See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/257022696

Real Time Face Detection

Article · January 2012

CITATIONS READS
12 2,325

4 authors, including:

Mudassar Raza
COMSATS University Islamabad
132 PUBLICATIONS   3,070 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Removal of pectoral muscle for reducing the false positive rate in detection of breast cancer View project

Image Processing View project

All content following this page was uploaded by Muhammad Sharif on 30 May 2014.

The user has requested enhancement of the downloaded file.

 Sindh Univ. Res. Jour. (Sci. Ser.) Vol.44 (4) 597-600 (2012)

SINDH UNIVERSITY RESEARCH JOURNAL (SCIENCE SERIES)

Real Time Face Detection

M. SHARIF, K. AYUB, D. SATTAR, M. RAZA
Department of Computer Sciences, COMSATS Institute of Information Technology, Wah Cantt-Pakistan
E-mail: [email protected], [email protected], [email protected]
Received 07th December 2011 and Revised 03th May 2012

Abstract: In this paper, the difficulty of face detection in real time scenario is considered and an appropriate solution is being
proposed for this problem. The proposed idea is to detect human face in color images. For this purpose, HSV color space is used in
the experiments to detect faces. The proposed algorithm detects face in a real time video in two steps. In the first step, it applies
statistical model to get H (Hue) and S (Saturation) ratios for skin region. Secondly, it applies these defined ratios for scene width and
height to get approximation of face location in an image with respect to the detected skin region. Finally, an eye template matching
algorithm is applied on the previous roughly detected skin region to verify the face from this region. The proposed model has been
tested in real time environment with fairly acceptable performance.
Keywords: Face, Detection, RGB, (Red, Green, Blue), HSV, (Hue, Saturation, Value) Skin.

1. INTRODUCTION Research for face detection problem is in

Face detection (Feraud et. al. 2001, Hjelmas process and a wide variety of approaches had been
et. al.. 2001), (Hsu et. al. 2002, Jones et. al. 1999, proposed for solving the detection problems. Face
Rajagopalan et. al. 1998, Yang et. al. 2002) is the main detection is difficult in video sequences as compared to
module for any face recognition (Singh et. al. 2002, the still images. Video sequences are largely affected
Takacs et. al. 1998) system which outlines and mark by internal and external factor changes and thus
boundaries of face regions from messy images. The maximize the complexities for detection module.
image data to be processed may be obtained from video Previous research defines a large number of skin ratios
sequence or it may be a still image data. Most of the on the basis of different races (Yang et. al. 1996, Yoo.
present face recognition systems require a facial image et. al. 1999). One can detect skin by making use of these
to be processed and it ultimately requires human skin ratios. However, it is not always good to use this
computer interaction. But in real time scenarios, it is approach because in messy images faces get close to
difficult to get facial image to be taken as an input for one another and the skin region gets confused.
surveillance or point automation systems. Face This paper proposes an idea of extracting skin
detection module should have the capability to find, region from converting an image from RGB color (Zarit
outline and extract the exact faces from messy images et. al. 1999) space to HSV color space. The H and S
to provide faces as an input to the face recognition ratios are defined for extracting skin region. H and S
module. Face detection module should have the values are proposed and used in our experiments for
following properties: extracting skin. Some global parameters are defined for
(1) It should be capable of outlining exact face area in an image. By using these parameters the
faces as fast as possible. (2) It should be capable of algorithm roughly detects the face region and then it
handling exact borders of malfunction faces. (3) The uses eye template matching to verify the face region in
probability of false face identification should be nil. (4) an image. The proposed algorithm has been tested and
Constraints of the system in context of human its results have been mentioned in this paper.
interaction should be low. (5) It should have the Comparison of these results with the previous
capacity of noise and background removal. techniques (given in the results section) proves the new
proposed algorithm quite satisfactory and efficient.
For a human brain, it is easy to detect and
analyze face but it‟s not so easy for computer yet 2. MATERIAL AND METHODS
because of the reason that human face expressions get This section provides the detailed description
changed by different internal and external factors like of proposed algorithm in three steps. It describes the
lighting, pose variation, mustache, bear, glasses HSV color model and conversion from RGB to HSV
background colors etc and computer can‟t detect these color model, skin verification process and face
types of changes easily. detection.
++Corresponding author M. SHARIF, e-mail : [email protected] Cell.No.+92-3335188872
M. SHARIF,et al., 598

2.1. Face Detection This proposed model has been tested and its results
Our proposed face detection system is mainly have been mentioned in results section.
comprised of two major steps. The initial step is to
h  i , j 1 ( i , j  h min   i , j  h max)
m, n
categorize the skin pixels. For classifying skin pixels, (1)
the equations are defined which satisfy the presence of where,  ,h m in ,h m ax represent the hue value, minimum
skin in an image. The second step is to use the defined
and maximum threshold values for hue respectively.
global parameters for scene height and width for
(  i , j  s m in   i , j  s m ax) (2)
m, n
approximating the face region in an image. Then the s   i , j 1
eye template matching is used to verify the face in an Where, p  ,s m in ,s m ax represent the saturation value,
image.
minimum and maximum threshold values for
2.2. HSV Model and RGB to HSV saturation.
Low complicated but precise methods such as HSV   h  s (3)
modeling and color histogram are favored. To section
human skin areas from a consistent skin color model where, ө gives the resultant detected skin region.
that is flexible to various skin colors and to diverse Along with these HSV ratios the proposed algorithm
illumination conditions is required. The basic RGB also used the defined ratios of chromic blue and
color model is not appropriate for skin color chromic red which are as under:
classification because the triplet (R, G, B) represents Cb  (xR)  (xG)  (xB)  C (4)
not only color but also lighting which varies Cr  (xR)  (xG)  (xB)  C (5)
transversely the person‟s face due to the ambient Where, Cb and Cr represent chromic blue and chromic
illumination. HSV color model is a nonlinear red respectively.  ,  ,  , ,  are conversion constants
renovation of the RGB color space. The HSV model is
user-oriented and is based on artist‟s philosophy of hue, from RGB to CbCr color space.
shade, and tone, with self-determining values for Hue, 2.4. Face Detection
Saturation, and Value, corresponding respectively to In this step the task is to detect face area in a
wavelength, excitation, and intensity. detected skin region. The skin region detected may be
rough up to some extent. As it is must to clean the skin
Black has HSV = (0, 0, 0). Thus the facial HSV
region for verifying face in this region, therefore, image
model of Africans is clustered about the source of the
cleaning on above results is applied using a defined
coordinate arrangement. Across the African continent,
global parameter „noise_face_max‟. It removes all
unassuming variations occur which do not leave
connected components (objects) from a binary image
significantly from this value. Most color pictures are
that have less than a defined number of pixels. Now
recorded as (R, G, B) triplets. Given a color distinct by
label the cleaned image. The proposed algorithm uses
(R, G, B) where R, G, and B are normalized to 0.0 to
defined ratios for scene height and width for searching
1.0, an equivalent (H, S, V) color is determined by the
a bounding box in an image. The bounding box must
following set of formulas. (Sandro et. al.)
GB
satisfy the conditions for height, width, portal and

0  MAX  MIN * 60, ifR  MAX landscape ratios for a suitable bounding box for face
 detection. The proposed bounding box used in our
 BR
H  2  * 60, ifG  MAX
 MAX  MIN approach is of size 100x100. Then the proposed
 R G algorithm applies region propping on the basis of
 4  MAX  MIN * 60, ifB  MAX
 labeled image and the characteristic of bounding box.
MAX  MIN Next it calculates the length of a propped region. The
S  ; V  MAX
MAX length of a propped region tells the predicted number of
Considering MAX to be the maximum of the (R, G, B) faces in an image. The final step is to search all the
values and MIN the minimum values, the model is: faces in an image with the help of eye template
where H, S and V vary from 0.0-360 and 0.0-1.0 matching.
respectively. Traditional template matching approaches are
2.3. Skin Verification very slow. Since the proposed algorithm is being
Following mathematical model is proposed to designed for real time scenarios, therefore, timing is a
classify skin pixels in an image. The model uses the crucial factor in it. So, template matching is applied in
values as: frequency domain instead of using time domain
skin_color_H_min = 0.10; because in frequency domain both images must have
skin_color_H_max = 0.90; the same size for applying template matching. So a
skin_color_S_min = 0.30; resizing phenomenon is defined here for resizing an eye
skin_color_S_max = 0.99; template according to the parameters of bounding box
 Real Time Face Detection 599

of predicted face region. Then fast Fourier transform is

applied on both images separately. To analyze the
results of matching it must go back again in time
domain because correlation does not work in frequency
domain. As a result inverse of Fourier transform is
applied. Following are the formulae used for FFT (Fast
Fourier Transform) and IFFT (Inverse Fast Fourier
Fig. 5. Hue Element Detected Fig. 6 Detected Skin region in binary
Transform). (Xioguanglu et. al.), (Frigo et. al. 1998)
N Setp 5: Convert binary image back to RGB color space
X k   x j ( j 1)( k 1)
N
and use morphological filling for better results.
j 1
N
x j  (1 / N ) X k  N ( j 1)( k 1)
k 1

where, N  e
( 2  i )/ N
is an Nth
root of unity, N=number of samples. Fig. 7 RGB skin Detected Fig. 8Morphological Filled image

For template matching a mathematical model Step 6: In this step, eye template matching for face
used is represented as under: detection is used. If the corelation is higher then 0.65
Tprows Tpcols then it satisfies the face otherwise it is a non face
SAD( x, y)    Diff ( x  i, y  j, i, j)
i 0 j 0
(6) region. The template for eye and the detected
results are as follows:
face

Save the correlation results and analyze them. If

the value of correlation co efficient comes out to be
greater than 0.15 then it satisfies to be a face. The
above proposed method is quite efficient for real time
scenarios and has been tested. The detailed
experimental results are provided in the results section.
Fig. 9 Eye Template Fig. 10 Detected Face Region
3. PROCESSING OF ALGORITHM
The algorithm is processed as follows: 4. RESULTS AND DISCUSSION
Following results for skin detection are taken
Setp 1: Following image is taken as test (input) image on the basis of true point (T.P) and false point (F.P)
for applying the proposed algorithm on it. respectively. Comparison with different techniques is
shown in the following table. True points are calculated
on the basis of hit ratio and false points on the basis of
miss ratio.
Table 1 Skin detection comparison Existing results taken from
(Aznaveh et. al. 2008)

Fig. 1 Input Image Fig. 2 HSV Color Space Image Method TP FP

Step 2: Following figure shows the converted image in Bayes SPM in RGB 80% 8.5%
[Jones and Rehg 1999] 90% 14.2%
to HSV color space. Bayes SPM in RGB
Step 3: Following figures show the results of extraction [Brand and Mason 2000] 93.4% 19.8%
of chromic red, chromic blue and hue elements Maximum Entropy
respectively. Model in RGB [Jedynak et al.2002] 80% 8%

Gaussian Mixture 80% ~9.5%

Models in RGB [Jones and Rehg 1999] 90% ~15.5%

SOM in TS [Brown et al. 2001] 78% 32%

Elliptical boundary
90% 20.9%
Model in CIE-xy [Lee and Yoo 2002]
Fig. 3 Choromic Red Element Fig. 4 Choromic Blue Elemen
Detected Detected Single Gaussian in
Step 4 : Following figure shows the result of combining CbCr [Lee and Yoo 2002] 90% 33.3%
all three regions/elements produced above and it gives Gaussian Mixture in IQ
90% 30.0%
the required skin region in binary format. [Lee and Yoo 2002]
 M. SHARIF,et al., 600

Thresholding [BrandandMason 2000] 94.7% 30.2% REFERENCES:

A New Color Based Method for Skin 83.3% ~15.6%
Aznaveh, M.M., H. Mirzaei E. Roshan and M. Saraee
Detection [18] 90.7% ~13.3% (2008) “A new and improves skin detection method
Proposed Technique 95.5% ~14.6% using RGB vector space “ 5th International Multi-
Conference on Systems, Signals and Devices, 2008.
Results for face detection are taken on the
IEEE SSD 2008. 1-5.
basis of hit and false percentage of an algorithm.
Following table shows the comparison of results of Feraud, R., O. J. Bernier, J. Viallet, and M. Collobert.
proposed algorithm with the previously defined (2001) “A fast and accurate face detector based on
techniques. Comparison shows that the proposed neural networks”, IEEE Transactions on Pattern
algorithm produces better results than the previous ones Analysis and Machine Intelligence, (23): 42–53.
taken by existing techniques.
Frigo, M., and S. G. Johnson, (1998) "FFTW: An
Table 2 Face detection comparison Adaptive Software Architecture for the FFT,"
Proceedings of the International Conference on Signal
Methods Hit False
The Proposed Technique 93.2% 6.8% Processing, Vol. (3): 1381-1384.
M2HD (Sim et. Al., 1999) 52.95% 47.05% Hjelmas, E. (2001) Face detection: Survey, Computer
MEHD (Huttenlocher et. Al.1993) 66.3% 33.7%
Vision and Image Understanding, (83): 236–274.
HD (Saber et. Al.1996) 35.3% 64.7%
YES+HD (Srisuk et. Al.2000) 82.36% 17.64% Hsu, R. L., M. A. Mottaleb, and A. K. Jain. (2002)
HSV+AMHD (Srisuk et. Al.2000) 88.24% 11.76% “Face detection in color images”, IEEE Trans. Pattern
AMHD (Xioguanglu et. Al.) 64.71% 35.29%
Analysis and Machine Intell., (24): 696–706.
Following bar chart shows the comparison of
hit and miss ratios of proposed technique with the Huttenlocher, D. P. and W. J. Rucklidge, (1993)
existing techniques like M2HD, MEHD, HD, “Comparing image using the hausdorff distance”, IEEE
YES+HD, HSV+AMHD and AMHD. Y-axis shows the Trans. on PAMI, Vol. (15): 850-863.
ratios in percentage of hit and misses for face detection. Jones, M. J. and J. M. Rehg, (1999) “Statistical color
models with application to skin detection”, Computer
100 vision and Pattern Recognition, 274–280.
Hit Rajagopalan, A. N., K. Kumar, J. Karlekar,
0
Miss P. Poonacha, and S. Chaudhuri, (1998) “Finding faces
in photographs”, In Proceedings sixth IEEE Int‟l
Conference on Computer Vision, 640–645.
Fig. 11. Face Detection comparison charts Saber, E. (1996) “Face detection and facial feature
Following shows the results of proposed face detection extraction using color, shape and symmetry-based cost
algorithm on single person images. functions”, IEEE Int. Conf. 654-658.
Sandro, D. S., Johnson I. A., (2004), “Human face
segmentation algorithms for mobile phones,” Proc.
South African Telecommunications Networks and
Applications Conference (SATNAC) Vol. (2): Cape
Town, South Africa, 373 – 374.
Singh, R. K. (2002) “Background learning robust face
recognition” In Int‟l Conference on Pattern
Recognition. Vol. (3): 305-325.
Srisuk, S. and W. Kurutach, (2000) “Fast detection of
scalable and multiple human faces” Int. Conf. on Intell.
Fig. 12. Single Face Detection Tech., Bangkok, Thailand, 351-357.
5. CONCLUSION
Takacs, B. and H. Wechsler (1998) “Fast searching of
The HSV color space is used in this paper for
digital face libraries using binary image metrics”, IEEE
face detection On the basis of above experiments and
Int. Conf. on Pattern. Recog. 1235-1237.
results taken, it is found that the proposed technique
marks the faces well in images as compared to the Xiaoguang L. (2003) "Image Analysis for Face
previous techniques. In future, this technique can be Recognition – A brief survey," personal notes, May
used in face recognition and surveillance systems, 2003 Dept. of com science and Engg, Michigan state
where high accuracy is required. university, East learning, MI 48824.

View publication stats