Skin Color Information Analysis And Detection

Human skin information analysis and detection is an hot research topic involved with optical information processing, digital image processing, computer vision and pattern analysis and recognition. In recent years, many studies have been done on human skin analysis and detection, and many useful conclusions have been obtained. This laid a sound foundation for researchers to understand the properties of human skin colors. In this thesis, some key problems on skin color detection are analyzed and studied soundly. Considering the reality that the appearance of human skin colors is closely related with the biometric properties of human skin and illumination conditions, an illumination corrected method is introduced firstly that has a high ability in weaken illumination affect on images. Then, the biometric and natural properties of the human skin are analyzed scientifically, and a non-linear color space transformation is proposed to described the cluster properties of the human skin colors. Finally, an skin color classification method and an adaptive skin color detection method are proposed respectively. To the above aspects, we investigated them from an scientific and reasonable angle, and present their responding solutions. The main contributions of this thesis include(1) An improved Gamma correction method is proposed. This method use nonlinear methods instead of linear method to modify the Gamma values. This can enlarge the interval of Gamma values effectively and make the Gamma correction method more suitable to the changes of highlights, transition regions and shadows in images. So the proposed Gamma correction method is improved to have a more adaptability in illumination changes, and can weaken illumination affects on images effectively.(2) An non-linear color transformation is proposed for describing the cluster property of human skin colors. Based on the cluster property of skin colors, an non-linear difference relations between color channels is constructed. These relations is then expressed in polar coordinates through coordinate transformation. The distributions of skin colors and non-skin colors are described using histogram models based on a large scale statistical training of skin and non-skin pixels. The experimental results show that the proposed non-linear transformation has a good performance in exhibiting the cluster property of human skin colors.(3) an skin color classifier with fusion of spatial and probabilistic information of pixels. Based on statistical results on a large scale, we obtain the distributions of skin and non-skin colors and use a piecewise closed curve to delimit the skin color cluster region. Considering the overlap between skin and non-skin distributions, probabilistic analysis is done to the overlap region. Thus, a nonlinear skin classifier is built using the fusion of the spatial and probabilistic distributions of skin colors.