Comparative performance of different skin chrominance models and chrominance spaces for the automatic detection of human faces in color images

This paper presents an analysis of the performance of two different skin chrominance models and of nine different chrominance spaces for the color segmentation and subsequent detection of human faces in two-dimensional static images. For each space, we use the single Gaussian model based on the Mahalanobis metric and a Gaussian mixture density model to segment faces from scene backgrounds. In the case of the mixture density model, the skin chrominance distribution is estimated by use of the expectation-maximisation (EM) algorithm. Feature extraction is performed on the segmented images by use of invariant Fourier-Mellin moments. A multilayer perceptron neural network (NN), with the invariant moments as the input vector, is then applied to distinguish faces from distractors. With the single Gaussian model, normalized color spaces are shown to produce the best segmentation results, and subsequently the highest rate of face detection. The results are comparable to those obtained with the more sophisticated mixture density model. However, the mixture density model improves the segmentation and face detection results significantly for most of the un-normalized color spaces. Ultimately, we show that, for each chrominance space, the detection efficiency depends on the capacity of each model to estimate the skin chrominance distribution and, most importantly, on the discriminability between skin and "non-skin" distributions.