International Journal of Scientific & Engineering Research, Volume 5, Issue 6, June-2014

ISSN 2229-5518

1138

Face Detection in Color Images Using Skin Color

Md. Mehedi Hasan, Jag Mohan Thakur, Prajoy Podder

Abstract— Because of the increasing demands of security for the present society, intelligent biometric identification such as face dete c- tion has got more application. Human face detection plays an important role in applications which are personal identification, face model- ling, fitness, face reconstruction, and face animation, facial expression analysis, video surveillance, control systems and security purpose. Face can be detected automatically with the help of computer but it is a challenging task for various face position, face sha pe, orientation, lighting condition, colour etc. In this article, a new assistive frame work has been introduced for fast and efficient detection of face. The goal of this paper is to detect the face by skin colour segmentation technique. Skin colour segmentation process helps to avoid the challenges of face colour, size and orientation. The brightness problem has been reduced by YCbCr colour space conversion. The experimental result shows that the proposed method has reliable performance than the existing methods. The accuracy of the proposed fame work is 99.27%.

Index Terms— Color space, Face detection, Knowledge base approach, Morphological technique, Region localization, Skin color segmentation, YCbCr color space .

1 INTRODUCTION

——————————  ——————————

ECURITY in our complex world is a vital issue. Network banking, financial abuse of bank, Credit cards sabotage makes the security topic more important. So, at present intelli- gent biometric system has been issued for the security pur- pose. Face detection is a biometric identification system, which

has been used in individual identification area.
During the last few years, many methods have been proposed
for face detection. Face detection methods are classified into three categories. They are Knowledge based methods, Tem- plate matching methods and Appearance based methods. Knowledge based methods are also called rule based methods, used to get image position of a single face. Knowledge based methods has been classified into two types which are Top down methods and Bottom up methods [1]. Top down meth- ods used different rules and conditions to get the facial fea- tures of human face. A human face consists of mouth, nose and two eyes which are symmetric to each other. Features re- lationship can be obtained by using relative position and dis- tances of image. Bottom up methods uses different facial fea- tures, multiple features, texture, skin colour etc. to detect face [2]. Feature invariant approaches also called structural fea- tures, use random labelled graph matching and colour infor- mation to locate faces [3]. Template matching methods uses different rules and constraints to template face. Template matching method has been classified into two sub categories, which are predefined templates and deformable templates.

————————————————

 Mehedi Hasan is currently pursuing B.Sc. degree program in electronics & comcmunication engineering in Khulna University of engineering & technology (KUET), Bangladesh, PH-008801790853171. E-mail: mehedihasanece@gmail.com

 Jag Mohan Thakur is currently pursuing B.Sc. degree program in electronics & communication engineering in Khulna University of engineering & technology (KUET),Bangladesh,PH-008801675850718.E-mail: jagmohanthakur43@yahoo.com

 Prajoy Podder is currently pursuing B.Sc. degree program in electronics & com- munication engineering in Khulna University of engineering & technology (KUET),Bangladesh,PH-008801714078499.E-mail: prajoypodder@gmail.com

Predefined template works in two steps. Firstly face is located and separated from image using templates. Secondly the ex- istence of face is determined by focusing the areas of face [4]. Deformable Template also called parameterized template, which are used to determine different facial features. Edges of the input images, peaks, valleys are parameters of the tem- plate and used to describe energy function. An energy func- tion of the different parameters is minimized to get elastic model [5]. Appearance based methods is a set of training im- ages, which is used to capture the variation of facial appear- ance. Machine learning and statically analysis [6] has been used to determine the relevant features of face and non-face images. This method has been divided into two types, which are Neural Networks [7] and Support Vector Machines [8]. Neural network is used to detect faces from anywhere of an image, at any image locations. In order to detect faces which are larger than 20x20 pixels, input image is sub scaled repeat- edly and at each scale network is applied. Multi-layer neural network has been used to get face and non-face patterns from face and non-face images. A neural network is a first compo- nent of this method to get a 20x 20 pixel of an image region. And the output score ranges from -1 to 1. According to given test pattern, the trained neural network uses output -1 to rep- resents non-face and 1 is used for face pattern. In support vec- tor machine (SVM) approach [8], polynomial function, radial base function and neural networks classifier is trained to get desired result. Training classifier methods has been used to minimize the training error. Structural risk minimization uses induction principle to minimize the upper bound of an ex- pected generalization error.
Skin colour is a good feature for detection of the human face. There are two main approaches in face detection based on skin colour. Pixel-Based Model is the first approach, which is used to detect all parts of human skin colour by processing the pix- els of skin. Each pixel is processed independently to detect whether it is skin colour or not. Skin colour detection has clas- sification problem and primary step to select suitable colour

IJSER © 2014

http://www.ijser.org

International Journal of Scientific & Engineering Research Volume 5, Issue 6, June-2014

ISSN 2229-5518

1139

space. So that, colour space can easily discriminate skin and non-skin pixels. Second approach used to determine the status of the region of the image. Necessary effort has been made to separate face from the given image. And after using the knowledge and information of previous image, it is decided that the given image belongs to face or not. Many different colour space RGB (Red Green Blue), HSV [9] (Hue Saturation Value), YES, YCbCr, CIE Lab, TSV, HIS (Hue Saturation Inten- sity) [10], TSL are used for face detection. In this paper, a pix- els based skin colour segmentation process has been intro- duced. YCbCr [11] is the main colour space. The YCbCr is a colour space that has red, blue components. In case of YCbCr colour space [11], [12] Cb is smaller than Cr components. RGB colour image has been used as input for the proposed method. After the skin colour segmentation, there have been some small noise. This noise has been reduced by image erosion, image dilation through morphological process. Image filling operator fills the unwanted holes in face region and exterior boundary points of face region has been traced to determine the top, bottom, left and right sides.
The rest of the paper is oriented as follows. Color models for skin colour are described in section– II. The overview of the proposed algorithm is described in section– III. Under this section face detection process is explained clearly. The exper- imental results and comparison with different colour space are explained in section-IV. Finally, section-V concludes the pa- per.

2 SKIN COLOR CLASSIFICATION

2.1 RGB colour space

An RGB colour space is mainly used due to its simplicity and easiness of implementation. Red, Green and Blue chromaticity has been used to produce any primary colour. The mixing of chrominance and luminance data is not suitable choice for colour analysis because of non-uniformity and high correla- tion between the channels. An RGB colour model has been used to represent digital images. RGB output has been used by most of the image display devices. It is mainly used in all computer systems, videos, cameras, etc. RGB and Adobe RGB are the mostly used RGB colour spaces [13]. Adobe Wide Gamut RGB is another colour space recently developed by Adobe.

Fig.1. RGB colour cube

2.2 HSV colour space

HSV [14] stands for Hue, Saturation and Value. The level of brightness has been shown by the Value. HSV colour space is much simpler and can be linearly transformed from RGB. HUE is defined by the dominant colour of the area such as Red, Purple and Yellow. Saturation is provided by the colour- fulness of the area, which is in proportion to the brightness of the image. Chrominance and luminance separation has been obtained in this space. Invariant to highlight, surface orienta- tion, etc. are the most important properties of colour segmen- tation. This makes this colour space most popular. Discrimi- nating information has been providing by H and S, which is related to skin. H, S, and V values for face and non-face pixels have been plotted with the help of reference image to detect any valuable trends.

Fig.2. HSV colour cube

The conversion of RGB to HSV is provided by the following equations,

2.3 TSV and TSL colour space

TSV and TSL colour space has been used for skin detection. This is complex perceptual colour Spaces used in place of HSV. TSL colour space is best choice for Gaussian Skin colour modelling [15].
T= (5)
where, =r-1/3, =g-1/3

IJSER © 2014

http://www.ijser.org

International Journal of Scientific & Engineering Research Volume 5, Issue 6, June-2014

ISSN 2229-5518

S =(6) V= & L = 0.299R + 0.587 G + 0.114 B (7)

1140

2.4 HSI colour space

HSI [16] stands for Hue, Saturation and Intensity, which is used to describe colour. HSI colour space can be easily pro- duced without knowing the percentage of Blue or Green. By adjusting Hue, we can easily produce any type of colour. Ad- justing the saturation, deep red to pink can be easily changed. Altering the intensity, lighter or darker colour can be easily made. HSI colour model has many applications. HSI colour space is used for Machine vision to identify the colour of dif- ferent objects. In image processing, intensity image are operat- ed by convolutions, intensity transformations and histogram.

Fig.4. YCbCr colour cube

2.6 Normalized RGB colour space

Normalized RGB has been used to reduce the dependencies of RGB values by changing its luminance. It is clear that from this equation, (r+g+b=1). Third components can be easily ob- tained by knowing any two components of r, g and b [20]. Color space detection has been used to separate chrominance from luminance. N-RGB is the mostly used colour space among researchers because of its simplicity in transformation and all its advantages. Normalized RGB is obtained by the following equations,

Fig.3.HSI colour cube

2.5 YCbCr colour space

The YCbCr colour space also known as family of colour spaces because the Chroma components Cr and Cb can be easily cal- culated. Luminance is denoted by Y. Blue difference and red difference Chroma components are denoted by Cb and Cr. The three components of YCbCr can be easily calculated by linearly combinations of R, G and B components of image. In order to get the skin region, it must satisfy the following equa- tions [17],
r = R/R+G+B, (8) g =G/R+G+B, (9) b =B/R+G+B. (10)

2.7 CIE LUV and CIE Lab colours space

CIE LUV and CIE Lab colour space [22] are the nonlinear transformation of CIE XYZ. CIE LUV provides much better perceptual uniformity in the comparison to its predecessors. Both colour spaces are device independent. Chrominance and Luminance can be easily separated in this colour space. These colour spaces are not suitable for skin detection due to its complexity and computational expensiveness.
140 <= Cr <= 165 AND
140 <= Cb <= 195
The pixels related to skin regions of human faces have similar characteristics as Cb and Cr components. Skin colour is mainly determined by the darkness or fairness of the skin. The differ- ence in brightness of colour mainly determines the Y compo- nent rather than Cb and Cr components [18]. In order to get skin regions, some restrictions are made on these two compo- nents and Hue. Hence, Skin colour can be easily detected by Chrominance and luminance colour due to all this property and its simplicity [19].

Fig.5. CIELab colour cube

3 OVERVIEW OF PROPOSED ALGORITHM

Face detection is a challenging task for different face structure, face position, orientation, facial expressions and skin colour. Here an easy frame work has been introduced for face detec- tion.
Fig.6 describes the total evaluation process of the activity of the proposed method. Firstly, face image is used as input then the face area is localized from the input face image. After area localization, face is detected from that localize region. The to-

IJSER © 2014

http://www.ijser.org

International Journal of Scientific & Engineering Research Volume 5, Issue 6, June-2014

ISSN 2229-5518

1141

tal processes of the proposed algorithm are explained as fol-

lows.

Fig.6. Block diagram of proposed method

4 FACE DETECTION BY PROPOSED METHOD

Skin colour segmentation is used in face detection procedure. Which is knowledge based approach that helps to avoid chal- lenges of face size, colour and orientation. The impact of brightness problem is reduced by the conversion of RGB to YCbCr. Besides Processing of skin colour is much faster than processing other facial features for detection of face. Lumi- nance component (Y) of YCbCr is independent of the colour, so it is used to solve the illumination variation problem. Fol- lowing conditions are applied to detect face region,

Cb>=80 AND Cb<=120
Cr<=173 AND Cr>=140 AND (11) Y<=255 AND Y>=60
After skin colour segmentation there remains some small noise. Those are reduced by using image erosion through morphological structure. Image erosion shrinks the object. The binary erosion of A by B, denoted A Θ B, is defined as the set operation.
In other words, it is the set of pixel locations z, where the
structuring element translated to location z overlaps only with foreground pixels in A.
Then binary image conversion is done to help image filling
operation that fills unwanted holes in face region. Now it is helpful for exterior boundary detection. Large area of face re- gion is achieved through maximum connected area. Exterior boundary points of the face region are traced to determine left, right, top and bottom side points. Exterior boundary of an object is obtained by first eroding the object by a structuring element and then taking the set difference of the object and its erosion. Boundary of a set A is denoted by β (A).
Β (A) = A-(A-B) (13)

Where, (AΘB) denotes the erosion operation. Fig. 7(a) illus- trates the mechanism of boundary extraction top, bottom,
a b

Fig.7. Exterior boundary extraction

Left and right side points are obtained from following equa- tions. Bounding box is obtained from those side points shown in fig 7 (b).
X1= maximum (x_coordinates of boundary). (14) X2= maximum (x_coordinates of boundary). (15) Y1= maximum (y_coordinates of boundary). (16) Y2= maximum (y_coordinates of boundary). (17)

AΘB = {z|Bz ⊆ A} (12)

Fig.8. Face detection process by proposed method

IJSER © 2014

http://www.ijser.org

International Journal of Scientific & Engineering Research Volume 5, Issue 6, June-2014

ISSN 2229-5518

1142

Fig. 8 shows the total process of face detection by proposed method. Fig. 8(a) is the input image, fig. 8(b) represents the skin colour segmentation image, fig. 8(c) represents after mor- phological process image, fig. 8 (d) represents image after used image filling operator, fig. 8(e) represents the boundary detection image and fig. 8(f) represents the detected face im- age.

5 EXPERIMENTAL RESULT

For performance measurement the proposed methods has been experimented through matrix laboratory software (MATLAB). The images which are used as input is obtained by the Samsung company digital camera. The proposed algo- rithm has been experimented on 275 face images. The face in the images of the experimented people was different face posi- tion, face structure, pose, facial expression, colour condition and orientation. All the face images have been used as input to the previous existing methods and proposed method.
Fig.9 shows face detection result at different pose, brightness and facial expression. Fig. 9 (b, h, k, o, r, s, u) represents differ- ent pose of detected face, fig. 9 (c, d, f, h, j, k, n) represents var- ious brightness of detected face and fig. 9 (e, h, k, m, o, s, t) represent various facial expression.

TABLE 1

FACE DETECTION BY THE PREVIOUS SYSTEM

TABLE 2

FACE DETECTION BY THE PROPOSED SYSTEM

Table 1 show the outcome of the previous system, where
RGB, HIS, CIELab, LCCS colour space are used. In case of RGB, the experimental results are not very much friendly with face detection based on skin colour. The face detection rate is
56.46%. HIS colour space shows that the face detection by this colour segmentation is 82.18%. CIELab colour space face de- tection rate is 85.8% and Log-Chromaticity Color Space
(LCCS) shows that the face detection rate is 89.8%. Fig.9. Results of face detection using proposed method at various pose brightness and facial expression.

IJSER © 2014

http://www.ijser.org

International Journal of Scientific & Engineering Research Volume 5, Issue 6, June-2014

ISSN 2229-5518

1143

a

Fig.11. Comparative chart between previous (RGB, HIS, CIELab, LCCS)

and proposed (YCbCr) method.

6 CONCLUSION

b

Fig.10. Multiple faces detection at same image

Fig. 10 (a) and fig. 10 (b) shows, multiple face detection results at same image
Fig. 11 shows the comparative chart between the previous and proposed colour space methods. Where RGB detection rate is
56.46%, HIS detection rate is 82.18%, CIELab detection rate is
85.8%, LCCS detection rate is 89.8% and YCbCr detection rate is 99.27%.
Table 2 shows the accuracy of the proposed system, where
YCbCr colour space has been used. The face detection rate is
99.27%, which is better than other existing colour space.

TABLE 3

COMPARATIVE RESULTS BETWEEN THE PROPOSED METHOD WITH

PREVIOUS METHOD

Table 3 shows the comparisons result between the existing methods and proposed method.
At present, facial information is used for various applications. So, in this paper, Author proposed to detect face from image with varying lighting conditions and complex background with the help of color spaces and color models of people. The practical model for human skin color has been discussed. We can easily obtain the threshold value of each color component with the help of color space model. The face detection algo- rithm is based on YCbCr color space method with lighting compensation technique and nonlinear color transformation. At first skin region is detected from image and then face area are found from grouping skin region. This proposed system works well on wide range of facial variation in color, position, scale and orientation with photo collection including both in- doors and outdoors. The experimental result shows that the proposed method is much better than the other existing meth- ods.

REFERENCES

[1] M.Yang, D.J.Kriegman, N.Ahuja, “Detecting Face in Image”, IEEE Transaction on Pattern Analysis and Machine Intelligence, Vol.24, No.1, January 2002.

[2] T.Barbu,”An Automatic Face Detection System for RGB Images”, Int.

J. of Computers, Communications & Control, Vol. VI, No. 1, ISSN

1841-9836, E-ISSN 1841-9844, pp. 21-32, March. 2011.

[3] M. Krishnan Nallaperumal, Ravi Subban, R.K.Selvakumar, A. Lenin Fred, C. Nelson KennadyBabu, S.S. Vinsley, C. Seldev,”Human Face Detection in Color Images using Mixed Gaussian Color Models”, In- ternational Journal of imaging science and engineering (IJISE),Vol.2,No.1, ISSN:1934-9955 January. 2008.

[4] Ghassan Marwan Abdulfattah, Mohhammad Nazir Ahmad,” face localization based template matching approach using new similarity measurements”, Journal of Theoretical and Applied Information Technology Vol. 57, No.3, ISSN: 1992-8645 , E-ISSN: 1817-3195, No- vember. 2013.

[5] T. Barbu,”An Automatic Face Detection System for RGB Images”, Int.

J. of Computers, Communications & Control, Vol. VI, No. 1, ISSN

1841-9836, E-ISSN 1841-9844, pp. 21-32, March. 2011.

IJSER © 2014

http://www.ijser.org

International Journal of Scientific & Engineering Research Volume 5, Issue 6, June-2014

ISSN 2229-5518

1144

[6] Y. F. a. R. E. Schapire, "A Decision-Theoretic Generaliztion of On- Line Learning and an Application to Boosting," Journal of Computer and System Sciences, no.55, pp. 119-139, 1997.

[7] Henry A. Rowley, Shumeet Baluja,T akeo Kanade,”Neural Network-

Based Face Detection”, Computer V ision and P attern Recognition,

1996,FLEXChip Signal Processor (MC68175/D), Motorola, 1996.

[8] Elena Casiraghi, Raffaella Lanzarotti, Giuseppe Lipori,” A face detec- tion system based on color and support vector machines”, volume

2859, Springer, 2003.

[9] Y. Z. Jie Yang, Xufeng Ling and Z. Zheng, “A face detection and recognition system in color image series,” Mathematics and Comput- ers in Simulation, pp. 531–539, 2008.

[10] Nicolas Gourier Daniela Hall James L. Crowley,” Facial Features

Detection Robust to Pose, Illumination and Identity”, IEEE, pp. 617-

622. 2004.

[11] Varsha Powar, Aditi Jahagirdar, Sumedha Sirsikar, “Skin Detection in YCbCr Color Space”, International Journal of Computer Applica- tions, pp. 0975 – 8887, 2011.

[12] H. X. Xinyu Wang, Xi Chen and Heng Li,"Fast and Robust Face De- tection with Skin Color Mixture Models and Asymmetric AdaBoost," Proc. of SPIE Vol. 7496 749618-1, 2009.

[13] Qieshi Zhang, Jun Zhang,”RGB color analysis for face detection”, Advances in computer science and IT, pp. 119-126, December. 2009.

[14] Douglas Chai, King N. Ngan,”Face Segmentation Using Skin-Color Map in Videophone Applications”, IEEE Transactions on circuits and systems for video technology, VOL. 9, NO. 4, JUNE. 1999.

[15] Elena Sikudova, “Comparison of color spaces for face detection in digitized paintings “, 2006.

[16] Chun-Liang Chien, Din-Chang Tseng,” Color image enhancement

with exact HIS color model” ,international journal of innovative

computing, information and control, Vol 7, Number 12, ISSN: 1349-

4198, pp. 6691-6710, December. 2011.

[17] R.Pooja, G.Suresh “Facial Expression Recognition under Different Color Transformations using Indian Face Database”, International journal of professional engineering studies, Volume -II, JAN.2014.

[18] S. Chitra , G. Balakrishnan “Comparative Study for Two Color Spac- es HSCbCr and YCbCr in Skin Color Detection”, Applied Mathemat- ical Sciences, Vol. 6, no. 85, pp.4229 – 4238, 2012.

[19] Rehanullah Khan , Zeeshan Khan , Muhammad Aamir , Syed Qasim Sattar ,”Static Filtered Skin Detection”, IJCSI International Journal of Computer Science Issues, Vol. 9, No 3, ISSN (Online): 1694-0814, March. 2012.

[20] Vandana S. Bhata, Jagadeesh D. Pujaria, “Face detection system us- ing HSV color model and morphing operations”, International Jour- nal of Current Engineering and Technology, Special Issue1, ISSN

2277 – 4106, sept.2013.

[21] Jian Yang, Chengjun Liu, Lei Zhang, “Color space normalization: Enhancing the discriminating power of color spaces for face recogni- tion”, Elsevier, Pattern Recognition 43, pp. 1454–1466, 2010.

[22] Amanpreet Kaur, B.V Kranthi, “Comparison between YCbCr Color Space and CIELab Color Space for Skin Color Segmentation”, Inter- national Journal of Applied Information Systems (IJAIS), Vol 3, Number 4, ISSN: 2249-0868, July. 2011.

[23] Mrs. Sunita Roy, Mr. Susanta Podder,” Face detection and its appli- cations”, International Journal of Research in Engineering & Ad- vanced Technology, Volume 1, Issue 2, ISSN: 2320 – 8791, April-May,

2013.

[24] S.Ravi, S.Wilson,” Face Detection with Facial Features and Gender Classification Based On Support Vector Machine”, International Journal of Imaging Science and Engineering, ISSN: 1934-9955, 2010.

IJSER © 2014

http://www.ijser.org