Graphics Reference
In-Depth Information
0.9
+ (1−
α
)
0.9
α
1.09
1.07
1.05
1.03
1.01
0
0.2
0.4
0.6
0.8
1
α
Figure 2.15.
The cost function in Equation (
2.43
) as a function of
α
, with
ρ
= 0.9.
(a)
(b)
Figure 2.16.
(a) The eight nearly binary matting components computed using spectral matting
for the image in Figure
2.14
a. (b) The four selected matting components are summed to give an
estimate of the full matte.
2.4.6
Learning-Based Matting
Zheng and Kambhamettu [
579
] described a generalization to the color line model
described previously that enables what they called
learning-basedmatting
. Suppose
we revisit the assumption about how the
α
values and image colors are related in a
window, that is, that
a
I
i
+
α
i
=
b
(2.46)
In closed-form matting, we eliminated
a
and
b
from the estimation problem
entirely; that is, we never directly estimated or recovered these values. On the other
hand, suppose that we knew
α
and
I
within a window
w
j
of pixel
j
; we could compute
a
and
b
directly:
a
I
i
+
2
argmin
a
,
b
w
j
(α
i
−
(
b
))
(2.47)
i
∈
