Skin Detection in Luminance Images using Threshold Technique

Skin Detection in Luminance Images using Threshold Technique Skin Detection in Luminance Images using Threshold Technique Hani. K. Almohair, Abd Rahman Ramli, Elsadig A. M., Shaiful J. Hashim Department of Computer and Communication Systems Engineering Faculty of Engineering University Putra Malaysia, 43400, Upm Serdang, Selangor, Malaysia Abstract With the development of the chromatic skin color detection models, the threshold value is a useful and forceful element to discriminate between the skin and non-skin pixels. This paper reports an experimental investigation, which aims at using different threshold values to classify the human skin. The chromatic skin color model was used to model skin color in the r-g chromaticity space. This model based on testing database containing many of human images. It has been found that the threshold value is an important tool for increasing the human skin regions detected in color images containing luminance. Keywords: Skin chrominance models, Pixels classification, Skin detection. 1. Introduction It has been reached by researchers that skin color models can be used for detecting human skin in various computer vision applications. (Yang et al, 1998) suggested a real-time face tracker using an adaptive skin color model. They modeled the rgchromaticities of skin as a 2D Gaussian distribution. (Jones and Rehg, 1999) compared a Bayesian approach to a mixture of Gaussian in RGB space. This model includes nearly 1 billion labeled training pixels obtained from Internet. (Jedynak et al, 2002) considered a sequence of three models for skin detection built from a large collection of labeled images. Each model is a maximum entropy model with respect to constraints concerning distributions. Several models were proposed for detecting human skin in various lighting conditions. (Soriano et al, 2000) presented a face tracking method that is adaptive to varying illumination color. (Storring and Andersen, 1999) proposed model for detecting human skin under changing lighting conditions. (Cho et al, 2001) presented a new skin color filter that is useful and capable of finding skin regions in images and adaptively adjust its threshold. While (Caetano et al, 2003) compared between single model and several Gaussian mixture models used to improve skin detection. This paper presents the impact of adjusting the threshold value in the chromatic skin color model for improving skin detection in images that contain luminance. 2. Chromatic Model The various color spaces can be easily converted to a RGB representation, A triple [r,g,b] doesn t represent only color but also International Journal of The Computer, the Internet and Management Vol. 15#1 (January April, 2007) pp 25-32 25

Hani. K. Almohair, Abd Rahman Ramli, Elsadig A. M., Shaiful J. Hashim brightness, due to that property it was used to represent skin color in there chromatic color space. Chromatic colors, which represented by (Yang et al, 2002) known as pure color in the absence of brightness, are defined by: R r= R+ G+ B G, g= R+ G+ B A color histogram is the distribution of colors in the color space; figure (1- a) shows )1( a various people s images, and corresponding areas for histogram analysis are shown in Fig (1-b). The histograms of the skin color region selected are illustrated in figure (1-c). The color distribution of the skin color is clustered in small area of the chromatic color space. Figure 2 shows a skin color distribution of fifty people of different races with different skin colors in the chromatic color space. a b c Fig.1 various people images (A), Skin area For histogram analysis (B), histograms for Selected skin area (C). Figure 2: Skin color distribution of different humans is clustered in small area in the chromatic color-space 26

Skin Detection in Luminance Images using Threshold Technique The skin color distribution can be represented by Gaussian model (2) by substituting the mean (3) and covariance (4) parameters of the skin color distribution in chromatic color space into the gaussian distribution model. The mean values represent the center of the skin color distribution and the variance values represent its shape. By using an optimal threshold value, the skin and non-skin pixels are classified. The skin pixels represented with the white color. The optimal threshold value was determined by a conducted test for many images to prove the formula (5). 2 ( ) 2 ( g g r ) i r i 1 2 2 2 2σ g f( r, g) e r σ = e 2π 1 r = r i i= 1 1 g = i= 1 g i 2 1 2 σ g = ( g g ) i 2 1 2 σ r = ( r r ) i TP+ F =1 The procedure for creating the Chromatic Model is as follows: (3) (4) 1. Select manually the sample skin colored region from the images download from Internet. 2. Create a color histogram of the samples selected. 3. Estimate the mean and covariance of the color distribution in chromatic color space based on (3) and (4). (5) (2) 4. Substitute the estimated parameters into the Gaussian distribution model. 5. By using the threshold Technique, the skin and non-skin pixels classified where the skin pixels appears with the white color. 3. Threshold Technique The final step in the detection of human skin in JPEG images is creating a skinsegmented image by using a threshold value of probability. If the probability of a pixels in the input images (image under test) is more or equal to the estimated threshold value. It would be supposed that this pixel represents skin color, if it doesn t, then it would be supposed that it doesn t represents skin color. The skin color pixels are white and the other ones are black in skin-segmented image, the threshold value used in skinsegmented image is called an optimal threshold value, where the minimum increase in region size is observed while decreasing the threshold value (Chang H., and Robes, 2000). As a suitable compromise to discriminate between the skin and non-skin pixels, an optimal threshold value should be obtained. It is obtained through the proportion of the true positive TP over the ensemble of regions of skin, which is approximately equal to the proportion of true negative T over the ensemble of regions when that proportion doesn t contain skin (Terrillon et al, 2000). Equivalently, when the proportion of false negative F equals the proportion of false positive FP. See formula (5) Since this method is better for illumination variations, and also for different skin colors. We threshold the pixels as skin if International Journal of The Computer, the Internet and Management Vol. 15#1 (January April, 2007) pp 25-32 27

Hani. K. Almohair, Abd Rahman Ramli, Elsadig A. M., Shaiful J. Hashim the value is greater than the estimated optimal threshold value. Several authors have been used different threshold method and values to discriminate between the skin and non-skin pixels,( Saber et al,1996) used a standard threshold value, (Shin et al, 2002) used a fixed threshold. And (Filipe et al, 2003) thresholded the pixels as skin if the value is greatest than 0.7. In this model, the threshold value selected start from 1 and decremented the threshold value in steps of 0.1 until 0.0 is reached. At each threshold value the TP and F are determined and applied to the formula (8).if true positive plus false negative values are equal approximately to one. This value is considered as the optimal threshold value. The optimal threshold value was determined by a conducted test designed for testing the data images to prove the formula (8). The testing data includes many images of different humans of different races. Besides, the skin and non-skin pixels are classified according to each image. Figure (3) shows how the optimal threshold value in our model was determined. After the estimated optimal threshold value is applied on the ICSC model, the testing image is segmented into skin and non-skin pixels, figure (3-b). The number of skin and non-skin pixels is determined; the skin pixels detected is called the real and non-real skin pixels detected, from this segmented-image, the number of the real skin-pixels is computed, and illustrated in figure (3-c). By using one of the image processing programs, the real skin pixels of the input (image under test) is determined. See figure (3-a). Table (1) shows the number of the skin and non-skin pixels for all the above cases Original image a b c Figure 3: first column the original image, real skin pixels (a), real and non-real skin pixels detected (b), and the real skin pixels detected (c). 28

Skin Detection in Luminance Images using Threshold Technique Table 1: umber of real and non-real skin pixels detected Real skin pixels (A) 128557 Real and non-real skin pixels detected (B) 128509 Real skin pixels detected (C) 125945 TP 0.979 FP 0.019 F 0.02 From the images and table above, the TP and F were estimated. The TP is recognized from the proportion of pixels detected as skin which are really skin pixels to the total number of real skin pixels, while the F identified from the proportion of pixels detected as skin which are not real skin pixels to the total number of real skin pixels [23]. It is wroth mentioning that these TP and F values prove the formula (5). The obtained optimal threshold value is equal to 0.1. This optimal value is estimated after a conducted test for the training data with different threshold values. The following paragraph will show the difference in the results that occur when the threshold value is 0.7. a b Figure 4: left the original image, right the segmented image International Journal of The Computer, the Internet and Management Vol. 15#1 (January April, 2007) pp 25-32 29

Hani. K. Almohair, Abd Rahman Ramli, Elsadig A. M., Shaiful J. Hashim From the figure (4), it can be noted clearly that the segmented-images is used when threshold value is (0.7). Besides the area of skin-detected is too small, in a way it leads to a high false detection rate. 4. Result and Discussion The Chromatic Skin Color model was applied to a set of testing data of people images download from the Internet belonging to people of different races contain luminance, The model applied with two threshold values, the threshold values are 0.1 and 0.01. In the figure below, some images contain luminance. The first column is representing the original images (images under test). Figure (5-b) shows the skin regions detected when the threshold value is 0.1.In the figure (5-c) the skin regions detected are belongs to the threshold value 0.01. a b c Fig.5 Shows the original images (a), Skin region detected when threshold value 0.1 (b), the skin regions detected when the threshold value is 0.01(c). 30

Skin Detection in Luminance Images using Threshold Technique These results confirm that the threshold value is very important tool for improving the skin region detected in color images. The figure (6) shows the detection rates of the images when the threshold value is 0.1. The detection rates ranging from 85% to 95%. Figure (7) shows the detection rates of the regions detected would increase, when the threshold value 0.01 was selected. Figure 6. Detection rates, threshold value is 0.1 Figure 7. Detection rates, threshold value is 0.01 4. Conclusion This paper has presented the model to detect the human skin in color images, and also presented the impact of the threshold value on the chromatic skin color model in the images that contain luminance. The experiment proved that if the threshold value is decreased, the amount of segmented skin region would increase. Accordingly, if the threshold value is a high, all the non-skin pixels will be classified in a correct way. Yet almost all of them will be classified as non-skin pixels. Moreover, if the threshold value is low, almost all the skin pixels will be classified correctly, and some of non-skin pixels will also be classified as skin pixels. International Journal of The Computer, the Internet and Management Vol. 15#1 (January April, 2007) pp 25-32 31

Hani. K. Almohair, Abd Rahman Ramli, Elsadig A. M., Shaiful J. Hashim References B. Jedynak, H. Zheng, M. Daoudi, and D. Barret. Maximum entropy models for skin detection. Technical Report publication IRMA,Volume 57,number XIII, universite des Sciences et Technologies de Lille, france,2002 Caetano, T.S., Olabarriaga, S.D., and barone, D.A.C., Do mixture models in chromaticity space improve skin detection? Pattern Recognition, 36(12), pp.3019-302, 2003. Chang H., and Robes U. face Detection, May 2000 http://www-cs-students.stanford.edu/ ~robles/ee368/main.html Cho, k.m., jang, J.H., and Hong, K.S. Adaptive skin color filter, Pattern recognition, 34(5), pp:1067-1073., 2001 E., Saber, A. M. Tekalp, R. Esxhbach, and K. Knox, Automatic image annotation using adaptive color classification, Graph. Models and Image Proc. 58(2), pages 115-126, 1996. Filipe T., Tiago C., and Hamid S., Improved automatic skin detection in color images. In proc.viith digital Image Computing: Techniques and Applications, Sunc, Talbot,Sydney, 2003. Jones, M. J., and Rehg, J. M. Statistical color models with application to skin detection. In Proc. of the CVPR 99, vol. 1,pp: 274 280,1999. M. Storring, H. Andersen, E. G.. Skin color detection under changing lighting conditions. In Araujo and j. Dias (ed.) 7 th Symposium on Intelligent Robotics systems, pp: 187-195.,1999 Shin, M.C., Chang, K.I., and Tsap, L.V. Does colorspace transformation make any diference on skin detection? In Sixth IEEE Workshop on Applications of Computer Vision (WACV 2002), pages 275{279, orth Carolina Univ., Charlotte, C, USA, 2002. Soriano, M., Huovinen, S., Martinkauppi, and Laaksonen,. Skin detection in video under changing illumination conditions, In Proc. 15 th international Conference on pattern Recognition,vol. 1, pp. 839 842,2000 Terrillon J.C., Shirazi, M.., Fukamachi, H., and Akamastu Comparative performance of different skin chrominance models and chrominance spaces for the automatic detection of human faces in color images. In Proc. of the International Conference on Face and Gesture Recognition, 54 61,2000. Yang, J., Lu, W., and Waibel, A.. Skincolor modeling and adaptation. In Chin, R. and Pong, T.C., editors, 3rd Asian Conf. on Computer Vision, volume 1352 of LCS, pp: 687-694,1998. M.H. Yang, D. kriegman,. Ahuja, Detecting faces in images: a survey, IEEE trans. Pattern Anal, pp. 236-274,2002. 32