Adaptive mouth segmentation using chromatic features

The automatic segmentation of the mouth from its facial background is a very difficult computer vision problem due to the low grayscale distinction between classes. Recently chromatic based segmentation has enjoyed some popularity for the purposes of mouth tracking due to its ability to distinguish between the two classes. Such systems have to be highly adaptive due to problems with colour constancy. In this paper a technique for adaptive segmentation is investigated using an unsupervised clustering technique incorporating the expectation maximisation (EM) algorithm across a variety of chromatic features. Results are presented from the M2VTS database across a number of subjects.