Procedia Computer Science

61 downloads 3361 Views 487KB Size Report
“Evaluation of RGB and HSV models in Human Faces Detection”. http://www.cg.tuwien.ac.at/hostings/cescg/CESCG-. 2004/web/Sedlacek-Marian/. Mars 2011.
Procedia Computer Science

Available online at www.sciencedirect.com

Procedia Computer Science 00 (2011) 000–000

www.elsevier.com/locate/procedia

WCIT-2011

Skin Detection Using Gaussian Mixture Models in YCbCr and Hsv Color Space Khammari Mohammed a *, Bencheriet Chemesse- Ennehar b, Tlili Yamina c b

a LRI, Université Badji Mokhtar, [email protected], Annaba, Algeria LAIG, Université 8 Mai 45 de Guelma, [email protected], Guelma, Algeria c LRI, Université Badji Mokhtar, [email protected], Annaba, Algeria

Abstract Many human-machine interaction systems such as face detection and hands tracking used skin detection as the main step in the system. Seen the importance of skin detection problem in the quasi-totality of faces detection and faces recognition systems several studies were proposed these last years but a reliable solution (giving a high rate of detection and a low rate of false alarms) and rapids (with a short processing time with any size of images or scene to analyze) is always looked for. In this paper we propose the use of a Gaussian mixture model, built from three skin colors between the most popular in the world (black, white, brown), that characteristics (of Gaussian mixture model) were fixed after a careful ethnic study made by our care. Knowing the importance and influence of the colors space representation on the quality of the obtained results we have and in comparative title created the Gaussian mixture model in two color spaces chosen between the most used: YCbCr and HSV color space. Tests, made on the international faces databases BAO and CALTECH, showed that the best rates of skin detection were obtained by the Gaussian mixture under the YCbCr color space.

Keywords: skin detection, segmentation, skin color, Mixture Gaussian model, color space.

1. Introduction Recently, many systems have been developed to face detection. Generally in the color images these systems are based on the search of skin color that will allow a minimization in the faces space search where research is restricted to segments of the detected skins. In the literature several methods have been proposed to discriminating between skin and non-skin pixels. These methods can be divided in two categories: parametric and nonparametric one. The parametric method used a model defining the distribution of skin color in a given color space. Nonparametric methods estimate the skin color histogram of learning database without the prior creation of a skin model [1]. Much research has been done on skin detection [2][3][4][5][6] since the results obtained are encouraging, around 90% but the remaining 10% is still causing problems in the steps below. So to ameliorate these results and in comparative title we propose in this paper the creation of the Gaussian mixture model in two color spaces chosen between the most used: YCbCr and HSV [7][8]. Tests made on the international faces databases BAO and CALTECH [13][14]. The paper is divided as follows: an overall introduction of skin color and the state of the art about skin detection methods in Section 1. The section 2 reserved to skin color and ethnic study. The different steps of adopted method

E-mail address: [email protected]

Khammari Mohammed / Procedia Computer Science 00 (2011) 000–000

will be developed in Section 3, follow by results and their interpretation in section 4, and terminate with a conclusion in section 5. 2. Skin color and ethnic studies The skin color human is widely used and proved that is an efficient feature in many applications, including, face detection and hands tracking. Although skin color can vary widely, recent research shows that the main difference come from intensity and not from the chrominance [9]. Unfortunately, the earliest attempts of human species classifications based on anatomical or cultural practices continue today to increase racist theories. According our careful ethnic study we concluded that: 30% of the world population has a black skin color, 30% of the world population has a white skin color and 40% of the world population has brown skin color [10]. These calculations will be used in Gaussian mixture model. 3. Detection steps of skin color The skin detection process must begin by converting the color image to the YCbCr or HSV color spaces after this we will apply the Gaussian mixture model, we'll get a grayscale image called generally likelihood image (where each gray scale represent the probabilities of pixel-skin membership), next this image will be binarized by dynamic thresholding given a skin binary image. The process of skin detection is summarized in figure 1 3.1. Color space Conversion Skin detection depends on the chosen color space. So to make the skin color segmentation, it is necessary to convert the image into the appropriate color space [7] [8]: 3.1.1. RGB to YCbCr YCbCr is often used in image compression; luminance (also called Luma) is separated from the color presented by the Y value witch calculated by a weighted sum of the R, G and B components. Y = 0,299 * R + 0,587 * G + 0,114 * B

(1)

The other two components of this color space represent the color information and are calculated by the following formulas: Cb = − 0,169 * R − 0,441 * G + 0,5 * B ; Cr = 0,5 * R − 0,418 * G − 0,081 * B

(1)

3.1.2. RGB to HSV HSV is a representation model called natural because of its resembling to the color physiological perception of human eye. It consists in breaking the color according to physiological criteria where: • The hue (H): corresponding to the color perception • The saturation (S): describing the purity of color, that is to say his character alive or dull • The value (V): indicating the amount of light color, that is to say it’s light or dark if B > G ,

Khammari Mohammed / Procedia Computer Science 00 (2011) 000–000

(2)

, 3.2. Segmentation by Gaussian mixture model

For segmentation by Gaussian mixture model , we must first create the Gaussian model for each color (black, brown and white). 3.2.1. Model Creation The creation of the model begins with the preparation of skin samples color from multiple color images. Original image

Converting image to the chosen color space

Likelihood Image

Application of Gaussian Mixture Dynamic thresholding

Binary image Fig. 1. Skin detection process

Fig 2. A set of samples from hands, forehead, neck, plays etc.

The following steps summarize the creation of the model: 1. Loading images of the skin database (Fig 2). 2. Transformation from RGB to YCbCr or HSV color space. 3. Averaging: m = E {X} where X = (C1 C2)T.

(3) T

4. Calculation of covariance matrix: C = E {(X - m) (X - m) }

(4)

5. Using the values of the mean and covariance, the skin color model can be adapted to a Gaussian model [11]: P(x) = exp [ - 0.5 (X-m)T C-1 (X-m) ] Where C1 = Cb or H, C2 = Cr or S, m is the average, C is the covariance matrix The following figure illustrates the Gaussian models of skin in space YCbCr and HSV:

(5)

Khammari Mohammed / Procedia Computer Science 00 (2011) 000–000

(a)

(b)

Fig 3: (a) Gaussian model of skin in space YCbCr ; (b) in the HSV space.

3.2.2. Applying the model With this Gaussian model, we can now obtain the probability of belonging to the skin for any pixel in an image. Given a pixel with a color value (C1, C2), the probability that this pixel is a skin, can then is calculated as follows [11]: Probability = P(C1,C2) = exp [ - 0.5 (X-m)T C-1 (X-m) ]

(6)

In this case we have three probabilities for each pixel, the first for the Gaussian model of black skin (p_ black) , the second for the Gaussian model of brown skin (p_ brown), and the third for the Gaussian white skin (p_ white) : P (x) = 0.3 * p_ black + 0.4 * p_ brown + 0.3 * p_ white

(7)

Therefore, this model can transform a color image into a grayscale image where the value of each pixel corresponds to its probability of belonging to the skin. 3.3. Thresholding With a dynamic thresholding (where the threshold is the mean of the grayscale image), images in gray level can then be converted into binary images showing areas of skin and non skin regions. Since the regions of skin are brighter than other parts of the image, these regions can be segmented by thresholding the rest. It then produces a binary image on which the "1" represent the skin color pixels and "0" for other pixels.

(a)

(b) (c) Fig.4: Application of Gaussian mixture (a) Original image (b) Skin likelihood image (c) Skin binary image

Khammari Mohammed / Procedia Computer Science 00 (2011) 000–000

4. Results and interpretations Skin color detection may be designated with several color space, why we chose to make a comparison between detection with YCbCr and HSV spaces using Gaussian mixture model. Performance is expressed by two values [12]:

(8) (9)

4.1. Results on BAO Image database: BAO [13] is a new database made by a Korean team under different conditions designed especially for detection tests divided into two groups of images; the first group consists of 140 images of one face and the 2nd group of 221 images with several faces. Group 1: Number of images = 149 Number of faces per image = 1 Table 1. Results of skin detection with basic BAO Detection method

Gaussian mixture

Color Space

YCbCr

Detection rate

95,97 %

91.33 %

HSV

False detection rate

39,59 %

35.37 %

Group 2: Number of images = 221 Number of faces per image = variable (between 1 and 32) Table 2. Results of skin detection with basic BAO Detection method Color Space

Gaussian mixture YCbCr

HSV

Detection rate

97,28 %

99.09 %

False detection rate

60,18 %

44.83 %

From the tests results performed on BAO database and present in Table 1 and 2: we obtained a detection rate of 96.63% with YCbCr and 95.21% with HSV space detection rates which we consider as significant and similar in both color spaces. We can conclude that we have increased the detection rate. 4.2. Results on CALTECH Image database CALTECH [14] designed by Markus Weber an institute Technology of California; it consists of 450 images of one face with backgrounds and complex lighting conditions. Number of images = 450 Number of faces per image = 1 Table 3. Result of skin detection with basic CALTECH Detection method Color Space

Gaussian mixture model YCbCr

HSV

Detection rate

85,77 %

42.95 %

False detection rate

60,22 %

68.15 %

Khammari Mohammed / Procedia Computer Science 00 (2011) 000–000

From the results presented in Table 3, performed on CALTECH image database we obtained a detection rate of 85.77% with YCbCr which exceeds detection rate of HSV color space 42.95 %. We can conclude that the YCbCr space is much better than the HSV color space in skin color detection using our mixture Gaussian model. Unfortunately it is clear that in our results we obtained a very high false detection rates; we can explain this by changing the lighting, the complexity of images background used in the tests and the richness of learning skin database using in the creation of Gaussian model. 5. Conclusion Skin color detection in color images is a preliminary step in most applications such as video conferencing, video indexing and especially video surveillance, the performance of these applications depends directly of the results obtained in this step. In this paper we presented a comparative study demonstrating the skin color detection using Gaussian mixture model applied on two different color spaces YCrCb and Hsv. From the obtained results the best color space is that which gives a high detection rate, the better in our case is YCbCr color space, which will be used in skin detection for several systems such as face detection, tracking hands or video surveillance.

6. References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14.

V.Vezhnevets, V.Sazonov, and A.Andreeva.”A Survey on Pixel-Based Skin Color Detection Techniques”. In Proc. 13th International Conference on the Computer Graphics and Vision, pages 85-92, Moscow, Russian, September 2003. J.Zhang, Q.Zhang, and J.Hu.” RGB Color Centroids Segmentation (CCS) for Face Detection”. ICGST-GVIP Journal, ISSN 1687-398X, Volume (9), Issue (II), April 2009. J.Yang, Z.Fu, T.Tan and W.Hu.” Adaptive Skin Detection using Multiple Cues”. http://kusu.comp.nus.edu.sg/proceedings /icip04/defevent/papers/cr1716.pdf. Mars 2011. H.Almohair, A.Ramli, A.Elsadig, J. Hashim.” Skin Detection in Luminance Images using Threshold Technique”. International Journal of The Computer, the Internet and Management .Vol. 15#1 pp 25 -32. April 2007. A. Conci, E Nunes, J. J. Pantrigo, A. Sanchez, “Comparing color and texture-based algorithms for human skin detection”. Computer Interaction. vol.5 pp. 168-173, 2008. H.Yu.“Face Recognition for Mobile Phone Using Eigenfaces”. www.eecs.umich.edu/~silvio/teaching/EECS598_2010/progress. / Hao.pdf . Mars 2 011. PHUNG, S. L., L.BOUZERDOUM, and D.CHAI."A novel skin color model in ycbcr color space and its application to human face detection". In IEEE International Conference on Image Processing .vol. 1, 289–292, (ICIP’2002). V.Nabiyev, and A.Günay.”TOWARDS A BIOMETRIC PURPOSE IMAGE FILTER ACCORDING TO SKIN DETECTION”. The Second International Conference “Problems of Cybernetics and Informatics”, Baku, Azerbaijan. Section #2. September 2008. H.HUYNH-HUU, J.MEUNIER, J.SEQUEIRA, M.DANIEL.” La détection et le suivi de régions d’intérêt pour la vidéosurveillance de la prise de médicaments”. http://documents.irevues.inist.fr/bitstream/handle/2042/28973/huynhhuu_413.pdf. Mars 2011 S.J.Gould. “Des races humaines suivant la couleur de la peau”. http://www.hominides.com/html/dossiers/race.html, Mars 2011. M.Sedláček. “Evaluation of RGB and HSV models in Human Faces Detection”. http://www.cg.tuwien.ac.at/hostings/cescg/CESCG2004/web/Sedlacek-Marian/. Mars 2011. A.A.R.Nusirwan, K.Wei and J.See. “RGB-H-CbCr Skin Colour Model for Human Face Detection”. www.johnsee.net/papers/m2usic2006rgb.pdf. Mars 2011. Frischholz R. Bao face database at the face detection homepage. Available: http://www.facedetection.com. Markus Weber an institute Technology of California, http://www.vision.caltech.edu/html-files/archive.html.