Wavelet Transform Architectures,
LLH LHL LHHLLL
3.4 Dyadic or Octave Band Subband Tree Structure
Logarithmic Wavelet Transform.
The dyadic or octave b an d tree is a special irreg u lar tree structure. It
splits only the low er half of the spectrum into tw o equal b an d s at an y level of
the tree. Therefore the detail or higher-hand com ponent of the signal at any
level of the tree is decom posed no further. The dyadic tree configuration an d its
co rresponding frequency resolution for L = 3 are given in Figure 3.3.
LH
LLL LLH K K K
i T ~2
Figure 3.3: The Dyadic Subband Tree Structure - Logarithmic Wavelet.
K
A n exam ination of this analysis-synthesis stru ctu re show s th a t a half
resolution frequency step is used at each level. Therefore it is also called the
octave-band or constant-Q subband tree structure. First, low (L) an d h ig h (H)
signal ban d s are obtained here. W hile the b an d (L) p ro v id es a coarser version of
th e original signal, b an d (H) contains the detail inform ation. If the low spectral
com ponent or b a n d (L) is interpolated by 2, the detail inform ation or the
in terp o latio n error is com pensated by the interp o lated version of b a n d (H).
C onsequently, the original is perfectly recovered in this one-step dyadic tree
com ponent of the higher-level n ode of the tree. N ote th at the total d ata rate for
the sub b an d signals at the o u tp u t of the analysis stage equals , the d a ta rate
of the source signal. Therefore the dyadic tree is also critically sam pled.
This m ultiresolution (coarse-to-fine) signal decom position idea w as first
p ro p o sed for 2D signals by Burt an d A delson for vision an d im age coding
p roblem s [34]. In fact the orthonorm al w avelet transform also em ploys this
dyadic subband tree. In th at case, the coefficients in the m u ltireso lu tio n w avelet
decom position of a continuous-tim e signal, are calculated usin g the discrete
tim e dyadic sub b an d tree p resented here.
3.5 Discrete Wavelet Transform (DWT) Algorithm
W e have seen th at the 's, /:=0,...,2 A^- 1, constitute the im pulse
response of the QMF Fq and th at the im pulse response of the QMF is related
b y
For the m om ent, an d for ease of notatio n w e w ill restrict ourselves to sim plest,
alth o u g h the m ost localised, m em ber of the class of w avelets discovered by
D aubechies [36] [38], often called DAUB4 w hich has only four coefficients
Co,Cj ,C2,C3. In other w o rd s w e require to satisfy the ap p roxim ation condition of
o rd e r N = 2 . The nu m b er of coefficients increases by tw o each tim e N is
increased by one.
Let us consider the follow ing transform ation m atrix acting on a colum n vector
Cq C j C j C 3 C-, — Cr, C, — C,0 C , Cr, C3 C 2 Cj Cq Cz C3 C l - C o ^ 0 ^ 2 ^3 C3 - C2 Cl - Cq C q Cl C l — Cr, (3.8)
Blank entries to the above m atrix signify zeros. The stru ctu re of the m atrix is
such that, the o d d row s generate one com ponent of the data co nvoluted w ith
the filter Co,q,C2,C3, an d the even row s perform a different convolution w ith
coefficients Cj^,-C2,c^,-Cq. The last tw o row s w ra p -aro u n d like convolutions
w ith periodic b o u n d ary conditions.
The b ehaviour of the m atrix, is to perform tw o related convolutions, th en to
decim ate each of them by half, by th row ing aw ay half the values, an d finally to
interleave the rem aining halves.
This results in the o u tp u t of Fq, reduced b y 50%, accurately rep resen tin g the
d a ta 's "approxim ation" (trend, sm ooth) inform ation. The o u tp u t of F ,, also
decim ated, is referred to as d ata's "detail" (fluctuation, difference) inform ation.
It m u st be possible to reconstruct the original d ata vector of length L from its
L/2 detail or d - com ponents. That is achieved b y req u irin g the above m atrix to
be orthogonal, so th at its inverse is just the tran sp o sed m atrix. In oth er w ords,
w e require the W avelet Filter Coefficients of this m atrix to satisfy all the
conditions th at w e have seen before in Section 2.4.
"approxim ation" vector of length L /2 , th en to the "approxim ation-
approxim ation" vector of length L /4 , an d so on u n til only a trivial n u m b er of
"approxim ation - ... - approxim ation" com ponents (usually 2) rem ain. The
p ro ced u re is som etim es called a P yram id A lgorithm , for obvious reasons. The
o u tp u t of the DWT consists of these rem aining com ponents an d all the "detail"
com ponents th at w ere accum ulated along the w ay. A d iag ram sh o u ld m ake the
pro ced u re clear:
- a , - a , ' A , ' A , ' a ; ' ^ 2 a j 0 , A a A a ' X3 f l j a , ^ 2 A 3 A , X4 d . a . A A 4 A a ^ 5 « 3 a ; A 3 A A ^ 6 d . a s A A D a a # a . A 4 A A X g W FCM ^ d . Permute ^ W FC M ^ D . Permute ^ O 4 W F C M ^ A a ; d , d , d . r f . • ^ 1 0 d s d . 4 4 A X i i a & d . d . 4 ^ 3 ^ 1 2 d . 4 d . d . d . • ^ 1 3 a , d s 4 d s 4 d . d . d . < • ^ 1 5 a s d , d , d . A . ^ 1 6 . d i _ d s _ A . _ d g _ _ d g _
If the length of the d ata w ere a h igher p o w er of tw o, there w o u ld be m ore
stages of ap plying the w avelet coefficients m atrix an d p erm uting. A t the en d
there w ill alw ays be a vector w ith tw o A " s an d a sequence of A 's, D 's, d 's,
etc. W e can notice from the above m atrix, th a t once d 's are generated, they
The value d . , of any level is term ed a "w avelet transform detail coefficient" of
the original signal vector. The final values A,", should strictly be called
"w avelet transform m other-approxim ation coefficients", alth o u g h the term
"w avelet coefficients" is often used loosely for b o th d ' s an d final A " s . Since
the full p ro ced u re is a com position of orthogonal linear operations, the w hole
DWT is itself an orthogonal linear operator. In o rd er to in v ert the DWT, w e can
sim ply reverse the procedure, starting w ith the sm allest level of the hierarchy
a n d w orking the above pro ced u re from the rig h t to the left. The follow ing
inverse m atrix is u sed instead of the first m atrix (3.8).
C0 C3