Information Technology Reference
In-Depth Information
Fig. 7
A schematic representation of temporal weighting in MASK for upscaling the
-
th block (
y
,
t
) of the frame at time
t
. First, we locate the neighboring blocks (
y
,
i
for
i
=
−
2
,−
1
,
1
,
2) indicated by the motion vectors (
m
,
i
). Then, we compute the motion re-
liability (η
,
i
) based on the difference between the
-th block at
t
and the neighboring blocks,
and combine the temporal penalization by
K
t
with the spatial kernel function
K
.
More specifically, we define
η
,
Δ
t
and
K
t
as
η
,
Δ
t
=
y
,
t
−
y
,
t
+Δ
t
F
M
,
(31)
1
1 +
η
,
Δ
t
h
t
K
h
t
(
η
,
Δ
t
)=
(32)
where
h
t
is the (global) smoothing parameter, which controls the strength of tempo-
ral penalization,
y
,
t
is the
th
block of the frame at time
t
,
t
+
Δ
t
is a neighboring
frame,
M
is the total number of pixels in a block, and
·
F
is Frobenius norm.
We replace the temporal kernel in (28) by (32). This temporal weighting technique
is similar to the Adaptive Weighted Averaging (AWA) approach proposed in [25];
however, the weights in AWA are computed pixel-wise. In MASK, the temporal ker-
nel weights are a function of radiometric distances between small pixel blocks and
are computed block-wise.
4.3
Quantization of Orientation Map
The computational cost of estimating local spatial steering (covariance) matrices is
high due to the SVD. In this section, using the well-known technique of
vector quan-
tization
[5], we describe a way to obtain stable (invertible) steering matrices without
using the SVD. Briefly speaking, first, we construct a look-up table which has a cer-
tain number of stable (invertible) steering matrices. Second, instead of computing
the steering matrix by (18), we compute the naive covariance matrix (15), and then
find the most similar steering matrix from the look-up table. The advantages of using
the look-up table are that (i) we can lower the computational complexity by avoid-
ing singular value decomposition, (ii) we can control and trade-off accuracy and
