2014 Fifth International Conference on Intelligent Systems Design and Engineering Applications

Fast Face Detection based on Wavelet Transform in the Color Image

Niu Ai-min
College of Information and Engineering, Shandong Yingcai University, Jinan 250104, China
[email protected]

Abstract—We have presented a fast wavelet transform based automatic face detection system on the basis of the previous
face detection algorithm. Firstly nonlinear transform has been research on face detection.
done against image, then using wavelet transform extract The paper is organized as follows. In the next section, the
facial high frenquency components. At the same time the principle of wavelet transform algorithm is given. In Section
hidden function of neural is repalced by wavelet kernel
3, wavelet transform algorithm with neural network is
function. The result of experiment shows that the new
proposed. In Section 4, in order to test the performance of
algorithm can reduce the detection time and can get the high
detection accuracy. It can be referenced by the practical face
the algorithm, this paper test performance of different
detection system. algorithms. Finally, we conclude our paper in section 5.
II. THE PRINCIPLE OF WAVELET TRANSFORM
Keywords- Wavelet Transform; face detection; color image;
ALGORITHM
neural network

(t ) is square integrable and its Fourier transformation

( w)
2
I. INTRODUCTION
Recently, human face detection, as a key technology in
is
 dw   ,
(t ) is called base wavelet.
R w
human face information processing, is becoming a big
1 tb
problem that gathers more attention in the field of pattern
a ,b (t ) 
( )
recognition and computer vision[1,2]. And now it has been a a
applied into other application such as entrance security,
video coding, video surveillance and tracking and content
a, b are constants. a  0 . Base
(t ) is shifted and is
based image retrieval etc[3,4]. The study on face detection in transformed by scale transformation. a, b take different
color images is becoming an active research subject with the values and a class of function
a , b ( t ) is generated. The
development of computer technology in recent years.
Compared with gray scale images, color images sequence continuous transformation of x (t ) is
provides more information. However, it should be more 1 tb
robust to different illumination conditions, complex
background, face occlusion, expression change and so on[5].
WTx ( a, b) 
a  x(t )
 (
a
)dt
The study on face detection in color images is still a
challenging task. Face processing can be divided into face
  x(t )
a, b (t )dt 
x(t ),
a , b (t ) .
detection, face recognition and face tracking. Face In the actual application, discrete signal need to deal
recognition and face tracking have been used in many with. So wavelet transformation is defined discretely.
applications such as wrongdoer recognition, security systems, 
m
controlling systems [6]. We should consider that accuracy of a  a0 , b  nb0 a0 ,
m ,n (t )  a
(a0  mt  nb0 )
m m 2
these processing is directly related to accuracy of face

detection. Although many different algorithms exist to
perform face detection, each has its own weaknesses and W f (m, n)  

f (t )
m ,n (t )dt
strengths [7].
Until recently, many researchers concerned the research The discrete wavelet transformation of two dimensional
of face detection that plays important role in applications image f ( x, y ) is defined as
such as face recognition, video surveillance, human
computer interface and face image database management. Wf ( x, y )   2 f (u, v)
( x  u , y  v)dudv
But, up to now, due to the complexity of the purpose such as R .
expression, illumination, age, pose, the quality of the images Digital image are broken into LL, LH, HH, and HL four
and glass, hairstyle, beard, mustaches may or may not be sub-bands after a wavelet transformation, as shown in Figure
present and so on, many researchers can not resolve these 1. LL sub-band is similar to original image, and the noise is
problems completely even if they have studied it for long removed. LH is obtained after horizontal direction low-pass
time[8-12]. Learned from the domestic and international filtering and vertical high-pass filtering. HH is the result of
discourse and research papers concerning face detection and the horizontal and vertical high-pass filtering. HL is the
facial feature location in recent years, firstly we give a result of horizontal high-pass filtering and vertical direction
systematic summary of previous work. And we developed an low-pass filtering. After the second floor wavelet

978-1-4799-4261-9/14 $31.00 © 2014 IEEE 474

DOI 10.1109/ISDEA.2014.114
transformation, the face high frequency detail component is space is from RGB color space linear transformation, its
obtained, which is used for face detection. brightness component is not completely independent of
chrominance information and skin color clustering regions
have nonlinear relationship with Y, which leads to
searching result of clustering area poor robustness in the
space of CbCr color. Transformation process is as follows
(Y  K h )  (118  108)
MCb (Y )  108  if ( K h  Y )
Ymax  K h
( K i Y )  (118  108)
MCb (Y )  108  if (Y  K i )
K i  Ymin
Figure 1. A wavelet decomposition diagram (K i Y )  (154 114)
MCr (Y )  154  if (Y  Ki )
Ki  Ymin
(Y  K h )  (154 114)
MCr (Y )  154  if (Kh  Y )
Ymax  Kh
Cb (Y ) is central axis of skin area in Y  Cb , and Cr (Y )
is central axis of skin area in Y  Cr .
(Y Ymin )(WCi WLCi )
WCi (Y) WLCi  if (Y  Ki )
Ki Ymin
(Y Y )  (WCi WHCi )
WCi (Y )  WHCi  max if (Kh  Y )
Ymax  Kh
Figure 2. The improved algorithm block diagram
K i  125 K h  188
, .
III. COMBINED WITH ARTIFICIAL NEURAL NETWORK TO WCi (Y ) Y  Ci . i is b or
is the width of skin area in
IMPROVE THE WAVELET TRANSFORM r . W C b  46.97 , WCr  38.76 , WLCb  23 ,
Currently when the MLP neural network is used for face WLCr  20 , WHCr  10 , WHCb  14 .
detection, mostly in advance the number of neurons in
hidden layer is determined, which is the deficiency of neural if (Y  K i )or ( K h  Y )
,
network algorithm to detect human face. It can be imagined
that the MLP neural network hidden layer function is WCi
replaced by wavelet function, through the study iteration of
Ci ' (Y )  (Ci (Y )  MCi (Y ))   MCi (Y )
neural network, until the error of the neural network output
WCi (Y ) .
and the expected value meets the requirements of the if (Y [ K i , K h ])
numerical value to determine the number of neurons in ,
hidden layer. The improved algorithm block diagram is Ci (Y )  Ci (Y ) .
'
shown in Figure.2.
The process of improved algorithm is shown as follows. Ci ' (Y ) represents the value of Ci (Y ) after nonlinear
The choice of color space and color segmentation is as
follows. One of the most significant features of human face transformation in space YCb’Cr’. The value in the ellipse
( x  Ecx ) 2 ( y  Ec y )
2
is skin color. In the color image, skin color detection is
very important and effective. Because skin color is the
2
 2
 r2 belongs to skin
relatively concentrated, stable region in the image not a b
relying on the details of facial features and is not pixel point. The binary standard of image is obtained up to
sensitive to changes in posture, rotation, expression or now.
other changes. It's easy to distinguish human face from the
background colors. Therefore, it can narrow the search  x   cos  sin   Cb'  cx 
space and speed the detection in the complex  y     sin   .
cos   Cr'  c y 
background obviously. However, the effect that only using
the skin color detection is not good enough as well as
  
the false alarm rate and the accuracy. As the YCbCr color

475
cx  109.38 , c y  152.02 ,   2.53 , Ecx  1.6 , V. CONCLUSIONS
We have presented a fast wavelet transform based face
a  25.39 , b  14.03 , r  1 .
detection algorithm. The proposed classification algorithm
Four boundaries are used to limit clustering region, so it can has been tested on Caltech Webfaces database and found to
well adapt to areas of high brightness or too darkness and provide good performance. We intend to implement the face
the skin color model robustness is greatly increased. A two- detection algorithm on real mages and to map the algorithm
dimensional discrete wavelet transform is done on image onto a hardware platform for our future work.
and the wavelet basis function is Daubechies compactly
supported orthogonal wavelet. The results of wavelet
transform is sent to MLP neural network directly, and
hidden layer function of neural network is replaced by
wavelet basis function.
mj

j 
(a0 x  n j b0 )
.
The output of neural network is
H
L H

fk  wjk
j  wjk uija0 j x njb0  .
m

j1 j1  i1 

uij represents weight of input layer neuron to hidden layer
neuron. wij represents weight of hidden layer neuron to
output layer neuron. The number of input layer neuron is L .
The number of hidden layer neurons is set to 1, the output
of the neural network is calculated. Compare the error of
expectation value and the output value. If it does not meet
the requirements of error, the number of hidden layer Figure 3. Detection result of improved algorithm
neurons is increased by 1 for iterative learning until error
meets the requirements. If it meets the error condition, the Table 1. Performance comparison of different algorithms
number of hidden layer neurons is fixed. Algorithm Right Missing Average Detection
IV. THE EXPERIMENTAL RESULTS detection rate
time
Experimental conditions are as follows. CPU is 3.00 ANN 188 12 1.426s 85%
GHz, with 1GRAM, the Windows XP system and SVM 188 20 2.139s 90%
MATLAB7.0 version. Two kinds of the face databases are The 186 14 0.984s 93%
used to test the improved algorithm, a database of 250 proposed
pieces of positive face from Caltech Webfaces and 200 algorithm
pieces of positive face from the internet. The size, color and
expression of the images are different etc. Part of the test
results are shown in Figure. 3 to Figure.5. In order to
validate the efficiency of the algorithm, using ANN face
detection algorithm, the SVM face detection algorithm test
200 faces of the CMU database, the results are shown in
table 1.
Right detected images of ANN algorithm is 188 pieces,
right detected images of SVM algorithm is 186 pieces, and
Right detected images of proposed algorithm is 188 pieces.
Average detection time of ANN algorithm is 1.426s,
average detection time of SVM algorithm is 2.139s, and
Average detection time of proposed algorithm is 0.984s.
Detection rate of ANN algorithm is 85%, and detection rate
of proposed algorithm is 93%.
It can be seen that the algorithm is better than that of
artificial neural network algorithm and SVM algorithm in
the average detection time of each image, and the detection
accuracy is better than the other two kinds of algorithms. Figure 4. Detection result of improved algorithm

476
 social communication,” in Signal and Image Processing Applications
(ICSIPA), 2009 IEEE International Conference on. IEEE, 2010, pp.
279-284

Figure 5. Detection result of improved algorithm

REFERENCES
[1] Ming-Hsuan Yang, David J. Kriegman, and Narendra Ahuja,
“Detecting Faces in Images: A Survey” IEEE Transactions on pattern
analysis and machine intelligence, vol. 24, no. 1, January 2002
[2] Ganhua Li, Xaunping Cai, Xianshuai Li, Yunhui Liu, “An Efficient
Face Normalization Algorithm Based on Eyes Detection,”
Proceedings of the 2006 IEEE/RSJ,International Conference on
Intelligent Robots and Systems, October 9-15, 2006, Beijing, China.
[3] Yunyang Yan, Zhibo Guo and Jingyu Yang, “Multi-view Face
Detection Based on the Enhanced AdaBoost Using Walsh Features,”
Eighth ACIS International Conference on Software Engineering,
Artificial Intelligence, Networking, and Parallel/Distributed
Computing, Qingdao, China, 2007, vol. 1, pp.200-205.
[4] P. Kakumanu, S. Makrogiannis, N. Bourbakis. “A survey of skin-
color modeling and detection methods,” Pattern Recognition, 2007,
vol. 40(3), pp. 106-1122.
[5] H.T. Shandiz, B. Yousefi, S.M. Mirhassani ,"Face Detection in
Digital Image Based on Fuzzy Method" The International Conference
on Engineering, Applied Science and Technology ICEAST 2007,
November 21-23 2007 Thailand. 978-974-8308-56-2 KMITL, pp 64-
66
[6] A. Mohamed, Y. W. Jianmin Jiang, and S. Ipson, Face detection
based neural networks using robust skin color segmentation, 5th
International Conference on Multi-systems, Signals and Devices,
IEEE SSD 2008.
[7] D. Masip, M. Bressan, and J. Vitri` a, “Feature extraction methods
for real-time face detection and classification,” EURASIP Journal of
Applied Signal Process., vol. 2005, pp. 2061–2071, 2005
[8] C. Garcia and M. Delakis, “Convolutional face finder: A neural
architecture for fast and robust face detection,” IEEE Transactions on
Pattern Analysis Machine Intelligence, vol. 26, no. 11, pp. 1408-1423,
2004
[9] L. Ding and A. Martinez, “Precise detailed detection offaces and
facial features,” in Computer Vision and PatternRecognition, 2008.
CVPR 2008. IEEE Conference on. IEEE,2008, pp. 1-7.
[10] C. Xu, Y. Zheng, and Z. Wang, “Semantic feature extractionfor
accurate eye corner detection,” in Pattern Recognition,2008. ICPR
2008. 19th International Conference on. IEEE,2009, pp. 1-4.
[11] J. van de Kraats and D. van Norren, “Directional and nondirectional
spectral reflection from the human fovea,” Journal of Biomedical
Optics, vol. 13, p. 024010, 2008.
[12] J. Chen and O. Lemon, “Facial feature detection and tracking in a
new multimodal technology-enhanced learning environment for

477