Digital Signal Processing Reference
In-Depth Information
α
α
= 0. The remaining unconstrained pixels are
marked as unknown. The goal of a digital matting algorithm is then to compute the
values of
= 1 or as known background with
z
z
,
F
and
B
for all pixels labeled as unknown in a trimap.
Matting techniques proposed previously can be categorized into sampling-based
methods [4,5], affinities-based methods [6], or a combination of the two [7,8]. The
combined methods became recently popular, where local affinities are employed in
their optimization steps for solving or refining the matte. Although various successful
examples [3] have been shown for these approaches, their performance rapidly de-
grades, when foreground and background patterns become complex, due to two key
factors.
First, sampling pairs of foreground and background pixels is an important factor
for accuracy of matting because in many images the foreground and background re-
gions contain significant textures and/or discontinuities, thus direct color sampling
may become erroneous. Second, once the pairs of foreground and background are
sampled, determining
α
z
α
for each unknown pixel requires the solution of a large
linear system, whose size is directly proportional to the number of the unknown va-
riables, making the matting a time-consuming procedure. One of the solutions to im-
prove the efficiency of matting is to develop a method to choose the best pairs of
foreground and background samples for unknown pixels, which can promote the mat-
ting accuracy and improve the time-complexity at the same time.
To this goal, Wang et al. [7] proposed a robust matting method by optimizing color
samples. They proposed an evaluation function to sample best color pairs of fore-
ground and background for a specific unknown pixel, so the matte only needs to be
(re)computed for a small portion of the image at a time. But, long or complex fore-
ground's boundaries are still monotonous and time-consuming to trace. In [8], Gastal
et al. carefully sampled the pairs of foreground and background in a small neighbor-
hood, pronouncing that their approach could generate high-quality mattes up to 100
times faster than the previous techniques. However, the accuracy of their method was
shown to decrease largely when the assumption that the true foreground and back-
ground colors of the unknown pixels can be explicitly estimated by analyzing nearby
known pixels did not hold. This indicates that it is not enough just to decrease the size
of the sampled color pairs in order to improve the performance of matting both in
accuracy and the speed.
In this paper, we propose a modified matting method named IFENRM based on the
Robust Matting algorithm proposed in [7]. The IFENRM exploits an artificial im-
mune system [9], known as immune feature extraction network (IFEN) proposed in
[10], to address the accuracy-efficiency tradeoff in the image matting.
One of the key factors which make the Robust Matting robust and efficient method
is to use an optimized color sample. However, the optimized color samples are se-
lected from a large original sample set, which are collected spreading foreground and
background samples along the boundaries of known foreground and background re-
gions. This original sample set is surely a large one, so the optimization will be time-
consuming and the optimized samples probably become the simple averages of the
full sample space [7].
z
