ORIGINAL ARTICLE
Alpha matting with image pixel correlation
Xueming Yan
1
•
Zhifeng Hao
2
•
Han Huang
3
Received: 16 September 2015 / Accepted: 8 August 2016
Ó Springer-Verlag Berlin Heidelberg 2016
Abstract Alpha matting aims at estimating the foreground
opacity matte in an image. It is critical to find the best
known samples for foreground and background color of
unknown pixels in color sampling-based matting approa-
ches. Most matting approaches can only select color sam-
ple for each pixel each time, which prevent them from
maintaining the same correlation in image pixels. In par-
ticular, they may fail to collect appropriate samples from
complex images and thus lead to artifacts. In order to solve
the problem, we present a correlation-based sampling
method in which the image pixel correlation is employed in
color sampling and optimal sample selection for the alpha
matte estimation of the image. First, the foreground and
background colors of sample set can completely cover the
color of unknown pixels to avoid missing the true samples.
This is accomplished by artificial immune network adap-
tively learning the image correlation in unknown pixels.
Besides, we propose the sample selection process as a
global optimization problem with image correlation. All
unknown pixels are treated as a high-dimensio nal input
variable, particle swarm optimization algorithms is
employed to solve the global optimization problem
selecting the best sample pairs for all unknown pixels. The
experimental study on images dataset shows that image
pixels correlation is effective to improve matting, and that
our matting results are comparable to some recent
approaches.
Keywords Image pixels correlation Complex im age
Artificial immune network Particle swarm optimization
Alpha matte estimation
1 Introduction
Image matt ing refers to the problem of extraction and
compositing the foreground objects in image and video
editing tasks [1]. Matting technical aims at decomposing an
image into two layers as the foreground colors (F) and
background colors (B) of an image, along with the opacity
values (a), which is a convex combination under the
compositing equation:
I
z
¼ a
z
F
z
þð1 aÞB
z
ð1Þ
where I
z
is the observed color of pixel z, and it is expressed
as a linear combination of unknown color F
z
and unknown
color B
z
with the unknown alpha matte a
z
. There are three
unknowns for each pixel z with only one constraint, and
this is a severely under-constrained problem [1]. Therefore,
an user-specified trimap (specifying known foreground/
background and unknown pixels) is often required.
Prior matting methods can be divided into four cate-
gories: propagation-based, sampling-based, combination of
sampling-based and propagation-based, and learning-based
methods. Propagation-based methods do not explicitly
& Han Huang
hhan@scut.edu.cn
Xueming Yan
csyan_xm@scut.edu.cn
Zhifeng Hao
mazfhao@scut.edu.cn
1
School of Computer Science and Engineering, South China
University of Technology, Guangzhou 510006, People’s
Republic of China
2
School of Mathematics and Big Data, Foshan University,
Foshan 510006, People’s Republic of China
3
School of Software Engineering, South China University of
Technology, Guangzhou 510006, People’s Republic of China
123
Int. J. Mach. Learn. & Cyber.
DOI 10.1007/s13042-016-0584-1