Comparative Performance of Different Chrominance Spaces for Color Segmentation and Detection of Human Faces in Complex Scene Images

Color is a powerful fundamental cue that can be used at an early stage to detect objects in complex scene images. This paper presents an analysis of the performance of nine different chrominance spaces in the specific problem of automatically detecting and locating human faces in twodimensional still scene images. For each space, we use a skin color model based on the Mahalanobis metric to segment faces from the scene background by thresholding. We perform feature extraction on the segmented images by use of fully translation-, scaleand in-plane rotation-invariant moments that are derived from the Fourier-Mellin transform, and apply a multilayer perceptron neural network with the invariant moments as the input vector to distinguish faces from distractors. We show that for each chrominance space, the detection efficiency is critically dependent on the goodness of fit of the skin chrominance distribution to the proposed model, and to a lesser extent on the discriminability between skin and "non-skin" distributions. Also, normalized color spaces are shown to yield the best segmentation results, and subsequently the highest rate of detection of faces with a large variety of poses and against relatively complex