J Sign Process Syst (2014) 74:59–67
DOI 10.1007/s11265-013-0809-4
Colorization for Gray Scale Facial Image
by Locality-Constrained Linear Coding
Yang Liang ·Mingli Song ·Jiajun Bu ·Chun Chen
Received: 15 January 2013 / Revised: 9 June 2013 / Accepted: 11 June 2013 / Published online: 12 July 2013
© Springer Science+Business Media New York 2013
Abstract Colorization for gray scale facial image is an
important technique in various practical applications. How-
ever, the methods that have been proposed are essentially
semi-automatic. In this paper, we present a new proba-
bilistic framework based on Maximum A Posteriori (MAP)
estimation to automatically transform the given gray scale
facial image to corresponding color one. Firstly, the input
image is divided into several patches and non-parametric
Markov random field (MRF) is employed to formulate the
global energy. Secondly, Locality-constrained Linear Cod-
ing (LLC) is employed to learn the color distribution for
each patch. At the same time, the simulated annealing algo-
rithm is employed to iteratively update the patches chosen
by LLC to optimize the MRF by decreasing global energy
cost. The experimental results demonstrate that the pro-
posed framework is effective to colorize the gray scale facial
images to corresponding color ones.
Keywords Colorization · MAP · MRF · LLC
Y. Liang · M. Song () · J. Bu · C. Chen
Zhejiang Provincial Key Laboratory of Service Robot,
College of Computer Science, Zhejiang University,
Hangzhou 310027, China
e-mail: brooksong@zju.edu.cn
Y. Liang
e-mail: liangyang@zju.edu.cn
J. Bu
e-mail: bjj@zju.edu.cn
C. Chen
e-mail: chenc@zju.edu.cn
1 Introduction
Human face conveys lots of information such as identity,
appearance, etc. during social communication or in security
systems. However, sometimes only gray scale images or
portraits can be obtained whose facial details like skin color,
lip luster, etc. are lost. For example, because of the qual-
ity of surveillance cameras or storage space restriction, only
gray scale images may be obtained. The same for archeolo-
gists, they may only obtain historical gray scale images. So
colorization for gray scale facial image is useful.
In 1970 Wilson Markle [1] introduced the term coloriza-
tion to describe the computer assisted process of adding
color to a gray scale image. Several approaches have been
presented towards successful image colorization. In general,
the existing colorization approaches can be classified into
two groups, one is example-based method and the other is
optimization-based method.
The example-based method colorizes the gray scale
image by choosing a similar image as a reference, and trans-
ferring its colors to the input gray scale image [2, 3]. Welsh
in [2] transfers the entire color mood by matching the lumi-
nance and texture of the example image to the target gray
scale image pixel by pixel. But it always fails on those
images that are hardly segmented by luminance or texture.
Kekre and Thepade in [3] searches the reference color image
for the same palette of a certain scale pixel window by
matching the luminance values of reference color image to
target gray scale image. However, a suitable reference color
image whose color mood and composition are required to be
similar to the target gray scale image may take effort to find
and it may fail on those images that are hardly segmented
by luminance or texture.
The optimization-based method colorizes the image
based on the color label priors offered by users [4–6]. Levin