The Effect of Scale, Facial Expressions, and Partial Occlusion on Skin Detection JAMAL AHMAD DARGHAM, ALI CHEKIMA, PAULRAJ MURUGESA PANDIYAN School of Engineering and Information Systems, University Malaysia Sabah, Locked Bag 2073, Teluk Likas, 88999 Kota Kinabalu, Sabah. Tel: 088-320000 Ext. 3014, Fax: 088-320349 ABSTRACT In this paper, the effect of scale, facial expressions, and partial occlusion by glasses on skin detection is investigated. Two pixel-based skin detectors were used. The first skin detector is based on the simple histogram thresholding and the second uses feed-forward neural networks. The neural networks were trained using only the red chrominance of the normalized rgb colour scheme. Eighty percent (80%) of skin data and a similar number of non-skin pixels from all image of scale factor 1 with neutral expression were used to train the neural networks. It was found that skin is almost invariant to scale, facial expressions and minor occlusion by glasses. However, the correct skin detection rate increases as the scale factor increases. Keywords: Skin Detection, Neural Network, rgb Colour Scheme, Histogram Technique. 1. Introduction Automatic skin detection is an important primitive in a wide variety of human-related images processing systems such as face recognition [1, 4, 13], lips reading [8], hand recognition and tracking [14], driver’s drowsiness [3], and pornography filtering [6]. A skin detector divides an image into two distinct classes one representing skin regions and the other non-skin regions. Since human-skin does not have a particular geometrical shape, the only attributes that can be used for skin detection are texture and colour. Almost all researchers use the skin colour defined in a given colour scheme as the attribute for the skin detector rather than texture. This might be due to the fact that colour significantly improve the performance of recognition over texture (greyscale) in low-resolution images when geometrical features become less apparent. In addition, it was suggested by [15] that colour may aid in low-level tasks such as segmentation. In addition, Brand et al. [4] attributed the wide use of skin detection as a low-level process in higher-level image processing systems to the facts that skin colour is a pixel level attribute and as such can be very fast. Skin colour distributions are clustered in the chromatic colour space, and under constant lighting condition, the chromaticity of skin-colour is assumed to be invariant to changes in size, orientation and partial occlusion [7]. In this paper, the effect of scale, facial expressions, and partial occlusion by glasses on skin detection is investigated. Two pixel-based skin detectors were used. The first skin detector is based on the simple histogram thresholding and the second uses feed- forward neural networks. 2. Previous Work Son et al. [12] developed a skin colour model in the Cr and Cb chrominance of the YCrCb colour scheme from a database of 200 images of various skin colours. They have found that the skin colour distributions for the three main skin colours: yellow, black, and white fall almost in the same space in the Cr - Cb chromatic space. Since the boundaries of these distributions are irregular, they used a multi- layer perceptron (MLP) as an adaptive classifier to