Detection of Human Faces in Color Images and Performance Analysis of Different Skin Chrominance Spaces and Skin Chrominance Models

The overall objective of the project is to build a system that detects human faces in a given color image and to compare performance of different chrominance models and chrominance spaces. For face detection, approach is to detect skin color and segment given image into skin and non-skin regions. In a skin segmented image, the region of skin whose height to width ratio falls under well-known Golden ratio = (1 +√5)) / 2 ± some tolerance, the probability of that region to be considered as face is very high. In this implementation, evaluation of this assumption has been performed. To be able to find out what skin looks like, system has to generate statistical model of skin color. A training set of 54 skin samples (37,020 pixels) and 28 background samples (23,229 pixels) used to generate such models. This paper discusses two types of statistical models - Single Gaussian Model and Gaussian Mixture Model. Performance of both these model has been compared and the best fit model for given dataset has been used. Training dataset was collected using University of Stirling’s face database and several images from internet is used. Since dataset comes from different sources, it might result in some unknowns in dataset. One way to eliminate such unknowns is to separate color information from intensity or try to reduce effect of illumination by normalizing color information. To reduce effect of unknowns, this paper uses normalized-rgb (Red, Green, Blue) (illumination is normalized across three color, so effect is reduced), HSV (Hue, Saturation, Value) space (intensity and chromaticity part are independent) and CIE-xyz (Commission Internationale de l'Elcairage) (Machine independent) color spaces. Chrominance models in all three color spaces have been generated and compared to find which color space best suits the selected dataset.