Graphics Reference
In-Depth Information
x-1, y-1
x-1, y
x-1, y+1
x-1, y-1
x-1, y
x-1, y+1
x-1, y-1
x-1, y
x-1, y+1
x, y-1
x, y
x, y+1
x, y-1
x, y
x, y+1
x, y-1
x, y
x, y+1
x-1, y
x-1, y+1
x-1, y-1
x-1, y
x-1, y-1
x-1, y+1
x, y-1
x, y+1
x, y-1
x, y+1
x, y-1
x, y+1
Case 1
Case 2
Case 3
Figure 3.32. Pixel configurations introduced by seam removal. The shaded seam pixels in the
top row are removed to produce the new pixel arrangements in the bottom row. New neighbors
are indicated by bold lines in the bottom row. The bold lines in the top row represent the edges
cut in a graph-cut formulation of the problem.
the forward energy avoids introducing visual artifacts. For this reason, the forward
energy is usually preferred in implementations of seam carving.
3.5.3
Patch-Based Methods
As we saw in Section 3.4.2 , patch-basedmethods enabled high-quality image inpaint-
ing; as we discuss here, they also enable high-quality image retargeting. Furthermore,
they naturally enable image recompositing, inwhich elements of one or more images
can be seamlessly combined, deleted, and rearranged.
The key concept of patch-based methods is called bidirectional similarity , pro-
posed by Simakov et al. [ 451 ]. Suppose we want to retarget an image I into a new
image I . Simakov et al. argued that a good retargeting should have two properties.
First, the new image I should be complete : it should contain as much of the visual
information from I as possible. Second, the new image should be coherent : it should
not contain any visual information not present in the original image. The idea is
illustrated in Figure 3.34 .
If we consider each image as a collection of overlapping patches at different scales,
this means that every patch of I should be present in I (for completeness) and every
patch of I should be present in I (for coherence). This leads to a simple bidirectional
similarity measure for the pair
I , I )
(
:
N
1
1
N
I , I ) =
) +
)
D
(
min
d
(
,
min
I d
(
,
(3.40)
I
I
I
is a patch in I ; these patches are effectively
generated at multiple scales, as described next. We assume N patches are created in I
Here, each
is a patch in I and each
 
Search WWH ::




Custom Search