Gaussian mixture model and its application on colour image segmentation

Gaussian Mixture Model (GMM) is a sophisticated way in modelling the histogram of a given signal. It is a method of fitting different scales of different Gaussians, and the point-wise sum of these Gaussians approximates the actual histogram. The parameters of GMM are obtained by expectation-maximisation (EM) algorithm. In general, a sequence of images taken from an object in a short time interval is affected by the lightings and the illuminant colours. A preprocessing procedure known as comprehensive colour image normalisation is used to make the shape of the histogram more stable, so that the variance of the EM fitting will be reduced. After fitting, the next step is to assign a cost function in classifying which pixel belongs to which Gaussian model. This procedure is called pixel clustering and the determination of the cost function is described. This paper also briefly discusses the situation when there are two classes.