The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
proposed method will be explained. The fourth chapter will
present the experimental results.
2. THE FRAMELET TRANSFORM
As it is well known, except for the Haar filterbank, two-band
finite impulse response (FIR) orthogonal filterbanks do not
V j =Span{f(2 j t-k)}
k
W i j = Span{tp‘(2 J t-k)}, i = 1,...,N-1
k
with
allow for symmetry. In addition, imposition of orthogonality
for the two-band FIR filterbanks requires relatively long filter
support for such properties as a high level of smoothness in the
resulting scaling function and wavelets, as well as a high
and the corresponding scaling function and wavelets satisfy the
following multiresolution equations:
approximation order. Symmetry and orthogonality can both be
obtained if the filterbanks have more than two bands.
Furthermore, due to the critical sampling, orthogonal filters
suffer a pronounced lack of shift invariance, though the
desirable properties can be achieved through the design of tight
frame filterbanks, of which orthogonal filters are a special case.
№ = ^K(.k)<t>(2t-k)
k
ip‘(t) = s[2£/?,.(k)<f>(2t-k), i = \,...,N-l.
k
In contrast to orthogonal filters, tight frame filters have a level
of redundancy that allows for the approximate shift invariance
Now, the frame bound A takes on the value
behavior caused by the dense time-scale plane. Besides
producing symmetry, the tight frame filterbanks are shorter and
result in smoother scaling and wavelet functions. Before
proceeding further, let me describe the basic concepts of frame
theory and oversampled filterbanks
1 N ~ l
^ i=0
2.1 A Tight Wavelet Frame
2.2 Oversampled Filterbanks
A set of functions jy/in a square integrable space,
L 2 (□ ) , is called a frame if there exist A > 0, B < oo so that, for
any function / £ Z, 2 (D ),
The frame condition can be expressed in terms of oversampled
polyphase filters. Given a set of N filters, we define them in
terms of their polyphase components:
S||/||\ (1)
¿=1 /60 ieO
Hfz) = H j0 (z 2 ) + z-'H.fz 2 ), i = 0,..., N -1,
where
= 0,1.
where A and B are known as frame bounds,
{f,g) := J f(*)g(x) dx , and ||/|f := (/,/).
Now define the polyphase analysis matrix as
The special case of A - B is known as tight frame. In a tight
frame we have, for all / ei 2 (0),
ZEl|(/>'< 2 '-*))f='<|/f
1=1 /60 *e0
' H oA z ) H 0 fz) '
H XQ (z) Hyi(z)
H (z) = . .
K H N ., 0 (z) H N _ xx (z)^
which implies
If we now define the signal X(z) in terms of its polyphase
components, we then have
/=1 jé0 £eQ
x(z) = (X 0 (z) X x (z)) T ,
In order to have a fast wavelet frame (or framelet) transform,
multiresolution analysis is generally used to derive tight
wavelet frames from scaling functions.
where the equation X t (z), / = 0,1 is defined in terms of the
time domain signal, x(ri), as follows:
Now, we obtain the following spaces: