Skin Color Based Face Detection Using Multi-Layer Perceptron
Taro Hosei
Faculty of Computer and Information Sciences
Hosei University
E-mail: xxxxxxxx@cis.k.hosei.ac.jp
Abstract
This paper addresses the very complex, challenging
problem of human face detection in color images with
complex background and a variety of lighting conditions.
We propose a new, skin-color-based face detection
method using a multi-layer perceptron (MLP) as applied
to a two-class classification problem between “skincolor” and “non-skin color.” The proposed method has
two key ideas. One is effective skin color detection using
MLP trained in a combined HLS and YCbCr color space.
The other is a hierarchical, sequential process of face
detection: segmentation and grouping of skin regions,
face candidate selection, and verification of face
candidates. This process utilizes not only image
processing techniques such as connected component
analysis, morphological operation, and contour tracing
but also the top-down knowledge about facial
characteristics such as the aspect ratio of a face and the
existence of eyes as holes in a skin region. Experimental
results using our personal photo collection demonstrate
that the proposed method achieves a high and
encouraging face detection rate of 80.79% while
suppressing a false positive detection rate.
1. Introduction
Human face detection is a crucial step in a wide variety
of personal identification applications such as human
computer interface, face recognition, image retrieval using
face images, face image database management and video
surveillance [1].
Various approaches to face detection are proposed.
For recent surveys on face detection, see the excellent one
[2]. Conventional techniques are classified into tow major
approaches: appearance-based methods versus featurebased methods. The former includes template matching
using deformable models and probabilistic/statistical
pattern matching using principal component analysis [3],
eigenface [4], neural network, support vector machines,
and hidden markov model. The latter utilizes facial
features, texture, and skin color [5] together with the rules
Supervisor: Prof. Toru Wakahara
about the relationships between facial features to verify
the existence of a face.
Our approach belongs to the feature-based methods, in
particular, focusing on skin color information. The key
idea of the proposed method is application of multi-layer
perceptron to a two-class classification problem between
“skin color” and “non-skin color” in a combined color
space of HLS and YCrCb. Moreover, we propose an
effective, sequential procedure of face detection making
full use of the empirical knowledge about facial
characteristics such as the aspect ratio of a face and the
existence of eyes ;as holes in a skin region.
Experiment results using our personal collection of
photos (both indoors and outdoors) demonstrate that the
proposed method is robust against a wide variation of
facial color, position, scale, orientation, expression, and
lighting conditions.
Input image
Output image
1) Skin color detection
4) Face area determination
2) Noise rejection
3) Face candidate detection
Figure 1. Total flow of face detection in the
proposed method.
2. Face image database
We use our personal photo collection. Our collection
includes 152 photos of friends and landscape. Each of
these images has zero or more faces with various
directions in complex backgrounds. 177 faces are
contained in this photograph collection in total. All
images we used have 640×480 pixels in a PPM color
format.
One example of our photo collection is shown in Figure
1. Figure 1 shows the total flow of the proposed method
of face detection explained in the following sections.
7. Conclusion
We described a novel face detection method for color
images, using MLP and sequential procedure of face
determination. Experimental results demonstrate that the
proposed method achieves a high and encouraging face
detection rate while suppressing a false positive detection
rate, and it is robust against a wide variation of face. It is
easy to customize for the purpose because the control of
the performance is possible by decreasing the step of the
face area determination.
References
[1] A. Jain , R. Bolle, and S. Pankanti, Eds., BIOMETRICS 
Personal Identification in Networked Society , Kluwer
Academic Publishers, 1999.
[2] M.-H. Yang, D. Kriegman, and N. Ahuja, “Detecting faces
in images: a survey,” IEEE Trans. Pattern Anal. Machine
Intell., vol. 24, pp. 34-58, 2002.
[3] B. Menser and F. Muller, “Face detection in color images
using principal components analysis” in Proc. Image
Processing and its Applications, vol. 2, pp. 620-624, 1999.
[4] B. Moghaddam and A. Pentland, “Probabilistic, view-based,
and modular models for human face recognition,” in
Handbook of Image and Video Processing, pp. 837-852,
Academic Press, 2000.
[5] R.-L. Hsu, M. Abdel-Mottaleb, and A. K. Jain, “Face
detection in color images,” IEEE Trans. Pattern Anal.
Machine Intell., vol. 24, pp. 696-706, May 2002.
[6] S. Madhvanath, G. Kim, and V. Govindaraju, “Chaincode
contour processing for handwritten word recognition,” IEEE
Trans. Pattern Anal. Machine Intell., vol. 21, pp. 928-932,
1999.
[7] K. Miki, “Face detection using skin color and correlogram,”
in Graduation Theses 2003: Extended Abstracts, Faculty of
Computer and Information Sciences, Hosei University,
pp.423-426, 2004.