Information Technology Reference
In-Depth Information
Ta b l e 1
Optimal parameter settings for variational TSR. The eleventh parameter of the algo-
rithms is the convergence threshold set to 10
−
7
in all tests.
Flow without GCA
Flow with GCA
Multiresolution
Scale factor
1.04
1.04
Levels
55 or 75
55 or 75
Flow
Fixed point iterations
10
5
Relaxation iterations
40
20
λ
3
in (6)
30
100
λ
2
in (7)
1
1
γ in (7)
0
100
Intensities
Fixed point iterations
5
5
Relaxation iterations
10
10
λ
s
in (5)
1
1
λ
t
in (5)
50
50
judging from stills, there where minor degradations, thus we recommend
λ
t
:
λ
s
=
50:1.
The number of flow iterations needed are higher than the number of intensity
iterations needed, which illustrates the larger complexity of flow calculations.
Table 1 also shows that without GCA in the flow we need more iterations to get
optimal flows. This is because fewer point give reliable flow information when only
evaluating brightness constancy, which increases the need for flow diffusion by the
regularization term on the flow,
E
3
in (6). The
E
3
-term is weighed 30 times over
BCA when we do not use GCA, but is given the same weight as GCA when GCA
is used (with the BCA nearly neglected due to the
= 1:100 ratio).
The number of levels in the multiresolution pyramid is set to either 55 or 75
depending on the frame size of the given image sequence with a 1.04 (coarse to fine)
scale factor between the levels. The low scaling factor ensures good information
transfer down through the pyramid, but increasing it would give a speedup (the
effect on quality have not been thoroughly investigated).
We also conducted experiments using zero as initial values for both flows and
intensities in the new frames at the top level. Given the many levels we use, the
error introduced was corrected down through pyramid, showing great robustness
against bad initializations.
λ
2
:
γ
3.2
Evaluation Methodology
As the human visual system is the final judge when evaluating the enhancement
achieved, we will focus on subjective results (although we have not used a stan-
dardized and complex evaluation as described in [33]). To give a broader valida-
tion, we have also given objective results. Still images are an borderline accept-
able and easy way to evaluate the quality of frame doubling, but it is imperative to
