Spatial clustering of pixels in the mouth area of face images

Abstract

We propose a method of image segmentation using a Gaussian mixture model of the colour image histogram. The model construction is based on the model validation philosophy of architecture selection (Kittler et al., 2001). In contrast with the k-means clustering approach, the number of segments in the proposed scheme is determined completely automatically. We show that the modelling method can be strengthened by incorporating spatial contextual information. The proposed approach speeds up the modelling process by a factor of three. The advocated methodology is successfully applied to the problem of lip pixel segmentation in face images.