Chinese skin detection in different color spaces

Skin detection is an important preliminary process in various computer vision applications. It is typically performed as follows: color constancy for light compensation, transformation from RGB to a non-RGB color space, dropping the luminance component and using the chrominance components only, finally classifying image pixels into skin or non-skin by an appropriate skin color modeling technique. However, there is not a common criterion for the choice of the best color space, which is the focus of our study, to approach this binary classification problem. We have adopted the Gray-Edge assumption for image color correction, evaluated 15 most used color models for color space transformation, and employed an explicit threshold algorithm with smoothed bivariate histogram for skin color classification. To perform detailed comparisons among the selected color spaces, we have manually generated 30 ground truth images, in which non-skin regions have been removed, thus we can compare at pixel level with an accurate and objective criterion. Results show that most appropriate color spaces for Chinese skin color detection are CIE-Lab and CIE-Luv, respectively, with and without the luminance component.