Lip Modelling and Segmentation

Which Is The Best Color Space?

The main issue is to determine the most appropriate color space to make the difference between skin pixels and lip pixels. The problem of skin or non skin pixels characterization has been extensively studied in the context of face detection. But the point here is to be able to distinguish skin pixels from lip pixels (all these pixels belonging to a face). Concerning region based lip segmentation algorithms, the optimal color space is the one in which skin pixels and lip pixels are represented by two compact and distinct groups of pixels (low intra-class variances and high inter class variance). Regarding contour based lip segmentation algorithms, the optimal color space is a space in which the gradient information between lip and skin pixels is accentuated.

We propose a summary of the different color spaces that have been used for lip and skin pixels separation. We first focus on current color spaces such as RGB, HSV, and YCbCr which appear not completely efficient. Then we describe some specific color information that has been specially introduced in order to increase the difference between skin and lip pixels. We try to demonstrate the discriminative power of each considered color space by comparing the repartition of some skin pixel samples and some lip pixel samples. A specific database of skin and lip pixel samples has been built. All these samples have been manually extracted from a database of 150 frames representing 20 subjects acquired with the same camera and with the same conditions of illumination. For the purpose of comparison the dynamic of all the chromatic measures has been normalized to the range [0, 1] and the intra-class and inter-class variances have been computed.

In a second step, we present a set of specific luminance and/or chrominance cues that have been developed in order to accentuate lip boundaries gradient.