Color Model Based Real-Time Face Detection with
AdaBoost in Color Image
Yuxin Peng
∗
, Yuxin Jin
∗
,Kezhong He
∗
,Fuchun Sun
∗
, Huaping Liu
∗
,LinmiTao
∗
∗
Department of Computer Science and Technology, Tsinghua University, BeiJing, China
Email: {pyx05; jyx05}@mails.tsinghua.edu.cn
{hkz-dcs; fcsun; hpliu; linmi}@mail.tsinghua.edu.cn
Abstract— Face detection plays an important role in Computer
Vision. In the past few years, much effort has been put into this
field. Some approaches are based on Color Model which is highly
expensive. Viola and Jones proposed a real-time approach based
on AdaBoost, but it suffered from the risk of overfitting. In this
paper, we propose a method to take advantage of Color Model
and AdaBoost. A novel Face Region Model will be proposed
so as to eliminate a large quantity of non-face regions in only
several milliseconds. Skin Strong Classifiers will be trained based
on color. Thereafter, the remained regions will be presented
into AdaBoost System for refining. Experiment shows that our
approach is even faster. Its error rate is lower than or at least
equal to Viola and Jones’.
Index Terms— Color Model, Integral Skin Map, Face Region
Model, Face Detection, AdaBoost.
I. INTRODUCTION
Face Detection is a basic and important problem. It can be
applied into face recognition, User Interface (UI), security and
surveillance etc.. In these applications, robust and real-time
face detector is required.
Various approaches focused on the skin color due to its
special attributions. By clustering a number of skin pixels, the
candidate localizations of face can be identified. [3] proposed
a color model based on skin and non-skin pixels’ histogram
and concludes that the model based on histogram is better than
on Gaussian Model. However, its result was not satisfactory.
[1] introduced light compensation and extraction of eyes,
mouths features. But it was highly expensive and not suitable
for detecting profile view. [2] proposed a method for skin
detection based on Markov Random Field. But the false alarm
rate was high.
With the development of pattern recognition, AdaBoost was
introduced to face detection. In 2001, Viola and Jones [5]
proposed a robust and real-time scheme for face detection.
Based on their contribution, [6] extended the scheme into
multi-view face detection. Unlike the fact that color model is
easily influenced by illumination, current AdaBoost is robust
to the change of illumination because it is based on grayscale
texture ignorant of the color information. However, due to the
loss of information in color, more weak classifiers should be
trained to classify the faces from non-faces. In result, such
method encounters serious overfitting problem according to
[4].
Our approach takes the advantage of both color information
and AdaBoost.
Fig. 1. Frame Work of our approach. The Candidate Region will be firstly
filtered by FRM, and then passed to AdaBoost System consisting.
Before AdaBoost System, all sub-windows will be filtered
by a novel Face Region Model (FRM) based on a Skin Pixel
Model (SPM). A novel Skin Strong Classifier will be trained
under AdaBoost System. These methods are used to make a
fast elimination of a large quantity of sub-windows with the
help of Integral Skin Map (ISM). Experiment shows promising
result in high detection rate, low false positive rate and great
efficiency.
In Section 2, Color Model including SPM, FRM, and ISM
will be introduced. Section 3 will give a brief introduction
of our AdaBoost training strategy and structure. Experiment
results are given in Section 4. Conclusion is drawn in Section
5. Fig. 1 shows the frame work of our approach.
II. C
OLOR MODEL
Unlike many previous works [1], our skin model does not
aim to act as a classifier distinguishing faces from non-faces.
We just use the skin model to eliminate as many apparent non-
face regions as possible and run in a sufficiently short time.
Thus, our approach does not include the step to compensate the
illumination’s effect and the illuminated pixels will be chosen
as ordinary training examples. For real-time application, based
on a large quantity of skin and non-skin pixels, we construct
a fast Look-Up-Table (LUT) of discriminant function value
of single pixel. As to face detection, it is required to identify
which region can be viewed as face candidate.
Our Color Model includes two sub-models: Skin Pixel
Model (SPM) and Face Region Model (FRM). To further
reduce the computation load, Integral Skin Map (ISM) similar
to Integral Image [5] will be proposed at the end of this
section. The s urvived regions after elimination by FRM will
be feed forwarded to AdaBoost Face Detection.
978-1-4244-1625-7/07/$25.00
c
2007 IEEE 40
Authorized licensed use limited to: Guangdong Univ of Tech. Downloaded on May 11,2010 at 10:09:54 UTC from IEEE Xplore. Restrictions apply.
