Progressive Data Access for Regular Grids 159
Equation 8.22 sheds some light on the choice of wavelet. As discussed ear-
lier (see Eq. 8.17), a wavelet possessing a higher number of vanishing moments
will permit the compaction of more energy (information) into fewer expansion
coefficients. Figure 8.4 qualitatively compares the compression of a signal by
a factor of 8, by using wavelets with different numbers of vanishing moments.
For ease of comparison, Figures 8.4a and 8.4b reproduce Figures 8.2a and 8.2b,
respectively. Results using the Haar (Fig. 8.4c) and Daubechies D4 (Fig. 8.4d)
wavelet are shown with one and two vanishing moments, respectively. Note
the “blockiness” resulting from the box shape of the Haar wavelet. Also note
that while the multiresolution approximation of Figure 8.4b may appear vi-
sually more appealing than Figure 8.4c, many extreme values in the original
signal are lost, which is the result of the cascade of applications of the low-pass
scaling filter.
Finally, unlike the pure multiresolution approaches based on Z-order curve
and wavelets, an arbitrary number of approximations are now produced. In
the extreme case, the approximation is refined one coefficient at a time.
8.4.5 Boundary Handling
Finite length signals present a couple of challenges that have been ignored
until this point. Consider the application of the DWT to a signal x[n]witha
low-pass filter h of length L = 4. By Equation 8.18, the computation of c[0]
is given by
c[0] =
˜
h[3]x[−3] +
˜
h[2]x[−2] +
˜
h[1]x[−1] +
˜
h[0]x[0].
But, for a finite signal, the samples x[−3], x[−2], and x[−1] do not exist! One
simple solution is to extend x[n] so that it is defined for n<0andn ≥ N.
There are a number of choices for the extension values, such as zero padding
or repeating the boundary samples, each with its own set of trade-offs. In the
general case, however, if the input signal, x, is extended, creating a new input
signal, x
e
of length N
e
>N, the resulting output after filtering and downsam-
pling no longer has exactly N non-zero and non-redundant approximations
and detail coefficients. Moreover, perfect reconstruction of x
e
requires all of
the N
e
coefficients. Due to downsampling in the case of the DWT, each iter-
ation of the filter bank requires an extension of the incoming approximation
coefficients. For a 1D signal, this overhead may not be significant, but later,
when moving on to 2D and 3D grids, the additional coefficients can consume
a substantial amount of space.
A filter bank that is nonexpansive is more preferable, as it preserves the
number of input coefficients on output, and does not introduce discontinuities
at the boundaries. The solution is the employment of symmetric filters com-
bined with a symmetric signal extension. Only the case of signals of length
N =2
n
, and odd length filters where the symmetry is about the center bound-
ary coefficients, is considered here. For an extensive discussion on symmetric