Rob Reeves
directly from the DCT of the original sequence. This and related properties have been reported previously (Reeves, 1999
Reeves and Kubik, 1998), and are summarized below. ME
We define the forward discrete transform and its inverse as
N—1
G(m) = T{g(n)} = X g(n)fa(m) i
n=0
and
N-1
g(n) = T HG (m)} = S G (m)rg (n) ()
m=0
where f,,(m) represents the forward transform kernel, r,, (1) represents the reverse transform kernel, and n and m are
integers from 0 to N — 1. The type-2 DCT (Rao and Yip, 1990) is defined by
v2
falm) = ran) = A cos((2n + 1)mx/2N) 0
with c(m) — A form = 0 and c(m) = 1 otherwise. We assume that the sequence g(n) is derived by sampling,
in accordance with the Nyquist criterion, a band-limited continuous signal g(x) at points x = n,n = 0...N — 1, wi
sampling interval taken as one without loss of generality. By considering a half-range cosine series expansion of g(x)
around x = — 1, we can show that,
N—1
glx) = So ra (z)G(m) (4)
m=0
This expression is equivalent to the reverse discrete transform, except that it gives the continuous band-limited function
g(x) instead of sequence g(n), and the discrete variable n has been replaced by continuous x in the reverse transfom
kernel. If g(z) satisfies Dirichlet’s conditions, it’s derivative can be computed by a term by term differentiation of its
Fourier series (James et al., 1993). Differentiating both sides of Equation 4 gives
d p N—1 7
1,79 = 2 Glmirm(a) = 2. G(m) rm (a) 6
Adopting the notation ¢'(n) to mean the sampled derivative of g(x) at z = n, and r^, (n) to refer to the sampled deriva
of Taiz) at T = n gives
N-1
gn) — S G (m)r;, (n) (6
m=0
It follows from Equation 1 that
N—1 N—1
Tig} = M Gp fot.) 0
p=0 n=0
where p is used as an additional index for the transform coefficients. Equation 7 represents a simple linear transfom
which computes T | 9 (n)) from the values of 7'{g(n)}. This property can be extended similarly to the second and higher
derivatives, and to any linear operation on g(z), including the terms involving derivatives required to formulate the les
squares problem in the DCT domain. This single linear transform is equivalent to reconstructing the continuous signi.
differentiating it, sampling it, and taking the DCT. The extension to two dimensions is straightforward.
3 LEAST SQUARES MATCHING
Our work is based on the formulation of Ackerman (Ackermann, 1984), who uses an affine transformation to model the
transformation of left image patch to right image patch as follows,
gi(z,y) = ho + h1g(ao + a1x + a2y, bo + biz + bay) + M (x,y)
and N
g2(x,y) = g(x,y) + na(x,y) 0
762 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
wher
a; an
Using
where
discre
and gt
The s!
with tl
where
betwe
Since |
re-soh
4 LS
By che
matrix
where
TA ar
of Equ
images
order c
(low vz
1991).
order I
reduce
frequer
depend
the qua
in the 1
involvi
5 EX
À pixel
derivati
and co]
parame
constru
conduc
constru
—