supported wavelets. Communications of Pure and Applied
Mathamatics, XLV:485–560, 1992.
[2] Wei Dong, Tian Jun, Wells Raymond O., and Burrus C. Sydney. A new
class of biorthogonal wavelet systems for image transform coding. IEEE
trans. on IP, 7(7):1000–1013, 1998.
[3] Daubechies I. Othonormal bases of compactly supported wavelets.
Comm. Pure Appl. Math., 41:909–996, 1988.