International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
2. THE 1-D DUAL-TREE COMPLEX WAVELET
TRANSFORM
It is well-known that the real biorthogonal wavelet transform
can provide PR and no redundancy, but it is lack of shift variant.
Then Kingsbury (Julian Magarey! 1 1998;N. Kingsbury, 1998a ;
Nick Kingsbury! 1998b; Serkan Hatipoglu, 1999) has developed
a dual-tree algorithm with a real biorthogonal transform, and an
approximate shift invariance can be obtained by doubling the
sampling rate at each scale, which is achieved by computing
two parallel subsampled wavelet trees respectively.
For one dimension signal, we can compute two parallel wavelet
trees. There is one sample offset delay between two trees at
level 1, which is achieved by doubling all the sample rates. The
shift invariance is perfect at level 1, since the two trees are fully
decimated. To get uniform intervals between two trees beyond
level 1, there have to be half a sample delay. The term will be
satisfied using odd-length and even-length filters alternatively
from level to level in each tree. Because we use the decimated
form of a real discrete wavelet transform beyond level 1, the
shift invariance is approximate.
The transform algorithm is described as following. Its process is
illustrated by fig.1.
1D COMPLEX WAVELET TRANSFORM:
> Atlevel 1, there is one sample offset between the trees.
(at), (a * a7). dy, =" zg"),
Gi an], O 1 rase
(aj), *(a * h^) (di), 2 (a * g^,
> Beyond level 1, there must be half a sample difference
between the trees.
ju > e +1 j e
(a; Yan = (a) *h Yan (d) > = (a; "o )an
je = | 3 o +1 j o | |
(a, I? =z (ay + h dant (dj, ). = (aj > 8 )»a 2
: — Ah“ | > di}
dh Ja —a— E
Tree A j T 2 ,
pi» a +. g | 2 ——d;
— 8 | 2 | d', >
qi up on mer
is — hl | | 2° Ld.
E h ]2 — a,—
— ; bend — g^ | 2" dl
Tree B . ud ! L i
pps | +de
Level 1£ j z16 Level 2f j=2©
Figure 1. The unidimensional dual tree complex wavelet
transform
The details 4, and g, can be interpreted as the real and
imaginary parts of a complex process z = d, +id,- The essential
property of this transform is that the magnitude of the step
response is approximately invariant with the input shift, while
only the phase varies rapidly (see (Nick Kingsbury! 1998a) for a
good illustration).(A. Jalobeanu , 2000)
It is not really a complex wavelet transform, since it does not
use any complex wavelet. It is implemented with real wavelets.
Classical complex-valued wavelet transforms can provide appr-
oximate shift invariance and good directionality, but PR and
good frequency characteristics cannot be obtained using compl-
ex filters in a single tree. At the same time, it is different from a
real wavelet transform because of the variety of filters. At level
1, the filter in tree A are odd-length filter, is same to tree B.
Beyond level 1, the filters in two trees are different, and they are
different between different levels in each tree. Hence the
wavelet functions varies continuously from level to level, whic-
h is quite different from the classical multi-resolution analysis.
Reconstruction is performed independently in each tree, and the
results are averaged to obtain 4° at level 1, for symmetry
between the two trees. This is illustrated by the following
algorithm and fig 2.
1D INVERSE COMPLEX WAVELET TRANSFORM:
€ Level; (j>0)
(a^), sb! * 5), «di *g*),
i m T. o ) + 0 7 f
(ag), = (; rot rudi tg), 3
e At j 20:
E ; C - 3. 141.1
atis E * h^), tr (d'* g^), + *h^), (dL * 8"),)
where { ER yg n-2p , x, if n=2p+l
, |o i nz2pil ^" 0 if n=2p
a, —- 21 h —
; $—a,, — 2 | nn
Dp | 4| i « bes j i | Tree A
€ RD 24 r g —— 3 [em
d'—5 2155 8$" —
ewe Sy
itn STE rir
ay—s 27 pe Mt ui |
i j | | |
; n ga. —— 21 LE nma
did 21 g — = LE
- 1 | |
d, 21 Eug Tree B
Level 2£ j 26 Level 1£ j=10
Figure 2. The unidimensional dual tree inverse complex wavelet
transform
530
A MO en