![](https://csdnimg.cn/release/download_crawler_static/10931098/bg3.jpg)
ZHANG et al.: DATA-DRIVEN SYNTHESIS OF CARTOON FACES USING DIFFERENT STYLES 3
Fig. 2. The overall pipeline of our framework.
learn a feature dictionary from photo patches and replace
these patches with a sparse parametric representation during
the searching process. Zhang et al. [32] develop a framework
to synthesize face sketches trained on only one template
sketch use a multi-feature-based optimization model to select
candidate image patches.
The above approaches can create sketches that closely match
an input photo, but it is difficult for their frameworks to
generate cartoons in given artistic styles, especially some
smooth and fancy styles. Furthermore, in most cases they also
do not consider the beautification of stylized output, while our
framework can simultaneously handle different styles, portrait-
cartoon similarity and visual attractiveness.
General Face Stylization: Meng et al. [33] render artistic
paper-cut of human portraits by considering it as an inhomo-
geneous image binarization problem. Zhao and Zhu [11] use
templates to fit a mesh to a face and transfer a set of strokes
from a specific painting to a given photograph. Rosin and
Lai [34] generate highly abstracted yet recognisable portraits
by fitting facial models. However, their objective is not to learn
a style of an artist so their scheme cannot be applied to our
problem.
III. D
ATA COLLECTION AND ANALYSIS
Fig. 2 shows the overview of our optimization approach
for data-driven cartoon face synthesis. The input comprise a
portrait image, a realistic facial component database D
r
and
a cartoon facial component database D
c
of a pre-define style.
The output is a cartoon face with the same style of D
c
.
A. Database Collection
We start the approach with building three databases: one
for realistic full faces, one for realistic facial components and
one for stylized facial components. The full face database D
f
consists of representative portrait photos downloaded from
the Internet and contains 500 frontal or near frontal faces
for male and another 500 for female subjects. We select
photos to ensure that we cover enough different kinds of facial
components for creating various shapes. Next we extract all the
facial components from the faces in D
f
to build the database
of realistic facial components D
r
. We then manually pick
representative components of 20 chins, 30 eyebrows, 30 eyes,
16 noses, 30 mouths and 75 types of hair for both male and
female photographies from D
r
and use them for building the
database of cartoon facial components D
c
by asking artists
to draw a stylized version for each representative realistic
component. For each real eye we separately draw a stylized
version of a single-fold eyelid and a double-fold eyelid. Finally
we ask artists to compose cartoons of these faces in D
f
,by
selecting and composing stylized facial components which are
similar to the components of the realistic faces (Fig. 3).
We label every facial component in D
r
with the closest
stylized facial component from D
c
. Additional styles other
than D
c
can be easily added by asking artists to draw stylized
facial components in new styles for the elements of D
r
.
Since there are significant visual differences between different
races, e.g., Asian and Caucasian faces, we build databases for
different races respectively.
B. Learning the Selection of Components
Previous works all focused on enhancing the attractiveness
of real faces [35], [36]. In the contrast, studies on enhancement
of facial attractiveness of cartoons have not yet been reported,
even though the neural responses related to the evaluation of
the attractiveness of cartoon faces have been investigated [37].
Generally, facial components and their composition contribute
most for creating stylized faces. A face differs from another
not only in terms of different components, but also in its
attractiveness which is determined by details and the implicit
semantic relationships of all the components. Therefore, it is
not meaningful to evaluate the attractiveness of a face just
through each individual component. An expedient way to
increase attractiveness is to select stylized facial components
only according to predefined rules. However, it is not easy
to define such rules, because if we consider all the possible
combinations, the rules will quickly become intractable to
maintain with a growing number of facial components. On the
other hand, a restriction to a small subset of possible facial
components can avoid awkward synthesis but will result in
a limited variety and common artifacts such as creating the
same stylized faces for different inputs.
One possibility to encode various relationships and define
attractive combinations of facial components is to adopt a
data-driven approach based on observational data. Data-driven
approaches, including probabilistic graphical models have
recently proven to be successful in modeling abstract semantic