Introduction to The Lifting Scheme

Wavelet Transforms

• Two approaches to make a wavelet

transform:
– Scaling function and wavelets (dilation equation
and wavelet equation)
– Filter banks (low-pass filter and high-pass filter)
• The two approaches produce same results,
proved by Doubeches.
• Filter bank approach is preferable in signal
processing literatures
Wavelet Transforms

H0 (z) 2 2 G0 (z)
X'0 (z) X0 (z) Y'0 (z) Y0 (z)
X(z) + Y(z)

H1 (z) 2 2 G1 (z)
X'1 (z) X1 (z) Y'1 (z) Y1 (z)

|H0(w)|, |G0(w)| | H 1 ( w ) |, | G 1 ( w ) |

  / 2  /2    / 2  /2 
Ideal low-pass filter Id e a l h ig h - p a s s f ilt e r

Practical Filter
Understanding The Lifting Scheme

signal

d k( 0 ) Splitting

sk( 0 )
Predicting
(1)
d k
Updating
(1)
s k
…
Transmitting
sk(1)

d k(1)
Inverse Updating
sk( 0 )
Inverse Predicting
d k( 0 )
Merge

signal
Lifting Scheme in the Z-Transform Domain

Update stage

 
S (z) S '( z )
 
2 Low band signal
X (z) s p lit P (z) U (z)
High band signal
-
D (z) D '( z )

 

Prediction stage
Lifting Scheme in the Z-Transform Domain

Inverse update stage

 
S '( z ) S (z)
-
com -
U (z) P (z) X (z)
b in e

D '( z ) D (z)  
2

 

Inverse prediction stage

Four Basic Stages

• A spatial domain construction of bi-

orthogonal wavelets, consists of the
following four basic operations:

• Split : sk(0)=x2i(0), dk(0)=x2i+1(0)

• Predict : dk(r)= dk(r-1) – pj(r) sk+j(r-1)

• Update : sk(r)= sk(r-1) + uj(r) dk+j(r)

• Normalize : sk(R)=K0sk(R), dk(R)=K1dk(R)

Two Main Stages

• Prediction and Update

d i-2 (k - 1 ) d i- 1 (k - 1 ) d i(k - 1 ) d i+ 1 (k - 1 ) d i+ 2 (k - 1 ) d i+ 3 (k - 1 )

s i- 2 (k - 1 ) s i-1 (k - 1 ) s i(k - 1 ) s i+ 1 (k - 1 ) s i+ 2 (k - 1 ) s i+ 3 (k - 1 )

... p 1 (k ) p 0 (k) p 0 (k ) p 1 (k ) ...

d i- 2 (k ) d i-1 (k ) d i(k ) d i+ 1 (k ) d i+ 2 (k ) d i+ 3 (k )

... u 1 (k ) u 0 (k) u 0 (k ) u 1 (k) ...

s i- 2 (k - 1 ) s i-1 (k - 1 ) s i(k - 1 ) s i+ 1 (k - 1 ) s i+ 2 (k - 1 ) s i+ 3 (k - 1 )
Prediction Stage

• A prediction rule : interpolation

– Linear interpolation coefficients: [1,1]/2
• used in the 5/3 filter
– Cubic interpolation coefficients: [-1,9,9,-1]/16
• used in the 13/7 CRF and the 13/7 SWE

x i- 1
x i- 1 xp x i+ 3
xp x i+ 3
x i+ 1 x i+ 1
x i- 3 x i- 3

x i- 2 x i+ 2 x i- 2 x i+ 2
-3 -1 0 1 3 -3 -1 0 1 3
Update Stage

• An update rule :
preservation of average (moments) of the
signal
– The update coefficients in the 5/3 are [1,1]/4
– The update coefficients in the 13/7 SWE are
[-1,9,9,-1]/32
– The update coefficients in the 13/7 CRF are
[-1,5,5,-1]/16
Example

• The 5/3 wavelet

– The (2,2) lifting scheme

xi2 xi1 xi xi1 xi2 xi3 xi4

s p lit

p0 p0 p0 p0 p r e d ic t

d k 1 dk h ig h p a s s

u0 u0 u p d a te

sk lo w p a s s
Example
• We have p0 = 1/2 by linear interpolation and the
detailed coefficient are given by
1
d k  xi 1  ( xi  xi  2 )
2
• In the update stage, we first assure that the average
of the signal be preserved
1
 sk 
k

2 i
xi

• From an update of the form, we have

s k  xi  A(d k 1  d k )
 s   x  2 A d  (1  2 A) x  2 A x
k
k
i
i
k
k
i
i
i
i 1

• From this, we get A=1/4 as the correct choice to

maintain the average.
Example

• The coefficients of the corresponding high

pass filter are {h1} = ½{-1,2,-1}

• The coefficients of the corresponding low

pass filter are {h0} = ⅛{-1,2,6,2,-1}

• So, the (2,2) lifting scheme is equal to the

5/3 wavelet.
Example
• Complexity of the lifting version and the
conventional version
– The conventional 5/3 filter
• X_low = ( 4*x[0]+2*x[0]+2*(x[-1]+x[1])-(x[2]+x[-2]) )/8
• X_high = x[0]-(x[1]+x[-1])/2
• Number of operations per pixel = 9+3 = 12
– The (2,2) lifting
• D[0] = x[0]- (x[1]+x[-1])/2
• S[0] = x[0] + (D[0]+D[1])/4
• Number of operations per pixel = 6
Conclusions
• The lifting scheme is an alternative method of
computing the wavelet coefficients
• Advantages of the lifting scheme:
– Requires less computation and less memory.
– Easily produces integer-to-integer wavelet
transforms for lossless compression.
– Linear, nonlinear, and adaptive wavelet transform is
feasible, and the resulting transform is invertible
and reversible.